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.pmed.1002803 | Women’s and girls’ experiences of menstruation in low- and middle-income countries: A systematic review and qualitative metasynthesis | Attention to women’s and girls’ menstrual needs is critical for global health and gender equality. The importance of this neglected experience has been elucidated by a growing body of qualitative research, which we systematically reviewed and synthesised.
We undertook systematic searching to identify qualitative studies of women’s and girls’ experiences of menstruation in low- and middle-income countries (LMICs). Of 6,892 citations screened, 76 studies reported in 87 citations were included. Studies captured the experiences of over 6,000 participants from 35 countries. This included 45 studies from sub-Saharan Africa (with the greatest number of studies from Kenya [n = 7], Uganda [n = 6], and Ethiopia [n = 5]), 21 from South Asia (including India [n = 12] and Nepal [n = 5]), 8 from East Asia and the Pacific, 5 from Latin America and the Caribbean, 5 from the Middle East and North Africa, and 1 study from Europe and Central Asia. Through synthesis, we identified overarching themes and their relationships to develop a directional model of menstrual experience. This model maps distal and proximal antecedents of menstrual experience through to the impacts of this experience on health and well-being. The sociocultural context, including menstrual stigma and gender norms, influenced experiences by limiting knowledge about menstruation, limiting social support, and shaping internalised and externally enforced behavioural expectations. Resource limitations underlay inadequate physical infrastructure to support menstruation, as well as an economic environment restricting access to affordable menstrual materials. Menstrual experience included multiple themes: menstrual practices, perceptions of practices and environments, confidence, shame and distress, and containment of bleeding and odour. These components of experience were interlinked and contributed to negative impacts on women’s and girls’ lives. Impacts included harms to physical and psychological health as well as education and social engagement. Our review is limited by the available studies. Study quality was varied, with 18 studies rated as high, 35 medium, and 23 low trustworthiness. Sampling and analysis tended to be untrustworthy in lower-quality studies. Studies focused on the experiences of adolescent girls were most strongly represented, and we achieved early saturation for this group. Reflecting the focus of menstrual health research globally, there was an absence of studies focused on adult women and those from certain geographical areas.
Through synthesis of extant qualitative studies of menstrual experience, we highlight consistent challenges and developed an integrated model of menstrual experience. This model hypothesises directional pathways that could be tested by future studies and may serve as a framework for program and policy development by highlighting critical antecedents and pathways through which interventions could improve women’s and girls’ health and well-being.
The review protocol registration is PROSPERO: CRD42018089581.
| A growing body of qualitative research has highlighted the importance of menstrual experiences for the health and well-being of women and girls in low- and middle-income countries (LMICs).
Qualitative research has identified an array of factors contributing to experiences but has not developed clear theory to direct intervention and evaluation.
We systematically searched and critically appraised the body of qualitative studies of menstrual experience in LMICs.
We identified overarching themes and mapped the relationships between them to develop a directional model of menstrual experience.
Women and girls reported impacts of negative menstrual experiences on physical and psychological health, education, employment, and social participation.
Both resource limitations and the sociocultural context contribute to menstrual experience.
Women’s and girls’ menstrual experiences are complex and multifaceted, but across LMICs, consistent factors contribute to experiences.
Findings advance the development of problem theory in menstrual health, and the developed model can be used as a framework for developing interventions and evaluation.
Future interventions should seek to address identified antecedents of menstrual experience, including knowledge, social support, restrictive behavioural expectations, and the physical and economic environment.
| Each day, more than 300 million women are menstruating [1]. There is increasing recognition that this natural process is experienced negatively and presents a barrier to health and gender equality in low- and middle-income contexts [2]. A growing body of qualitative research has been critical to highlighting this issue. Early studies focused on adolescent girls reported that menstruation was experienced with discomfort and fear [3–5]. Access to clean, reliable materials to absorb menses, supportive sanitation infrastructure, and biological and pragmatic information about menstruation were highlighted as core challenges [6, 7]. Studies suggested that these challenges negatively impacted school participation [4, 8, 9], health, and well-being [10, 11, 12, 13]. Fewer studies of adult women have highlighted that they too lack resources and support [14, 15], which may contribute to stress and absence from employment [16, 17].
In response to growing advocacy, programs and policies seeking to address menstrual needs have emerged rapidly. Development of these interventions has drawn largely on qualitative literature, with limited high-quality quantitative research and controlled trials available [18, 19]. However, this use of qualitative evidence has been anecdotal and situated, drawing on individual studies that may be restricted in scope and have limited generalizability beyond their context. Thus, the first objective of this review was to undertake a systematic search and synthesis of extant qualitative studies to draw out common themes, as well as appraise the coverage and quality of existing research.
The second objective of this review was to draw on the rich body of qualitative evidence to advance understanding of menstrual health. Program-orientated scoping research has often identified lists of factors important for menstrual experience, alongside lists of consequences for health and education. However, pathways between the variety of contributors and their impacts remain poorly understood [20]. Menstrual health research has tended to draw on programmatic models rather than detailed problem theory. One common model includes a Venn diagram with three circles: knowledge, menstrual products, and sanitation (e.g., [21, 22]). This may also include cultural norms and social factors as a broad concept surrounding the three core circles. These Venn models provide an intuitive high-level picture and pillars to inform programming. However, they do not provide a detailed problem theory, mapping relationships and aetiological pathways to hypothesised impacts on health and education. For example, we would expect different causal pathways contributing to reproductive tract infections than to mental health. The need for detailed problem theory becomes more salient when, in responding to the many challenges for women and girls, actors develop increasingly complex, multicomponent interventions, with many desired outcomes.
A lack of clarity around core concepts, or terminologies, in menstrual research has further complicated the development of problem theory. Early efforts were united around ‘menstrual hygiene’. This was defined as ‘women and adolescent girls using a clean menstrual management material to absorb or collect blood that can be changed in privacy as often as necessary for the duration of the menstruation period, using soap and water for washing the body as required, and having access to facilities to dispose of used menstrual management materials’ [6], providing a target for improving the effective and hygienic management of menses. However, because this definition does not include other menstrual needs highlighted by qualitative studies, more recent versions have expanded the definition, adding that ‘[women and adolescent girls] understand the basic facts linked to the menstrual cycle and how to manage it with dignity and without discomfort or fear’ [23]. Dissatisfied with the coverage and physical focus of ‘menstrual hygiene’, more recent efforts have discussed ‘menstrual health’, suggested as ‘an encompassing term that includes both menstrual hygiene management (MHM) as well as the broader systemic factors that link menstruation with health, well-being, gender, education, equity, empowerment, and rights’ [21]. This term may be useful to describe a broad field of research and practice and facilitates the consideration of menstrual disorders (e.g., dysmenorrhea, endometriosis). However, although the use of ‘menstrual health’ or expansions to ‘menstrual hygiene’ may be beneficial for more comprehensive advocacy, neither addresses the underlying lack of unified problem theory. Thus, through an interpretive approach in this review, we seek to evolve problem theory and concept definitions by providing a more nuanced model and iteratively identifying aspects of menstrual experience.
In sum, this systematic review and metasynthesis had 2 overarching objectives: (1) synthesise and appraise the extant qualitative research on women’s and girls’ menstrual experiences and (2) interpret findings across studies to develop an integrated model of menstrual experience mapping relationships between contributing factors (antecedents), menstrual experience, and the consequences of poor experiences for health and well-being (impacts).
The review protocol is registered on PROSPERO: CRD42018089581 and is reported according to PRISMA guidance (S1 PRISMA Checklist).
The search strategy was designed to capture studies reporting on women’s and girls’ experiences of menstruation. Searches were undertaken in 11 databases (Applied Social Science Index and Abstracts, Cumulative Index of Nursing and Allied Health Literature, ProQuest Dissertation and Theses, Embase, Global Health, Medline, Open Grey, Popline, PsycINFO, Sociological Abstracts, WHO Global Health Library) using a prespecified, piloted strategy reported in Table 1. Searches were completed in January 2019 with no language of publication or date restrictions applied. Comprehensive grey literature searching and hand searching were undertaken. Organisations attending to menstrual health were identified through participation in reports [7, 21], stakeholder meetings [2], and online searches. Websites (see list in S1 Text) were searched using relevant terms (e.g., ‘menstrual’, ‘menstruation’). Citations of included studies and reference lists of large menstrual health reports were searched [7, 21]. Results were exported into EPPI-Reviewer 4 (EPPI-Centre; https://eppi.ioe.ac.uk/cms/Default.aspx?tabid=2914). Two authors (JH, AS) independently screened titles and abstracts, followed by full-text screening to determine eligibility (JH).
Studies were eligible if they reported qualitative analysis of the menstrual experiences of women and girls residing in low- or middle-income countries (LMICs) as defined by the World Bank [24]. Studies that included women from LMICs now residing in high-income countries, or that combined populations from LMICs with those in high-income settings, were excluded. While these experiences also deserve increased attention, this review sought to synthesise the large set of studies situated in LMICs to inform evolving policy and practice in these regions. Studies exclusively concerning the acceptability of menstrual suppression were excluded. Studies focussed on puberty more broadly, or the use of sanitation infrastructure, were only included when they reported on experiences of menstruation. For example, studies that included lists of puberty education needs that referenced menstruation but did not report on women’s or girls’ lived menstrual experiences were not included. Similarly, studies focussed on menopause, premenstrual syndrome, or polycystic ovary syndrome were not eligible for inclusion. Studies capturing the menstrual experiences of populations with menstrual disorders (e.g., dysmenorrhea, endometriosis) were eligible. Menstruating women and girls were the target population; thus, studies were excluded if they focused exclusively on girls’ premenarche, key informants, or males. Where key informant interviews were analysed alongside women’s and girls’ experiences, studies were included, but analysis focused on the experience of the target population. Qualitative and mixed-methods studies reported in peer-reviewed or grey literature were eligible for inclusion. Studies were excluded if they did not report any qualitative analysis or results (e.g., qualitative responses were back-coded for quantitative description).
Following full-text screening, study research questions were extracted and iteratively grouped. Three groupings emerged: studies broadly focused on menstrual experiences, studies of experiences of menstruation for those with dysmenorrhea or disorders, and studies of experiences of menstrual interventions or products. Because the review aimed to provide a synthesis of menstrual experience and advance problem theory rather than explore the role of interventions, the third grouping was excluded from the present review but was retained for analyses reported elsewhere.
Included studies were appraised using the EPPI-Centre checklist [25]. This checklist assesses methodological quality across recommended domains: sampling, data collection, analysis, interpretation, and privileging of participant perspectives (see S1 Table) [26, 27]. The checklist also facilitates generation of overall trustworthiness taking into account the first 4 ratings (sampling, data collection, analysis, and support for interpretation), as well as relevance to the review (relevance to the review question, privileging and involvement of participant perspectives), rated as low, medium or high. One author (JH) appraised the quality of all studies; 10% were independently appraised by a second author (GJMT), with 100% agreement.
Reflecting the research questions, we employed a thematic synthesis approach [28], or an approach that seeks to describe themes as recurrent ‘units of meaning’ within and across studies, to synthesise insights from individual studies [29, 30]. We then used these themes to generate thematic networks to facilitate repeated comparison of the relationships between these themes. The use of thematic networks was appropriate as it includes (a) the explicit step of evidencing themes against specific statements and quotes from included studies and (b) the development of relationships between themes to form a network. The use of thematic networks reflected a lines-of-argument approach to meta-ethnography [31]; that is, drawing on the synthesis of qualitative research (meta-ethnography), we sought to go beyond merely describing a phenomenon to interpreting and understanding the social processes underlying it as part of an integrated scheme, such that no one study captures the entirety of the phenomenon (lines of argument).
Analysis was undertaken in 4 steps:
The findings of high- and medium-quality studies were thematically coded using NVivo 12 (QSR International; https://www.qsrinternational.com/nvivo/nvivo-products) by the first author. Throughout initial coding, multiple mappings of the relationships between themes were generated. Representative quotations and core sections of text corresponding to each theme were recorded for each included study.
Emergent themes were discussed among the authorship team, and multiple mappings of the relationships between themes were generated until authors reached consensus on a representation with the greatest explanatory power.
Findings of low-quality studies were assessed for their fit with the proposed themes and integrated model. Low-quality studies supported the primary analysis, and no new constructs emerged. This is akin to the ‘sensitivity analysis’ common in systematic reviews of qualitative research, in which themes are checked to see if they rely on low-quality studies alone [30].
Using Nvivo 12, all studies were recoded against the consensus themes and integrated model for validation.
The review flowchart is presented in Fig 1.
Study characteristics, quality, and relevance are presented in Table 2. Included studies captured experiences from 35 countries and over 6,000 participants. India was the country with the highest number of studies (n = 12), followed by Kenya (n = 7) and Uganda (n = 6).
Using World Bank regional groupings [24], most studies were conducted in sub-Saharan Africa (n = 45); this was followed by 21 studies undertaken in South Asia, 8 in East Asia and the Pacific, 5 in Latin America and the Caribbean, 5 in the Middle East and North Africa, and 1 study from Europe and Central Asia (Kyrgyzstan).
Focus Group Discussions (FGDs) were the most common data collection method used, particularly when working with adolescent girls. Eleven studies reported using participatory activities such as asking groups to design an ‘ideal latrine’, or rank priorities for improvements to explore experiences and preferences. A total of 55 studies included adolescent girls (school aged), with only 21 studies including exclusively university students or adult women. Most studies reported undertaking a thematic approach to analysis.
Overall study quality and relevance ratings are reported in Table 2. Classifications according to each item in the EPPI-Centre checklist, along with justifications, are displayed in S1 Table. Most studies were situated within a public health discipline and focused on cataloguing antecedents and impacts of menstruation on participants’ lives, rather than sociological analysis. Study quality was varied, with 18 studies rated as high, 35 as medium, and 23 as low trustworthiness. Lower-quality studies were characterised by poor reporting of participant selection and analysis, with many having no defined or described analytic approach. After initial analysis of high- and medium-quality studies, analysis and integration of lower-quality studies did not elicit any new themes nor changes to the integrated model.
Fig 2 presents an integrated model of menstrual experience. The figure summarises the themes and subthemes identified through analysis and the relationships between them. Studies varied in their coverage of included themes, with none reflecting the full picture gained through metasynthesis.
In identifying themes and mapping the relationships between themes as part of our analysis, differences in author orientations across the body of literature were apparent. We found that studies varied along a continuum in the extent to which they approached menstrual experiences as a result of the sociocultural context, particularly menstrual stigma, or focused on resource deficits in the environment. These orientations influenced the focus of studies, topics covered, and author interpretation of the relationships between antecedents, experience, and impacts.
Studies with a more stigma-centric orientation conceptualised menstrual experiences and negative outcomes as resulting from the stigmatised and gendered nature of menstruation. These studies were often those using individual interviews or engaging adult women to provide reflexive insights into their experiences. Themes centred around the pressure to suppress evidence of menses and control the body. Studies focused on stigma, knowledge, and social support as antecedents of experience. Where resource limitations were discussed in these studies, they were considered problematic because they failed to support socially proscribed discretion. Menstrual experiences were characterised by shame and a lack of confidence to engage socially when menstruating. Stigma-centric models of menstruation rarely elaborated on physical health consequences; rather, they detailed psychological and educational concerns and consequences for empowerment. Others did not address the impacts of menstruation on such outcomes, exploring only the experience itself.
Other studies were orientated towards poverty and resource limitations as causes of menstrual experiences and impacts. These studies viewed the resources available to women and girls as inadequate to meet their basic biological needs. This included an emphasis on menstrual hygiene practices such as the use of materials poorly suited to absorbing menses. Knowledge and cultural practices were important in resource-centric studies but were more often linked with a failure to support effective management practices. Menstrual hygiene, the hygienic and effective management of menstrual blood, lay at the centre of experience in these studies, which frequently noted concerns around infection and consequences for school attendance.
Our final integrated model synthesised studies along this continuum. As displayed in Fig 2, the model views resource limitations and the sociocultural context, including menstrual stigma, as determinants of menstrual experience and negative impacts. Themes are described from this perspective.
Table 3 summarises the individual studies contributing to each theme according to their level of trustworthiness.
Resource deficits contributed to economic and physical environments that limited women’s choices for menstrual practices and shaped their experiences of menstruation. Quotations are presented in Box 2 and contributing citations in Table 3.
Illustrative quotations of menstrual practices and perceptions are provided in S2 Table; Box 3 shows themes of containment, confidence, and shame.
Illustrative quotations are presented in Box 4, and a summary of contributing citations is displayed in Table 3.
This systematic review had 2 overarching objectives: (1) to synthesise extant qualitative studies of women’s and girls’ menstrual experience in LMICs and (2) to integrate findings across studies to develop a directional model of menstrual experience to advance problem theory in menstrual health research. Despite different settings and populations, the narratives and lived experiences that emerged reflected consistent themes, with manifestations that differed by context. Mapping relationships between themes highlighted the multidimensional nature of menstrual experience. The integrated model produced illustrates pathways through which distal and proximal antecedents influence menstrual experience and ultimately result in impacts on physical and psychological health, education, employment, and social participation.
Metasynthesis identified multiple components of menstrual experience. Menstrual practices, such as the type of material used, represented only one of these and directly contributed to physical health outcomes alone. Women’s and girls’ perceptions of their menstrual practices and environments emerged as a critical independent theme connecting practices to other impacts such as school absenteeism. In addition, confidence to manage menstrual bleeding and to undertake other activities during menses, as well as experiences of shame and fear, were salient contributors to psychological, social, and educational outcomes. Distal antecedents—the sociocultural context and resource limitations—underlay more proximal factors. Experiences captured in qualitative studies revealed knowledge of the biological, reproductive, and practical aspects of menstruation as a source of confidence, shame, and perceived acceptability of practices and environments. Social support enhanced or diminished experiences. This was closely connected to behavioural expectations that were placed upon women and girls and were enforced through discipline and cultural practices and ideals of cleanliness or femininity. Such expectations and broader menstrual stigma were internalised, influencing experience and self-concept. Poorly supportive physical infrastructure, such as a lack of water and sanitation facilities, made it difficult for women and girls to undertake their preferred menstrual practices in privacy and safety; in addition, the economic environment restricted access to preferred materials, soap, and pain relief.
Included studies represented a broad range of contexts, although the majority attended to the experiences of school-aged girls. Amongst this group, we achieved early saturation of impacts and antecedents. Qualitative studies identified different relative contributions of antecedents such as the physical environment, knowledge, or behavioural expectations (particularly explicit restrictions) for girls in different settings, which have implications for programming in those areas. However, researchers should reconsider burdening future populations to identify these same factors and should focus on more detailed research questions. Greater depth in understanding the transmission of social norms and behavioural expectations, as well as confidence during menstruation, requires more attention. Experiences of adult women, particularly in the workplace, were under-researched. The experience of pain and menstrual disorders was poorly integrated in included studies and did not reach saturation. Most studies relied on FGDs. This approach was often selected to source priorities for intervention but meant that studies more often captured shared experiences of the school environment, rather than deep understanding of individual experiences that may yield greater sociological or psychological insights.
The model presented differs from previous nonsystematic summaries that have generated lists of considerations to improve menstrual experience but have not specified directional relationships. Systematic searching identified a large body of literature, facilitating clarity in the identified themes. Metasynthesis remains an evolving methodology for the review of qualitative literature [30, 110–112]. This work draws on current best practice guidance [26, 27, 29, 30, 112]. As in all reviews, the findings are limited by available studies. The majority of included studies came from sub-Saharan Africa, and findings may be more representative of this context. We identified no studies from China, and studies from the Middle East and North African countries were more frequently undertaken in urban, more highly educated populations, though this reflects the global focus on menstrual hygiene research in specific geographical areas. Despite extensive searching, eligible studies may have been missed. A small number of studies in languages other than English were included; however, searching was only undertaken in English and may not have identified all available publications. Study quality was appraised using a tool designed for qualitative studies. This may have resulted in poorer scoring of mixed-methods studies that were considered only based on their qualitative components. Finally, as with any systematic review of qualitative research, a different team of researchers may have generated different insights. However, our focus on consensus themes and auditable methods lends credibility to our results as presented.
This review included a large volume of studies. This presented the opportunity to draw themes and an integrated model across a broad range of contexts and populations. Region-specific reviews, or reviews with more targeted research questions, may have space to provide deeper insights into specific questions or region-specific challenges.
Findings of our synthesis have implications for research and practice. To date, developed interventions have sought to improve menstrual experience by investing in education (to improve knowledge) and menstrual products (to improve the economic environment), two of the core pillars of existing menstrual health frameworks [19, 113]. Social support, the physical environment, cultural restrictions, and the perception of menses as dirty and needing to be concealed were all highly salient in included studies, yet these risk factors have received limited attention in interventions. Furthermore, these factors likely represent important covariates that should be considered when not the focus of interventions. The integrated model presented here offers a more nuanced framework to inform theories of change in program development and to assist evaluation strategies.
Awareness of multiple, interconnected constructs also aids the identification of unanticipated harms. Models focused exclusively on menstrual practices have assessed only physical harms such as infections. Our integrated model suggests that changes to girls’ perceptions of their practices may cause distress as they struggle to adopt new practices advised in education interventions in an unsupportive physical environment. Interventions that focus on products may maintain menstrual stigma and reinforce behavioural expectations that concealing menses is paramount. Such interventions may have positive impacts on containment in the short term but risk greater harms if access to more reliable menstrual materials is unsustained.
Results have implications for the design of quantitative studies. To date, these have evaluated links between menstrual practices and impacts (e.g., sanitary pad use and school attendance) [18, 114]. Our integrated model suggests that this is an indirect relationship and is unlikely to capture the true effects of menstrual experience on psychological, educational, and social outcomes without attention to the other components of experience that mediate this relationship. This integrated model should inform hypotheses for more detailed quantitative studies to test the pathways emerging from qualitative research and subsequently inform intervention research. Furthermore, it may serve to inform measure development for identified components of menstrual experience, antecedents, and impacts.
The definition ‘menstrual hygiene’ outlined in the Introduction fit poorly with the integrated model resulting from this metasynthesis. This definition provides an unclear list of menstrual practices, incorporates one mention of perceptions of the menstrual environment but not practices, and refers to physical facilities for disposal only. Recent expansions on this definition to include menstrual knowledge conflates menstrual experience with antecedents [115]. As presently defined, ‘menstrual hygiene’ may reflect an incomplete set of menstrual rights or needs, or it could be amended to capture menstrual practices as have emerged here. Findings of this review suggest that terminology and construct definition in menstrual health research may need to be expanded to recognise the many components of menstrual experience and contributing factors.
In sum, extant qualitative studies have identified consistent negative impacts for health and social participation resulting from poor menstrual experience. This large body of qualitative evidence emphasises the need for practitioners and policymakers to attend to menstruation to improve the physical and psychological health, educational attainment, and social participation of women and girls. The integrated model presented advances the development of problem theory in menstrual health research and highlights important factors to consider in future research and practice. Menstrual experience is characterised not only as the hygiene practices undertaken to manage menstrual bleeding but by women’s and girls’ perceptions of these practices, their confidence to manage menses and engage in other activities while menstruating, and their experience of shame and containment. Through synthesis, we elucidate antecedent pathways and highlight the multiple components of menstrual experience that must be considered for effective interventions and comprehensive quantitative evaluation.
|
10.1371/journal.pcbi.1003615 | Axonal Noise as a Source of Synaptic Variability | Post-synaptic potential (PSP) variability is typically attributed to mechanisms inside synapses, yet recent advances in experimental methods and biophysical understanding have led us to reconsider the role of axons as highly reliable transmission channels. We show that in many thin axons of our brain, the action potential (AP) waveform and thus the Ca++ signal controlling vesicle release at synapses will be significantly affected by the inherent variability of ion channel gating. We investigate how and to what extent fluctuations in the AP waveform explain observed PSP variability. Using both biophysical theory and stochastic simulations of central and peripheral nervous system axons from vertebrates and invertebrates, we show that channel noise in thin axons (<1 µm diameter) causes random fluctuations in AP waveforms. AP height and width, both experimentally characterised parameters of post-synaptic response amplitude, vary e.g. by up to 20 mV and 0.5 ms while a single AP propagates in C-fibre axons. We show how AP height and width variabilities increase with a ¾ power-law as diameter decreases and translate these fluctuations into post-synaptic response variability using biophysical data and models of synaptic transmission. We find for example that for mammalian unmyelinated axons with 0.2 µm diameter (matching cerebellar parallel fibres) axonal noise alone can explain half of the PSP variability in cerebellar synapses. We conclude that axonal variability may have considerable impact on synaptic response variability. Thus, in many experimental frameworks investigating synaptic transmission through paired-cell recordings or extracellular stimulation of presynaptic neurons, causes of variability may have been confounded. We thereby show how bottom-up aggregation of molecular noise sources contributes to our understanding of variability observed at higher levels of biological organisation.
| The fundamental signal of the nervous system is the action potential: an electrical spike propagated along neurons and transmitted between them via synapses. Once triggered, action potentials are generally assumed to be robust to noise, and the variability observed at all levels of the nervous system is primarily attributed to synapses. However, this view is based on data from classically studied axons, which are very large compared to the average diameter of axons in the mammalian nervous system, and even larger when compared to the thinnest axons. As the effects of thermodynamic noise affecting the proteins responsible for the initiation and propagation of action potentials are much bigger in thin axons, the assumption does not necessarily hold for very thin axons. We show that the action potentials waveform in thin axons is subject to random variability. Fluctuations in this waveform result in fluctuations in synaptic ionic currents, and account for a significant portion of the variability observed at the synapse.
| The great majority of axons use action potentials (APs) to transmit information reliably to synapses. Once the AP arrives at the synapse the characteristics of its waveform are fundamental in determining the strength and reliability of information transmission, as was extensively shown in the central and peripheral nervous system of both vertebrates and invertebrates [1]–[14]. Although the nervous system exhibits stochastic variability (noise) at all levels (see [15] for a review), it is generally assumed that little random variability affects the AP waveform as it travels from the soma along the axon to the synapse. However, recent understanding of biophysics and experimental methods prompt us to reconsider this common assumption.
The AP is mediated by voltage-gated ion channels, which control the flow of ionic currents through the membrane. Thermodynamic fluctuations in voltage-gated ion channels result in probabilistic gating, producing random electrical currents called channel noise [16]. In thin axons, the behaviour of individual ion channels can have significant effects on the membrane potential dynamics due to the higher input resistance of those axons [17]–[19]. Fewer channels sustain AP conduction and fluctuations in individual ion channels have a larger impact on the membrane potential in thinner axons. Faisal et al. [20] have shown that channel noise sets a lower limit to reliable axonal communication at 0.08–0.1 µm diameter, a general limit matched by anatomical data across species. Above this limit, in axons of 0.1–0.5 µm diameter, channel noise causes variability in the rising phase of the AP and the resting input resistance of axons. Therefore APs are jittered, shifted, added and deleted in a history-dependent way along the axon [18]. Thus, noise in axons affects the timing of APs and therefore reduces the information capacity of the neural code. Here, we are going to investigate how noise in axons affects the waveform of APs, and produces random variability in the responses of synapses, with implications for information transmission and learning.
Attempts at investigating the impact of axonal noise on the synapse have so far been limited to rather large diameter axons (≥1 µm diameter) [21], [22]. However, many unmyelinated axons are very thin (0.1–0.3 µm diameter [23]). Examples include cerebellar parallel fibres (average diameter 0.2 µm [24]), C-fibres implicated in sensory and pain transmission (diameter range 0.1–0.2 µm [25]) and cortical pyramidal cell axon collaterals (average diameter 0.3 µm [26], making up most of the local cortical connectivity [26]). The variability of the AP waveform in all these axons is unknown. Basic biophysical considerations suggest that axonal noise sources are bound to introduce fluctuations [20], [27] in the shape of the travelling AP waveform in thin axons with immediate consequences for synaptic transmission and reliability [28].
Intracellular recordings from such thin axons are difficult to obtain. Extracellular stimulation offers only limited signal resolution and stimulus control, and tiny intracellular volumes limit the application of imaging methods to quantify AP waveforms accurately. This motivated the study presented here which uses biophysically detailed stochastic simulations of travelling APs in thin axons and basic biophysical theory. Our goal is to investigate the mechanisms behind the observed synaptic variability; specifically how much variability can be explained by channel noise in axons. We quantify waveform fluctuations of single propagating APs in terms of standard synaptic efficacy measures, namely AP width and height. We explain how channel noise causes AP waveform fluctuations and show that these fluctuations scale with axon diameter according to an inverse power law, i.e. the finer an axon the bigger the impact. We then investigate the AP waveform fluctuations for propagating spike trains and predict the post-synaptic response variability axonal noise would cause in two ways: 1. by using models of synaptic dynamics and vesicle release and 2. by using direct experimental data linking AP waveform to post-synaptic response. Thus, we will be able to estimate the influence of axonal channel noise on synaptic variability.
We find that single APs propagating in central and peripheral nervous systems (CNS and PNS), mammalian and invertebrate axons of up to 1 µm diameter display large random variability in their waveform as they propagate. We visualize this by measuring the AP waveform (membrane potential versus time) at various positions along the axon (Figure 1.A,B) and then align the waveforms at the instant of half-peak crossing (Figure 1.D). As a control, we simulated a deterministic axon, i.e. one that had the same set of biophysical parameters and received the same stimuli but where we modelled the ion channels using deterministic kinetics instead of the corresponding stochastic kinetics [19], [29]. APs in all our deterministic simulations, starting from the same initial condition and receiving the same trigger input, exhibit no waveform variability across repeated trials. Since the stochastic kinetics of ion channels are the only source of variability given that all other parameters and stimuli are controlled by our simulation, the variability of the travelling AP waveform observed must be due to channel noise and thus entirely random in nature. Crucially, the AP waveform is not only variable across repeated trials with identical stimulus, but also varies as the same AP propagates along the axon.
Comparing the variability of the AP waveform in axons with identical biophysical parameters and ion channels but varying axon diameter from 0.1 µm to 1 µm, shows that the waveform fluctuations become larger as the axon becomes thinner. This is true for both models of squid giant axons, rat hippocampal interneuron and C-fibre axons. The general structure of the variability profile remains preserved across diameters. The width of the propagating AP varies as it travels down a thin axon in the order of a tenth of a millisecond (Figure 2.A,C,E,G). Similarly, AP height varies in the order of 1 to 10 millivolts (Figure 2.B,D,F,H). The variability is more pronounced the thinner the axon is (Figure 3.A,B). The variations between proximal and distal AP shape are, as expected uncorrelated (R2<<0.2 across all diameters and axons for both AP heights and AP widths). This implies that both AP width and height become decorrelated with themselves (autocorrelation decreases) and between each other (cross-correlation decreases) the further the AP propagates down an axon (Supplementary Figure S1).
To measure how channel noise affects the propagating waveform, one has to track the relevant quantities at corresponding points of the moving AP. To this end, the time series recorded at closely spaced axonal positions (here, corresponding to a cylindrical membrane compartment of the axon model) are superimposed after having been aligned at the instant when the membrane potential crosses its half AP peak value. Thus, the individual quantities and their variability at corresponding points of the travelling AP are displayed at corresponding points (Figure 6).
The variability of the waveform has a characteristic structure that is conserved across different axons types and diameters as it is caused by the basic mechanism of the AP itself. The first maximum in waveform variability is reached in the late rising phase of the AP (between half peak and peak depolarisation, see Figure 4.A,B). The location of this peak is not an artefact of our aligning of APs (at 50% AP height, c.f. alignment at 20% AP height in Supplementary Figure S2). This first peak is due to fluctuations in the number of opening Na+ channels and Na+ current (red curves in Figure 6.B,C), as the first peak of Na+ current variability is reached at half-peak membrane depolarisation (Figure 6.C, shortly after 0 ms). The variability of the depolarising Na+ current accounts for the variations in AP height because the number of Na+ open channels and their inactivation prior to reaching Na+ reversal potential (the upper limit to AP peak) determine how much driving current is depolarising the membrane capacitance. K+ channels begin to open later, and thus Na+ channels carry most of the net membrane current in this initial phase of the AP and are responsible for the initial variability (Standard deviation (SD) profile of Na+ current, red curve, and K+ current, blue curve, with net membrane current, green curve, in Figure 6.E).
The second, broad peak in waveform variability is reached in the repolarizing phase (Figure 6.A and Figure 4.A,B beginning at 1 ms and increasing up to 2.5 ms). As the rate of repolarization (here, <50 mV/ms) is much slower than that of depolarization (here, >200 mV/ms), variability in the height of the AP waveform translates into much larger changes of AP width. Note, AP width is measured between the up and down crossings of any given membrane potential level, here chosen to be half-peak depolarization. Thus, AP width variability is mainly generated in the repolarizing phase of the AP and caused by a long period of large fluctuations in net membrane current (Figure 6.D, between 0.75 and 2.25 ms). Variability is generated initially by K+ current noise and then by Na+ current noise (Na+, red, and K+, blue, in Figure 6.B,C). After K+ channels begin to open in the early repolarizing phase, K+ current fluctuations peak as K+ channel opening probabilities increase and the variance of the number of open channels becomes larger. The increase in variance can be understood, if one considers that a population of ion channels with open probability follows a binomial distribution for the number of open channels. The variance in the number of open channels is given by and thus has a maximum as the open probability approaches 0.5 from all channels closed () or all channels open (). By the time the maximum K+ channel open probability is reached (which is not necessarily for many voltage-gated ion channels), electro-motive forces are lower than near AP peak and, membrane potential fluctuations due to K+ currents have consequently lower amplitudes. An equally large and broad maximum in the fluctuations is due to Na+ channel inactivation in the late repolarizing phase for analogous reasons, following a similar binomial argument, when Na+ electro-motive forces are large.
Thus, AP height variability (Figure 5.A,C,E,G) is mainly caused by the fluctuating number of open and inactivating Na+ channels during the upstroke of the AP. AP width variability (Figure 5.B,D,F,H) is predominantly caused by the noisy repolarizing phase of the AP, where both K+ and Na+ channels contribute to large fluctuations in the rate of repolarization. Having described how channel noise affects a single AP's waveform, the question arises whether AP waveforms are more variable in spike trains, as APs may influence each other.
Using a naturalistic white noise current stimulus protocol [32] (1 kHz cut-off frequency, see methods), we elicited spike trains for a period of 10 minutes in a 0.2 µm diameter axon (average cerebellar parallel fibre diameter) using the rat hippocampal interneuron model. Note that interspike intervals and AP triggering currents therefore varied as successive APs were triggered (f = 40.8 Hz±42.8 Hz, mean ± SD). At the axon's distal end (measured at approx. 95% of the axon's total length to exclude boundary effects) waveforms showed considerable variation in the AP shape (Figure 7.A).
Plotting pairs of an AP's width measured at the mid and distal position revealed an uncorrelated structure (correlation coefficients 0.04 and 0.03 for AP width and height), as in the case of single APs. AP widths measured at half-peak had a coefficient of variation (CV = SD/average) of 6% (0.7 ms±0.04 ms) and AP amplitude (resting potential to peak) had a CV of 3% (93.7 mV±2.5 mV). AP waveform variability in spike trains was larger than in the case of individual spikes propagating. The standard deviation of change in AP height after propagating for 1 mm in the axon was 3.5 mV (0.05 ms for the width, N = 2000), compared to 1.6 mV (0.04 ms for the width, N = 250) for the single spike protocol. The profile of waveform variability (Figure 7.A) peaks close to AP threshold and, at a higher level, in the late repolarising phase. Random waveform variability has matching profiles in the spike train and the single AP protocol as it is caused in both cases by the AP mechanism itself.
This illustrates that AP waveform variability is a constantly acting random process, occurring independent of AP initiation or stimulus. Stochastic waveform variability is non-existent in identical simulations where we replace stochastic ion channel models by the equivalent deterministic Hodgkin-Huxley type conductance models. Thus, all axonal variability observed here must result from the effects of the only source of noise modelled – channel noise in Na+ and K+ channels. We have previously shown that the memory of voltage-gated ion channels causes an increased effect on membrane potential noise, affecting the speed of propagation [18]. The same mechanism is also acting on the waveform.
We have thus quantified the impact of axonal channel noise on AP waveform variability and explained the biophysical mechanisms acting in thin axons. This previously overlooked effect will only become relevant at the neural circuit and behavioural level if it can influence synaptic transmission. Therefore, we modelled next the synaptic transmission process from arrival of the AP to the post-synaptic response.
Synaptic transmission follows a general sequence of events leading to a post-synaptic response. An AP propagates down the axon and causes the opening of voltage-gated Ca++ channels resulting in the influx of Ca++ at the pre-synaptic terminal. Ca++-sensitive proteins trigger the fusion of vesicles, which release neurotransmitters into the synaptic cleft. These transmitters diffuse and trigger the opening of ion channels in the post-synaptic cell, producing a voltage response. Thus, AP waveform variability could perturb post-synaptic responses [33].
We estimate the synaptic impact of waveform variability for spike trains propagating down a 0.2 µm diameter axon using two distinct approaches: first, we model the individual stages of synaptic transmission in a synapse driven by our thin axons using biophysical models. Second, we use experimental data relating AP width and height to post-synaptic response amplitude to estimate directly how the variability of the AP would transform into response variability.
Axons are often thought of as fast, reliable transmission channels for electrical impulses. This is mainly because our understanding of axons and action potentials comes from studies of large axons, where noise sources have little impact. However in the many thin (<1 µm diameter) axons in our body, e.g. the dense wiring of cortex or the hundreds of kilometres of C-fibres in our PNS, axonal noise will have significant impact as dictated by basic biophysics [15(review)], [18],[20]. This theoretical work and recent advances in experimental methods have prompted us to reconsider these assumptions [27], [39], [40], and linked to the potential role of the axon as an information processing unit in its own right (reviewed in [41], [42]). Axonal information processing is closely related to the question of how APs are translated at the synapse. Do synapses consider incoming APs as unitary events, or do they use information contained in the waveform (regardless of its origin) to modulate the release of neurotransmitters? Can different synapses sitting on the same axon transmit information differentially? Does the size of axons influence the variability of downstream post synaptic potentials (PSPs)? Understanding these questions has important consequences for the computational capacity of neural circuits, the rate of information transmission between neurons and thus the metabolic efficiency of neurons [43]. We show here that the answer to this question is likely to depend on the diameter, i.e. the anatomy, of axons and is crucial to understand the design constraints of densely wired neural circuits.
The AP that drives the synapse has to travel along an axon, yet the impact of axonal noise sources on the AP waveform in thin axons was, complete propagation failures set aside [44], not considered in previous studies. Thus, synaptic response reliability and variability [45]–[48] have been in general attributed to mechanisms inside the synapse alone [44]. The results presented here show that in thin unmyelinated axons below 1 µm diameter, commonly found in the CNS and PNS, the travelling waveform of an AP undergoes considerable random variability. This random variability is caused by axonal Na+ and K+ channel noise, which continuously acts during propagation and thus accumulates with distance [18]. The variability of AP width and amplitude, key parameters linked to synaptic efficacy, dramatically increased (the CV increasing by a factor of approx. 4, see Figure 5) as diameter decreased from 1 µm to 0.2 µm. We predict this change by deriving a scaling relationship which is the direct result of the geometry and general biophysics of axons and thus independent of specific channel kinetics or other biophysical parameters [20]. Invariably, channel noise is bound to increase as diameter decreases to the point that it affects the waveform of the AP. Therefore, we can observe the effects of this variability in CNS and PNS axons, in both vertebrates and invertebrates.
The range of the waveform fluctuations is about 4 to 6 times the SD, thus we found that AP widths vary by 0.1–1 ms in axons between 0.2 and 1 µm diameter. AP width fluctuations result mainly from K+ channel noise and inactivating Na+ channels during the repolarising phase of the AP. While Na+ channel noise principally effects AP propagation speed and thus spike timing reliability [18], K+ channel noise has more impact on waveform variability (Although variability in Na+ and K+ channels partially compensate each other [20]). This fits well with genetic knock-out studies where one type of K+ channels was removed from the central nervous system, and which showed increased temporal response jitter [49].
Activity-dependent modulation mechanisms specific to the pre-synaptic terminal are well-known and provide neurons with means for positive or negative feedback regulation of pre-synaptic Ca++ influx through regulation of the AP width at the synapse [50]–[53]. One example of such modulatory mechanism, the broadening of APs during spike trains due to slow deactivation of A-type K+ channels in mossy fibres has been observed at the level of the synapse [54], and postulated in the axon [41]. This mechanism can be disrupted by random opening of Na+ channels in the repolarising phase (which broadens the AP) or random opening of K+ channels (which shortens the AP), independently of the spiking history.
In general, the observable variability in synaptic responses could be due to two sources: (1) noise and/or (2) very complex mechanisms that appear random. We can distinguish to which extent these two sources of variability are present at the cellular level, by aggregating the effect of random variability generated by identified molecular stochastic processes (such as thermodynamic fluctuations in molecular conformations, reviewed in [15]).
Here, we considered axonal noise as a source of synaptic variability due to channel noise in axons. We modelled a Calyx-of-Held synapse and used data on the Cerebellar Granule-to-Purkinje synapse to estimate the effects of AP waveform noise on synaptic responses in the absence of detailed models for small synapses. Quantitative measurements and models of the mechanistic level of synaptic transmission are limited in small synapses by the technical difficulties to record from thin axon terminals (<1 µm diameter [27]) and the need to look at very short range connections (<500 µm). Therefore, we ignored pre- and post-synaptic activity dependent effects – which may reduce the effects of waveform variability – and used simplified synaptic transmission models.
Care has to be taken when extrapolating results from these synapses to small CNS synapses [55], [56], and extrapolating from any type of synapse to another – even synapses from the same parent axon – may be difficult when details are considered [57]. Bearing that in mind, individual active zones in the Calyx are known to be ultra-structurally similar to those found in small, bouton-like CNS synapses [58]–[60] and the Calyx's functional organization corresponds to a parallel arrangement of several hundred conventional active zones in a single – bouton-like – terminal [61].
Mapping our AP waveform variability data for parallel-fibre like axons onto the empirical relationship between AP width and EPSC amplitude [11] showed a CV of approx. 30% for EPSC amplitude. Other synapses also display this common power-law relationship between AP width and synaptic response, suggesting that axonally induced random variability of the waveform would scale accordingly [5], [6], [11]. The detailed allosteric model of vesicle release rate for Calyx-type synaptic transmission produced comparable amplification of the AP waveform noise (CV increased from 6% to 25%).
Empirical synaptic response CV is typically between 20 and 60%. In all cases modelled here the extrapolated post-synaptic variability is considerable (CV 10 to 30%) and suggests that the observed synaptic variability could be partially explained by axonal noise. Axonal variability will show more impact in synapses placed 1 mm and more down the axon; yet, due to the technical difficulties of finding cell pairs at these distances, their variability is little studied.
The Hodgkin-Huxley axon model and related deterministic axon models allow information about the stimulus to be retained in the AP waveform, e.g. stimulus strength is correlated with AP height [62]. It has been shown in vitro that APs triggered and measured at the soma of the same cell can indeed encode information about the stimulus [63], [64]. Changes in the width of APs, whether due to a depolarized soma [65] or application of glutamate [27], have been shown to influence post-synaptic potentials (PSPs). Moradmand et al. [31] studied the deterministic transformation of propagating AP waveforms in a paired-pulse framework and showed that the second AP waveform in the pair becomes increasingly stereotyped due to refractory interaction with the first AP. Thus, even in bursts, only the first spike would be the likely candidate to carry stimulus information in the waveform over long distances. Kole et al. [40] have shown that changes in the waveform of APs operated at the AIS are conserved along the axonal arbour for relatively large diameter axons. Here, we show that for both single APs and spike trains, channel noise decorrelates the waveform (within the limits of the AP's regenerative dynamics) as the AP propagates even over short distances of less than 0.5 mm. Thus, any en-passant synapses along the path of an AP will be driven by randomly differing waveforms and produce different responses (even if the synapses were identical [66]).
Synapses innervated by thin axons may have developed mechanisms to circumvent the problem of axonal variability. A simple solution would be to treat an incoming noisy AP waveform as a unitary event. Taking this view, small synapses on thin axons should treat APs as unitary signals and adjust their transduction mechanisms accordingly to be robust to axonal noise effects. This is, in addition to spontaneous APs, another way in which noise constraints affect neural coding [67].
Both axonal and synaptic noise will affect post-synaptic response amplitude variability and vesicle release probability. The above results suggest that the amount of random variability depends on axonal and synaptic morphology. Impedance matching and volume conservation in densely packed CNS tissues suggest that the diameter of axons should be closely related to the size of the synapse (If the synapse was much larger, the impedence mismatch would cause a drop in the membrane potential, preventing incoming APs from triggering vesicle release.) Although there exists little data relating axon length and diameter to synaptic morphology and function, one can see that the input resistance of synapses is proportional to their surface area. Similarly, the voltage-sensitivity to an incoming AP increases with the inverse of the squared synaptic diameter. Thus, a smaller synaptic geometry supra-linearly amplifies variations in Ca++ current and influx, and due to the smaller volume, variations in Ca++ concentration as well. Thus, if the properties and kinetics of signal transduction were invariant to synaptic size, an increase in synaptic response variability due to noise in pre-synaptic signal transduction would be expected for smaller synapses. Variability in the waveform of action potentials is known to also affect synaptic latency [68]. However, a careful investigation of this phenomenon requires stochastic simulations of both Ca++ channels and the 5-state vesicle release model.
Axons play an important active role for information processing that may be comparable to that of dendritic computation [41]. However, axonal variability has traditionally not been considered as a source of neuronal variability [69] because the AP mechanism was considered highly reliable by extrapolating from classic studies in large (3 orders of magnitude larger diameters) fibres such as squid giant axons [70]. Yet, in densely connected central neural circuits the APs become sensitive to channel noise [18]. The effects of channel noise will inescapably increase non-linearly as diameter decreases due to the very nature of the AP mechanism [20]. Axonal channel noise will affect the reliability (<0.1 µm diameter) and cause considerable variability to both timing (<0.5 µm diameter) [18] and, as shown here, the shape of the AP in axons below 1 µm diameter. Thin unmyelinated axons typically innervate large numbers of small CNS synapses [71] and are associated with, and required for, the high level and density of circuit miniaturisation encountered in the cortex and the cerebellum [20], [72].
In sensory and motor nervous systems, reliability is typically achieved by averaging over many release sites and high release rates. The corresponding large synapses are associated with large axons. However, in the cerebral cortex, hippocampus and cerebellum the dense connectivity within a restricted space limits the diameter of axons, the number of redundant axonal connections and the size of the synaptic contact areas. This makes synaptic transmission prone to the effects of axonal channel noise in thin axons innervating small synapses. The results presented here prompt careful experimental consideration, because paired-cell measurements and optical methods do not offer the control and resolution necessary to determine the source of PSP variability. More generally, we show how molecular noise sources can explain observed variability at higher levels of biological organisation.
Simulations were based on biophysical data and reproduced physiological data such as the amplitude and width of APs (but see [73], for possible shortcomings and other models). Computations were carried out using the Modigliani stochastic simulator [74], available from http://www.modigliani.co.uk, on a Linux PC using an Intel core i7 processor with the binomial algorithm [19]. Simulations were carried out using Markov models of squid giant axon channels (Na+ channel expressed by gene GFLN1, K+ channel expressed by gene SqKv1.1) as several independently constructed kinetic models exist for these channels. We used the original models given by Hodgkin 18 and Huxley [75] as well as a more recent model with delayed opening [76] with little difference in results. These ion channel models captured the corresponding ion channel kinetics from patch-clamp experiments. To account for differences between the squid giant axon and mammalian axon [77], [78], we confirmed the results using models of rat hippocampal interneurons with Markov models of rodent ion channels [79] shown to have little overlap between Na+ and K+ currents [78], and a model of rodent C-fibre axons (Nav1.8) [80]. We chose to simulate all model axons at the temperature at which their channel kinetics were experimentally recorded. The parameters for each model are summarised in Table 2.
We used two sets of simulation protocols. In the first protocol, 0.1 µm, 0.2 µm, 0.3 µm, 0.5 µm diameter (1 cm long) and 1 µm diameter (2 cm long) axons were stimulated in a single spike per trial framework (N = 250 trials per diameter) to allow for fast parameter exploration. In the second protocol, we simulated 10 minutes long spike trains. To this end, a 0.2 µm diameter (2 mm long) axon was stimulated with a zero-mean white noise current (SD = 0.01 nA, 1 kHz corner frequency) injected at the proximal end. Membrane properties were set to Ra = 70 Ωcm, Rm = 20000 Ωcm2, typical for cortical cells [81]. All axons had a resting potential of −65 mV. After visual inspection of the data, we used a threshold discriminator detecting AP height and aligned their waveforms at the rising half-peak potential crossing time. We measured voltage-traces of the AP waveforms at regular intervals (typical distance 10% of total axon length, see Figure 1.A) between 5% and 95% of the axon's length (0% being the axon's proximal end) to avoid measuring stimulus artefacts or boundary effects and to measure the evolution of the AP shape along the axon. The height and width of APs were defined according to Figure 1.C.
We estimated the impact of action potential waveform variability on synaptic transmission using two approaches. First, synaptic transmission was modelled using data and deterministic models from the Calyx of Held synapse (reviewed in [34]). We drive the Calyx- of-Held synapse with noisy spike train waveforms directly (voltage clamping the synapse) to circumvent any potential issues of impedance mismatch. We then compute Ca++ currents evoked by the AP waveform by integrating the dynamics of a Hodgkin-Huxley type conductance-based Ca++ channel model of Calyx synapses [82]. We describe the Ca++ channel behaviour using a conductance-based Hodgkin-Huxley type model with two identical gating particles (denoted m) with channel opening probability . The corresponding rate functions are and , with dynamics [83]. Here we modelled the voltage-gated Ca++ channel deterministically, and calculated the waveform of the incoming Ca++ current using a reversal potential of . This Ca++ current model simplifies the heterogeneity in both the biophysical and the pharmacological properties of Ca++ currents found in Calyx-type synapses [84], [85], but was shown to capture sufficient detail for quantitative modelling of synaptic transmission [82].
The transient encountered by vesicles in the proximity of Ca++ channels is shown to follow a time course similar to that of the Ca++ current [86], [87], i.e. the Ca++ concentration rapidly declines due to effects such as buffering [88]. We use the dynamics of Ca++ channels and modify them slightly so that the rise time is conserved, but the width at half-height becomes approx. 100 µs longer [86]. We then scaled the resulting waveform to have a peak of approx. 12 µm. This value was chosen based on an approximate reading of figure 6.B in [89]. The resting was 50 nm. Finally, we modelled the impact of on transmitter release using an allosteric “5 state” model [90], which allows us to derive the instantaneous vesicle release rate (Figure 9).
Synaptic transmission dependence on the AP waveform was also modelled using experimental data for the much smaller rodent cerebellar Granule cell-to-Purkinje cell synapse. In this synapse the width of the pre-synaptic AP waveform, pre-synaptic Ca++ entry and the resulting post-synaptic currents were directly measured [11]. To obtain an estimate of how AP waveform variability would affect this synapse, we passed the simulated APs' widths through the experimentally characterised relationship between postsynaptic response and AP width. We then computed the variability of the post-synaptic response over all APs.
|
10.1371/journal.pntd.0005029 | Colorimetric Detection of Plasmodium vivax in Urine Using MSP10 Oligonucleotides and Gold Nanoparticles | Plasmodium vivax is the most prevalent cause of human malaria in the world and can lead to severe disease with high potential for relapse. Its genetic and geographic diversities make it challenging to control. P. vivax is understudied and to achieve control of malaria in endemic areas, a rapid, accurate, and simple diagnostic tool is necessary. In this pilot study, we found that a colorimetric system using AuNPs and MSP10 DNA detection in urine can provide fast, easy, and inexpensive identification of P. vivax. The test exhibited promising sensitivity (84%), high specificity (97%), and only mild cross-reactivity with P. falciparum (21%). It is simple to use, with a visible color change that negates the need for a spectrometer, making it suitable for use in austere conditions. Using urine eliminates the need for finger-prick, increasing both the safety profile and patient acceptance of this model.
| To control malaria, there is an urgent need for applying innovative diagnostics and new technologies. Nanoparticles can augment detection of malaria at lower parasite levels while providing fast and simple methodology. Novel use of MSP10 and gold nanoparticles to identify Plasmodium vivax’s DNA in urine can be utilized as screening tool with global application potentials. The proposed test could impact the control of the most common species of malaria in low resource settings as it could present a simple, fast, cheap and easy to interpret test. Furthermore, utilizing urine instead of blood eliminates the need for finger-prick which would increase safety profile and likely increase participation rate in mass screening programs.
| Malaria is the most common infectious disease in the tropics and subtropics [1]. Currently, P. vivax is endemic across Asia, the South Pacific, North Africa, Middle East, and South and Central America [2], and has recently reappeared in regions where it had previously been eradicated, including North America and Europe [3]. Currently, an estimated 2.9 billion people live at risk of P. vivax infection [4]. Research in malaria has been primarily focused on P. falciparum, the most fatal form of malaria. However, P. vivax can also cause severe illness with serious complications and costs, especially in children, in whom it has a major impact on growth [5–7].
Relying on microscopic identification of malaria species jeopardizes malaria control due to its limitations [8,9]. The WHO recognizes a need for rapid, accurate, and easy diagnostic tools in order to control malaria and need for such a test is mounting in developing countries [10]. Currently available rapid diagnostic tests (RDTs) are controversial due to their sensitivity and specificity, and differentiate poorly between plasmodium species [11–13].
With the expanding use of nanotechnology in the biomedical arena, nanoparticles can play a role in low cost, innovative diagnostics [14]. Gold nanoparticles aggregate and change their color from red to purple-blue upon exposure to single stranded DNA in aqueous solution while double stranded DNA stabilize them to preserve their red color and thus present an opportunity to develop a fast and easily interpreted diagnostic test [15–20].
Merozoite Surface Protein 10 (MSP10) is an immunogenic protein encoded by a single copy gene (in P. falciparum, GenBan Accession PF3D7 0620400; in P. vivax, PVX_114145) which is expressed in the asexual blood stages of Plasmodium falciparum and P. vivax [21]. One of at least 10 epidermal growth factor domain-containing proteins, the role of MSP10 in the biology of Plasmodium parasites has yet to be determined. Plasmodium MSP10 proteins have been identified as being subject to positive selection for amino acid-changing polymorphisms at the population genomic level [22,23]. Population genomics studies identify signatures of global dispersal and drug resistance in Plasmodium vivax [24].
Blood based tests can discourage screening both because of the pain associated with finger-prick and because of social and cultural beliefs about blood sampling. Less painful, more culturally sensitive, and safer tools for malaria diagnosis should encourage participation in mass screening programs and improve public health [25]. Urine contains circulating P. vivax DNA in detectable quantities [26–28] and can therefore serve as a less invasive and more acceptable sample for malaria screening and diagnosis. Also, urine contains less interfering proteins and inhibitors than blood which allows easier DNA extraction [29]. Furthermore, urine provides lower risks to healthcare personnel, with reliable amounts of malaria DNA found in urine despite being substantially lower than blood samples [30]. Additionally, urine color is not expected to obscure color change of gold nanoparticles.
In this pilot study, we tested the hypothesis that a colorimetric system using gold nanoparticles and MSP10 DNA detection in urine would be useful as a safe diagnostic and surveillance tool for P. vivax. Such a tool is needed for improving malaria control in the endemic setting.
Citrate reduced gold 15nm nanoparticles, and KCl were purchased from Sigma Aldrich (St. Louis, MO, United States). PBS was obtained from Invitrogen (Grand Island, NY, United States). NaCl and NaOH were acquired from Merck Millipore (Kenilworth, NJ, United States).
The two MSP10 oligonucleotides utilized in this study were a generous gift from Professor Mirko Zimic (Universidad Peruana Cayetano Heredia, Lima, Peru) and designed by Dr. Joseph Vinetz (University of California San Diego, United States). Crafted to represent the C-Terminal segment of MSP10, the first oligonucleotide has a sequence of 5´CACCATGGAACAGTTTATCCTGAAGAC3'. The other oligonucleotide was used as a representative of the N-terminal segment of MSP10. It has a sequence of 5´AGCCATGGAACGTGCTAAGTGCAACA3’.
Archived urine samples positive for P. vivax and P. falciparum were collected from Iquitos in Peru and Ghana, respectively. Negative control urine samples were collected from volunteers who were blood smear negative in Iquitos, Peru and in Ghana, as well as in Lima, Peru, which is a non-endemic site. All urine samples were collected by clean catch procedures. Ghana urine samples were pelleted in the field and shipped on dry ice, pH was adjusted and samples were refrozen at -80°C as described earlier [31]. Peru urine samples were stored initially at -20°C prior to freezing at -80°C (Table 1). Peru’s urine samples were stored for 8 months while all Ghana samples were stored for more than one year.
During the epidemiological surveys on the communities, the field microscopist reports whether a slide is positive or negative, and identifies the species, P. vivax and P. falciparum. They read 300 microscopy fields before the slide is reported as negative.
In the laboratory, a second reader (an experienced microscopist working for research projects for more than 15 years) read the slide to report species and parasite density assuming a white blood cell count of 6,000/μl. The research microscopist read until 500 microscopy fields, before the slide is reported as negative.
For quality control, 10% randomly selected slides (positive and negative) were reexamined by two blinded, expert microscopists at a reference laboratory in Loreto from Peru’s Ministry of Health. From the quality control examinations, the level of concordance varies between 98–100% for species and parasite density.
Urine was thawed at 25°C. Once urine was at room temperature, dipsticks were carried out to determine urine pH and the presence of protein. Each urine sample was centrifuged at 15,000 rpm for 5 minutes to remove sediments and then filtered using a 0.2 mm membrane (Minisart, Bohemia, NY, United States) to remove possible confounding particulates. The urine samples were diluted 1:16 with PBS. Diluted samples’ pH was adjusted to reach ≈ 6.4 using pH meter, and HCl and NaOH solutions. 50 uL of each diluted urine sample was heated at 95°C for 30 seconds using a thermocycler. Samples were cooled at room temperature for 10 minutes and 10 uL of either C-Terminal or N-Terminal MSP10 oligonucleotides and 20 uL of 0.25 M NaCl were added. The sample was heated at 59°C for two minutes and allowed to cool to room temperature for ten minutes. Finally, 50 uL of citrate reduced AuNPs were added. Two minutes later, the system was read visually and by spectrophotometer.
Urine and blood were collected for previous studies that were approved by institutional review boards of Universidad Peruana Caytano Heredia and University of Ghana, respectively. Written informed consents were obtained prior to storing samples as anonymous and unidentified.
Up to 84% of P. vivax positive samples stabilized the gold nanoparticles and maintained a red color while 97% of negative controls induced aggregation and allowed color change to purple-blue. This color change was distinctly distinguished by naked eye (Fig 1). Additionally, the color difference was well defined by spectrophotometer, with positive samples at wavelengths of 520 and negative samples exhibiting wavelengths of 610–630 (Fig 2).
The colorimetric system was able to detect P. vivax with variable sensitivity. The sensitivity was dependent on which segment of MSP10 was used. The N-terminal segment distinguished 26 of 31 positive samples (84%) while the C-terminal segment distinguished only 20 of 31 samples (65%).
Both the N-terminal and C-terminal segments of MSP10 had an overall specificity of 97% in urine, with only one false positive out of 45 control samples. The false positive was from a laboratory control in Lima (Fig 3).
Parasitemia level was determined by blood smears collected at the same time as the urine samples. Data were shared after running the colorimetric system on urine to minimize investigator bias. There was no correlation between parasitemia level in blood and false negativity. The lowest parasitemia level observed in blood smear that also had a positive colorimetric test was 12 parasite/uL. However, there were two false negative samples with average parasitemia level in blood of 2510 parasite/uL (Table 2).
N-Terminal MSP10 oligonucleotide had a lower cross-reactivity (21%). C-Terminal MSP10 oligonucleotides showed high cross-reactivity with P. falciparum in urine utilizing colorimetric system (36%) (Table 3).
It took an average of 45 minutes from collecting urine to reading the test. Cost of the raw materials for each test is $0.20.
Currently available RDTs suffer from limited sensitivity, specificity, genetic instability, the inability to differentiate Plasmodium species, expense, and limitation in sample types (blood only) [11–13]. Blood tests also cannot be performed in many field settings since expertise, temperature control, storage, and laboratory equipment are unavailable, [11]. Urine has been reported as an accessible and reliable source of malaria DNA with lower detection threshold than blood; however, it has lower levels of detectable antigen and current RDTs have a minimum detection level of 100 parasites per microliter [26,32]. However, urine contains smaller double stranded DNA of 150-to 250-nucleotide size that can interact with nanoparticles and still be utilized as a valid source of microorganisms’ DNA [28,33].
P. vivax presents unique diagnostic challenges because of its genetic and geographic diversities [34,35]. The emergence of sequencing technologies and malaria sequences is providing us with a greater understanding of conserved regions, which must be targeted for broad-applicability of diagnostic tests [36]. Merozoite Surface Protein 10 (MSP10) is one of the asexual stage proteins of P. vivax linked to erythrocyte invasion [37]. It has two prominent EGF-like domains at the C-terminus, which are highly conserved and carry close homogeneity among all Plasmodium species [38,39].
We used two segments in our study. The N terminal segment showed superior sensitivity and equal specificity when compared to the C terminal segment. It could be due to its higher adenine and guanine contents as they both have higher adsorption rate to AuNPs’ surface [40,41].
The sensitivity of MSP10 oligonucleotides in detecting DNA in urine samples is dependent on quality of urine. Stored urine is known to have less DNA than fresh urine and age could play an important role in determining sensitivity of MSP10 oligonucleotides [28].
MSP10 has little similarity among Plasmodium species apart from the two EGF-like domains [21], which could explain lower cross-reactivity of N-terminal segment with P. falciparum. To be able to lower cross-reactivity, optimization of MSP10 oligonucleotides is required in future work.
Currently, most commercially and widely used malaria RDTs employ monoclonal antibodies to identify histidine-rich protein two (HRPII) [32]. Although HRPII based RDTs reported variable sensitivity and specificity, HRPII was subject to genetic and geographical diversities along with gene polymorphism [42]. Additionally, monoclonal antibodies against HRPII might cross react with other proteins [42]. In Peru, approximately 30% of P. Falciparum had HRPII gene deletions, which might lead to false negative RDTs results [43]. Additionally, other RDTs employed malaria markers that have been subject to controversy. Aldolase, an isoenzyme widely used to diagnose P. falciparum and P. vivax, was criticized for its low sensitivity and genetic variability in diagnosing P. vivax [44]. P.vivax’s lactate dehydrogenase (pLDH), another common antigen used in malaria RDTs, was found to be affected by its gene polymorphism [45]. Furthermore, pLDH requires whole blood sampling and declines very fast, with clearance of asexual parasitemia [45]. This study provides evidence that MSP10 DNA can serve as a marker for malaria at a global scale, and may deliver innovative tools to aid in malaria control.
In developing other diagnostic tests, nanoparticles have been utilized both for concentrating and for detecting samples with low concentration of target molecules in austere settings [46–48]. Jeon et al in 2013 found that AuNPs could be used to recognize P. vivax DNA in diluted blood samples using pLDH with detection levels as low as 74 parasites/uL [46]. We were unable to determine our detection level due to lack of correlation between Plasmodium species’ blood and urine DNA levels, which has been previously reported [30]. Despite our ability to detect a positive urine sample in a patient with low blood parasitemia, we had false negative samples with very high blood parasitemia levels indicating poor correlation between our test and parasitemia levels in blood.
In spite of lack of correlation between blood and urine DNA, this test has many other advantages. With further simplification of the process, it has potential to represent a simple, rapid test that does not require extensive lab skills, which will make it suitable as point of care test for low resource settings [49].
To our knowledge, this test is the first RDT utilizing urine samples rather than blood and employing nanoparticles. The colorimetric assay using AuNPs and MSP10 oligonucleotides to detect P. vivax in urine holds potentials to provide a safe, simple, rapid, and cheap tool to diagnose one of the most common form of malaria. Innovative use of MSP10 as a marker for P. vivax has potential for global application in mass screening programs.
|
10.1371/journal.ppat.1007543 | Merkel cell polyomavirus Tumor antigens expressed in Merkel cell carcinoma function independently of the ubiquitin ligases Fbw7 and β-TrCP | Merkel cell polyomavirus (MCPyV) accounts for 80% of all Merkel cell carcinoma (MCC) cases through expression of two viral oncoproteins: the truncated large T antigen (LT-t) and small T antigen (ST). MCPyV ST is thought to be the main driver of cellular transformation and has also been shown to increase LT protein levels through the activity of its Large-T Stabilization Domain (LSD). The ST LSD was reported to bind and sequester several ubiquitin ligases, including Fbw7 and β-TrCP, and thereby stabilize LT-t and several other Fbw7 targets including c-Myc and cyclin E. Therefore, the ST LSD is thought to contribute to transformation by promoting the accumulation of these oncoproteins. Targets of Fbw7 and β-TrCP contain well-defined, conserved, phospho-degrons. However, as neither MCPyV LT, LT-t nor ST contain the canonical Fbw7 phospho-degron, we sought to further investigate the proposed model of ST stabilization of LT-t and transformation. In this study, we provide several lines of evidence that fail to support a specific interaction between MCPyV T antigens and Fbw7 or β-TrCP by co-immunoprecipitation or functional consequence. Although MCPyV ST does indeed increase LT protein levels through its Large-T Stabilization domain (LSD), this is accomplished independently of Fbw7. Therefore, our study indicates a need for further investigation into the role and mechanism(s) of MCPyV T antigens in viral replication, latency, transformation, and tumorigenesis.
| Merkel cell carcinoma (MCC) is a very aggressive and deadly neuroendocrine skin cancer. Merkel cell polyomavirus (MCPyV) contributes to the development and maintenance of 80% of MCCs through the expression of its truncated large tumor antigen (LT-t) and small tumor antigen (ST). MCPyV ST is thought to be primarily responsible for transformation and tumorigenesis through many mechanisms including stabilization of MCPyV LT-t and other cellular proteins involved in proliferation such as c-Myc. As c-Myc is a known target substrate, and MCPyV LT-t is a proposed target substrate, of the ubiquitin ligase Fbw7, it is currently thought that ST stabilizes these proteins and transforms cells through binding and perturbing the function of Fbw7. However, neither MCPyV LT-t nor ST contain a canonical Fbw7 degron sequence necessary for this interaction. MCPyV LT-t, found in MCCs, does not bind to, nor is targeted by, Fbw7. However, an ill-defined, unidirectional interaction between MCPyV LT, ST, and Fbw7 was observed, but had no functional consequence. Therefore, this study calls for further investigation into the mechanism(s) by which MCPyV ST leads to the development and maintenance of MCC.
| Merkel cell carcinoma (MCC) is an extremely rare, but aggressive, neuroendocrine skin cancer with an incidence of 0.7 cases/100,000 person-years in the United States, and a less than 45% 5-year survival rate, making MCC almost three times as lethal as melanoma [1, 2]. Although MCC was first described in 1972, it wasn’t until 2008 that a previously undescribed polyomavirus was found to be clonally integrated into 80% of MCC tumors and was thus termed Merkel cell polyomavirus (MCPyV) [3, 4]. MCPyV is a small, circular, dsDNA virus that utilizes alternative splicing of its early region (ER) to generate four ER proteins including the Large Tumor antigen (LT), Small Tumor antigen (ST), 57kT, and Alternate Large-T Open reading frame (ALTO) [5]. However, in MCCs only ST and a truncated form of the LT antigen (LT-t) are expressed. The tumor specific truncation of LT occurs as a consequence of a premature stop codon or deletion in the viral genome downstream of the Rb binding site and before the LT helicase domain [6]. Selection of this truncation is thought to be driven through the prevention of several activities and consequences deleterious for tumorigenesis [7, 8].
Unlike other oncogenic polyomaviruses, such as simian vacuolating virus 40 (SV40) and murine polyomavirus (MPyV), the dominant transforming protein of MCPyV appears to be ST; however, LT-t has also been found to be necessary for tumor maintenance through the activity of its Rb binding domain [9–11]. Several mechanisms of MCPyV ST mediated transformation have been proposed including, but not limited to, increased cap-dependent translation through 4E-BP1 phosphorylation [9, 12], a more migratory phenotype through dysregulation of actin and microtubule dynamics [13, 14], and transcriptome perturbations through binding of L-Myc and EP400 [15, 16].
It has also been proposed that MCPyV ST can perturb the functions of several cellular ubiquitin ligases, most notably SCFFbw7, to accomplish transformation [12, 17–19]. SCFFbw7 is a multi-protein complex responsible for binding and ubiquitinating target proteins for proteasomal degradation [20]. Fbw7 is the component of the SCF (Skp1, Cul1 and Fbox protein) ubiquitin ligases responsible for both substrate recognition and recruitment of the ubiquitination machinery. There are three isoforms of Fbw7 which localize to different subcellular compartments: Fbw7α is nucleoplasmic, Fbw7β is cytoplasmic, and Fbw7γ is nucleolar, with Fbw7α being thought to perform the majority of Fbw7 functions [21, 22]. All three Fbw7 isoforms contain three functional domains: 1) the Fbox domain binds Skp1, thereby linking Fbw7 to the SCF ubiquitination machinery, 2) the WD40 domain forms a β-propeller with two phosphate binding pockets that recognize target proteins containing phosphorylations within a conserved Cdc4 phospho-degron (CPD), named after the budding yeast Fbw7 homolog CDC4, and 3) the dimerization domain mediates the formation of Fbw7 dimers [20, 22–25] (S1A Fig). Most Fbw7 substrates share a dually phosphorylated degron with both phosphates spaced four amino acids apart (e.g. pTPPxpT/S), while the “+4” position can also be a negatively charged amino acid, and the “0” position is preceded by several hydrophobic amino acids (S1B Fig) [22]. Of note, this CPD sequence is conserved among all currently known Fbw7 targets [24]. Three arginine residues within the WD40 domain of Fbw7 (R465, R479, and R505) make direct contacts with degron phosphates and are most critical for substrate interactions [20, 25]. Fbw7 is frequently mutated in cancers due to its tumor suppressive function of regulating protein levels of several cellular oncoproteins such as c-Myc and cyclin E [21, 22, 26–28]. Interestingly, Fbw7 function has also been shown to be perturbed by the LT protein of SV40. SV40 LT contains a decoy-CPD at its extreme C-terminus (pTPPPE), which can directly interact with the WD40 domain of Fbw7 and consequently decrease turnover of normal Fbw7 target substrates (S1B Fig) [29].
In contrast to SV40 LT, Kwun et al. have proposed MCPyV LT to be targeted for proteasomal degradation through interaction with the WD40 domains of several ubiquitin ligases including Fbw7 and Fbw1 (β-TrCP) [17, 18]. As MCPyV LT and LT-t play an important role in viral replication and tumor maintenance, respectively, their rapid turnover would be deleterious [17, 18]. However, amino acids 91–95 of MCPyV ST have been shown to increase LT protein levels, and was thus termed the Large-T Stabilization Domain (LSD) [17]. Currently, the ST LSD is thought to increase LT protein levels by binding and sequestering several ubiquitin ligases, including Fbw7, from LT and their other cellular targets, such as c-Myc, thereby decreasing their turnover (S1C Fig) [17]. Such an activity, similar to SV40 LT, would suggest a possible role of the ST LSD in transformation and tumorigenesis, as increased concentrations of a MCPyV tumor antigen and cellular oncoproteins could lead to aberrant cellular proliferation, increased translation, and genomic instability [12, 17, 19]. Indeed, mutation of the LSD has been shown to ablate the ability of MCPyV ST expressing cells to form colonies in soft agar and promote epithelial hyperplasia in pre-term transgenic mouse embryos [17, 30]. Furthermore, as MCPyV LT is necessary for viral replication, the activity of the ST LSD has also been proposed to initiate viral replication and exit from viral latency [18].
In this study we further investigated the proposed interaction between MCPyV T antigens and Fbw7, as we find targeting of MCPyV LT by Fbw7 conceptually difficult to rationalize, and because of the absence of a canonical Fbw7 CPD in both MCPyV LT and ST. Contrary to the interactions and model proposed by Kwun et al., MCPyV LT, LT-t and ST failed to specifically interact with, or be destabilized by, Fbw7 and Fbw1 [17, 18]. Furthermore, the MCPyV ST LSD was capable of increasing LT protein levels independent of Fbw7. Therefore, this study calls for further investigation into the molecular mechanisms by which MCPyV ST contributes to the development and maintenance of MCPyV positive MCC.
To first validate an interaction between Fbw7 and a well-established target, SV40 LT, co-immunoprecipitations were performed (Fig 1A and S1 Table) [29, 31, 32]. In 293A cells FLAG tagged Fbw7α was co-transfected with HA tagged SV40 LT and immunoprecipitated with an anti-HA antibody. As expected, Fbw7α readily co-immunoprecipitated with SV40 LT (Fig 1A, lane 4). SV40 LT was also found to co-immunoprecipitate the Fbw7α ΔFbox mutant, which is unable to ubiquitinate its substrate protein but is still capable of interacting with its substrate through an intact WD40 domain (Fig 1A, lane 6). The interaction between SV40 and Fbw7α also displayed degron dependence as the SV40 LT CPD mutant, in which the central phosphorylated threonine of the CPD was substituted with an alanine, was incapable of co-immunoprecipitating wild-type Fbw7 (Fig 1A, lane 5). Additionally, the interaction was disrupted by an Fbw7 WD40 mutant, in which a crucial arginine of the Fbw7 WD40 domain is mutated to a leucine (Fbw7α ΔFb R505L) (Fig 1A, lane 7).
Having confirmed the reliability of the immunoprecipitation (IP) protocol utilized, the interaction between the CPD deficient MCPyV T antigens and Fbw7 was investigated. 293A cells were co-transfected with either HA-tagged MCPyV full-length LT, MCPyV truncated LT (LT-t), untagged MCPyV ST, or HA-SV40 LT, in combination with Fbw7α. Despite the fact that 293A cells express adenovirus E1A and E1B proteins, they were chosen as our experimental model to replicate the studies done by Kwun et al. [17, 18] and other Fbw7 substrate studies [21, 29]. Fbw7 proteins were pulled-down with an antibody recognizing a FLAG tag on the N-terminus of Fbw7, reciprocal to the IP performed in Fig 1A. Surprisingly, only SV40 LT co-immunoprecipitated with Fbw7α in this IP, consistent with it being a bona-fide Fbw7 binding partner, whereas none of the MCPyV T antigens (LT, LT-t, and ST) interacted with Fbw7α even though their expression was confirmed in the cellular lysates (Fig 1B). It should be noted that the alternatively spliced form of the full-length MCPyV LT, 57kT, was preferentially pulled-down in these assays; however, neither full-length LT nor 57kT are expressed in MCC tumors [6]. Also of note, the tumor expressed MCPyV LT-t was found to decrease Fbw7α protein levels, as seen in the protein lysate (Fig 1B, lane 7). This decrease was not found to be a result of LT-t decreasing Fbw7α transcription by qRT-PCR (S2A Fig). Currently, the mechanism underlying this observation is not understood, and could either be a physiologically irrelevant consequence of artificial LT-t expression, or lead to the discovery of a novel role for MCPyV LT-t.
To ensure that the inability of FLAG tagged Fbw7 to pull-down the MCPyV T antigens is not a consequence of the tag used, we performed an identical experiment using Myc tagged Fbw7. Consistent with Fig 1B, Myc tagged Fbw7 was capable of co-immunoprecipitating SV40 LT, but not the MCPyV T antigens (Fig 1C), suggesting that the tag used to pull-down Fbw7 is not responsible for the inability to observe an interaction between the MCPyV T antigens and Fbw7.
It is possible that the HA tags found on MCPyV LT and LT-t could pose a steric hindrance to the interaction with Fbw7, although this would not explain the inability of MCPyV ST to bind Fbw7, as ST was untagged in these experiments. A similar co-immunoprecipitation to Fig 1B was performed by transfecting 293A cells with FLAG tagged Fbw7, and untagged SV40 and MCPyV T antigens, followed by pulldown of Fbw7 and XT10 immunoblotting to detect co-immunoprecipitated T antigens. Again, untagged SV40 LT was capable of co-immunoprecipitating with Fbw7; however, none of the untagged MCPyV T antigens co-immunoprecipitated (Fig 1D). Therefore, the N-terminal HA tag is not responsible for the inability of the MCPyV T antigens to co-immunoprecipitate with Fbw7.
In an effort to understand the discrepancies between our data and those of Kwun et al. [17, 18], a reciprocal co-immunoprecipitation was performed with an antibody that recognizes the J-domain of both SV40 and MCPyV T antigens, XT10. Immunoblotting for Fbw7 (FLAG) revealed an interaction between MCPyV LT and Fbw7α, and a very weak interaction between MCPyV ST and Fbw7α (Fig 1E). Furthermore, we were surprised to find that MCPyV LT and ST bound equivalently to wild-type Fbw7α and the Fbw7α WD40 mutant, R505L (Fig 1E, lanes 6 and 9), which is unable to bind any other known Fbw7 substrates, including SV40 LT (Fig 1A, lane 7). Although the interaction between MCPyV ST and Fbw7 R505L appears to be slightly fainter than the already weak interaction between Fbw7 and ST, this is most likely a consequence of less ST pulled-down in this lane (Fig 1E, lane 9).
To determine whether XT10 was responsible for the contradictory binding results between co-immunoprecipitations performed in different directions, immunoprecipitations were performed with additional antibodies specific for the region shared by MCPyV LT, LT-t, 57kT and ST (common T–Ab5) (S3A Fig), or LT, 57kT and LT-t only (CM2B4, Ab3) (S3B Fig) (S1D Fig). MCPyV LT co-immunoprecipitated Fbw7 with each MCPyV T antigen antibody used for the pulldown (S3A and S3B Fig, lane 3). Similarly, MCPyV ST again very weakly co-immunoprecipitated Fbw7 (S3A Fig, lane 7). As seen with XT10 pull-down (Fig 1E), LT was found to co-immunoprecipitate with the Fbw7 WD40 mutant (Fbw7 ΔFb R505L) (S3A and S3B, lane 5). Taken together, it appears that MCPyV LT and ST interact with Fbw7 nonspecifically when pulled-down with antibodies specific for the MCPyV T antigens, and this interaction is not reproducible when pulled-down with an Fbw7 anti-tag antibody (FLAG or Myc).
The LSD of MCPyV ST has been proposed to be responsible for binding Fbw7α even though this domain does not contain a recognizable conserved CPD (S1B Fig) [17]. To assess the role of the ST LSD in the weak interaction between ST and Fbw7, a ST ΔLSD mutant was constructed and tested in similar co-immunoprecipitation experiments. In our hands, the ST ΔLSD mutant was still capable of interacting with Fbw7, and appeared to bind more strongly to Fbw7α than wild-type ST, most likely a consequence of greater amounts of ST ΔLSD mutant being expressed (Fig 2A, lane 4). Similar results were found when the ST ΔLSD mutant was immunoprecipitated with Ab5 (S3A Fig, lane 8).
The current model proposes MCPyV ST preferentially binds and sequesters Fbw7 away from MCPyV LT and its other cellular targets, suggesting ST has a higher affinity for Fbw7 than LT (S1C Fig) [17, 18]. However, since the unidirectional interaction between MCPyV ST and Fbw7α was found to be weaker than that of LT (Figs 1E and 2A, S3A Fig), and independent of the ST LSD (Fig 2A, S3A Fig), we examined whether ST was capable of increasing MCPyV LT protein levels in the absence of Fbw7. Knockdown of Fbw7 in 293A cells was performed with lentiviruses containing either an shControl or shFbw7, and the knockdown was confirmed by pulldown and immunoblotting of endogenous Fbw7 (Fig 2B). To investigate the described perturbation of Fbw7 by MCPyV ST, a downstream target of Fbw7, c-Myc, was analyzed in 293A cells expressing shControl, shFbw7, and/or MCPyV ST. As expected, Fbw7 knockdown led to increased c-Myc protein levels (Fig 2C, lane 3); however, MCPyV ST expression in shControl 293A cells did not increase endogenous c-Myc protein levels (Fig 2C, lane 2). Therefore, MCPyV ST’s inability to stabilize a downstream Fbw7 target, c-Myc, is contrary to the model of ST sequestration of Fbw7. As has been described previously, MCPyV LT protein levels were increased by co-expression of MCPyV ST, but not the ST ΔLSD mutant (Fig 2D, lanes 1–3) [17]. However, in Fbw7 knockdown cells MCPyV ST also increased LT protein levels (Fig 2D, lane 5). Therefore, the LSD sequence in ST is capable of increasing LT protein levels independent of Fbw7, supporting the lack of functional significance of the LSD sequence in the unidirectional interaction between MCPyV ST and Fbw7. Taken together, MCPyV ST does not perturb the function of Fbw7, and thus opens the door for further investigation into the mechanism by which the ST LSD increases LT protein levels.
In MCPyV positive MCC only ST and LT-t are expressed [4, 6]. Since ST sequestration of Fbw7α away from LT-t is a proposed mechanism of ST induced transformation and tumorigenesis [17], we next sought to determine whether LT-t is bound and destabilized by Fbw7α. MCPyV LT-t or SV40 LT were immunoprecipitated with either the XT10, Ab3, or Ab5 antibodies from whole cell lysates of 293A cells transfected with the T antigens and Fbw7α. Unlike SV40 LT, MCPyV LT-t was unable to co-immunoprecipitate with Fbw7α (Fig 3A, lane 5; S3A and S3B Fig, lane 6). The interaction between MCPyV LT-t and Fbw7α has been described as extremely transient due to rapid degradation and is, therefore, only observable after treatment with the broad proteasome inhibitor MG132 [17]. Similar to Kwun et al., we were able to see an interaction between MCPyV LT-t and Fbw7α when treated with MG132; however, an even stronger interaction was observed between our negative control, SV40 LT-T701A, and Fbw7α with MG132 treatment, suggesting that pleiotropic effects of MG132 treatment can lead to false positives (compare Fig 1A with S2B Fig) [17]. A more direct and reliable way to assess binding partners of Fbw7α uncoupled from turnover is through the utilization of the degradation incompetent Fbw7α ΔFbox mutant. MCPyV LT-t was also unable to co-immunoprecipitate with Fbw7α ΔFbox, further suggesting the inability of these two proteins to interact, even in the absence of turnover (Fig 3A, lane 6).
To determine if MCPyV LT-t is destabilized by co-expression of Fbw7α, cells were co-transfected with identical amounts MCPyV LT-t and increasing amounts of either Fbw7α or the degradation incompetent Fbw7α ΔFbox mutant. Consistent with Fbw7α not binding LT-t, co-expression of Fbw7α did not reduce LT-t protein levels (Fig 3B). This is in contrast to c-Myc, which was readily destabilized by Fbw7α, but was not destabilized by the Fbw7α ΔFbox mutant (Fig 3C). Together, these results suggest that the LT-antigen found in MCC, LT-t, is neither bound nor destabilized by Fbw7α.
Kwun et al. mapped the domain of MCPyV LT-t targeted by Fbw7α to residues around S239, although this domain does not resemble a canonical Fbw7 CPD (S1B Fig) [18]. Because we did not detect an interaction between LT-t and Fbw7, the S239A mutation was constructed in full-length LT; however, this mutant was still capable of co-immunoprecipitating Fbw7α (Fig 4A; S3A and S3B Fig, lane 4). Therefore, since MCPyV LT does not contain a canonical degron, and the proposed degron, S239, was found to be dispensable, it is unknown how or where the unidirectional interaction between MCPyV LT and Fbw7 is occurring.
The MCPyV T antigen specific antibodies used in these immunoprecipitations can pull-down both MCPyV LT and 57kT since they are identical in sequence with the exception of a spliced-out intron of the 57kT (S4A Fig). Therefore, it is unclear whether or not the observed interaction between MCPyV LT/57kT and Fbw7α is specific to MCPyV LT, 57kT, or a domain shared by both. However, a construct that could only express the 57kT also readily co-immunoprecipitated with Fbw7 independent of the WD40 domain (Fig 4B).
Since MCPyV LT and 57kT, but not LT-t, were capable of binding Fbw7α, the Fbw7α binding domain of MCPyV LT and 57kT must reside in the C-terminal 100 amino acids common to both proteins (S4A Fig). Although the decoy-CPD of SV40 LT is found at its extreme C-terminus [29], the entire full-length MCPyV LT does not contain a canonical Fbw7 CPD. To identify the domain within the C-terminal 100 amino acids of MCPyV LT responsible for binding Fbw7α, an alanine scan in which sequential 5 amino acids were substituted with alanines was performed. Every stable mutant of the alanine scan was capable of co-immunoprecipitating Fbw7α (S4B–S4E Fig), though we cannot rule out interactions with sequences that when mutated destabilized LT. It is also possible that there are multiple weak degrons in the C-terminus of MCPyV LT, though this type of binding has been refuted [33, 34] and would nevertheless be inconsequential to MCPyV transformation as this region of LT is not expressed in MCCs. Therefore, the MCPyV LT domain responsible for the unidirectional interaction observed between LT and Fbw7α is still unresolved.
After unsuccessful identification of the MCPyV LT domain responsible for binding Fbw7α, we decided to also investigate the domain of Fbw7α responsible for binding MCPyV LT. As was shown in Fig 1E, MCPyV LT and ST were capable of interacting with an Fbw7α WD40 domain mutant, R505L, which has not been reported for any other known Fbw7 substrates. In contrast, the proposed interaction between MCPyV LT and Fbw7α was reported to be WD40 dependent, as an R465C mutation ablated this interaction [17]. Since all three arginine residues (R465, R479, R505) are necessary for the function of the Fbw7α WD40 domain, experiments were conducted to determine whether an Fbw7 R465C mutation would interfere with binding to LT. HA-tagged Fbw7α and the mutant R465C (gifts from Patrick Moore and Yuan Chang) in addition to FLAG-tagged Fbw7α and R505L were studied for their ability to interact with MCPyV LT or LT-t. Both wild-type Fbw7α constructs, regardless of having an HA or FLAG tag, were able to co-immunoprecipitate with MCPyV LT, but not the tumor-specific LT-t with XT10 pull-down (Fig 5A, lanes 5,6, 9, and 10). However, we found both WD40 domain mutants, R505L and R465C, to co-immunoprecipitate with MCPyV LT, further supporting that the unidirectional interaction observed between MCPyV LT and Fbw7 occurs independently of residue R465 in the WD40 domain (Fig 5A, lanes 7 and 8) [17]. Furthermore, the interaction between Fbw7α and MCPyV LT did not lead to destabilization of LT (Fig 5B). Thus, the interaction observed between MCPyV LT and Fbw7α is independent of the WD40 domain, and is consistent with the inability of Fbw7 to degrade LT, as the WD40 domain is responsible for positioning the substrate for ubiquitination and subsequent degradation [20, 22, 25].
The region of Fbw7α responsible for binding MCPyV LT mapped to the common C-terminal region shared by all Fbw7 isoforms (S5A and S5B Fig), which is known to contain three functional domains: the dimerization, Fbox, and WD40 domains [22]. As shown thus far, the Fbw7 Fbox and WD40 domains are dispensable for this interaction (Figs 1E,4B and 5A, S3A and S3B Fig); therefore, the dimerization domain was also assessed. A double Fbw7α mutant containing a deletion of both the Fbox and dimerization domain (Fbw7α ΔFD) was still capable of co-immunoprecipitating with both MCPyV LT and ST (S5C Fig). Therefore, it is unclear which domain of Fbw7α is responsible for the unidirectional interaction observed with MCPyV LT and ST.
In addition to Fbw7, it has also been proposed that MCPyV LT is targeted for destruction by several other ubiquitin ligases including another Fbox containing E3 ubiquitin ligase, Fbw1(β-TrCP) [18]. The phospho-degron for Fbw1 recognition is also well defined and contains the sequence DpSGXXpS/pT [35], which is present at the extreme C-terminus of MCPyV LT. However, the reported domain of MCPyV LT responsible for the interaction with Fbw1 was mapped to S147 [18], which is missing several components of the canonical Fbw1 phospho-degron sequence (S1B Fig). Further, it has been hypothesized that MCPyV ST promiscuously binds several E3 ubiquitin ligases through its LSD to facilitate LT stability, including Fbw1 [12, 17–19], despite the fact that the ST LSD also shares no resemblance to the canonical Fbw1 degron (S1B Fig). To investigate the proposed interaction between MCPyV LT, LT-t, ST and Fbw1, co-immunoprecipitations were performed in which the T antigens were pulled-down and co-immunoprecipitated Fbw7 or Fbw1 was detected. MCPyV LT bound both Fbw7α and Fbw1; however, LT-t and ST failed to bind Fbw1 (Fig 6A). Furthermore, an alanine MCPyV LT mutant, S147A [18], was still capable of co-immunoprecipitating Fbw1 (Fig 6B). Similar results were found when the immunoprecipitation was performed with additional MCPyV T antigen specific antibodies (S6A and S6B Fig). Similar to Fbw7, MCPyV LT was capable of binding to both Fbw1 WD40 and ΔFbox mutants (Fig 6C lane 3, 6D lane 5), again suggesting the interaction between MCPyV T antigens and Fbw1 to be non-specific. Of note, MCPyV LT-t was also capable of decreasing Fbw1 protein levels, as was seen with Fbw7, (compare Fig 1B to Fig 6A) even though MCPyV LT-t was not found to bind either proteins. Therefore, this either novel function of MCPyV LT-t, or irrelevant consequence of artificial LT-t expression, is broadly acting.
Integration and expression of MCPyV T antigens accounts for 80% of MCC cases [4]; therefore, elucidating the mechanisms by which LT-t and ST accomplish transformation and tumorigenesis is paramount for the design of novel therapeutics to treat MCPyV positive MCCs. Furthermore, such insight could also have implications not only for understanding polyomavirus oncogenesis and disease, but also for cellular pathways, homeostasis, and broad disease pathology. MCPyV ST has been shown to increase protein levels of MCPyV LT, LT-t, and several other cellular oncoproteins involved in proliferation, such as c-Myc and cyclin E [17, 30]. This has been proposed to be a consequence of ST sequestration of several ubiquitin ligases, including Fbw7 and β-TrCP, and underlies the proposed mechanisms of ST enhancement of viral replication, and induction of transformation and tumorigenesis (S1C Fig) [12, 17–19].
Since many cellular pathways, processes, and fates are sensitive to, and directed by, the amount of a given protein, and ubiquitin-mediated proteolysis is irreversible, this process is tightly regulated [36]. To be targeted by Fbw7, and subsequently ubiquitinated for proteasomal degradation, a protein must 1) contain a specific and highly conserved amino acid sequence called a CPD, and 2) include multiple site-specific phosphorylations performed by other tightly regulated kinases within the cell [20, 25–27]. For these reasons, it was surprising that both MCPyV LT-t and ST have been reported to bind Fbw7, as neither viral protein contain the well-known, conserved, Fbw7 CPD.
The concept of viral oncoprotein perturbation of ubiquitin ligases is well-established [29, 37–41]. Although evolutionarily related, several significant differences lie between the interaction of SV40 LT and Fbw7, and the proposed interaction between MCPyV LT-t and Fbw7. SV40 LT binds the WD40 domain of Fbw7 through its canonical CPD sequence, leading to the stabilization of SV40 LT and Fbw7 targets, and mislocalization of the nucleolar Fbw7γ isoform [29]. In contrast, MCPyV LT-t has been proposed to also bind the WD40 domain of Fbw7, leading to its destruction despite the absence of a canonical CPD [17, 18]. While not impossible, it is difficult to rationalize how MCPyV LT-t is binding to the WD40 domain of Fbw7 in the absence of a CPD. Also, why MCPyV LT-t, a protein necessary for tumor viability, would retain sequences that would ultimately lead to its destruction, and therefore rely on the activities of another viral protein, ST, to avoid degradation is counter-intuitive [17, 18].
The interaction between MCPyV T antigens and ubiquitin ligases have been implicated in regulating both viral latency in normal, asymptomatic, viral infection (LT and ST), and transformation in the setting of MCC (LT-t and ST) [17, 18]. In this report, we provide several lines of evidence that oppose both proposed models.
It has been proposed that MCPyV ST promiscuously binds and sequesters several E3 ubiquitin ligases through its LSD; however, as each of these ubiquitin ligases recognize distinct phospho-degrons, it is difficult to imagine how one domain could interact with several different highly specific proteins [12, 17, 18, 20, 35]. In our hands, weak ST binding to Fbw7 was only observed unidirectionally and independent of the Fbw7 WD40 and ST LSD domains. Moreover, an interaction between MCPyV ST and Fbw1 was never observed. Therefore, we hypothesized that MCPyV ST, which lacks a canonical CPD and does not bind to the WD40 domain of Fbw7, may possess an alternative mechanism for binding and perturbing the function of Fbw7 independent of the WD40 domain. For instance, binding to another component of the SCF complex, as has been proposed for the adenovirus E1A protein [41], could explain the proposed ubiquitin ligase binding promiscuity of MCPyV ST. However, this hypothesis is also not supported by our data, as MCPyV ST also unidirectionally co-immunoprecipitated the Fbw7 ΔFbox mutant, which cannot recruit the remaining SCF complex. Furthermore, others have hypothesized that MCPyV ST may inhibit the formation of dimers to perturb Fbw7 function, as the dimerization domain of Fbw7 has been found to enhance binding to low affinity substrates [17, 23]. However, as we have also shown MCPyV ST capable of binding a dimerization domain Fbw7 mutant, it is unlikely that MCPyV ST utilizes this mechanism to perturb Fbw7 targeting of low affinity substrates, such as LT. Therefore, thorough investigation of several mechanisms by which MCPyV ST could perturb the function of ubiquitin ligases failed to uncover or support the proposed promiscuity of MCPyV ST ubiquitin ligase perturbation.
The proposed model of MCPyV ST mediated transformation suggests that MCPyV LT-t and other cellular oncoproteins are stabilized through MCPyV ST sequestration of Fbw7, leading to aberrant cellular proliferation and transformation [17]. Although our studies confirmed the ability of ST to increase LT protein levels through the LSD, this was accomplished independently of Fbw7. This is supported by the fact that MCPyV LT-t was never found to interact with, or be destabilized by, Fbw7. This suggests that MCPyV LT-t is not targeted by Fbw7, and is consistent with MCPyV ST not specifically binding to, or perturbing the function of Fbw7 to increase LT-t protein levels. Furthermore, MCPyV ST had no effect on c-Myc protein levels, a bona-fide Fbw7 substrate, further confirming no association between MCPyV ST and Fbw7. Thus, it can be concluded that the mechanism(s) by which ST increases LT-t protein levels, and transforms cells, does not involve sequestration and/or perturbation of Fbw7.
MCPyV LT destabilization by ubiquitin ligases has been proposed to play a role in maintaining viral latency in a normal infection, as reduced LT protein levels decreases viral replication [18]. Unlike tumor expressed MCPyV LT-t, we were able to observe a weak interaction between MCPyV LT and Fbw7; however, this interaction was only observed unidirectionally, was independent of proposed LT binding domains and the Fbw7 WD40, did not lead to destabilization of LT, and has no relevance to transformation and tumorigenesis. Although neither MCPyV LT nor ST contain a canonical Fbw7 CPD, one could hypothesize that the unidirectional interaction observed between LT, ST and Fbw7 could occur through a novel degron sequence found on these T antigens. However, such an interaction would still be dependent on an intact Fbw7 WD40 domain, which we found to be dispensable for binding. In addition, we were unable to identify the domains of the T antigens nor Fbw7 responsible for the unidirectional interaction through both direct mutagenesis of the proposed binding domains and comprehensive mutational screens. Thus, the proposed model of ST sequestration of ubiquitin ligases to stabilize LT protein levels and induce viral replication in the setting of a normal viral infection can not be confirmed at the level of the interaction itself or functionally.
To directly evaluate the discrepancies between our data and those of Kwun et al., plasmids were exchanged and tested in parallel. Here, both Fbw7 expressing constructs were capable of co-immunoprecipitating with MCPyV LT, but not the tumor expressed MCPyV LT-t. Also contrary to previous reports, the interactions observed between MCPyV LT and the provided Fbw7 constructs were independent of the WD40 domain. Thus, it is unlikely that the observed weak, unidirectional, destabilization and CPD/WD40 independent interaction observed between MCPyV T antigens and Fbw7 is relevant to the role of the T antigens in either the viral life-cycle or tumorigenesis.
In conclusion, we thoroughly investigated the specific interaction and downstream consequences of the proposed involvement of MCPyV T antigens and Fbw7. A likely artifactual interaction between MCPyV LT, ST and Fbw7 was observed, but not with LT-t, thereby dismissing the relevance of the proposed model in MCC tumorigenesis. It should be noted that ST is capable of increasing LT protein levels through its LSD; however, the mechanism by which this is accomplished by MCPyV ST is not fully understood, as Fbw7 was found to play no role in this interaction. In conclusion, we propose MCPyV ST to be capable of transformation and tumorigenesis independent of ubiquitin ligase perturbation, and therefore open the door for further investigation into the mechanisms by which MCPyV ST and the ST LSD contribute to viral replication, latency, and induce transformation and tumorigenesis.
Adenovirus-transformed human embryonic kidney cell lines (HEK293A) (Thermo Fisher Scientific) were cultured in DMEM supplemented with 10% (vol/vol) fetal bovine serum, penicillin-streptomycin (Life Technologies), GlutaMAX (ThermoFisher Scientific), and Non-Essential Amino Acids (ThermoFisher Scientific). All cell lines were maintained at 37°C in 5% CO2. For transient transfection experiments, cells were plated in 10cm plates and transfected the next day using TransIT-293 Transfection Reagent (Mirus) at ~80% confluence. Cells were harvested 36–48 hours after transfection for co-immunoprecipitation and/or western blot analysis. Plasmids used are summarized in S1 Table. Fbw7 knock-down was performed by transducing 293A cells with lentiviruses containing shControl or shFbw7 (CAGAGAAATTGCTTGCTTT), followed by puromycin selection.
pCMV24-3xFLAG-Fbw7α wild-type and mutants, pCS2-5xMyc-Fbw7α, pCMV24-3xFLAG-Fbw1 wild-type and mutants, pCS2-HA-SV40 LT, and pCS2-HA-c-Myc have been described [21, 23, 24, 27, 29, 31, 32, 42]. MCPyV T antigens were subcloned into pCS2, and/or pCS2-HA vectors using the In-fusion HD Cloning Kit (Clontech). The pCS2-MCPyV.57kT plasmid was generated by GENEWIZ TurboGENE Gene Synthesis. The MCPyV LT alanine scan mutagenesis was performed by ThermoFisher Scientific GeneArt. MCPyV T antigen mutants were generated using New England BioLabs Q5 Site-Directed Mutagenesis Kit, with primers designed using the NEBase Changer tool (New England Biolabs). pCGN.HA-Fbw7 and pCGN.HA-Fbw7 R465C were provided by Patrick Moore and Yuan Chang (University of Pittsburgh).
Co-immunoprecipitations were performed using a modified protocol from Jianxin You (University of Pennsylvania). Briefly, 293A cells were harvested 36–48 hours after transfection and lysed with NP40 lysis buffer containing cOmplete, EDTA-free Protease Inhibitor Cocktail (Sigma-Aldrich) on ice. Cell lysates were lightly sonicated and the lysates were rotated at 4°C for 1 hour. The BCA Protein Assay Kit (Pierce) was used to quantify the protein concentrations. 500μg of normalized cell lysates were pre-cleared with 25μl of equilibrized Protein A/G Magnetic Beads (Pierce) for 1 hour at 4°C with rotation. The A/G Magnetic Beads were removed from the cellular lysates and discarded, followed by incubation with 10μg of the immunoprecipitation antibody overnight at 4°C with rotation (XT10-donated from Chris Buck, CM2B4 (Santa-Cruz, sc-136172), Ab3-donated from Jim DeCaprio, Ab5-donated from Jim DeCaprio, anti-HA (BioLegend-MMS-101P), 9E10 (made in house), anti-FLAG M2 (Sigma-Aldrich, F1804), or anti-Fbw7 (Bethyl Labs, A301-720). 25μl of protein A/G magnetic beads, per sample, were also blocked in 1% BSA overnight at 4°C with rotation. The next morning, antibody/cellular lysates were added to the 25μl of pre-blocked A/G magnetic beads, and incubated at room temperature, with rotation, for 1 hour. Immunoprecipitated samples were washed several times with KCL buffer, resuspended in SDS sample buffer, boiled, separated by electrophoresis, and transferred to an Immobilon-P PVDF Membrane (Millipore) in parallel with 30μg of whole cell lysate. Membranes were blocked in 4% milk overnight at 4°C, followed by incubation with the primary antibodies described above, in addition to the 2T2 antibody which binds an epitope common to both MCPyV LT and ST (provided by Chris Buck; National Cancer Institute). After washing, a mouse IgG light chain specific HRP conjugated secondary antibody (Cell Signaling-D3V2A, #58802) was incubated with the membranes in 4% milk for one hour followed by washing and chemiluminescent detection with a ChemiDoc Imaging System (Bio-Rad). Actin immunoblotting was performed with β-Actin (Cell Signaling-13E5, #5125). c-Myc immunoblotting was also performed (Cell Signaling Technology, D84C12). A more detailed co-immunoprecipitation protocol may be viewed at: dx.doi.org/10.17504/protocols.io.v6ke9cw
Total RNA was isolated from 293A cells 36–48 hours after transfection using a PureLink RNA Mini Kit (Thermo Fisher Scientific) with DNase I treatment. The SuperScript VILO cDNA Synthesis Kit (Thermo Fisher Scientific) was used to synthesize single-stranded complementary DNA (cDNA) from 1μg total RNA. Fbw7α, and GAPDH expression were evaluated in triplicate using diluted cDNA as template, gene-specific forward and reverse primers (0.3μM), and Power SYBR Green Master Mix (Thermo Fisher Scientific) in an Applied Biosystems StepOnePlus Real-Time PCR system (Thermo Fisher Scientific).
|
10.1371/journal.ppat.1003824 | Identification of the Virulence Landscape Essential for Entamoeba histolytica Invasion of the Human Colon | Entamoeba histolytica is the pathogenic amoeba responsible for amoebiasis, an infectious disease targeting human tissues. Amoebiasis arises when virulent trophozoites start to destroy the muco-epithelial barrier by first crossing the mucus, then killing host cells, triggering inflammation and subsequently causing dysentery. The main goal of this study was to analyse pathophysiology and gene expression changes related to virulent (i.e. HM1:IMSS) and non-virulent (i.e. Rahman) strains when they are in contact with the human colon. Transcriptome comparisons between the two strains, both in culture conditions and upon contact with human colon explants, provide a global view of gene expression changes that might contribute to the observed phenotypic differences. The most remarkable feature of the virulent phenotype resides in the up-regulation of genes implicated in carbohydrate metabolism and processing of glycosylated residues. Consequently, inhibition of gene expression by RNA interference of a glycoside hydrolase (β-amylase absent from humans) abolishes mucus depletion and tissue invasion by HM1:IMSS. In summary, our data suggest a potential role of carbohydrate metabolism in colon invasion by virulent E. histolytica.
| Entamoeba histolytica is an intestinal parasite which displays diverse phenotypes with respect to pathogenesis in the human colon. Trophozoites can remain as commensal, without causing evident intestinal damage, or they can destroy the colonic mucosa leading to amoebiasis. Using human colon explants and transcriptome analysis, we investigated the gene expression profile of two E. histolytica strains (virulent and non-virulent) during their contact with the intestinal mucus to gain insights into the molecular basis responsible for amoebic divergent phenotypes. Our results suggest that the virulent E. histolytica, when in contact with the intestinal barrier, specifically increases the rate of gene transcription for enzymes necessary to exploits the carbohydrate resources present in the human colon. Using RNA interference methodologies to knockdown gene expression, our data revealed the potential role of amoebic β-amylase (a glycosydase) in colon invasion and mucus depletion. Our data implies that the ability of an E. histolytica strain to exploit the carbohydrate resources might affect its ability to invasion the intestine.
| In the human colon, mucus acts as a lubricant facilitating the passage of digestive content, protects the underlying epithelium from mechanical stress, and provides a protective barrier against pathogens. Mucin 2 (MUC2) is the major component of the mucus layer. MUC2 is a heavily glycosylated protein, containing more than 100 different glycan chain variants which are responsible for approximately 80% of the MUC2 mass [1]. The extensive glycosylation of MUC2 provides protection to resist proteolytic activities. The MUC2-related glycans also represent a potential carbon source for microbiota nutrition, mainly in the distal colon where the availability of free carbohydrates is limited. For instance, intestinal commensal bacteria express genes involved in the biodegradation of complex sugars and glycans present in dietary fibers [2] or genes important for degrading the endogenous pool of host glycans, the last offering a permanent nutrient source for the gut microbiota [3]. During infection, pathogens and resident microbiota compete for nutritional metabolites present in the intestinal lumen and therefore changes in carbon availability may alter the equilibrium in the colon ecosystem contributing to the susceptibility to infection.
Entamoeba histolytica is a protozoan parasite residing in the human colon where it feeds on bacteria. In some cases, trophozoites invade the tissue leading to intestinal amoebiasis and, in rare cases, to hepatic amoebiasis. E. histolytica infection is a persistent and worldwide disease that is the third leading cause of mortality due to a protozoan [4]. Most infections with this parasite are asymptomatic since only ∼20% of the cases develop intestinal amoebiasis, which are characterized by colonic mucosa invasion and tissue destruction. Trophozoites have been isolated from symptomatic and asymptomatic patients. E. histolytica HM1:IMSS, isolated from a patient suffering from amoebic dysentery, is a virulent strain routinely used to reproduce the main features of intestinal [5] and hepatic amoebiasis [6] in experimental models. Another strain, E. histolytica Rahman, was isolated from an asymptomatic carrier, it is unable to growth in animals due to its inherent phenotype and is classically referred as a non-virulent strain [7]. Analysis of the 5.8S rRNA sequences indicates that Rahman belongs to E. histolytica species [8], nonetheless, the Rahman strain presents a reduced cytotoxicity towards epithelial cells in vitro [9], does not form liver abscesses in animal models [10], exhibits defects in phagocytosis, and shows significantly reduced virulence in a human intestinal xenograft model of amoebic colitis [11]. A genomic hybridization study comparing HM1:IMSS and Rahman strains revealed that only 5 out of 1817 genes studied are significantly divergent [12]. At the protein level, important differences in Rahman have nevertheless been described including a truncated glycan chain of the proteophosphoglycan coating the surface [13], a decreased level of both peroxiredoxin [11] and the light subunit of the Gal/GalNAc lectin [9]. Several studies have also attempted to identify genes whose expression correlates with a virulent phenotype by comparing the transcriptomes of both strains under culture conditions [14], [15]. Although they highlighted changes in multiple pathways during parasite axenic growth, no clear explanation has been given to account for their differences in virulence.
To gain insights into the molecular basis of phenotypic differences between E. histolytica HM1:IMSS and Rahman during their interaction with the intestine, we took advantage of the human colon ex vivo model of amoebiasis [5]. Their interaction with the human colon explants was investigated by analysing the morphological changes of the mucosa architecture. We then performed a gene expression analysis for each strain and made comparisons between their transcriptomes. We identified (i) genes that are constitutively expressed in each strain in the two different environments (i.e. in axenic culture or human colon explants), (ii) transcripts specifically upregulated in each strain upon contact with human colon explants, and (iii) transcripts commonly modulated in both strains upon contact with human colon explants.
Genes encoding glycolytic enzymes, carbohydrate catabolism enzymes, and genes characterized as virulent factors were identified and exclusively upregulated in HM1:IMSS upon contact with human colon explants. In particular, one of the most upregulated genes in HM1:IMSS is β-amylase, a glycoside hydrolase absent in humans. The potential role of β-amylase in colon invasion was further investigated by knocking down of its encoding gene using double-stranded RNA (dsRNA). Parasites treated with dsRNA were unable to deplete the mucus and subsequently invade the human colon explants. Altogether, our data provides a novel view of how E. histolytica crosses the intestinal barrier and suggests new avenues to understand amoebic pathogenicity.
To investigate the phenotypic differences between the virulent HM1:IMSS and non-virulent Rahman strains during their interactions with the human colon explants, we monitored the ex-vivo invasion of human colon explants from six patients. After 1 or 7 hours (h) of incubation with trophozoites, colon fragments were fixed for histological analysis and longitudinal sections of the tissues were prepared and examined for mucus integrity (Figure 1A) and tissue invasion (Figure 1B). After 1 h of incubation, the protective mucus layer remains intact in all three conditions (Figure 1A). After 7 h of incubation, we observed tissue penetration by the HM1:IMSS trophozoites with a strong depletion of the mucus layer (Figure 1A). Trophozoites were then localized by immunostaining for the Gal/GalNAc lectin (Figure 1B). The HM1:IMSS trophozoites degraded the intestinal epithelium and penetrated into the mucosa as described previously [5], [16]. In contrast to these findings, after 7 h of incubation with the Rahman strain, trophozoites were still at the surface of the explant, no penetration of the mucosa was observed, and the tissue structure remains intact (Figure 1B). We utilized video microscopy to monitor Rahman trophozoites on the explants to ensure that they were still viable during the incubation (Video S1).
To identify gene expression specifically modulated in HM1:IMSS and Rahman strains upon contact with the human colon explants, total RNA was purified from trophozoites in axenic culture and after contact for 1 h with the human colon explants, a time where virulent trophozoites begin to penetrate through the mucus layer of the colonic tissue [5]. The experiment was conducted on six independent human colon explants from six patients and therefore six biological replicates. Each explant was cut into three pieces, the first was incubated with HM1:IMSS trophozoites, the second was incubated with Rahman trophozoites and the third was incubated without trophozoites (as a control for pathophysiology). RNA was purified from the amoebic samples in contact with the tissue as well as from amoebic samples growing in in vitro culture. As a control, RNA was also purified from trophozoites of both strains incubated in Krebs buffer only (the medium for incubation with the human colon explants). RNA was then reverse-transcribed and hybridized with whole genome cDNA microarrays (EH-IP2008, Agilent technologies) as described previously [17]. A total of 54 hybridizations were performed. For transcriptome data analysis we adopted a step-by-step strategy. First, we performed pairwise comparisons (Figure 2A) with statistics computed for each gene and each condition to identify the transcriptome differences between HM1:IMSS and Rahman strains under axenic culture (Comparison 1) and upon contact with the human colon explants (Comparison 4). The pairwise comparisons also identified transcriptome responses specific for each strain when comparing the human colon explant to the respective profiles in axenic culture (Comparisons 2 and 3). The genes that were commonly modulated in TYI or Krebs buffer only were eliminated from the analysis (Table S1 and S2).
In the second part of the analysis, we used a nested statistical approach (Limma package [18]) where values were tested across the comparisons as indicated in Figure 2B. The gene expression profile of ubiquitously expressed genes for HM1:IMSS was defined by the genes upregulated during both in axenic culture and upon contact with the human colon explants, compared to Rahman (Conditions (A/C+B/D)). Notice that in this analysis genes downregulated in HM1:IMSS were also considered since these genes became upregulated in Rahman. Similarly, the gene expression profile of ubiquitously expressed genes for Rahman was obtained by detecting genes upregulated in axenic culture and upon contact with the human colon explants compared to HM1:IMSS (Conditions (C/A+D/B)). Since both strains bind to the mucus, we searched for a gene expression profile reflecting their common responses the mucus contact by comparing the upregulated genes shared by both strains upon contact with the human colon explants compared to axenic culture (Figure 2B, Conditions (D/C+B/A)). Furthermore, we established the gene expression profile of HM1:IMSS genes specifically expressed during mucus contact, composed of upregulated genes in HM1:IMSS upon contact with the human colon explants compared (i) to the axenic culture and (ii) to Rahman upon contact with the human colon explants. We removed genes upregulated in HM1:IMSS in axenic culture compared to both Rahman in axenic culture and Rahman upon contact with the human colon explants (Conditions (B/A+B/D)−(A/C and D/C)). Analogously, we obtained the gene expression profile of Rahman genes specifically expressed during mucus contact, composed of upregulated genes in the Rahman upon contact with the human colon explants as compared (i) to the axenic culture and (ii) to HM1:IMSS upon contact with the human colon explants. We removed genes that were upregulated in HM1:IMSS in the axenic culture and upon contact with the human colon explants (Conditions (D/C+D/B)−(C/A and B/A)). Overall the combined analysis established the gene expression profiles characteristic of the virulent and non-virulent phenotypes.
Statistical evaluation by Principal Component Analysis (PCA) of the expression data showed that each comparison segregates as a distinct pool (Figure 3) indicating that (i) the biological replicates within each comparison showed similar gene expression patterns and (ii) the differences between the comparisons were higher than the individual variability, thereby validating our experimental settings. A stringent statistical threshold for the microarray data analysis was used, detecting a total of 614 genes with significantly modulated expression (Fold Change (FC) ≥2, Bonferroni adjusted p value≤0.05) (Figure 4). Eighty-one upregulated and 59 downregulated transcripts were different between HM1:IMSS and Rahman in axenic culture (Comparison 1) (Figure 2B, Table S3). Upon contact with the human colon explants (Comparison 2), 63 genes were upregulated and 56 were downregulated in HM1:IMSS, compared to axenic culture (Table S4). Comparison 3 indicates that 75 genes are upregulated and 95 downregulated in Rahman upon contact with the human colon explants, compared to axenic culture. Following mucus contact an additional 77 genes are upregulated in Rahman compared to HM1:IMSS (Table S5). Finally, the comparison between HM1:IMSS and Rahman upon contact with the human colon explants (Comparison 4), reveals 133 genes upregulated and 53 genes downregulated (Table S6). The 614 modulated genes were then manually classified into functional categories based on the gene annotation in AmoebaDB. These categories include, adhesion - cell surface molecules, translation - protein maturation, stress response, DNA-RNA regulation, cell signalling, nucleic acid metabolism, subcellular trafficking, oxidoreduction activities, proteolysis, carbohydrate metabolism, lipid metabolism, and cytoskeleton (Table 1).
Genes ubiquitously expressed in Rahman strain (n = 17, Table 2), were defined as those transcripts upregulated in both axenic culture and upon contact with human explants compared to that of HM1:IMSS. In particular, two α-1,3-mannosyltransferases (ALG2) and two Cysteine protease, CP-A8 and CP-A3 were included. CP-A3 has already been associated with non-virulent phenotypes as it is upregulated in the Rahman strain [11] and the non-virulent species, E. dispar [19]. Five genes encoding enzymes involved in lipid biogenesis were also present, including three lecithin:cholesterol acyltransferases that convert free cholesterol into cholesteryl ester, a START lipid binding domain containing protein, and a 1-O-acylceramide synthase. One gene belongs to the cytoskeleton functional category, coronin, as well as 2 genes encoding signalling molecules, were characteristic for the Rahman strain.
We identified 37 genes in Rahman trophozoites specifically upregulated only upon contact with the human colon explants (Table 3). The largest functional group is composed of factors involved in cell signalling such as several genes encoding phosphatases, kinases, a guanine exchange factor, a GTPase, and calcium binding protein 1 (CaBP1). The cysteine protease, CP-A4, is specific to this profile. Five genes encoding proteins that regulate lipid metabolism were found, including a gene encoding a long-chain-fatty-acid-CoA ligase (also called Fatty acyl-CoA synthetase) which catalyses the formation of fatty acyl-CoA, a substrate for β-oxidation and phospholipid biosynthesis [20] and another allele of lecithin:cholesterol acyltransferase. We also observed an increased expression of genes encoding proteins involved in DNA-RNA regulation, including a DNA repair and recombination protein, a DNA-directed RNA polymerase II, Piwi, a 5′-3′ exonuclease, and 1 RNA binding protein.
Upon contact with the human colon explants, the response common to both E. histolytica strains, was defined by 13 genes (Table 4). The adhesion-cell surface molecules class includes the intermediate subunit 2 of the Gal/GalNAc lectin (Igl-2) and 1 newly identified protein containing a fibrinogen-binding domain (EHI_098440). Two cysteine protease-encoding genes were also identified for both strains as being upregulated. CP-A7 and an unannotated CP (EHI_010850) belonging to the peptidase C1A subfamily. Concerning energy metabolism, a gene implicated in lipid metabolism (long chain fatty acid CoA ligase) and 2 genes involved in carbohydrate metabolism were found (α-amylase and UDP-glucose 4-epimerase). A member of the Myb transcription factor family (EHI_008130) and several transcripts encoding signalling molecules were also found.
The specific signature of HM1:IMSS in axenic culture and upon contact with the human colon explants is characterized by 39 transcripts (Table 5). This signature includes several surface associated proteins [21] namely the Gal/GalNAc lectin light subunits Lgl-1 and Lgl-5, the lysine- and glutamic acid- rich protein 1 (KERP1), the serine/threonine/isoleucine-rich protein (STIRP), and the cysteine protease CP-A5. The presence of CP-A5 is important to highlight since its activity is necessary for invasion of the human colon [5], [16]. The fact that we found well-known virulence factors associated with the HM1:IMSS gene signature confirms the relevance of the integrated analysis performed here.
Genes encoding proteins were identified to be important for the amoebic stress response and include heat shock proteins-70 (HSP-70) and HSP-101, a calcium binding protein involved in signalling, two calmodulins, and several GTPases from the AIG protein family. Three proteases-encoding genes also characterized the gene expression profile specific for the HM1:IMSS strain, namely an unannotated Cysteine protease containing a C1-A peptidase domain and a metalloprotease MP-1. Several genes implicated in carbohydrate metabolism, including 5 genes encoding glycolytic enzymes, phosphofructokinase, fructose 1–6 aldolase, and aldose reductase, and two genes encoding β-amylase were found.
HM1:IMSS trophozoites specifically upregulate 40 genes upon contact with human colon explants (Table 6 and 7) and two points are worth to notice in particular. First, it is the upregulation of 6 genes encoding proteins annotated as regulators of nonsense transcripts. They all contain a RNA helicase domain belonging to the super family 1 (SF1). This RNA helicase domain promotes structural transitions of RNA or RNA-protein complexes. We further found Myb 13 (EHI_053000) that belongs to the MybR2R3 family of transcription factors and which has been reported to bind a DNA consensus Myb recognition element in vitro [22]. Second, it is the upregulation of proteins involved in signalling, including a phosphatase, a kinase, 2 Rab GTPases, 3 Ras GTPases, and a cyclin. Genes linked to the stress response were identified and include 2 HSP-70 genes and 2 ubiquitin genes. Furthermore, the 2 genes implicated in sugar catabolism were also upregulated and they encode a starch binding protein (EHI_074010) and another allele of β-amylase (EHI_035700) respectively.
The 5 profiles established above were combined to depict the transcriptomic landscape associated with HM1:IMSS (virulent and intestinal invasive) and the Rahman strain (non-virulent and intestinal non-invasive) phenotypes of E. histolytica. We highlighted in Figure 5 the well-known virulent factors and the metabolic pathways herewith identified. The specific signature for the mucus-invading HM1:IMSS strain is composed of the following 3 profiles: the HM1:IMSS ubiquitously expressed genes (common to culture and mucus), the common gene expression profile of both strains in response to mucus contact, and the gene expression profile induce in response to colon invasion (exclusive to HM1:IMSS and inherent to mucus invasion). Thus the virulent phenotype of E. histolytica associated with HM1:IMSS is characterized by the expression of genes involved in adhesion (Lgl-1, Lgl-5, Igl-2, KERP1, STIRP, putative fibrinogen binding protein), proteolytic activities (MP-1, CP-A5, CP), and carbohydrate metabolism (phosphofructokinase, aldose reductase, fructose aldolase, β-amylase, α-amylase, UDP-glucose isomerase, triosephosphate isomerase, glucose-4-epimerase, 4-α-glucanotransferase, and oligosaccharide-glycosyltransferase).
The specific signature associated to the non-virulent Rahman strain consists of the ubiquitously expressed gene profile (common to culture and mucus) in addition to the common gene expression profile in response to mucus contact and the gene expression profile specifically induce in Rahman in response to colon contact (exclusive to Rahman and inherent to mucus contact). Thus the non-virulent phenotype of E. histolytica associated to Rahman strain is characterized by the independence from adhesion molecules, the activation of genes encoding proteases activities (CP-A3 and CP-A8) distinct from virulent trophozoites, and the importance of lipid metabolism (lecithin: cholesterol acyl-transferase, START protein, 1-O-acylceramide synthase, fatty acid elongase, long chain fatty acid-CoA synthase, serine palmitoytransferase, and Niemann-Pick C1 protein). Since the Rahman strain expresses this particular set of genes, we conclude that it does not favour colonic mucosa invasion.
A striking result of this study is the discovery of specific distinction concerning energy metabolism pathways activated by non-virulent and virulent strains when they are in contact with the mucus layer. Rahman strain is characterized by an increased expression of genes related to lipid metabolism (Table 2 and 3), whereas HM1:IMSS strain is characterized by upregulation of genes encoding proteins involved in carbohydrate metabolism (Table 5 and 7). To test for functional enrichment in genes upregulated in HM-1:IMSS versus Rahman strains during colon invasion, we performed a hyper-geometric test for gene ontology enrichment [23] and gene set enrichment analysis [24] for KEGG pathway [25]. In the hyper-geometric test, carbohydrate catabolic process (GO:0016052), among other carbohydrate metabolism related gene ontology terms, was significantly enriched (Table S7). Moreover, in gene set enrichment analysis for KEGG pathway, Glycolysis/Gluconeogenesis (ehi00010) and, Fructose and mannose metabolism (ehi00051) were also significantly enriched (Table S8).
The results from gene enrichment tests prompt us to take a closer look at the carbohydrate metabolism genes that are significantly upregulated in the HM1:IMSS strain (without fold-change cut-off), 39 additional genes involved in carbohydrate metabolism were found and listed in Table S9. In particular we identified genes encoding enzymes that are potentially involved in carbohydrate retrieval from MUC2: β-galactosidase (EHI_170020) and β-N-acetylhexosaminidase (EHI_148130) and 3 genes involved in the production of glucose-1-phosphate - glycogen phosphorylase (EHI_110120), 2 other alleles of β-amylase (EHI_098200, EHI_148800), and UDP-glucose pyrophosphorylase (EHI_000440). Glycogen phosphorylase catalyses the rate-limiting step in glycogen degradation by releasing glucose-1-phosphate from the terminal α-1,4-glycosidic bond, β-amylase releases maltose from the polysaccharide chain by hydrolysis of α-1,4-glucan linkages, and UDP-glucose pyrophosphorylase that catalyses the formation of glucose-1-phosphate and UDP from UDP-glucose. In addition, among the 11 enzymes involved in the glycolytic pathway, 7 were specifically induced in the HM1:IMSS strain: phosphoglucomutase, aldose reductase, glucose-6-phosphate isomerase, phosphofructokinase, fructose-1,6-biphosphate aldolase, triosephosphate isomerise, and phosphoglycerate mutase (Table S9 and Table 5). A global view of potential activities of these metabolic enzymes accounting for MUC2 degradation by HM1:IMSS during the invasive process is presented in Figure 6.
Based on the sharp increase in β-amylase transcript level (EHI_192590, fold change up to 25) in HM1:IMSS strain and the enrichment of carbohydrate metabolism genes, we opted to further investigate the role of β-amylase during human mucus invasion. The predicted 3D structure of E. histolytica β-amylase using LOMETS software [26] reveals a strong structural homology to the crystal structure of Glycine max β-amylase. Analysis of the β-amylase amino acid sequence by BLAST reveals similarity (42% pairwise identity, E value = 9e−119) with β-amylase of G. max. Importantly, two glutamic acids residues (E185 and E378) involved in the catalytic activity are present at the homologous position in the E. histolytica enzyme (Figure 7A). An additional trans-membrane domain was predicted for E. histolytica β-amylase (Figure 7B). Based on significant protein homologies with β-amylase from plants, we took advantage of an existing antibody against β-amylase, which recognizes these enzymes. The specificity of this commercial antibody was confirmed by expressing the amoebic β-amylase encoding gene (EHI_192590) in bacteria and western blot analysis (Figure S1). We observed that the protein is localized both on the cell surface and at focused locations in cytoplasm by using immunofluorescence on trophozoites (Figure 7 C).
Entamoeba histolytica possesses 8 copies of the β-amylase encoding gene (EHI_009020, EHI_035700, EHI_049700, EHI_148800, EHI_058340, EHI_118440, EHI_192590, EHI_098200) whose protein lengths range from 436 to 444 amino acids. We confirmed this information by taking advantage of RNA-Seq analysis recently performed in our laboratory [27] that the most highly expressed β-amylase genes in HM1:IMSS were EHI_192590, EHI_098200, and EHI118440 (Figure S2). In our microarray experiments, EHI_192590 is the most upregulated compared to Rahman (mucus and culture conditions) and in addition EHI_035700 is only overexpressed in HM1:IMSS during colon invasion. Levels of expression were very low in the Rahman strain and we confirmed by western blot that β-amylases were indeed present in cultured HM1:IMSS strain and highly reduced in Rahman (7.7 fold decrease at the protein level, Figure 7D). In order to gain insights into the role of β-amylase during mucus invasion, we knock down the expression of β-amylase in the HM1:IMSS strain using a dsRNA-based RNA interference approach [28]. We designed a specific dsRNA targeting the transcripts of all 8 copies (see material and methods). Total protein extracts were analysed using western blot after 24 h and 48 h of incubation with the specific β-amylase dsRNA or a control dsRNA (i.e GFP dsRNA). After 48 h of incubation, the β-amylase quantity was decreased by 75.5% (SEM ± 4.6%; n = 3) in comparison to the control (GFP dsRNA-treated trophozoites) without impacting the growth of theses trophozoites (Figure 8A). The viability of dsRNA treated trophozoites was determined upon an hour incubation in Krebs buffer by trypan bleu exclusion test (percentage of cell death was for dsGFP = 11.2±3.4 sd, and for ds β-amylase = 10.8±2.9 sd., n = 3). HM1:IMSS trophozoites with reduced levels of β-amylase were then challenged for human colon invasion. We observed by histological analysis that after 7 h of incubation, tissue invasion by β-amylase dsRNA-treated trophozoites was abolished, while these trophozoites were still associated with the mucus layer (Figure 8B and 8C). In contrast, GFP dsRNA treated parasites (used as a control) depleted the mucus layer and penetrated the lamina propria, as wild-type HMI:IMSS trophozoites. Measurement of the mucus thickness after 7 h incubation in the presence of β-amylase dsRNA-treated trophozoites (133.7 µm SEM ± 2.33 µm) was comparable to the mucus thickness of the tissue control incubated without trophozoites (132.7 µm SEM ± 3.29 µm). However, in the presence of GFP dsRNA treated parasites the mean thickness of the mucus layer was significantly decreased to 13.58 µm (SEM ± 1.35 µm, p<0.0001).
E. histolytica colonizes the human gut mainly as a parasite. Only 1 in 5 infections leads to disease [29]. The classical view of amoebic infection outcome is that the virulence of E. histolytica is the consequence of the interactions between host, parasite, and environmental factors. Although the evidence supporting the phenotypic conversion of a strain from non-virulent to virulent is currently lacking, it is admitted that a latent period between infection and disease is due to parasite adaptation to the host via modifications in gene expression [30]. However, E. histolytica strains isolated from healthy asymptomatic carriers do not reproduce infection in animals implying that there is an unidentified mechanism regulating gene expression in addition to adaptation. Using the human colon explant model [5], we compared the transcriptome modulation upon mucus contact of E. histolytica strains isolated from asymptomatic (Rahman) or symptomatic (HM1:IMSS) patients. Notice that only one representative virulent strain (HM1-IMSS) and only one representative non-virulent strain (Rahman) were compared in this study. Trophozoites from these isolates has been in culture for decades and likely may harbour differences unrelated to virulence, however these represent the best characterized amoebic isolates from genomics and biological point of views. Indeed, non-virulent Rahman trophozoites bound to the mucus but neither depleted the protective barrier nor invaded and destroyed the tissue, in contrast to HM1:IMSS virulent trophozoites. The transcriptome analysis identified genes: (i) ubiquitously expressed in each strain, (ii) common to the 2 strains interacting with human mucus, and (iii) specifically expressed in response to colon contact. The transcriptome of amoebae able to invade the mucus was characterized by several virulence factors (the Gal/GalNAc lectin, STIRP, KERP1 and CP-A5) already described for their participation in the pathological process or over-expressed in virulent amoebic strains [21]. Also identified were proteins such as the SHAQKYF (Myb 13) transcription factor regulating the expression of genes related to signal transduction, vesicular transport, heat shock response and virulence [22] as well as transcripts linked to stress responses and to signalling pathways, including the GTPase AIG1 known to be expressed during colonization of the mouse intestine [31] and in pathogenic E. histolytica [32].
Besides the involvement of the above cited virulence factors, the remarkable feature of colon explant invasion concerns the changes in expression of genes encoding enzymes involved in the carbohydrate metabolism. In addition to several enzymes implicated in the production of glucose-1-phosphate, upregulation of genes encoding the majority of enzymes involved in glycolysis was characteristic to mucus depletion. Therefore, we hypothesized that carbohydrate metabolism might play a role in sustaining the invasive behaviour of the virulent strain during intestinal invasion. Indeed when accessibility of polysaccharides in the lumen is decreased and glucose levels are low, virulent E. histolytica might be able to adapts its transcriptome to proficiently utilized host mucus glycans as its carbon source. Here we proposed a sequential mode of MUC2 degradation, involving the release of oligosaccharides from MUC2 by glycosidases (e.g. beta-galactosidase and beta-N-acetylhexosaminidase, upregulated in virulent strain during colon invasion), and followed by cleavage of the exposed protein backbone by proteases (Figure 6 and Figure 9). We speculate β-amylase might play a role in breaking down the already released oligosaccharides into sugars as carbon sources for energy production. Thus, the reduced β-amylase activity in the dsRNA treated strain might hamper the utilization of MUC2 as the carbon source for glycolysis. The upregulation of multiple genes in the glycolytic pathway in the virulent strain during colon invasion correlates with this speculation and we interpret the upregulation of these genes in the virulent strain as the consequence of utilization of MUC2 as the carbon source. This hypothesis supports our previous findings showing that E. histolytica virulence increased when in the presence of a low glucose environment [33]. This scenario also fits well with previous findings indicating that E. histolytica depletes colonic mucin oligosaccharide side chains by using a glycosidase activity [34]. Following the breakdown of MUC2 oligosaccharides, the protein backbone is no longer protected and may be degraded by specific amoebic proteases as has been previously demonstrated [35], [36].
In this work we highlighted β-amylase, because it is a protein absent from the mammalian kingdom proteome and is strongly overexpressed (25-fold) in HM1:IMSS strain. The enzyme β-amylase acts on the α-1,4 glycosidic bonds and catalyses the breakdown of starch into maltose (a glucose dimer). Using a dsRNA-based strategy, we decreased β-amylase protein levels in HM1:IMSS strain, and resulted in reduced mucus layer depletion and mucosa invasion. The β-amylase activity and its substrate in the invasive process have yet to be determined. The fact that β-amylase does not exist in the human genome makes this enzyme a potential therapeutic target to inhibit amoebic intestinal invasion.
Entamoeba histolytica typically feeds on bacteria in the intestinal lumen. Microbial inhabitants of the gut, which can also have an influence on metabolic processes, such as energy extraction from food and host mucus glycan, can be considered as an environmental factor that contributes to amoebic maintenance in the colon lumen and further in the pathology. Our hypothesis (Figure 9) is in line with findings obtained from bacteria resident in the mucus layer, in which they are capable to adapt their gene expression to gut diet content. For example gene expression profiling of Bacteroides thetaiotaomicron, revealed that rich polysaccharide diets are associated with a selective upregulation of glycoside hydrolases (e.g. xylanases, arabinosidases, and pectate lyase). These bacteria also upregulate genes encoding enzymes involved in delivering glucose to the glycolytic pathway [3]. When these bacteria are in the presence of a unique glucose diet devoid of polysaccharides, the induction of a different subset of glycoside hydrolases is activated including enzymes necessary for retrieving carbohydrates from mucus glycans, as well as enzymes that increase accessibility to host glycans [3]. We proposed that the ability of virulent E. histolytica trophozoites to exploit carbohydrate resources derived from the human mucus might be one of the factors powering intestinal amoebiasis.
Healthy segments of human colon were obtained from patients undergoing colon surgery. Patient-written informed consent was obtained at Foch Hospital and the data were analysed anonymously at the Pasteur Institute. Tissues were processed according to the French Government guidelines for research on human tissues and the French Bioethics Act, with the authorization from the “comité de protection des personnes, Ile de France VII” and the “Institut Pasteur Recherche Biomedicale” investigational review board (RBM./2009.50).
RNAseIII-deficient Escherichia coli strain HT115 (rnc14::ΔTn10) was grown in LB-broth containing ampicillin (100 µg/ml) and tetracycline (10 µg/ml). Entamoeba histolytica HM1:IMSS is a virulent strain and E. histolytica Rahman is a non-virulent strain [7]. The HM1-IMSS strain was isolated in 1967 from a colonic biopsy of rectal ulcer from adult human male with amebic dysentery, Mexico City, Mexico. The HM1-IMSS was deposited in the American strain collection (ATCC® 30459™) and it is a gift of Professor Ruy Perez Tamayo (UNAM, Mexico). To maintain virulence, the HM1-IMSS strain has been passed through the liver of hamsters (Male Syrian golden hamsters Mesocricetus auratus) (roughly 174 passages since isolation until experiments were done). The procedure applied for animal infection was previously described [6], trophozoites were isolated from the liver abscesses after 7 days of intraportal inoculation (4 animals), mixed and further growth in axenic conditions. The Rahman strain is non-virulent [7] and is unable to growth in animals due to its inherent phenotype. The Rahman strain has been maintained in axenic culture since isolation in 1978 (with undetermined periods of frozen preservation) and it is a gift of Professor David Mirelman (Weizmann Institute, Israel). Trophozoites of both strains were grown axenically in TYI-S-33 medium at 37°C [37] and harvested during the exponential growth phase.
Previous experimental published conditions were used for handling human colon pieces [5]. Briefly, 1.6×105 trophozoites were added to the luminal face of the colon and incubated in Krebs buffer at 37°C for 1 and 7 h. After 1 h of incubation, mucus interacting trophozoites were collected by pipetting the mucus layer and 1 ml of Trizol was added. After 7 h of incubation, tissue fragments were fixed either in Carnoy fixative or in PFA (4%) and included into paraffin. PFA-fixed tissue sections were immunostained with a 1∶200 diluted rabbit antibody recognizing the Gal/GalNAc lectin [5] Sections from Carnoy-fixed tissue were stained with Alcian blue to visualize the mucus layer [38]. For each experiment, a representative histology image was taken.
For the measurement of mucus layer thickness, transverse sections were stained with Alcian blue stain. Light microscope images (NIKON, Eclipse E800) were analysed with ACT-1 software (NIKON). The mucus layer thickness was measured at three points of twenty different sections for three different patients (60 measurements for each condition). The mean of these measurements was considered as the mucus thickness for each condition. Statistical analysis was performed using GraphPad Prism software version 5.0b (GraphPad Software Inc). An unpaired, two-tailed student T-test was performed. Differences being considered as significant if P<0.05. Data are expressed as mean ± SEM.
Entamoeba histolytica HM1:IMSS or Rahman trophozoites (1.6×105) grown in axenic culture were lysed with Trizol reagent (Invitrogen), and total RNA isolated according to the manufacturer's protocol. RNA from mucus-interacting trophozoites was purified by gently scratching-off the mucus layer containing the trophozoites after 1 h of incubation. Trizol was added to the samples and RNA purification was performed. RNA was analysed for integrity and the concentration determined by capillary electrophoresis using the Agilent Bioanalyzer 2100 RNA nanochip Assay (Agilent Technologies). RNA from mucus-interacting trophozoites showed a mixture of amoebic and human RNA (up to 30%). Thus RNA isolated from human epithelial cells was used as a control to evaluate potential cross-hybridization of human transcripts in the subsequent experiments. Agilent microarrays EH-IP2008, scanning the entire amoebic genome, were used as previously described [17]. Six biological replicates were performed with amoebic strains grown in culture or incubated with the colon explants. Dye swap hybridizations were performed for the six biological replicates leading to a total of 12 hybridizations for each of the four conditions: Rahman in colon vs culture, HM1:IMSS in colon vs culture, Rahman vs HM1:IMSS in the colon, and Rahman vs HM1:IMSS in culture. In addition, one technical replicate was performed for one of the biological replicates and two self-self hybridizations were conducted. The resulting fluorescence signals were used to tune the scanner for the set of arrays. Probes cross-hybridizing to human RNA were identified and removed from the analysis (data not show). In addition, since prior to colon mucus contact the parasites were incubated in Krebs buffer we also determined gene expression changes in Krebs buffer; the modulated genes in each strain were removed before the analysis (Data in Table S1 and S2). The experiment finally yielded 54 competitive hybridizations. The whole data set was submitted to the ArrayExpress database (Accession number: E-MTAB-1201).
A Principal Component Analysis of the whole microarray dataset was first carried out with Partek (http://www.partek.com/software) on the raw data. Microarray data statistical analyses were carried out with the R software (http://www.R-project.org) and Bioconductor packages (http://www.bioconductor.org). Our experiment follows a multifactorial design that includes two strains (HM1:IMSS and Rahman) in two different growth conditions (colon and culture). Linear models are well suited for the analysis of such designs, since they allow a global analysis of the whole dataset. Global effects, such as strain or growth condition effects can be measured, as well as differences between particular pairs of combinations of factors called contrasts, for example, the difference between Rahman and HM1:IMSS in colon condition. As Limma implements linear models for microarray data analysis, it was chosen for the present study (Limma package [19]). A Loess normalization was first performed on the 48 microarrays in order to render expression ratios comparable. The full experimental design was described through a design matrix (as explained in the Limma vignette) which is a binary matrix composed of (0, 1, −1) used by the linear model. The matrix makes a formal correspondence between arrays and pairs of conditions that have been hybridized. Then, a contrast matrix was created. It contains the list of comparisons that we wish to test with the linear model, namely HM1:IMSS – colon vs culture, Rahman – colon vs culture, HM1:IMSS vs Rahman – colon, HM1:IMSS vs Rahman – culture, HM1:IMSS vs Rahman, and colon vs culture. The moderated t-test associated with the empirical Bayes method (33was first applied to the hybridization value of each probe and the resulting p-values were further adjusted using a Bonferroni correction [39]). Finally, a median log-ratio was computed taking all probes in consideration in the case of genes represented by more than one probe on the array. An equivalent analysis was performed on a gene basis using the same design and contrast matrices and the same p-value adjustment. Only genes with an adjusted p-value lower than 0.05 and a fold change higher than 2 were considered for further analysis. Notice that according to this microarray analysis, upregulated and downregulated genes were taken into consideration. Thus the final fold changes values correspond to the ratio of changes between the two strains (i.e numbers from HM1:IMSS versus numbers from Rahman and vice versa). In other words genes appearing upregulated for HMI:IMSS strain are down regulated for Rahman strain counterpart and conversely genes upregulated for Rahman are downregulated for HM1:IMSS.
Gene ontology and KEGG pathway annotations were retrieved from AmoebaDB v3.0 [40] and KEGG database [25]. To test for gene ontology enrichment, genes that are significantly upregulated (FDR<0.05, without fold-change cut-off) in HM1:IMSS comparing with Rahman during colon invasion were used as the foreground to test against the whole gene background using the hyper-geometric test implemented in FUNC package [23]. To test for KEGG pathway enrichment, the moderated fold-change of all genes in HM1:IMSS versus Rahman during colon invasion was used as the input into GSEA package [24]. Statistical significance was determined according to the default false discovery rates of the packages (5% in FUNC and 25% in GSEA).
The structure of β-amylase from E. histolytica (Accession number EHI_192590) was predicted using LOMETS [41] which identifies β-amylase from Glycine max (Accession number: BMY1; 547931 BMY1) as the best-hit template (Z score = 102, 377). Protein domains were identified with SMART and defined EHI_192590 as a member of glycoside hydrolase family 14 which comprises β-amylase (EC 3.2.1.2). The amino acid sequence of the full-length homolog in Entamoeba was aligned by CLUSTALW software with β-amylase from G. max. The N-terminal tail of E. histolytica β-amylase was predicted as a transmembrane domain using TMHMM plugin of the Geneious software.
To construct the dsRNA expression vectors, DNA fragments of the E. histolytica β-amylase gene (position +694 to position +1187, GenBank Accession number: EHI_192590) and the entire green fluorescent protein (GFP) coding sequence (GenBank Accession number: U73901) were amplified by PCR and subcloned into the TA-cloning vector pCR2.1-TOPO (Invitrogen). DNA inserts were excised from these constructs with restriction enzymes (KpnI and BamHI for GFP; KpnI and Bgl II for β-amylase) and cloned into the MCS of the L4440 plasmid vector that is bidirectionally flanked by T7 promoters. The resulting plasmids construct L4440-β-amylase and L4440-GFP were verified by restriction analysis and DNA sequencing. To purify dsRNA and perform soaking experiments we followed the procedure described previously [28].
A polyclonal rabbit anti-β-amylase antibody raised against the full-length β-amylase of Ipomoea batatas (sweet potato) was purchased from Abcam (ab6617). The specificity of this antibody was assessed by expression of amoebic β-amylase encoding gene (EHI_192590) in Escherichia coli (BL 21 strain). To this end the gene was amplified from the amoebic genome (forward primer: TACCATGGATGTTATTAACACTATGTTTTATATCAATAGC; reverse primer: ATCTCGAGTCTCATTGAATTAACAAATGAACAA) and cloned in MCS of pET28 vector. The insert was verified by DNA sequencing and upon expression in bacteria the recombinant protein was identified by western blot (Figure S2). For western blot analysis of amoebic extracts, the loaded protein amounts were normalized using an anti-actin C4 monoclonal antibody (ref: 08691001, MP Biomedicals) and secondary HRP-antibodies (MP Biomedicals) were used. Trophozoites submitted to dsRNA soaking experiments were collected to prepare crude extracts as previously described [42]. Crude extracts (4×104 cells/lane) were resolved by SDS-PAGE, transferred to PVDF membranes and incubated with specific antibodies and ECL Plus reagent (GE Healthcare Bio-sciences) for chemiluminescence detection. Semi-quantitative analysis of light emission from probed nitrocellulose membranes was carried out from scanned autoradiographs using Quantity one software (BioRad) and protein abundance was normalized with actin values.
Trophozoites were grown axenically in TYI-S-33 medium at 37°C and then centrifuged for 5 min at 550× g during the exponential growth phase. The pellet was fixed in 4% paraformaldehyde at 37°C for 15 min and permeabilized or not with Triton X-100. Cells were incubated in 1% PBS/BSA to avoid non-specific labelling. The primary antibody againstβ-amylase 1/1000 (Abcam® (ab6617)) was then deposited onto the coverslip and incubated in a humid chamber for 2 h at 37°C. The coverslips were washed in 1% PBS/BSA and the secondary antibody coupled to Alexa-568 (Molecular Probes, Invitrogen) 1/200 was added to the coverslip and incubated in a humid chamber for 30 min at 37°C. The coverslips were washed and the slides were then mounted using VectaShield mounting medium, sealed and conserved at 4°C until confocal microscopy analysis. The slides were analysed using a Zeiss LSM 710 Confocal Microscope and LSM software.
|
10.1371/journal.pcbi.1001108 | A Compact Statistical Model of the Song Syntax in Bengalese Finch | Songs of many songbird species consist of variable sequences of a finite number of syllables. A common approach for characterizing the syntax of these complex syllable sequences is to use transition probabilities between the syllables. This is equivalent to the Markov model, in which each syllable is associated with one state, and the transition probabilities between the states do not depend on the state transition history. Here we analyze the song syntax in Bengalese finch. We show that the Markov model fails to capture the statistical properties of the syllable sequences. Instead, a state transition model that accurately describes the statistics of the syllable sequences includes adaptation of the self-transition probabilities when states are revisited consecutively, and allows associations of more than one state to a given syllable. Such a model does not increase the model complexity significantly. Mathematically, the model is a partially observable Markov model with adaptation (POMMA). The success of the POMMA supports the branching chain network model of how syntax is controlled within the premotor song nucleus HVC, but also suggests that adaptation and many-to-one mapping from the syllable-encoding chain networks in HVC to syllables should be included in the network model.
| Complex action sequences in many animals are organized according to syntactical rules that specify how individual actions are strung together. A critical problem for understanding the neural basis of action sequences is how to derive the syntax that captures the statistics of the sequences. Here we solve this problem for the songs of Bengalese finch, which consist of variable sequences of several stereotypical syllables. The Markov model is widely used for describing variable birdsongs, where each syllable is associated with one state, and the transitions between the states are stochastic and depend only on the state pairs. However, such a model fails to describe the syntax of Bengalese finch songs. We show that two modifications are needed. The first is adaptation. Syllable repetitions are common in the Bengalese finch songs. Allowing the probability of repeating a syllable to decrease with the number of repetitions leads to better fits to the observed repeat number distributions. The second is many-to-one mapping from the states to the syllables. A given syllable can be generated by more than one state. With these modifications, the model successfully describes the statistics of the observed syllable sequences.
| Complex action sequences in animals and humans are often organized according to syntactical rules that specify how actions are strung together into sequences [1], [2]. Many examples are found in birdsong. Songs of birdsong species such as Bengalese finch [3]–[5], sedge warbler [6], nightingale [7], and willow warbler [8] consist of a finite number of stereotypical syllables (or notes) arranged in variable sequences. Quantitative analysis of the action syntax is critical for understanding the neural mechanisms of how complex sequences are generated [1], [3], [5], [9], [10], and for comparative studies of learning and cultural transmissions of sequential behaviors [11].
Pairwise transition probabilities between syllables are widely used to characterize variable birdsong sequences [3], [4], [7], [8]. This is equivalent to using the Markov model to capture the statistical properties of the syllable sequences. The Markov model is a generative statistical model of sequences, and consists of a set of states. Here the states are mathematical abstractions; they can correspond to concrete neural substrates in specific neural mechanisms of birdsong generation. There is a start state and an end state, which correspond to the start and the end of the sequences, respectively. For each syllable, there is one corresponding state. A state sequence starting from the start state and ending at the end state is produced through probabilistic transitions from one state to the next, and the corresponding syllable sequence is generated. The transition probabilities between the states depend only on the state pairs, and are set to the observed pairwise transition probabilities of the associated syllables. More sophisticated models allow chunks of fixed syllable sequences to be associated with state transitions, with a possibility that a syllable appears in different chunks [5], [12], [13]. However, no detailed statistical tests of these state transition models have been performed, and their validity as quantitative descriptions of the birdsong syntax remains unclear.
In this paper, we analyze the songs of Bengalese finch. We demonstrate that the Markov model fails to capture the statistical properties of the observed sequences, including the repeat number distributions of individual syllables, the distributions of the N-grams (sequences of length N) [14] and the probability of observing a given syllable at a given step from the start of the sequences. We introduce two modifications to the Markov model and show that the extended model is successful in describing the syntax of the Bengalese finch songs. The first modification is adaptation. Syllable repetitions are common in the Bengalese finch songs. Allowing the repeat probabilities of syllables to decrease with the number of repetitions leads to a better fit of the repeat number distributions. The second modification is many-to-one mapping from the states to the syllables. A given syllable can be generated by more than one state. Even if the transitions between the states are Markovian, the syllable statistics are not Markovian due to the multiple representations of the same syllables. The resulting model, which we call a partially observable Markov model with adaptation (POMMA), has history-dependent transition probabilities between the states and many-to-one mappings from the states to the syllables. The POMMA successfully describes the statistical properties of the observed syllable sequences. It is consistent with the branching chain network model of generating variable birdsong syntax, in which syllable-encoding chain networks of projection neurons in the premotor song nucleus HVC are connected in a branching topology [10], [15].
Spontaneous vocalizations of two Bengalese finches were recorded in an acoustic chamber using a single microphone over six (Bird 1) and five (Bird 2) days, respectively. Vocal elements (, Bird 1; , Bird 2) were isolated from the recorded pressure waves (Materials and Methods). In the following, we first present the analysis of Bird 1 and then Bird 2.
For Bird 1, the vocal elements were clustered into 25 types according to the similarities of their spectrograms (Materials and Methods). We identified seven types of vocal elements as song syllables (Figure 1a, for syllables A to G, respectively). The rest were call notes (14 types; 7 examples shown in Figure 1b; C1 and C2 were the the most frequent call notes with , respectively) and noise. The song syllables were distinguished by rich structures in the spectrograms and tight distributions of the durations (), (Figure1a), and frequently appeared together in long sequences (sequence length mean ) with small inter-syllable gaps () (Figure 1c–d). The gaps between the consecutive syllables were filled with silence or small noisy fluctuations; no call notes or unidentified vocal elements were in them. In contrast, the call notes had broad or simple spectra and more variable distributions of the durations (), and appeared in short sequences (sequence length mean ). All consecutive sequences of the song syllables with inter-syllable gaps smaller than 200ms were assigned as song sequences. Additionally, syllable E (Figure 1a), which predominantly appeared at the start of the sequences obtained above, was assigned as a start syllable such that whenever syllable E appeared for the first time and was not following another E, a new song sequence was started. Thus, a long sequence containing k non-continuous E's in the middle was broken into k+1 song sequences. Altogether, we ended up with 1921 song sequences. Sequences of call notes can precede or follow song sequences, and these call notes were considered to be introductory notes.
A simple statistical model of the song sequences is the Markov model, which is completely specified by the transition probabilities between the syllables. For each syllable, there is a corresponding state; additionally, there is a start state (symbol s) and an end state (symbol e), as shown in Figure 2a. We computed the transition probability pij for the state Si associated with syllable i to the state Sj associated with syllable j, from the observed song sequences as the ratio of the frequency of the sequence ij over the total frequency of syllable i. Transitions with small probabilities (pij<0.01) were excluded.
To evaluate how well the Markov model describes the statistics of the observed song sequences, we generated 10000 sequences from the model, and compared three statistical properties of the generated sequences and the observed sequences. The method of sequence generation is as follows. From the start state, one of three states associated with syllables C, E, D can follow with probabilities , , , respectively (Figure 2a). A random number r is uniformly sampled from 0 to 1. If , is selected (the state following the start state is ), and the generated sequence starts with C. If , is selected (), and the sequence starts with E. If , is selected (), and the sequence starts with D. From the selected state , the next state can be selected similarly according to the transition probabilities from . This process of sampling random numbers and selecting the next state and syllable is continued until the end state is reached, generating a specific syllable sequence. Examples of the generated syllable sequences are shown in Figure 2b.
The first statistical property to be compared was the distribution of the syllable repeats. Except syllable F, all syllables appeared in repetitions, and the number of repeats were variable. For each syllable, we constructed the probability distribution of the repeat numbers by counting the frequencies of observing a given number of repeats in the observed song sequences. The distributions are shown as black curves in Figure 3a. We also constructed the repeat number distributions from the sequences generated from the Markov model. These are shown as cyan curves in Figure 3a. For syllables E and G, the comparisons are favorable. However, for other syllables the distributions clearly disagree. To quantify the difference between two distributions and , we defined the maximum normalized difference d, which is the maximum of the absolute differences divided by the maximum values in the two distributions, i.e. . The d-values for syllables A, B, C, D, E are , respectively. The major difference is that, for syllables A, C, D, the observed distributions peak at repeat number 4, 2, 2, respectively, while the generated distributions are decreasing functions of the repeat numbers. Indeed, if the probability of returning to state from itself is a constant , the probability of observing repeats of the associated syllable is , which is a decreasing function of . Therefore the Markov model is incapable of producing repeat number distributions having maxima at .
The second statistical property to be compared was the N-gram distribution. An N-gram is a fixed subsequence of length . For example, syllable sequences EC and AA are 2-grams; ECC and AAA are 3-grams; etc. We constructed the probability distributions for 2- to 7-grams in the observed song sequences by counting the frequencies of a given subsequence. The results are shown in Figure 4a as black curves, with the N-grams sorted according to decreasing probabilities. We also computed the probability distributions of the corresponding N-grams in the generated sequences. The results are shown in Figure 4a as cyan curves. The distributions for 2-grams agree very well, which is expected, since the Markov model was constructed with the transition probabilities, which are equivalent to the 2-gram distributions. The distributions are quite different for 3- to 7-grams, with -values ranging from 0.26 to 0.93 (Figure 4a).
The final statistical property to be compared was the step probability of the syllables, which is defined as the probability of observing a syllable at a given step from the start. The step probabilities for all syllables computed from the observed song sequences, as well as the step probability of the end symbol , which describes the probability of observing that a sequence has ended at or before a given step, or equivalently, the cumulative distribution function of the sequence length, are plotted as black curves in Figure 5a; and those from the generated sequences are plotted as cyan curves. The comparison for syllable E is quite good (). But the differences between the probabilities for other syllables and the end symbol are large, as indicated by the d-values ranging from 0.11 to 0.61.
Because the number of the observed song sequences is finite, even a perfect statistical model that would exactly reproduce the Bengalese finch songs cannot lead to zero d-values when compared to the observed distributions. One way of assessing the goodness of fits is to use benchmarks for the d-values created from the observed syllable sequences. We split the observed sequences into two groups by randomly assigning each sequence with a probability 0.5. One group is considered as generated by a perfect statistical model and compared against the other group. For each group we computed the repeat number distributions, the N-gram distributions, and the step probability distributions. The distributions from the two groups were compared to obtain the d-values. We performed the random splitting 500 times and constructed distribution profiles for each d-value. These profiles characterized the fluctuations of the d-values due to the finite number samplings of the observed sequences. For each d-value, we chose the point in the profile as the benchmark. This means that the probability that the d-value is smaller than the benchmark is 0.95. The benchmarks are plotted as gray vertical bars in Figure 6. A good statistical model of the syllable sequences should produce d-values smaller than the benchmarks or close to them. The d-values obtained from the Markov model, plotted as the cyan curves in Figure 6, are mostly far beyond the benchmarks. It is clear that the Markov model fails to capture the statistical properties of the songs of Bird 1.
One way of extending the Markov model is to allow the transition probabilities to change depending on the state transition history. There are many possible formulations of such dependence. Adaptation, in which the transition probabilities are reduced as the state transitions are repeatedly revisited, is one formulation motivated by the observation that repeated activations of synapses and neurons reduce their efficacy [16]–[18].
Ideally, all transition probabilities should be subject to dynamical changes depending on the histories of the state transitions in the Markov model. But such a model is difficult to analyze. We therefore considered a simple model in which only the return probabilities of the states from themselves are adaptive. In particular, the return probability of a state is reduced to after repetition of the associated syllable. The transition probabilities to all other states are mutiplied by a factor to keep the total probability normalized. Here is the adaptation parameter, and is the return probability when . The probabilities recover to original values once the dynamics moves on to other states. In this Markov model with adaptation, the probability of observing repetitions is given by (Materials and Methods). We fitted the parameters and for the states with self-transitions in the Markov model (Figure 2a) using the repeat number distributions in the observed song sequences. The resulting model is shown in Figure 2b, which is identical to the Markov model (Figure 2a) except that the return probabilities for the states associated with syllables A, C, D, E are adaptive, with , respectively. Fittings for syllables B and G did not lead to an adaptive model (), so the associated return probabilities are unchanged.
To evaluate the Markov model with adaptation, we again generated 10000 song sequences and compared the repeat number distributions, the N-gram distributions, and the step probabilities to the observed song sequences. The generation procedure was the same as in the original Markov model, except that the return probabilities were adaptive as prescribed above. The repeat number distributions, shown as green curves in Figure 3b, are much improved compared to the Markov model. In particular, the peaked distributions of syllables A, C, D are well reproduced. This demonstrates that the adaptation is capable of producing peaked repeat number distributions. Adaptation did not improve the comparisons of the N-gram distributions (Figure 4b). Adaptaion improved the comparisons of the step probabilities for syllables C, D, F but not for syllables A, B, D and the end symbol (Figure 5b). The d-values (green curves in Figure 6) compared to the benchmarks confirm these observations. The Markov model with adaptation is a better statistical model for song sequences of Bird 1 than the Markov model; however, it is still not capable of accurately describing all statistical properties.
In the Markov model and its extension with adaptation, each syllable is associated with one state. Hence the number of states is equal to the number of the syllables, plus two if we count the start and end states (we will exclude the start and end states when we count the number of states in a model). However, it is possible that there is more than one state corresponding to one syllable. This many-to-one mapping from the states to the syllables enables the state transition models to describe more elaborate statistical properties of syllable sequences [10]. With the many-to-one mapping, the number of states can be larger than the number of syllables. When this is the case, some of the states are “hidden”, and cannot be simply deduced by counting the number of syllable types. This kind of model is often referred to as “partially observable Markov model” (POMM) [10], [19], and is a special case of the hidden Markov model (HMM) in which each state is associated with a single symbol. We tested whether introducing many-to-one mapping in addition to the adaptation, which leads to a “partially observable Markov model with adaptation” (POMMA), would better explain the statistical properties of the observed song sequences.
To derive a POMM from observed sequences, we developed a state merging method, in which the sequences are translated into a POMM with tree transition structure, and the states are merged if they have equivalent statistical properties and deleted if they are rarely reached (Materials and Methods). To incorporate adaptation to syllable repetitions, we first derived a POMM with the non-repeat versions of the song sequences, in which the repeats of syllables were ignored but the number of repeats were recorded. For example, the non-repeat version of a song sequence ECCDDFBBGBAA is E(1)C(2)D(2)F(1)B(2)G(1)B(1)A(2), where the repeat numbers are in the parenthesis. While creating the tree-POMM and merging the states, the repeat numbers were kept track of, so that the repeat number distribution for each state could be constructed. After following the POMM derivation procedure, there were 18 states in the model. The resulting model was evaluated by generating 10000 sequences following the state transitions from the start state. If a state with no repeat syllable was reached, the syllable associated with the state was generated. If a state with repeat syllables was reached, a repeated sequence of the syllable was generated with the repeat number sampled from the repeat number distribution associated with the state. The sequence stopped if the end state was reached. The generated sequences were compared with the observed sequences for the repeat number distributions of each syllable, the N-gram distributions, and the step probabilities of each syllable and the end symbol. We further tested deletion of each state and mergers of all pairs of states with the same syllables, while monitoring the d-values of the three statistical properties. The deletions and mergers were accepted if the d-values fell below the benchmarks or they were less than the corresponding d-values of the model with the 18 states. The resulting POMM, shown in Figure 7a, has 11 states. Syllables B, C, D, G are associated with two states each, and syllables A, E, F have one associated state each.
We next modeled the repeat number distributions in each state with the adaptation model described previously. For some states, the adaptation model was not adequate to fit well the repeat number distributions (cosine-similarity of the distributions with best fitting parameters; Eq.(1) in Materials and Methods). In such a case, the state was split into two serially connected states . The transitions and associated probabilities to were set to , and and emitted to the same states and probabilities as . has a self-transition with probability and adaptation parameter , while has no self-transition but has a transition probability to . The repeat number distribution with these parameters is given by (Materials and Methods). The parameters were fit with the nonlinear least square fitting procedure. Each state-splitting thus introduced one more state and one more parameter to the model, and was adequate to fit well the observed repeat number distributions when necessary. The resulting POMMA is shown in Figure 7b. Three states associated with syllables A, C, D were split. Altogether, there are 14 states, and the number of states for syllables A to G are 2, 2, 3, 3, 1, 1, 2, respectively.
We generated 10000 syllable sequences from the POMMA (examples shown in Figure 7c), and compared with the observed song sequences the repeat number distributions (Figure 3c), the N-gram distributions (Figure 4c), and the step probabilities (Figure 5c). The comparisons are excellent. All d-values fall below or close to the benchmarks, as shown with the red curves in Figure 6. In contrast, the d-values for the Markov model are mostly far beyond the benchmarks, as shown with the cyan curves in Figure 6. The d-values for the Markov model with adaptation are also larger than those for the POMMA, as shown with the green curves in Figure 6. In particular, the d-values for the N-gram distributions are far beyond the benchmarks and the d-values of the POMMA. Thus, the POMMA is a much better model than the Markov model or the Markov model with adaptation.
We repeated the analysis for songs of Bird 2. The vocal elements were clustered into 7 types, with 6 types identified as song syllables (Figure 8a, for syllables A to F, respectively) and one type identified as the introductory note (Figure 8a, C1, ). The song sequences occurred in long sequences (mean length s.d.), with the gaps between consecutive syllables smaller than . The introductory note appeared with repeats preceding the song sequences, and had much smaller volume compared to the song syllables. Less call notes were recorded for Bird 2 than for Bird 1 since the song sequences could be distinguished from the calls based on the lengths of the consecutive sequences of vocal elements with the gaps . A total of 845 song sequences were used for deriving the models.
We derived the POMMA for Bird 2 using the same procedure as for Bird 1. The POMM derived with the non-repeat versions of the song sequences has 10 states (Figure 9a). There are two states associated with syllable A, three states with syllable C, and one state with all other syllables. The states in the POMM with syllable repeats were replaced with states with adaptive self-transition probabilities and additional states when necessary to derive the POMMA (Figure 9b). Syllable A is associated with state 12 and state 10 of the POMM. In state 12, the number of repetitions of syllable A ranges from 2 to 16 and the repetition distribution peaks at 6. We modeled this distribution by replacing state 12 with two serially connected states , each with adaptive self-transitions (Materials and Methods). The self-transition probabilities and the adaptation parameters are for , and for . The transition probability from to is . The inward transitions to state 12 of the POMM were set to with the probabilities intact. The outward transitions from state 12 were transferred to and , with the transition probabilities scaled to make sure that total transition probabilities out from and were normalized including the self-transitions and the transitions from to . The resulting repeat number distribution with these parameters was fitted with the observed distribution using the nonlinear least square procedure (Materials and Methods), and the cosine-similarity of the fitted and the observed distributions reached 0.98. We tested simpler models of the repeat number distribution for state 12, including one state with adaptive self-transition probability and two serial states with only one state with adaptive self-transition probability, but they did not work as well. In state 10 of the POMM, syllable A repeats twice more than 99.7% of the time, with the rest being single repeats. We modeled this by replacing state 10 with two serial states with no self-transitions, and with a small probability of not transitioning from to to account for the rare case of single syllable A. The inward transitions to state 10 were transferred to , and the outward transitions from state 10 were transferred to and , similarly as for the case of state 12. The situation is similar for syllable B in state 11, which predominantly has two repeats (90%). State 11 was replaced with two serial states with no self-transitions. The number of repetitions for syllable C in state 6 ranged from 1 to 6 and peaked at 3. This repetition number distribution was model with one state with adaptive self-transition probability. All other states with more than one repeat were accurately modeled by adding self-transitions as in the Markov model. The cosine-similarities of the fitted and the actual repeat number distributions were all greater than 0.95. The resulting POMMA, shown in Figure 9b, has 13 states ( for syllables A to F, respectively).
The POMMA accurately describes the statistical properties of the syllable sequences of Bird 2. We generated 10000 song sequences using the POMMA, and compared to the observed sequences the repeat number distributions, the N-gram distributions, and the step probability distributions. The comparisons are excellent (Figure 10a–c). The -values between the model and the observed distributions are below or very close to the benchmarks obtained from the observed sequences as in the case of Bird 1 (Figure 10d–f, red curves). In contrast, the Markov model and the Markov model with adaptation, derived and evaluated following the same procedure as for Bird 1, fail to describe the statistical properties of the observed sequences (Figure 10d–f, cyan and green curves). The Markov model with adaptation cannot accurately model the repeat number distribution of syllable A, which has double peaks as shown in the first graph in Figure 10a, even though the model can accurately describe the repeat number distributions of other syllables. This contributed significantly to the inaccuracy of the Markov model with adaptation in the N-gram distributions and the step probability distributions.
In the POMM, different states can be associated with the same syllable type. One possible piece of evidence of such many-to-one mapping from states to syllables can be the subtle differences that might exist in the instances of the same syllable associated with different states. For Bird 1, there are two states for syllables B,C,D,G in the POMM shown in Figure 7a. We compared the duration distributions of the same syllable types associated with different states, as shown in Figure 11a. The distributions are clearly distinctive for syllables B, C, G (, shuffle test of the significance that the difference of the means of the two distributions is none-zero; the null-distribution of the difference of the means was generated using 500 pairs of randomly shuffled distributions, and the -value is the two-tailed probability of the difference of the means greater than the observed value given the null-distribution). There is no clear evidence of distinctions for syllable D (). Despite the significant differences in the durations for syllables B in the two states, the spectrograms of the syllables in the two states are very similar, as shown in Figure 11b. The same is true for other syllables.
For Bird 2, the duration distributions of the same syllable types associated with different states are mostly distinctive ( in three cases and in one case), as shown in Figure 11c, while spectrally the syllables are very similar (examples shown in Figure 11d). Most interestingly, durations of the syllables associated with the same state in the POMM can also be distinctive depending on the positions of the syllables in the repetition. In Figure 12a we show three cases. The first is syllable B associated with state 11 in the POMM. The durations of syllable B in the first position of repetition is significantly longer than in the second position of the repetition (). The second is syllable A associated with state 10. The durations of syllable B in the first position of repetition is clearly shorter than those in the second position (). Spectrally, these sets of syllables are indistinguishable (Figure 12b for syllable B and 12c for syllable A). Both states were replaced with two serial states in the POMMA. Weaker evidence () also exists for syllable A associated with state 12 in the POMM (Figure 12a), which is replaced with two serial states both with adaptive self-transition probabilities in the POMMA. The systematic variations of syllable durations on the positions in repetition supports the idea of using multiple states to model repeat number distributions associated with single states in the POMM.
Taken together, the results on syllable durations provide some evidence for the validity of the many-to-one mapping from the states to the syllables.
Bengalese finch songs consist of variable sequences of a finite number of syllables. We have shown that the statistical properties of the sequences are well captured by a state transition model, the POMMA, in which the repeat probabilities of the syllables adapt and many-to-one mappings from the states to the syllables are allowed. The Markov model, which has been commonly used in studies of characterizing variable birdsong sequences, is clearly inadequate for the Bengalese finch songs. The POMMA is an extension of the Markov model. As in the Markov model, each state is associated with a single syllable, and the state transitions are characterized by the transition probabilities. However, unlike the Markov model, many states are allowed to be associated with the same syllable, and the state transition probabilities can vary depending on the history of the state transitions dynamics. These extensions are motivated by considerations of the neural mechanisms of birdsong generation.
The premotor nucleus HVC (used as a proper name) is a critical area in songbird brain for song production [20]. Firing of HVC neurons that project to RA (the robust nucleus of the arcopallium) drives singing [21], [22]. Experimental evidence suggests that a syllable is produced by the bursts of spikes propagating in a chain network of HVC projection neurons [22]–[25]. A set of HVC projection neurons reliably drive the RA neurons [22], which in turn drive downstream motor neurons to produce sound. Such a chain network in HVC could be a neural representation of a single state in POMMA. Thus, the association of a state to a single syllable is a reflection of the reliability of a chain network driving the production of a syllable.
The connections from HVC to RA are learned [26]–[29]. This makes it possible that different sets of HVC projection neurons are set up during learning to drive acoustically similar syllables. In zebra finch, different neural activity in HVC has been observed during vocalizations of acoustically similar syllables [21], [30], supporting the possibility of multiple sets of HVC neurons driving the same syllable. Such a possibility of many-to-one associations from the neural sets in HVC to syllables motivates introduction of many states corresponding to one syllable in the POMMA. It is conceivable that the same syllable driven by different sets of HVC neurons have subtle differences in the acoustic features due to imperfections of learning. Indeed, we found that instances of the same syllable associated with different states in the POMMA can have significantly different duration distributions (Figure 11 and Figure 12). A recent study has shown that the acoustic features of Bengalese finch syllables can shift systematically depending on the sequences around the syllables [31], which is in agreement with our observation. There can be alternative explanations to our observations that do not require separate sets of HVC neurons to encode the same syllable. One possibility is that the sequence-dependent differences in the acoustic features are due to the history dependence of the activations of the unique set of HVC neurons driving the syllable. Another possibility is that the differences are due to the inertia of the motor periphery rather than the variations in neural activity [31]. Finally, the differences can be due to sequence dependent activations of neurons in other areas in the song system, such as RA [31]. More direct experiments, such as single unit recordings in HVC of singing Bengalese finch, are required to test unambiguously whether the many-to-one mapping from HVC to RA exits.
The POMMA can be directly mapped onto the branched chain network model of the Bengalese finch song syntax [10]. Each state of the POMMA corresponds to a syllable-encoding chain network of HVC projection neurons, and each transition in the POMMA corresponds to the connection from the end of the synaptic chain corresponding to to the start of the synaptic chain corresponding to . The POMMA and the network model thus have identical branching connection patterns. In the network model, spike propagation along a chain drives the production of a syllable. At a branching point, spike propagation continues along one of the connected chain networks with a probability that depends on a winner-take-all competition and noise [10], [15]. The success of the POMMA in capturing the statistical properties of the Bengalese song sequences supports the branched chain network model of Bengalese finch song syntax. A critical prediction for the network model is that, for some syllables, HVC projection neurons should burst intermittently, bursting during some instances of the syllables but not in others. This is markedly different from the case of zebra finch, in which HVC projection neurons burst reliably for each production of the song sequence [22], [25]. The prediction can be tested with electrophysiological experiments.
Adaptations are widely observed in neural systems. Continuous firing can reduce neuron excitability [18], and excitatory synapses can be less effective when activated repeatedly [16], [17]. In zebra finch, consecutive singing increases the durations of the song syllables [32]. It is possible that the slow-down of the song tempo is due to some adaptive processes in HVC. In the branched chain network model of the Bengalese song syntax, weakening connection strength from one chain network to another at a branching point reduces the transition probabilities between them [10]. These observations suggest that the transition probabilities might not be fixed. Introducing adaptive processes in the neural excitability and synaptic efficacy should lead to adaptive transition probabilities in the branched chain network model, especially for the repeated activations of a chain network, which correspond to the reduction of the self-transition probability. It remains to be seen experimentally whether HVC projection neurons or the excitatory synapses between them have the adaptive properties. It might be also possible to see the signatures of adaptation by analyzing the burst intervals of HVC projection neurons during syllable repetitions, or the burst intervals of RA neurons. The observation that burst intervals in RA neurons steadily increase with song sequence repetition in zebra finch [32] suggests that similar effect could be observed in Bengalese finch.
We emphasize that adaptation is important for reducing the complexity of the state transition model. It is possible to include syllable repetitions in the POMM, with no adaptations of the transition probabilities, and accurately describe the statistical properties of the Bengalese finch songs (Materials and Methods; supplementary Figures S2–S4). However, compared to the POMMA with adaptation, the number of states is larger. While the POMMA has 14 and 13 states for Bird 1 and Bird 2 (Figures 7b and 9b), respectively, the POMM has 20 and 18 states (Figures S2 and S3). In the POMM, many states are needed to produce the peaked repeat number distributions such as that of syllable A in Bird 2 (Figure 10a). The difference of the number of states in the POMM and the POMMA should increase with the number of syllables with peaked repeat number distributions. It is the significant reduction of the model complexity that motivates our choice of the model with adaptation (the POMMA) rather than the non-adapting model (the POMM).
We have used multiplicative reduction of the repeat probabilities. It remains to be investigated whether other formulations of the adaptation can be similarly or even more effective. In our approach, only the repeat probabilities are adapted. A more consistent model should allow adaptation and recovery in all transition probabilities, such that the state transition dynamics depends on the history of the entire syllable sequence, not just the syllable repetitions. This approach might be important if there are repeats of short sequences such as ABABABAB, in which the transition probabilities from A to B and B to A might need to be adapted. But such a model is difficult to derive from the observed sequences. In our data, repetitions of short sequences were rarely seen, hence adapting only the repeat probabilities of single syllables was adequate. We have shown that adaptation alone is not sufficient to augment the ability of the Markov model to describe the Bengalese finch songs, and the many-to-one mapping from the states to the syllables is necessary. However, we cannot rule out the possibility that the more consistent model with all transition probabilities adaptive, and perhaps with more complex forms of adaptation, can eliminate the requirement for the many-to-one mapping.
The POMMA is closely related the hidden Markov model (HMM) [33], which is widely used to model sequential structures in human languages [14], [33], [34] and genomes [35], [36]. In the HMM, the transitions between the states are as in the Markov model, but each state is allowed to emit all symbols (or syllables in birdsong case) with some probability dependent on the state. The flexibility of the state and the symbol associations makes the HMM much more capable of capturing statistical properties of sequences than the Markov model. To apply the HMM to birdsong, however, it makes more sense to require that a state can be associated with a single syllable only, if the correspondence between the model and the neural dynamics of birdsong generation is considered [10]. HVC neurons reliably activate RA neurons [22], and there is no evidence that activation of the same sets of HVC or RA neurons can probabilistically produce multiple syllables. The HMM with the restriction that one state emits one symbol is the POMM [10], [19]. The POMM is distinguished from the Markov model in that a syllable can be associated with multiple states (many-to-one mapping from the states to the syllables). Even though the transitions between the states are Markovian, the syllable statistics can be non-Markovian due to the multiple representations of the same syllables [10]. The HMM with no one-to-one restriction does not lead to a more compact model than the POMMA for the Bengalese finch songs (Materials and Methods). To achieve the level of the accuracy of the POMMA, the HMM needs close to 18 states for both Bird 1 and Bird 2 (Figures S7), which is similar to the POMM. Indeed, most states in the HMMs predominantly emit one syllable (Figures S5 and S6), and the structures of the HMMs and the POMMs are similar for both birds.
There are previous efforts of describing Bengalese finch song sequences with state transition models [12], [13]. Chunks of syllable sequences, which are fixed sequences of syllables, were extracted from the observed sequences and used as the basic units of the state transition models [12], [13]. A syllable can appear in many chunks, hence these models implicitly contain the many-to-one mapping from the states to the syllables. But the chunk extractions and the state models were not derived from the statistics of the observed sequences. Furthermore, the models were not tested against the observed song sequences for statistical properties. In contrast, the POMMAs were derived from and tested with the observed song sequences.
Although there is a close connection between the POMMA and the branched chain network model of how HVC generates variable syllable sequences in Bengalese finch [10], [15], the POMMA or the POMM can be compatible with alternative neural mechanisms, including feedback control of sequences through RA to HVC projections [31], syntax generation in other nuclei upstream to HVC or RA in the song system [12], [37], [38], noisy recurrent networks in HVC [39], and branched chain networks of inhibitory HVC interneurons [40]. It is also possible that different statistical models can be derived from these mechanisms. More detailed analyses of the alternative mechanisms are needed to see whether they can produce syllable sequences with statistics compatible to the observed Bengalese finch songs.
There should be a family of equivalent POMMAs for the songs of a Bengalese finch. For example, the same repeat distributions can always be modeled with more states. The POMMA that we have derived is the simplest model that is consistent with the data. Given this insight, we expect that the neural representation of the syntax should be similar to the derived POMMA but most likely not identical. We have developed a state merging method for deriving the POMM from the observed syllable sequences. It is possible to use the well-established methods of training the HMM [33] to derive the POMM. We observe that our method is faster than the training methods of the HMM. A more detailed analysis of the state merging method is needed to quantify its speed and convergence properties.
In conclusion, we have derived a compact POMMA that successfully describes the statistical properties of Bengalese finch songs. Our approach can be useful for modeling other sequential behaviors in animals and statistical properties of sequences in general.
Acoustic recordings were performed with a boundary microphone (Audio-Technica PRO44). Microphone signals were amplified and filtered (8th-order Bessel high-pass filter with and 8th-order Bessel low-pass filter with kHz, Frequency Devices). The filtered signals were digitized with a 16-bit A/D converter (PCI-6251, National Instruments) with a sampling rate of kHz.
Vocal elements were defined as continuous sounds bounded by silent periods. Thresholding the amplitudes of the pressure waves is a common approach of isolating vocal elements in birdsongs [31], [41], [42]. We developed a similar method. From the pressure wave of a vocalization, the oscillation amplitude at time was obtained by finding the maximum of in the interval of one oscillation cycle that contains . The amplitude was further transformed to , where is a smoothing function that uses the second order Savitzky-Golay filter with window (801 data points). Vocal elements were isolated by detecting continuous regions in that were above a threshold function . The threshold function was obtained in moving windows (step size ); in each window, the threshold was set at the 0.3 point from the floor of to the local maximum of in the window. The floor is the characteristic value of in the regimes with no sound, and was identified as the position of the lowest peak in the histogram of the values of for all . A detected region was excluded if the total area above was smaller than multiplied by the difference between the maximum value and ; or if the maximum value of in the region minus was smaller than ; or if the width of the region was less than . These exclusions ensured that most noisy fluctuations were not counted as vocal elements. The results of the vocal element isolations were manually checked and adjusted by plotting out the waveforms in conjunction with the boundaries of the vocal elements to ensure that no obvious mistakes were made. The parameters used in the above procedure were empirically determined to yield the best results in our dataset. They should be adjusted if the procedure is used for other recordings of birdsong.
The waveform of an isolated vocal element was transformed into a spectrogram , which is the energy density at frequency and time . The frequency was restricted to to . The spectrogram was computed with the multi-taper method [43] (time-bandwidth product, 1.5; number of tapers, 2) with window size and step size (software from http://chronux.org). The frequency was discretized into grids with between adjacent points. To exclude silent periods at the beginning and the end of the vocal element, the time span of the spectrogram was redefined to the region in which the total power in the spectrum at each time point exceeded 5% of the maximum of the total powers.
We used a semi-automated procedure to cluster the vocal elements into separate categories. Similarities between the vocal elements were defined and used in a clustering algorithm. The final results were visually inspected and adjusted by plotting the spectrograms of all vocal elements in the clusters.
The similarity between the vocal elements was defined as follows. The spectrogram was considered as a sequence of spectra at the discrete time points. The spectrum at each time point was smoothed over the frequency domain using the second order Savitzky-Golay filter with window size of 5 frequency points. The smoothed spectrum was further decomposed into a slowly varying background by smoothing with the second order Savitzky-Golay filter with window size of 20 frequency points; and peaks by subtracting out . The relative importance of the peaks compared to the background was characterized by the weight , where is the standard deviation of the distribution over the frequency domain.
The spectrum at of was compared to the spectrum at of by computing which is the weighted sum of the cosine-similarities between the peaks and between the backgrounds. Here and are the peaks and and are the backgrounds of and , respectively. The cosine-similarity of two vectors (or distributions) was defined as (1)where and are the means and is the norm. is the maximum of the weights across all time points of the two syllables. If , the two spectra and were considered the same (denoted ). Otherwise the two spectra were defined as distinctive.
The similarity between two syllables was characterized by the longest common subsequence (LCS) between them. A common subsequence was defined by a set of time points in syllable and a set in syllable , such that the spectra at corresponding time points are the same, i.e. , , ..., . There was an additional restriction that corresponding time points did not differ by more than , i.e. for all . The length of the common subsequence is . LCS is the common subsequence with the maximum length. A long LCS indicates that the two syllables are similar, while a short LCS indicates they are dissimilar. We defined the similarity score of two syllables as the length of LCS divided by the mean of the lengths of the two syllables.
Types of vocal elements were identified by clustering 4000 vocal elements using a core-clustering algorithm, modified from the algorithm described in Jin et al [44]. The algorithm is based on the distance between vocal elements, defined as one minus the similarity score, and consists of the following steps. (1) For each vocal element, find the list of nearby vocal elements with distances less than 0.1. (2) Among the vocal elements that are not yet part of a cluster, select the one with at least 5 nearby vocal elements and the smallest mean distances to its nearby vocal elements as the core point of a new cluster. (3) Assign all unclustered vocal elements that are in the nearby-list of the core point to the new cluster. All vocal elements that are in the nearby-list but already clustered are reassigned to the new cluster if their distances to the core points of their respective clusters are larger than their distances to the new core point. (4) Repeat steps (2–3) until no new cluster could be created. (5) Merge clusters. Two clusters are merged if at least 5% of the vocal elements in each cluster had small distances () to the vocal elements in the other cluster. (6) Assign vocal elements that are not yet clustered. A vocal element is assigned to the cluster that had the maximum number of members whose distances to the vocal element are less than 0.15. In some cases, individual clusters contained separate vocal element types that had subtle differences but distinguishable. Such clusters are split into new clusters.
Once the types of vocal elements were identified with the clustering algorithm, we used the following procedure to classify all vocal elements that were not already clustered. (1) Identify the center of each cluster as the vocal element that has the minimum mean distances to all other vocal elements in the cluster. (2) Compute the distances from the vocal element to be assigned to the cluster centers. The three clusters with the lowest distances are selected. (3) Compare the durations of the vocal elements in the selected clusters to the duration of the candidate vocal element, and select 20 (or less if the cluster size is smaller than 20) from each selected cluster that are closest. (4) Compute the distances from the candidate vocal element to the selected vocal elements. (5) Assign the vocal element to the cluster to which the most of the selected vocal elements with the distances smaller than 0.2 belong. (6) If none of the selected vocal elements have distances less than 0.2, do not assign the candidate vocal element. The unclustered vocal elements were grouped into 2000 blocks, and their mutual distances were computed. The clustering and identifying procedures were repeated until no more clusters emerge. During this process, clusters were merged if they were subjectively judged as similar by inspecting the spectrograms and the mutual distances between the members of the clusters. Individual vocal elements were reassigned to different clusters if necessary.
The final results of the clustering of the vocal elements were validated and adjusted by visual inspections of the spectrograms.
In the case of a state with self-transition, the transition probability is initially but is reduced to after repetitions of the state, where is the adaptation parameter. The probability of having repeats is then
More complex repeat distributions can be modeled with more states. One model has two serial states . Both are associated with the same syllable, and only has self-transition. The transition probability from to is , and the self-transition probability of is initially but undergoes adaptation with the adaptation parameter . The probability of observing one repeat is given by
The probability of observing repeats is given by
Another model with two serial states allows both and to have self-transitions with parameters for and for . The probability of transitioning to after leaving is . The probability of observing one repeat is
The probability of observing two repeats is in which the first and the second terms are the probabilities of the state sequences and , respectively. Similarly, for , the probability of observing repeats is given by where and
Here and are the probabilities of repeating and times, respectively.
The cases above were all we needed to model the Bengalese finch songs in this study. More complex models with more states can be necessary for other Bengalese finch songs, and the repeat number distributions can be similarly derived.
We used a state-merging method to derive the POMM from the observed syllable sequences. The process is illustrated with an example in Figure S1 with a simple case of two syllables 1 and 2. From 5000 observed sequences (Figure S1a), a tree Markov model is constructed (Figure S1b). For each sequence, the tree model contains a unique path of state transitions from the start state. This is achieved by starting with the start state and the end state only, and adding new states as needed by finding the paths for the sequences. For example, consider the first sequence 12. At this point no states are emitted from the start state. A new state with syllable 1 is added and connected from the start state; a new state with syllable 2 is added and connected from ; finally, connects to the end state. With the additions of the two states, the sequence is mapped to a state transition path . Now consider the second sequence 121. State transitions generate the first two syllables in the sequence. To generate the last 1, a new state with syllable 1 is added, and is connected from and also to the end state. Now branches into and . This process continues, until all observed sequences are uniqued mapped into the paths in the tree model. The transition probabilities from a state to all connected states are computed from the frequencies of the transitions observed in the sequences. The tree model is a simple POMM that is a direct translation of the observed sequences; it contains all observed sequences. However, the tree model is incapable of generating novel sequences that are statistically consistent with the observed sequences. Moreover, since each transition probability can be considered as a parameter, the number of parameters in the tree model is enormous, severely restricting its predictive power. To reduce the number of parameters, a more concise POMM is derived by merging the equivalent states in the tree model. If two states are associated with the same syllable, and the probability distributions of subsequent sequences of length 15 or smaller are similar (cosine-similarity ), the two states are merged. This is done until no further mergers are possible. Finally, state transitions with probabilities smaller than 0.01 are eliminated, and all states that are reached less than 0.005 times in all observed sequences are also eliminated. These merging and pruning procedures lead to a concise POMM with five states for the simple example, as shown in Figure S1c. There are two states for syllable 1, which is an example of the many-to-one mapping. Indeed, the observed sequences in Figure S1c was generated with a POMM with structure identical to the one in Figure S1c and with equal transition probabilities to all connected states from a given state. The example demonstrates that the state merging method can lead to a concise POMM from observed sequences. The procedure was used to derive the POMMs for Bird 1 and Bird 2 using the non-repeat versions of the syllable sequences and keeping track of the number of syllable repetitions in each state, as described in the main text. The accuracy of the POMM from the state merging procedure was tested by generating 10000 sequences (see the main text for the generation procedure) and comparing with the observed sequences the repeat number distributions, the N-gram distributions, and the step probability distributions. The -values were computed and compared with the benchmarks derived from the observed syllable sequences as discussed in the main text. The number of states in the POMM was further reduced by testing mergers of all states associated with the same syllables and testing deletions of all states. The mergers and deletions were accepted if the -values of the resulting POMM fell below the benchmarks or they were smaller than the -values of the original POMM. The state merging and subsequent reduction of the number of states was fully automated. The POMM derived from the above procedure were morphed into the POMMA by replacing each state associated to repeating syllables with one or more states with adaptive self-transition probabilities. Various adaptive models for the repeat number distributions were tested as described in the main text. The process of morphing the POMM to the POMMA was not automated.
To derive the POMM from the syllable sequences but include the syllable repetitions without introducing adaptation, each state associated with repeating syllables in the POMM derived with the non-repeat versions was replaced by its own POMM. The replacing POMM was derived from the repeat sequences of the syllable using the HMM training method described below. In this case, since there is only single syllable in the repeat sequences, the HMM is equivalent to the POMM. We increased the number of states in the replacing POMM until the repeat number distribution of the syllable could be reproduced with the cosine-similarity . The in and out transitions in the POMM from the non-repeat versions were retained in the replacements. The resulting POMMs for Bird 1 and Bird 2 are shown in Figures S2 and S3. Direct applications of the state merging procedure did not lead to concise POMMs using the syllables sequences with repetitions. The main reason was that the syllable repetitions, especially when the mean repetition number was larger, required more sequences than available to accurately judge the statistical equivalence of the states for merging in the tree POMM.
We used the Baum-Welch algorithm for training the HMM from the observed sequences [33]. A number of states is chosen for the HMM. There is a start state and an end state, which only emit the start and the end symbols, respectively. All other states can be associated with any of the syllables with the emission probabilities. The transitions from the start state to the end state and from all states to the start state were excluded. All transition and emission probabilities were set randomly initially, and adjusted with the observed sequences using the Baum-Welch algorithm until they converged (errors of the probabilities below 0.001). To avoid local minima in deriving the HMM, we repeated the training process 20 times, and selected the HMM with the maximum log-likelihood for the observed sequences. The derived HMM was evaluated by generating 10000 sequences and comparing the statistics with the observed sequences. The generation method is the same as in the Markov model, except that at each state, the syllable generated is determined from the emission probabilities at that state. The number of states in the HMM was systematically varied. The results for Bird 1 and Bird 2 are shown in Figures S5–S7.
|
10.1371/journal.pgen.1004141 | Age, Gender, and Cancer but Not Neurodegenerative and Cardiovascular Diseases Strongly Modulate Systemic Effect of the Apolipoprotein E4 Allele on Lifespan | Enduring interest in the Apolipoprotein E (ApoE) polymorphism is ensured by its evolutionary-driven uniqueness in humans and its prominent role in geriatrics and gerontology. We use large samples of longitudinally followed populations from the Framingham Heart Study (FHS) original and offspring cohorts and the Long Life Family Study (LLFS) to investigate gender-specific effects of the ApoE4 allele on human survival in a wide range of ages from midlife to extreme old ages, and the sensitivity of these effects to cardiovascular disease (CVD), cancer, and neurodegenerative disorders (ND). The analyses show that women's lifespan is more sensitive to the e4 allele than men's in all these populations. A highly significant adverse effect of the e4 allele is limited to women with moderate lifespan of about 70 to 95 years in two FHS cohorts and the LLFS with relative risk of death RR = 1.48 (p = 3.6×10−6) in the FHS cohorts. Major human diseases including CVD, ND, and cancer, whose risks can be sensitive to the e4 allele, do not mediate the association of this allele with lifespan in large FHS samples. Non-skin cancer non-additively increases mortality of the FHS women with moderate lifespans increasing the risks of death of the e4 carriers with cancer two-fold compared to the non-e4 carriers, i.e., RR = 2.07 (p = 5.0×10−7). The results suggest a pivotal role of non-sex-specific cancer as a nonlinear modulator of survival in this sample that increases the risk of death of the ApoE4 carriers by 150% (p = 5.3×10−8) compared to the non-carriers. This risk explains the 4.2 year shorter life expectancy of the e4 carriers compared to the non-carriers in this sample. The analyses suggest the existence of age- and gender-sensitive systemic mechanisms linking the e4 allele to lifespan which can non-additively interfere with cancer-related mechanisms.
| Discovering genetic origins of healthspan and lifespan could lead to breakthroughs in increasing the years of healthy and long life. In this paper we characterize the association of the e4 allele of the well-studied ApoE gene with lifespan in two generations of participants of large longitudinal studies, the Framingham Heart Study and the Long Life Family Study, and investigate the role of major human diseases such as cardiovascular disease, cancer, and neurodegenerative disorders in this association. This wide range of systemic analyses is possible given the large sample with directly genotyped ApoE polymorphism available from these studies (N = 9841, with 2557 deaths). The analyses show that women's lifespan is more sensitive to the e4 allele than men's in these populations. However, the strongly adverse effect of the e4 allele is not observed for all women, but only for those 70 to 95 years old. Cardiovascular disease, cancer, and neurodegenerative disorders do not mediate the association of the e4 allele with lifespan. However, cancer, but not cardiovascular and neurodegenerative diseases, non-additively enhances this effect resulting in 4.2 years of difference in mean lifespan for the e4 allele carriers compared to the non-carriers.
| The Apolipoprotein E (ApoE) common polymorphism (e2, e3, and e4) is one of the most studied genetic variants in humans. The interest in this polymorphism is two-fold. First, the functional diversity of the ApoE polymorphism appears to be a unique signature of humans with no coding variation in this gene even in human's closest ancestries in which the monomorphic ApoE sequence resembled human's e4 allele [1], [2]. Understanding the functional diversity of the ApoE gene, thus, can help in gaining insights on human evolution. Second, the ApoE polymorphism is of fundamental interest for geriatrics and gerontology because of its profound role in human diseases in late (post-reproductive) life and lifespan.
Most consistent associations were reported for the detrimental effect of the e4 allele on Alzheimer disease [3]–[5]. Studies also mostly documented a detrimental role of the e4 allele in cardiovascular health [6], [7] although a protective role of this allele was also reported [6], [8]. The e4 allele was associated with human lifespan and longevity in a number of studies [9]–[19]; some studies reported, however, no significant effect [20]–[22] (see also http://genomics.senescence.info/longevity). Studies of the role of the e4 allele in human longevity were mostly limited to comparing frequencies of genotypes in long-living individuals and younger controls [23], a strategy which has limitations [24]. Studies examining survival of older individuals carrying the e4 allele are rare (notably, [17], [18]). Sexual dimorphism of the ApoE gene in human survival has not been widely studied so far (see [17] and references therein).
Since the e4 allele may be involved in regulation of such common diseases in the elderly as dementia and cardiovascular diseases (CVD), it is often assumed that the detrimental effect of the e4 allele on human longevity is mediated by these diseases (e.g., [11], [14], [25]). Studies of the systemic effect of the e4 allele and major human diseases on lifespan in the same samples are rare [16], [17] primarily because they require large samples of genotyped individuals followed for a long period of time to have sufficient number of events.
Despite the detrimental role of the e4 allele in human health and longevity, this allele continues to be widespread in human population [26]. The persistence of this allele has been proposed to be a result of balancing selection implying that the e4 allele should be also evolutionarily advantageous with a beneficial role in early life [27]–[30].
In this work we examine three inter-related problems which, taken together, address the systemic role of the e4 allele in human lifespan. First, we investigate gender-specific effects of the ApoE4 allele on survival in a wide range of ages starting from midlife to extreme old ages. Second, we examine whether major human diseases such as CVD, cancer, and neurodegenerative disorders (ND) can explain (i.e., mediate) the effect of the e4 allele on survival. Third, we investigate whether these diseases can modulate the e4-specific survival non-additively. This wide range of systemic analyses is possible given the large sample with directly genotyped ApoE polymorphism available for the analyses and selected from the Framingham Heart Study (N = 5182) and the Long Life Family Study (N = 4659) followed longitudinally for up to 60 years with a total of 2557 deaths.
The proportions of the ApoE4 allele carriers (see Methods) and the allele-specific proportions of deaths, CVD, cancer, and ND are given in Table 1. Table 1 shows that the proportion of the e4 allele carriers is about the same regardless of gender in FHS and FHSO cohorts at the time of biospecimens collection, i.e., 20.6% FHS men, 22.3% FHSO men, 22.8% FHS women, and 22.7% FHSO women carry the e4 allele. The FHSO sample of genotyped survivors (mean age of about 50 years) was, however, about 20 years younger than that in the FHS at the time of biospecimens collection (Table 1) indicating no strong e4-specific survival selection by that time in the FHS and FHSO survivors. The proportion of the e4 allele carriers in the LLFS was the largest in spouses; it was (significantly [31]) smaller in children of the long-living individuals compared to spouses; it declined in the selected population of long-living individuals compared to younger populations.
Our empirical analysis showed no consistent detrimental effect of the e4 allele across ages on the survival of men either in the FHS or FHSO cohorts (Figures 1A and 1C). Contrary to men, the e4 female carriers have shorter lives than the non-carriers (Figures 1B and 1D). An important result is that the role of the e4 allele in survival can change with age. Specifically, there is no e4-specific difference in survival of either the FHS men or women at ages 95 years and older (Figures 1A and 1B). The e4 allele does not affect survival at ages 70 years and younger (Figure 1D) either.
Analysis of survival age patterns of the LLFS male and female offspring/spouses directly supports these observations. Specifically, the LLFS female offspring and spouses carrying the e4 allele show worse survival than those who do not carry this allele (Figure 2D). Survival of the LLFS male offspring and spouses is not sensitive to this allele (Figure 2C).
To better understand survival age patterns of the LLFS participants from the parental generation (Figures 2A–B), one should keep in mind that this is a population selected for its exceptional chances to live a long life based on family history and their own survival to old ages (see Methods). Accordingly, this population resembles the subpopulation of individuals who survive to the very old ages in the FHS original cohort rather than the entire sample of a normal population in this cohort. Then, an important result is that the LLFS women selected for their chances of exceptional longevity (Figure 2B) and the long-living women in the FHS original cohort (represented in Figure 1B by a tail of the survival age pattern) have the same lifespan regardless of whether they carry the e4 allele. When analyzing survival age patterns one should also consider the possibility of survival selection in aging cohorts; if this selection is sensitive to a specific genetic variant then we may have biased empirical age patterns for carriers of genotypes from this variant particularly at advanced ages. Then, although the lifespans of the long-living LLFS men may be sensitive to the e4 allele (Figure 2A), further analyses are necessary (see next subsection) to determine whether this effect is real.
Thus, Figures 1B and 2B document an important result that survival of long-living women participating in the FHS (see upper tail in Figure 1B) and LLFS is insensitive to the e4 allele. Figures 1D and 2D show another remarkable result that the effect of the e4 allele on survival in the FHSO and LLFS offspring/spouses is pronounced: (i) starting at the same age, 70 years and (ii) in women only.
We evaluated the sensitivity of the survival of the long-living LLFS men to the e4 allele seen in Figure 2A. Evaluation of the relative risk (RR) of death for the e4 allele carriers using a model without adjustment for birth cohorts supported the presence of the effect (RR = 1.52, p = 6.9×10−3) in this sample (Table 2). However, adjustment for birth cohorts entirely explained this association (RR = 1.17, p = 0.319; Table 2), suggesting that this sensitivity was likely due to differential survival of the e4 carriers and non-carriers in different birth cohorts in the LLFS.
The relative risks obtained from the data in Figures 1 and 2 revealed the presence of a significant detrimental effect of the e4 allele on survival in women in the FHS (RR = 1.25, p = 0.027), FHSO (RR = 1.59, p = 2.4×10−4), and LLFS offspring/spouse (LLFS_O+S; RR = 2.23, p = 5.2×10−3) samples (Table 2, all). No significant effect was seen in men in either sample or in long-living women in the LLFS (Table 2, all). Pooled data from the FHS and FHSO slightly improved the significance of the estimates for women, RR = 1.36, p = 1.3×10−4 (Table 2, FHS+FHSO, all).
However, given the empirical evidence on the substantial role of age-related heterogeneity (Figures 1 and 2), analyses of the relative risks using the Cox proportional hazards regression model, which disregards such heterogeneity, likely underestimate the effects. A more appropriate way to address the impact of age-related heterogeneity is to consider more homogeneous groups of individuals for whom the variation of the hazards is proportional over age. Empirical evidence from independent FHS and LLFS cohorts (Figures 1 and 2) suggests selecting more homogeneous groups of individuals who died or were censored at ages: (i) younger than 95 years in the FHS (note that there were virtually no genotyped individuals with lifespans less than 70 years in this sample), (ii) 70 years and older in the FHSO and LLFS_O+S (note, virtually all genotyped participants in these samples had lifespans less than 95 years), and (iii) 70 to 95 years in the pooled sample of the FHS and FHSO.
Table 2 shows that individuals from these more homogeneous groups in each sample are at substantially larger risk of death compared to the entire sample. For example, we observe 9% increment (from RR = 1.36 to RR = 1.48) in the risk of death in the more homogeneous 70–95 year group of the FHS and FHSO women. Correspondingly, the significance of the estimate also sharply increases from p = 1.3×10−4 to 3.6×10−6.
Importantly, the analyses also confirm the lack of a significant effect of the e4 allele on survival in the groups of individuals who did not belong to the selected more homogeneous groups (Table 2). Specifically, no significant effects were observed in: (a) the groups of individuals with lifespans less than 70 years in the FHSO and LLFS_O+S, (b) individuals with exceptional survival including the entire sample of the LLFS long-living men and women (LLFS_P), and (c) individuals who were aged 95 years and older in the FHS. The lack of significant effects cannot be explained by the sample size differences (Table 2).
To address this question, we focused on the more homogeneous groups of participants of the FHS original and FHSO cohorts defined in the previous subsection (the LLFS sample is underpowered for such analyses) in order to diminish bias attributable to disproportionality of hazards when using the Cox regression model. Given slightly smaller samples of the FHS participants with known ND status (Table 1), these analyses were limited to individuals with missing information on ND excluded (sample sizes are provided in the respective tables along with the effect estimates).
Additive adjustments of the Cox regression models estimating the risk of death for carriers and non-carriers of the e4 allele by (i) CVD, (ii) CVD and cancer, and (iii) CVD, cancer, and ND, reveal that CVD and cancer do not explain the observed associations. Contrarily, CVD and cancer tend to improve the estimates in each sample with a more pronounced role for cancer (Figure 3). ND plays at most minor mediating role in the associations of the e4 allele with survival of either men (Figure 3A) or women (Figure 3B). Thus, none of these diseases explain the association of the e4 allele with risks of death (see Supplementary Information, Table S1).
Given no qualitative difference in the additive role of CVD, cancer, and ND in the e4-specific risks of death across the FHS samples, we evaluated the risks in the largest more homogeneous pooled sample of the FHS and FHSO participants in disease-stratified analyses (see Methods). Figure 4 and Table 3 show that the risks of death for women are the same regardless of CVD or ND status, i.e., neither CVD nor ND increase mortality of the e4 female carriers nonlinearly even after adjustment for alternative diseases. These diseases do not non-additively modulate men's survival either.
A striking result was that non-skin cancer significantly (p = 0.029 for multiplicative interaction of cancer with ApoE) differentiated the e4-specific risks of death for women from the more homogeneous group (with moderate lifespans of 70 to 95 years) increasing them by 52% from RR = 1.36 (p = 3.8×10−3) for women who did not have cancer to RR = 2.07 (p = 5.0×10−7) for women who had cancer (Figure 4B and Table 3). The high risk of death for women with moderate lifespan who had cancer explained the 3.2-year shorter life expectancy for the e4-allele carriers compared to the non-carriers (Table 4). The same trend on the e4-specific excess in the risks of death was seen for male cancer patients compared to non-patients (Figure 4A). Cancer increases risks for the e4 allele carriers compared to the non-carriers making them marginally significant, RR = 1.31 (p = 0.080) (Table 3).
The available sample size allowed us to gain some insights on potential differences between cancer sites (other than skin) in these associations. In these analyses we excluded major sex-specific sites, i.e., prostate in men and breast in women. Figure 5 and Table 3 show that relative risks of death for men without non-sex-specific cancers (RR = 1.11) increases compared to men without cancers (RR = 1.03) but it declines for men having non-sex-specific cancers (RR = 1.17) compared to men having cancers (RR = 1.31). This pattern suggests that the potential modulating effect of cancer in men is likely not sensitive to cancer site. Contrary to men, Figure 5 and Table 3 show that modulating role of cancer in women is entirely attributed to non-sex-specific cancers. The relative risk of death for women with moderate lifespan who had non-sex-specific cancers became much more pronounced (RR = 2.51, p = 5.3×10−8). This high risk explained the 4.2-year difference in life expectancy for the e4-allele carriers and non-carriers in this group (Table 4).
Analysis of genotyped offspring in the FHS revealed that the e4 allele is irrelevant to survival in mid to early-old life, up to about 70 years (Table 2). This result appeared to be corroborated in an independent population of the LLFS offspring and spouses (Table 2). The e4 allele changed its role from neutral in mid to early-old life to detrimental at older ages. This change was found in independent samples of the FHS Offspring cohort (Figure 1D) and the LLFS offspring and spouses (Figure 2D). Moreover, this change occurred concordantly in the FHSO and LLFS: (i) at about the same age of 70 years and (ii) in women only. The detrimental effect of the e4 allele at old ages (until 95 years of age) was also found in a sample of the FHS women (Figure 1B; note that virtually no individuals with lifespan less than 70 years were genotyped in this cohort).
At extreme ages (95 years and older) we concordantly observed a neutral role of the e4 allele in each gender in the FHS (Figures 1A and 1B). Analysis of the long-living individuals in the LLFS corroborated these findings (see the “Empirical Age Patterns of Survival of the FHS and LLFS Men and Women” and “Risks of Death of the FHS and LLFS Men and Women” subsections).
Overall, these analyses demonstrated a strong detrimental effect of the e4 allele on survival which was mostly attributed to women with moderate lifespans of 70 to 95 years in the FHS, FHSO, and LLFS. For example, the e4 allele increased the risks of death of the FHS and FHSO women by about 48% (RR = 1.48) with very high confidence, p = 3.6×10−6 (Table 2).
Although our study provided robust evidence of a women-specific detrimental effect of the e4 allele on lifespan in three different samples of mostly North-American population (i.e., FHS, FHSO, and LLFS, see Methods), there is also robust evidence of a detrimental effect of this allele in Swedish men but not women [17]. Further, although our results on the neutral role of the e4 allele at extreme ages (95 years and older) are in agreement with some meta-analyses [e.g., 32], there is also evidence of a significant detrimental effect of the e4 allele at those ages in the Danish population [18]. The results by Rosvall et al. [17], Jacobsen et al. [18], and ours explicitly show that the effect of the e4 allele on lifespan may not be the same in different populations. These robust evidences from different populations illustrate that the concept of replication of the same effect of the same allele on the same complex phenotype characteristic for post-reproductive period has inherent limitations [33]–[36].
The e4 allele is a major susceptibility allele for Alzheimer disease (which is a subtype of the ND in this study) particularly in Caucasians [4] (but may be not in Hispanics [37]). Despite that, our well-powered analyses show that ND explains at most a tiny part in the association of the e4 allele with survival (Figure 3). The results of our analyses do not support the hypothesis that the lack of a mediating effect of ND can be due to potential ND misclassification. This is evidenced in Figure 3 by: (i) the tiny reduction of the effect size attributed to ND (Figure 3) despite the large prevalence of ND (particularly in the FHS as the older cohort, Table 1), and (ii) the role of cancer as a nonlinear modulator of the effect of the e4 allele on survival (Figure 4). Additive contributions of the e4 allele and dementia to survival was also observed in other studies [16] although the attenuation of the effect size by dementia varied [17].
Despite the associations of the e4 allele with CVD [6], [7] and with CVD-free life [19], [38], our analyses show that CVD does not explain the effect of the e4 allele on women's survival (Figure 3). Recent analyses support these results by showing independent associations of the e4 allele and various characteristics of cardiovascular health and CVD with survival [16], [17], [39].
Several studies reported on a role of the ApoE gene in cancer [40]–[43]. It has been also shown that the e4 allele can increase cancer-free lifespan in the FHS and FHSO men [19], [38]. The analyses in this study show no mediating role of cancer in the association of the e4 allele with women's survival; the additive contribution of cancer, however, can modulate the effect of the e4 allele, increasing the strength of this association (Figure 3).
CVD and cancer are the most common causes of death in humans and ND is fast growing cause of death in the elderly. CVD and ND are the diseases which have been most consistently associated with ApoE4. Despite that, these diseases do not explain the detrimental role of the e4 allele in lifespan. This finding implies the existence of a mechanism linking the e4 allele with lifespan which is largely independent of the mechanisms affecting susceptibility to CVD, cancer, and ND. Given also that the e4 allele may not be associated with frailty [14], [44], it is likely that this allele can be directly involved in regulation of human aging through intrinsic biological mechanisms. One potential mechanism could be associated with inflammation which may be involved in aging through two main pathways associated with “immunosenescence and synergies with chronic diseases that have inflammatory components” [29]. Given no mediating role of CVD, cancer, and ND observed in our study and that these diseases (particularly CVD and ND) can have e4-specific inflammatory etiology [8], [45], it might well be the case that the e4 allele affects survival through immunosenescence whereas it affects the risks of diseases through disease-specific inflammatory component.
Neither CVD nor ND non-additively (i.e., nonlinearly) modulates the detrimental effect of the e4 allele on women's survival, i.e., the relative risks of death for the e4 allele carriers are the same regardless of women's CVD and ND statuses (Figure 4B). This result is in line with findings by Little et al. [16]. However, the e4 allele was shown to be mostly associated with dementia-caused deaths by Newman et al. [39].
We found that cancer showed a significant nonlinear modulating effect in the association of the e4 allele with women's survival (Figure 4B). The e4-positive female cancer patients have about a two-fold increased risk of death at ages between 70 and 95 years compared to the non-e4 allele carriers (RR = 2.07) which is highly significant, p = 5.0×10−7 (Table 3). Such a strong effect results in a 3.2-year shorter life expectancy of the e4 carriers compared to the non-carriers in this sample (Table 4). Further analyses show that this effect is attributed to non-sex-specific cancer sites, it substantially increases, i.e., RR = 2.51, p = 5.3×10−8 (Table 3), and it explains the large 4.2 year differential in the life expectancy (Table 4). Women without cancer carrying the e4 allele are still at significant risk of death. The same non-additive role of cancer was found in the effect of the e4 allele on men's survival, i.e., this allele negatively affected cancer survivorship (Figure 4A). The diminished role of cancer as a nonlinear modulator of the effect of the e4 allele on survival in men compared to women can be attributed to a protective role of this allele in susceptibility to risk of cancer in men but not in women [19], [38], i.e., protection against risks of cancer may well explain modest risks of cancer survivorship of male e4 carriers.
The cancer-sensitive non-additive effect of the e4 allele on human lifespan suggests that mechanisms associated with cancer survivorship (i.e., with its progression and/or treatment) can interfere with a mechanism linking the e4 allele to lifespan. Our findings are particularly in line with inflammatory pathways [29], [43] which may overlap for aging and cancer survivorship as a result of the compromising of the immune system with age [46] (see also next subsection). Thus, the non-additive role of cancer in the effect of the e4 allele on lifespan and the lack of this role for CVD and ND likely underscores the synergism between cancer and aging.
Given the persistence of the e4 allele in humans, it may be beneficial in early life and, thus, be subject to balancing selection [27]–[30]. Indeed, several studies provided support for a beneficial role of the e4 allele in early life. For example, it was shown that the proportion of the e4 allele was significantly smaller in spontaneously aborted embryos than in adults [47]. The proportion of the e4 allele was also found to be significantly larger in healthy liveborn infants compared with stillborn infants and with adults [48]. These findings suggest that the e4 allele can benefit early survival. Then, given the detrimental role of this allele for survival in old ages, we should expect a neutral role at some point in between. Our finding of a neutral role of the e4 allele in survival in mid to early-old life of the genotyped FHS and LLFS participants supports this logic.
Studies also show that ApoE4 may protect against early life infectious diseases such as, e.g., diarrhea [49] and liver damage caused by the hepatitis C virus infection [50], [51]. A putative protective mechanism may be associated with an enhanced function of the immune system in early life [25] with a role of ApoE as an immunomodulator [52]. At old ages immunosenescence may be a factor favoring neoplasia [53]. Then, if ApoE4 boosts the immune system in early life, this may naturally lead to prematurely exhausting this system later in life which may affect cancer survivorship for carriers of this allele (and, thus, implying antagonistic pleiotropy). This hypothesis is supported by our findings of a strong non-additive modulating role of cancer in survival of female e4 allele carriers (Figure 5), by the very high proportion of deaths (80%) among female e4 carriers with non-sex-specific cancer by age 95 years (44 deaths among 55 carriers; Table 4), and by the 150% excess risk of death for such women compared to the non-e4 carriers (RR = 2.51, p = 5.3×10−8; Table 3). These high death rates can, in part, explain the diminishing detrimental effect of ApoE4 at very advanced ages (95+ years) in the FHS.
The lack of an association of ApoE4 with survival at extreme ages (95+) in the FHS and in an exceptional population of the LLFS long-living participants suggests that the detrimental effect of ApoE4 can be counterbalanced in some individuals. Potential factors can include buffering mechanisms (by other genes [54]) and/or environmental modulations of genetic effects [36]. Given large samples of long-living individuals in the LLFS, this study could be highly promising for revealing such mechanisms.
Analyses of the association of the ApoE4 allele with lifespan in three populations of the FHS, FHSO, and LLFS participants showed that women's lifespan was more sensitive to the e4 allele than men's. The adverse role of the e4 allele was limited to women with moderate lifespans of about 70 to 95 years; no survival disadvantage is seen for women with lifespans less than 70 or more than 95 years. The highly significant association of the e4 allele with lifespan was not explained by major diseases including CVD, ND, and cancer, whose risks can be sensitive to this allele, in large FHS samples. Non-skin cancer non-additively increased mortality of the FHS women with moderate lifespans increasing the risks of death of the e4 carriers two-fold compared to the non-carriers. High and highly significant risks of death of the e4-allele carriers in this sample explained their 3.2 year shorter life expectancy. The results suggest a pivotal role of non-sex-specific cancer as a nonlinear modulator of survival in this sample of women that increased the risk of death of the ApoE4 carriers by 150% (p = 5.3×10−8) compared to the non-carriers and explained the 4.2 year differential in life expectancy in this group. Our results suggest the existence of age- and gender-sensitive systemic mechanisms linking the e4 allele to lifespan which can non-additively interfere with cancer-related mechanisms.
We use data on longitudinally followed FHS/FHSO and LLFS participants to characterize the role of the ApoE4 allele (e2/4, e3/4, and e4/4) and non-e4 genotypes (e2/2, e2/3, and e3/3) in the lifespans of men and women separately.
Associations of the e4 allele with risks of death were characterized by the Kaplan-Meier estimator and the Cox proportional hazards regression model. The time variable in the analyses was age at death or age at the end of follow up. The model adjustments were explicitly stated when applicable.
To examine whether or not major human diseases can shape the association of the e4 allele with survival, we considered additive and nonlinear roles of CVD (diseases of hearth and stroke combined), cancer, and ND (dementia and Alzheimer disease combined) in this association. We considered all non-skin cancers unless explicitly stated. CVD and ND were chosen because they were most consistently associated with the ApoE polymorphism [3], [6], [7], [64]. Recent studies also showed that the ApoE polymorphism can be associated with cancer [e.g., 41]. These analyses were conducted using rigorously ascertained information on diseases in the FHS/FHSO only because the LLFS data are currently underpowered for such analyses.
To address nonlinear effect of diseases on the association of the e4 allele with survival, we conducted disease-stratified analyses. Each disease group included individuals who were diagnosed with the disease (or died from it) prior to death or the end of follow up in 2008. Otherwise, individuals were included in the complementary non-disease group [65].
We used the robust sandwich estimator of variances in the Cox model to account for potential clustering (e.g., familial). Statistical analyses were conducted using SAS (release 9.3, Cary, NC, USA).
This study used de-identified data from the FHS and LLFS. The FHS data are available from the NHLBI through dbGaP. No new data were collected in this work. As such, this study does not require either ethics committee approval or written consent.
|
10.1371/journal.pcbi.1000760 | Robustness under Functional Constraint: The Genetic Network for Temporal Expression in Drosophila Neurogenesis | Precise temporal coordination of gene expression is crucial for many developmental processes. One central question in developmental biology is how such coordinated expression patterns are robustly controlled. During embryonic development of the Drosophila central nervous system, neural stem cells called neuroblasts express a group of genes in a definite order, which leads to the diversity of cell types. We produced all possible regulatory networks of these genes and examined their expression dynamics numerically. From the analysis, we identified requisite regulations and predicted an unknown factor to reproduce known expression profiles caused by loss-of-function or overexpression of the genes in vivo, as well as in the wild type. Following this, we evaluated the stability of the actual Drosophila network for sequential expression. This network shows the highest robustness against parameter variations and gene expression fluctuations among the possible networks that reproduce the expression profiles. We propose a regulatory module composed of three types of regulations that is responsible for precise sequential expression. This study suggests that the Drosophila network for sequential expression has evolved to generate the robust temporal expression for neuronal specification.
| Cell fate specification is of key importance in the development of multicellular organisms. To specify various cell fates correctly, genetic networks precisely coordinate spatial and temporal gene expression patterns during various developmental stages. One central question in developmental biology is to elucidate the relationship between the pattern formation and the network architecture. During embryonic development of the Drosophila central nervous system, the neural stem cells express a group of genes in a definite order, which is responsible for the diversity of neural cells. To elucidate the underlying mechanism of the process, we analyzed the structure and dynamics of the genetic network for the temporal changes occurring in the Drosophila neural stem cells. Searching all the possible regulatory networks of these genes using a computer program, we detected the requisite regulations that reproduce observed gene expression profiles. By comparing the stability of the dynamics among the functional networks, we uncovered the robust nature of the actual Drosophila network against environmental and intrinsic fluctuations. These results indicate that the genetic network for sequential expression has evolved to be robust under functional constraints. Our study proposes regulatory modules that are responsible for the precise sequential expressions, which might exist in genetic networks for other temporal patterning processes.
| Precise coordination of cell fate decisions is crucial in the development of multicellular organisms. In the developmental processes, where a series of events occurs at a specific place and time, gene regulatory networks are responsible for implementing reliable biological functions [1], [2]. To obtain system-level understanding of such processes, it is necessary to integrate the molecular machinery of each regulation with architecture and dynamics at the regulatory network level. Biological functions achieved by gene networks are generally expected to possess robustness, i.e., insensitivity of system properties against a variety of perturbations that might originate from fluctuations during development and mutations through evolution. Recent investigations have addressed the questions of how robust biological functions are achieved through underlying molecular network architecture and its dynamic properties [3], [4], [5], [6], [7]. An illustrative example in developmental systems on this subject is segmentation of Drosophila melanogaster, which has been studied both experimentally and theoretically [8], [9], [10]. The requisite regulations or architecture of this system have been discussed at the network description level [10], [11], [12], [13], , and it is suggested that the underlying gene network has evolved to perform its processes in a robust manner [15], [16], [17].
Besides spatial patterning, temporal profiles of gene expression also play important roles in development [18], [19], [20]. Several computational studies have analyzed temporal expression profiles in biological processes such as the midgut development of sea urchin [21], [22] and vulval development of C. elegans [23]. These studies have shown relevant regulatory interactions and predicted unknown regulations for cell-fate specification.
The development of the Drosophila central nervous system (CNS) also manifests the importance of temporal patterning mechanism in development. Drosophila neural stem cell-like progenitors, called neuroblasts (NBs), generate a variety of neural cell types. During the embryonic development of the Drosophila CNS, NBs in the ventral nerve cord express certain transcription factors, i.e., Hunchback (Hb), Krüppel (Kr), Pdm1/Pdm2 (Pdm), and Castor (Cas), in a definite order (Fig. 1A–C) [24],[25],[26],[27]. In addition, the fifth factor, Seven-up (Svp), is expressed in the time window between Hb and Kr expression [28]. In association with this sequential expression, NBs divide asymmetrically to bud off a series of ganglion mother cells (GMCs). Each GMC undergoes an additional division to typically generate two postmitotic neurons. Depending on the transcription factors expressed in NBs at each division, postmitotic neurons acquire different cell fates. Thus, the sequentially expressed transcription factors control the cell-fate specification, thereby establishing the diversity of neurons in the Drosophila CNS. While neuronal specification process and generated cell types also depend on the spatial position [29], [30], [31] and lineage [32], [33] of NBs, the sequential expression is observed in a majority of ventral nerve cord NBs [34].
Isolated NBs exhibit sequential expression in vitro and differentiate into various neurons in a manner similar to that observed in vivo [35], [36]. Hb expression is switched off by Svp in a mitosis-dependent manner, while the subsequent expression of Kr, Pdm, and Cas proceeds in a mitosis-independent manner [28], [37]. These observations suggest that sequential expression of the genes is regulated cell-autonomously and occurs through mutual interactions among the factors.
In this study, we address the robustness of the gene network for sequential expression in the Drosophila CNS. One of the promising approaches to obtain insights into the system-level properties of biological systems is to compare the robustness of the actual network with that of other possible network architectures. Wagner considered how network architecture and robustness are related by studying circadian oscillation networks [38], although these networks lack a direct biological counterpart. Ma et al. studied the robustness of the Drosophila segmentation network [39], in which they had to arbitrarily eliminate components to reduce the size of the entire network. From theoretical and computational points of view, one advantage of studying temporal patterning in the Drosophila CNS is that the number of system components is so small that we can perform a comprehensive analysis of network architecture without any loss of biological relevance.
First, we explored the regulatory networks to reproduce the observed expression patterns in both wild-type (WT) and mutant embryos. We did not confine ourselves to only known regulations for sequential expression, but rather searched all possible networks that could reproduce the observed expression patterns. Studying the common structure of the specified genetic networks, we detected requisite regulations and predicted an unknown factor to reproduce known expression profiles. Second, we compared the robustness of the actual Drosophila network with that of the other networks reproducing the expression profiles. As a measure of robustness, we analyzed the stability of sequential expression against parameter variations and gene expression fluctuations. We found that the Drosophila network is highly robust and stable among possible functional networks. By further investigating the regulations necessary for the Drosophila network to be robust, we detected the responsible regulations. We propose a regulatory module composed of three kinds of regulations that is responsible for precise sequential expression of the Drosophila network.
Expression profiles of temporal transcription factors (hb, Kr, pdm, cas, and svp) in Drosophila NBs are summarized in Figure 1D for WT, loss-of-function, and overexpression embryos [25], [26], [28], [36], [40], [41]. It has been considered that these sequential expressions are produced (or at least modulated) by mutual regulations among the temporal transcription factors [24], [25]. We reconstructed the gene network for sequential expression in Drosophila NBs from the literature as shown in Figure 1E and F (for references, see Table 1).
First, we searched for regulatory networks that reproduce the sequential expression patterns of both WT and mutants. To investigate gene expression dynamics, we adopted a Boolean-type model [6] (see Materials and Methods for details of the model and the following analysis):(1)where represents the expression state of gene i () at the t-th time step and takes either 1 (ON) or 0 (OFF). Regulation from gene j to gene i is either positive (Jij >0), negative (Jij <0), or zero (Jij = 0), which corresponds to activation, repression, or absence of regulation, respectively. The state of gene i at the next step () is 1 when the sum of regulatory inputs is positive () or 0 when the sum is negative (). When the sum equals zero (), takes the default expression state : . In this study, the value of Jij is supposed to take one of the discrete values . The large negative value (−5) of Jij signifies that the expression of a gene is completely shut off in the presence of a repressor. This choice of large negative value comes from experimental observations of mutants. In experimentally observed expression patterns (Fig. 1D), genes are not activated when both repressors and activators are expressed. For example, in Kr++ and pdm++ embryo (here “++” means overexpression of the gene), pdm and cas expression is not observed in hb-expressing time window, although their activators are overexpressed. This indicates that the repressive effect from hb is dominant over pdm activation by Kr and cas activation by pdm.
Initial expression state of genes is set to 0, except for Hb, which emulates the NB gene expression in the first stage of sequential expression [24], [25]. Thus far, the only known function of Svp during the early stage is downregulation of Hb. There is no evidence that Svp regulates or is regulated by other temporal transcription factors during the expression series: Kr Pdm Cas [28]. In addition, Hb is only regulated by Svp and not by the other three factors (Kr, Pdm, and Cas). Thus, in the model, we assumed a pulsed expression of Svp as an input to the system, resulting in downregulation of Hb at the next time step. The temporal expression dynamics of Kr, Pdm, and Cas follow Eq. (1) with assigned values of Jij (Fig. 1F).
Based on the above formulation, we investigated whether the reconstructed Drosophila gene network (Fig. 1E and F) is sufficient to reproduce the sequential expression observed in WT, as well as all the known single loss-of-function and overexpression mutants, i.e., hb−, Kr−, pdm−, cas−, hb++, Kr++, pdm++, and cas++ (Fig. 1D, Table 2). Presently, we cannot specify the value of the parameters , and from empirical data; thus, each value could be arbitrarily chosen from (). We studied all 23 combinations of and found that the dynamics coincide with the expression profile in WT but not in some mutants for each choice of parameters (examples shown in Fig. 2). Depending on the parameter values, the expression dynamics changed to some extent, but none of the possible combinations reproduced the expression profiles of all of the mutants. For example, in case of , , and , the dynamics of the network for hb− and Kr− did not agree with the experiments (Fig. 2A), and in case of , , and , the dynamics of hb− and pdm− did not (Fig. 2C).
We then investigated whether networks other than the Drosophila network can reproduce the observed expression profiles by checking all the 312 ( = 531,441) combinations of Jij values. The dynamics agreed with the expression profile in WT for a large number of networks (39,391 out of 531,441), but any networks composed of hb, Kr, pdm, cas, and svp did not reproduce the profiles in both WT and mutants.
Preceding results indicate the difficulty of reproducing the observed expression patterns only with known constituents. We therefore introduced an additional presumptive regulator (x). The expression state of x was assumed to start in the ON state and change into OFF, or vice versa at () (see Materials and Methods). Including this assumption, we reinvestigated the dynamics of all 315 ( = 14,348,907) possible regulatory networks with all the possible switching timings of x. In the case that the expression of x switches OFF to ON, none of the networks conformed to the expected expression profiles. On the other hand, in the case that the expression of x switches ON to OFF, we found that 384 networks (<0.003%) reproduced the expression profiles of both WT and mutants. We refer to the detected networks as “the functional networks” hereafter in the study.
Comparing the regulatory interactions of the functional networks, we found that the regulations shared among all the functional networks are coincident with experimentally verified regulations (colored as black in Fig. 3A). In addition, activation of Kr and repression of cas by a presumptive factor x appear in all of the functional networks (colored as brown in Fig. 3A). The genetic network composed of these common regulations is a minimum network to reproduce the expression profiles of WT and mutants. To quantify the similarity among the functional networks, we measured the distances of the 384 functional networks from the actual Drosophila network (Fig. 3C); the distances are defined by the number of different regulations (see Materials and Methods). As a reference, we also performed the same analyses of distance measurement for all possible networks and the networks that are randomly reconnected from functional networks (see Materials and Methods). For all possible networks, the frequency distribution of the distances shows that the network architectures are different from the actual Drosophila network by 7.81.5 regulations. The reconnected networks yield similar results, albeit with slightly decreased distances (7.01.7 regulations). In contrast, the architectures of the functional networks differ by only 2.41.1 regulations. The architectures of the functional networks resemble that of the actual Drosophila network. These indicate that the gene networks that reproduce the known sequential expression patterns are highly constrained in their topologies.
Because there are multiple network architectures that explain the observed expression profiles as shown above, we then investigated the characteristics of the actual Drosophila network among the functional networks. From the biological point of view, the sequential expression in NBs should proceed reliably despite developmental disturbances such as cell-to-cell variation and intracellular fluctuations. We thus evaluated the stability of sequential expression for each of the detected functional networks and compared the properties of the actual Drosophila network to those of the other networks. To address the problem quantitatively, we extended the previous Boolean model into a model of ordinary differential equations with fluctuations in gene expression, where the concentrations of mRNAs {Mi(t)} and proteins {Pi(t)} obey the following equations [42], [43] (see Materials and Methods for the details of the model and the following analysis):(2)Here i refers to one of each gene: . The variables {Mi(t)} and {Pi(t)} take continuous values, unlike the previous Boolean description. The precise function form of promoter activities {Fi({Pj(t)})} is dependent on the regulatory interactions of the genetic networks and the default promoter activities {Si}, corresponding to the Boolean model. The time-dependent variables represent the noise in promoter activities. Here we have assumed that the expression noise comes from the transcription process (noise is incorporated only in the dynamics of {Mi(t)}). One reason is the practical convenience in the numerical calculations. In addition, recent quantitative analyses of gene expression have indicated that the gene expression noise mainly arises from transcription [44], [45], [46]. However, we should note that the result and conclusion obtained from the following analysis does not change even if we incorporate noise in the dynamics of {Pi(t)} as well (data is not shown).
Typical dynamics of the Drosophila network are shown in Figure 4, where sequential expression of WT is reproduced. The dynamics of the model are largely dependent on the parameter values and the noise intensities, and coincide with the experimental observations only under appropriate conditions. Therefore, such sensitivity to parameter variation is important for the development to proceed under environmental and individual fluctuations.
To characterize sensitivity, we measured the fraction of successes; that is, the fraction of the parameter sets that can reproduce the expression profile of WT among all the trials of random parameter assignments [15], [39]. To judge whether the dynamics coincide with the expression profile in Drosophila NBs, the dynamics of the protein concentrations {Pi} were discretized to 1 (0) for Pi > Pth (Pi < Pth). The threshold Pth was set as Pth = 0.2. The temporal dynamics of a network were accepted when the discretized dynamics satisfied the condition for WT in Table 3. To obtain the effect of parameter variation, we carried out the simulation without stochastic terms in Eq. (2). In each network, we repeated the simulations with random assignment of parameter values and calculated the fraction of successes (Fig. 5A). The Drosophila network scored the highest fraction of successes among the functional networks, and the networks closer to the Drosophila network tended to have higher scores.
We also investigated the dynamical stability of the gene networks against fluctuations. In this case, we performed the stochastic simulations in Eq. (2) with expression noise. To evaluate stability against noise, we chose the parameter values with which the expression profile is reproduced in the absence of noise. We then measured the relative fraction of successes under fluctuation. As is shown in Figure 5B, the fraction of successes under expression noise increased with the similarity to the actual Drosophila network as the fraction of successes under parameter variations. Thus, the Drosophila network lies at the top level of the functional networks in terms of robustness against these perturbations.
Because the Drosophila network has several other regulations in addition to the minimum functional network (gray arrows in Fig. 3A), these regulations might be responsible for the robustness shown above. We compared the robustness among the networks with or without the additional regulations. The fraction of successes against parameter variations for these networks is plotted in Figure 6A. The minimum network reproduces the sequential expression under the appropriate parameters, but the robustness is much lower than that of the Drosophila network. The scores of networks that lack one of the regulations fall between the minimum and the Drosophila network. Stability to expression noise was also evaluated by changing noise intensity, and similar results were obtained (Fig. 6B). The fraction of successes decreased as the noise intensity became larger, but the effect of noise on the Drosophila network was less severe than that on the minimum network. Thus, each of these regulations contributes to the robustness of the system.
To elucidate the roles of these regulations, we tried random parameter assignments for each of these networks and sampled successful parameter sets that reproduce WT sequential expression profile (Fig. 7). In the Drosophila network (Fig. 7A), wide ranges of parameter values are allowed, indicating that this network reproduces the required profile without quantitative tuning of parameters, and thus, shows high robustness. For other networks (Fig. 7B–E), the ranges are narrower for some parameters (as clearly seen in Spdm and Scas), and the numbers of successful parameter sets are less than those obtained for the Drosophila network.
How is the robust nature of the Drosophila network implemented by these regulations? As seen above, the parameter values of Spdm and Scas (default promoter activities of pdm and cas) are most influenced by the loss of these regulations. Because expression of a gene is induced by either the activity of the default promoter or the activators (see Materials and Methods), additional regulations in the Drosophila network (gray arrows in Fig. 3A) might compensate for the loss of default activities. To verify this possibility, we measured the dependence of the fraction of successes on the strength of regulations (, , and ) and default promoter activities (Spdm and Scas) (Fig. 8A–C).
Figure 8A shows the fraction of successes for random assignments of parameter values under given strengths of and Spdm. To score high reproducibility, Spdm must be large for small , but need not to be large for sufficiently large . This indicates that activation of pdm expression by Kr indeed compensates for the loss of default promoter activity of pdm. Thus, for the network lacking this regulation, the default promoter activity is necessary because inductions from other factors are absent. A similar relationship is found between and Scas (Fig. 8B).
As for repression of cas by hb, the role for robustness seems to be different from the above two. When the absolute value of is small, Scas must be small to achieve a high fraction of successes (Fig. 8C). As becomes larger, a higher value of Scas is allowed. This is because the repression from hb to cas reduces the mis-expression of cas in the early stage of sequential expression. Grosskortenhaus et al. suggested the direct repression from hb to cas [26], although there is no confirmative evidence to our knowledge. This regulation possibly contributes to the robustness of the actual system.
Through the present analyses, we obtained 384 functional networks that reproduce the sequential expression of both WT and mutants. The detected functional networks exhibit high similarity in regulatory interactions among the transcription factors (Fig. 3). This exemplifies the importance of the regulations in the minimum network for the sequential expression. In addition, the actual Drosophila network scores quite high on reproducibility of the WT sequential expression among all the functional networks (Fig. 5 and 6). Below, we discuss the biological implications of the temporal patterning of Drosophila NBs drawn from our numerical analyses.
In this study, we introduced an additional presumptive factor x to obtain networks that reproduce the sequential expression of both WT and mutants. Because x is hypothetical, we discuss its validity here.
Because the loss-of-function mutant of any one gene has only minor effects on the expression sequence (Fig. 1D), several previous reports suggested the existence of either unknown regulators or an additional clock mechanism that regulates the sequential expression [25], [26]. Our assumption is feasible for explaining experimental results because it does not need any other clock mechanism or superfluous multiple regulators. It is notable that our analysis indicates that the possible regulations of the presumptive factor are highly restricted; the expression of x switches ON state to OFF state (Fig. 4), and all the functional networks have activation of Kr and repression of cas by x (Fig. 3A). Thus, our assumption can be tested in future experiments in vivo.
We should note that while the regulator x is needed to explain the mutant profiles under our modeling assumptions, the mutual regulations of only known factors also reproduce the WT sequential expression (Fig. 1D). Therefore, the regulations among hb, Kr, pdm, and cas would play a primary role as discussed below.
An effective way to capture network function is to focus on the specific substructures (network motifs or modules) [1], [13], [14], [16], [39], [47]. Comparing all the functional networks, we detected the minimum structure for the sequential expression, which contains two successive regulatory loops (Fig. 3A and 9A); one is composed of hb, Kr, and pdm, and the other of Kr, pdm, and cas. In each loop, one gene represses the previous and the second next factor. The repressions of the second next factors (hb to pdm and Kr to cas) define the induction timing of the regulated factors, since they are kept repressed until the regulators are switched off. The feedback repression of the previous factors (pdm to Kr and cas to pdm) ensures their downregulation, which promotes the progress of the sequential expression. These coincide with the observations by Kambadur et al., who experimentally showed that the repressions from hb and cas define the temporal window of Pdm [24]. These repressive regulations and the activation from hb to Kr compose the minimum network for sequential expression (Fig. 9A). Although they are enough to reproduce the sequential expression under appropriate conditions, the expression profiles could be easily perturbed by parameter variations or increase of noise (Fig. 5 and 9A).
In the two loops of the Drosophila network, activations from one gene to the next (Kr to pdm and pdm to cas) exist in addition to the repressive regulations. Other functional networks do not necessarily have these activations, but the activations can compensate for the loss of default promoter activities (Fig. 8A and B). These regulations achieve precise expression by enhancing the correlations among the factors and heightening the stability against fluctuations (Figs. 5B and 6B). From these results, we conclude that three types of regulations (activation of the next factor, feedback repression, and repression of the second next factor) compose a regulatory module for precise temporal expression, as summarized in Figure 9B. The feature of this network module embodies the robustness of the Drosophila network.
Do the previous discussions have any implications on other developmental processes? In the studies of spatial patterning in Drosophila segmentation, it was claimed that the frequent substructure feed forward loop (FFL) can set the positions of expression domains [13], and mutual feedback repressions between the gap genes also have a pivotal role in the formation of expression domains with steep boundaries [12], [47]. In case of the Drosophila network for sequential expression, preceding genes activate the next ones, while these genes repress the preceding ones. Similar regulatory interactions are reported in the yeast cell cycle by Lau et al. [48]. Thus, such asymmetric mutual regulations would be a general mechanism that serves as precise switches in the process of temporal patterning.
We showed that the temporal specification network of Drosophila NBs contains not only the regulations necessary for generating sequential expression, but also additional regulations to achieve higher precision in the expression. In each hemisegment of Drosophila embryo, 30 different NBs are generated through spatial heterogeneity [29]. To guarantee sequential expression of common temporal transcription factors despite their differences in Drosophila NBs, the robustness of the system might be important.
The robust nature of the Drosophila temporal network could be the consequence of evolutionary optimization in the reproducibility of the sequential expression under functional constraint. In future, we expect that experimental manipulation of corresponding enhancers will be able to clarify the relevance of each regulation to temporal patterning and stability.
Here we describe the details of the Boolean model (Eq. (1)). The expressions of svp and x occur as inputs to the system. A pulse of svp expression always occurs at t = 1. Expression of x switches either from ON to OFF state, or from OFF to ON state at (). Once we assign the switching time of x expression (), its value becomes fixed through the analysis of expression patterns for all the genotypes. Because the autonomous pulsed expression of svp results in hb downregulation, we set Jhb, svp = −5, Jhb, j = 0 (j = hb, Kr, pdm, cas, or x), and Jk, svp = 0 (k = Kr, pdm, or cas) throughout this study. The time step at which we finish the simulation () was set as .
We thus investigated the behaviors of the remaining three factors (Kr, pdm, and cas) under the given regulatory interactions {Jij}. The total number of combinations of the parameters is 3M23 (the number of possible network architecture {Jij} multiplied by the number of default expression states ), where M is the number of regulations. To simulate the dynamics for mutants, we always set the expression state of the corresponding gene to 0 (OFF) for loss-of-function or to 1 (ON) for overexpression. We then examined whether the temporal dynamics of the genetic networks are coincident with the expression profiles of each mutant (Fig. 1D and Table 3).
In order to measure the similarity between the functional networks and the actual Drosophila network, we used two types of network ensembles as references. One is the ensemble of the possible network architectures. The other is a set of reconnected networks generated from the functional networks by iterative random reconnections of the matrix elements (1,000 iterations). The numbers of positive and negative regulations are preserved in the iterations.
To count the number of different regulations between functional networks and the actual Drosophila network, we neglected the regulations from x and positive self-feedbacks because the existence of those is uncertain from the experimental data.
We introduced the continuous model with stochasticity as shown in Equation (2). The promoter activity of gene i (i = hb, Kr, pdm, cas, or x) is described as follows,Regulatory interactions are continuous equivalents of {Jij} in the Boolean model, and g(x) is a piece-wise linear function such that g(x) = x for x>0 and g(x) = 0 for x<0. The parameters {Si} are the default activities of the promoters. Transcription of a gene is induced when the total regulatory inputs become positive (), and is suppressed when they become negative (). In order to consider the effect of fluctuations on the expression dynamics, we introduced additive white Gaussian noise : (Eq (2)), where is the noise intensity of gene i.
The expression of hb and x is induced only by the default promoter activities because all the regulations are absent for these two (). To describe the expression change of hb and x, the promoter activities of these two are set as Shb >0 for (Sx >0 for ) and Shb = 0 for (Sx = 0 for ), respectively. The promoter activities of the others are always assumed to exist (SKr, Spdm, and Scas >0). The noise intensities are also set as (>0) for and for (i = hb, x). Those of the other genes are (>0) (j = Kr, pdm, cas), Here we simply assume that the noise intensities of the genes take the same value . The noise intensity is set as in Figure 4, and in Figure 5. Noise intensity (horizontal axis) in Figure 6B means the value of .
For the continuous model, we considered two different types of robustness: (1) the reproducibility of the sequential expression against parameter variations and (2) dynamical stability against temporal fluctuations. To analyze the former, the default promoter activities {Si} were assigned randomly within the defined ranges. The values of the matrix were set to 0 when the corresponding regulations were absent (the corresponding element of the Boolean model takes Jij = 0) or assigned randomly when they are present (Jij0). In order to confine our attention to the properties of network architectures, the other parameters (, , , and ) were fixed throughout the analysis. The ranges and the fixed values of the parameters are listed in Table 4. Robustness against temporal fluctuations is measured as explained in the main text.
In the simulations, we found that the existence of positive self-regulation enhanced the fraction of successes in many cases, but hardly affected the sequential expression. To focus on the contributions of mutual regulations of genes to robustness, we neglected the positive self-feedback regulations and confined the analysis to 120 out of 384 functional networks.
|
10.1371/journal.pntd.0001591 | A Somatically Diversified Defense Factor, FREP3, Is a Determinant of Snail Resistance to Schistosome Infection | Schistosomiasis, a neglected tropical disease, owes its continued success to freshwater snails that support production of prolific numbers of human-infective cercariae. Encounters between schistosomes and snails do not always result in the snail becoming infected, in part because snails can mount immune responses that prevent schistosome development. Fibrinogen-related protein 3 (FREP3) has been previously associated with snail defense against digenetic trematode infection. It is a member of a large family of immune molecules with a unique structure consisting of one or two immunoglobulin superfamily domains connected to a fibrinogen domain; to date fibrinogen containing proteins with this arrangement are found only in gastropod molluscs. Furthermore, specific gastropod FREPs have been shown to undergo somatic diversification. Here we demonstrate that siRNA mediated knockdown of FREP3 results in a phenotypic loss of resistance to Schistosoma mansoni infection in 15 of 70 (21.4%) snails of the resistant BS-90 strain of Biomphalaria glabrata. In contrast, none of the 64 control BS-90 snails receiving a GFP siRNA construct and then exposed to S. mansoni became infected. Furthermore, resistance to S. mansoni was overcome in 22 of 48 snails (46%) by pre-exposure to another digenetic trematode, Echinostoma paraensei. Loss of resistance in this case was shown by microarray analysis to be associated with strong down-regulation of FREP3, and other candidate immune molecules. Although many factors are certainly involved in snail defense from trematode infection, this study identifies for the first time the involvement of a specific snail gene, FREP3, in the phenotype of resistance to the medically important parasite, S. mansoni. The results have implications for revealing the underlying mechanisms involved in dictating the range of snail strains used by S. mansoni, and, more generally, for better understanding the phenomena of host specificity and host switching. It also highlights the role of a diversified invertebrate immune molecule in defense against a human pathogen. It suggests new lines of investigation for understanding how susceptibility of snails in areas endemic for S. mansoni could be manipulated and diminished.
| Schistosomiasis, a neglected tropical disease, owes its continued success to freshwater snails that support production of prolific numbers of human-infective cercariae. Encounters between schistosomes and snails do not always result in the snail becoming infected, in part because snails can mount immune responses that prevent schistosome development. Understanding the factors important for snail resistance to schistosome infection will facilitate new lines of investigation to 1) understand the underlying basis of compatibility between schistosomes and snails in endemic areas and how this affects transmission dynamics and control efforts; and 2) to reveal ways to manipulate natural snail populations to enhance their resistance to schistosome infections. Here, we present the first evidence that a snail immune molecule, fibrinogen related protein 3 (FREP3), is important for successful defense against schistosome infections in Biomphalaria snails. In addition, we demonstrate that FREP3 is a target suppressed by trematode parasites to facilitate their establishment within the snail.
| Schistosomiasis is one of the world's most tenacious neglected tropical diseases, infecting an estimated 207 million people, mostly children [1]. The persistence of schistosome parasites stems in part from their use of freshwater snails for their larval development and transmission. Snails are often abundant and difficult to control, and it is in snails that the cercariae infective to humans are produced in prolific numbers. It takes only a single schistosome miracidium to establish a snail infection capable of producing hundreds of cercariae on a daily basis for months [2]. The amplification of schistosomes that occurs within snails creates a reoccurring problem for control efforts and is a significant obstacle for sustained prevention. It highlights the importance of understanding the dynamics of schistosome infections in snails and is the reasoning behind studies focused on characterizing the mechanistic basis for snail resistance to schistosome infection. If we could understand the underlying factors that enable snails to resist schistosome infection, then we could better understand the basis of compatibility in field snails. The level of compatibility exhibited will directly influence both transmission dynamics and control efforts. We could also potentially exploit resistance to favor development of more sustainable control strategies that go beyond today's largely one-dimensional control programs that depend primarily on treatment of infected people with praziquantel [3].
Not all snails are created equal: some are susceptible and some resistant to schistosome infection. Resistance is genetically controlled and affects immunological factors [4], [5] that vary among snail species, strains or age categories. For example, the human parasite Schistosoma mansoni infects only certain species of Biomphalaria (such as B. glabrata). Furthermore, only some strains of B. glabrata are compatible with this parasite. Many studies have focused on characterizing the transcriptional profiles of schistosome resistant strains compared to susceptible counterparts, and have identified a number of putative resistance-associated factors in the process [6], [7]. Amongst these molecules are the fibrinogen-related proteins (FREPs), members of a multi-gene family that undergo somatic diversification and point mutation events. FREP proteins couple together fibrinogen and immunoglobulin superfamily domains, to generate a protein that is unique as far as presently known to gastropod molluscs [8]. FREPs are capable of precipitating secretory/excretory products from digenetic trematode sporocysts [9], and binding to diversified glycoproteins produced by parasites [10]. One individual FREP, FREP3, has been singled out for further study because of its role in the snail defense response against the trematode Echinostoma paraensei [4]. FREP3, like other FREPs, is a lectin-like molecule that recognizes a number of monosaccharides and is able to enhance the phagocytic uptake of targets, acting as an opsonin [11]. Knockdown of FREP3 in a normally resistant snail phenotype, and subsequent challenge of those snails with E. paraensei resulted in a significant proportion of the snails becoming infected with E. paraensei [11].
Trematode infection of a snail host is achieved, in part, by evading and suppressing the snail defense response. This provides a window for establishment of infection and then preventing the immune response from interfering with parasite development. These immune-evasion strategies can be observed in vitro [12], and also by transcriptional analysis [7], which suggests that many of the transcripts expressed by resistant snails during successful defense are suppressed in susceptible snails that become infected [11]. Immunosuppression is especially strong following exposure to E. paraensei, a parasite that can alter snail hemocyte morphology and interfere with hemocyte function [12], and that can suppress the expression of important immune molecules almost immediately upon entry into the snail [7]. One of the factors we identified as being suppressed by E. paraensei during infection is FREP3 [11]. This observation prompted us to use, in one of the experiments described below, a protocol first employed by Lie and Heyneman [13] in which pre-exposure of schistosome-resistant snails to E. paraensei was used to abrogate resistance to subsequent schistosome infection. We hypothesized specifically that this treatment would interfere with FREP3 expression (and likely with expression of other immune components as well), as compared to schistosome-resistant control snails not exposed to E. paraensei.
In this study, we report on the results of two different manipulations undertaken with the intention of abrogating resistance to S. mansoni in the naturally resistant BS-90 strain of B. glabrata. We first examined the effects of knocking down FREP3 using RNAi on the subsequent ability of BS-90 snails to support S. mansoni development. Secondly, we also expressly repeated the classic experiment of Lie et al. (1977) [14], using both BS-90 snails and accompanying microarray monitoring for the first time. We first exposed BS-90 snails to radiation-attenuated miracidia of E. paraensei, and then assessed their resistance level to S. mansoni as compared to snails not pre-exposed to E. paraensei. Radiation-attenuated E. paraensei parasites do not establish long-term, proliferative infections in snails, avoiding the potential complication that persistent larvae of this species would prevent the potential development of S. mansoni.
It is known, however, that irradiated E. paraensei larvae, during their brief lifespan, exert a potent immunosuppressive effect just as do normal E. paraensei larvae [7], [11], [13]. Infection with S. mansoni of FREP3 knockdown snails and those first exposed to irradiated E. paraensei was also assessed by histological examination as well as by checking for shedding S. mansoni cercariae, which were tested for infectivity to mice. We compared the transcriptional profiles of BS-90 snails exposed only to irradiated E. paraensei to those exposed to irradiated E. paraensei and then challenged with S. mansoni. Our study seeks to demonstrate the involvement of a specific molecule in snail resistance to S. mansoni infection, and to provide a plausible natural mechanism by which trematode-mediated immunosuppression of the defense responses of a snail could facilitate infection by a parasite that it would normally successfully resist.
BS-90 and M-line strain Biomphalaria glabrata (B.g.) snails, and Schistosoma mansoni (S.m.) and Echinostoma paraensei (E.p.) were maintained as previously described [15].
Four independent 27 nucleotide oligos were designed to specific regions of FREP3 that displayed high conservation within the known diversified FREP3 transcripts. These oligos were combined and diluted in sterile snail saline at a final total concentration of 2 µg/µl, which was then injected into snails in a 5 µl volume. BS-90 snails were separated into two groups, the first to be injected individually with FREP3-specific siRNA oligos, and the second as a control, with GFP-specific oligos [16], siRNA oligo design and injection techniques have been previously described [11]. Four hours later, all snails were exposed individually to 30 S.m. miracidia. Snails were collected for histology at 2, 8, 18, 21, and 28 dpe. Snails were examined for the presence of infection (presence or absence of primary and secondary sporocysts) as described above for signs of infection at 21, 28, 34, 41, 48, and 54 dpe. Snails that shed cercariae were collected for histology and the rest were dissected to look for infections.
Knockdown of FREP3 was confirmed by RT-PCR and western blot analysis both of which have been previously described [11]. Specific knockdown of FREP3 protein levels was confirmed by probing the same samples with a FREP4 specific antibody. For both Western blot analyses, 100 µg of cell free plasma was loaded into each well of an SDS acrylamide gel. FREP3 was detected using a FREP3 specific antibody, and the Western blot was developed using the Supersignal West Femto Chemiluminescent Substrate (Pierce). FREP4 was detected using a FREP4-specific antibody and the Western Blot was developed using alkaline phosphatase. Injection of siRNA oligos and challenge of both FREP3 knockdown and GFP knockdown snails with S. mansoni resulted in similar mortality in both groups of snails. 36% of the FREP3 knockdown snails and 31% of the GFP knockdown snails died as a result of treatment.
In addition to RT-PCR and Western blot confirmation of FREP3 knockdown using FREP3-specific siRNA oligos as previously described [11], we confirmed the specificity of FREP3 knockdown using microarray analysis. BS-90 snails (4–8 mm) were injected with either FREP3 or GFP-specific siRNA oligos, and 2 hours later exposed to 30 S.m. miracidia. At 2 and 4 dpe, ten snails from each group were collected, RNA was extracted and then used to generate template for the microarray as previously described [15]. Ten arrays were completed. Each array was probed with template from an individual FREP3 knock-down snail labeled with Cy5 and an individual GFP knock-down snail labeled with Cy3. Hybridization, scanning, and analysis of the microarrays were previously described [7], using a significance cutoff of +/−log 1.5, and a false detection rate of 5%. Microarray results were submitted to GEO under the accession number GSE33525. The microarray revealed that indeed FREP3 expression was reduced at the transcriptional level by 2.4 fold at 2 dpe and by 5.1 fold at 4 dpe. The only other significant results from that array revealed a slight reduction in FREP13 expression by 1.2 fold at 2 dpe and 2 fold at 4 dpe and a slight up-regulation of TGFR-1 at 1.8 fold at 2 dpe and 2.6 fold at 4 dpe (Fig. S1). 18 other transcripts including FREP2 and FREP6 displayed slight alterations in expression however these changes were not considered statistically significant; none of these other 16 transcripts were FREPs.
To confirm viability of the cercariae produced from the FREP3 knock-down BS-90 snails that shed at 31 dpe we collected all cercariae produced (∼150), and exposed one mouse, using standard procedures as previously described [17]. Seven weeks post-exposure, the mouse was injected with a heparin solution and perfused by cutting the hepatic portal vein and injecting a standard RPMI medium into the heart. S. mansoni adult worms were collected and the liver was homogenized to collect S. mansoni eggs. The presence of adult worms confirmed the cercariae isolated from BS-90 snails were viable and the miracidia hatched from the eggs were also viable, being able to infect M-line B. glabrata snails (data not shown).
Size (4–8 mm shell diameter) matched snails were distributed into five groups: 1). BS-90 exposed to 25–30 irradiated E.p. miracidia at day 0 and secondarily challenged 4 days later with 15 S.m. miracidia, 2). BS-90 exposed to 25–30 irradiated E.p. only at day 0, 3). BS-90 unexposed control, 4). BS-90 exposed to only 15 S.m. miracidia at day 4, and 5). M-line exposed to only 15 S.m. miracidia at day 4 to confirm S.m. infectivity. For groups 2 and 3, RNA was collected at 1, 2, and 4 days post-exposure (dpe) to E.p. RNA was collected from groups 1, 2 and 4 at 1, 2, 4, and 8 dpe to S.m. and snails were collected for histology from groups 1 and 4 at 2, 8, 18 and 28 dpe to S.m. Snails from group 5 at 18 dpe to S.m. were also collected for histology. At days 18, 28, and 34 post-exposure to S.m., all remaining snails were placed into large tissue culture wells with artificial spring water and examined for the presence of developing primary sporocysts in the head foot or mantle, or secondary sporocysts in the mantle or digestive gland/ovotestes. Snails that were shedding S.m. cercariae were collected for histology. All snails that did not shed cercariae, were individually placed in snail saline, dissected and examined with the aid of a dissecting microscope for any signs of infection (sporocysts, germ balls, cercarial embryos) which dissection of known infected snails indicates can be seen under the 40× magnification used. Irradiation of E.p., and RNA extraction were previously described [15].
RNA was collected from whole snails at 2 and 4 dpe to 15 S.m. miracidia from groups 1 and 2 above, and was used to generate template for the microarray as previously described [15]. Each array was probed with RNA from an individual snail from the experimental group (1 from above), labeled with Cy5 and with RNA from a snail from control group 2, labeled with Cy3. There were twelve arrays in total, six from 2 dpe and six from 4 dpe, as a previous study revealed a great differentiation in transcription between these two time points [7]. Hybridization, scanning, and analysis of the microarrays were previously described [7], using a significance cutoff of +/−log 1.5, and a false detection rate of 5%. Microarray results were submitted to GEO under the accession number GSE28293.
Snails were collected and placed whole into tubes containing Railliet-Henry's fixative (930 ml H2O, 50 ml formalin, 20 ml acetic acid, and 6 g NaCl) to both fix the tissue and dissolve the shell. Any remaining shell was removed before the tissue was transferred into 10% buffered formalin. All tissue processing, sectioning, mounting, and hematoxylin and eosin staining was performed by TriCore Reference Laboratories in Albuquerque, New Mexico. The images generated from these sections were taken using a Nikon D5000 SLR camera attached to a Zeiss Axioskop compound microscope with an MM-SLR adapter and T-mount by Martin Microscope Company.
Specific siRNA-mediated suppression of FREP3 expression in BS-90 snails was confirmed at both the transcriptional (Fig. 1A) and protein levels (Fig. 1B) using RT-PCR and Western blot respectively. To assess whether FREP3 participated in an anti-S. mansoni defense response a total of 70 S. mansoni-resistant BS-90 strain snails were injected with FREP3-specific siRNA oligos to assess the impact of FREP3 knockdown on the subsequent ability of S. mansoni to develop. Knockdown of FREP3 resulted in cercariae-producing S. mansoni infections in 15 (21.4%) of these normally resistant snails (Fig. 1C). In contrast, none of 64 control BS-90 snails receiving GFP specific siRNA oligos shed cercariae. As a check of the viability of the S. mansoni miracidia used in this experiment, over 85% of schistosome-susceptible M-line snails exposed to infection in both trials became infected, a level of infection typical for exposure of such snails (Fig. 1C).
Histological observations revealed that S. mansoni miracidia penetrated snails receiving either FREP3 or GFP siRNA oligos. The early stage mother sporocysts (from 2 to 4 days post-infection) we observed were not conspicuously encapsulated in either group of snails. In most of the FREP3 knockdown snails that shed cercariae, shedding was light and intermittent over a 1–2 week observation period, after which they were fixed for histology at 31 days post-exposure to S. mansoni. Histological examination of S. mansoni-challenged FREP3 knockdown BS-90 snails revealed a small number of large sporocysts in the head-foot of each of these snails (Fig. 2 B, C). No disseminated daughter sporocysts were found in the digestive glands of these snails, however (Fig, 2A). The head-foot sporocysts had clearly grown considerably in size beyond that of young mother sporocysts, and whether they represented mother, or ectopic daughter sporocysts could not be determined. They were not encapsulated by hemocytes, nor were hemocytes prominently found near them. Developing cercariae were not seen within them but the sporocysts were of a size that easily could have supported cercariae development.
One of the infected FREP3 knock-down snails more persistently released cercariae over a 2 week observation period. Histological examination revealed this snail to have daughter sporocysts disseminated throughout the digestive gland. Hemocytes were conspicuous around them and encapsulation responses were noted (Fig. 2D). Only one of eight control BS-90 snails injected with GFP-specific siRNA oligos and sectioned at 28 days post-exposure to S. mansoni was observed to contain S. mansoni sporocysts, but they had not grown and did not contain germ balls.
BS-90 snails exposed to irradiated E. paraensei miracidia were challenged with S. mansoni miracidia 4 days later. After another 35 days, the snails were checked for shedding of viable S. mansoni cercariae, an indication that the infection was successful. Of 48 snails, 22 (46%) shed S. mansoni cercariae, compared to 0% (n = 35) of control BS-90 snails exposed to only S. mansoni (Fig. 3). To confirm the infectivity of the S. mansoni used, 22 M-line B. glabrata were challenged and 82% were successfully infected (Fig. 3). Histological comparison of S. mansoni cercariae-shedding BS-90 snails to normal resistant control BS-90 snails (Fig. 4A) showed they had disseminated S. mansoni sporocysts throughout the digestive gland (Fig. 4B, C) typical of normal infections. Snails exposed to irradiation-attenuated E. paraensei only did not develop disseminated E. paraensei infections, as expected. Degenerating E. paraensei sporocysts could be observed in the hearts of the sensitized snails, including those subsequently challenged with S. mansoni (Fig. 4E). To confirm the viability of the E. paraensei cohort used, BS-90 snails were exposed to non-irradiated control miracidia from the same cohort that was irradiated and were successfully infected by 28 dpe, as expected (not shown).
BS-90 snails first exposed to irradiated E. paraensei miracidia were challenged four days later with S. mansoni miracidia. Microarray analysis was then undertaken on individual snails either 2 or 4 days post-exposure to S. mansoni. Schistosome-specific markers on the array were used to indicate whether each snail had been successfully infected with S. mansoni, or if it had resisted the challenge. At both 2 and 4 days post S. mansoni challenge, 50% of the snails assayed using the array were positive for S. mansoni infections (3 positive, and 3 negative for S. mansoni for each time point). Immunosuppression (as indicated by the greater number of down-regulated than up-regulated features) resulting from exposure to irradiated E. paraensei miracidia was noticeable for all 12 snails studied with the arrays (Fig. 5A).
However, snails negative for S. mansoni markers displayed increased expression of a variety of known and putative defense-related factors (Fig. 5B). For some factors (FREP3, Dermatopontin, Heat shock protein 70, Superoxide dismutase 1 Cu-ZnA, Serpin B4, and Matrilin-1A) increased expression in snails negative for S. mansoni was contrasted by a suppression of expression in snails positive for this parasite. Other factors (FREP2, Coagulation factor XI, Dual oxidase, Galectin 4, Migration inhibition factor, Peroxiredoxin, and SOD Cu-Zn B) increased in expression in snails not infected by S. mansoni, but remained unaltered as compared to control values in snails that were successfully infected. FREP4 expression differed from other putative resistance molecules in that it was increased compared to control levels in both snails positive or negative for S. mansoni (Fig. 5B).
Schistosome parasites, including those that infect people, continue to thrive the world over, in no small measure owing their success to their productive use of snails as intermediate hosts. Particularly given that schistosome infection is harmful to the snail and results in its castration [18], it is reasonable to expect that the snail would mount defense responses to prevent infection. Although schistosomes obviously frequently prevail and establish long-term, infections, it is likely that many schistosome-snail encounters in the field result in failed infections. Such failures go overlooked but may well have a significant impact on transmission. Furthermore, the efficacy of present-day chemotherapy-based control operations could potentially be enhanced if we could also exploit snail resistance responses to further limit the number of new snail infections that arise. After all, it is in snails where cercariae - the source of reinfections in people that so frustrate control efforts - are produced in such prodigious numbers. To fully understand the potential impact of snail defenses on schistosome transmission to people, we need to achieve a better understanding of the mechanistic basis of snail defenses to infection, and how these defenses are overcome by schistosomes. In the process, we will also learn a great deal about the general nature of invertebrate (snail) defense mechanisms and the intimate interplay between host and parasite.
With respect to the immune responses of snails, our studies have lead us to focus on fibrinogen related proteins, or FREPs. One of the most noteworthy aspects of their biology is that two FREPs (first shown for FREP3, then FREP2) have been shown to undergo somatic diversification driven by gene conversion events and point mutations, creating a diversity of expressed sequences from a limited number of germ-line source sequences [10], [11], [19]. Recently, functional assessment of FREP3 demonstrated that it is capable of binding to carbohydrates and acts as an opsonin to enhance phagocytosis of targets by snail hemocytes. RNAi-mediated knockdown of FREP3 in snails resistant to the digenetic trematode Echinostoma paraensei resulted in an abrogation of resistance, resulting in one third of the snails developing established E. paraensei infections. Additionally, this study identified that FREP3, while increased in expression in resistant snails challenged with S. mansoni or E. paraensei, was suppressed in snails that were successfully infected by either parasite [11]. FREP2, another FREP that has the capacity for diversification, has been co-immuno-precipitated with S. mansoni polymorphic mucins, suggesting that this complex family of diversified parasite molecules may be the targets for FREPs [10].
Building on these earlier studies, here we demonstrate that FREP3 also plays a role in defense against S. mansoni infection. Knockdown of FREP3 resulted in 21% of the resistant BS-90 strain B. glabrata snails becoming successfully infected (shedding cercariae) with S. mansoni. In contrast, none of the 64 snails injected with GFP siRNAs shed cercariae. As previously hypothesized, FREP3 is likely working in combination with other defense mechanisms to manifest the resistant qualities of the BS-90 snails. However, this study clearly demonstrates that it is an important component of defense against S. mansoni.
Examination of sectioned snails revealed that S. mansoni miracidia penetrated both control GFP and experimental FREP3 knockdown snails, but observations of sporocysts at 2 and 8-days post-infection did not yield obvious evidence in either group of snails of sporocysts under conspicuous attack by hemocytes including within multilayered hemocyte capsules. Rather, sporocysts were found with only loose aggregates of hemocytes in their vicinity. This is compatible with observations reported by Galvan et al., 2000 [20] who noted that mother sporocysts of S. mansoni could remain viable in BS-90 snails for as long as 33 days. However, none of the S. mansoni-exposed BS-90 snails that they observed, nor any that we have observed over the years prior to this study, have ever shed cercariae. Our observations suggest that the inability of S. mansoni to thrive in BS-90 snails - at least in some cases - may be more dependent on inhibitory humoral factors than on overt hemocyte aggression and dismemberment. For example, humoral factors might serve to inhibit S. mansoni larval development or nutrition acquisition.
In all but one of the BS-90 FREP3 knockdown snails from which cercariae were shed, cercariae production must have originated from a small number of sporocysts in the head-foot of the snail. This likely explains why cercariae were produced by them in small numbers and intermittently. The daughter sporocysts producing these cercariae were either within or adjacent to the mother sporocyst that produced them. As these snails were fixed for histology, it is not clear how long they might have persisted in shedding cercariae. We suggest FREP3 knockdown in these snails allowed sporocysts to persist and enlarge, but was insufficient to enable them to proliferate and establish disseminated infections in the digestive gland. Hemocytes were not prominent around the head-foot sporocysts suggesting they had acquired some ability to protect themselves from attack. In snails receiving GFP siRNAs, in only one snail examined could sporocysts be found. They were small and showed no evidence of germ ball development. Based on these results, one possibility is that FREP3 plays a role in suppressing development of S. mansoni sporocysts in BS-90 snails, and if its effects are temporarily reduced, sporocysts may be released from this inhibition sufficiently well to enable some sporocyst development and multiplication to occur. As the knock-down effects inevitably wane, then the sporocysts may be prevented from further development such that proliferative infections do not usually result.
For the one FREP3 knockdown snail noted to have a disseminated infection, hemocytes accumulated in the digestive gland and in some cases were seen to be encapsulating daughter sporocysts. This is reminiscent of what was noted by Lie et al. [21] in some of the 10-R2 B. glabrata snails they observed in which resistance to S. mansoni had been broken down by pre-exposure of these snails to irradiated miracidia of E. paraensei. In about 30% of these snails, “self-cure” was eventually noted, characterized by hemocyte reactions to daughter sporocysts. Results of both experiments imply that snails inherently resistant to S. mansoni can reinvigorate an effective resistance response later in the course of infection, even though their ability to prevent establishment and development of infection had been earlier compromised by experimental manipulation. This suggests that the machinery for generating resistance is still intact. Furthermore, even though their collective biomass is large, daughter sporocysts may not be as effective as newly-penetrated (and much smaller) mother sporocysts in preventing effective responses.
As noted in the previous paragraph, and initially documented in studies by Lie and co-workers (13,25), both normal and irradiated sporocysts of E. paraensei have a potent ability to interfere with the resistance of B. glabrata to trematode infection. Their classic work has since stimulated a number of studies to reveal the underlying mechanisms of immunosuppression. Hemocytes collected from B. glabrata infected with E. paraensei exhibited reduced adhesive, spreading and phagocytic capacity compared to uninfected controls [22], [23]. B. glabrata hemocytes exposed to live E. paraensei sporocysts in vitro actively move away from the parasite [12], and eventually lose adherence to the substrate if exposed to parasite excretory/secretory (ES) products [24]. Furthermore, (ES) products of a related parasite, Echinostoma caproni, significantly impact the functional capacity and behavior of snail hemocytes, including a loss of adhesion, spreading and phagocytosis [25]. As these effects seem to be specific to suitable snail hosts, not extending to echinostome-resistant snail strains [25] or species [12], [26], the mechanism of their effects must be tailored to specific aspects of the defense system of compatible snails.
To pursue the molecular basis of E. paraensei-induced immunosuppression, we have followed the transcriptional responses of exposed snails using microarrays. By as early as 12 hours post-exposure, E. paraensei provokes down-regulation of snail defense responses, including FREP3 expression [7]. Many of the targets of E. paraensei immunosuppression are putative or known resistance-associated factors such as FREPs 1, 3, 5, 8, 9, and 10 [8], migration inhibition factor [27], dermatopontin [28], alpha-2- macroglobulin receptor [29], mannose receptor [30], peroxiredoxin [31], and galectins 4 and 7 [32]. Reduction in the presence of these factors theoretically would impact many aspects of defense function such as activation and phagocytosis by hemocytes (FREPs, mannose receptor, alpha-2- macroglobulin receptor), intra- and extra-cellular killing (peroxiredoxin), hemocyte adhesion and encapsulation (dermatopontin, migration inhibition factor), and coagulation (galectins).
Based on these results, we sought to repeat the basic design of the experiment of Lie et al. [14], to see if we could use pre-exposure of irradiated miracidia of E. paraensei to interfere with resistance of BS-90 snails to S. mansoni. Our experiment represents the first repeat of their classic experiment that has been accompanied by molecular (microarray) studies, and it is the first experiment to employ the naturally resistant BS-90 snails as hosts as opposed to other resistant B. glabrata snails of the 10-R2 or 13-16-R1 strains that were bred and selected for resistance [14].
We show that irradiated E. paraensei sporocysts suppress the BS-90 defense response sufficiently to allow 46% of the snails so treated to develop patent S. mansoni infections. When snails that permitted S. mansoni development were compared with those that did not using the B. glabrata microarray, we observed a number of immune-relevant transcripts that exhibited expression patterns indicating S. mansoni contributed to the suppression as well. FREP 2 and 3, coagulation factor IX, dermatopontin, dual oxidase, galectin 4, MIF, peroxiredoxin, superoxide dismutase Cu-Zn, and heat-shock protein 70 all exhibited increased expression in snails that successfully resisted infection compared to those that were infected by S. mansoni. Thus, we confirm previous hypotheses [7] suggesting that S. mansoni also utilizes a program targeted at suppressing the expression of important defense factors involved in killing larval parasites. This past work indicates that S. mansoni and E. paraensei differ in the timing and targets suppressed, E. paraensei beginning aggressive immunosuppression by 12 hours post challenge, S. mansoni beginning between 2 and 4 days post challenge [7].
Our results also suggest that irradiated echinostome larvae are more effective than our FREP3 knockdown protocol in protecting S mansoni sporocysts in resistant snails. This may be because irradiated echinostomes provide more persistent down-regulation of FREP3, and also have effects on other immune factors as well. The irradiated echinostome experiment indicates that if S. mansoni sporocysts are sufficiently protected, they can reliably develop disseminated infections in resistant snails. This may be because irradiated echinostomes provide S. mansoni sporocysts a longer interval to acquire and express their own immunosuppressive effects.
Our studies indicate that both echinostomes and schistosomes employ means of immunosuppression to colonize snails, and that this property can be manipulated to increase the breadth of strains of a single species, B. glabrata, that can be colonized. This work also bears on two important related general issues in parasitology, host specificity and host switching. Even though most digenetic trematodes are very host specific with respect to their choice of snail hosts, phylogenetic studies suggest that host-switching with respect to snails has been common in the history of trematodes like schistosomes [33]. The suppression we document offers one potential mechanism to resolve this apparent paradox: down-regulation of defense responses by one parasite may open the door for colonization of another parasite normally incompatible with that host. Field studies indicating the ability of one trematode to facilitate infection with another are consistent with this possibility [34], [35]. Cercariae produced in this study, both from BS-90 B. glabrata infected by S. mansoni due to reduced FREP3, or E. paraensei-mediated immunosuppression, were viable and able to infect mice. Thus, there is the potential for continuation of a trematode life cycle from a normally resistant snail host. It remains to be seen whether eggs produced from these mice have improved success at infecting BS-90 B. glabrata. We suggest that this study provides proof of principle that parasite-induced immunosuppression improves the chances that normally incompatible parasites can be successful in new, and hostile host environments. Furthermore, it provides a specific mechanism and molecules to target for future studies aimed at experimentally studying host specificity and host switching.
Another potential application of this work relates to the role of FREP3 in resistance of wild B. glabrata to infection with S. mansoni. Although it is clear that other factors are involved in resistance, this line of work suggests efforts to up-regulate FREP3 expression in snails from natural populations could have the effect of diminishing S. mansoni infections. We now must focus our efforts on understanding whether snails from endemic areas mount FREP3 responses following exposure to natural schistosome infections. It also raises the question as to whether snails differ in their inherent FREP3 responsiveness, and if this trait can be manipulated or favored to diminish natural schistosome infections.
|
10.1371/journal.pgen.1000485 | Additions, Losses, and Rearrangements on the Evolutionary Route from a Reconstructed Ancestor to the Modern Saccharomyces cerevisiae Genome | Comparative genomics can be used to infer the history of genomic rearrangements that occurred during the evolution of a species. We used the principle of parsimony, applied to aligned synteny blocks from 11 yeast species, to infer the gene content and gene order that existed in the genome of an extinct ancestral yeast about 100 Mya, immediately before it underwent whole-genome duplication (WGD). The reconstructed ancestral genome contains 4,703 ordered loci on eight chromosomes. The reconstruction is complete except for the subtelomeric regions. We then inferred the series of rearrangement steps that led from this ancestor to the current Saccharomyces cerevisiae genome; relative to the ancestral genome we observe 73 inversions, 66 reciprocal translocations, and five translocations involving telomeres. Some fragile chromosomal sites were reused as evolutionary breakpoints multiple times. We identified 124 genes that have been gained by S. cerevisiae in the time since the WGD, including one that is derived from a hAT family transposon, and 88 ancestral loci at which S. cerevisiae did not retain either of the gene copies that were formed by WGD. Sites of gene gain and evolutionary breakpoints both tend to be associated with tRNA genes and, to a lesser extent, with origins of replication. Many of the gained genes in S. cerevisiae have functions associated with ethanol production, growth in hypoxic environments, or the uptake of alternative nutrient sources.
| Genomes evolve in structure as well as in DNA sequence. We used data from 11 different yeast species to investigate the process of structural evolution of the genome on the evolutionary path leading to the bakers' yeast S. cerevisiae. We focused on an ancestor that existed about 100 million years ago. We were able to deduce almost the complete set of genes that existed in this ancestor and the order of these genes along its chromosomes. We then identified the complete set of more than 100 structural rearrangements that occurred as this ancestor evolved into S. cerevisiae and found that some places in the genome seem to be fragile sites that have been broken repeatedly during evolution. We also identified 124 genes that must be relatively recent additions into the S. cerevisiae genome because they were not present in this ancestor. These genes include several that play roles in the unique lifestyle of this species, as regards the intensive production and consumption of alcohol.
| Inferring the genome organization and gene content of an extinct species has the potential to provide detailed information about the recent evolution of species descended from it. If we know what was present in the genome of an ancestor, we can deduce how a current-day descendant differs from it. We can then ask questions about how it came to be different. The most recent changes in a genome are often the most interesting ones, because they reflect the most recent (or even current) evolutionary pressures acting on that genome [1],[2].
Yeast species offer the potential for the precise reconstruction of ancestral genomes, because many genomes have been sequenced and they show extensive colinearity of gene order among species [3]–[6]. As the number of sequenced genomes from related species rises, so does the precision with which we can reconstruct their history. In this study we compare the genomes of a group of species in the subphylum Saccharomycotina, spanning an evolutionary time-depth that is comparable to that of the vertebrates [7]. A whole-genome duplication (WGD) event occurred during the evolution of this subphylum [8], and we can compare the genomes of several species (including S. cerevisiae) that are descended from this event to the genomes of several species that branched off before the WGD occurred. We focus on an ancestor that existed approximately 100–200 Mya, at the point immediately before the WGD occurred. The evolutionary period beginning with this ancestor corresponds to a time during which the S. cerevisiae lineage became increasingly adapted to rapid fermentative growth [9],[10] and extensive rearrangement of the genome occurred (including the deletion of thousands of redundant copies of duplicated genes) [11].
Previous studies in other systems have employed both manual and computational approaches to reconstructing ancestral genomes. One of the most successful applications of computational methods has been the estimation of the ancestral order of orthologous genes in the common ancestor of 12 Drosophila species [12],[13]. Ancestral reconstruction is more difficult when ancient polyploidizations are present [14]. In studies of the 2R duplications in vertebrates, for example, the emphasis has been on establishing the ancestral gene content of paralogous chromosomal regions rather than on their precise gene order [15],[16]. We chose to use a manual, parsimony-based, approach to reconstructing the yeast ancestor at the point of WGD. The manual approach has the attractions of being tractable (whereas computational methods are still under development [17],[18]), of providing an independent result to which computational results can be compared, and of forcing us to examine every rearrangement event without prejudice as to what mechanism might have caused it.
Sankoff and colleagues [14],[17],[18] have developed computational methods that aim to reconstruct ancestral gene order in datasets that include polyploidizations. In recent work [18], they evaluated their ‘guided genome halving’ (GGH) algorithm by comparing its results to ours, using a preliminary version of the manually-derived ancestral yeast gene order that we report here as a ‘gold standard’. As currently implemented, the GGH algorithm can only consider input from a single post-WGD genome and 1–2 non-WGD outgroups, and only considers genes that are duplicated in the post-WGD genome.
Inferring the set of genes that existed in a yeast ancestor, and the order of those genes along the chromosomes, is of interest from both genome-evolutionary and organismal-evolutionary standpoints. Knowing the ancestral gene order enables us to trace all the inter- and intra-chromosomal rearrangements that occurred en route from this ancestor to the current S. cerevisiae genome, which is informative about the molecular mechanisms of evolutionary genome rearrangement and is also phylogenetically informative. Knowing the ancestral gene content allows us to identify genes that have been added to, or lost from, the S. cerevisiae genome during the past 100 Myr. Previous studies have shown that changes in gene content can provide a strong indication of changing evolutionary circumstances, either in cases of gene loss (such as the losses of GAL, DAL and BNA genes in Candida glabrata [1],[19],[20]) or in cases of gene gain (such as the ADH2 and URA1 genes of S. cerevisiae [9],[21],[22]). Even though it may not be possible to conclude that any particular gene gain was adaptive, the clear links between the functions of the gained genes ADH2 and URA1 and the adaptation of S. cerevisiae to a fermentative lifestyle [23] suggested to us that a systematic search for all the genes that were gained by S. cerevisiae since WGD would be worthwhile.
We used a manual parsimony approach to reconstruct the gene order and gene content of the yeast ancestor that existed immediately prior to WGD (Figure 1). The reconstruction was made by visually comparing the local gene orders in every region of the genome, stepping through the genome in overlapping 25-gene windows using the Yeast Gene Order Browser [YGOB; 6]. Initially, during 2007–08, we compared data from five post-WGD species (S. cerevisiae, S. bayanus, C. glabrata, Naumovia castellii and Vanderwaltozyma polyspora) and three non-WGD species (E. gossypii, Kluyveromyces lactis and Lachancea waltii) and inferred an ancestral genome based on these data. Later, in 2009, we added the genomes of three more non-WGD species (Zygosaccharomyces rouxii, L. thermotolerans and L. kluyveri [24]) and re-examined the whole genome window-by-window using YGOB. This process confirmed that our initial ancestral reconstruction was largely correct, but identified a few places where the gene content or local gene order in the ancestor needed to be revised. In particular, by adding data from more non-WGD species we were able in some cases to detect non-WGD orthologs of S. cerevisiae genes that are short and rapidly-evolving, which previously appeared to be unique to S. cerevisiae (for example, YLR146W-A).
The gene order and content of the ancestor were inferred as shown in Figure 1B,C. We first established the gene content, and then examined the adjacency relationships among these genes. Within any post-WGD species such as S. cerevisiae, most of the genome can be sorted into pairs of sister regions that have a double-conserved synteny (DCS) relationship with any non-WGD species such as L. waltii [25],[26]. Breaks in the DCS pattern correspond to two types of event, called single-breaks and double-breaks of synteny [26]. For each single-break of synteny (Figure 1B), because we have genome sequences from multiple post-WGD species, and because the endpoints of the chromosomal rearrangements in different species generally do not coincide, we can infer the species and chromosomal track on which the break happened. This inference also tells us the ancestral gene order across the site of breakage: in general, for a single break, the ancestral order has been disrupted in one track in one post-WGD species, but it is still conserved in the same track from the other post-WGD species, in the sister track from all the post-WGD species, and in the non-WGD species. Similarly for each double-break of synteny (Figure 1C), because we have multiple genome sequences from non-WGD species we can in general identify the break as having occurred in one particular non-WGD species. A small number of double-breaks of synteny are caused by situations where all the non-WGD species show one gene order but both of the tracks from all the post-WGD species show a different order. These breaks correspond to rearrangements that occurred on the branch between the Z. rouxii divergence and the common ancestor of the post-WGD species (before the WGD happened). We do not include these breaks in our analysis because we are only interested in events that occurred after the WGD.
Manual reconstruction by this method resulted in an inferred ancestral genome with eight chromosomes, containing 4703 protein-coding genes. The ancestral gene set represents the intersections of orthologous genes between non-WGD and post-WGD species, and between ohnologs (paralogs formed by WGD) across the post-WGD species. The ancestral genome is listed in Table S1 and can be browsed using YGOB (http://wolfe.gen.tcd.ie/ygob). Genes in this genome were given names such as Anc_1.125, meaning the 125th gene on chromosome 1 of the ancestor. The ancestral gene set accounts for 5158 (92%) of the 5601 genes currently present in S. cerevisiae (1088 ohnologs and 4070 single copy genes), which covers all genomic regions in S. cerevisiae except for the subtelomeric regions (discussed below). The S. cerevisiae genome can be mapped onto the inferred ancestral genome in 182 DCS blocks that tile together in an unambiguous 2∶1 fashion across the ancestral genome (Figure 2). Similarly, the other post-WGD species and non-WGD species can be mapped onto the ancestral genome, with 2∶1 and 1∶1 mappings, respectively, by the numbers of blocks shown in Figure 1A. The C. glabrata genome is much more rearranged (582 blocks) than S. cerevisiae as previously noted [19],[27]. The L. kluyveri genome is remarkably unrearranged, with the whole genome mapping into just 57 blocks relative to the ancestor.
Our inferred ancestral genome is incomplete in some regards:
Using the breakpoints between S. cerevisiae synteny blocks in the ancestral genome, we inferred the large scale chromosomal rearrangements that have occurred in the S. cerevisiae lineage since the WGD. Most rearrangement events could be classified as either reciprocal translocations (Figure 3) or inversions. Note that it is impossible to count inversions and reciprocal translocations with absolute precision, because if a genomic region that contains one endpoint of a reciprocal translocation subsequently undergoes inversion, the result is identical to one that could be produced by two successive reciprocal translocations (Figure S1). We counted these situations as two reciprocal translocations, so we have probably misclassified some inversions as reciprocal translocations. Inversions were defined as events where the two endpoints of the rearrangement were on the same ancestral chromosome and on the same post-WGD track.
In total we inferred 73 inversion events and 66 reciprocal translocations events on the evolutionary path from the ancestor to S. cerevisiae (Table 1). Five of the inversions have endpoints at telomeres. There were also five non-reciprocal translocations, which we call ‘telomeric translocations’ because they involved an exchange between a telomere and an internal region of another chromosome, which moved the end of an arm from one chromosome to another (one of these events occurs at a shared inversion/translocation breakpoint). The data indicate that some intergenic regions were re-used as breakpoints in more than one rearrangement event. We classified the rearrangements as consisting of 34 simple inversion events (not overlapping other inversions or reusing breakpoints), 39 complex inversion events (overlapping other rearrangements and/or reusing breakpoints), 44 simple reciprocal translocation events, and 22 reciprocal translocation events involving breakpoint reuse (Figure 4). These results are in reasonable agreement with our estimate from a decade ago of 70–100 rearrangement events, based only on S. cerevisiae data [33].
If some post-WGD species share a rearrangement relative to the ancestor but others retain the ancestral gene order, the rearrangement event is a phylogenetically informative character [34],[35]. We searched for rearrangements shared by any pair of post-WGD species. As described below, we found many that support the branching order of the post-WGD species shown in Figure 1A (Table S2). We did not find any shared rearrangements supporting alternative topologies. This result supports our previous conclusion, based on shared patterns of gene losses, that N. castellii is an outgroup to a clade containing C. glabrata and S. cerevisiae [11],[36]. In contrast, phylogenies based on sequence analysis tend to place C. glabrata outside N. castellii and S. cerevisiae [1],[37],[38], a result that we believe is an artifact.
Given this phylogeny, the post-WGD species define four temporal intervals for rearrangements (Figure 1A): (i) no rearrangements are shared by all the post-WGD species relative to the ancestor; (ii) 8 rearrangement events are shared by N. castellii, C. glabrata and S. cerevisiae (6 inversions, 1 reciprocal translocation, 1 telomeric translocation); (iii) 19 rearrangements are shared only by C. glabrata and S. cerevisiae, with N. castellii and V. polyspora retaining the ancestral organization (13 inversions, 6 reciprocal translocations); and (iv) 117 rearrangements are unique to S. cerevisiae or shared by this species and S. bayanus (54 inversions, 59 reciprocal translocations, 4 telomeric translocations). Most of the rearrangements that are specific to S. cerevisiae are temporally ambiguous relative to each other. We did not subdivide the group of 117 events into those that occurred before and after the S. bayanus divergence because the S. bayanus genome assembly is quite fragmented. The above analysis does not include gene transpositions, which we find to be relatively rare in yeast genomes but which are difficult to count precisely because to identify a transposed gene in a particular species, we need to be certain that it is orthologous to a gene at a non-syntenic location in the ancestral genome.
The lack of rearrangements in the first time interval is notable because it indicates that V. polyspora separated from the other lineages soon after the WGD. We also found that no rearrangements occurred on one genomic track of all the post-WGD species, relative to the other track and the ancestor, prior to this speciation. This observation argues against the possibility that the WGD event was an allopolyploidization rather than an autopolyploidization (see ref. [11] for discussion): if it was an allopolyploidization, then the two hybridizing genomes must have been completely colinear.
It was necessary to infer breakpoint reuse at the ends of some synteny blocks. Reused breakpoints appear as cycles in the map of breakpoint pairs (Figure 4). The evolutionary re-use of breakpoints has previously been identified in studies on mammals and Drosophila [13],[39]. We find that for both reciprocal translocations and inversions, there are fewer breakpoints than expected if every event had unique ends (Table 1). The average number of breaks per used site is 1.12 for reciprocal translocations and 1.16 for inversions. Some sites were used as endpoints of both an inversion and a reciprocal translocation, and if we pool these two categories there are only 228 unique breakpoints instead of the expected 278, implying an average of 1.22 breaks per site (Table 1).
We identified 96 sites in the ancestral genome at which genes are inferred to have been gained subsequently in the lineage leading to S. cerevisiae. The total number of gained genes is 124, because some sites contain groups of consecutive gained genes (Figure 5). We were surprised to find that 33 (34%) of these ‘gene gain’ sites are beside tRNA genes. tRNA genes have previously been linked to sites of genomic rearrangement between E. gossypii and S. cerevisiae [26]. Furthermore, it is known that origins of replication in yeast are often located near tRNA genes [40], and it seems plausible that origins might be fragile sites for evolutionary breakage and/or integration of new DNA. We used computer simulation to test the significance of the associations among tRNA genes, origins of replication, evolutionary breakpoints, and sites of gene gain (see Methods). tRNA genes are present at breakpoints and gain sites about three times more often than expected by chance (Table 2, rows 2 and 3), and origins are present about twice as often (Table 2, rows 4 and 5). It should be noted however that the locations of all the tRNA genes are known whereas it is probable that many origins have not yet been identified [41].
There are several plausible mechanisms by which tRNA genes could precipitate genomic rearrangements. tRNA genes exist in multiple near-identical copies in the genome, so illegitimate recombination between these sequences could result in reciprocal translocations [42],[43]. Ty retroelements tend to integrate beside tRNA genes and provide long sections of near-identical sequence scattered around the genome that could be substrates for ectopic recombination, as seen in S. cerevisiae irradiation experiments [44]. Ty LTRs, tRNA genes, and origins of replications have also all been associated with the endpoints of spontaneous segmental DNA duplications in S. cerevisiae [45]. Replication forks tend to stall near highly-expressed genes (such as tRNA genes), and sites of replication fork collapse are hotspots for chromosomal rearrangements [46],[47]. It is also possible that the Ty-encoded reverse transcriptase has played a direct role in the integration of new genes into sites beside tRNA genes, similar to the way that cDNA fragments of transcribed genes are sometimes captured at sites of double-strand break repair in S. cerevisiae experiments [48].
We identified 124 genes, excluding those in subtelomeric regions, that are inferred to have been gained on the lineage leading to S. cerevisiae during the time since WGD (Figure 5). The S. cerevisiae gene set that we used in this study consists only of genes that are conserved between S. cerevisiae and at least one of the other Saccharomyces sensu stricto species (dN/dS ratio<1 in the analysis of Kellis et al. [49]), or that are duplicates of other genes in S. cerevisiae (again with dN/dS<1), so we can be confident that the all the gains we identity are real genes and not annotation artifacts. Some of the gained genes are unique to S. cerevisiae and sensu stricto species, while others are shared by the other post-WGD species (Figure 5).
The 124 gained genes range from those with high similarity to another gene in the S. cerevisiae genome to those with no similarity to any known gene from any organism. We classified the gained genes into nine groups as described in Figure 5, and then into three larger categories according to their apparent mechanism of formation. The three large categories are:
Analysis of the functions of the gained genes should provide insight into the evolutionary pressures that have acted on S. cerevisiae in the period since WGD but, remarkably, there is no functional information in the Saccharomyces Genome Database (SGD) for almost half of the recently gained genes. None of the 124 genes is essential when deleted, according to SGD. The non-essentiality of gained genes is not surprising because they were gained by an organism that was already fully functional in its environment before they were gained. It is particularly notable that only 16 of the 51 orphans in Figure 5 have been assigned genetic names, which would indicate that something is known about their function.
In the sections below, we discuss some of the functional groups of gained genes. The gene information in these sections is derived primarily from summaries in the SGD and YPD databases [52],[53], and from a MIPS (Munich Information Centre for Protein Sequences) catalog analysis.
Reconstructing the content and gene order of the ancestral yeast genome just prior to WGD has provided a mechanism for studying the structural rearrangements that occurred subsequent to WGD. Our reconstruction is dependant on the set of extant genomes available for comparison, so it is likely that our list of candidate gene gains includes some false positives that will turn out to have been present at the time of WGD. As more genome sequences become available the ancestral gene set will become progressively more complete and the list of gains may shrink.
From a biological perspective, the main shortcoming of our work is that we were unable to reconstruct the telomeric regions of the genome, corresponding to the last ∼10 genes on each arm of each chromosome in S. cerevisiae. These regions turn over so dynamically that synteny breaks down almost completely between the species considered here. This is unfortunate because many of the most interesting evolutionary events such as the gain of genes by horizontal gene transfer (HGT) from other species [22], seem to occur preferentially near telomeres. Our set of candidate gene gains in S. cerevisiae contains only two possible cases of gene gain by HGT at internal chromosomal sites (YLR011W/LOT6 and YLR012W; we did not study these in detail), whereas Hall et al. [22] found eight examples of apparent transfer of bacterial genes into telomeric sites. A second shortcoming is that we relied on sequence conservation (dN/dS<1) among the sensu stricto species as a way of distinguishing between genuine S. cerevisiae genes and annotation artifacts, which had the inadvertent effect that we overlooked any genes that may have been gained by S. cerevisiae in the time since it diverged from the other sensu stricto species; one such case is BSC4, which appears to have been formed de novo in S. cerevisiae [87].
The set of genes inferred to have been gained on the S. cerevisiae lineage is relatively small (2% of the gene set) and their functions point squarely towards increasing adaptation to the ‘fermentative lifestyle’ [23]. They indicate increasing throughput of the glycolysis and fermentation pathways, and adaptation towards growth in conditions with little oxygen, including modifications to the cell wall and the bypass of biochemical pathways that require molecular oxygen by importing substances from outside the cell. There are also many gained genes in our set that we have not discussed in detail here because they did not fall into larger functional groups. Further analysis of these gains on an individual basis may reveal insights into the evolution of S. cerevisiae and the other species in the WGD clade.
In this paper we have adopted the revised genus nomenclature proposed by Kurtzman [88]: Saccharomyces castellii becomes Naumovia castellii; Kluyveromyces polysporus becomes Vanderwaltozyma polyspora; Ashbya gossypii becomes Eremothecium gossypii; Kluyveromyces waltii becomes Lachancea waltii; Kluyveromyces thermotolerans becomes Lachancea thermotolerans; and Saccharomyces kluyveri becomes Lachancea kluyveri. In this scheme each genus name refers to a monophyletic group, whereas previously Saccharomyces and Kluyveromyces were polyphyletic. We did not change any gene names, even though in many species the gene names have a prefix that is an acronym of the obsolete species name.
The numbers of double-conserved synteny (DCS) and synteny blocks between the reconstructed ancestor and each other species (Figure 1A) were counted automatically using an algorithm that smoothes over small inversions and other interruptions in cases where endpoints are ≤20 genes apart in the ancestral genome. For S. cerevisiae our manual analysis identified 228 breakpoints (Table 1 and Figure 4), which subdivide the 16 linear chromosomes into 244 segments. The discrepancy in numbers between these 244 segments and the 182 DCS blocks in S. cerevisiae (Figures 1A and 2) is due to the use of the smoothing algorithm.
We described the inferred ancestral gene order in terms of synteny blocks of current Saccharomyces cerevisiae genes. We manually identified intrachromosomal rearrangements (inversions) between the ancestor and S. cerevisiae and reversed them, revising our synteny blocks, in order to more easily identify the endpoints of reciprocal translocations. For each synteny block end not at a telomere, the location is at a position in the ancestral genome that underwent a reciprocal translocation in its transition towards the current S. cerevisiae genome. Each synteny block end in the ancestral genome is bordered by another synteny block found elsewhere in the current S. cerevisiae genome. The two breakpoints at the ends of two ancestral synteny blocks now adjacent in the current S. cerevisiae genome were created by a reciprocal translocation event that joined them together from different ancestral locations. Concurrently the other synteny blocks that border each breakpoint in the ancestral genome were joined together by the same event. We ordered the synteny blocks in the manner in which they are found along each chromosome in S. cerevisiae, thus inferring all the interchromosomal rearrangements between the ancestral polyploid genome and S. cerevisiae (Figure 4). To confirm that the inferred reciprocal translocation events were correct, we found the location in S. cerevisiae of the other synteny block ends joined by each event. In cases where these synteny blocks are not adjacent in the S. cerevisiae genome, we found the ancestral breakpoint locations of the blocks that are adjacent to each of these blocks in S. cerevisiae, inferring another reciprocal translocation event. If the synteny blocks bordering each breakpoint were again not adjacent, this process was repeated.
To obtain a set of likely gene gains (Figure 5) we subtracted the set of S. cerevisiae genes represented in the ancestral genome from the curated set of 5601 S. cerevisiae genes currently used in YGOB. YGOB's S. cerevisiae gene set is based on the SGD annotation (‘verified’ and ‘uncharacterized’ protein-coding loci only) with some additional manual curation. It omits loci that failed Kellis et al.'s test of reading frame conservation among sensu stricto species [49]. S. cerevisiae genes that are present in YGOB set but absent from the inferred ancestral set are candidates for having been gained in the S. cerevisiae lineage after the WGD. We did not include subtelomeric genes from the YGOB set, as orthologous relationships across species break down at the telomeres [49]. This candidate set of gains was then manually checked to ensure that there were no possible non-syntenic homologs that were ancestral but missing from our ancestral genome reconstruction due to a breakdown of synteny information. Any cases where a good candidate non-syntenic homolog was found were removed from the gained set and flagged as a likely transposition event. It is possible that the set of candidate gained genes may also contain ancestral genes that were lost in all of the non-WGD species used here but have orthologs in more distantly related outgroups.
We compiled lists of the 245 S. cerevisiae intergenic regions that contain one or more tRNA genes [from SGD; 52], the 228 intergenic regions that contain evolutionary breakpoints on the S. cerevisiae lineage, the 96 sites of gene gain in S. cerevisiae, and 267 intergenic regions that contain an origin of replication in S. cerevisiae (from OriDB [41]; we included origins that overlap with genes). We counted the numbers of intergenic regions that contain combinations of multiple types of site.
We then used computer simulation to estimate the significance of the observed numbers of coinciding sites (Table 2). In each of 1 million replicates we simulated a genome with 5100 intergenic spacers (the estimated number of intergenic spacers between S. cerevisiae genes that are at an ancestral locus). We placed the same numbers of tRNA genes, origins, breakpoints, and gene gain sites as above into randomly chosen spacers in the simulated genome. Each type of site was placed randomly and independently of the other types of site. We then counted the numbers of spacers containing all possible combinations of types of site in the replicate. Finally, we compared the observed numbers of coinciding sites in the real data to the distribution of results from the simulation (Table 2). The proportion of simulated genomes in which the number of sites with a particular colocalization pattern matches or exceeds the observed number of such sites in the real genome is an empirical measure of the statistical significance of the observation, under the null hypothesis of a random distribution of sites. We then applied a false discovery rate correction to these empirical P-values.
|
10.1371/journal.ppat.1005215 | Identification of the Mechanisms Causing Reversion to Virulence in an Attenuated SARS-CoV for the Design of a Genetically Stable Vaccine | A SARS-CoV lacking the full-length E gene (SARS-CoV-∆E) was attenuated and an effective vaccine. Here, we show that this mutant virus regained fitness after serial passages in cell culture or in vivo, resulting in the partial duplication of the membrane gene or in the insertion of a new sequence in gene 8a, respectively. The chimeric proteins generated in cell culture increased virus fitness in vitro but remained attenuated in mice. In contrast, during SARS-CoV-∆E passage in mice, the virus incorporated a mutated variant of 8a protein, resulting in reversion to a virulent phenotype. When the full-length E protein was deleted or its PDZ-binding motif (PBM) was mutated, the revertant viruses either incorporated a novel chimeric protein with a PBM or restored the sequence of the PBM on the E protein, respectively. Similarly, after passage in mice, SARS-CoV-∆E protein 8a mutated, to now encode a PBM, and also regained virulence. These data indicated that the virus requires a PBM on a transmembrane protein to compensate for removal of this motif from the E protein. To increase the genetic stability of the vaccine candidate, we introduced small attenuating deletions in E gene that did not affect the endogenous PBM, preventing the incorporation of novel chimeric proteins in the virus genome. In addition, to increase vaccine biosafety, we introduced additional attenuating mutations into the nsp1 protein. Deletions in the carboxy-terminal region of nsp1 protein led to higher host interferon responses and virus attenuation. Recombinant viruses including attenuating mutations in E and nsp1 genes maintained their attenuation after passage in vitro and in vivo. Further, these viruses fully protected mice against challenge with the lethal parental virus, and are therefore safe and stable vaccine candidates for protection against SARS-CoV.
| Zoonotic coronaviruses, including SARS-CoV, Middle East respiratory syndrome (MERS-CoV), porcine epidemic diarrhea virus (PEDV) and swine delta coronavirus (SDCoV) have recently emerged causing high morbidity and mortality in human or piglets. No fully protective therapy is still available for these CoVs. Therefore, the development of efficient vaccines is a high priority. Live attenuated vaccines are considered most effective compared to other types of vaccines, as they induce a long-lived, balanced immune response. However, safety is the main concern of this type of vaccines because attenuated viruses can eventually revert to a virulent phenotype. Therefore, an essential feature of any live attenuated vaccine candidate is its stability. In addition, introduction of several safety guards is advisable to increase vaccine safety. In this manuscript, we analyzed the mechanisms by which an attenuated SARS-CoV reverted to a virulent phenotype and describe the introduction of attenuating deletions that maintained virus stability. The virus, engineered with two safety guards, provided full protection against challenge with a lethal SARS-CoV. Understanding the molecular mechanisms leading to pathogenicity and the in vivo evaluation of vaccine genetic stability contributed to a rational design of a promising SARS-CoV vaccine.
| Coronaviruses (CoVs) are pathogens responsible for a wide range of existing and emerging diseases in humans and other animals [1]. A novel coronavirus causing the severe acute respiratory syndrome (SARS-CoV) was identified in Southeast China in 2002. SARS-CoV rapidly spread worldwide to more than 30 countries within six months, infecting 8000 people and leading to death in approximately 10% of the cases [2, 3]. While SARS-CoV has not reappeared in humans, CoVs including those similar to SARS-CoV, are widely disseminated in bats circulating all over the world, making future SARS-CoV outbreaks possible [4–7]. Furthermore, in September 2012, a novel coronavirus infecting humans, the Middle East respiratory syndrome coronavirus (MERS-CoV), was identified in two patients with severe respiratory disease in Saudi Arabia [8, 9], again indicating that emergence of other highly pathogenic CoVs is likely. Thus, development of efficacious and safe vaccines and anti-virus therapies for these pathogens is essential.
SARS-CoV is an enveloped virus with a positive sense RNA genome of 29.7 kb that belongs to the Coronavirinae subfamily, genus β [2]. The virion envelope contains embedded three structural proteins, spike (S), envelope (E), and membrane (M) and several group specific proteins: 3a, 3b, 6, 7a, and 7b [10–12]. The S protein, which mediates virus entry into host cells, the 3a protein and the M proteins, induce neutralizing antibodies, with those specific for S protein being most protective [13–16]. The SARS-CoV S and N proteins trigger T cell responses [17], which are also important for protection and enhance the kinetics of virus clearance [18, 19].
SARS-CoV E protein is a small integral membrane protein of 76 amino acids that contains a short hydrophilic amino-terminus followed by a hydrophobic transmembrane domain and a hydrophilic carboxy-terminus [20]. E protein oligomerizes to form an ion-conductive pore in membranes [21–23], and contains a PDZ-binding motif (PBM) formed by its last four carboxy-terminal amino acids [24, 25]. PDZ domains are protein-protein recognition sequences, consisting of 80 to 90 amino acids that bind to peptide sequences (PBMs) [26–28]. These protein-protein interactions modulate cellular pathways important for viral replication, dissemination in the host and pathogenesis [29]. We previously demonstrated that a SARS-CoV lacking the E gene (SARS-CoV-∆E) was attenuated in different animal models [30–34], indicating that SARS-CoV E protein is a virulence factor. SARS-CoV lacking the E protein fully protected both young and elderly BALB/c mice against challenge with virulent mouse-adapted SARS-CoV [32]; therefore, rSARS-CoV-∆E is a promising vaccine candidate. Live attenuated vaccines are considered highly effective because of their ability to replicate within host cells, resulting in high levels of antigenic stimulation, and robust long-term immunological memory [35, 36].
However, a major safety concern with live attenuated vaccines is the possibility of reversion to a pathogenic form. CoVs are prone to RNA recombination and mutation in tissue culture and during animal infection [37], so it is crucial that rSARS-CoV-∆E be thoroughly studied after serial passage. In this study, we show that passage of viruses lacking all or part of the E protein in Vero E6 cells and mouse DBT-mACE2 cells [38] led to the incorporation of compensatory insertions.
Interestingly, passage of SARS-CoVs lacking the E protein PBM led to regeneration of viral proteins containing PBMs, either by incorporation of novel chimeric proteins or by the insertion of new PBMs into existing proteins, such as the 8a protein or the E protein if E was only partially deleted. Strikingly, these modifications were not observed after passage of SARS-CoVs with mutated E protein that retained the PBM. In fact, we have shown that if instead of deleting the full-length E protein, only small deletions of 8–12 amino acids were introduced into the carboxy-terminus of E protein with retention of the PBM, the SARS-CoVs generated were attenuated and genetically stable both in cell culture or in mice. While this partial E protein deletion resulted in virus stability, we also augmented vaccine safety by introducing mutations into the SARS-CoV nsp1 protein. The nsp1 protein of CoVs suppresses host gene expression by inducing host mRNA degradation and inhibiting protein translation [39–44], and is an IFN antagonist [45–47]. Nsp1 deletions resulted in attenuated murine coronaviruses that fully protected against the challenge with parental virus [48, 49]. In this manuscript we show that small deletions within SARS-CoV nsp1 protein resulted in virus attenuation, associated with reduction of inflammation and higher levels of IFN-β and interferon-stimulated genes (ISGs). Vaccination with mutated nsp1 variants protected against challenge with the virulent mouse-adapted SARS-CoV (rSARS-CoV) virus. To generate safer vaccine candidates, viruses incorporating deletions in both the nsp1 and E proteins were constructed. These double mutants were protective against virulent virus challenge, and were genetically stable.
To determine the stability of SARS-CoV-∆E, or of virus containing deletions of the E protein and several group specific genes including 6, 7a, 7b, 8a, 8b and 9b (SARS-CoV-∆[E,6-9b]), we infected Vero E6 and DBT-mACE2 cells with rSARS-CoV, rSARS-CoV-∆E or rSARS-CoV-∆[E,6-9b]. Supernatants were serially passaged 16 times and the distal third of the genome, from the S gene to the 3´ end (around 8 kb), was sequenced using specific primers (S1 Table). In all cases, an insertion consisting of a partially duplicated M gene fused to the SARS-CoV leader RNA sequence, a 5´ sequence common to coronavirus mRNAs [50–52] was detected upstream of the native M protein (Fig 1A). In contrast, no chimeric proteins were detected after serial passage of the parental virus. All MCH genes encoded the amino terminus and the three transmembrane domains of M and also different PDZ-binding motifs at the carboxy-terminus of the protein (Fig 1B). Genomic evolution occurred rapidly, as the chimeric genes were already detected within 5 passages in both cells lines, Vero E6 and DBT-mACE2.
The expression of viral sgmRNAs corresponding to the chimeric genes was characterized by RT-PCR (S1A Fig). We used RNA harvested from infected cells after serial passage, plaque purification and amplification along with specific primers (S2 Table). PCR products corresponding to specific MCH sgmRNAs were identified in MCH-Vero and MCH-DBT-infected cells (S1B Fig). Expression of the chimeric proteins encoded by these sgmRNAs was confirmed using an antibody specific for all M and MCH proteins and a second one that recognized the MCH-DBT protein. Vero E6 cells were mock infected or infected with different recombinant viruses (rSARS-CoV, rSARS-CoV-∆E, rSARS-CoV-∆E-MCH-Vero and rSARS-CoV-∆E-MCH-DBT) at a moi of 0.3. The expression of native M and MCH was confirmed at 24 hpi by Western blot analysis (Fig 1C and 1D). These results indicated that, after serial passages of SARS-CoVs lacking the E protein in cell culture, a similar type of chimeric membrane protein was generated in three independent experimental settings.
To test whether the presence of MCH genes conferred a replication advantage to SARS-CoV-∆E in vitro, the growth kinetics of SARS-CoV-∆E-MCH-Vero (MCH-Vero) and SARS-CoV-∆E-MCH-DBT (MCH-DBT) were analyzed. Vero E6 and DBT-mACE2 cells were infected with the recombinant viruses (moi of 0.001) and viral titers were determined at the indicated hpi (Fig 2). MCH-Vero and MCH-DBT viruses showed lower titers at 24 hpi in Vero E6 cells compared to rSARS-CoV but both virus titers of the two chimeric viruses and rSARS-CoV were similar at 72 hpi. In contrast, a 100-fold decrease in viral growth was observed in rSARS-CoV-∆E-infected cells (Fig 2). Interestingly, chimeric proteins seemed to be specific for each cell type, as a ∆E virus containing a chimeric protein generated in Vero E6 cells (MCH-Vero) grew better in this cell line than in DBT-mACE cells, and the MCH-DBT virus generated in DBT-mACE2 cells specifically enhanced its growth in this cell line. This result indicated that the MCH protein provided a growth advantage for the virus, which partly compensated for the lack of the E protein.
To determine the effect of the MCH protein in pathogenesis, BALB/c mice were intranasally infected with the recombinant viruses using 100,000 pfu, and weight loss and survival were monitored for 10 days (Fig 3A). Mice infected with the parental virus (wt) showed signs of clinical disease at 2 days post infection (dpi), reflected by ruffled fur, shaking, loss of mobility and weight loss, resulting in the death of all mice at 6 dpi (Fig 3A). In contrast, mock-infected mice or mice infected with the ∆E virus, independently of whether the chimeric proteins (∆E, MCH-Vero and MCH-DBT) were present or absent, did not lose weight and all of them survived without symptoms of disease (Fig 3A).
To analyze the effect of the MCH protein on virus growth in vivo, BALB/c mice were intranasally inoculated with the recombinant viruses and euthanized at 2 and 4 dpi. Virus titers in the lungs were determined (Fig 3B). Viruses lacking the E protein, in the presence or absence of MCH protein, grew to lower titers in lungs at 2 and 4 dpi, as compared with those observed in mice infected with rSARS-CoV. Notably, rSARS-CoV-∆E replicated in the lung to higher levels than those containing the chimeric proteins (MCH-Vero and MCH-DBT) at both days p.i. Chimeric proteins only increased fitness in a cell type-dependent manner, i.e., viruses with the chimeric protein only grew better in the cell system in which this chimeric protein was generated. The virus containing a chimeric protein generated in Vero E6 cells grew better in this cell line, and the virus containing a chimeric protein generated in DBT-mACE2 cells showed an increased growth in these cells (Fig 2).
Lungs of mice infected with rSARS-CoV were highly edematous and showed profuse hemorrhagic areas at 2 and especially at 4 dpi (S2A Fig), leading to a significant increase in lung weight at 4 dpi (S2B Fig). Lung sections from mock-infected mice or mice infected with rSARS-CoV-∆E, MCH-Vero and MCH-DBT (Fig 3C) showed minimal damage at 2 and 4 dpi. In contrast, analysis of the lungs of mice infected with rSARS-CoV revealed extensive inflammatory cell infiltration and edema in alveolar and bronchiolar airways (Fig 3C).
The chimeric protein MCH was generated after rSARS-CoV-∆E passage in cell culture, but not when full-length E protein was present, suggesting that it compensated for functions originally performed by E protein. To identify such E protein functional domains, a set of recombinant SARS-CoVs (Fig 4A), with mutations or deletions in different regions of E protein [23, 24, 53], was passaged 16 times in Vero E6 cells (Fig 4). In Mut 1, several amino acid substitutions were introduced at the E protein amino-terminal region. rSARS-CoV deletion mutants ∆2, ∆3, ∆4, ∆5 and ∆6 included sequential or partially overlapping small deletions of 6 to 12 amino acids in the carboxy-terminus of E protein. Interestingly, the last 6 amino acids within the Δ6 virus (YSRVKN; Fig 4) revealed an alternative PBM at the carboxy-terminal domain of the protein [53]. In recombinant ∆PBM, the last 9 amino acids of E protein were deleted, truncating the carboxy-terminus and eliminating the E protein PBM. In mutPBM, the PBM was abolished by mutating the last 4 amino acids to glycine, maintaining the full-length E protein. In contrast, in altPBM, 4 amino acids within E protein carboxy-terminal region were mutated to alanine, maintaining an active PBM domain [24]. In SARS-CoV N15A and V25F mutants, ion channel activity of E protein was abolished by one point mutation in the transmembrane domain (Fig 4A). Vero E6 cells were infected with each of the recombinant viruses at a moi of 0.5 and supernatants were serially passaged for 16 times, and the presence of MCH gene was determined by sequence analysis. The results indicated that the MCH was not generated when the E protein contained a PBM sequence (rSARS-CoV, Mut 1, ∆2, ∆3, ∆4, ∆5, ∆6, altPBM, N15A and V25F) (Fig 4B). When the PBM was absent (∆PBM and mutPBM), a new PBM containing the original sequence was added to the carboxy-terminal end of mutated E protein in all cases, reinforcing the importance of the PBM domain during infection. A virus incorporating the chimeric MCH protein was only generated when E protein was completely deleted (∆E), i.e., when the restoration of a PBM on the E protein was not possible.
To further analyze whether SARS-CoV requires a transmembrane protein containing a PBM, we generated two recombinant rSARS-CoV-∆E that contained artificial chimeric proteins (Fig 4C). In SARS-CoV-∆E-MCH-EPBM (MCH-EPBM), a chimeric protein containing the first transmembrane domain of the M protein fused to the last nine amino acids of E protein, encompassing the PBM, was introduced. In SARS-CoV-∆E-3aCH-3aPBM (3aCH-3aPBM), the chimeric protein was formed by the first transmembrane domain of 3a protein and a PBM composed by the last nine amino acids of 3a protein (Fig 4C). Both viruses were passaged 16 times in Vero E6 cells and compensatory mutations were not detected after sequencing. All these data indicated that the virus requires a transmembrane protein displaying a PBM and that novel proteins with a PBM compensate for the loss of the E protein PBM.
MCH-Vero and MCH-DBT exhibited an attenuated phenotype (see above). To analyze the genetic stability of recombinant SARS-CoV-∆E in BALB/c mice, virus was passaged every 48 hours by intranasal inoculation. A partial duplication of 45 nucleotides was found within 8a gene (Fig 5), leading to the incorporation of a fragment of 15 amino acids at the carboxy-terminus of 8a protein, and generating a novel 8a protein (8a-dup) with an internal PBM (CTVV) (Fig 5A and 5B). To determine the virulence of this novel virus (SARS-CoV-∆E-8a-dup), we infected a new cohort of BALB/c mice and assessed survival and weight loss, and measured virus titers. Mice infected with SARS-CoV-∆E did not lose weight and all survived. In contrast, SARS-CoV-∆E-8a-dup grew to titers similar to those of rSARS-CoV and developed profound weight loss, developing signs of illness and death by 7 dpi (Fig 6A and 6B). Histological examination of lungs from SARS-CoV-∆E-infected mice showed absence of lung damage at both dpi. In contrast, lungs of mice infected with rSARS-CoV or SARS-CoV-∆E-8a-dup showed substantial perivascular, peribronchial and interstitial cellular infiltration and edema at 2 and 4 dpi (Fig 6C). To further confirm the relevance of the partial duplication within 8a gene in the induction of virulence, a recombinant SARS-CoV-∆E with an 8a-dup gene was generated (rSARS-CoV-∆E-8a-dup). Virulence during infection with rSARS-CoV-∆E-8a-dup was evaluated as described above (Fig 6D). The engineered virus (rSARS-CoV-∆E-8a-dup) was as virulent as the SARS-CoV-∆E-8a-dup generated after SARS-CoV-∆E passage in vivo (Fig 6D). These results indicated that SARS-CoV-∆E regained virulence after serial passage in mice.
SARS-CoV infection is associated with p38 mitogen-activated protein kinase (MAPK) activation and elevated levels of pro-inflammatory cytokines [24, 54–56]. To begin to determine the basis of increased virulence exhibited by SARS-CoV-∆E-8a-dup, p38 MAPK activation and pro-inflammatory cytokine expression were analyzed during infection (Fig 7). p38 MAPK activation was analyzed by Western blot at 24 hpi using a phospho-p38 MAPK (p-p38) specific antibody. Antibodies recognizing the total endogenous p38 MAPK and actin were used as controls. A significant increase in p38 MAPK activation, assessed at 24 hpi using a phospho-p38 MAPK (p-p38)-specific antibody and Western blot analysis, was observed in SARS-CoV-∆E-8a-dup-infected compared to SARS-CoV-∆E or mock-infected cells (Fig 7A and 7B).
To test whether pro-inflammatory cytokine expression was induced during SARS-CoV-∆E-8a-dup infection, we analyzed the expression of several genes previously associated with SARS-CoV pathology [24, 57] including: chemokine (C-X-C motif) ligand 10 (CXCL10), chemokine (C-C motif) ligand 2 (CCL2) and interleukin 6 (IL6) (S3 Table). 18S ribosomal RNA was used to normalize the data, as previously described [58, 59]. BALB/c mice were mock-infected or infected with 100,000 pfu of recombinant SARS-CoV, and lungs were collected at 2 dpi. A significant increase in the expression of all pro-inflammatory cytokines in the lungs was observed during infection with virulent viruses (rSARS-CoV and SARS-CoV-∆E-8a-dup) (Fig 7C). In contrast, infection with the recombinant virus lacking E protein at passage 0 (SARS-CoV-∆E) did not induce the expression of pro-inflammatory cytokines. Activation of p38 MAPK and pro-inflammatory cytokines expression during infection with rSARS-CoV-∆E-8a-dup were evaluated as described above. The engineered virus (rSARS-CoV-∆E-8a-dup) induced similar p38 MAPK activation (Fig 7D and 7E) and overexpression of proinflammatory cytokines (Fig 7F) as compared with the SARS-CoV-∆E-8a-dup generated after SARS-CoV-∆E passage in vivo. These data indicated that reversion of SARS-CoV-∆E-8a-dup to a virulent phenotype was associated with an exacerbated immune response similar to that triggered during infection with the rSARS-CoV.
SARS-CoV-ΔE reverted to a virulent phenotype after serial passages in mice. To increase the genetic stability of the vaccine candidate, we introduced small attenuating deletions in the E gene, instead of deleting of the full-length E protein [53] as described above. In addition, to increase vaccine biosafety, we introduced additional attenuating mutations within the SARS-CoV nsp1 gene. To identify domains within the SARS-CoV nsp1 protein that could contribute to virulence, we compared the sequence to that of the MHV nsp1 (Fig 8A), with the goal of identifying conserved regions that could be functionally important. Based on this information, four mutant viruses (rSARS-CoV-nsp1*) were generated by introducing deletions of 8 to 11 amino acids into the nsp1 protein (rSARS-CoV-nsp1-∆A, -∆B, -∆C and -∆D Fig 8A). All the deletion mutants grew to similar titers as rSARS-CoV in Vero E6 cells (Fig 8B). However the ∆A and ∆B mutants grew to lower titers in DBT-mACE2 cells.
To analyze the effects of these deletions in vivo, mice were intranasally infected with mutants rSARS-CoV-nsp1-∆A, -∆B, -∆C and -∆D, and daily monitored for 10 days. SARS-CoV-nsp1-∆C and -∆D infected mice transiently lost a small amount of weight and all mice survived. In contrast, mice infected with SARS-CoV lost weight, and all died by day 5 (Fig 9A). Mice infected with rSARS-CoV-nsp1-∆A or rSARS-CoV-nsp1-∆B lost 20 and 15% of their initial weight by day 3, with survival reduced to 60 and 80%, respectively (Fig 9A). These data indicated that deletion of regions C and D within nsp1 protein led to attenuated mutants, whereas deletion of regions A and B were only partially attenuating. With the exception of rSARS-CoV-nsp1-∆A at day 2 p.i., virus titers were reduced compared to rSARS-CoV-infected mice (Fig 9B). There was not a strict correlation between virus titers and virulence, possibly because nsp1 is involved in the countering IFN production after infection.
No significant changes on gross inspection of lungs or in their weight were observed when the lungs of mock-infected and SARS-CoV-nsp1-∆C and -∆D-infected mice were compared (Figs 9C and S3). In contrast, lungs from mice infected with the rSARS-CoV and, to a much lower extent SARS-CoV-nsp1-∆A and -∆B-infected mice lungs, showed evidence of hemorrhage (S3 Fig). In addition, rSARS-CoV-infected mice showed lung weight increase, not observed with the lungs of SARS-CoV-nsp1*-infected mice. Compared to mock-infected mice, lungs from rSARS-CoV-infected mice showed clear inflammatory infiltrates and alveolar and bronchiolar edema (Fig 9C). In contrast, mice infected with the viruses rSARS-CoV-nsp1* showed no (SARS-CoV-nsp1-∆C and -∆D), or minimal (SARS-CoV-nsp1-∆A and -∆B) lung damage. These data correlated well with the virulence observed for the SARS-CoV-nsp1* mutants, showing that the most attenuated viruses were those that induced less lung pathology in vivo.
Since nsp1 has anti-interferon activity, we next measured expression of IFN and IFN-stimulated genes after infection with rSARS-CoV, rSARS-CoV-nsp1-∆C and -∆D. We focused on the ∆C and ∆D viruses, because these viruses were fully attenuated and had an efficient growth in vivo. SARS-CoV-nsp1-∆C and -∆D induced higher levels of IFN-β and ISGs (IRF1, DDX58, and STAT1), compared to mock-infected and rSARS-CoV-infected cells (Fig 10A and 10B). This effect was specific, as the expression of control 18S rRNA was the same in virus-infected cells or mock-infected cells (Fig 10B). These results indicated that deletion of regions C and D of nsp1 restored IFN responses, leading to virus attenuation.
To analyze whether SARS-CoV-nsp1-∆C and -∆D induced protective immune responses, mice were intranasally vaccinated with SARS-CoV-nsp1-∆C and -∆D, and challenged 21 days later with rSARS-CoV. After challenge, mock-vaccinated mice rapidly lost weight, and all mice died by day 6 (Fig 11A and 11B). In contrast, mice immunized with SARS-CoV-nsp1-∆C and -∆D viruses, did not significantly lose weight, and 100% survived the challenge (Fig 11A and 11B).
In order to develop a safe vaccine candidate, mutant viruses with deletions in both nsp1 and E genes were engineered. A rSARS-CoV deleted in the nsp1 D domain and the E protein (SARS-CoV-nsp1ΔD-ΔE), and a second mutant virus with deletions of the nsp1 D domain coupled with a small deletion (E∆3) (Fig 4A) in the E protein (SARS-CoV-nsp1ΔD-EΔ3) were generated. EΔ3 deletion mutant was selected for further studies because this deletion led to a virus that grew to titers similar or higher than the SARS-CoV-∆E, in cell culture or in mice, respectively (Fig 12A and 12B). More importantly, the E∆3 virus was genetically stable after 10 passages in cell culture [53] or in vivo, maintaining its attenuated phenotype (Fig 12C), in contrast to the ∆E virus (Figs 1 and 5). This deletion was combined with another one in SARS-CoV nsp1 protein (nsp1ΔD), which was fully attenuating. The resulting virus grew to relatively high titers in vivo (Figs 9 and 13A). Viruses were rescued in Vero E6 cells, cloned and sequenced to confirm the presence of the desired mutations.
To analyze the stability of the viruses in tissue culture cells, SARS-CoV-nsp1ΔD-ΔE and SARS-CoV-nsp1ΔD-EΔ3 were passaged 10 times in Vero E6 cells, followed by sequencing of the nsp1 and E genes. The deletions introduced in both nsp1 and E genes were conserved, suggesting that these deletions were genetically stable in vitro. SARS-CoV-nsp1ΔD-ΔE and SARS-CoV-nsp1ΔD-EΔ3 at passage 1 reached peak titers at 72 hpi (5·105 pfu/ml and 5·104 pfu/ml in Vero E6 and DBT-mACE2 cells, respectively) (Fig 13B). Decreased virus growth was likely due to deletions in the E gene, as previously described [53] (Fig 8B). After 10 passages, viruses showed a slight increase in titer (Fig 13B), suggesting incorporation of additional mutations but not in E or nsp1.
To analyze the pathogenicity of SARS-CoV-nsp1ΔD-ΔE and SARS-CoV-nsp1ΔD-EΔ3 mutants at passages 1 and 10 (p1 and p10C, respectively), BALB/c mice were intranasally inoculated with recombinant viruses. All mice infected with these viruses maintained their weight and survived (Fig 14A). In contrast, all mice infected with rSARS-CoV died (Fig 14A). These results indicated that viruses including deletions in both nsp1 and E proteins were attenuated in vivo. SARS-CoV-nsp1ΔD-EΔ3, especially the p10C virus, grew more efficiently than SARS-CoV-nsp1ΔD-ΔE (Fig 14B). Nevertheless, no obvious gross lesions or changes in weight were observed in the lungs of mice infected with any of these doubly mutant rSARS-CoV (Figs 14C and S4). In contrast, mice infected with rSARS-CoV showed lung injury and a marked increase in the weight of lungs, as described above (S4 Fig). Histological examination of lungs from SARS-CoV-nsp1ΔD-ΔE and SARS-CoV-nsp1ΔD-EΔ3 (p1 and p10C viruses)-infected mice showed only minimal evidence of damage or leukocyte infiltration at days 2 and 4 post-infection (Fig 14C) while rSARS-CoV-infected mice showed extensive cellular infiltration and edema. The rSARS-CoV-nsp1∆D-E∆3 was selected for further study because this virus showed higher titers in vivo as compared with the rSARS-CoV-nsp1∆D-∆E virus (Fig 14B), therefore it could promote a higher immunization. In addition, ∆E mutation led to unstable viruses that incorporated new chimeric proteins in cell culture, or a novel 8a protein in vivo, causing reversion to a virulent phenotype (Fig 6). In contrast, viruses containing the E∆3 mutation remained stable after 10 passages and maintained their attenuated phenotype (Fig 12).
To analyze the stability of rSARS-CoV with mutations in nsp1 and E in mice, SARS-CoV-nsp1ΔD-EΔ3 was passaged 10 times (p10M), followed by sequencing from the S gene to the 3´ end of the genome. Only two changes were observed in the viral sequence, one in the E gene (A26250T N→I), and a second one in the M gene (A26450G Q→R). The deletions introduced in the nsp1 and E genes were conserved, suggesting that this virus was essentially genetically stable in vivo. To analyze whether the E and M mutations could be compensatory mutations, virus titers at p1 and p10 were compared in Vero E6 and DBT-mACE2 cells. Titers of SARS-CoV-nsp1ΔD-EΔ3 (p10M) at different times post-infection were the same as those observed for rSARS-CoV (Fig 15A), indicating that the mutations increased virus replication.
To evaluate whether the compensatory mutations restored the pathogenicity of the virus, BALB/c mice were intranasally inoculated with rSARS-CoV and SARS-CoV-nsp1ΔD-EΔ3 (p10), and were daily monitored for 10 days. Mice infected with rSARS-CoV started to lose weight by day 2, and died by day 6 (Fig 15B). In contrast, although mice infected with the SARS-CoV-nsp1ΔD-EΔ3-p10M mutant initially lost 10% of their weight, at day 5 the mice started to regain weight, fully recovered from the disease, and 100% survived (Fig 15B). SARS-CoV-nsp1ΔD-EΔ3-p10M grew to similar titers as rSARS-CoV (Fig 15C). Nevertheless, unlike the parental virus, lungs of SARS-CoV-nsp1ΔD-EΔ3-p10M-infected mice presented no significant increase of inflammatory cytokines. In addition, no obvious lung lesions, weight increases, nor substantial inflammatory cell infiltration, as determined by macroscopic and histological examination, were observed (S5 Fig). To further support the stability and attenuation of the double mutant virus, additional passages (up to 20) were conducted in mice. Evaluation of the passaged virus virulence showed that rSARS-CoV-nsp1∆D-E∆3 remained attenuated (Fig 15D). These results indicated that despite the mutations that the virus incorporated after their passage in mice, the virus maintained the in vivo attenuated phenotype.
To determine whether SARS-CoV-nsp1-ΔD-EΔ3 confers protection against challenge with rSARS-CoV, BALB/c mice were either immunized with SARS-CoV-nsp1ΔD-EΔ3-p1, -p10C and -p10M or non-immunized, as a control. At 21 days postimmunization, mice were challenged with rSARS-CoV administered by the same route. Non-immunized mice lost weight and all died on day 6 after the challenge (Fig 16A and 16B). In contrast, vaccination with the attenuated mutant viruses completely protected mice against challenge with rSARS-CoV (Fig 16A and 16B), indicating that the double mutant virus is a promising vaccine candidate.
We have previously shown that deletion of SARS-CoV E gene leads to an attenuated virus that is a promising vaccine candidate [30–34]. However, since safety and stability are main concerns of live attenuated vaccine candidates, we focused on rSARS-CoV-∆E stability in vitro and in vivo and on the generation of a safe vaccine candidate by identifying the mechanisms of reversion to virulence.
Unexpectedly, serial passage of rSARS-CoV-∆E in cell culture resulted in the generation of chimeric proteins composed of a partial duplication of the membrane gene fused to a part of the leader sequence. Our results are in agreement with a recombinant MHV lacking the E protein (rMHV-∆E) that was viable but its replication was drastically impaired [60]. rMHV-∆E replicated to 10,000-fold lower titers than the parental virus and remarkably, evolved similarly to SARS-CoV-∆E after serial passage in tissue culture cells [61]. Despite that the generation of a chimeric protein was previously observed in MHV when E gene was deleted [61], the presence of a PBM motif within the inserted chimeric sequence and its main role in providing genetic stability to SARS-CoV-ΔE during passage described in this manuscript, were not previously noticed. Alteration of coronavirus genome was not unexpected due to the high frequency of RNA recombination and mutation described for these viruses, both in cell culture and in animals [37, 62–64].
All chimeric SARS-CoV M proteins generated in cell culture were expressed, enhancing virus growth in cell culture compared to rSARS-CoV-∆E. Similarly, rMHV-∆E that contained chimeric proteins also showed significant increases in viral yields [61, 65]. In contrast, mice infected with rSARS-CoVs containing chimeric proteins generated in cell culture showed a decrease in viral titers in the lungs of infected mice even when compared with rSARS-CoV-∆E virus. Similarly, extensive passage in cell culture of other CoVs, including porcine epidemic diarrhea virus (PEDV) and transmissible gastroenteritis virus (TGEV), led to less pathogenic strains compared to wild-type viruses, possibly due to the emergence of deletion mutants that lost sequence domain not needed for their growth in cell culture, but that influenced their tropism and in vivo replication [66–68]. Chimeric proteins containing a PBM inserted in the viral genome after passage increased viral fitness in cell culture in a tissue specific manner, i.e., the chimeric protein inserted into viral genome during passage in monkey cells promoted virus growth in cells from this species, whereas the one inserted during passage in murine cells specifically increased virus fitness in cells from mice. Furthermore, these chimeric proteins did not enhance viral growth nor virulence in vivo. These data indicated that the insertion of chimeric proteins specifically adapted the virus for an optimum growth in cell culture but did not enhance in vivo growth nor virulence. In this context, it is also important to note that the activity of PBMs is dependent on their specific sequence, and also on the sequence context in which they are inserted [27, 28, 69, 70].
Serial passage of rSARS-CoV-∆E in mice introduced a partial duplication of 45 nucleotides in the 8a protein, resulting in its reversion to a virulent phenotype. This phenotype was associated to its ability to activate p38 MAPK and to the induction of inflammatory cytokine expression and increased lung damage, as previously described [24, 71].
8a protein is a short transmembrane protein composed of 39 amino acids that forms cation-selective ion channels [72]. SARS-CoV variants with deletions in 8a ORF, have been transmitted and maintained in humans in the late phases of SARS-CoV epidemic [73, 74]. Interestingly, an 8a protein mutant generated during virus passage in vivo contained a new potential PBM (CTTV) localized in the internal region of the carboxy-terminal domain of the protein (Fig 5). Despite the CTVV sequence was already present in the original 8a protein, it most likely does not represent a functional PBM, as it is located within the transmembrane domain of the 8a protein. Active PBMs are in general located in exposed regions of the proteins, usually the end of the carboxy-terminus or, exceptionally, in internal positions within the carboxy-terminal domain, allowing their interaction with PDZ domains, such as it has been observed in the NS5 proteins of tick-borne encephalitis virus (TBEV) and Dengue virus [75, 76]. However, PBMs forming part of a transmembrane domain are not accessible to PDZ-containing proteins and have not been described [29]. Therefore, the new CTVV sequence placed in an exposed environment may constitute a novel and active PBM. The PBM insertion within 8a protein after passaging in vivo could be due to the fact that the ORF8 is one of the regions where most variations were observed between human and animal isolates of SARS-CoV [4]. In fact, a complete genome sequence of SARS-like coronaviruses in bats isolates showed the presence of a PBM within the ORF8 [5]. As mentioned above, species of bats are a natural host of coronaviruses closely related to those responsible for the SARS outbreak.
PDZ domains are among the modules most frequently involved in protein-protein interactions found in all metazoans [69]. In the human genome, there are more than 900 PDZ domains in at least 400 different proteins [77]. Many pathogenic viruses produce PDZ ligands that disrupt host protein complexes for their own benefit, such as hepatitis B virus, influenza virus, rabies virus and human immunodeficiency virus, influencing their replication, dissemination in the host, transmission and virulence [29]. Generation of new proteins containing a PBM after passaging in cell culture and in mice may affect their interaction with a wide range of cellular PDZ-containing proteins, affecting diverse biological functions with high relevance in pathogenesis. E protein PBM participates in two different and independent issues, virus stability and virulence. Our data suggest that when the PBM was present in a proper environment at the end of the E protein, either a native or a mutant protein, viruses remained stable. An independent observation is that the presence of a PBM within E protein confers pathogenicity to the virus [24]. This virulence is prevented either by PBM removal [24] or by the introduction of small deletions within the carboxy-terminus of E protein [53], which by themselves may cause attenuation or, alternatively, by indirectly affecting the PBM. Our results highlighted the critical requirement of viral proteins containing a PBM in the generation of CoVs with virulent phenotypes, and opened up new approaches for the rational design of genetically stable vaccines.
Maintaining the attenuated phenotype of the vaccine candidate after passage in vitro was crucial to avoid the reversion to a virulent phenotype during the design and production of a genetically stable vaccine candidate. To this end, the identification of the relevance of the presence of a functional PBM motif at the carboxy-terminus of a transmembrane protein of the virus has been instrumental in the development of a stable SARS-CoV vaccine candidate. To minimize the risk of regain of virulence after passage, we engineered viruses with small deletions in E gene, instead of deletion of the entire E gene. By preserving the PBM, we observed no evidence for the development of chimeric proteins and thus no gain in virulence. As additional measures to ensure safety of this live attenuated vaccine candidate, we incorporated attenuating mutations into nsp1, in the context of the rSARS-CoV-ΔE or EΔ3. Nsp1 was chosen as a second attenuation target because this gene is located at a distant site (>20 kb) from that of the E gene in the viral genome, making it very unlikely that a single recombination event with a circulating wt coronavirus could result in the restoration of a virulent phenotype.
To analyze the role of SARS-CoV nsp1 in the pathogenesis of the virus, recombinant viruses encoding four different small deletions were generated. Deletion of amino acids 121–129 and 154–165, in the carboxy terminal region of nsp1 led to virus attenuation, indicating that nsp1 enhanced virus pathogenicity, as was previously shown for MHV [45, 48, 49]. Interestingly, these attenuated mutants grew in mice to lower titers than rSARS-CoV, probably by inducing higher IFN responses, indicating that these regions of nsp1 are critical for IFN antagonism. The induction of a higher innate immune response by the nsp1 deletion is most probably responsible for the decrease in SARS-CoV-nsp1* virus titers observed in mice and, to a lesser extent, in DBT-mACE–2 cells. In fact, a rSARS-CoV lacking the nsp1 protein grew poorly in IFN competent cells, but replicated as efficiently as the wt virus in IFN deficient cells [46], consistent with our findings. Similarly, titers of MHV deleted in nsp1 are restored almost to wild type levels in type I IFN receptor-deficient mice [48].
Immunization with singly deleted rSARS-CoV protected mice against challenge with rSARS-CoV, as it was previously shown with MHV nsp1-deletion mutants [48, 49]. SARS-CoV-nsp1ΔD-EΔ3, which contained deletions in nsp1 and E protein, maintained its attenuated phenotype after passage in Vero E6 cells and in mice. In addition, immunization with this double mutant fully protected mice from challenge with the parental virulent virus, indicating that it is a promising vaccine candidate in terms of both stability and efficacy.
Both humoral and cellular responses are relevant to protect from SARS [18, 19, 78, 79]. The viruses generated in this work express all viral proteins, except for small regions deleted in the E and nsp1 proteins, therefore have the potential of inducing both antibody and T cell responses, making this type of live vaccine more attractive than subunit or non replicating virus vaccines. Understanding of the molecular mechanisms by which an attenuated SARS-CoV reverted to a virulent phenotype could also be applied to the development of other relevant CoVs vaccines, such as MERS-CoV.
Animal experimental protocols were approved by the Ethical Committee of The Center for Animal Health Research (CISA-INIA) (permit numbers: 2011–009 and 2011–09) in strict accordance with Spanish National Royal Decree (RD 1201/2005) and international EU guidelines 2010/63/UE about protection of animals used for experimentation and other scientific purposes and Spanish national law 32/2007 about animal welfare in their exploitation, transport and sacrifice and also in accordance with the Royal Decree (RD 1201/2005). Infected mice were housed in a ventilated rack (Allentown, NJ).
The mouse-adapted (MA15) [71] parental virus (wt), and recombinant viruses were rescued from infectious cDNA clones generated in a bacterial artificial chromosome (BAC) in our laboratory [32, 33, 53, 80].
Vero E6 and BHK cells were kindly provided by E. Snijder (University of Leiden, The Netherlands) and H. Laude (Unité de Virologie et Immunologie Molecularies, INRA, France), respectively. The mouse delayed brain tumor (DBT) cells expressing the murine receptor (ACE2) for SARS-CoV (DBT-mACE2) were generated in our laboratory [38]. In all cases, cells were grown in Dulbecco's modified Eagle's medium (DMEM, GIBCO) supplemented with 25 mM HEPES, 2 mM L-glutamine (SIGMA), 1% non essential amino acids (SIGMA) and 10% fetal bovine serum (FBS, Biowhittaker). Virus titrations were performed in Vero E6 cells as previously described [33].
8 week-old specific-pathogen-free BALB/c Ola Hsd mice females were purchased from Harlan Laboratories. BALB/c mice were maintained for 8 additional weeks in the animal care facility at the National Center of Biotechnology (Madrid). For infection experiments, mice were anesthetized with isoflurane and intranasally inoculated at the age of 16 weeks with 100,000 plaque forming units (pfu) of the indicated viruses. All work with infected animals was performed in a BSL3 laboratory (CISA, INIA).
Mutant viruses (SARS-CoV-nsp1*) with small deletions covering different regions of nsp1 protein (SARS-CoV-nsp1-∆A, -∆B, -∆C and -∆D), were constructed using an infectious cDNA clone. cDNA encoding the genome of SARS-CoV-MA15 strain was assembled in a bacterial artificial chromosome (BAC) (plasmid pBAC-SARS-CoV-MA15) [32, 57, 80]. DNA fragments containing nucleotides 8142 to 9211, comprising the nsp1 gene of the SARS-CoV genome were generated by overlap extension PCR using as template the plasmid pBAC-SARS-CoV-MA15 and the primers indicated in S4 Table. The final PCR products were digested with the enzymes AvrII and BstBI and cloned into the intermediate plasmid pBAC-SfoI-MluI-SARS-CoV that contains the first 7452 nucleotides of the SARS-CoV infectious cDNA clone [80], to generate plasmids pBAC-SfoI-MluI SARS-CoV-nsp1* (pBAC-∆A, -∆B, -∆C, and -∆D) [80]. The plasmids pBAC-SfoI-MluI-SARS-CoV-nsp1* were digested with the restriction enzymes SfoI and MluI and the fragments were inserted into the plasmid pBAC-SARS-CoV-MA15, digested with the same restriction enzymes, to generate pBAC-SARS-CoV-MA15-nsp1* plasmids. Mutant viruses SARS-CoV-nsp1ΔD-ΔE and SARS-CoV-nsp1ΔD-EΔ3 were generated using the plasmids pBAC-SARS-CoV-MA15-ΔE and -EΔ3 [33, 53]. The plasmids were digested with the enzymes BamHI and RsrII and the digested fragments were exchanged with the fragment of plasmid pBAC-SARS-CoV-MA15-nsp1∆D, to generate pBAC-SARS-CoV-MA15-nsp1ΔD-ΔE and pBAC-SARS-CoV-MA15-nsp1ΔD-EΔ3 plasmids. Two fragments representing the nucleotides containing the chimeric proteins MCH-EPBM and 3aCH-3aPBM were chemically synthesized (BioBasic Inc) to generate SARS-CoV mutants. The final PCR products and synthesis fragments were digested with enzymes BamHI and MfeI and cloned into the intermediate plasmid psl1190+BamHI/SacII-SARS-CoV to generate the plasmids psl1190-∆E-MCH-EPBM and psl1190-∆E-3aCH-3aPBM. The plasmid psl1190+BamHI/SacII SARS-CoV contains a fragment corresponding to nucleotides 26045 to 30091 of the SARS-CoV infectious cDNA clone engineered into plasmid psl1190 (Pharmacia) [80]. These constructs were cloned in the infectious pBAC-SARS-CoV-MA15-∆E with the enzymes BamHI and SacII. Mutant virus rSARS-CoV-∆E-8a-dup with small duplication of 8a protein was constructed using an infectious cDNA clone. cDNA encoding the genome of SARS-CoV-MA15-∆E strain was assembled in a bacterial artificial chromosome (BAC) (plasmid pBAC-SARS-CoV-MA15-∆E) [32, 57, 80]. DNA fragments containing nucleotides 27779 to 27898, comprising the 8a gene of the SARS-CoV genome were generated by overlap extension PCR using as template the plasmid pBAC-SARS-CoV-MA15 and the primers indicated in S4 Table. The final PCR products were digested with the enzymes XcmI and NheI and cloned into the intermediate plasmid pBAC-BamHI-NheI-SARS-CoV that contains the nucleotides (nt) 26044 to 28753 nucleotides of the SARS-CoV infectious cDNA clone [80], to generate plasmid pBAC-BamHI-NheI-SARS-CoV-∆E-8a-dup [80]. The plasmid pBAC-BamHI-NheI-SARS-CoV-∆E-8a-dup was digested with the restriction enzymes BamHI and NheI and the fragments were inserted into the plasmid pBAC-SARS-CoV-MA15, digested with the same restriction enzymes, to generate pBAC-SARS-CoV-MA15-∆E-8a-dup plasmid. The viruses were rescued in BHK and Vero E6 cells as previously described [33]. Viruses were cloned by three rounds of plaque purification.
Subconfluent monolayers (90% confluency) of Vero E6 and DBT-mACE2 on 12.5 cm2 flasks were infected at a multiplicity of infection (moi) of 0.001 with the indicated viruses. Culture supernatants were collected at 0, 4, 24, 48 and 72 hpi and virus titers were determined as previously described [33].
BALB/c mice were anesthetized with isoflurane and intranasally inoculated with 100,000 plaque forming units (pfu) of virus in 50 μL of DMEM. Weight loss and mortality were evaluated daily. For protection experiments mice were immunized intranasally with 6000 pfu of the attenuated viruses, and then challenged with an intranasal inoculation of 100,000 pfu of SARS-CoV at 21 days post-immunization. Mice were monitored daily for weight loss and mortality. To determine SARS-CoV titers, lungs were homogenized in PBS containing 100 UI/ml penicillin, 0.1 mg/ml streptomycin, 50 μg/ml gentamicin, and 0.5 μg/ml amphotericin B (Fungizone), using a gentleMACS dissociator (Miltenyi Biotec) and virus titrations were performed in Vero E6 cells as described above. Viral titers were expressed as pfu/g tissue.
Mice were sacrificed at 2 and 4 dpi. Lungs were removed, fixed in 10% zinc formalin for 24 h at 4°C and paraffin embedded. Histological examination was performed using hematoxylin and eosin staining of sections.
BALB/c mice were anesthetized with isoflurane and intranasally inoculated with 100,000 pfu of the indicated recombinant viruses in 50 μL of DMEM. Two days after inoculation, mice were euthanized, and their lungs were removed and homogenized as previously described. The lung homogenate was clarified by low-speed centrifugation at 3,000 rpm for 12 min, and 100 μL of the supernatant was administered intranasally to naive mice. Intranasal inoculation of BALB/c mice with clarified supernatants of lung homogenates collected 2 dpi was repeated 10 times.
Cell lysates were resolved by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE), transferred to a nitrocellulose membrane by wet immunotransfer and processed for Western blotting. The blots were probed with monoclonal antibodies for p38 MAPK (dilution 1:500; Cell Signaling), phospho-p38 MAPK (dilution 1:500; Cell Signaling) and actin (dilution 1:10,000; Abcam) or polyclonal antibodies specific for M (dilution 1:1000; Biogenes) and MCH-DBT (dilution 1:1000; Biogenes) proteins. Both polyclonal antibodies recognizing the parental SARS-CoV M protein or the MCH-DBT protein were generated by Biogenes (Germany) as previously described [81] using synthetic peptides corresponding to the residues RTRSMWSFNPETNILLNVPLRGTIVTRPLM and PLMNLSLVL, respectively. Bound antibodies were detected with horseradish peroxidase-conjugated goat anti-rabbit (dilution 1:30,000; Cappel) and the Immobilon Western chemiluminescent substrate (Millipore).
DBT-mACE2 cells were infected with SARS-CoV, SARS-CoV-nsp1-ΔC and -ΔD at a moi of 0.125. Total RNAs from DBT-mACE2 infected cells were extracted at 48 hpi using the Qiagen RNeasy kit according to the manufacturer’s instructions. Quantitative reverse transcription-polymerase chain reaction (qRT-PCR) reactions were performed at 37°C for 2 h using the High Capacity cDNA transcription kit (Applied Biosystems) and 100 ng of total RNA and random hexamer oligonucleotides. Cellular gene expression was analyzed using TaqMan gene expression assays (Applied Biosystems) specific for Mus musculus genes (S3 Table). Data were acquired with an ABI PRISM 7000 sequence detection system (Applied Biosystems) and analyzed with ABI PRISM 7000 SDS version 1.0 software. Gene expression in mock-infected cells and SARS-CoV, SARS-CoV-nsp1-ΔC and -ΔD-infected cells was compared. Quantification was achieved using the 2-ΔΔCt method, which analyzes relative changes in gene expression in qPCR experiments (Livak and Scmittgen, 2001). The results of three independent experiments were analyzed. All experiments and data analysis were MIQE compliant [82].
Lung sections from infected animals were collected at 2 dpi and homogenized using gentleMACS Dissociator (Miltenyibiotec). Then, total RNA was extracted using the RNeasy purification kit (Qiagen). Reactions were performed at 37°C for 2 h using a High Capacity cDNA transcription kit (Applied Biosystems) with 100 ng of total RNA and random hexamer oligonucleotides. Cellular gene expression was analyzed using TaqMan gene expression assays (Applied Biosystems) specific for mouse genes (S4 Table). Data representing the average of three independent experiments were acquired and analyzed as previously described [57]. All experiments and data analysis were MIQE compliant [82].
The computer modeling of 8a protein structure was performed with the raptorX server http://raptorx.uchicago.edu [83]. The predicted structures were visualized using Pymol (http://www.pymol.org/).
Student´s t test was used to analyze differences in mean values between groups. All results are expressed as means ± standard errors of the means. P values of <0.05 were considered statistically significant.
The UniProt (http://www.uniprot.org/) accession numbers for genes and proteins discussed in this paper are: SARS-CoV E protein, P59637; SARS-CoV 8a protein, Q19QW2; SARS 3a protein, P59632; SARS M protein, P59596; mouse IFN-β, P01575; mouse IRF1, P15314; mouse DDX58, Q6Q899; mouse STAT1, P42225; human p38 MAPK, Q16539; human ACE2, Q9BYF1; mouse ACE2, Q8R0I0; mouse CXCL10, P17515; mouse CCL2, P10148; mouse IL6, P08505; mouse 18S, O35130; human actin, P60709.
|
10.1371/journal.pcbi.1000930 | Evolution and Optimality of Similar Neural Mechanisms for Perception and Action during Search | A prevailing theory proposes that the brain's two visual pathways, the ventral and dorsal, lead to differing visual processing and world representations for conscious perception than those for action. Others have claimed that perception and action share much of their visual processing. But which of these two neural architectures is favored by evolution? Successful visual search is life-critical and here we investigate the evolution and optimality of neural mechanisms mediating perception and eye movement actions for visual search in natural images. We implement an approximation to the ideal Bayesian searcher with two separate processing streams, one controlling the eye movements and the other stream determining the perceptual search decisions. We virtually evolved the neural mechanisms of the searchers' two separate pathways built from linear combinations of primary visual cortex receptive fields (V1) by making the simulated individuals' probability of survival depend on the perceptual accuracy finding targets in cluttered backgrounds. We find that for a variety of targets, backgrounds, and dependence of target detectability on retinal eccentricity, the mechanisms of the searchers' two processing streams converge to similar representations showing that mismatches in the mechanisms for perception and eye movements lead to suboptimal search. Three exceptions which resulted in partial or no convergence were a case of an organism for which the targets are equally detectable across the retina, an organism with sufficient time to foveate all possible target locations, and a strict two-pathway model with no interconnections and differential pre-filtering based on parvocellular and magnocellular lateral geniculate cell properties. Thus, similar neural mechanisms for perception and eye movement actions during search are optimal and should be expected from the effects of natural selection on an organism with limited time to search for food that is not equi-detectable across its retina and interconnected perception and action neural pathways.
| The brain has two processing pathways of visual information, the ventral and dorsal streams. A prevailing theory proposes that this division leads to different world representations for conscious perception than those for actions such as grasping or eye movements. Perceptual tasks such as searching for our car keys in a living room requires the brain to coordinate eye movement actions to point the high resolution center of the eye, the fovea, to regions of interest in the scene to extract information used for a subsequent decision, such as identifying or localizing the keys. Does having different neural representations of the world for eye movement actions and perception have any costs for performance during visual search? We use computer vision algorithms that simulate components of the human visual system with the two separate processing streams and search for simple targets added to thousands of natural images. We simulate the process of evolution to show that the neural mechanisms of the perception and action processing streams co-evolve similar representations of the target suggesting that discrepancies in the neural representations of the world for perception and eye movements lead to lower visual search performance and are not favored by evolution.
| Neurophysiology studies of the macaque monkey [1]–[3] support the existence of two functionally distinct neural pathways in the brain mediating the processing of visual information. The behavior of patients with brain damage has led to the proposal that perception is mediated by the ventral stream projecting from the primary visual cortex to the inferior temporal cortex, and that action is mediated by the dorsal stream projecting from the primary visual cortex to the posterior parietal cortex [4]–[6] (Figure 1a). Although there has been debate about whether this separation into ventral/dorsal streams implies that the brain contains two distinct neural representations of the visual world [7]–[12], there has been no formal theoretical analysis about the functional consequences of the two different neural architectures on an animal's survival. Visual search requires animals to move their eyes to point the high-resolution region of the eye, the fovea, to potentially interesting regions of the scene to sub-serve perceptual decisions such as localizing food or a predator. What is the impact of having similar versus different neural mechanisms guiding eye movements and mediating perceptual decisions on visual search performance for an organism with a foveated visual system? We consider two leading computational models of multiple-fixation human visual search, the Bayesian ideal searcher (IS) [13]–[15] and the ideal saccadic targeting model (maximum a posteriori probability, MAP [16], [17]) for a search task of a target in one of eight locations equidistant from initial fixation (Figure 1b). The ideal searcher uses knowledge of how the detectability of a target varies with retinal eccentricity (visibility map) and statistics of the scenes to move the fovea to spatial locations which maximize the accuracy of the perceptual decision at the end of search [13] (Figure 1b). The saccadic targeting model (MAP) makes eye movements to the most probable target location [6], [17] which is optimal if the goal was to saccade to the target rather than collect information to optimize a subsequent perceptual decision [1] (Figure 1b). Depending on the spatial layout of the possible target locations and the visibility map, the IS and MAP strategies lead to similar (Figure 1c) or diverging eye-fixations (Figure 1d–e). For example for a steeply varying visibility map (Figure 1c) both models make eye movements to the possible target locations while for a broader visibility map (Figure 1d–e) the ideal searcher tends to make eye movements in between the possible target locations attempting to obtain simultaneous close-to-fovea processing for more than one location. Covert attention allows both models to select possible target locations and ignore locations that are unlikely to contain the target when deciding on saccade endpoints and making perceptual search decisions [18], [19]. Perceptual target localization decisions for both models are based on visual information collected in parallel over the whole retina, temporally integrated across saccades, and based on the location with highest sensory evidence for the presence of the target. Critically, we implemented the models to have two processing pathways, one determining where to move the fovea and the other stream processing visual information to reach a final perceptual decision about the target location. Rather than having a single linear mechanism or perceptual template (Figure 1b), each pathway in the model had its own neural mechanism which is compared to the incoming visual data at each possible target location. Likelihood ratios [20] of the observed responses for each of the mechanisms under the hypothesis that the target is present or absent at that location are used to make decisions about where to move the eyes and perceptual decisions (see Materials and Methods).
We used a genetic algorithm as a method to find near-optimal solutions for perception and action mechanisms but also to simulate the effects of the evolutionary process of natural selection on the neural mechanisms driving saccadic eye movements and perceptual decisions during search. The computational complexity of the ideal Bayesian searcher makes it difficult to virtually evolve the model (see note 1 in Text S2) and thus we used a recently proposed approximation to the ideal searcher that is computationally faster (Entropy Limit Minimization, ELM [15], [21]). The ELM model chooses the fixation location that minimizes the uncertainty of posterior probabilities over the potential target locations. The decision rule can be simplified to choose the fixation location with the maximum sum of likelihood ratios across potential target locations, each weighted by its squared detectability given the fixation location [15]. The ELM model can be shown to approximate the fixation patterns of the ideal searcher [15] and capture the main characteristics of the fixation patterns of the IS for our task and visibility maps (Figure 1c–e; ELM) (see note 2 in Text S2). The process of virtual evolution started with the creation of one thousand simulated individuals with separate linear mechanisms for perception (ventral) and eye movement programming (dorsal; Figure 2a). Each pathway's template for each individual was created from independent random combinations of the receptive fields of twenty four V1 simple cells. Each simulated individual was allowed two eye movements (see note 3 in Text S2) before making a final perceptual search decision about the location of the target. Performance finding the target in one of eight locations for five thousand test-images (one thousand for natural images) was evaluated and the probability of survival of an individual was proportional to its performance accuracy. A new generation was then created from the surviving individuals through the process of reproduction, mutation and cross-over (Figure 2a). The process was repeated for up to 500 generations.
We first evolved the ideal searcher approximation (ELM model) for different shape luminance targets (isotropic Gaussian, vertical elongated Gaussian and cross pattern consisting of a positive and negative polarity elongated Gaussian) embedded in 1/f noise and a steep visibility map (Figure 1c). Irrespective of the target shape, virtual evolution led to converging perception (ventral) and saccade (dorsal) mechanisms that are similar to the target (Figure 2b; see Video S1, Video S2, and Video S3 for virtual evolution). To further investigate the generality of the result we evolved the ELM model to search a circular Gaussian target added to backgrounds with different statistical properties: white noise, 1/f noise and importantly, a calibrated set of natural image backgrounds [22]. Figure 3 (2nd row) presents the distribution of perceptual decision accuracies across individuals in a generation and shows that perceptual performances of simulated individuals in the population improve with generations and then converge to an asymptote. We characterized the similarity between the perception and saccade mechanisms by computing the correlations between the 2 dimensional linear mechanisms for each individual in each generation. Figure 3 (3rd row) shows that the distribution of correlations across individuals in the population evolves to unity irrespective of the background type. To visualize in detail the shape of the evolved templates, we analyzed the radial profile of the templates of the highest performing simulated individuals in the last generation (Figure 3; 4th row). For all three backgrounds the saccade and perception templates converge to similar shapes (perception and saccade 2-D template correlations for the best performing templates in the last generation: 0.990±0.006, 0.986±0.013, 0.982±0.013). In addition, the linear mechanisms for the 1/f noise and natural scenes are narrower than those for the white noise and show an inhibitory surround (Figure 3).
These previous results were based on a visibility map that steeply declines with eccentricity and rely on the assumption that humans are near-ideal searchers. We, thus, evolved the mechanisms for the case of a broader visibility map that is similar to that measured for human observers in 1/f noise [15] (Figure 4a) and showed that the convergence of neural mechanisms generalizes to different visibility maps (Figure 4a) and also to a model in which eye movement planning is assumed to follow a saccadic targeting strategy (MAP) rather than approximating an ideal strategy (Figure 4a). Furthermore, Figure 4b shows that there is nothing particular about the symmetry of the eight location configuration search task since similar convergent evolution is observed for an asymmetric four location task (Figure 1e).
We also evaluated whether our results would change if the model included the increasing size of V1 receptive fields and lower frequency tuning with retinal eccentricity (see note 4 in Text S2). Figure 5a (right graph) shows the center frequency and bandwidth (standard deviation) of the oriented Gabor receptive fields as a function of retinal eccentricity. The computational time demands of this simulation restricted us to evaluate this model for a fixed set of receptive field weights across eccentricities (see note 5 in Text S2) and limited set of scenarios: 1/f noise, steep visibility map and two targets: a low frequency Gaussian (Figure 5b; left) and a Difference of Gaussians (DoG) with a center frequency of 8 c/deg (Figure 5b; right). Due to the fixed set of weights across eccentricity, in this model the spatial profile of the linear combination of receptive fields scales up with eccentricity. Thus, for each retinal eccentricity category there was a pair of evolved template profiles. Figure 5c shows that convergent evolution still results when receptive field size increases with eccentricity and irrespective of the spatial frequency of the target. Figure 5d presents the similar radial profiles of the of evolved perception and saccade mechanisms for the fovea and a sample peripheral retinal location (perception and saccade 2-D template correlations for the best performing templates in the last generation averaged across retinal eccentricities were: Gaussian target: 0.963±0.008; DoG target: 0.961±0.004).
Do all scenarios lead to converging evolution of the perception (ventral) and action (dorsal) pathways? No, if we take a case in which the sought target is equally detectable across the retina (flat visibility map), the results show the correlations between the perceptual and saccade templates do not converge to unity (Figure 6a). A second example is a case in which the organism makes a decision after eight eye movements rather than two eye movements (Figure 6b). Because the organism gets to fixate on all eight target locations, there is little added benefit of an efficient saccadic system and the co-evolution is much slower (Figure 6b). A third scenario of partial convergence results if we adopt a strong model of two visual processing streams which spatially pre-filter the visual input based on the properties of the cells in the parvocellular and magnocellular lateral geniculate nucleus (LGN) ([23]; see Figure 6d) and assume no further interaction across pathways. The differential spatial frequency filtering of the two pathways can introduce constraints in the frequency content of the evolved mechanisms preventing a full convergence of the templates (Figure 6e; perception and saccade 2-D template correlations for the best performing templates in the last generation for: 1/f noise: 0.603±0.082). A similar simulation with the same target but white noise instead of 1/f noise also resulted in partial convergence (perception/saccade 2-D template correlation of 0.856±0.046).
We used an approximation to an Ideal Bayesian Searcher (Entropy Limit Minimization model; ELM) to virtually evolve separate linear mechanisms for eye movements and perceptual decisions during visual search for a variety of targets embedded in various synthetic and natural image backgrounds. Evolved templates contain similarities to the target but for the 1/f and natural images they are narrower than the target and contain a subtle inhibitory surrounding not present in the signals but often present in monkey neuronal receptive fields and human behavioral receptive fields [9], [19] (see blue outline in Figure 2b). A previous study has shown that such inhibitory surrounds serve to suppress high amplitude noise in the low frequencies and optimize the detection of spatially compact signals in natural images [24]. The current result extends previous results [24] to show the optimality of inhibitory surrounds during visual search in natural images for an organism with a foveated visual system and saccadic eye movements.
Central to this paper, the mechanisms for perception and saccades evolved to similar representations. This result is robust across different types of backgrounds, signals, visibility maps, and spatial distributions of possible target locations. Due to computational constraints we did not investigate the more general case of allowing the target to appear at any location within the image but there is no particular reason to suggest that our result would differ for this latter general case. In addition, similar convergence between mechanisms was found for what arguably are the most common contender algorithms to model how humans plan eye movements during search: an approximation to the ideal searcher, ELM and a saccadic targeting model; MAP model; [13]. For simplicity our original models did not include receptive fields that increased with retinal eccentricity but an implementation of such a model led to similar convergent evolution for a low and a higher spatial frequency target.
The scenarios for which we did not find full convergent evolution of the linear mechanisms were for cases for which the target was either equi-detectable across the retina or the organism had enough time to fixate all of the possible target locations. Note, however, that for both cases, performance of the evolved individuals does improve with increasing generations (Figure 6a–b) through the evolution of the perceptual template to a target-like structure. Yet, there is no performance advantage for evolving a neural mechanism for saccades that encodes target information because, for these cases different eye movement patterns have little or no impact on perceptual performance. A third scenario which resulted in partial convergence was a two stream model with pathway-specific pre-filtering of the visual input. A strong assumption that there are no interconnections between the two pathways would result in processing constraints based on the early stages of visual processing of both pathways. Inclusion of pre-filtering properties of the parvocellular and magnocellular LGN cells restricted the full convergence of the evolved mechanisms. These finding suggest that if we adopt a strict separation of pathways and take into account properties of LGN cells we should not always expect similar mechanisms driving perception and saccadic decisions during search. The specific circumstances for which we will not find convergent evolution and the degree of similarity between evolved templates will depend on the spatial frequency of the target and background statistical properties (see results for 1/f noise vs. white noise). Yet, is the strict separation of pathways and constraints to the filtering properties of parvocellular (perception) and magnocellular (action) LGN cells tenable for the case of eye movements and perceptual decisions during search? A recent psychophysical study [9] used the same Gaussian target as in the simulations and reverse correlation to show that estimated underlying templates mediating human saccadic actions and perceptual search decisions are similar. Thus, these psychophysical findings would suggest that the strong assumption of no interconnections across pathways and constraints by the early LGN processing might not hold at least for the case of perception and eye movements during visual search.
Together, our present results suggest a theory of why evolution would favor similar neural mechanisms for perception and action during search [9] and provide an explanation for the recent study finding similar estimated underlying templates mediating human saccadic decisions and perceptual decisions. Our findings and theory do not necessarily imply either that one pathway mediates both perception and action nor are they incompatible with the existence of separate magnocellular and parvocellular pathways. Instead, our theory would be consistent with the idea that pathways for perception and oculomotor largely overlap, leading to significant sharing of visual information across pathways [8], [12], [25], [26]. For the case of saccadic eye movements, visual cortical pathways through the frontal eye fields [27] and the lateral intra-parietal cortex [28] play critical roles, as well as brainstem and cortical pathways through the superior colliculus [29]. In addition, studies have related areas in the ventral stream (V4) to target selection of saccades [30], [31]. In addition, the results do not prohibit small differences in visual processing for perception and saccadic action but provide functional constraints on how much discrepancy can exist between neural mechanisms without jeopardizing the survival of the organism.
In the larger context, the similar neural mechanisms for perception and saccade actions should be understood as another effective strategy implemented in the brain, in addition to guidance by target properties [13], [14], [32], [33], optimal saccade planning [15], contextual cues [34], [35] and miniature eye movements [36] to ensure successful visual search. Finally, the approach of the present study demonstrates how the rising field of natural systems analysis [37], [38] can be used in conjunction with virtual evolution and physiological components of the visual system to evaluate whether properties of the human brain might reflect evolved strategies to optimize perceptual decisions and actions that are critical to survival.
We assumed a viewing distance of 50cm for the models. Search targets for simulations were: a) A Gaussian target with 0.5539 square root contrast energy (SCE) and a standard deviation of 0.1376 degrees (Figure 1c; 2b left column; 3); b) An elongated Gaussian with 0.9594 SCE, standard deviations of 0.4128 deg. in the vertical direction and 0.1376 degrees in the horizontal direction (Figure 2b center column, Figure 4); c) The difference of a vertically oriented and a horizontally oriented elongated Gaussians with 0.8581 SCE (Figure 2b, right column). The white noise root mean square contrast (rms) was 0.0781. The same rms was used for white noise filtered with the 1/f function (1/f noise). Possible target locations were equidistant 7 degrees from the center fixation cross. Independent external and internal noise samples were refreshed with each saccade for the white and 1/f noise. For the natural images the external backgrounds were fixed but the internal noise refreshed across saccades.
Here, we briefly describe the models implementations (see Text S1 for detailed mathematical development and details). The initial stage of all three models investigated (ideal searcher, IS; entropy limit minimization, ELM; and saccadic targeting, MAP) is the dot product of a perceptual and saccade template (w) with the image data (g) at all possible target locations, where r is the resulting scalar response and w and g are expressed as 1-D vectors. The templates for the perceptual decisions and saccade planning were independent and random linear combinations of 24 Gabor functions that spanned the targets: spatial frequencies, 0.5, 1, 2, 4 cycles/degree for 6 different orientations, 30 degrees apart, and with octave bandwidths. A subset of simulations (Figure 6) also modeled pre-processing of the image by separate LGN cells corresponding to the magnocellular (dorsal) and parvocellular (ventral) cells. The filtering was done using DoG functions with different center frequencies (see Text S1 for mathematical details) prior to the processing by the Gabor functions.
Use of a larger number of Gabor functions did not significantly change the evolved templates for the targets considered but required prohibitively longer computational times due to the dimensionality explosion. For the template derived for the case of the isotropic Gaussian target we used an additional constraint of equal weighting for all orientations of the Gabor functions for a given spatial frequency. Most of the simulations used the fixed 24 Gabor functions irrespective of retinal eccentricity. A subset of simulations (see Figure 5) used sets of 24 Gabor functions that increased linearly in size and also decreased in the central frequency tuning with retinal eccentricity (see details in effects of retinal eccentricity section). Template responses were integrated across saccades. Calculation of likelihood ratios use Gaussian probability density functions which depend on the image parameters for the white and 1/f Gaussian noisy images. For the natural images, the likelihood calculation required estimating the probability density function from a training set of 3000 images and fitting the probability density functions with Laplacian distributions convolved with a Gaussian distribution representing the internal noise (see Text S1).
Two methods were used to model the detrimental effect of retinal eccentricity on the detectability of the target. The first method which is similar to Najemnik and Geisler [13] was implemented by adding internal noise to the scalar template response: , where the additive internal noise scalar value is sampled from a Gaussian distribution which standard deviation () is dependent on the distance ( i.e. retinal eccentricity) between the tth fixation and the template response location i out of m possible target locations. Also the internal noise was proportional to the template's response standard deviation resulting from the external image variability. The visibility maps referred to as steep and broad (see also Figure S1) were obtained with internal noise standard deviations given by:(1a)(1b)where σo is the standard deviation of the template response due to external noise, e is the eccentricity in degrees, and the subscripts k refer to the fixation location, and i to the possible target location. For all models, independent samples of internal noise were used for each saccade and pathways.
The second method to model the effects of retinal eccentricity included internal noise (see above) and also varied the sets of 24 Gabor functions with retinal eccentricity.
The size of Gabor functions increased with the retinal eccentricity (e) so that the standard deviation of the spatial Gaussian envelope is given by:(2)where is the bandwidth and is the center frequency of Gabor function in the fovea. Thus, the standard deviation in the frequency domain of each Gabor function (Figure 5a; right graph) decreases as:(3)
The center frequency tuning of the Gabor functions (s) linearly decreased with retinal eccentricity: .
The saccadic targeting or maximum a posteriori probability model (MAP) chooses the location of the next fixation with the maximum product of likelihoods ratios () across previous and present fixation (t = 1,…, T):(4)
For the case of white noise and 1/f Gaussian noise the expression can be simplified to the sum of log-likelihood ratios:(5)where Δμ is the difference in mean response of the template to the signal plus background and background only and all other symbols are defined above.
The ideal searcher selects as the next fixation the location that will maximize the probability of finding the target after the eye movement is made:(6)where is the proportion correct (PC) given that the target location is i, and the next fixation is . The term is the prior that the ith location contains the target given the sensory evidence collected up to the present fixation: and m is the number of possible target locations. For white noise and 1/f noise Gaussian noise, becomes:(7)where is the probability density function of the Gaussian function in Equation (9a), the cumulative density function of the Gaussian function in Equation (9b), and , are the log-likelihood ratios which are known scalar values based on acquired visual information,(8a)(8b)while and are random variables describing log-likelihoods after the next fixation and described by normal probability density functions:(9a)(9b)where () is the detectability at target location i (j), given fixation at location . The present formulation is identical to that of Najemink and Gielser [13] but uses likelihood ratios rather than product of posteriors.
The entropy limit minimization model chooses as the next fixation the locations that minimize the expected entropy and can be approximated by maximizing the expected information gain. This can be shown to be approximated by calculating for each potential fixation location, , a sum of the posterior probability for each location weighted by the squared detectability given the fixation location [15]:(10)where is the Shannon entropy of , and is the information gain.
For all models, the final perceptual decision about the target location was obtained by combining the likelihood ratios for each possible target locations across all fixations and choosing the location with the highest product of likelihood ratios:(11)where the likelihoods of the responses given the background only and the target are given and which are the probability density functions (pdf) assumed to be Gaussian (white noise and 1/f noise) or empirically estimated from samples (see next section) for the natural images.
The distribution of template responses for the natural image dataset [22] were estimated from 24,000 image patches extracted from the eight possible target locations for 3000 natural images. We fit the distribution of these responses for each template of each simulated individual with a Laplacian distribution:(12)where is the mean parameter and is a scale parameter. To take into account the effect of additive Gaussian internal noise on the probability density function of the template responses we convolved the Laplacian distribution with the Gaussian distributions:(13)where and are Gaussian and Laplace probability density functions respectively (see Figure S2).
We used the Genetic Algorithm Optimization Toolbox (GAOT) [39]. Arithmetic crossover parameter was set to operate 50 times per generation, and uniform mutation to operate 50 times per generation. The selection process used a real-valued roulette wheel selection [38]. A generation consisted of 1,000 individual parameter settings. All individuals were randomly initialized, and allowed to evolve over 500 generations (see Text S1 for additional details). Reported results for each scenario/model were averages across ten simulated evolution runs.
|
10.1371/journal.pcbi.1002860 | Adding Protein Context to the Human Protein-Protein Interaction Network to Reveal Meaningful Interactions | Interactions of proteins regulate signaling, catalysis, gene expression and many other cellular functions. Therefore, characterizing the entire human interactome is a key effort in current proteomics research. This challenge is complicated by the dynamic nature of protein-protein interactions (PPIs), which are conditional on the cellular context: both interacting proteins must be expressed in the same cell and localized in the same organelle to meet. Additionally, interactions underlie a delicate control of signaling pathways, e.g. by post-translational modifications of the protein partners - hence, many diseases are caused by the perturbation of these mechanisms. Despite the high degree of cell-state specificity of PPIs, many interactions are measured under artificial conditions (e.g. yeast cells are transfected with human genes in yeast two-hybrid assays) or even if detected in a physiological context, this information is missing from the common PPI databases. To overcome these problems, we developed a method that assigns context information to PPIs inferred from various attributes of the interacting proteins: gene expression, functional and disease annotations, and inferred pathways. We demonstrate that context consistency correlates with the experimental reliability of PPIs, which allows us to generate high-confidence tissue- and function-specific subnetworks. We illustrate how these context-filtered networks are enriched in bona fide pathways and disease proteins to prove the ability of context-filters to highlight meaningful interactions with respect to various biological questions. We use this approach to study the lung-specific pathways used by the influenza virus, pointing to IRAK1, BHLHE40 and TOLLIP as potential regulators of influenza virus pathogenicity, and to study the signalling pathways that play a role in Alzheimer's disease, identifying a pathway involving the altered phosphorylation of the Tau protein. Finally, we provide the annotated human PPI network via a web frontend that allows the construction of context-specific networks in several ways.
| Protein-protein-interactions (PPIs) participate in virtually all biological processes. However, the PPI map is not static but the pairs of proteins that interact depends on the type of cell, the subcellular localization and modifications of the participating proteins, among many other factors. Therefore, it is important to understand the specific conditions under which a PPI happens. Unfortunately, experimental methods often do not provide this information or, even worse, measure PPIs under artificial conditions not found in biological systems. We developed a method to infer this missing information from properties of the interacting proteins, such as in which cell types the proteins are found, which functions they fulfill and whether they are known to play a role in disease. We show that PPIs for which we can infer conditions under which they happen have a higher experimental reliability. Also, our inference agrees well with known pathways and disease proteins. Since diseases usually affect specific cell types, we study PPI networks of influenza proteins in lung tissues and of Alzheimer's disease proteins in neural tissues. In both cases, we can highlight interesting interactions potentially playing a role in disease progression.
| The advent of high-throughput techniques to measure and perturb molecular species in a systematic way has enabled researchers to assess the different layers of cellular metabolism under different experimental conditions. Protein-protein interaction (PPI) networks created by a variety of methods including yeast-two-hybrid (Y2H), mass-spectrometry (MS) and computational predictions [1], [2] are valuable research resources, and have been used heavily in the last decade. However, a major drawback of these data is that the artificial expression systems used to reconstruct PPI networks do not take into account two of the many factors that are essential to understand the biology of the cell: first, the time-point at which the proteins are expressed (e.g., cell-cycle or developmental stage) and second, the tissue or intracellular compartment where the proteins are expressed or located (different organs and tissues have very specific protein compositions). Therefore, two proteins may be reported as interaction partners, although they are expressed in different tissues or at different time-points. While high-throughput studies acknowledge these caveats, PPI databases collect these data without mechanisms explicitly directed to discern the biological plausibility of a reported interaction. Therefore, the selection of proteins expressed in a specific cell type or compartment would allow the generation of subnetworks that more realistically represent biological processes in the respective cell types or cellular compartment.
Several attempts have been made to investigate the tissue-specific binding behavior of single proteins and the spatio-temporal dynamics of PPI networks [3], [4], [5], [6], [7], [8]. In a recent study evaluating the characteristics of publicly available PPI databases, we demonstrated that the use of subnetworks (which include only interactions of proteins expressed in the same tissue) identifies potential mechanisms or pathways that would remain obscured if the complete PPI database was used [9].
In addition, many proteins have multiple functions, carried out in cooperation with distinct sets of interacting partners. Networks of interacting proteins with coherent function have been termed context networks [10]. Here, we adopt this notion of context and extend it to PPIs or networks of proteins being expressed in the same tissue or cooperatively transmitting signal flow.
There is a lack of studies testing systematically the potential of adding context information to PPI networks in recovering meaningful PPI subsets and, although there are a few approaches that allow to add expression or functional information to PPI data [11], [12], [13], convenient methods for the creation of such context-specific subnetworks are generally missing.
Here, we introduce an approach to add context to PPI networks using annotations and relations between the interacting partners and demonstrate that context-specific PPI networks are enriched in high-confidence interactions. We use this approach to investigate how the proteins of the human influenza virus interfere with the immune response of the host cell in a tissue-specific manner, finding novel potential regulators of influenza virus pathogenicity, and to study the brain-specific signaling pathways that play a role in Alzheimer's disease, identifying a pathway involving the altered phosphorylation of the Tau protein. Thereby, we illustrate how the addition of context to PPI networks can guide researchers in the discovery of meaningful interactions and pathways, which would otherwise be obscured by the vast amount of irrelevant (for a specific question) and partly erroneous amount of PPI data.
Our approach to add context-specific information to human PPI data was implemented in the HIPPIE database [14]. HIPPIE is an integrated PPI database that currently contains more than 101,000 interactions of ∼13,500 human proteins. HIPPIE is regularly updated by incorporating interaction data from major expert-curated experimental PPI databases (such as BioGRID [15], HPRD [16], IntAct [17] and MINT [18]) in an automated manner using the web service PSICQUIC [19]. All interactions have an associated confidence score based on the sum of cumulative supporting experimental evidence.
Individual proteins were associated with tissues, subcellular locations and biological processes in the following manner. First, proteins were associated with tissues (based on their gene expression profiles retrieved from BioGPS [20] and using the method defined in [9]) or defined as housekeeping (using a list from [21]). Next, associations with biological processes and subcellular locations were determined according to the EBI Gene Ontology (GO) annotation (release from October 28, 2011; reduced to GO slim terms) [22], and to MeSH terms belonging to “Diseases” (class C) or “Tissues” (class A10) that annotate the biomedical references associated to them in MEDLINE (release 2012; gene2pubmed at NCBI ftp site).
We associated an interaction with a tissue when both interactors are expressed in the same tissue (e.g. “lung”). Given a term of a functional ontology, we associated an interaction with this function when both interactors are annotated with either the given functional term or with children of it in the hierarchy of the ontology. For example, the GO term “transport” would be associated with an interaction between a protein annotated as involved in “vacuolar transport” and another protein annotated as involved in “nucleocytoplasmic transport”. Functional terms considered were either GO terms or MeSH terms. We excluded the rather unspecific top-level terms ‘biological process’, ‘cellular component’ and ‘cell’. Additionally, we ignored categories that are associated to less than 20 interactions.
Our approach includes a method to infer directed PPIs. This inference of interaction (edge) directionality needs sets of proteins predefined as sinks and sources. As default sources and sinks, we connected all proteins annotated with the GO terms ‘receptor’ and ‘sequence-specific DNA binding transcription factor activity’, respectively, in the UniprotKB [23]. This is done assuming that signal pathways follow the transmission of information through interacting proteins starting in cell surface receptors that collect external cues and ending in transcription factors as final effectors on gene regulation, following [24]. To infer edge directionality, all pairwise shortest paths between proteins of the source and the sink sets present in the generated output network are calculated. We do not consider edge weights and, hence we are able to determine each shortest path in linear time via a breadth-first search. An edge of the network is considered to be directed if at least one shortest path goes through that edge. The direction of the path (from source to sink) determines the direction of the edge. Edges with conflicting orientations of passing paths are not assigned directionality.
For the evaluation of the influenza virus host factor network generation we performed pathway enrichment analysis with ConsensusPathDB (run on August 30, 2012; [25]). We used a cut-off of 0.05 on the q-value, which is the false discovery rate (FDR) adjusted equivalent to the p-value. The background control for the tests was the complete list of proteins annotated as expressed in the given tissues (and with PPI information in HIPPIE).
We retrieved the preprocessed microarray data described in [26] measuring gene expression changes over multiple time points in a lung adenocarcinoma cell line (Calu-3) infected with influenza A/Netherlands/602/2009 (H1N1). To select steadily up-regulated genes we filtered for probes differentially expressed at the last three time-points in the time series (30, 36 and 48 h) with a q-value less than 0.01 and a log2 fold change greater than 1.
To generate a list of PPIs related to Alzheimer's and protein phosphorylation, first, we used the webserver MedlineRanker [27] to retrieve a list of ranked PubMed abstracts (corresponding to manuscripts published within the last 5 years) according to their relevance to the search term “Alzheimer phosphorylation”, which relates loosely to the question of interest. Next, we input the top 50 abstracts from MedlineRanker into the webserver PESCADOR [28], which extracts a network of potential PPIs based on a set of PubMed abstracts. In our example, PESCADOR outputs 10 interaction pairs (type 2; co-occurrence of genes or proteins within a sentence containing a biointeraction term), of which only 4 pairs existed in HIPPIE as scored interactions (PSEN1:PSEN2, GSK3B:MAPT, APP:BACE1, PPP2R4:SET). These confirmed PPIs were then used as input for further analysis.
We inferred context information for all interactions in the human PPI database HIPPIE [14]. This database collects human PPIs for which there is experimental evidence. The amount and quality of the experimental evidence supporting each PPI is evaluated with a confidence score that ranges from 0 to 1. In a first step, we associated all 13,477 proteins in HIPPIE with the following attributes: tissue-expression, GO biological process and cellular compartment, and inferred annotations for the MeSH categories disease and tissue. We then inferred context associations to the PPIs according to the annotations of the interacting proteins and taking into account the hierarchical structure of GO and MeSH terms (see Materials and Methods for details).
By assuming that a large fraction of signaling events transmits information from proteins sensing environmental changes to effector proteins altering the cellular state, we computed shortest paths from membrane-bound receptors to transcription factors (TF) through the network. From the predicted information flow we assigned edge directionality to interactions on these paths (see Materials and Methods for details).
Overall, we were able to associate context to more than 97,000 of the 101,131 interactions of the current version of HIPPIE. Interactions for which we inferred or collected annotations had significantly better experimental evidence (Figure 1A). This suggests that annotated interactions might have higher biological significance than non-annotated ones.
As expected, we observed that more specific context categories were associated to interactions with higher experimental reliability: while the confidence scores of interactions with rather unspecific and ubiquitous terms resemble the overall confidence score distribution, interactions with highly specific terms usually have a higher than average confidence score (Figure 1B-C). For example, the 43,372 interactions associated with the GO category ‘cytoplasm’ (of depth 1 in the GO hierarchy) have an average confidence score of 0.675 as compared the average of 0.670 over all interactions. On the other hand, the 159 interactions associated with the (depth 3) GO category ‘ribonucleoprotein complex assembly’ have an average confidence score of 0.754. We observed a similar tendency for more specific MeSH terms to have a higher experimental reliability.
To demonstrate that our automated context association approach allows identification of relevant interactions, we tested if networks of interactions of our inferred MESH-based disease-annotation are enriched in well-known disease proteins. Therefore, we repeatedly generated disease-context networks around a set of canonical disease proteins. As a canonical disease protein specification, we retrieved the manually curated UniProt Knowledgebase disease protein annotation. For each of the canonical disease proteins, we generated two types of networks: (a) disease networks consisting only of interaction partners of the disease proteins that we had associated with the equivalent MeSH disease term and (b) unfiltered PPI network consisting of all interaction partners of the disease protein from HIPPIE. We did this for all disease proteins where the disease was associated with at least two disease proteins in UniProt and at least two interactions that we had associated with this disease. To quantify the enrichment of disease proteins in these networks we repeatedly calculated the F1 score, the harmonic mean of precision and recall (F1 = 2*precision*recall/(precision+recall)). A one-sided Mann-Whitney-test comparing the distribution of F1 scores between the disease networks and the non-filtered networks indicated that the F1 scores for the disease networks were significantly larger (p<0.05) proving an enrichment of disease proteins in the disease filtered networks (without losing sensitivity by removing disease proteins in the filtering step). The mean precision on the filtered networks was 0.47 and on the unfiltered networks 0.21. The mean recall for the filtered networks was 0.14 and for the unfiltered networks 0.15. This illustrates that in exchange for a small decrease in recall the precision can be more than doubled by applying the MeSH disease filter.
We then investigated the potential of edge directionality inference based on the shortest paths between membrane-bound receptors and TFs through the PPI network to recover known pathways. We retrieved pathway annotations (extracted from WikiPathways download March 29, 2012) and computed the shortest paths through HIPPIE between all pairs of receptors and TFs within the same pathway (excluding only pairs that directly interact or could not be connected by any path). We counted the number of proteins of each pathway found on the shortest paths. We found for 3163 of the 5063 pairs that this approach correctly identified proteins of the selected pathway. The mean precision (the fraction of proteins on the paths that indeed belonged to the correct pathway) over all combinations of receptors with transcription factors was 0.20. The mean recall (the fraction of the pathway that was recovered by considering the paths between one receptor and one transcription factor) was 0.02.
To assess if the agreement between shortest paths and canonical pathways was larger than expected by chance, we generated a background distribution by computing repeatedly the shortest paths between a receptor and a TF from different pathways and computed the overlap between the proteins on the shortest paths to either the TF- or the receptor-containing pathway. We found that the overlap distribution was significantly higher when the receptor and the TF were members of the same pathway (p<0.001; Mann-Whitney-test) proving the potential of shortest paths to recover the signal flow between TFs and receptors when functionally related pairs of receptors and transcription factors are chosen.
We wondered if we could further increase the overlap between the shortest paths and the canonical pathways by filtering the networks for tissue expression. To associate pathways with tissues, we determined for each pathway which tissues were enriched among the genes of the pathway (Supplementary Table S1 lists pathway that are associated to more than 2-fold enriched tissues). Inspection of the tissues enriched among proteins forming a pathway revealed that in many cases they indeed reflect plausible locations for pathway activity. For example, immune response pathways were enriched among blood cells and pathways associated with neurodegenerative diseases and addiction in brain-related tissues.
We repeated the computation of shortest paths linking receptors to transcription factors in tissue-specific networks for combinations of pathways and tissues listed in Supplementary Table S1 and for all pairs of receptors and transcription factors that were expressed in the respective tissue. Indeed, we observed an increase of the mean precision to 0.24, which indicates that we could increase the amount of meaningful interactions by additionally filtering for tissue expression. The recall remained low (at 0.03), which is not surprising since many pathway-related proteins were not present in the considered tissue-specific networks and, hence, could not be detected. Again, the amount of pathway proteins on the tissue-specific shortest paths between receptors and TF from the same pathway was significantly larger as compared to shortest paths between receptors and TF from different pathways (p<0.05).
To further investigate if the described context-associations can help to extract pathway information from networks, we compared the frequency of protein pairs being member of the same pathway (as defined by WikiPathways) among tissue-specific PPIs (both proteins where required to be co-expressed in at least one tissue) and compared this frequency to PPIs between proteins that are not expressed in the same tissue. We observed that interacting protein pairs that are expressed in the same tissue are indeed more likely to be in the same pathway as compared to interacting protein pairs that are expressed in disjoint sets of tissues (p<0.001). This, again, demonstrates that the annotations have captured properties related to pathways and suggests that the filtering helps revealing pathway information.
In the next sections we use the context-associated PPI network to obtain novel insights into the mechanisms of human disease: we perform a targeted study of the PPI network surrounding the human proteins that interact with influenza virus proteins to find potential regulators of viral pathogenicity, and we explore the question of whether and how altered protein phosphorylation might be a cause of Alzheimer's disease.
We analyzed PPI data of human proteins that interact with influenza virus proteins. Influenza viruses infect bronchial epithelial tissue and many cell types in the lung, sometimes resulting in viral pneumonia [29]. We started by obtaining a list of 87 human proteins that have been shown to interact with at least one influenza virus protein in a previous study [30]. From this list, we observed that 23 proteins were expressed in bronchial epithelial tissue (BET), in whole lung, or in both tissues - we refer to these proteins as first layer host factors. We created the second layer by filtering tissue-specific proteins (expressed in BET or whole lung) that interact with members of the first-layer (Figure 2A). Together, the first and second layers compose the tissue-specific PPI subnetworks.
Next, we identified known pathways enriched in the BET- and lung-specific PPI subnetworks, and found both similarities and differences in the cellular functions of each (see Materials and Methods for details on the enrichment analysis and a full list of enriched pathways in Supplementary Table S2). Both subnetworks showed enrichment for processes related to programmed cell death and eukaryotic translation. These results are consistent with functions known to be activated or disrupted by influenza virus infection [31], [32], [33]. In addition, proteins in the BET subnetwork exhibited a stronger signature in processes involved with transcriptional regulation, sumoylation, and the regulation of mRNA stability (in particular, the stability of AU-rich element-containing mRNAs). Although these processes tend to be associated with general housekeeping functions, we point out that many cytokine and interferon mRNAs contain AU-rich elements [34]. This observation suggests, hypothetically, that influenza virus proteins may function to dysregulate cytokine mRNA stability in BET, a function that could impact influenza virus pathogenesis through modulation of immune cell infiltration and function. In relation to sumoylation, it has been noted recently that influenza virus can gain protein functionality during infection by interacting with the sumoylation system of the host cell [35]. On the other hand, the lung subnetwork was uniquely enriched for processes related to cell-substrate adhesion (pathway “signaling events mediated by focal adhesion kinase”). Because cell adhesion is important for maintaining cellular viability and epithelial barrier function, it is possible that influenza virus protein-mediated interference with this process could impact both the amount of virus-inflicted damage upon the lung and dissemination of influenza virus into extra-pulmonary sites.
Cells respond to influenza infection by producing cytokines and chemokines [36], [37], while viral proteins counteract this innate immune response. One example of a viral protein that directly interferes on the protein level with cellular immune pathways is NS1 (its involvement in immune response suppression is reviewed in [38]). Here, we noted that the lung PPI subnetwork – which was centered on viral protein-host protein interactions – was enriched for several curated pathways involving Toll-like receptor (TLR) and IL-1 receptor (IL-1R) signaling (e.g., “TLR JNK”, “TRAF6 mediated IRF7 activation in TLR7/8 or 9 signalling”, “IL-1 JNK”, “TLR ECSIT MEKK1 JNK” and “IL1-mediated signaling events”). Although these pathways are expected to be activated in response to viral infection, no previous study has identified any role for any influenza virus protein in perturbing TLR or IL-1R signal transduction. Several host proteins were consistently observed in most/all of the enriched TLR/IL-1R pathways from the influenza PPI lung subnetwork, including IRAK1, TOLLIP and MyD88. Under normal conditions, the IRAK1 kinase associates with TOLLIP (an inhibitory molecule), and upon receptor stimulation, IRAK1 is recruited to the TLR/IL1R-receptor complex through its interaction with MyD88 (reviewed in [39]). Recruitment results in activation of IRAK1 kinase activity and subsequent activation of MAP kinase pathways, NF-κB-dependent gene expression and interferon α induction. Altogether, these observations suggest the novel possibility that influenza virus proteins interfere with TLR/IL-1R signaling in lung – possibly by accessing a critical regulator of TLR/IL-1R signal transduction (i.e., IRAK1) – an observation that may have implications for the regulation of pathogenesis associated with influenza virus infections.
A recent study demonstrated that signaling through the IL-1 receptor has a protective effect in mice infected with the pandemic 1918 influenza virus [40]. Another study reported that IL-1 receptor-deficient mice succumbed more easily than wild-type mice to infection with an H5N1 virus of low pathogenicity (A/Hong Kong/486/1997) [41]. Moreover, IL-1 receptor-deficient mice showed reduced inflammatory pathology upon infection with A/Puerto Rico/8/34 (H1N1) influenza virus [42]. Several studies also established that influenza virus infection is sensed by TLR7 in plasmacytoid dendritic cells [43], [44], [45], [46], [47], [48]. However, none of these studies addressed the significance of IRAK1 in influenza virus pathogenicity. Our study thus exemplifies how our network analysis can identify potential regulators of influenza pathogenicity for experimental testing, for example, by assessing influenza virus infections in IRAK1-deficient cells or mice.
Next, we aimed to predict more specific novel interference mechanisms by constructing directed and tissue-specific protein networks linking the viral proteins with proteins whose corresponding transcript was up-regulated after influenza virus infection. We selected steadily up-regulated transcripts from a microarray experiment measuring gene expression changes over time in a lung epithelial cell line infected with a 2009 pandemic H1N1 virus [26] (228 transcripts were selected in total; see Materials and Methods for more details). As expected, all ten most strongly enriched known pathways among the selected transcripts were involved in infection and the immune response. For example, the most highly overrepresented pathway was interferon alpha-beta signaling (p<10e-20).
We constructed BET- and lung-specific networks connecting the viral proteins with the 228 up-regulated factors by shortest paths. From the shortest paths we assigned directions to the edges on these paths. The directed networks consisted of 577 (BET) and 1056 (lung) PPIs. To examine if these networks might reveal relevant information on how viral proteins interfere with the cellular immune response, we tested for enrichment of known pathways in the directed networks. We found that the directed networks were strongly enriched in immune response-related pathways (especially cytokine-related) even after excluding the 228 up-regulated transcripts, indicating that enrichment was independent of the high fraction of immune response factors in the transcriptomics data (Supplementary Table S3). For example, we observed a significant enrichment in both the directed BET- and lung-specific networks for proteins related to IL-2 and IL-6 signaling and focal adhesions (q-values<0.05). This suggested that we, indeed, might have captured relevant crosstalk between the viral proteins and immune pathways. The full networks are included in the File S1.
To mine the directed networks for interactions that are involved in interference mechanisms of the viral proteins with the cellular immune response, we concentrated, again, on layer one and two host factor proteins on the shortest paths. From the list of curated pathways enriched in both the BET and the lung directed networks (Supplementary Table S3), we selected several cytokine-related pathways (marked in Supplementary Table S3) and filtered for interactions where the second layer protein was in one of these pathways but the layer one protein was not (to specifically detect novel, indirect interference mechanisms). This resulted in a comprehensive BET network consisting of 49 interactions and a lung network formed by 67 interactions including viral proteins and host factors up to layer two (see Supplementary Table S4 for the comprehensive networks and Figure 2 for a manually curated subset of these networks).
Close inspection of these comprehensive cytokine-related networks in both BET and lung revealed several points of potential viral protein-mediated interference with inflammatory pathways (Figure 2). For example, the BET network showed interactions between viral polymerase complex proteins (i.e., PB1 and PB2) and BHLHE40, a transcriptional regulator that cooperates with HDAC1 to repress STAT1 activity [49] (Figure 2B). STAT1 is essential for the activation of interferon stimulated genes, which repress viral replication, and while influenza virus has an established ability to impair STAT1 [50], no such function has been assigned to any of the viral polymerase complex subunits. BHLHE40 also interacts with TOLLIP, a suppressor of TLR signaling [51] (see also the discussion of lung-specific inflammatory pathways above). This implies that the BHLHE40 protein could act as an important access point for influenza virus-mediated interference with host antiviral and inflammatory regulation in BET, and further that viral polymerase subunits may have an important – yet unappreciated – role in this activity.
As in BET, lung-specific cytokine-related networks revealed that influenza virus proteins interface with TOLLIP (Figure 2C). However, it is notable that, in lung, this interaction occurs through BHLHE40 and two additional routes (i.e., MAGED1 and RBPMS), potentially involving up to four viral proteins: (i) the aforementioned polymerase complex subunits, PB1 and PB2; (ii) the viral ion channel protein, M2; (iii) and the viral RNA-binding nucleoprotein, NP. Thus, access to TOLLIP might be particularly important in lung. The PB1/PB2-BHLHE40 interaction is maintained in this tissue type, although the nature of the interaction may differ compared to BET. Specifically, BHLHE40 may favor interaction with STAT3 (Figure 2C), and previous evidence indicates that BHLHE40 stimulates STAT3 activity rather than inducing inhibition [52]. Thus, analysis of context-specific PPIs – in combination with influenza virus-induced changes in the cellular transcriptome – reveal important, putative tissue-specific differences in the ability of viral proteins to interact with cellular immune response signaling networks. Additional experiments will be necessary to further establish the functions of these interactions.
Assuming no prior expert knowledge on a given topic, we applied a systematic protocol which can, in principle, be used to interrogate the PPI network about the involvement of protein interactions in a complex biological question according to current knowledge. In general, altered states of protein phosphorylation affect the PPI network and can lead to pathogenesis. Our goal in this example was to investigate the possible role of protein phosphorylation in Alzheimer's disease (AD), the most common form of dementia. AD is a degenerative disease manifesting in the brain, and its cause has been hypothesized to be the formation of protein aggregates leading to neuron death, in particular related to the abnormal phosphorylation of the microtubule-associated protein tau [53].
First, we need to input a list of proteins related to the topic. Using a literature mining protocol (see Materials and Methods for details) we generated a list of PPIs related to Alzheimer's and protein phosphorylation: PSEN1:PSEN2, GSK3B:MAPT, APP:BACE1, and PPP2R4:SET. We then studied the network surrounding these interactors (Figure 3).
The initial PPI network contained 727 interactions (Figure 4A). Interactions could be further filtered on the basis of reasonable criteria, namely by tissue filtering for housekeeping and genes expressed in the brain (we selected “whole brain” and “prefrontal cortex”), and filtering for genes related to the GO term “cell death”, reflecting that AD is characterized by death of neural cells (Figure 4B). Finally, to reveal potential signal transduction pathways we used the inference of edge directionality from receptors to TFs described above (Figure 4B).
Within the resulting network, we highlighted the following path (Figure 4): LRP6-GSK3B-MAPT-AATF. The low density lipoprotein receptor-related protein 6 (LRP6) interacts with glycogen synthase kinase 3B and attenuates the kinase's ability to phosphorylate microtubule associated protein tau (MAPT) [54]. Tau protein can contribute to AD in different ways: 1) the hyperphosphorylation of tau protein can affect microtubule stability, leading to a disassociation of tau protein from the microtubule, possibly followed by the aggregation of phosphorylated tau into neurofibrillary tangles, which are observed in the brains of AD patients [55]; 2) mediated by protein phosphatase 1 and GSK3 activity, Tau filaments interfere with axonal transport in the neuron, which is consistent with deficiencies in axonal transport in AD [56]. Tau protein has been found to co-localize in the cytoplasm with Che-1 (AATF), which is an evolutionarily conserved RNA polymerase II binding protein that accumulates in the cell upon DNA damage [57]. It appears that Che-1/Tau proteins dissociate during neuronal cell death [58]; however, the function of Che-1 in the cytoplasm is unclear, as Che-1 is a nuclear protein that is involved in gene regulation of E2F1 targets and p53 and has pro-proliferative and anti-apoptotic functions [59]. Together, these interactions suggest a complex interplay whereby the Tau phosphorylation state and structure, and context-dependent protein distribution within the cell may contribute to neuronal cell death and AD pathology. An unbiased search for protein phosphorylation in relation to cell death in AD pointed us to this interesting pathway.
The incorporation of tissue-specific expression information to create PPI subnetworks is a useful method to elucidate biological processes that cannot be observed when using the complete PPI network. Here we have shown an approach for the inference of associated context for PPIs based on the annotations of the interacting partners, which enhances the relevance of the annotated interactions. Interactions between proteins expressed in the same location (e.g. lung) or at the same time or developmental stage (e.g. embryo development) can then be selected. Directed pathways can be inferred and highlighted in the filtered network according to sets of sources and sinks corresponding to receptors and transcription factors. Using this approach we were able to identify novel, tissue-specific interactions between influenza virus proteins and cellular inflammatory signaling pathways that may regulate pathogenesis associated with infection, and to describe a brain-specific protein phosphorylation pathway relevant for Alzheimer's disease.
Several methods exist to create subnetworks of the human interactome based on context criteria. For example, POINeT [11] integrates the major PPI databases and allows the creation of tissue-specific networks. To our knowledge we are the first to combine edge directionality, gene expression and functional information for the detection of meaningful interactions. Some approaches exist that infer information flow in a network from the shortest paths (or ‘lowest costs’ if costs are associated with edges) that connects a set of source nodes with sink nodes. Cytoscape plug-ins such as BisoGenet [60] and GenePro [61] find the shortest paths between nodes of the gene and protein network and represent properties of the nodes. SPIKE [62] includes curated pathway data and also calculates pathway inference. The task of identifying signaling events from PPI data and functional protein annotation alone has been addressed in several studies [24], [63], [64] and implemented in tools (e.g. ANAT [65]). Here, we proposed a protocol for edge directionality prediction based on calculating the shortest paths between sources and sinks. This protocol is runtime-efficient, which allowed us to provide it as a web tool that is the first to combine both PPI analysis for inference of edge directionality and PPI filtering by tissue and function (available from http://cbdm.mdc-berlin.de/tools/hippie/).
In summary, we have presented and made available an approach to associate context to PPI networks, which provides novel biological insight into mechanisms of disease. The continuing generation of PPI data and further incorporation into databases, and an increasing quality of annotations attached to genes and proteins will result in further improvements of our methodology.
|
10.1371/journal.pntd.0007092 | Cutaneous leishmaniasis and co-morbid major depressive disorder: A systematic review with burden estimates | Major depressive disorder (MDD) associated with chronic neglected tropical diseases (NTDs) has been identified as a significant and overlooked contributor to overall disease burden. Cutaneous leishmaniasis (CL) is one of the most prevalent and stigmatising NTDs, with an incidence of around 1 million new cases of active CL infection annually. However, the characteristic residual scarring (inactive CL) following almost all cases of active CL has only recently been recognised as part of the CL disease spectrum due to its lasting psychosocial impact.
We performed a multi-language systematic review of the psychosocial impact of active and inactive CL. We estimated inactive CL (iCL) prevalence for the first time using reported WHO active CL (aCL) incidence data that were adjusted for life expectancy and underreporting. We then quantified the disability (YLD) burden of co-morbid MDD in CL using MDD disability weights at three severity levels. Overall, we identified 29 studies of CL psychological impact from 5 WHO regions, representing 11 of the 50 highest burden countries for CL. We conservatively calculated the disability burden of co-morbid MDD in CL to be 1.9 million YLDs, which equalled the overall (DALY) disease burden (assuming no excess mortality in depressed CL patients). Thus, upon inclusion of co-morbid MDD alone in both active and inactive CL, the DALY burden was seven times higher than the latest 2016 Global Burden of Disease study estimates, which notably omitted both psychological impact and inactive CL.
Failure to include co-morbid MDD and the lasting sequelae of chronic NTDs, as exemplified by CL, leads to large underestimates of overall disease burden.
| Cutaneous leishmaniasis is a highly prevalent vector-borne disease affecting large parts of Latin America and the Middle East, as well as parts of Northern Africa. There are several types of Cutaneous leishmaniasis, almost all of which have an active phase characterized by a disfiguring lesion (typically on exposed parts of the body), which then becomes a permanent scar (the inactive phase). We recently published an article highlighting the impact of the inactive scarring phase of CL on affected individuals, which is associated with high levels of stigma. Nevertheless, this aspect of the disease is not considered in its own right when calculating the overall disease burden by the Global Burden of Disease (GBD) Studies. In this article we estimate the prevalence of depression (major depressive disorder) in cutaneous leishmaniasis, in both the active and inactive forms. We then show the contribution of inactive CL to the overall disease burden estimates when included, which is due to the large psychological impact it has on those affected by it. We also highlight the importance of further similar efforts for other NTDs which have a chronic course, and which are also not sufficiently included in disease burden calculations at present.
| Cutaneous leishmaniasis (CL) is the most prevalent form of leishmaniasis and 1 of 22 highly prevalent neglected tropical diseases (NTD) [1]. Current disease classifications differentiate aspects of the active (nodular, ulcerative or plaque) CL lesion in terms of its transmission route (“zoonotic” vs “anthroponotic”), geographical location (“New World” vs “Old World”), and extent of its dermatological manifestations (“diffuse” vs “localised”) [2]. However, none capture the characteristic stigmatisation and psychological sequelae of life-long residual CL scarring that accompanies active infection in almost all cases. As such, we recently expanded the spectrum of CL disease by introducing new terminology—active (aCL) and inactive (iCL) scarring cutaneous leishmaniasis—to describe the dermatological changes of CL in relation to its disease activity [3]. Such a classification is also inclusive of long-term sequelae such as mucocutaneous leishmaniasis (MCL), which develops in a minority of CL cases (~4%) [4] mainly in the Americas and East African regions and which may represent a reactive form of CL [5].
The stigmatisation resulting from visible active and inactive CL lesions can be traced back centuries and was probably a major driver in establishing the ancient practice of leishmanisation [6]. Nevertheless, this defining psychosocial aspect of cutaneous leishmaniasis has been almost completely overlooked by successive disease burden studies [7–10]. Furthermore, the prevalence of inactive CL has not previously been estimated and as such is not presently incorporated into burden estimates. This unfortunately underlines a habitual lack of consideration for the chronic sequelae of NTDs. Regrettably, as CL is not a life-limiting infection, policy-makers often neglect CL as a priority disease [11–13] despite its importance to endemic communities and its links to poverty [14]. This oversight is particularly problematic given the increasing CL incidence in highly endemic conflict zones of Afghanistan, Iraq, the Syrian Arab Republic, and Yemen, creating a major public health problem [15,16].
Major Depressive Disorder (MDD) is the most prevalent form of mental disorder, affecting 4.4% of world’s population [17]. The diagnosis of MDD is symptom-based and follows the Disease Statistical Manual (DSM). MDD is one of two depressive disorders that account for the fifth largest cause of disability (years of life lived with disability; YLD) in the latest 2016 Global Burden of Disease (GBD) Study [18]. There is also a growing recognition by the global mental health community of the importance of adopting a more inclusive approach to mental health and disease, from wellness to subclinical distress to clinical “disorder”, known as the staged model of depression [19].
The psychological impact of NTDs is an area that has only recently been emphasised in the NTD community [20]. For example, mental ill health was not included in recent calculations of disability-adjusted life years (DALYs) by NTD programmes, suggesting that the psychological impact of these conditions is not a primary outcome of such programmes [21]. It is therefore unsurprising that previous global burden of depression studies appear to exclude NTDs from their prevalence and burden estimates [17,22,23] This omission is highly significant for two reasons: Many NTDs are uniquely stigmatizing [20], and collectively, WHO estimates that NTDs affect over 1 billion (or 1 in 6) people worldwide [1].
In summary, CL is often ignored at the policy level due to its lack of mortality, and is therefore a prime example of a stigmatising, prevalent NTD whose associated mental illness is disregarded. The aims of the present study are two-fold: 1) To conduct a systematic review of the psychological impact of cutaneous leishmaniasis; 2) To quantify the burden of co-morbid major depressive disorder in this highly prevalent and stigmatising condition for the first time.
Our study reflects the current approach to disease burden estimates, which are based upon MDD as classified by the DSM [22]. We have also adopted the staged model of depression to use additional evidence from psychological and quality of life studies. These latter studies were used to calculate stages of subclinical distress associated with CL and to quantify its overall psychosocial impact.
There are four steps to calculating the burden of co-morbid depression (in DALYs) due to CL. Firstly, we conducted a systematic review of the psychosocial impact of all forms of CL (including MCL). To quantify the overall impact of iCL as part of the burden of CL, we also had to generate estimates of iCL prevalence for the first time. Following these first two steps, we then estimated the prevalence of MDD co-morbidity and its severity in aCL and iCL patients. We did not calculate the burden of co-morbid MCL as the associated mortality rate is not known and therefore prevalence estimates could not be reliably calculated. Finally, we multiplied the prevalence of aCL and iCL with co-morbid MDD by the disability weight (DW) for MDD at three severity levels (mild, moderate, and severe) following the methodology of Ton et al (2015) [24] (see Fig 1).
The search strategy queried four Ovid databases–Medline [25], EMBASE [26], Global Health [27], and PSYCInfo [28]–as well as LILACS [29], using English, French, Spanish, and Portuguese search terms on 4th December 2017. Additional searches through Google Scholar [30] were performed in Arabic and English, along with back referencing of relevant articles and a grey literature search. The search strategy accounted for common terms and abbreviations for cutaneous leishmaniasis (e.g. “CL” and “cutaneous leishmaniasis”), and combined these with key words for major depressive disorder and its symptoms, as well as general psychological impact (e.g. “psych*”, “major depressive disorder”, “distress”). We included all relevant psychological studies in CL patients and those with reliable knowledge of their experiences (i.e. their caregivers and their care providers) (Fig 2). As such, community studies were excluded from our final analysis except to further contextualise our findings. Please see S1 Appendix for further details of the search strategy and individual terms queried. Please see S2 Appendix for our inclusion and exclusion criteria, and S3 Appendix for the reasons for excluding studies from final analysis.
Twenty-nine studies were included in the final analysis of the psychosocial impact of CL (see S4 Appendix). The large majority (25/29) of studies were based in middle-income countries (18/29 UMIC, 7/29 LMIC) [32]. Similarly, most studies took place in the highest burden world regions (12/29 in the Eastern Mediterranean Region (EMR) and 11/29 in the Americas Region (AMR)), and included 11 of the 50 highest burden countries for CL in the world [9].
Studies that quantified an MDD diagnosis or symptoms using both validated (e.g. SCID-1; BDI) and unvalidated tools (e.g. self-reported depression symptoms) were used to determine rates of co-morbid MDD in both aCL and iCL (See Table 1). Additional quality of life, stigma, socioeconomic, and qualitative studies were used to generate an estimate of subclinical “distress” as per the staged model of depression (see Tables 2 and 3).
A diagnosis of MDD was consistently reached within the mean or one standard deviation of the mean in CL patients [33,34,36,38], equating to MDD rates of 30–50%. Meanwhile, quantification of symptoms of MDD mostly relied upon self-reporting. As such, symptoms of low mood and depression in CL patients ranged from 12.5–90.9% [34,40–45] aCL patients had significantly higher rates of MDD compared to controls in both children and adults [36] aCL was also found on multivariate analysis to be an independent risk factor for mental disorder in the primary care setting [33]. It is therefore unlikely that these results are a product of significant selection bias.
Equally, whilst rates of MDD were not measured for children with iCL, significantly higher rates of MDD were found in adults compared to controls. iCL patients were also at significantly higher suicide risk than controls [34]. In the only study to measure co-morbid MDD in both aCL and iCL, CL scarring was associated with non-statistically significantly higher MDD scores [38]. These findings are important, as considerably more patients are in the inactive (scarring) phase of CL than in the active phase. Although the data suggest that rates of MDD in iCL are at least equal to those found in aCL patients, the majority of studies (16/29) focused exclusively on aCL.
More broadly, quality of life was found to be significantly decreased in CL patients compared with controls. Stigma was a characteristic feature of CL in most quantitative and qualitative studies, whilst psychological distress was found to be between 50–90% [46,55]. Similarly, issues of disfigurement and reduced capacity to work affected the majority of sufferers (see Table 2). Interestingly, the psychological burden extended to CL caregivers, who were also found to have significantly elevated depression rates [36] and diminished quality of life [36,49] compared to controls.
Overall, CL is associated with a high degree of psychological morbidity irrespective of country, age, and disease activity. We present two other important patient- and disease-specific variables considered during our analysis: patient sex and lesion location. These were chosen due to multiple reports linking them with increased psychosocial impact. Indeed, despite findings of qualitative studies that facial lesions are the most psychologically damaging [42,45,63,67], none of the four quantitative studies [34,46,52,54] providing subgroup analysis demonstrated a statistically significant association with facial lesions and worsening psychological outcomes. Moreover, facial iCL scars were actually associated with lower rates of depression and suicidality than those located on other parts of the body [34]. Instead, it may be more appropriate to differentiate the visibility of lesions in future studies.
A significant number of studies focused solely on women (5/29) on the basis that women are generally at greater risk of depression [17]. It is therefore important to consider possible sex differences in MDD rates given that men have more reported cases of CL than women in most endemic countries [4] Interestingly, women-only studies were found to have comparable MDD rates to mixed sex studies, although differences in self-reported symptoms of MDD were noted in some countries [43,44]. The reasons for these findings could perhaps be explained by community [68], socio-economic [62], and qualitative studies [67]. For example, whilst women are commonly more concerned by bodily appearance and marital prospects, a roughly equal impact is placed upon men through incapacity to work and perform leadership responsibilities [52] due to the disease.
Based on the available evidence, we conservatively estimate that 70% of individuals with both active and inactive CL will experience some degree of psychological morbidity. This ranges from subclinical “distress” (50%) to clinical “disorder” (20%), in accordance with the staged model of depression [19] As such, 30% of CL patients fall into the “wellness” category of the model, in view of regional differences in psychosocial impact [55,65] and the small number of countries and endemic communities in which CL is less stigmatizing [59] and perceived as less severe [69] (see Table 4).
The 2016 GBD Study provides CL prevalence estimates that account solely for aCL and that also include MCL within them unseparated. As such, the prevalence of inactive (scarring) CL has not been previously estimated, and is not incorporated formally into the GBD burden estimates for CL. The methodology for calculating the prevalence of inactive CL has been previously described [3]. In short, our calculations are derived from the latest reported aCL incidence data from WHO spanning 2006–2015 [70] that have been adjusted for underreporting [10,71] and the presence of MCL within them [72–74] (see Table 5). We assume zero CL-associated mortality and a life expectancy of 30 years with scarring; this is a conservative longevity estimate considering the life expectancy of at-risk populations in high burden countries [74]. For further information on this methodology, please see S5 Appendix.
GBD Studies differentiate the severity of episodes of MDD at three levels—mild, moderate, and severe–each with its own disability weight [22]. Therefore, it is necessary to calculate the severity of co-morbid MDD in CL patients to calculate the disability burden (YLD) component of the DALY.
In the studies we identified, the mean depression scores of CL patients equated to mild MDD, with moderate MDD scores being reached within one standard deviation in most studies. Furthermore, in a study of depression in inactive CL using Beck’s Depression Inventory, ~70% of cases with depression scored in “mild” severity [34]. Due to the relatively small sample sizes and difficulties in comparing MDD severity from different measurement tools, we used data from the 2010 GBD study on depressive disorders to help inform our estimates (see Table 6). In that study, the patient MDD cohort was classified accordingly: 72.7% with Mild severity; 16.5% with Moderate severity; and 10.8% with Severe MDD [22].
Applying the estimate for MDD severity to our prevalence estimates for cutaneous leishmaniasis, the following YLDs were calculated: 200,000 for active CL, and 1.7 million for inactive CL (combined total 1.9 million YLDs for CL) (see Table 7 and Table 8). We assumed no mortality burden associated with MDD co-morbid to cutaneous leishmaniasis, and as such our YLD figures equalled the overall DALY figures (see S6 Appendix for in-depth calculations). These figures only represent the impact of co-morbid MDD in this condition and do not account for the impact of other mental disorders such as anxiety disorders or the subclinical state of distress as per the staged model of depression [19].
The results presented here challenge the most recent GBD estimates for the overall burden of CL given the prevalence of mental illness reported in the literature for the condition. We highlight the lack of reliable prevalence estimates on which GBD figures are based. We further emphasise that, despite the increased recognition of NTDs through their inclusion within the UN Sustainable Development Goal (SDG) health targets, the burden of mental health associated with stigmatising and chronically disabling NTDs is not appropriately factored into the calculations of overall global mental health estimates. We stress the importance of residual disease on the continuing suffering of those with NTDs using the example of inactive CL.
Indeed, inclusion of iCL increases the CL prevalence estimate 10-fold, which substantially increases the CL disease burden in itself. However, factoring in the burden of co-morbid MDD for both aCL and iCL further increases its overall burden to 2.2 million DALYs. This is approximately eight times greater than the previous DALY estimate reported in the 2016 GBD study that accounted for aCL alone [76]; this is despite our conservative estimate of only a 30 years of life expectancy post-lesion acquisition (see Table 8). Significant increases in burden estimates were calculated previously for lymphatic filariasis [24], indicating that mental illness is grossly unaccounted for in the NTD GBD estimates.
These findings come at a crucial time for those affected by CL, a growing number of whom continue to be affected by war and displacement in current conflict zones. The inclusion of iCL into prevalence estimates for CL, we argue, is necessary to enact changes at the policy level that reflect the importance of CL to affected individuals and their communities. Moreover, the studies we have highlighted show a clear benefit for psychological as well as physical therapies on quality of life [41,47,51] as well as rates of depression [41] in CL patients; sadly, inability to access any form of treatment is a commonly cited major concern for patients [45,55,63,64]. As such, there is a very clear opportunity for national NTD programmes and partner international NGOs to incorporate mental health care into their activities and to provide appropriate services to tackle this growing public health problem.
Overall, the stigma and depression linked to NTDs represent areas of global health that have only recently been highlighted [21]. From our literature review, the previous GBD estimates for depression (which predict depressive disorders as a leading cause of DALYs) do not incorporate MDD (or any other mental illness) associated with NTDs. Omitting NTDs from such consideration of global mental health burden is significant as NTDs have been estimated by WHO to affect over 1 billion (1 in 6) people worldwide [1].
In the latest 2016 iteration of the GBD study, the psychological impact of CL scarring has been incorporated into the disease burden estimates for the first time via a modification of disability weights (DW) (IHME personal communication). As such, the disability burden of CL has increased from 41,500 [77] to 273,000 [18] YLDs. Despite this modification, relying upon DWs to capture the unique psychosocial aspects of NTDs has unfortunately led to some of the most stigmatising (namely CL and leprosy) diseases yielding some of the lowest disability (YLD) estimates of all the NTDs in past iterations [18,77–79]. CL is currently viewed as a “level two disfigurement”, meaning that its DW reflects “a visible physical deformity that causes others to stare and comment. As a result, the person is worried and has trouble sleeping and concentrating”. This corresponds to a DW of 0.067 in GBD 2016, where 0 indicates perfect health and 1 indicates death [75] Thus, we can be confident that our findings represent an unrecognized mental disease burden of CL.
Instead, we strongly recommend that inactive (scarring) CL be included with active CL infection in future CL prevalence estimates, and that MCL and aCL estimates be presented separately for further information. We have shown that with inactive CL, such a large increase in prevalence (10-fold higher) and burden of co-morbid MDD (8-fold increase) is not sufficiently accounted for by simply altering the DWs for active CL given the evidence of mental illness in patients with residual scarring. As we have only included the “disorder” stage of depressive burden in our YLD estimates, our estimate of CL-related distress (50%) using the staged model approach to depression is not accounted for. Here adjustments to DWs for both aCL and iCL would be justified, as a large proportion of affected individuals with both forms of CL experience some degree of quantifiable distress or socially adverse consequences.
Finally, it is important to highlight that the 2016 GBD Study estimates of aCL incidence [18] are almost half those of previously accepted incidence estimates published in 2012 [71]. This is despite the marked increase in CL incidence due to ongoing conflict and displacement in the Middle East [15]. Similarly, our aCL burden estimates are based upon the 2016 GBD Study estimates of aCL prevalence to allow for comparisons to be made. However, it is unclear why these prevalence estimates are almost seven times lower than the annual incidence of aCL [18] when the majority of cases of aCL self-heal within 6–12 months [2]. For these reasons, we did not include GBD estimates in our calculations of iCL prevalence.
Although our study is the first to generate prevalence estimates of inactive (scarring) CL, we were cautious of the life span of patients with iCL lesions, which is currently unknown. Whilst the majority of CL infections occur in older children and young adults [4] we took a conservative approach to our iCL prevalence estimates by assuming just 30 years lived with residual scars. Nevertheless, given that the majority of aCL cases occur in the young and working adult populations, this figure could be significantly higher. We also conservatively assume no mortality burden with CL, yet suicidal risk and ideation has been noted in both aCL and iCL patients [34,63].
Secondly, we acknowledge our failure to include prevalence and isolated burden estimates for co-morbid MDD in mucocutaneous leishmaniasis (MCL). As discussed, MCL prevalence (and YLD burden) has not been separated from that of aCL in GBD Studies. A further complicating factor is the mortality rate of MCL, which has not been established and consequently prevented us from generating reliable MCL prevalence estimates from WHO incidence data. Nevertheless, the experience of shame in CL patients [45,59] was surprisingly higher than that found in a study of mixed MCL and CL patients [61]. However, in a study of MCL patients alone [80], notably those with severe disease, rates of social exclusion and reduced quality of life were comparable to those found in CL patients [45,52,54,62]. It is possible that the prevalence of co-morbid MDD in MCL patients is similar to that of aCL patients (~20% of cases), meaning that our aCL burden estimates may be relatively unaffected by the presence of MCL cases within them.
This is the first study to estimate the burden of a co-morbid mental disorder in aCL and iCL. One major limitation of our estimates is the evidence underpinning them. We recognize that our 29 studies represent only a relatively small proportion of the global CL caseload. Nevertheless, our systematic literature review has identified the most evidence of psychological impact in CL patients to date, and doubled the evidence of previous recent attempts [81]. Moreover, these studies represent a range of geographically diverse populations across several levels of economic development. In our analysis, studies quantifying MDD using robust and internationally recognised criteria (i.e. DSM) were given the most weight in generating our final estimates of MDD co-morbid to CL. We were also selective and chose to only utilize studies of CL patients and their care providers. In order to minimize the effects of bias we accounted for patient- and disease-specific variables such as sex, age, lesion location, and country of study. As results for co-morbid MDD were comparable when these variables changed, we were confident that none of these variables could have significantly biased our overall estimates.
Finally, whilst depressive disorders represent the most prevalent form of mental disorder worldwide, CL patients are affected by a range of other mental disorders, which have not been included in our estimates. Indeed, CL patients may be at even greater risk of multiple mental disorders [33]. These include generalised anxiety disorder, which may predominate in the active CL phase [36,38] post-traumatic stress disorder [33] and mixed anxiety and depressive disorder [41], the latter of which is not independently considered within the GBD framework at present.
Social stigma, disfigurement, and patient suffering are some of the most identifiable features of NTDs, as emphasized by the case of cutaneous leishmaniasis. However, the suffering of those with active infection as well as those who remain disfigured by NTDs post-infection is not adequately factored into NTD programmes or burden estimates. We reason that there is value in striving for both goals by placing the individual at the centre of such programmes to achieve the holistic care of individuals affected by NTDs. After all, focusing on the disease alone ignores the characteristic disability associated with NTDs such as cutaneous leishmaniasis, leprosy, and filariasis, and risks leaving affected individuals behind.
|
10.1371/journal.pcbi.1004892 | Single-Cell Co-expression Analysis Reveals Distinct Functional Modules, Co-regulation Mechanisms and Clinical Outcomes | Co-expression analysis has been employed to predict gene function, identify functional modules, and determine tumor subtypes. Previous co-expression analysis was mainly conducted at bulk tissue level. It is unclear whether co-expression analysis at the single-cell level will provide novel insights into transcriptional regulation. Here we developed a computational approach to compare glioblastoma expression profiles at the single-cell level with those obtained from bulk tumors. We found that the co-expressed genes observed in single cells and bulk tumors have little overlap and show distinct characteristics. The co-expressed genes identified in bulk tumors tend to have similar biological functions, and are enriched for intrachromosomal interactions with synchronized promoter activity. In contrast, single-cell co-expressed genes are enriched for known protein-protein interactions, and are regulated through interchromosomal interactions. Moreover, gene members of some protein complexes are co-expressed only at the bulk level, while those of other complexes are co-expressed at both single-cell and bulk levels. Finally, we identified a set of co-expressed genes that can predict the survival of glioblastoma patients. Our study highlights that comparative analyses of single-cell and bulk gene expression profiles enable us to identify functional modules that are regulated at different levels and hold great translational potential.
| With the development of single-cell sequencing, an increasing number of biological insights were revealed at the single-cell resolution. Here we integrated the expression profiles from single cells and bulk tissues to discover that a majority of gene pairs were specifically co-expressed at single-cell and bulk levels. Our comparative analysis reveals co-expressed functional modules at different levels, and suggests a distinct regulatory mechanism in which single-cell co-expressed genes are regulated through physical interactions from different chromosomes. Moreover, we found a set of co-expressed genes to predict patient survival. This study suggests that single-cell and bulk co-expression analysis could provide novel biological insights and great clinical potential.
| Gene expression is often coordinated to carry out cellular activities and biological functions [1]. If the expression levels of two genes rise and fall together across different conditions, they are likely to be members of the same protein complex or participate in the same biological pathways. Therefore, co-expression analysis has been widely used to predict protein-protein interactions (PPIs) or annotate functions of uncharacterized genes [2–4]. Built upon co-expression relationships, co-expression networks were often constructed to reveal the functional modules consisting of genes with functional relationships [5–7]. Furthermore, co-expression relationships are often considered to be the consequence of co-regulation that is governed by the same regulatory machinery. Therefore, regulatory elements could be predicted based on the co-expression relationships [8–10]. In addition, co-expression analysis has been applied to cancer biology. For example, co-expressed gene sets could reveal interaction modules in tumor progression [11], or serve as molecular signatures to classify tumors into different subtypes, which often showed distinct clinical outcomes [12,13].
Previous co-expression analyses were mainly conducted at the bulk level in which a large population of cells was profiled as a whole. Recently, single-cell sequencing has emerged as a powerful tool to investigate cellular variability and intratumor heterogeneity [14–16]. However, it remains elusive whether co-expression analysis at the single-cell level will provide novel biological insights into the molecular principles of transcription regulation that would be otherwise hidden at the bulk level. For example, can the same set of co-expressed genes be identified both at the single-cell and bulk levels from the same tissue origin? Will the comparative co-expression analysis reveal functional modules that are regulated at different levels? Do the co-expression relationships detected at the single-cell and bulk levels reflect the same regulatory mechanisms?
To address these important questions, we developed a computational approach to perform comparative co-expression analysis between single-cell and bulk samples, and discovered that the majority of the co-expressed gene pairs were unique. Multiple lines of evidence suggest that the discrepancy between the two analyses is not due to technical artifacts. Interestingly, the co-expressed genes in bulk tissues tend to have the same biological functions, while the co-expressed genes in single cells encode proteins that are likely to interact with each other. Strikingly, members in different protein complexes are often predominately connected by one type of co-expression relationships. Furthermore, we find that the co-expression relationships in the single cells and bulk tissues might reflect distinct co-regulatory mechanisms. Interestingly, interchromosomal interactions are highly enriched for single-cell co-expression. Finally, we discover a set of co-expressed genes that can predict the clinical outcome of glioblastoma.
We used glioblastoma as a model system because both single-cell and bulk expression data are available. A dataset of single-cell RNA-seq was obtained from 430 individual cells of five glioblastoma patients [14]. Similarly, gene expression profiles of 120 glioblastomas as bulk tissues were obtained from TCGA consortium [17]. To compare co-expression patterns at single-cell and bulk levels, we calculated Pearson’s correlation coefficients (R) of gene expression for all possible gene pairs across the cells (or tumors).
Strikingly, the majority (> 90%) of co-expressed gene pairs were unique to either single-cell or bulk analysis. For instance, we observed that the expression profiles of two genes, ATP9B and MORC4, were highly correlated at the single-cell level (R = 0.97, Fig 1A); the correlation coefficients calculated separately from the five tumors were also consistent (S1 Fig). However, their correlation was not significant at the bulk level (R = 0.04). Conversely, the expression profiles of REST and ROCK2 were found highly correlated at the bulk level (R = 0.85), but not at the single-cell level (R = 0.00051, Fig 1A and S2 Fig). Globally, we separately identified the top 1,000 most correlated gene pairs at either single-cell or bulk level and cross-examined whether the same pairs were also correlated at the other level. Surprisingly, only 76 (7.6%) of the top 1,000 gene pairs are shared between the bulk and single-cell levels (Fig 1B). For example, RPL41 and RPS14 are co-expressed in both single cells (R = 0.75) and bulk tissues (R = 0.83) (Fig 1A and S3 Fig). However, most co-expressed gene pairs at the single-cell level have no or even negative correlation at the bulk level. Similar pattern was also observed for the top 1,000 correlations at the bulk level. It is worthy to note that the observation is not sensitive to the correlation measurement we choose. For example, if maximal information coefficient (MIC), which is able to capture non-linear relationships [18], was used, a consistent pattern was observed that 96.4% co-expressed gene pairs were specific at single-cell or bulk level (S4 Fig). These results suggested that distinct sets of co-expressed gene pairs were yielded at single-cell and bulk levels.
Several lines of evidence suggest that the discrepancy in co-expression analysis between bulk and single-cell levels is not due to technical artifacts. First, we checked whether expression correlation was sensitive to the samples chosen for analysis. We randomly partitioned the cells (or tumors) into two equal-sized sub-groups and separately calculated corresponding gene expression correlations. The top 1,000 co-expressed genes were highly consistent between the two sub-groups (Fig 1C). For example, 524 (52.4%) of the top 1,000 correlations were shared between the two sub-groups in the single-cell analysis. The remaining 47.6% of gene pairs are also highly correlated, even though they were not in the top 1,000. A similar observation was made for bulk-level analysis. This observation suggested that expression correlations were robust and not sensitive to the samples used for calculation.
Second, we examined whether the dissociation and processing of single cells introduced measurement errors, which could lead to the discrepancy of co-expression between single-cell and bulk levels. For the five glioblastomas with single-cell expression profiles, we averaged gene expression across the individual cells and then compared the average gene expression profiles with the genuine bulk expression profiles from the same glioblastomas. The comparison showed that the average gene expression was highly correlated with the expression in bulk tissue for each tumor (Fig 1D and S5 Fig). These results suggest that the procedure of isolating and harvesting single cells did not introduce much distortion in expression profiles. Furthermore, in comparison of the expression profiles of the five tumors for single-cell sequencing with the other 120 bulk tumors from TCGA, we found that the five samples were dispersed among the 120 glioblastomas (Fig 1D). This result suggests that the five tumors for single-cell analysis are not characteristically different to the 120 glioblastomas for bulk analysis, and both of the datasets were representative of primary glioblastomas.
Third, we explored whether the discrepancy of co-expression patterns between single cells and bulk tissues could be observed in other tissues. Similar analyses were performed using data obtained from prostate cancers. We compared the transcriptome of 122 individual prostate cancer cells with those of 398 bulk prostate cancers from TCGA [19]. The results showed that only 4% of the top 1,000 correlations were shared between single-cell and bulk levels (S6 Fig). Taken together, all of the above analyses suggest that the observation of distinct co-expressed gene pairs in single cells and bulk tissues was valid, and not due to technical artifacts.
In order to dissect the biological roles of the co-expressed genes at the single-cell level, we classified the co-expressed genes into three groups: single-cell specific, bulk specific, and shared at both levels (S7 Fig and see Materials and Methods for the details). In brief, we compared the distributions of expression correlation coefficients from real and randomly shuffled expression profiles to identify the thresholds of significantly positive correlations at single-cell or bulk levels. Using the obtained thresholds, 5,303, 107,851, and 12,584 gene pairs were classified as single-cell specific, bulk specific, and shared co-expressed gene pairs, respectively (S8 Fig).
Next, we attempted to discover distinct characteristics of these three groups of co-expressed genes. We first checked whether protein products of the co-expressed genes were enriched for known PPIs. By surveying the PPI networks of the BioGRID database [20] using the corresponding proteins of those co-expressed genes, we found that bulk specific co-expressed genes were slightly enriched for PPIs. Specifically, the protein products of 591 (0.55%) of the 107,851 co-expressed gene pairs specific to the bulk tissues have known PPI relationships, while only 0.34% was expected for randomized gene pairs (P = 3.8E-91, student’s t-test). In contrast, PPIs were much more enriched in single-cell specific co-expressed genes (90 of 5,303, 1.7%), which was a 5-fold enrichment compared to the expectation (P = 2.0E-247, student’s t-test) (Fig 2A). Strikingly, we observed that 1,167 of 12,584 (9.3%) shared co-expressed genes have PPIs (Fig 2A), a 27-fold enrichment compared to the expectation (P < 1.0E-500, student’s t-test). The enrichment was not due to relatively high correlation coefficients in the shared group. The same trend was also observed if we compared the three groups at the same range of correlation coefficients (Fig 2B). Furthermore, the enrichment for PPIs in co-expressed genes increased with the degree of correlation coefficients, suggesting the fidelity of the relationships between the co-expressed genes and PPIs.
Surprisingly, we observed that the three classes of co-expressions were not homogeneously distributed among annotated protein complexes. Instead, different protein complexes were enriched in different classes of co-expressions. Members of many protein complexes are co-expressed at bulk level, such as proteasome and CDC5L complex (Fig 2C). However, members in other complexes (e.g. condensing II) are co-expressed in both single cells and bulk tissues (Fig 2C). Perhaps the most striking examples are cytoplasmic and mitochondrial ribosomes. Of 1,816 co-expressed gene pairs that belong to the cytoplasmic ribosomal complexes, 1,791 (98.6%) were co-expressed at both single-cell and bulk levels. In contrast, among 329 co-expressed genes of the mitochondrial ribosomal complexes, all of them are bulk specific (Fig 2C). These results suggest that the synchronized expression of members in protein complexes is governed through different types of co-expression relationships, reflecting distinct regulatory mechanisms.
We next examined whether co-expressed genes tend to share similar biological functions. To this end, we calculated the semantic similarity of the biological process (BP) terms of gene ontology (GO) [21] between two genes using GOSemSim [22]. Our analyses demonstrated that the shared and bulk specific co-expressed gene pairs tend to have similar biological functions. Specifically, the fractions of shared and bulk specific co-expressed genes having the same functions were 4.6 and 1.5-fold higher than the expectation, respectively (Fig 3A). In contrast, the single-cell specific co-expressed genes were not enriched for function similarity (0.997-fold, Fig 3A). Nevertheless, gene pairs with the highest correlation coefficients (R > 0.4) at the single-cell level were also enriched for function similarity (Fig 3B).
The three groups of co-expressed gene pairs are enriched for different biological functions. Again, we checked the biological functions associated with the genes from the top 1,000 shared, single-cell specific, or bulk specific co-expression pairs. For single-cell specific co-expressed genes, GO terms including biological adhesion, and regulation of apoptosis were enriched (Fig 3C and S9 Fig). The shared co-expressed genes were associated with translational elongation, and oxidative phosphorylation. It is also interesting to note that these genes are also enriched in neurodegenerative diseases, such as Parkinson’s disease, given the neuronal origin of glioblastomas (Fig 3C and S9 Fig). The bulk specific co-expressed genes were significantly associated with oxidative phosphorylation and neurodegenerative diseases. These results further demonstrated that the different functional modules were associated with different types of co-expressed genes.
To determine whether the underlying regulatory mechanism of co-expression at single-cell level is different from those at the bulk level, we analyzed the possible regulatory relationships for the three groups of co-expressed genes. We examined two distinct and complementary mechanisms for co-expression (Fig 4A). First, we tested whether the co-expressed genes tend to have synchronized activity of their cis-regulatory elements across different physiological conditions. Second, we checked whether the cis-regulatory elements governing each pair of co-expressed genes more likely have physical contact in the three-dimensional nuclear space.
We computed the accessibility of gene promoters annotated by DNase I hypersensitive sites (DHSs) and corresponding DHS signal correlations across 125 human cell types and tissues [23]. Our analysis revealed that shared and bulk specific co-expressed gene pairs had significantly higher DHS correlation than the random expectation. For example, RPL9 and RPS6 belong to the shared co-expressed gene group, and the accessibility of their promoters was perfectly synchronized across the 125 cell types (R = 0.98, Fig 4B). Similarly, the accessibility of another pair, FRYL and RAPGEF6, bulk specific co-expressed genes, was also highly correlated (R = 0.98). Overall, the highest peak of the distribution of DHS correlation for shared and bulk specific co-expressed gene pairs were located at 0.81 and 0.79, respectively (Fig 4C). In contrast, the correlation coefficient for single-cell specific co-expressed genes was much lower than the other two groups (P < 1.0E-300, student’s t-test). For example, DHS signal of two genes, GIS and GSS, was not correlated (R = 0.068, Fig 4B). The correlation coefficients of single-cell specific genes were much broader distributed, with the highest peak located at 0.12 (Fig 4C).
We then calculated the probability that two genes physically interact with each other based on chromatin interaction data [24]. In IMR90 cell lines, we discovered that the single-cell specific and bulk specific co-expressed genes were more likely to have physical interactions than expectation (Fig 4D). In contrast, the shared co-expressed genes were not enriched for chromatin interactions (Fig 4D). The same observation was confirmed in an independent cell line of hESC (S10 Fig). Our results demonstrated that the datasets obtained from bulk tissues (e.g. DHS and chromatin interactions) could partially explain the co-expression at bulk and single-cell levels, and different types of co-expressions might be regulated by different mechanisms.
Previous co-expression studies at the bulk level have shown that genes within the same topological domain were more likely to interact with each other [24]. Here we asked whether a pair of co-expressed genes resided on the same chromosome or even within the same topological domain. Interestingly, for bulk specific co-expressed genes, we observed that 16.9% were found on the same chromosome, whereas only 5.3% of single-cell specific co-expressed genes were encoded on the same chromosome, which was almost the same as randomly selected gene pairs (average 5.4%, Fig 5A). If we only focused on the top 1,000 highest co-expressed gene pairs, the difference between two levels became even more significant, 47.5% and 5.5% of bulk and single-cell specific genes were located in the same chromosome, respectively (Fig 5B). We further asked to what degree the intrachromosomal co-expressed genes were from the same topological domain [24]. Our analysis revealed that 3–6% of shared and bulk specific intrachromosomal co-expressed gene pairs were located at the same topological domain (Fig 5C). By contrast, no single-cell specific gene pairs were from the same topological domain.
When we separated the co-expressed genes based on whether they were encoded on the same chromosomes, we found that the interchromosomal chromatin interactions were enriched for single-cell specific co-expressed genes (Fig 5D). This result suggests that many co-expressed genes in single cells were co-regulated through interchromosomal interactions, by which the cis-regulatory elements of genes were physically connected and co-regulated by common regulators such as enhancers (Fig 4A).
Recent studies demonstrated that network-based classification approaches provided more power in prediction of clinical outcomes than individual genes [25–27]. We searched the subnetworks within the three types of co-expression networks to identify a set of co-expressed genes that could stratify patients with most significantly different survival time. We classified 120 patients with RNA sequencing from TCGA into two groups based on the expression profiles of genes within each subnetwork (or combination of subnetworks) and compared the survival rates between the two groups (Fig 6A). By examining all subnetworks and the combination of the subnetworks, we discovered a combination that achieved the best separation of patient survival rates, which consisted of 4 shared and 2 single-cell specific co-expressed genes (Fig 6B). The two groups of patients were well separated based on the silhouette plot (S11 Fig). The survival rates of the two groups were significantly different (P = 3.9E-4, log-rank test, FDR < 0.1, Fig 6C). As comparison, we performed the same analysis to bulk co-expressed genes, but no subnetwork was found to classify the patients with significant difference in survival rates (S12 Fig), suggesting that single-cell expression profiles help to improve the prognosis of glioblastoma. Furthermore, we classified the patients into four subtypes according to TCGA classification scheme [28], and their survival rates were not significantly different (S13 Fig).
To confirm the classification power of co-expressed genes, we tested our gene signature using an independent set of 101 glioblastomas whose expressions were profiled using microarray from TCGA. The validation indicated that six-gene signature could significantly stratify poor and favorable survival of the patients (P = 0.013, log-rank test, Fig 6D). These results suggest that the co-expressed gene signature has a great potential to predict patient survival.
Our analysis revealed distinct characteristics for the co-expressed genes at single-cell and bulk levels. The stark difference between the two levels suggests that the single-cell expression profiles provide novel biological insights when they are compared with bulk expression profiles. Interestingly, the DHS and chromatin interaction datasets obtained from bulk tissues could partially explain the co-expression at single-cell level. Nevertheless, we are fully aware of the difference of gene regulation between bulk and single-cell levels. For example, two bulk co-expressed genes could have the same accessibility of regulators in their promoters, whereas the regulation of the two genes at single-cell level is independent to each other and could result in un-correlated accessibilities of the promoters (Fig 4A). If we could deconvolute the signal from the bulk tissues or obtain the datasets on gene regulation at single-cell level, we expect to obtain stronger connection between co-expression and co-regulation. Although a few DHS or ChIP-seq datasets at single-cell level are available [29–31], the data quality is still not optimal (e.g. low sequencing depth). One interesting observation is that majority of the single-cell co-expressed genes are located in different chromosomes, in line with a recent observation that co-expressed odorant receptor genes was not restricted to single chromosome at single-cell level [32]. While the current chromatin interaction analyses are mainly focused on intrachromosomal interactions [33,34], our analysis suggests that interchromosomal interactions are of biological interests.
In our analysis, a set of six co-expressed genes was used to stratify glioblastoma patients into two groups with significantly different survival. Although these genes were selected without prior knowledge of cancer biology, the genes are relevant to glioblastomas. For example, PLXNA4 (plexin A4) has been shown to promote tumor angiogenesis and progression of glioblastoma cells [35]. Similarly, NRXN3 (Neurexin 3) was involved neuron cell-cell adhesion and glioma cell migration [36]. Gene RAB3GAP2 (RAB3 GTPase activating non-catalytic protein subunit 2) was implicated in neurodevelopment and Warburg Micro syndrome [37], whereas PJA2 (praja ring finger ubiquitin ligase 2) degraded MOB1 to support glioblastoma growth [38]. Moreover, both GDI2 (GDP dissociation inhibitor 2) and C21orf33 (chromosome 21 open reading frame 33) was dysregulated in fetal Down syndrome brain [39,40]. All the genes were related to glioblastoma or neural diseases, suggesting their underlying function in tumorigenesis and progression of glioblastoma.
Single cell expression datasets were obtained from references [14,19]. For glioblastoma, 430 individual cells from 5 patients were sequenced for gene expression. For prostate cancer, 122 cells from 22 patients were sequenced. Bulk expression datasets were obtained from The Cancer Genome Atlas (TCGA, https://tcga-data.nci.nih.gov/tcga/). In total, 120 glioblastomas and 398 prostate adenocarcinomas were measured by RNA sequencing at the bulk level.
For bulk expression profile, we excluded the genes whose average expression was below 100 RPKM (Reads Per Kilobase per Million mapped reads). For single-cell gene expression, we excluded the genes if the expression levels across over two-thirds individual cells were equal to zero. Only the genes that were measured at both single-cell and bulk levels were included for further analysis. In total, 4,837 and 4,722 genes were analyzed for glioblastoma and prostate adenocarcinomas, respectively. We performed log2-transformation for RPKM. In order to avoid 0 value for invalid log2-transformation, we added 1 to RPKM value. We then performed global centralization by subtracting corresponding average expression across tissues or cells. Quantile normalization of the expression was further conducted across all samples. All analyses were performed in R platform (http://www.r-project.org/).
Pairwise correlations for all genes were calculated using Pearson correlation coefficient (R). The formula is as follows
R=∑i=1n(xi−x¯)(yi−y¯)∑i=1n(xi−x¯)2∑i=1n(yi−y¯)2
where x, y are gene pairs, and n is the sample size. All gene pairs were ranked according to R values. The hierarchical clustering of expression profiles took Pearson’s correlation coefficient as similarity measurement, and used complete linkage. Similarly, we also used MIC (maximal information coefficient) to measure expression correlation of gene pairs [18].
We then classified the co-expressed genes into three groups: bulk specific, single-cell specific, and shared. Since the distributions of correlation coefficients are quite different between single cells and bulk tissues, we could not choose a uniform cutoff to define the positive correlation. Instead, we developed a shuffled-expression-based algorithm to determine the cutoffs for single-cell and bulk expression separately. Firstly, we shuffled the expression for each gene across the samples, and generated a corresponding distribution of correlation coefficients. We then set the correlation coefficient at the top percentage of 10−6 as cutoff for positive correlation. After setting the cumulative probability of no correlation in random distribution to 0.3 for each side around zero correlation, we obtained the cutoffs of no correlation. The criteria for positive correlation and no correlation are very stringent here because we want to make sure the selected groups of gene pairs are indeed bulk specific or single-cell specific. Those positively correlated gene pairs at both single-cell and bulk levels were assigned to the group of shared co-expressed genes. Single-cell specific co-expressions were those gene pairs with positive correlation at the single-cell level but no correlation at the bulk level. Similarly, those gene pairs with positive correlation at the bulk level but with no correlation at the single-cell level were assigned to bulk specific co-expressed genes.
In order to associate gene co-expressions with protein-protein interactions (PPIs), we downloaded PPIs from BioGrid [20]. We calculated the fraction of co-expressions with PPIs in each type of co-expressed genes. Meanwhile, we generated one thousand sets of 1,000 pairs of genes randomly selected from all gene pairs as control gene pair sets. Each set of control gene pairs were associated with PPIs as well. To calculate the proportion for different ranges of expression correlations, we divided co-expressed genes into equal-interval groups with 0.1 bin size of the correlations.
The components of protein complexes were from CORUM database [41]. All shared and specific co-expressions were mapped to each protein complexes. The layout and view of co-expression network of protein complexes were performed in Cytoscape [42].
We used R package GOSemSim to calculate the semantic similarity of the biological process (BP) terms of gene ontology (GO) [22] between two genes. If similarity value of gene pair ≥ 0.5, the genes were called with GO similarity. Based on this criterion, we calculated the percentage of co-expressed gene pairs and randomly selected gene pairs with GO similarity.
To identify the enriched GO terms, we each chose the top 1,000 co-expressed gene pairs from three groups of co-expressions, respectively. We obtained 129 unique genes from top 1,000 shared co-expressions. After excluding the genes were overlapped between shared and single-cell specific co-expressions, we obtained 319 single-cell specific genes. Besides, 640 genes were unique to bulk specific co-expressions. These three groups of genes were separately performed function enrichment analysis through DAVID software [43]. According to the enriched functions, co-expression networks of top 1,000 correlations were organized into different modules. The genes were assigned to the most significant module if they were enriched in multiple functional modules.
DNase I hypersensitive sites (DHSs) in 125 human cells and tissues were downloaded from ENCODE project [23]. DHSs within the promoter regions (upstream 1,000 base pairs relative to transcription start sites (TSSs)) were associated to genes. If no DHS peaks were found within the promoter regions, the intensity of DHSs of genes was assigned to zero. We then calculated DHS correlations of gene pairs across 125 cell types.
To identify chromatin interaction of co-expressed genes, we used Hi-C data from previous publication [24]. The DNA regions across upstream 5,000, gene body, and downstream 5,000 were used to identify whether gene pairs have chromatin interaction.
Chromosomal relationships of co-expressed gene pairs were plotted using Circus [44]. Topological domains in genome were reported by a previous study [24]. According to the locations of TSSs, gene pairs were determined whether they were located in the same chromosome or topological domain.
We constructed three networks separately from single-cell specific, bulk specific and shared co-expressions. Using ‘Fast Modularity’ software [45], we then determined 56, 42, and 15 dense subnetworks within these three co-expression networks, which reflect the functionally related gene groups. For each subnetwork, we performed hierarchical clustering of patients based on the bulk expression levels of the genes within the subnetwork. The patients were classified into two groups according to the clustering and then compared of their survival using the Kaplan-Meier method [46]. The significance of differential survival between two groups of patients was assessed with a log-rank test. After testing all the subnetworks, 7 shared, 1 single-cell specific and 1 bulk specific subnetworks were found to be able to separate the patients with significantly different survival rates (P < 0.05, log-rank test). We then further examined the combination of at most three significant subnetworks using the same procedure and discovered one combination with the best performance for tumor prognosis. We then estimated the false discovery rate (FDR) using Benjamini and Hochberg approach [47]. The quality of partition of patients was assessed through silhouette graph [48].
In TCGA, another 123 glioblastomas were measured using microarray platform, of which 22 samples were also profiled with RNA sequencing. In order to make the expression levels comparable between microarray samples and sequencing samples, we used one of patients (TCGA-06-0156), which was measured both by RNA sequencing and microarray, for normalization. After log2-transformation of sequencing data, expression profile of each patient subtracted the average expression of TCGA-06-0156 from RNA sequencing. Similarly, expression profiles measured by microarray also subtracted the average expression of microarray-measured TCGA-06-0156. Using the gene signature obtained from the 120 glioblastomas, we then predicted the class of additional 101 microarray-measured glioblastomas (a validation set) through a nearest shrunken centroid [49].
|
10.1371/journal.pntd.0001181 | Experimental Transmission of Leishmania infantum by Two Major Vectors: A Comparison between a Viscerotropic and a Dermotropic Strain | We quantified Leishmania infantum parasites transmitted by natural vectors for the first time. Both L. infantum strains studied, dermotropic CUK3 and viscerotropic IMT373, developed well in Phlebotomus perniciosus and Lutzomyia longipalpis. They produced heavy late-stage infection and colonized the stomodeal valve, which is a prerequisite for successful transmission. Infected sand fly females, and especially those that transmit parasites, feed significantly longer on the host (1.5–1.8 times) than non-transmitting females. Quantitative PCR revealed that P. perniciosus harboured more CUK3 strain parasites, while in L. longipalpis the intensity of infection was higher for the IMT373 strain. However, in both sand fly species the parasite load transmitted was higher for the strain with dermal tropism (CUK3). All but one sand fly female infected by the IMT373 strain transmitted less than 600 promastigotes; in contrast, 29% of L. longipalpis and 14% of P. perniciosus infected with the CUK3 strain transmitted more than 1000 parasites. The parasite number transmitted by individual sand flies ranged from 4 up to 4.19×104 promastigotes; thus, the maximal natural dose found was still about 250 times lower than the experimental challenge dose used in previous studies. This finding emphasizes the importance of determining the natural infective dose for the development of an accurate experimental model useful for the evaluation of new drugs and vaccines.
| Leishmaniasis is a disease caused by protozoan parasites which are transmitted through the bites of infected insects called sand flies. The World Health Organization has estimated that leishmaniases cause 1.6 million new cases annually, of which an estimated 1.1 million are cutaneous or mucocutaneous, and 500,000 are visceral, the most severe form of the disease and fatal if left untreated. The development of a more natural model is crucial for the evaluation of new drugs or vaccine candidates against leishmaniases. The main aim of this study was to quantify the number of Leishmania infantum parasites transmitted by a single sand fly female into the skin of a vertebrate host (mouse). Two L. infantum strains, viscerotropic IMT373 and dermotropic CUK3, were compared in two natural sand fly vectors: Phlebotomus perniciosus and Lutzomyia longipalpis. We found that the parasite number transmitted by individual sand flies ranged from 4 up to 4.19×104. The maximal natural infective dose found in our experiments was about 250 times lower than the experimental challenge dose used in most previous studies.
| Leishmania are intracellular protozoan parasites that establish infection in mammalian hosts following transmission through the bite of an infected phlebotomine sand fly. Visceral leishmaniasis, caused by Leishmania donovani in the Old World and L. infantum in both the Old and New World, invariably leads to death if left untreated [1]. Despite the fact that parasites from the L. donovani complex are mainly associated with disseminated infection of the spleen and liver, it has been shown that L. infantum can also cause cutaneous lesions [2]–[5]. A novel focus of cutaneous leishmaniasis caused by L. infantum was recently described in the Cukurova region in Turkey [6].
During the natural transmission of Leishmania into the dermis, sand flies deposit pharmacologically active saliva [7] and egest parasite-released glycoconjugates, the promastigote secretory gel [8]. Both substances modulate the immune response of the bitten host and enhance the severity of infection (reviewed by [9]).
The ideal leishmaniasis model to test therapeutics and immunoprophylaxis candidates should reproduce the biological and immunological aspects of natural infection and disease. Different approaches regarding the parasite number and route of inoculation have been tested in order to develop an accurate experimental model for the L. donovani complex, most of them using subcutaneous, intraperitoneal or intravenous injections of millions of axenic promastigotes or amastigotes [10]–[11]. Although in some studies up to 107 parasites have been co-inoculated into the dermis with small amounts of sand fly saliva, is not clear how well these experiments mimic natural transmission [12]–[13].
The number of L. infantum parasites inoculated by infected vectors during natural transmission was not previously known, even though a determination of the natural infective dose is crucial for the development of an accurate experimental model to evaluate new drugs and vaccine candidates. In the L. major - P. duboscqi model, it was demonstrated that the number of promastigotes inoculated by individual sand flies ranged between 10 and 1×105 Leishmania [14]. The average number of L. infantum parasites egested was recently reported [15], but the technique used (feeding the pool of infected L. longipalpis through chick skin membrane on culture medium) did not allow an evaluation of the variation in numbers delivered by individual sand flies. Thus, the main aims of this work were to determine the transmission rate and the number of promastigotes inoculated into the skin of mice by individual sand fly females. Phlebotomus perniciosus and Lutzomyia longipalpis, two main vectors of L. infantum in the Mediterranean basin and in the New World, respectively [16], were experimentally infected by L. infantum dermotropic and viscerotropic parasites.
The following results summarize the data obtained in 15 and 10 independent experiments with both vectors and L. infantum strain combinations: 9 with P. perniciosus-IMT373, 6 with P. perniciosus-CUK3, 6 with L. longipalpis-IMT373 and 4 with L. longipalpis-CUK3.
The L. infantum strains studied developed well in both P. perniciosus and L. longipalpis, producing heavy late-stage infection and colonizing the stomodeal valve of the vectors, which is a prerequisite for successful transmission. For both L. infantum strains, the average parasite load in the sand fly midgut is summarized in Table 1. Quantitative PCR revealed that in P. perniciosus the intensity of infection was higher for the CUK3 strain (p = 0.01) while L. longipalpis harboured more IMT373 parasites (p<0.001). However, in both sand fly species the number of parasites transmitted was higher for the dermotropic strain CUK3 (p<0.001); see below.
Out of 88 P. perniciosus, females that bit mice, 62 (70.5%) were infected with CUK3; of these, 36 (58%) delivered parasites into the skin of the mice on days 10–14 post infective blood meal (Fig. 1a). Out of 114 biting L. longipalpis females, 86 (75.5%) were infected and 56 (65% of those infected) inoculated parasites into the mice on days 7–14 post infective blood meal (Fig. 1b).
Despite the fact that the intensity of infection was significantly higher in P. perniciosus (p<0.01), the percentage of transmission and number of inoculated parasites was comparable for both vectors. The parasite load delivered by P. perniciosus and L. longipalpis in the skin of mice ranged between 16 and 4.19×104 and between 4 and 1.11×104, respectively. The average number of CUK3 parasites inoculated into the skin of mice and the percentages of transmission are summarized in Table 1.
In L. longipalpis, the feeding time was positively correlated with the number of CUK3 parasites delivered into host skin (p<0.05), while in P. perniciosus females no such correlation was observed. On the other hand, there was a significant correlation between the pre-feeding load inside both sand fly species' midguts and the number of parasites transmitted (p = 0.0178 for L. longipalpis and p<0.001 for P. perniciosus).
Out of 101 P. pernicious females that bit mice, 73 (72%) were infected with IMT373, and of these 24 (33%) transmitted parasites into the mice's skin. Leishmania transmission occurred between days 9 and 16 post infective bloodmeal (Fig. 2a). From 190 biting L. longipalpis females, 159 (84%) were infected and 23 (14.5% on infected ones) inoculated parasites into the mice between days 7 and 14 post blood meal (Fig. 2b).
In contrast to above, the intensity of infection was significantly higher in L. longipalpis (p<0.001), but the transmission rate (i.e. percentage of transmitting females) and the number of parasites transmitted were significantly higher in P. perniciosus (p<0.01).
The number of parasites transmitted by P. perniciosus and L. longipalpis ranged from 8 to 513 and between 7 and 1240 promastigotes, respectively. The median number of IMT373 transmitted is summarized in Table 1.
For both sand fly species, there was no correlation between feeding time and the number of IMT373 parasites in each female (p = 0.1594), or between the time to take a blood meal and the number of parasites transmitted (p = 0.6666). Moreover, no correlation was observed between the pre-feeding load in each sand fly species and the number of Leishmania delivered (p = 0.1340 for P. perniciosus; p = 0.6473 for L. longipalpis).
For all Leishmania-sand fly combinations, ears were the preferential biting place for sand flies transmitting the parasites, followed by the paws and tail. A few specimens that fed in the nose and eyes were also able to transmit parasites.
Table 2 summarizes the feeding times for both sand fly species: L. longipalpis transmitting IMT373 completed their bloodmeals in times ranging from 2 to 27 minutes, while those transmitting CUK3 parasites needed between 3 to 55 minutes. The maximum and minimum feeding times for P. perniciosus transmitting CUK3 and IMT373 parasites ranged between 4–33 and 1–32 minutes, respectively. Infected sand flies transmitting CUK3 needed more time to feed than those that were infected but non-transmitting, while no differences in feeding time were observed between transmitting and non-transmitting females with IMT373 parasites.
For the first time, we have quantified the number of parasites belonging to L. infantum dermotropic and viscerotropic strains transmitted to the dermis of experimental mice by individual sand fly females. The only previous attempt to calculate the number of transmitted L. infantum parasites was performed just recently [15], with the average number of promastigotes inoculated by 63 L. longipalpis into culture medium through a chicken membrane skin being 457 parasites, with 95% (431 promastigotes) of these corresponding to metacyclic parasites. However, these results do not allow us to take into consideration the individual variability of parasite transmission by a single specimen. The wide range of parasites inoculated per individual sand fly in our study (from 4 up to 4.19×104 promastigotes) is in accordance to data previously obtained with other Leishmania-vector combinations [14], [17], although the approach using microcapillaries as artificial feeding systems [17] could have interfered with the normal sand fly feeding behaviour.
In our study, Phlebotomus perniciosus harboured more L. infantum dermotropic parasites of the CUK3 strain, while in L. longipalpis the intensity of infection was higher for the viscerotropic strain IMT373. However, in both sand fly species the parasite load transmitted was higher for the strain with dermal tropism. All but one sand fly female infected by IMT373 strain transmitted less than 600 promastigotes, the exception being a L. longipalpis female that inoculated 1240 parasites. On the other hand, 29% of L. longipalpis and 14% of P. perniciosus infected with the CUK3 strain transmitted more than 1000 parasites.
The majority of transmitting females inoculated less than 600 parasites. As most of these females were fully engorged by blood we may expect that their feeding pumps (the cibarial and pharyngeal pumps) and stomodeal valve were functioning normally. On the other hand, in those transmitting more than 1000 parasites there was a significant correlation between the pre-feeding load and the number of parasites transmitted. We suggest that these females with high dose deliveries regurgitated parasites because of impaired stomodeal valve function [18]. This would be consistent with previous studies [19], [20] which have demonstrated an opened stomodeal valve due to the physical presence of a parasite plug and damage of the chitin layer of the valve by Leishmania chitinase.
Infected sand fly females, and especially those that transmit parasites, feed longer on hosts than non-transmitting ones do. Lutzomyia longipalpis females transmitting dermotropic CUK3 strain parasites took an average of 1.5 times longer to complete a bloodmeal compared to specimens infected but not transmitting, and 1.8 times longer than uninfected females. Similarly, P. perniciosus infected by CUK3 and IMT373 take 1.5 and 1.2 times more time for a blood meal. Most of the infected sand flies exposed to anaesthetized mice did not demonstrate increased probing, but rather remained feeding for longer periods until either they were fully or partially engorged. This is in agreement with data previously published on the L. longipalpis-L. mexicana combination [21].
Although only one dermotropic and one viscerotropic L. infantum strains were evaluated, the significant variation in inoculum size between them allow us to hypothetise that the infectious dose delivered by vector sand flies may be an inherent character of each Leishmania strain. Moreover, the infectious dose might be a determining factor in the outcome of Leishmania infection. Local cutaneous lesions might result from a high-dose inoculum of dermotropic Leishmania resulting in a strong local immune response, whereas dissemination to internal organs might be the result of infected sand flies delivering a low number of parasites below the threshold required to produce/develop a localized and restraining immune response. This hypothesis corresponds with the data of Kimblin et al. [14] on the L. major-P. duboscqi combination. These authors evaluated the impact of inoculum size on infection outcome by comparing L. major infections with high (5×103) and low (1×102) dose intradermally inoculated by needle in the ears of C57BL/6 mice, and observed the rapid development of large lesions in the ears of mice receiving the high-dose inoculum. In contrast, the low dose resulted in only minor pathology but a higher parasite titre during the chronic phase [14]. Nevertheless, it will be necessary to evaluate more L. infantum strains with visceral and cutaneous tropism in order to determine if differences detected in our study were due to individual stock characteristics or if they are associated with parasite tropism in vertebrate hosts.
In conclusion, we have demonstrated that individual sand flies transmit Leishmania parasites in a wide dose range. However, the maximal natural dose found was still about 250 times lower than the challenge dose used for the L. donovani complex in most previous experimental works. This finding emphasizes the importance of determining the natural infective dose for the development of an accurate experimental model, which is crucial for the evaluation of new drugs and vaccine candidates against leishmaniasis.
The viscerotropic Leishmania infantum strain IMT373 MON-1 (MCAN/PT/2005/IMT373) and the dermotropic L. infantum strain CUK3 (ITOB/TR/2005/CUK3) were used in this study. CUK3 was isolated from Phlebotomus tobbi from a Cukurova focus of cutaneous leishmaniasis [6] while IMT373 was isolated from a dog with leishmaniasis and passaged through mice in order to keep its virulence [13], [22]. Promastigotes (with less than 12 in vitro passages since isolation) were cultured at +26°C in M199 medium (Sigma, USA) containing 10% heat-inactivated foetal calf serum (Gibco, USA), 50 mg/ml mikacin solution (Bristol-Myers Squibb, Czech Republic) and 1% sterile urine.
Lutzomyia longipalpis (originating from Jacobina, Brazil) and Phlebotomus perniciosus (originating from Murcia, Spain) colonies were maintained in an insectary under standard conditions as described by Volf and Volfova [23]. Five to six-day old female flies (200 P. perniciosus and 150 L. longipalpis females per experiment, respectively) were fed on heat inactivated rabbit blood containing promastigotes (107 parasites per ml of blood) through a chicken-skin membrane. Blood-engorged females were separated immediately and maintained on a 50% sucrose diet in >70% relative humidity at +26°C.
One group of females was dissected to study the development and localization of infection in the sand fly midgut two and ten days post blood meal, i.e., during early and late stage infection, respectively. Individual midguts were placed into a drop of saline buffer, and parasite numbers were estimated under a light microscope at 200X and 400X magnifications by an experienced worker. Parasite loads were graded as previously described [24] into four categories: negative, 1–100, 100–1000, and >1000 parasites per gut. A second group of females from the same batch was used for transmission experiments and parasite quantification by Real-time PCR (see below). Nine and six independent experiments were performed with P. perniciosus-IMT373 and P. perniciosus-CUK3 combinations, respectively, while six and four artificial infections were done with L. longipalpis-IMT373 and L. longipalpis-CUK3 combinations.
One hundred and eight BALB/c mice (41 for P. perniciosus-IMT373, 28 mice for L. longipalpis-IMT373, 23 for P. perniciosus–CUK3 and 16 for L. longipalpis-CUK3 combinations) older than 8 weeks of age were purchased from AnLab (Czech Republic) and housed at Charles University, Prague, under stable climatic and dietary conditions. Experiments were approved by the institutional Ethical Committee and performed in accordance with national legislation for the care and use of animals for research purposes. Mice were anaesthetized intraperitoneally with ketamine (150 mg/kg) and xylazine (15 mg/kg).
Sand fly females were allowed to feed on whole body of anesthetised mice in a rectangular cage (20×20 cm) for about one hour at various days post infective blood meal (7–14 days for L. longipalpis and 9–23 days for P. perniciosus). Each mouse was placed individually into a cage together with about 50 P. perniciosus or 10 L. longipalpis females (the difference was due to the fact that L. longipalpis were more aggressive and had higher feeding rate). Two people followed each experiment; one recorded biting place and feeding time while the second ensured that each sand fly female probed in different place and then collected engorged flies by an aspirator immediately after terminating their blood meal; the site of bite and time of feeding were recorded for each female. After exposure, mice were sacrificed, biting place was inspected under a stereoscope and excised. Both samples (skin biopsies and corresponding fed sand flies) were stored at −20°C until DNA extraction.
Extraction of total DNA from each bite site and the corresponding sand fly were performed using a DNA tissue isolation kit (Roche Diagnostics, Germany) according to the manufacturer's instructions. DNA was eluted in 100 µl and stored at −20°C. qPCR for detection and quantification of Leishmania sp. was performed in a Rotor-Gene 2000 from Corbett Research (St. Neots, UK) using the SYBR Green detection method (iQ SYBR Green Supermix, Bio-Rad, Hercules, CA). For adequate sensitivity, kinetoplast DNA was chosen as the molecular target, with primers as previously described [25] (forward primer 5′-CTTTTCTGGTCCTCCGGGTAGG-3′ and reverse primer 5′-CCACCCGGCCCTATTTTACACCAA- 3′). Two microliters of eluted DNA was used per individual reaction. PCR amplifications were performed in duplicate wells using the conditions described previously [26]. Briefly, 3 min at 95°C followed by 45 cycles of: 10 s at 95°C, 10 s at 56°C, and 10 s at 72°C. Reaction specificities were checked for all samples by melting analysis. Quantitative results were expressed by interpolation with a standard curve included in each PCR run. Mass cultures of L. infantum promastigotes were used to construct a series of 10-fold dilutions ranging from 105 to 1 parasite per PCR reaction. Diluted parasites were co-processed with mouse tissue or sand fly females for DNA extraction. DNA from uninfected sand flies and mice were used as a negative control.
For sand fly females transmitting promastigotes into mouse skin, the pre-feeding midgut load was calculated as the sum of parasites in the midgut after feeding and the number of parasites transmitted.
Statistical analysis was performed using the software STATISTICA. For each L. infantum strain, a nonparametric Kruskal-Wallis test was used to compare: (i) the intensity of infection in P. perniciosus and L. longipalpis, and (ii) the number of parasites transmitted by each sand fly species into mice's skin. Correlations between feeding time and the number of parasites (i) in each sand fly female and (ii) inoculated into the skin, as well as the correlation between pre-feeding load and the number of parasites transmitted, were determined by simple linear regression analysis. Differences were considered statistically significant for p values <0.05.
|
10.1371/journal.pmed.1002417 | HIV-1 persistence following extremely early initiation of antiretroviral therapy (ART) during acute HIV-1 infection: An observational study | It is unknown if extremely early initiation of antiretroviral therapy (ART) may lead to long-term ART-free HIV remission or cure. As a result, we studied 2 individuals recruited from a pre-exposure prophylaxis (PrEP) program who started prophylactic ART an estimated 10 days (Participant A; 54-year-old male) and 12 days (Participant B; 31-year-old male) after infection with peak plasma HIV RNA of 220 copies/mL and 3,343 copies/mL, respectively. Extensive testing of blood and tissue for HIV persistence was performed, and PrEP Participant A underwent analytical treatment interruption (ATI) following 32 weeks of continuous ART.
Colorectal and lymph node tissues, bone marrow, cerebral spinal fluid (CSF), plasma, and very large numbers of peripheral blood mononuclear cells (PBMCs) were obtained longitudinally from both participants and were studied for HIV persistence in several laboratories using molecular and culture-based detection methods, including a murine viral outgrowth assay (mVOA). Both participants initiated PrEP with tenofovir/emtricitabine during very early Fiebig stage I (detectable plasma HIV-1 RNA, antibody negative) followed by 4-drug ART intensification. Following peak viral loads, both participants experienced full suppression of HIV-1 plasma viremia. Over the following 2 years, no further HIV could be detected in blood or tissue from PrEP Participant A despite extensive sampling from ileum, rectum, lymph nodes, bone marrow, CSF, circulating CD4+ T cell subsets, and plasma. No HIV was detected from tissues obtained from PrEP Participant B, but low-level HIV RNA or DNA was intermittently detected from various CD4+ T cell subsets. Over 500 million CD4+ T cells were assayed from both participants in a humanized mouse outgrowth assay. Three of 8 mice infused with CD4+ T cells from PrEP Participant B developed viremia (50 million input cells/surviving mouse), but only 1 of 10 mice infused with CD4+ T cells from PrEP Participant A (53 million input cells/mouse) experienced very low level viremia (201 copies/mL); sequence confirmation was unsuccessful. PrEP Participant A stopped ART and remained aviremic for 7.4 months, rebounding with HIV RNA of 36 copies/mL that rose to 59,805 copies/mL 6 days later. ART was restarted promptly. Rebound plasma HIV sequences were identical to those obtained during acute infection by single-genome sequencing. Mathematical modeling predicted that the latent reservoir size was approximately 200 cells prior to ATI and that only around 1% of individuals with a similar HIV burden may achieve lifelong ART-free remission. Furthermore, we observed that lymphocytes expressing the tumor marker CD30 increased in frequency weeks to months prior to detectable HIV-1 RNA in plasma. This study was limited by the small sample size, which was a result of the rarity of individuals presenting during hyperacute infection.
We report HIV relapse despite initiation of ART at one of the earliest stages of acute HIV infection possible. Near complete or complete loss of detectable HIV in blood and tissues did not lead to indefinite ART-free HIV remission. However, the small numbers of latently infected cells in individuals treated during hyperacute infection may be associated with prolonged ART-free remission.
| Early initiation of ART following infection may limit the total body burden of HIV.
It is not known if starting ART extremely early after HIV infection will lead to ART-free remission or cure.
We studied 2 individuals who started ART an estimated 10 and 12 days after HIV infection with very low peak viral load measurement; extensive testing of blood and tissue for HIV persistence was performed.
One participant stopped ART in order to test if and when HIV would rebound.
No HIV could be definitively detected for up to 2 years in the participant who initiated ART approximately 10 days after HIV infection.
Intermittent, very low levels of HIV were detected in blood but not tissue in the participant who initiated ART an estimated 12 days following infection.
The participant with no detectable HIV following ART experienced viral rebound 225 days after stopping ART.
HIV relapsed despite initiation of ART at one of the earliest stages of acute HIV infection possible.
Near complete loss of detectable HIV in blood and tissues did not lead to indefinite ART-free HIV remission.
| The development of a cure for HIV infection is a major public health objective [1]. Despite the ability of antiretroviral therapy (ART) to significantly reduce disease-related morbidity and mortality in HIV-1 infection, viral reservoirs persist indefinitely in latently infected cells [2]. HIV persists during ART primarily within circulating and tissue-resident, long-lived memory CD4+ T cells that harbor integrated HIV DNA; these cells are not cleared with ART and are a source of viral rebound when treatment is discontinued [3]. A major HIV eradication strategy involves aborting the initial seeding of these long-lived reservoirs by the very early initiation of ART [4,5]. For example, initiation of ART in a perinatally infected infant at 31 hours of life led to significant reductions in the viral reservoir and a significant time off ART (>2 years) before eventual viral recrudescence [6,7]. However, the impact of extremely early ART on HIV persistence and seeding the viral reservoir with the potential to prevent establishment of lifelong infection in adults is unknown.
Antiretroviral drugs initiated before HIV exposure (pre-exposure prophylaxis [PrEP]) can be an effective method of preventing HIV acquisition [8–11]. PrEP programs involve HIV antibody testing before the initiation of prophylactic ART in individuals at high risk for acquiring HIV. PrEP is typically started following negative HIV antibody or combined antibody/antigen screening in the absence of clinical symptoms [8,11]. Because there is a delay between HIV infection and when an HIV antibody or combined antibody/antigen test is reactive, PrEP may be unknowingly started in an individual who has very recently been infected with HIV. As a result, a small number of individuals may begin 2-drug ART just prior to or after the development of detectable plasma HIV-1 RNA (the transition from the "eclipse phase" to Fiebig stage I of infection) and prior to the detection of HIV antigen or antibody [12,13]. PrEP programs are therefore ideal settings in which to identify individuals treated extremely early during infection for the longitudinal study of HIV-1 reservoir persistence in blood and various tissues [4,5,14,15].
The aims of this study were to determine the impact of extremely early initiation of ART on the size of the HIV reservoir in blood and various tissues and the potential for long-term ART-free remission. As a result, we studied 2 PrEP study participants who initiated ART during emergent, unrecognized HIV infection in Fiebig stage I, with 1 individual treated just as he was transitioning out of the “eclipse phase” of HIV infection [12,13]. We describe the results of extensive tissue and blood sampling in these individuals and the result of a highly monitored treatment interruption. Using this case and our previously described recipient of an allogeneic bone marrow transplant (hematopoietic stem cell transplantation [HSCT] Participant B) [16,17], we also describe potential biomarkers for HIV reactivation during prolonged states of viremia post-ART interruption.
The PrEP Demo Project was a prospective study of PrEP for men who have sex with men (MSM) in which participants were tested for HIV both by HIV antibody/antigen combination assay and by HIV RNA on the day of PrEP initiation [18]. There was no prespecified plan for the present analysis at that time. Two participants were identified in the study who had positive viral load tests performed on the day of initiation of PrEP (truvada/emtricitabine). PrEP was converted to standard ART once HIV infection was confirmed. The participants were enrolled in the longitudinal University of California San Francisco (UCSF) SCOPE study after providing written informed consent. The study was approved by the UCSF Committee on Human Research. A protocol and an analysis plan were in place prior to analytical treatment interruption (ATI). The protocol and STROBE checklist are included in the supporting information (S1 Protocol and S1 STROBE Checklist).
To more fully explore biomarkers associated with HIV reactivation after the interruption of ART, we also accessed peripheral blood mononuclear cells (PBMCs) from a previously described case of an HIV-infected adult who underwent an allogenic hematopoietic stem-cell transplantation (HSCT Participant B) [16]. All HIV reservoir studies for HSCT Participant B during ATI were conducted at commercial laboratories and cryopreserved PBMCs were not available.
PBMCs and plasma were collected longitudinally either by large-volume peripheral blood draws or by leukapheresis and purified by Ficoll-Hypaque (Sigma-Aldrich) density gradient centrifugation. Colorectal and ileal tissue were collected and processed as previously described [19,20]. A whole, inguinal lymph node was obtained from each participant by surgical excision for mononuclear cell isolation and downstream HIV reservoir characterization months following combination ART initiation. Total PBMCs, purified CD4+ T cells, or CD4+ T cell subsets [naive (TN), central memory (TCM), transitional memory (TTM), and effector memory (TEM)] from blood and tissues were collected via bead-based separation (Stem Cell Technologies) or by fluorescence-activated cell sorting using previously described methods [21]. Bone marrow biopsies were performed, followed by sorting and collection of 4 cell populations: myeloid cells (CD33+Lin+/−), CD3+CD4+ T cells (CD33−Lin+CD4+), multilineage CD4+ cells (CD33−Lin−CD4dim), and Lin− cells that are not CD4+ or myeloid cells (CD33−Lin−CD4−CD34+/−). Cerebrospinal fluid (CSF) was collected by lumbar puncture and centrifuged to separate the liquid and cellular fraction for downstream HIV-1 detection, and quantification was carried out as previously described [22,23].
HIV-1 RNA was isolated from baseline (pre-PrEP) and subsequent plasma samples followed by single-genome sequencing (SGS) of HIV-1 protease-pol (pro-pol) or a 1.5 kb portion of the envelope gene as previously described [24,25]. Population sequencing on subsequent timepoints was performed using TRUGENE (Siemens, Tarrytown NY). PBMCs and tissue-derived cells were analyzed for HIV-1 persistence using a variety of highly sensitive assays in up to 10 independent laboratories located at the UCSF, University of Montreal, University of Pittsburgh, University of California San Diego (UCSD), and Johns Hopkins University. Assays utilized were previously described and included highly sensitive quantitative PCR for total and integrated HIV-1 DNA (unspliced and total RNA) [26,27], droplet digital PCR (ddPCR) (HIV pol DNA and 2-LTR circles) [28], total virus recovery assay, quantitative viral outgrowth assay (qVOA) [29,30], Tat/Rev Induced Limiting Dilution Assay (TILDA) [31], and whole HIV genome sequencing [32]. In addition, large volume plasma was tested using the MEGA iSCA [33].
CD4+ T cells obtained by leukapheresis from both participants approximately 18 months following the initiation of ART were tested using the murine (humanized mouse) viral outgrowth assay (mVOA) as previously described [34]. Briefly, 530 million and 379 million CD4+ T cells were divided and injected intraperitoneally into 8–10 humanized female young adult NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ (NSG) mice from Jackson Labs (53 million and 50 million cells per mouse from the 2 participants). Plasma HIV-1 RNA testing was then performed following up to 5.5 weeks after engraftment on blood obtained by weekly mandibular sinus bleed not exceeding 0.5% of body weight. Mice were euthanized by CO2 inhalation and HIV-1 sequencing was attempted from plasma and spleen cells using cDNA and methods optimized for low HIV-1 RNA copy numbers [34]. The Johns Hopkins University Institutional Animal Care and Use Committee approved this research and it was conducted in accordance with the 8th edition of the Guide for the Care and Use of Laboratory Animals within fully AAALACi accredited animal facilities. Mice were group-housed with other mice xenografted from the same donor in microisolator caging (Allentown) and fed a commercial rodent chow (Harlan) and hyperchlorinated water ad libitum.
A carefully monitored ATI was performed on 1 participant (PrEP Participant A) who initiated PrEP an estimated 10 days following HIV infection and had no definitive HIV detected in blood or tissues for 32 months on ART. After ATI, viral load testing was performed twice weekly using the COBAS AmpliPrep/COBAS TaqMan V.2 assay (Roche) for the first 3 months, followed by weekly testing thereafter with close clinical observation. Large-volume blood draws were also obtained monthly for the isolation of PBMCs, and plasma was obtained for further reservoir and flow cytometric testing. ART was initiated immediately after confirmation of detectable plasma HIV-1 RNA.
Flow cytometric testing was performed on cryopreserved PBMCs isolated longitudinally just before, during, and after ATI for PrEP Participant A and HSCT Participant B. Thawed PBMCs were stained with LIVE/DEAD Fixable Aqua Dead Cell Stain Kit (ThermoFisher Scientific), Brilliant Violet 711-conjugated anti-CD4 (SK3) (BD Biosciences), Brilliant Violet 650-conjugated anti-CD3 (SP34-2) (BD Biosciences), allophycocyanin (APC)-conjugated anti-CD38 (HB7) (BD Bioscience), PE-conjugated anti-CD30 (BERH8) (BD Biosciences), APC-H7-conjugated anti-HLA-DR (L243) (BD Bioscience), CD8 Alexa Fluor 700-conjugated anti-CD8 (RPA-T8) (BD Bioscience), BV786-conjugated anti-CD16 (3G8) (BD Bioscience), PE-Cy7-conjugated anti-CD107a (H4A3) (BD Bioscience), FITC-conjugated anti-CD56 (NCAM 16.2) (BD Bioscience), and PerCP-conjugated anti-CD69 (L78) (BD Bioscience). Cells were then analyzed on a BD LSRII flow cytometer (BD Biosciences). Single stained beads (Life Technologies) were used for compensation, and fluorescence minus one (FMO) controls were used to set gates. Data for phenotyping were acquired on all events and analyzed in FlowJo V10 (TreeStar). Examples of gating strategies are shown in S1 Fig.
CD4+ and CD8+ T cell phenotypes were further characterized on live cells for PrEP Participant B by flow cytometry after exclusion of dead cells (Fixable Aqua dye, Molecular Probes) using the following fluorescently labeled antibodies: CD3 BV650 clone SK7, CD4 BV711 clone OKT4, CCR7 BV785 clone G043H7, CD45RA APC-Cy7 clone H100, Tbet PE clone 4B10, Eomesodermin PE-e610, PD-1 BV421 or BV605 clone EH12.2H7, CD160 PE-Cy7 clone BY55, Ki-67 FITC clone Ki-67, Granzyme B FITC clone GB11, Perforin PE-Cy7 clone B-D48 (all Biolegend), and CD8α APC-R700 clone RPA-T8 (BD Biosciences). Positive gates for activation markers and effector cell transcription factors were drawn based on expression of these markers in peripheral blood naive CD8+ T cells isolated from an HIV-uninfected donor. To limit experimental variability, flow cytometry was performed on the same day.
After thawing, PBMCs were incubated at 37°C overnight at a concentration of 2 × 106 cells/mL in RPMI medium containing 10% FBS and penicillin/streptomycin. The next day, 1 × 106 cells were stimulated with Gag pooled peptides (final concentration 1 μg/mL in DMSO provided by the NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH: HIV-1 Con B Gag Peptide Pool cat #12425), for 6 hours in the presence of brefeldin A and monensin (BD). The percentage of CD8+ or CD4+ T cells producing interferon gamma (PE-dazzle, clone 4S.B3, Biolegend), tumor necrosis factor alpha (Alexa fluor 700, clone mAb11, eBioscience) or interleukin-2 (BV785, clone MQ1-17H12, Biolegend), or degranulation as detected by CD107a expression (APC, clone H4A3, Biolegend), was measured by flow cytometry, and the frequency of positive cells was determined after subtraction of the frequency measured in wells incubated with DMSO alone. Examples of gating strategies are shown in S2 Fig.
Graphical analyses were performed using Prism v.6 (GraphPad software). Mathematical modeling of reservoir size, rebound rate, and probability of cure were performed using our previously described methods [35]. A summary of these methods and the input variables used is provided in S1 Text.
Participant A, a 54-year-old male, was HIV-uninfected at 2 PrEP pre-enrollment visits but continued to have ongoing sexual risk for HIV infection. He then initiated tenofovir/emtricitabine PrEP and usage was confirmed by testing drug levels. Seven days after PrEP initiation, results from the baseline (day 0 of PrEP) visit revealed a plasma HIV RNA level of 220 copies/mL (Abbott RealTime HIV-1, lower limit of detection [LLD] <40 copies/mL), and 69 copies/mL by the Single-Copy Assay (LLD 1 copy/mL); 4th generation EIA (Abbott) and rapid HIV-antibody (Clearview HIV 1/2 Stat-Pak) tests were negative. Based upon these results (and a negative pooled HIV RNA, 4th generation EIA, and rapid HIV-antibody tests at 2 pre-enrollment visits 21 and 13 days prior), it was determined that he had been in the transition from the eclipse phase to Fiebig stage I HIV infection (estimated 10 days after HIV infection [12,13]) at the time of initiating PrEP. PrEP was substituted with a 4-drug ART regimen (darunavir/ritonavir/raltegravir/tenofovir/emtricitabine) 7 days after the initiation of PrEP. This ART regimen was chosen due to concern about the potential for the development of drug resistance to tenofovir/emtricitabine. The participant was asymptomatic at the time. HIV western blot assays (BioRad) were repeatedly indeterminate (p55 only) and became nonreactive at an estimated 130 days after time of infection. SGS from the plasma sample from the date when PrEP was initiated confirmed that the individual was infected with subtype B virus without known drug resistance mutations.
Plasma HIV RNA levels subsequently declined following PrEP and combination ART initiation: 220 copies/mL (PrEP baseline), 120 copies/mL (7 days after initiation of PrEP), and <40 copies/mL but detectable (estimated 22 days after starting PrEP). All subsequent plasma HIV RNA levels were undetectable. Low-level cell-associated HIV RNA (3.2 copies/million CD4+ T cells) was detected 32 days after infection. However, sorted rectal CD4+ T cells were negative for HIV RNA and DNA (collected 1.9 months after infection), and leukapheresis-collected PBMCs enriched for total CD4+ T cells and sorted CD4+ T cell subsets (TN, TCM, TTM, and TEM) were negative for cellular HIV RNA, total HIV DNA (confirmed in 2 independent laboratories), integrated HIV DNA, and 2-LTR circles (collected 2.1 months after infection). Over the following 2 years, no further HIV (nucleic acid, viral outgrowth, total virus recovery, or whole genome sequences) could be detected, despite sampling from ileum, rectum, lymph nodes, bone marrow, CSF, and circulating CD4+ T cell subsets. A detailed timeline of clinical viral load measurements and results of longitudinal HIV-1 reservoir assays are shown in Fig 1 and Table 1. Chemokine receptor 5 (CCR5) genotyping revealed that the individual did not carry any CCR5 Δ32 mutations, and HLA typing revealed that he was HLAB5701 negative. We estimated that greater than 1 billion CD4+ T cells were eventually examined without any evidence of HIV infection.
PrEP Participant B, a 31-year-old male, was HIV-uninfected at a pre-enrollment visit 6 weeks before initiating tenofovir/emtricitabine PrEP but had ongoing sexual risk for HIV infection. On the date of initiation of PrEP his Clearview HIV 1/2 Stat-Pak, HIV-1 CMIA (Abbott Ag/Ab combo assay), HIV-1 IFA, western blot, and HIV 1/2 MultiSpot Rapid Test were all non-reactive. Six days after initiation of PrEP, results from the baseline (day 0 of PrEP) visit revealed a plasma HIV of 359 copies/mL and PrEP was discontinued. Based on these results and clinical information, it was determined that he was infected approximately 12 days prior to starting PrEP (Fiebig stage I). Eight days after starting PrEP, the plasma HIV-1 RNA level increased to 668 copies/mL. HIV genotyping was positive for the M184M/I resistance mutation (no mutation was detected by genotyping on PrEP day 0), and 3,343 copies/mL were measured 9 days after the initiation of PrEP. He started combination ART (darunavir/ritonavir/raltegravir/tenofovir/emtricitabine) 12 days after starting PrEP, with subsequent decline of HIV plasma RNA to 351 copies/mL and 39 copies/mL on post-PrEP days 16 and 44, respectively. All subsequent plasma HIV RNA levels were undetectable starting at post-infection day 91. He changed ART to tenofovir/emtricitabine/dolutegravir/rilpiverine 56 days after starting PrEP and then to abacavir/lamivudine/dolutegravir 29 months following initial PrEP (Fig 1). HIV antibody testing remained negative through day 219 after the initiation of PrEP, but a western blot was indeterminate (p24 weekly positive, all other bands negative) on day 258.
Three months following infection, Participant B had very low level detectable HIV-1 RNA (2.9 copies/106 cell) and DNA (14.3 copies/106 cells) in TTM and TEM CD4+ T cell subsets, respectively, and again 2.4 and 5.5 HIV RNA copies/106 total CD4+ T cells approximately 9 months after infection. No HIV could be detected from CSF, bone marrow cells or CD4+ T cell subsets from a whole excised inguinal lymph node 8–9 months following infection. No definitive HIV outgrowth could be detected with a qVOA performed 32 months after infection utilizing 20 million input total CD4+ T cells. Very low levels of HIV-1 RNA (<50 copies/mL) were detected from several of the viral outgrowth assay (VOA) wells, but no increases were observed in RNA over time, and no bulk or single genome HIV-1 sequences could be obtained (see Fig 1 and Table 2).
Given the very low level or lack of detectable HIV-1 from large numbers of cells from various tissues, large quantity plasma, or CSF, we performed a leukapheresis on both participants and used large numbers of CD4+ T cells for input into a previously reported mVOA [34]. We obtained 530 million total CD4+ T cells from Participant A approximately 18 months following infection on ART. One of 10 mice given 53 million input cells developed low-level HIV RNA in plasma (201 copies/mL) following in vivo activation with anti-CD3 antibody approximately 5.5 weeks after engraftment, but HIV was not detected in spleen tissue (<1 million spleen cells present). However, HIV-1 RNA and cDNA sequencing at an independent laboratory was unsuccessful from plasma and spleen cells, and true infection in cells from Participant A could not be verified. We obtained 500 million total CD4+ T cells from Participant B. Three of 8 mice developed detectable HIV RNA starting approximately 4.5 weeks following cell transfer (1,000; 5,000; and 11,000 copies/mL, respectively, from a total of 50 million CD4+ T cells per mouse).
Following 34 months of continuous ART, Participant A provided informed consent to undergo a carefully monitored ATI (Fig 2). The patient remained clinically well with undetectable plasma viral load measurements through 225 days of observation post-interruption. On day 225, he had a detectable, low-level plasma HIV-1 RNA of 36 copies/mL. Repeat testing 6 days later (day 231) confirmed rebound with virus increasing to 77,397 copies/mL. He initiated therapy (tenofovir/emtricitabine/ritonavir/darunavir/dolutegravir) on post-ATI day 231 prior to receiving the confirmatory test. He then had viral loads of 158 and 19 copies RNA/mL on days 245 and 252, respectively. Subsequent plasma viral load testing revealed no detectable HIV RNA, and he has remained clinically well. Single genome sequences of a 1,190 base-pair region of HIV-1 Pol were monoclonal and identical to the PrEP baseline sequences. The level of intra-sequence diversity was extremely low at both time points (<0.02%) but distinct from the consensus ancestral subtype B sequence. HIV-1 envelope single genome (1,557 base-pair region) analysis at the time of recrudescence revealed 2 unique but highly related sequences; HIV envelope sequences were not obtained at PrEP baseline (S3 Fig).
We used a collection of mathematical modeling approaches (see S1 Text) to better understand the dynamics of reservoir seeding and rebound in these participants. For PrEP Participant A, the estimated latent reservoir size immediately before treatment interruption, based only on the observed time of rebound (226 days) [36] was 0.0020 [0.00045, 0.0063] infectious units per million cells (IUPM) (brackets give 95% credible intervals). With this estimated reservoir size (including uncertainty), about 1% of identical individuals (i.e., having similar restrictions in reservoir size) would be expected to achieve lifelong (>70 year) ART-free HIV remission. These results are consistent with estimates of reservoir size from various assays described in Fig 1 and Tables 1 and 2. Based on the viral load level at initial diagnosis and a calibrated model of reservoir seeding during acute infection [37], we estimate that the reservoir size prior to the initiation of ART was 0.02 IUPM, with about 1 log uncertainty in either direction. qVOA results on day 166 suggest that there is a 95% probability that the reservoir size is below 0.075 IUPM [38]. If we assume the positive outgrowth in the mVOA is a true positive signal, then there is a 95% probability that the reservoir size is below 0.0057 IUPM. If we assume the outgrowth was not real, then the central estimate for the reservoir size is 0.0020 IUPM (95% credible interval 0.00028–0.014 IUPM). Assuming there are around 1011 total body CD4+ T cells [39,40], these estimates collectively suggest there were only approximately a few hundred cells infected with replication-competent HIV provirus prior to treatment interruption. Separately, we estimated the exponential growth rate of viral load during rebound to be 1.3/day, very similar to that seen in the HSCT Boston participants [36] and near the central value seen in acute infection [41–43], but much higher than that seen in rebound following chronic infection (excluding the prolonged time off ART prior to first detection of HIV-1) [44–46].
Flow cytometric characterization of surface markers of T and natural killer (NK) cell activation was performed on samples from PrEP Participant A before and during ATI and following HIV rebound in order to identify potential predictors of viral recrudescence. In addition, pre- and post-ATI samples were investigated from HSCT Participant B, an individual who lost detectable HIV-1 in blood and gut tissue following allogeneic HSCT for malignancy. Similar to PrEP Participant A, the HSCT participant experienced confirmed HIV rebound 225 days after interrupting ART and had similar exponential growth rate during recrudescence [16]. Interestingly, surface expression of CD30, a member of the tumor necrosis factor (TNF) super-receptor family and lymphoma tumor marker, increased on CD4+ and CD8+ T cells months prior to detectable plasma HIV in both of these participants (Fig 3). The frequency of CD30+ and CD69-expressing cells also increased on CD4+ and CD8+ T cells and CD56+ and CD16+ NK cells prior to viral recrudescence in PrEP Participant A. Overall, increases in the frequency of CD30-expressing cells appeared to be larger than those expressing CD69. However, no distinct patterns in lymphocyte HLA-DR/CD38 were observed in either participant, and only CD30 expression increased prior to rebound in the HSCT recipient.
Sufficient cells were available for further immunological phenotyping from PrEP Participant A, and we identified an increase in the frequency of eomesodermin expressing CD8+ T cells prior to viral rebound. However, no patterns emerged during ATI for expression of PD1, TIGIT, or Ki67. There were no changes in CD8 responses in CD107a expression, and no significant intracellular INF-gamma, TNF-alpha, or IL1 production was detected in response to overlapping HIV-1 gag peptide pools prior to, during, or following ATI.
We report 2 cases of extremely early HIV diagnosis and initiation of ART at the threshold when plasma viremia begins to expand exponentially (the end of the so-called “eclipse phase” and beginning of Fiebig stage I). These stages of HIV infection precede the time when acute HIV infection becomes clinically apparent and are theoretically the earliest time when ART can be initiated in an adult [12,13]. To the best of our knowledge, PrEP Participant A is the earliest documented case of adult HIV infection followed by immediate initiation of ART with the exception of successful post-exposure ART prophylaxis, and it would be very challenging to initiate therapy any earlier. Despite the complete or near-complete loss of detectable HIV in blood and a variety of tissues, HIV rebounded in this person 225 days following cessation of ART. Although identifying hyperacute infection is rare, our group has previously reported that 15.6% of patients referred to our HIV clinic for newly diagnosed infection had detectable plasma HIV-1 RNA but negative HIV-specific antibody test results [47]. These data include individuals taking part in a rapid treatment initiation study and may overestimate the incidence of acute infection identified during Fiebig stage I. Excluding participants in this study, 6.3% of individuals presented with positive plasma HIV-1 RNA and no detectable HIV-specific antibodies at the time of diagnosis [47].
The delayed timing of viral rebound in PrEP Participant A was similar to that observed in an HIV-infected individual who lost detectable HIV in blood and tissue following allogeneic HSCT (Boston Participant B). In each case, modeling predicted a residual HIV replication-competent reservoir of perhaps hundreds of infected CD4+ T cells throughout the body, which explains the lack of detectable HIV from blood or tissues despite massive sampling. Overall, allogeneic HSCT and extremely early ART initiation appear to have similar long-term effects on reducing both the residual HIV reservoir and immune responses. Nonetheless, small numbers of latently infected cells likely persisted in these individuals and became activated, leading to HIV rebound in the absence of ART. Modeling of these types of participants leads to wide confidence intervals, which are likely due to the varying size of the reservoir in different patients, uncertainty around model parameters, and the stochastic nature of reactivation of latently infected cells in vivo. It is also possible that a non-CD4+ T cell source of HIV persistence contributed the viral rebound; a majority of cells tested for persistence in blood and tissue were CD4+ lymphocytes. Of note, the median time to viral rebound in 8 Thai individuals treated with ART during later phases of Fiebig stage I was recently reported to be 26 days [48]. Although sample size is limited, these and our data suggest that differences in initiation of ART time of just a few days during Fiebig stage I infection may have a noticeable impact on the duration of ART-free remission.
After this delay, the rapid HIV rebound dynamics in PrEP Participant A were again similar to those observed in HSCT Participant B [16], consistent with rapid exponential growth seen with primary infection. Unlike the HSCT participant, however, the PrEP recipient was asymptomatic and the peak viral load may have been mitigated by the earlier reinitiation of ART. The observed rapid rebound kinetics also differ from those in individuals who achieved post-treatment HIV-1 control following early ART initiation [49].The rapid rebound kinetics and lack of post-treatment control observed in this study are likely secondary to the lack of HIV-specific immunity since exponential growth is lower in the setting of withdrawal of ART initiated during chronic infection when HIV-specific immunity is present. Concomitant immune-modifying therapies may be necessary in order to achieve ART-free HIV control in very-early treated individuals.
PrEP Participant B initiated ART later than PrEP Participant A and had higher levels of viremia at his baseline visit. In many ways, he is similar to the “Mississippi baby” [6,7] in that he started ART with moderate levels of viremia and subsequently had a difficult-to-detect reservoir. Individuals who initiate therapy during the earliest stages of infection will almost certainly need other “curative” interventions before ART might be interrupted with the expectation that a viral relapse will be avoided. It is also possible that higher peak plasma HIV RNA levels and subsequent low-level viral reservoir detection in samples from PrEP Participant B were a result of the appearance of an emtricitabine-resistant associated mutation and exposure to a single active antiretroviral drug during initial PrEP.
It is likely that there are rare individuals who start PrEP during the true eclipse phase, when HIV has the potential to establish a long-lived reservoir but before the development of detectable HIV in blood or tissues. Detecting a potential “cure” in such a setting will be challenging but such efforts are ongoing. Of note, non-human primate studies have demonstrated that, while ART initiated within 48 hours of SIV challenge is able to prevent infection [50–52], ART started 3 days following SIV infection leads to viral rebound following cessation of therapy [53]. As a result, the "curative window" between infection and the potential to abort infection after exposure is likely very small. It is also possible that there may be individuals infected just before the start of PrEP, but in whom infection would not be known for some time until they stop ART or until a 2-drug regimen is unable to suppress the virus.
Our study is one of the first to incorporate an mVOA to measure potential HIV persistence following an intervention that leads to loss of detectable HIV in blood and various tissues. One mouse became viremic (at low levels) and 3 mice became viremic following receipt of cells from PrEP Participants A and B, respectively. This positive finding with PrEP Participant A's PBMCs may be the only evidence that this individual had persistent HIV prior to ART interruption. Unfortunately, sequence verification of plasma and splenic HIV in mice from both PrEP participants could not confirm definitive viral outgrowth from participant cells. Sample volumes are limited in murine models of HIV infection and were exhausted during testing. Nonetheless, 2 published studies suggest that mVOAs may be more sensitive than traditional outgrowth assays to detect replication-competent HIV in individuals with undetectable HIV-1 DNA or RNA by traditional means [34,54]. If this is the case, our study suggests that sampling of hundreds of millions of PBMCs may, at times, be more sensitive than tissue-based studies for the detection of residual HIV infection since a much larger number of cells can be interrogated. Further studies comparing mVOAs with traditional ex vivo co-culture assays utilizing rigorous positive and negative controls are certainly warranted.
A major emphasis of HIV curative science has been to identify potential markers or correlates of HIV rebound before or after treatment cessation. CD30, a member of the TNF receptor superfamily, is expressed on very small percentages of lymphocytes and myeloid cells but is dramatically upregulated on Hodgkin and other lymphoma cells [55–59]. CD30 has been implicated in the activation, proliferation, and cell death of selected cell populations [55,57,60,61]; and infections with human T-cell lymphotropic virus (HTLV), Epstein-Barr virus (EBV), and poxviruses can lead to increases in CD30 expression [57,62,63]. In addition, increases in the plasma concentration of the soluble form of CD30 have been associated with HIV disease progression prior to ART initiation [57,58,60,64–71]. Although anecdotal, our data in PrEP Participant A and the HSCT participant suggest that upregulation of CD30 may occur prior to detectable plasma HIV-1 RNA. Furthermore, given a similar but somewhat less pronounced pre-rebound expression pattern of CD69 in the PrEP Participant A, it is possible that these markers represent more global lymphocyte activation, proliferation, or stress responses that are detectable prior to viral recrudescence. Further investigation of CD30 and related cell-surface markers as potential predictors of HIV rebound is urgently needed.
These cases also suggest that PrEP programs should test individuals for HIV using an HIV test with the narrowest window period possible, just before initiating PrEP. If cost permits and in settings with a high incidence of acute HIV (e.g., STI clinics), it is reasonable to test with a plasma HIV RNA before PrEP initiation and when reinitiating PrEP after an interruption. Fourth generation combination p24 antigen/antibody tests will identify most patients with acute HIV but miss those in the earliest Fiebig stages, such as those presented in this analysis. Plasma HIV RNA testing should also be considered when reinitiating PrEP after interruption [9]. PrEP programs should recommend an immediate switch to conventional ART if an individual is found to be newly HIV-positive. Potential benefits of such programmatic changes include (1) lower rates of missed acute HIV diagnoses, (2) decreased acquired drug resistance to tenofovir/emtricitabine, (3) individual clinical benefit [72], and (4) fewer subsequent transmission events [73].
This study was limited by its small sample size. The identification of individuals treated during hyperacute infection is rare given the rapid increase in plasma HIV-1 RNA during the early phase of Fiebig stage I infection and the fact that hyperacute infection may be asymptomatic. The limited number of individuals included in the analysis make it difficult to draw definitive conclusions about the impact of very early ART on restricting HIV reservoir size or prolonging ART-free remission. In addition, a larger number of individuals are required to validate potential cell-surface markers of HIV persistence or predictors of HIV rebound. Despite the challenges with including very early treated individuals in prospective studies, these cases provide valuable information as to the rapidity of seeding of the HIV reservoir and the impact of extremely early ART on HIV persistence.
In summary, we report 2 cases of extremely early initiation of prophylactic ART immediately following the “eclipse phase” in Fiebig stage I (at approximately 10 days of HIV infection). A very small residual HIV reservoir size was observed in these participants who started very early PrEP and subsequently converted to full ART. In 1 individual, we observed a prolonged period of ART-free remission similar in duration to the allogeneic HSCT Boston Participant B. However, although HIV persisted indefinitely in both of these PrEP cases, a continuum may exist across PrEP, post-exposure prophylaxis and curative early ART strategies. Further investigation in larger cohorts of individuals treated extremely early following HIV infection is warranted.
|
10.1371/journal.pcbi.1006800 | Changes in morphogen kinetics and pollen grain size are potential mechanisms of aberrant pollen aperture patterning in previously observed and novel mutants of Arabidopsis thaliana | Pollen provides an excellent system to study pattern formation at the single-cell level. Pollen surface is covered by the pollen wall exine, whose deposition is excluded from certain surface areas, the apertures, which vary between the species in their numbers, positions, and morphology. What determines aperture patterns is not understood. Arabidopsis thaliana normally develops three apertures, equally spaced along the pollen equator. However, Arabidopsis mutants whose pollen has higher ploidy and larger volume develop four or more apertures. To explore possible mechanisms responsible for aperture patterning, we developed a mathematical model based on the Gierer-Meinhardt system of equations. This model was able to recapitulate aperture patterns observed in the wild-type and higher-ploidy pollen. We then used this model to further explore geometric and kinetic factors that may influence aperture patterns and found that pollen size, as well as certain kinetic parameters, like diffusion and decay of morphogens, could play a role in formation of aperture patterns. In conjunction with mathematical modeling, we also performed a forward genetic screen in Arabidopsis and discovered two mutants with aperture patterns that had not been previously observed in this species but were predicted by our model. The macaron mutant develops a single ring-like aperture, matching the unusual ring-like pattern produced by the model. The doughnut mutant forms two pore-like apertures at the poles of the pollen grain. Further tests on these novel mutants, motivated by the modeling results, suggested the existence of an area of inhibition around apertures that prevents formation of additional apertures in their vicinity. This work demonstrates the ability of the theoretical model to help focus experimental efforts and to provide fundamental insights into an important biological process.
| Pollen is renowned for its ability to form beautiful and complex patterns on its surface. One of the most prominent patterns on the pollen surface is formed by apertures, the regions that lack deposition of the pollen wall exine and develop at precise locations which often vary between the species. How aperture patterns are created is an intriguing and poorly understood question. We developed a mathematical model that aims to explore the mechanisms responsible for the aperture patterning in the pollen of the model plant Arabidopsis. Our model showed that size of the pollen grain could be solely responsible for the increase in aperture number observed in the pollen of some Arabidopsis mutants. Additionally, kinetic parameters, such as diffusion and decay of aperture factors, could also influence aperture number. We coupled our mathematical modeling with a forward genetic screen of a mutagenized population of Arabidopsis. This screen discovered novel mutants with aperture patterns that had been predicted by our mathematical model. Further experiments on these mutants provided additional support to the modeling predictions. These results demonstrate that mathematical modeling could be a powerful tool for understanding the mechanisms responsible for patterning of pollen grains.
| The process of cell morphogenesis often depends on the ability of cells to form distinct domains of plasma membrane and precisely target deposition of extracellular materials. Pollen presents a powerful model to study the mechanisms that control formation of membrane domains and localization of extracellular structures. Pollen grains are surrounded by a complex extracellular structure, exine, that can assemble into thousands of elaborate, species-specific patterns, which make the pollen surface one of the most diverse structures found in nature [1–3]. These patterns are formed by precisely depositing exine at certain areas of the pollen surface and by preventing or reducing its deposition at other sites.
In most plant species the restriction of exine deposition at particular locations leads to the development of some of the most obvious patterning elements on the pollen surface. These characteristic areas which either lack exine completely or have decreased amounts of exine are called apertures [2, 4, 5] (Fig 1). Apertures are critical for pollen viability and function, as they often serve as portals through which pollen tubes exit during germination and as architectural details that help the pollen accommodate volume changes in response to changing hydration levels [5–9].
The patterns produced by apertures on the pollen surface are one of the major taxonomic features used for classification of flowering plants. The two major clades of flowering plants, eudicots and monocots, are each characterized by a prototypical aperture pattern, defined by aperture number and position. Monocots tend to have a single, polarly localized aperture, whereas eudicots most commonly have three apertures, equidistantly distributed like three meridians around the pollen equator. Although these are the most prevalent patterns for these clades, individual species exhibit wide variations on these common themes, with dramatic differences in aperture number, position, and morphology. Among the observed patterns are pore-like single apertures in grasses, two hole-like apertures at the opposite poles in many species of bromeliads, three equatorial apertures in tomato, four equatorial apertures in some species of tobacco, six apertures in species of mint and passion flower, multiple hole-like apertures in hibiscus and phlox, and many other possible patterns [1, 2, 4, 5]. In the cases when pollen has more than one aperture, these apertures (or their centers) tend to be equally distributed on the pollen surface or around the pollen equator. Yet, the molecular and cellular mechanisms that restrict exine deposition at these specific sites and contribute to the formation of aperture patterns are not understood in any species.
In the model plant Arabidopsis thaliana, like in many other eudicot species, pollen has three long and narrow apertures placed equidistantly around the equator of the pollen grains (Fig 1A, 1A’ and 1G). This patterning is very precise: essentially all pollen grains in the wild-type Arabidopsis develop this pattern. Formation of Arabidopsis apertures begins after male meiosis, when four meiotic products—the pollen precursors known as the microspores—are held together in a tetrad arrangement, surrounded and separated from each other by a callose wall [10–12] (Fig 1G).
Until recently, only a single molecular player, the product of the Arabidopsis INAPERTURATE POLLEN1 (INP1) gene, was known to act as an aperture-promoting factor. During the tetrad stage INP1 protein specifically aggregates at the areas of the microspore plasma membrane that will become apertures, while in the absence of functional INP1 apertures do not form [12, 13]. Although INP1 is likely a late-acting factor which does not by itself define the aperture pattern [14, 15], its specific localization to certain membrane sites suggests that the plasma membrane in microspores acquires polarity and forms distinct domains, which then become protected from exine deposition. Recently, we identified a second aperture factor in Arabidopsis, the protein kinase D6 PROTEIN KINASE-LIKE3 (D6PKL3), which appears to act upstream of INP1 and helps attract it to the aperture domains [16]. However, the mechanism by which these proteins select domains for aperture sites is not understood.
Previously, we demonstrated that the mechanism of aperture formation is sensitive to pollen ploidy or to factors tightly linked to ploidy, such as pollen size [13]. Compared to the normal haploid (1n) pollen grains in Arabidopsis with three equatorial apertures, the larger diploid (2n) pollen grains generated by several mechanisms (e.g. through a tetrad or dyad stage) commonly develop either four equatorial apertures or six apertures distributed along the edges of a tetrahedron (Fig 1B–1D’ and 1G), although some other patterns are also observed [13]. In turn, pollen of even higher ploidy (3n or 4n) develops complex aperture patterns with the larger number of apertures that often coalesce into ring-shaped structures (Fig 1E–1G) [13]. INP1 localization in higher-ploidy microspores recapitulates the changed number of apertures in the mature pollen [13, 14], suggesting that the number and placement of membrane aperture domains changes in these cells. However, it is not known what factors linked to ploidy (for example, biochemical—e.g. gene dosage and levels of gene expression, or mechanical—e.g. pollen size) are responsible for changes in membrane domains and aperture number [13, 17, 18].
In addition to pollen ploidy, factors connected to male meiosis and meiotic cytokinesis have been hypothesized to be involved in aperture formation [5, 10, 19–24]. In eudicot species which often have a tetrahedral arrangement of microspores in a tetrad and form three apertures on each microspore, these apertures tend to develop at the centers of the six centripetally growing division plates, close to the three points of last contact that each microspore has with its three sisters at the end of meiotic cytokinesis [5, 10, 24]. These observations led to the hypothesis that points of last contact may serve as patterning landmarks for aperture positions [5, 10, 24]. Although our recent results indicated that the points of last contact per se are unlikely to act as the determinants for aperture placement [13], the strong correlation between their positions and positions of apertures suggest a possibility that meiotic cytokinesis may provide some prepatterning cues for aperture sites.
To explore biochemical and geometric parameters that may be responsible for the observed aperture patterns but cannot yet be approached experimentally, we turned to mathematical modeling. As with other studies where little is known about the underlying interaction network [25], a pattern-forming model can be used to study broad properties of the biological system. The equally spaced distribution of the INP1-decorated membrane domains and apertures in Arabidopsis is reminiscent of patterns that can be generated with mathematical pattern-formation models. A simple reaction-diffusion system, first introduced by Alan Turing [26] and later developed into the Gierer-Meinhardt (GM) system of equations [27, 28], which simulates the interaction of two morphogens, a short-range activator and a long-range inhibitor (Fig 2), was used as the basis for our model. The GM model was chosen because it had been previously used to successfully model a variety of patterns, including some similar to aperture patterns [28–30]. Furthermore, a Turing-type model has been previously used to investigate pattern formation on the surface of a sphere, like pollen [31], although the patterns matching pollen aperture number and placement have not been explored in previous studies.
Here we focused on the specification of aperture number and placement of aperture centers on a single pollen grain. A deterministic model with stochastic initial conditions was able to reliably produce the three-aperture pattern characteristic of the wild-type Arabidopsis pollen. After the model was parametrized, we tested different variables that may have an effect on the number and positions of apertures, such as domain size, morphogen kinetics, initial conditions, and possible pre-patterns.
We found that the domain size and morphogen kinetics have the largest impact on the aperture patterns produced by our simulations. In addition, in parallel with modeling, we performed a new forward genetic screen in Arabidopsis and have identified novel aperture mutants. These mutants develop phenotypes that have not been previously observed in this species but were predicted by our model. We have then used these mutants to further test the model predictions by increasing ploidy of the mutant pollen. Through integration of computational and experimental approaches we have extended the framework for the process of pollen aperture formation and provided insight into the potential causes of mutant pollen patterning.
We implemented the coupled GM equations in FlexPDE, a finite-element partial differential equation (PDE) solving environment, as described in the Materials and Methods. To reproduce the aperture patterning conditions of Arabidopsis, a domain was modeled with a size that corresponded to the size of wild-type haploid (1n) pollen grains with an observed front-view surface area averaging 550 μm2 [13], hereafter referred to as the WT domain size. We used two distinct types of domains in the model: a one-dimensional (1D) domain representing the equator of the pollen grain and a three-dimensional (3D) domain representing the surface on the pollen grain. We explored parameter values that determine the model’s kinetics on the 1D WT domain to find values that would produce three equally spaced spikes, the areas with increased morphogen concentration, to match the aperture patterns of wild-type Arabidopsis pollen grains (Fig 3A; Table 1).
We then verified that these parameter values satisfy the conditions for Turing-type patterning [32]. The chosen parameter values were determined to be within the Turing pattern-forming regime. The molecular mechanisms responsible for regulation of the number of apertures and their positioning are still essentially unknown, thus precluding estimations of kinetics for the molecules involved. Therefore, to determine how much change our model could possibly tolerate while still producing Turing patterns, we tested a range of parameter values. We found that when any single parameter was varied between 40% and 270% from the value in Table 1, the criteria for the formation of Turing patterns [32] remained satisfied, indicating that the model is quite robust in its ability to generate patterns.
To verify that our model can reliably reproduce the patterning of Arabidopsis pollen grains, we then increased the domain to match the larger size of the diploid (2n) Arabidopsis pollen grains (hereafter referred to as the larger-size domain). 2n Arabidopsis pollen grains have an average front-view surface area of 750 μm2 and often develop patterns with four equatorial apertures [13]. With the base parameters from Table 1, our model produced mostly four-spike patterns on the larger-size domain (Fig 3B), similar to patterns seen in Arabidopsis mutants with 2n pollen. To test if a change of the same magnitude in the opposite direction would also change the number of spikes produced by our model, the domain size was decreased to correspond to a front-view surface area of 350 μm2. In this smaller-size domain the number of spikes was reduced to two (Fig 3C).
To test how our model behaved in 3D, we started out with the same parameter values as in 1D. Running initial simulations on the 3D domain, we observed that the WT domain produced mostly three-spike patterns, while the larger-size domain produced mostly four-spike patterns, consistent with the results of the 1D model. The three-spike patterns generated in both 1D and 3D had spikes equally spaced around the equator. However, the four-spikes generated by the 3D model were located at the corners of a tetrahedron. After these initial verifications, we used both the WT domain and the larger-size domain to run our simulations. These two domain sizes, corresponding, respectively, to the sizes of 1n and 2n Arabidopsis pollen grains, were thus capable of reproducing aperture patterns observed in vivo.
Based on the observations that different numbers of spikes can be produced strictly in response to changes in domain size, we decided to systematically determine the effects of domain size on the produced patterns. To accomplish this, we varied the domain size while keeping all of the model parameters at their original values (Table 1). We performed this domain-size analysis for both the 1D and 3D domains and cataloged the resulting patterns. For both domain types, increasing the domain size resulted in a higher number of spikes (Fig 4). The three-spike pattern associated with the wild-type 1n pollen was the most common pattern for 1D domains corresponding to a range of areas between 400 μm2 and 700 μm2 (95% of simulations, n = 175). The 3D domain, however, was more sensitive to deviations from the original size of the WT domain: the three-spike pattern was predominant (72%, n = 25) only for the area of 550 μm2, corresponding to the wild-type domain. The 1D domain produced patterns ranging from two to five spikes, while the 3D domain was capable of producing patterns with up to eight spikes. These results suggest that the domain size, in the absence of any other kinetic or geometric changes, could be responsible for differences in aperture number and patterning.
Higher-ploidy pollen grains have larger size, but they may also have other unknown changes responsible for the changes in aperture patterning. After observing that domain size affects the number of spikes produced, we tested what other mechanisms could be responsible for the aperture phenotypes associated with higher-ploidy pollen. To accomplish this, we performed a one-parameter-at-a-time analysis [33], varying the model parameters in both the WT and the larger-size domains. Parameter values from 20% to 300% of their original values (Table 1) were tested in increments of 20% and the resulting numbers of spikes were tabulated.
In 3D, our model tended to produce patterns with various numbers of spikes that were separated from each other by an area where the concentration of the activator was reduced to almost zero (Fig 7A–7E). In addition, in our kinetic experiments we discovered patterns in morphogen concentration that were composed not just of the typical spikes. One of these patterns was the ring-like pattern that had increased morphogen concentration all along a great circle of the spherical domain (Fig 7F). This pattern was observed on the larger 3D domain, in the cases when the diffusion of the activator was greatly increased (Fig 5) or the decay of the activator was sufficiently reduced (Fig 6). Upon further investigation, we determined that the ring-like aperture was created by merging two separate spikes that were initially located on opposite poles (S1 Video). The second type of unusual pattern consisted of a single spike located at a pole of the sphere, combined with a ring located on the opposite side of the sphere (Fig 7G). Additionally, we observed a pattern composed of six elongated spikes located at the edges of a tetrahedron (Fig 7H and 7H’). All of these more complex patterns were formed when the parameter values for the morphogen diffusion or decay were close to the edge of the pattern-forming regime (Figs 5 and 6).
To determine if patterns predicted in response to changes in model parameters could be observed in vivo in response to genetic perturbations, we performed a forward genetic screen on an ethyl methanesulfonate (EMS)-mutagenized population of Arabidopsis plants. We found mutants belonging to two complementation groups, macaron (mcr) and doughnut (dnt), which had pollen aperture phenotypes that have not been previously observed in Arabidopsis, but were predicted by our mathematical model to result from changes in microspore geometry or morphogen kinetics.
Instead of three apertures characteristic of the wild-type Arabidopsis pollen, pollen in the mcr mutants developed a single circular aperture that, like a belt, surrounded each pollen grain (Fig 8A and 8A’, S2 Video). To establish positional orientation of the circular aperture in relation to pollen poles and equator, we imaged tetrads of microspores with early signs of apertures and found that this aperture passed through the poles in each microspore (Fig 8C), indicating that the normal longitudinal orientation of apertures was not disrupted in the mcr mutant.
By crossing the DMC1pr:INP1-YFP reporter construct [14] into the mcr background, we determined that in this mutant the aperture factor INP1 assembled in a single circular line (Fig 8D, S3 Video). Moreover, through this imaging approach we established that the single aperture in the mature mcr pollen in fact originates as two apertures that form at the opposite sites close to the equator of each microspore, and then become connected at the poles (Fig 8E, S4 Video). Thus, the mcr mutation affected the number, but not the furrow-like morphology or the equidistant longitudinal distribution of apertures. The patterning behavior of these mcr mutant pollen matches the ring-like apertures that were observed in the mathematical model when the decay of the activator was low or the diffusion of the activator was high (Fig 7F, S1 Video). It could also potentially be considered similar to the two-aperture pattern since the ring aperture in the mcr mutants starts as two separate apertures.
Mutants belonging to the dnt complementation group also exhibited a novel aperture phenotype: instead of three furrow-like apertures, pollen developed two round, hole-like, apertures that were significantly wider and shorter than the wild-type apertures, had internal deposits of the exine material sporopollenin, and were located at the opposite sites on the pollen surface (Fig 8B and 8B’, S5 Video). By crossing DMC1pr:INP1-YFP into the dnt background, we established that, as with mcr, in this mutant INP1 became localized to the positions of new apertures. These apertures, however, no longer formed at the pollen equator but rather at its poles (Fig 8F, S6 Video; see S3 Fig for additional examples of dnt phenotypes). Therefore, dnt mutations simultaneously affected multiple aspects of aperture patterns—morphology, number, and position. The patterning of the dnt mutants qualitatively resembles the two-spike patterns observed in the model (Fig 7A).
Our previous findings that pollen ploidy and/or size have a strong influence on aperture number (Fig 1; [13]) prompted us to pay attention to these parameters in mcr and dnt. The size of both mcr and dnt pollen did not differ from the size of the normal 1n pollen produced by the 2n wild-type Arabidopsis (S4 Fig). In addition, crosses of mcr and dnt with the diploid Landsberg and Columbia accessions, as well as with some other diploid Arabidopsis strains, resulted in fully fertile plants that produced homogeneous pollen and themselves generated normal progeny, without any defects indicative of abnormal ploidy. Similarly, the results of ploidy manipulations that were later performed on mcr and dnt (see below) were also consistent with the original mutants behaving like the 2n plants and producing 1n pollen. We conclude, therefore, that the changes in aperture number in these mutants were not caused by a size- or ploidy-sensitive mechanism. Also, in both mcr and dnt mutants, pollen developed through a normal tetrahedral tetrad stage (Fig 8C–8F, S3, S4 and S6 Videos), consistent with the normal progression of meiosis and cytokinesis during microsporogenesis.
Our mathematical model predicted that kinetic changes, such as increased rates of diffusion of an activator (DA) or an inhibitor (DH), can result in the formation of wild type-sized pollen with two apertures (Fig 5). The model further predicted that, if similar changes in diffusion occurred on a larger domain (equivalent to changes in size from 1n to 2n pollen), the number of apertures should correspondingly change from four, typically produced by the 2n/larger-size pollen, to three.
Because both mcr and dnt pollen grains, at least initially, develop two apertures, we decided to use them to test the model’s predictions by creating mcr and dnt plants that would produce larger (2n) pollen and assessing the number of apertures in these pollen grains. By treating mcr plants with colchicine, as we did previously for wild type [13], we generated tetraploid (4n) mcr plants that produced 2n pollen through a tetrad stage. Additionally, both mcr and dnt mutants were crossed with the osd1 mutation, which results in the omission of the second meiotic division and leads to the formation of 2n pollen through a dyad stage [13, 34] (Fig 1G). In all cases, the size of the resulting 2n mcr and dnt pollen was similar to the size of other 2n pollen (S4 Fig), as well as to the size of the larger domain that was used in our simulations.
In wild type, such ploidy-increasing manipulations commonly lead to the development of pollen with four or six apertures (Fig 1B–1D’ and 1G; [13]). In the case of mcr, independently of the mechanism through which the 2n pollen was generated and consistent with the model’s predictions, the pollen switched to producing three equidistant apertures (Fig 9A–9B’). To determine positional orientations of these three apertures in the mcr osd1 pollen, we used INP1-YFP as a marker for aperture positions. We found that, in each microspore, INP1-YFP localized to three equidistant longitudinal lines, which were not aligned between the sister microspores (Fig 9C, S7 Video), unlike in wild-type tetrads where the lines of INP1 align between sister microspores [12–14].
In contrast to the 2n mcr pollen, in the case of the 2n dnt pollen (dnt osd1), pollen grains retained two apertures that were morphologically identical to the apertures of the 1n dnt pollen (Fig 9D and 9D’, S8 Video). Also, the INP1-YFP signal was found at the poles of microspores in the diploid dyads (Fig 9E, S9 Video), indicating that, like in their 1n counterparts, the two apertures in the 2n dnt osd1 pollen developed at the poles.
In Arabidopsis, pollen with ploidy higher than 2n commonly develops complex and irregular aperture patterns, often consisting of ring-shaped apertures (Fig 1E–1F’; [13]). To test what effect further ploidy increase will have on the mcr and dnt pollen, we took advantage of the fact that the osd1 mutation affects meiosis not only in pollen but also in eggs, thus leading to doubled ploidy after self-pollination [13, 34]. Using this approach, we generated tetraploid (4n) mcr osd1 and dnt osd1 plants that produced 4n pollen through the dyad stage. As with other versions of the mcr and dnt pollen, which were similar in size to the other pollen of the same ploidy level, the size of the 4n mcr and dnt pollen did not differ from the size of 4n osd1 pollen (S4 Fig) or from the other 4n pollen types that we previously studied [13].
In the 4n mcr osd1 pollen, the predominant aperture pattern (81%, n = 85) consisted of a ring-shaped aperture located on one side of the pollen grain and 1-2 small, hole-like apertures on the opposite side of the pollen grain (Fig 9F and 9F’; S5 Fig). The remaining grains had variations of this pattern: for example, a ring-shaped aperture and a furrow (S5C and S5C’ Fig), or two ring-shaped apertures on the opposite sides of a grain (S5D and S5D’ Fig). Interestingly, this unusual aperture pattern strongly resembled the ring-and-spike pattern produced by the model (Figs 6 and 7G). To determine where on the surface of 4n mcr osd1 pollen the ring and the hole-like apertures were located, we analyzed the late-stage dyads in which apertures were already recognizable. We found that the rings were located at the distal end of microspores, thus placing the hole-like apertures at the proximal end, near the intersporal callose wall (Fig 9H–9H”’, S10 Video.
In the 4n dnt osd1 pollen, the shape and size of apertures was still unchanged compared to the 1n dnt or 2n dnt osd1 (Fig 9G and 9G’); however, their numbers increased. Most of the pollen grains had either three (59%, n = 81) or four (36%, n = 81) hole-like apertures (Fig 9G and 9G’, S3A and S3A’ Fig), with some having up to eight apertures (S3B and S3B’ Fig, S11 Video).
It is not clear if some kind of pre-patterning information (e.g. related to meiosis or cytokinesis, such as positions of meiotic spindles or last contact points between sister microspores) influences the number and positions of apertures. It is also unknown if pre-patterns might act as transient or sustained stimuli. To test whether pre-patterning can dictate final patterns of spikes, we subjected the model to either a transient stimulus, applied as an initial condition, or a sustained stimulus which remained active through the entire simulation time.
In Arabidopsis thaliana and in many other eudicot species, aperture patterns on the pollen surface are influenced by mutations that affect size and ploidy of the pollen grains [13, 35–38]. It is, however, unknown if this aberrant patterning is caused by the increase in the pollen grain size or by changes in some other factor(s) linked to ploidy. To explore possible mechanisms responsible for pollen aperture patterning, a mathematical model based on the Gierer-Meinhardt equations was developed. This GM-based model successfully recapitulated the triaperturate equatorial patterning of the wild-type Arabidopsis pollen both in 1D and in 3D geometries.
Our simulations demonstrated that the domain size, as well as morphogen diffusion and decay, could greatly influence the patterns produced by the morphogens. When the kinetic parameters of the morphogens were kept constant, a domain with a larger surface area generated patterns with a larger number of spikes representing apertures: under these conditions more spikes could fit on the domain.
Areas between the morphogen spikes exhibited reduced concentrations of the two morphogens, the activator and the inhibitor. Importantly, in these areas the concentration of the activator went down almost to zero, whereas the concentration of the inhibitor, although decreased, still remained at a relatively high level (Fig 3). The size of these areas of inhibition depends on the kinetics of the system—in particular, on the diffusion and decay of the two morphogens. If the diffusion of a morphogen is increased, the morphogen can travel further through the domain, leading to the decrease in the number of spikes produced. When the morphogen decay rate is decreased, the morphogens are also able to travel further and a similar decrease in the number of resulting spikes was observed. These two types of kinetic changes both extend the morphogen’s range and increase the area of inhibition, thus limiting the number of spikes that can be formed. The opposite changes to the morphogen kinetics—decrease in the diffusion or increase in the decay—will decrease the area of inhibition and correspondingly increase the number of spikes. The other parameters used in the model, ρ, ρA, and ρH, do not influence the dispersion of the morphogens but instead control the rate of morphogen production at any given location. It appears that the distance the morphogens travel, rather than the amounts of morphogens, can affect the number and patterning of spikes. Formation of the areas of inhibition may be the basis of the mechanism acting in Arabidopsis pollen grains.
To explore how sensitive our model was to the influence of additional stimuli, we first tested transient stimuli, in the form of either patterned or unpatterned initial conditions, followed by continuous stimuli. We found that only a strong transient stimulus could influence the final pattern produced by the model, whereas the weak stimuli, both patterned and unpatterned, were unable to affect the final spike pattern. Similarly, very weak continuous stimuli were often unable to influence patterning. Yet, overall, continuous stimuli of low amplitudes were more successful in specifying patterns than the corresponding transient stimuli. We found that in the presence of either transient or continuous stimuli with three spikes and the amplitude above 10−4 μM, the larger-sized 3D domain produced only three-spike solutions instead of the typical four, indicating the strong effect of these types of stimuli on the resulting pattern. This suggests that if there exists some form of in vivo pre-patterning stimulus specifying three apertures, then the stimulus is likely to be small since pollen is capable of overcoming its influence and producing a different number of apertures (Fig 1G).
Our model was able to recreate elongated, furrow-like spikes resembling Arabidopsis apertures (Fig 7) only when we used extreme parameter values that were close to the levels at which the model reaches bifurcation and stops producing patterns. This indicates that the elongation is not robust to perturbations to the model, and that the formation of elongated apertures in vivo may be caused by other mechanisms that were not captured by our model.
Interestingly, the distribution of four apertures, equally spaced around the equator and commonly observed in the 2n Arabidopsis pollen, was not captured well in the 3D geometry. The majority of the four-spike patterns produced in 3D had the spikes centered on the corners of a tetrahedron (Fig 7C). In contrast, in 1D the model assumes that the process of initial aperture patterning is restricted to the equator of the microspores, and here four-spikes were the predominant pattern for larger domain sizes and for many of the tested parameter values. If the process that initially positions aperture factors is restricted to the pollen equator, then the 1D domain would be a good fit for the patterning expected in Arabidopsis mutants with larger pollen grains.
In addition to the patterns of three or more apertures that were previously observed in Arabidopsis pollen, our model also predicted that under some conditions pollen might be able to produce two apertures or a ring-shaped aperture that could be created by the fusion of two apertures. The aperture pattern of the mcr mutant, with a ring-like aperture produced by joining of the two initial apertures (Fig 8E, S4 Video), has a similar phenotype and appears to be generated in the same way as the ring-like patterns in the model (S1 Video). However, in the mathematical model, the ring-like patterns developed only on the larger-size domain and either when the morphogen diffusion was at an extreme value, close to the boundary of the pattern-forming region, or when the morphogen decay was sufficiently reduced. On the WT domain, the ring-like pattern was not formed with any choice of individually changed parameters, nor was it formed in the presence of a stimulus mimicking the ring-like pattern. Therefore, it is possible that in the mcr mutants, which have pollen of normal size, the mutation leads to multiple kinetic changes that together contribute to the production of the ring-like aperture. An alternative possibility is that the apertures in the mcr mutant correspond to the two-spike pattern, which was frequently produced by our model for normal-size domains (e.g. in the cases of increased diffusion or decreased decay for an activator). This hypothesis is consistent with the observation that apertures in the tetrad-stage mcr microspores start their development at two separate positions (Fig 8E, S4 Video) and later coalesce into the ring-like pattern. It is, therefore, conceivable that the mcr mutation could affect the kinetics of aperture factors and that separate mechanisms could be responsible for positioning the centers of apertures and for promoting aperture elongation that creates their final, furrow-like, morphology.
When pollen ploidy and size were increased in the mcr mutant, the behavior of its aperture patterns followed the model’s predictions. The model suggested that, instead of two apertures, three apertures should form if the domain size increased to that of the 2n pollen. Indeed, the 2n mcr pollen developed the triaperturate pattern—independently of the way this pollen was produced, through a tetrad or dyad. In the 4n mcr pollen, the most common aperture pattern consisted of a ring-shaped aperture on one side of the pollen grain and one or two hole-like apertures on the opposite side. Interestingly, our model was also able to generate this unusual pattern.
In palynological literature, the mcr -like aperture pattern is referred to as bi-syncolpate [6], zonacolpate [39], or ring-like [2]. The examples of pollen with this aperture pattern are relatively infrequent in nature: for instance, only 15 species out of almost 3,000 species represented in the palynological database PalDat exhibit this pattern [2]. Interestingly, in the case of one well-documented example, the genus Pedicularis includes multiple species with bi-syncolpate pollen, as well as species that produce triaperturate (tricolpate and tri-syncolpate) pollen grains [2, 40, 41]. This suggests a possibility that the variations in aperture patterns among the Pedicularis species could be the result of mutations affecting the homologs of mcr or related aperture factors.
The phenotype of the dnt mutant, with its two round apertures on the opposite poles of the pollen grain, mimics the two-spike patterns produced by the model. This indicates that the dnt mutation may affect the useful lifespan of the morphogens or the way they are transported within the microspores. The dnt mutation might increase the area of inhibition in microspores, so that only two spikes in the concentration of aperture factors will form.
If there are indeed kinetic changes in the Arabidopsis mutants, it is unlikely that they only affect one of the parameters in the model. To account for this, a multi-parameter kinetics analysis could be performed to test all possible combinations of parameter values. In addition, our mathematical model only considered a single microspore in the shape of a perfect sphere, whereas in vivo apertures are formed on the microspores that are not entirely spherical and are joined into tetrads. Differently shaped microspores could be modeled to determine if the surface morphology plays a role in the number and positions of apertures. The local geometric and biophysical properties, such as membrane curvature and pressure from the callose wall that envelopes tetrads, could potentially affect protein kinetics, movement, or positioning and influence such processes as aperture elongation. The identification of the genes affected in the mcr and dnt mutants will likely provide important insights into the mechanism of pollen aperture patterning that can be used to further refine the mathematical model.
The choice of the Gierer-Meinhardt model in this study was due to the fact that the patterns observed on pollen grains are equally spaced, similar to the patterns produced by the GM model. The GM model has been successfully used to study similar patterns in other biological systems [27, 28], and we applied that model here to better understand specification of aperture positioning. A similar model has previously been used to simulate pattern forming processes, where the underlying interaction network is unknown, to great effect [25]. As more information is gained about the proteins involved in aperture specification, more detailed and biologically realistic models can be developed to explore aperture formation in more detail. For example, we have recently identified the protein kinase D6PKL3 as an aperture factor [16]. This suggests that mechanisms involving protein phosphorylation and dephosphorylation might have to be taken into account in the future models.
In conclusion, we were able to show computationally that aperture formation in Arabidopsis could be controlled by kinetics of aperture factors and by pollen grain size. Following the model’s predictions, we found Arabidopsis plants with the previously unobserved mcr and dnt aperture phenotypes. These findings demonstrate that mathematical modeling is able to provide valuable insights even when the mechanisms behind the biological phenomena are unknown or when a direct connection between the biological patterning mechanism(s) and the mathematical patterning is not known to exist. Future work with similar qualitative modeling may help to explain the mechanisms responsible for the great variety of patterns that exist in pollen grains across species.
In the Gierer-Meinhardt equations [27, 28], a short-range activator (A( x → , t ) and a long-range inhibitor (H( x → , t ) are modeled as a system of partial differential equations. Eqs 1 and 2 describe the changes over time of the activator and the inhibitor, respectively. In the equations, DA and DH are the diffusion constants, μA and μH are the removal/decay constants, ρA and ρH represent a constant basal growth of each morphogen, and ρ1 and ρ2 are the rates of the interaction between the activator and the inhibitor. The magnitudes of ρ1 and ρ2 are the same value and they differ only in their units.
∂ A ∂ t = D A ∇ 2 A - μ A A + ρ 1 ( A + ρ A ) 2 H (1) ∂ H ∂ t = D H ∇ 2 H - μ H H + ρ 2 A 2 + ρ H (2)
Two types of domains were used to explore the patterning of the two morphogens. The first was a one-dimensional domain representing the equator of a microspore. The second was the surface of a three-dimensional sphere, representing the surface of a microspore. For all simulations, the initial conditions of the two morphogens were set to be near their diffusion-free steady-state values. Turing-type patterns require the initial conditions to have a level of random noise for patterns to form [26]. Therefore, we tested random noise levels between 10−15 μM to 102 μM, and found that the model produced identical patterns, indicating that within these limits the noise did not influence the resulting patterns. However, levels of noise below 10−6 μM and above 10−2 μM resulted in simulations taking longer to run. Therefore, we added a random value between ±5*10−4 μM to every point of our initial conditions.
The following assumptions and simplifications were made:
Simulations of the model (Eqs 1 and 2) were performed using the software FlexPDE (pdesolutions.com). Time and space steps were chosen to ensure numerical stability. To test for possible causes of the change in pollen aperture number and patterning observed in Arabidopsis mutants, the following in silico experiments were run:
Wild-type (Landsberg erecta) and mutant Arabidopsis plants were grown at 20 − 22° C with the 16-hour light: 8-hour dark cycle in growth chambers or in a greenhouse at the Biotechnology facility at OSU. The genetic screen that led to the isolation of the mcr and dnt mutants was performed similar to the previous screen [42]. In brief, M2 plants from eight pools of EMS-treated lines of Landsberg erecta background (~10,000 plants) were screened for the presence of morphological abnormalities in their pollen (e.g. in size, shape, light reflection, ease of pollen release from anthers) that were identifiable with standard dissecting stereomicroscopes (Zeiss Stemi-2000C and Nikon SMZ745) at 75-80X magnification. For primary screening, pollen did not undergo any treatment. Particular attention was paid to changes in pollen shape, known to be associated with aperture defects [12, 42]. Pollen of the candidates isolated in the primary screen was stained with auramine O as described [13] and observed for exine and aperture defects with confocal microscopy. Confirmed mutants were then backcrossed at least once. mcr and dnt mutants were crossed with the previously described DMC1pr::INP1-YFP line [14] and with osd1-2/+ plants that were identified as described [34].
To create plants of higher ploidy, shoot apical meristems of young plants were treated with colchicine as previously described [13], with minor modifications. Because we found that Landsberg plants were very sensitive to colchicine, the amount of colchicine and the number of treatments were reduced compared to the previous study. A 20 μl drop of 0.25% colchicine; 0.2% Silwet L-77 was applied once onto shoot apices before bolting. Conspicuous, larger-than-normal flowers (often found on thicker-than-normal stems) in colchicine-treated plants were allowed to self-pollinate and their seeds were harvested. These progeny were then analyzed for characteristic increase in the size of plant organs and of pollen, and for stable inheritance of these traits.
Samples for confocal microscopy were prepared as described [13]. Exine of mature pollen stained with auramine O was excited with a 488-nm laser and the emitted fluorescence was collected at 500-550 nm. Tetrads released from stage-9 anthers [43], were placed into Vectashield anti-fade solution (Vector Labs) supplemented with 0.02% Calcofluor White and imaged on a Nikon A1+ confocal microscope with a 100x oil-immersion objective (NA = 1.4) and 5x confocal zoom. The following settings were used to collect signals of fluorophores: YFP—514-nm excitation/ 522-555 nm emission; Calcofluor White and early exine on tetrad-stage microspores—405-nm excitation/ 424-475 nm emission. Z-stacks of tetrads were obtained with a step size of 500 nm and volume-reconstructed using NIS Elements v.4.20 (Nikon).
|
10.1371/journal.pcbi.1006217 | Quantitative theory of deep brain stimulation of the subthalamic nucleus for the suppression of pathological rhythms in Parkinson’s disease | Deep brain stimulation (DBS) of the subthalamic nucleus (STN) is modeled to explore the mechanisms of this effective, but poorly understood, treatment for motor symptoms of drug-refractory Parkinson’s disease and dystonia. First, a neural field model of the corticothalamic-basal ganglia (CTBG) system is developed that reproduces key clinical features of Parkinson’s disease, including its characteristic 4–8 Hz and 13–30 Hz electrophysiological signatures. Deep brain stimulation of the STN is then modeled and shown to suppress the pathological 13–30 Hz (beta) activity for physiologically realistic and optimized stimulus parameters. This supports the idea that suppression of abnormally coherent activity in the CTBG system is a major factor in DBS therapy for Parkinson’s disease, by permitting normal dynamics to resume. At high stimulus intensities, nonlinear effects in the target population mediate wave-wave interactions between resonant beta activity and the stimulus pulse train, leading to complex spectral structure that shows remarkable similarity to that seen in steady-state evoked potential experiments.
| Pathological 13-30 Hz (beta) oscillations within the basal ganglia are a characteristic feature of human Parkinson’s disease which seem to correlate with symptom severity. The origin of these oscillations and the suppressive mechanism of effective deep brain stimulation treatments remains to be shown. We formulate a physiologically based population model of the corticothalamic-basal ganglia system that produces 13-30 Hz oscillations in the neural circuit formed between the globus pallidus pars externa and the subthalamic nucleus and the hyperdirect corticothalamic-basal ganglia pathway. We then develop a model of deep brain stimulation applied to the corticothalamic-basal ganglia system that permits systematic determination of effective stimulus protocols, which have been estimated by trial and error to date. Our results demonstrate that high pulse frequency (>140 Hz) stimulation is required to effectively suppress the pathological oscillations, which agrees with clinically used values. Interactions between these oscillations and the applied stimulus also lead to complex spectral structure that shows remarkable similarity to that seen in steady-state evoked potential experiments.
| Deep brain stimulation (DBS) has become an effective treatment for a number of neurological disorders such as Parkinson’s disease (PD) and essential tremor [1, 2]. In Parkinson’s disease DBS treatments, a macroelectrode is chronically implanted in a target nucleus, typically either the globus pallidus internus (GPi), subthalamic nucleus (STN), or the ventral intermediate nucleus of the thalamus; this electrode delivers high frequency (>100 Hz) electrical stimulation as a series of pulses. More broadly, studies have also shown the efficacy of DBS treatments in dystonia [3], epilepsy [4], and obsessive-compulsive disorder [5].
Significant progress has been made exploring the influence of DBS on neural activity [6]. However, the efficacy of DBS treatments could be improved with a greater understanding of the underlying therapeutic mechanisms. Furthermore, it is unclear what stimulation parameters, electrode geometries, and electrode locations are most effective for the present and future uses of DBS technologies.
Electrical stimulation of the brain influences a variety of mechanisms involved in the function and signaling of neurons. The sensitivity of different contributing elements depends upon the amplitude and temporal properties of the stimulation [7], geometry of the stimulus field [8], target cell physiology and geometry [9], and the possible pathophysiology of different disease states [10]. It is known that distinct neuron types possess different types of ion channels and that these may have different voltage-sensitive activation and inactivation properties [11]. Thus, the effect of DBS on a single neuron’s dynamics may vary significantly between brain regions. However, by averaging over millimeter to whole-brain scales, a generalized description of DBS at a population level may provide insights into the net effect of these different mechanisms and allow prediction of effective stimulation protocols.
It was initially thought that deep brain stimulation had a predominantly inhibitory effect on the stimulated population due to similar therapeutic effects to lesioning [12, 13]. This inhibition hypothesis is supported by several experimental findings in STN-DBS of rats [14] and monkeys [15] and GPi-DBS and STN-DBS in humans [16, 17, 18]. Furthermore, [19] demonstrated an inhibitory response to GPi-DBS that was mediated by the GABA receptors and that this inhibition could be blocked via a GABA antagonist.
Seemingly in contradiction with the inhibition hypothesis, several experiments with recordings in efferent nuclei of the DBS target population indicate an increase in the stimulated populations activity [20, 21, 22], and an entrainment of local neural firing during DBS [23].
Several modeling approaches have been used to elucidate DBS effects across multiple scales. Finite element methods used in conjunction with multi-compartment neuron models have explored DBS responses in small neural assemblies [24, 25, 8], and the distribution of the applied electric field [26]. These single cell models have demonstrated a disassociation of activity at the soma relative to the axon during extracellular stimulation [24], and a systemic activation of axons both efferent and afferent to the stimulation site [27]. The variable nature of response to DBS between brain regions might then be understood by approximating DBS as an activation of intrinsic axons within a certain effective range of the electrode, and thus could explain observations supporting the contradictory excitation and inhibition hypotheses.
Parkinson’s disease (PD) is a neurodegenerative disorder characterized by motor dysfunction including akinesia, bradykinesia, tremor, and rigidity [28]. These clinical manifestations have been linked to dopaminergic denervation in the substantia nigra pars compacta (SNc) and synucleinopathy leading to Lewy bodies, and neurites in the SNc and other brain regions [29]. The firing pattern hypothesis regarding PD proposes that pathological oscillations and/or synchronization play a primary role in motor symptoms of the disorder. Single unit and local field potential recordings have shown enhanced activity within and between the basal ganglia (BG), thalamus and motor cortex at about 4–8 Hz and 13–30 Hz [30, 31, 32, 33, 34], which seems to correlate with significant coherence at these frequencies [35, 36, 37, 38, 39, 40]. It has thus been suggested that these pathological rhythms cause a disturbance of motor-related information processing in the BG [41], which could explain some PD symptoms.
The enhanced beta band oscillations (13-30 Hz) found in the STN of PD patients are thought to be related to symptom severity, based on direct correlation results [42], as well as observations of a reduction in beta power following treatments that ameliorate PD symptoms, such as dopaminergic supplementation [43] and deep brain stimulation [44]. Several experimental and modelling studies have suggested that the circuit formed between the STN and GPe may be responsible for beta activity generation [45, 46], and that cortical excitatory inputs to the STN amplify them [47]. However, the origin of parkinsonian beta activity is still a matter of debate.
Physiologically based mean-field models of the brain provide a tractable framework for the analysis of large-scale neuronal dynamics by averaging microscopic structure and activity [48, 49, 50, 51, 52]. Neural field theory incorporates realistic anatomy of neural populations, nonlinear neural response, interpopulation connections and dendritic, synaptic, cell-body, and axonal dynamics [48, 50, 52, 53, 54, 55, 56, 57, 58]. Neural field models have been successful in accounting for many characteristic states of brain activity including sleep stages, eyes-open, and eyes-closed in waking, nonlinear seizure dynamics, anesthesia and many other phenomena [52, 53, 55, 51, 59, 60, 61, 62].
In the particular case of Parkinson’s disease, neural field models of the corticothalamic-basal ganglia system have been able to account for several electrophysiological correlates of the disease including changes in population average activity, and ∼4-8 Hz and ∼13–30 Hz oscillations characteristic of EEG and LFP spectra [63, 64, 65]. However, the generative mechanisms of these characteristic PD rhythms is still a matter of debate.
The core aims of this work are to develop a population level description of DBS of the corticothalamic-basal ganglia (CTBG) system that can account for experimental observations and the results of other modeling studies. The work will explore parkinsonian states of the CTBG system and determine whether subthalamo-pallidal circuits can sustain characteristic beta oscillations in this framework. Finally, the effects of DBS on these parkinsonian states will be analyzed and provide insights into the efficacy of DBS treatments.
Fig 1 shows a schematic of the CTBG model. The system contains nine distinct neural populations across three brain regions. The cerebral cortex contains populations of excitatory pyramidal neurons, e, and inhibitory interneurons, i. The thalamus is divided into an excitatory population for the specific relay nuclei (SRN), s, and an inhibitory population for the thalamic reticular nucleus (TRN), r. The basal ganglia (BG) contains two inhibitory populations within the striatum, one expressing the D1 dopamine receptor, d1, and one expressing the D2 dopamine receptor, d2. The striatum projects to two inhibitory populations, the globus pallidus pars externa, p2, and a population representing the globus pallidus pars interna and substantia nigra pars reticulata, p1. The subthalamic nucleus (STN) is represented by an excitatory population, ζ. Finally, deep brain stimulation is defined as an input source, x, which is coupled to STN as well as to its projection sites. This is discussed in detail in a later section. The substantia nigra pars compacta (SNc) and ventral tegmental area (VTA) are not explicitly defined as a population within the model, however, they are included in Fig 1 as an indication of the neural pathways affected by dopamine.
The mean firing rate, Qa(r, t), of a neural population can be approximately related to its mean membrane potential, Va(r, t), by [66, 67]
Q a ( r , t ) = S a [ V a ( r , t ) ] , (1) = Q a max 1 + exp [ - { V a ( r , t ) - θ a } / σ ′ ] . (2)
where Eqs (1) & (2) define the sigmoidal mapping function Sa, Q a max is the maximal firing rate, Va is the average membrane potential relative to resting, θa is the mean neural firing threshold, and σ ′ π / 3 is the standard deviation of this threshold.
A number of experimental studies have revealed waves of neural activity spreading across the cortex [68, 49, 69, 70], which have been analyzed theoretically [71, 72, 73, 51, 74, 75, 52, 58]. This propagating activity is represented as a field of mean spike rates in axons, ϕa. A population a, with a mean firing rate Qa, is related to ϕa by the damped wave equation
D a ( r , t ) ϕ a ( r , t ) = Q a ( r , t ) , (3)
where
D a ( r , t ) = 1 γ a 2 ∂ 2 ∂ t 2 + 2 γ a ∂ ∂ t + 1 - r a 2 ∇ 2 . (4)
Here, γa = va/ra represents the damping rate, where va is the propagation velocity in axons, and ra is the characteristic axonal length for the population. The propagation of these waves is facilitated primarily by the relatively long-range white matter axons of excitatory cortical pyramidal neurons. Later in our model the simplifying local interaction approximation rb ≈ 0 is made for b = i, r, s, d1, d2, p1, p2, ζ due to the short ranges of the corresponding axons which implies ϕb(r, t) = Qb(r, t) for these populations [52, 57, 54, 55, 76, 77].
The mean soma potential Va of a population a at position r and time t is given by sum of the postsynaptic potentials (PSPs):
V a ( r , t ) = ∑ b V a b ( r , t ) , (5)
where Vab(r, t) is the postsynaptic potential generated by projections arriving at population a from population b. The influence of incoming spikes to population a from population b is weighted by a connection strength parameter, νab = Nabsab, where Nab is the mean number of connections between the two populations and sab is the mean strength of response in neuron a to a single spike from neuron b. The postsynaptic potential response in the dendrite is given by
D α β V a b ( r , t ) = ν a b ( r , t ) ϕ a b ( r , t - τ a b ) , (6)
where τab is the average axonal delay for the transmission of signals to population a from population b. The operator Dαβ describes the time evolution of Vab in response to synaptic input, and is given by
D α β = 1 α β d 2 d t 2 + ( 1 α + 1 β ) d d t + 1 . (7)
where β and α are the overall rise and decay response rates to the synaptodendritic and soma dynamics.
It has been shown that nominal brain activity is well characterized by perturbations about a mean value [55]. Hence, we first find the time independent states of the CTBG system. Following the approach of previous neural field models, excitatory and inhibitory synapses in the cortex are assumed proportional to the numbers of neurons [50, 78]. This random connectivity approximation results in νee = νie, νei = νii, and νes = νis, which implies Ve = Vi and Qe = Qi. Inhibitory population variables can then be expressed in terms of excitatory quantities and are thus not neglected even though they do not appear explicitly below.
The steady states are obtained by setting all time derivatives to zero in Eqs (3), (4) and (6). Using the connectivity shown in Fig 1, and excluding DBS, Eqs (5) and (6) give
V e ( 0 ) = ( ν e e + ν e i ) ϕ e ( 0 ) + ν e s ϕ s ( 0 ) , (8) V r ( 0 ) = ν r e ϕ e ( 0 ) + ν r s ϕ s ( 0 ) , (9) V s ( 0 ) = ν s e ϕ e ( 0 ) + ν s r ϕ r ( 0 ) + ν s p 1 ϕ p 1 ( 0 ) + ν s n ϕ n ( 0 ) , (10) V d 1 ( 0 ) = ν d 1 e ϕ e ( 0 ) + ν d 1 s ϕ s ( 0 ) + ν d 1 d 1 ϕ d 1 ( 0 ) , (11) V d 2 ( 0 ) = ν d 2 e ϕ e ( 0 ) + ν d 2 s ϕ s ( 0 ) + ν d 2 d 2 ϕ d 2 ( 0 ) , (12) V p 1 ( 0 ) = ν p 1 ζ ϕ ζ ( 0 ) + ν p 1 d 1 ϕ d 1 ( 0 ) + ν p 1 p 2 ϕ p 2 ( 0 ) , (13) V p 2 ( 0 ) = ν p 2 ζ ϕ ζ ( 0 ) + ν p 2 d 2 ϕ d 2 ( 0 ) + ν p 2 p 2 ϕ p 2 ( 0 ) , (14) V ζ ( 0 ) = ν ζ e ϕ e ( 0 ) + ν ζ p 2 ϕ p 2 ( 0 ) . (15)
The system’s steady states then can be determined by considering the simultaneous zeros of the five functions
F ( ϕ e ) = ϕ e - S e [ ( ν e e + ν e i ) ϕ e + ν e s ϕ s ] , (16) F ( ϕ s ) = ϕ s - S s [ ν s e ϕ e + ν s r ϕ r + ν s p 1 ϕ p 1 + ν s n ϕ n ] , (17) F ( ϕ d 1 ) = ϕ d 1 - S d 1 [ ν d 1 e ϕ e + ν d 1 s ϕ s + ν d 1 d 1 ϕ d 1 ] , (18) F ( ϕ d 2 ) = ϕ d 2 - S d 2 [ ν d 2 e ϕ e + ν d 2 s ϕ s + ν d 2 d 2 ϕ d 2 ] , (19) F ( ϕ p 2 ) = ϕ p 2 - S p 2 [ ν p 2 d 2 ϕ d 2 + ν p 2 p 2 ϕ p 2 + ν p 2 ζ ϕ ζ ] . (20)
where ϕr, ϕp1, and ϕζ can be determined from Eqs (9), (13) and (15), respectively, in conjunction with Eq (1). The roots of Eqs (16)–(20) are computed numerically using the MATLAB function fsolve() with a tolerance of 10−15 V.
A linearized form of the CTBG model can be used to derive the transfer function of the system [63, 64, 65]. This is a function of the internal gains of the system, which represent the additional activity generated in postsynaptic nuclei per additional unit input activity from presynaptic nuclei, and are [53, 55]
G a b = ρ a ν a b (21)
where
ρ a = d Q a d V a | V a ( 0 ) = ϕ a ( 0 ) σ ′ [ 1 − ϕ a ( 0 ) Q a max ] . (22)
All numerical simulations of the CTBG neural field model in this work are performed using the NFTsim code package detailed by [79]. This package is used to solve Eqs (1)–(7) numerically for the spatially uniform case where the ∇2 in (4) is zero. The solutions to these delay differential equations are found using a standard forth-order Runge-Kutta integration method with a time step of 10−4 s.
Nominal brain states have been found to exist near stable fixed points [55]. Hence, all simulations in this work are performed with the system initialized to the low firing steady state found in the previous section using the parameters given in Table 1, unless otherwise specified.
Many different stimulus protocols have been used in clinical DBS—with different pulse geometries (i.e. sinusiodal or square-wave), signal amplitudes, stimulation frequencies, and/or transient stimulation phases, followed by varied quiescent periods.
In this work we seek a general formulation of a neural populations response to fluctuations in an applied electric field that will allow for the effects of various stimulus protocols to be determined.
The minimum current necessary to stimulate a given neural element with a long stimulus duration is called the rheobase [80]. The minimum length of time required to activate a given neural element using a stimulus amplitude twice as large as the rheobase is called the chronaxie. Extracellular stimulation experiments have demonstrated a chronaxie time for the myelinated axons which is substantially smaller than the chronaxies of the cell body and dendrites [7, 81, 82]. Hence, our key assumption is that the net effect of fluctuations of an applied electric field is a stimulation of voltage-gated ion channels that induces transmembrane current flow predominantly in both afferent and efferent axons of a subset of neurons within the stimulated population.
A mean-field model has recently been used to describe population effects of transcranial magnetic stimulation [83, 84]. A modification of this approach is used by defining an external pulse rate ϕx(t) that consists of a train of pulses with a width twidth similar to time series used in DBS treatments. The applied stimulation is then given by
ϕ x ( t ) = ϕ x max ∑ j R ( t - t j p ) , (23)
where ϕ x max is the pulse amplitude and R(t) is a top-hat function of width twidth,
R ( t ) = { 1 , 0 < t < t width , 0 , otherwise. (24)
The time-integral of ϕx(t), Eq (23), over the pulse width twidth is the average number of additional spikes generated in the target axon by the applied stimulation. The external stimulus is then coupled to a target population a via a connection parameter νax with a pulse frequency fstim. In the case of STN-DBS, ϕx(t) is coupled to the STN, but also to the GPi and GPe populations as an approximation of the activation of axons terminals near the stimulation site.
An afferent spike rate to any population in the CTBG system induces a change in the dendritic membrane potential of that population with a time evolution described by Eq (6). Depending on the connection type, this change may be positive (excitatory) or negative (inhibitory). Each inter-population connection then produces a change in voltage which is integrated at the soma, as described by Eq (5).
In the case of DBS, ϕx(t), the mean voltage perturbation observed at the soma of neurons, can be shown by numerically convolving the stimulus time series with the normalized impulse response function given in differential form in Eq (7). Fig 2 shows the evoked response potential generated by a stimulus pulse train, which resembles typical 130 Hz clinical stimulation, and the resulting perturbation to the target population firing rate. The temporal parameters used for the stimulus in Fig 2(a) prescribes an inter-pulse quiescent period of about 7 ms. It can be seen in Fig 2(b) that during this period the impulse response function only decays to about 80% of its maximum value. Fig 2(c) can then be understood as showing small oscillations in the evoked response potential about a constant mean perturbation that results from stimulus time scales which are shorter than population response time scales. The evoked response potential is integrated at the soma of the target population along with intrinsic afferents from other populations within the network. A constant perturbation applied to the soma potential of a population changes its mean firing rate by moving the population along its corresponding sigmoidal response function, (1). Fig 2(d) demonstrates that in the case of inhibition mean soma potentials correspond to lower mean firing rates when compared with unperturbed values. In the case of excitation, the effect is reversed with mean soma potentials corresponding to higher mean firing rates when compared with the unperturbed values.
Enhanced activity at ∼13-30 Hz is a common feature of Parkinson’s disease patient LFP recordings in the GPi and STN which has been correlated with symptom severity [43, 44, 42]. Recent works have suggested that the neural circuit formed between the GPe and STN can generate these beta oscillations [45, 46] and that excitatory inputs from the cortex may facilitate their amplification [47].
Table 1 contains parameter estimates for parkinsonian states of the CTBG model adapted from [63]. Changes to the [63] connection strength estimates were made in order to explore the effects of a dominant GPe-STN-GPe pathway.
In Fig 3(a) and 3(b), power spectra of the STN firing rate demonstrate enhanced activity at ∼26 Hz, as well as at ∼6 Hz. By increasing STN-GPe coupling vp2ζ, damping of the GPe-STN-GPe loop is weakened and results in a strengthened hyperdirect pathway. Together these loops drive 26 Hz oscillations in the STN firing rate which project to the GPi population through STN efferents and then on to thalamic and cortical populations.
The 6 Hz STN oscillations observed in Fig 3(b) are weaker than the 26 Hz beta oscillations. However, the power spectrum for the cortical population shows an opposite relationship with stronger 6 Hz activity than the 26 Hz beta oscillations. This is an interesting result because tremor oscillations in PD patients measured via electromyography (EMG) are typically about 6 Hz and these correlate well with motor cortical activity measured via electroencephalography (EEG) [85]. However, STN LFP recordings about the 4–6 Hz tremor frequency have yet to be demonstrated as a reliable source for tremor detection [86]. Furthermore, other studies have suggested the thalamo-basal ganglia circuit as the origin of tremor oscillations [87].
The model configuration required to produce a dominant GPe-STN-GPe loop resonance involves both an increase in STN-GPe coupling vp2ζ, with respect to previous parameter estimates [63], as well as an increase in cortico-STN coupling νζe. This is consistent the findings of a recent conductance-based modelling study where cortical inputs amplified parkinsonian oscillations generated by the subthalamo-pallidal circuit [47], although the frequencies observed in that study were lower, 8–14 Hz, and represent dopamine depleted states of primates [45].
In this section the CTBG system is numerically simulated using parkinsonian parameters defined in Table 1. These parameters yield strong GPe-STN and hyperdirect loop resonances, which results in large amplitude ∼26 Hz oscillations in STN activity.
In Fig 4(a) parkinsonian ∼26 Hz STN activity is simulated for 30 s and then 150 Hz DBS is applied. Following the application of this stimulation, a damping of the ∼26 Hz oscillation is observed. A comparison of STN power spectrums pre-stimulus and during stimulation is given in Fig 4(b) and this shows peak power is reduced as a result of DBS.
Fig 5(a) demonstrates that increasing DBS pulse frequency strengthens the suppression of 13-30 Hz STN activity. As discussed in previous sections, coupling a DBS input to any population in the model results in an effective constant perturbation to the membrane potential Va of that population. Because DBS is coupled to the STN, GPe, and GPi populations via νζx = −1.2 mVs and v p 1 x , v p 1 x = 1.2 mVs, each corresponding mean membrane potential is perturbed by |ΔV|. This perturbation is −ΔV (inhibitory) for the STN population, and +ΔV (excitatory) for the GPi and GPe populations. In Fig 5(b) we compare power suppression at 13–30 Hz for two cases: In the first case, a direct constant perturbation is made to the membrane potentials of the STN, GPi, and GPe populations. In the second case, the stimulus input used to produce Fig 5(a) is convolved with an impulse response function, as discussed in a previous section. This allows the DBS input to be approximately represented as a constant perturbation to the membrane potential of a given population. Fig 5(b) shows how peak power between 13-30 Hz is effected by directly perturbing the mean membrane potential for the STN, GPi, and GPe populations relative to indirectly perturbing them with an oscillating DBS input. The suppression of pathological beta activity by DBS in our model is then largely attributable to this effective perturbation to the mean membrane potential.
Fig 5(a) and 5(b) also show constructive wave interactions for stimulus pulse frequencies equal to the 26 Hz beta oscillation and destructive interactions near the beta peak and its harmonic (52 Hz) and subharmonic (13 Hz). Studies have shown low-frequency stimulation may worsen PD motor symptoms [88, 89] as well as improve them [90].
The dependence of key network gains on the DBS pulse frequency is shown in Fig 6(a) and 6(b). The parkinsonian parameters define a pathologically strong STN-GPe-STN loop gain as well as a strong hyperdirect pathway. As the pulse frequency of the DBS inputs increases, so too does the net inhibition in the system. This is due to DBS inputs activating the inhibitory pallidal populations (GPe and GPi) more strongly. In contrast, the remaining DBS input inhibits the STN population, which is critical to the generation of a ∼ 26 Hz resonance.
It is important to note that the same suppressive effect can be achieved for a lower stimulus frequency if the stimulus amplitude is correspondingly increased. In DBS treatments, using larger signal amplitudes has the potential to increase the area directly affected by the applied stimulation, possibly incorporating non-motor projecting segments of the STN or even adjacent populations. Our results demonstrate that a high stimulus pulse frequency (fstim > 100 Hz) is necessary for beta suppression when the signal amplitude is constrained to be small relative to other STN inputs, e.g., for the Table 1 parameters DBS constitutes ∼6% of the connection weighted activity arriving at the STN over a time interval greater than several stimulus pulse widths.
Fig 7 shows the power spectrum of the STN firing rate time series as a function of DBS pulse frequency. Strong oscillations are seen at ∼26 Hz and its second harmonic ∼52 Hz. When the stimulus pulse frequency reaches about 140 Hz the ∼26 Hz power decreases by several orders of magnitude. Overall power within the 0–60 Hz frequency band also decreases and is redistributed to higher frequencies.
Additionally, Fig 7 shows peaks at the stimulus frequency fstim (1:1) and also at its harmonics fstim(N:1), however, the harmonics are much weaker and not clearly discernible in this plot.
Nonlinear wave interactions have previously been demonstrated in a neural field model of the corticothalamic system [91] which shows good agreement with human EEG studies [92]. In [91], a periodic nonlinear input was used to drive the CT system and resulted in nonlinear interactions between the drive frequency and an intrinsic alpha oscillation. Spectral peaks were found at frequencies equal to the sum and difference of the drive frequency and the intrinsic alpha frequency, f± = |fstim ± fα|, as well their respective harmonics. Our model also demonstrates these nonlinear interactions. In Fig 7, spectral peaks are seen at the sum and difference of the stimulus pulse frequency and the beta frequency f± = |fstim ± fβ| where fβ = 26 Hz. These nonlinear interactions are much more distinct in Fig 8 with peaks seen at f± = −fstim + 2fβ, 2fstim − fβ, and −2fstim + 3fβ. Additionally, Fig 8 demonstrates an entrainment of STN activity as a result of DBS inputs. The intrinsic parkinsonian beta peak is shifted to match the stimulus pulse frequency within the 25.5-26.2 Hz range. This result is consistent with experimental findings in human PD studies where a local entrainment of neural activity was observed during GPi-DBS [23].
In this work we have developed a novel description of deep brain stimulation and incorporated it into a neural field model of the corticothalamic-basal ganglia system. The model has enabled us to explore generative mechanisms for the pathological beta band activity observed in Parkinson’s disease and the influences of DBS on these oscillations. The main results of the paper are as follows:
Overall, our work provides insights into the generative mechanisms of pathological oscillations in human Parkinson’s disease and the population level effects of deep brain stimulation upon these oscillations. Furthermore, the model provides a framework for predicting effective stimulus protocols systematically rather than by trail and error, as has been the case to date.
Closed-loop adaptive DBS systems use feedback from local field potential measurements made via the implanted simulation electrode to modulate stimulus protocols [94]. Our model could be used in conjunction with an adaptive DBS system to increase the efficacy of clinical treatments. Cortical and subthalamic firing rate spectra in this model could be fitted to EEG and LFP spectra during an on-off DBS treatment cycle. The change in spectra corresponds to specific variable changes in the model and the trajectory of these changes could then be used as a detection method for parkinsonian states that are specific to the patient.
Several studies have observed antidromic activation as a result of deep brain stimulation [95], and activation of pallido-thalamic fibers during STN-DBS [96], which could be included in future generations of the model.
|
10.1371/journal.pcbi.1004920 | Learning from Heterogeneous Data Sources: An Application in Spatial Proteomics | Sub-cellular localisation of proteins is an essential post-translational regulatory mechanism that can be assayed using high-throughput mass spectrometry (MS). These MS-based spatial proteomics experiments enable us to pinpoint the sub-cellular distribution of thousands of proteins in a specific system under controlled conditions. Recent advances in high-throughput MS methods have yielded a plethora of experimental spatial proteomics data for the cell biology community. Yet, there are many third-party data sources, such as immunofluorescence microscopy or protein annotations and sequences, which represent a rich and vast source of complementary information. We present a unique transfer learning classification framework that utilises a nearest-neighbour or support vector machine system, to integrate heterogeneous data sources to considerably improve on the quantity and quality of sub-cellular protein assignment. We demonstrate the utility of our algorithms through evaluation of five experimental datasets, from four different species in conjunction with four different auxiliary data sources to classify proteins to tens of sub-cellular compartments with high generalisation accuracy. We further apply the method to an experiment on pluripotent mouse embryonic stem cells to classify a set of previously unknown proteins, and validate our findings against a recent high resolution map of the mouse stem cell proteome. The methodology is distributed as part of the open-source Bioconductor pRoloc suite for spatial proteomics data analysis.
| Sub-cellular localisation of proteins is critical to their function in all cellular processes; proteins localising to their intended micro-environment, e.g organelles, vesicles or macro-molecular complexes, will meet the interaction partners and biochemical conditions suitable to pursue their molecular function. Therefore, sound data and methods to reliably and systematically study protein localisation, and hence their mis-localisation and the disruption of protein trafficking, that are relied upon by the cell biology community, are essential. Here we present a method to infer protein localisation relying on the optimal integration of experimental mass spectrometry-based data and auxiliary sources, such as GO annotation, outputs from third-party software, protein-protein interactions or immunocytochemistry data. We found that the application of transfer learning algorithms across these diverse data sources considerably improves on the quantity and reliability of sub-cellular protein assignment, compared to single data classifiers previously applied to infer sub-cellular localisation using experimental data only. We show how our method does not compromise biologically relevant experimental-specific signal after integration with heterogeneous freely available third-party resources. The integration of different data sources is an important challenge in the data intensive world of biology and we anticipate the transfer learning methods presented here will prove useful to many areas of biology, to unify data obtained from different but complimentary sources.
| Cell biology is currently undergoing a data-driven paradigm shift [1]. Molecular biology tools, imaging, biochemical analyses and omics technologies, enable cell biologists to track the complexity of many fundamental processes such as signal transduction, gene regulation, protein interactions and sub-cellular localisation [2]. This remarkable success, has resulted in dramatic growth in data over the last decade, both in terms of size and heterogeneity. Coupled with this influx of experimental data, databases such as UniProt [3] and the Gene Ontology [4] have become more information rich, providing valuable resources for the community. The time is ripe to take advantage of complementary data sources in a systematic way to support hypothesis- and data-driven research. However, one of the biggest challenges in computational biology is how to meaningfully integrate heterogeneous data; transfer learning, a paradigm in machine learning, is ideally suited to this task.
Transfer learning has yet to be fully exploited in computational biology. To date, various data mining and machine learning (ML) tools, in particular classification algorithms have been widely applied in many areas of biology [5]. A classifier is trained to learn a mapping between a set of observed instances and associated external attributes (class labels) which is subsequently used to predict the attributes on data with unknown class labels (unlabelled data). In transfer learning, there is a primary task to solve, and associated primary data which is typically expensive, of high quality and targeted to address a specific question about a specific biological system/condition of interest. While standard supervised learning algorithms seek to learn a classifier on this data alone, the general idea in transfer learning is to complement the primary data by drawing upon an auxiliary data source, from which one can extract complementary information to help solve the primary task. The secondary data typically contains information that is related to the primary learning objective, but was not primarily collected to tackle the specific primary research question at hand. These data can be heterogeneous to the primary data and are often, but not necessarily, cheaper to obtain and more plentiful but with lower signal-to-noise ratio.
There are several challenges associated with the integration of information from auxiliary sources. Firstly, if the primary and auxiliary sources are combined via straightforward concatenation the signal in the primary can be lost through dilution with the auxiliary due to the latter being more plentiful and often having lower signal-to-noise ratio (see Fig H in S5 File for an illustration). Feature selection can be used to extract the attributes with the most distinct signals, however the challenge still remains in how to combine this data in a meaningful way. Secondly, combining data that exist in different data spaces is often not straightforward and different data types can be sensitive to the classifier employed, in terms of classifier accuracy.
In one of the first applications of transfer learning Wu and Dietterich [6] used a k-nearest neighbours (k-NN) and support vector machine (SVM) framework for plant image classification. Their primary data consisted of high-resolution images of isolated plant leaves and the primary task was to determine the tree species given an isolated leaf. An auxiliary data source was available in the form of dried leaf samples from a Herbarium. Using a kernel derived from the shapes of the leaves and applying the transfer learning (TL) framework [6], they showed that when primary data is small, training with auxiliary data improves classification accuracy considerably. There were several limitations in their methods: firstly, the data in the k-NN TL classifier were only weighted by data source and not on a class-by-class basis, and, secondly in the SVM framework both data sources were expected to have the same dimensions and lie in the same space. We present an adaption and significant improvement of this framework and extend the usability of the method by (i) incorporating a multi-class weighting schema in the k-NN TL classifier, and (ii) by allowing the integration of primary and auxiliary data with different dimensions in the SVM schema to allow the integration of heterogeneous data types. We apply this framework to the task of protein sub-cellular localisation prediction from high resolution mass spectrometry (MS)-based data.
Spatial proteomics, the systematic large-scale analysis of a cell’s proteins and their assignment to distinct sub-cellular compartments, is vital for deciphering a protein’s function(s) and possible interaction partners. Knowledge of where a protein spatially resides within the cell is important, as it not only provides the physiological context for their function but also plays an important role in furthering our understanding of a protein’s complex molecular interactions e.g. signalling and transport mechanisms, by matching certain molecular functions to specific organelles.
There are a number of sources of information which can be utilised to assign a protein to a sub-cellular niche. These range from high quality data produced from experimental high-throughput quantitative MS-based methods (e.g. LOPIT [7] and PCP [8]) and imaging data (e.g. [9]), to freely available data from repositories and amino acid sequences. The former, in a nutshell, involves cell lysis followed by separation and fractionation of the subcellular structures as a function of their density, and then selecting a set of distinct fractions to quantify by mass-spectrometry. These quantitative protein profiles are representative of organelle distribution and hence are indicative of their subcellular localisation [10]. Based on the distribution of a set of known genuine organelle marker proteins, pattern recognition and ML methods can be used to match and associate the distributions of unknown residents to that of one of the markers. There is thus a reliance on reliable organelle markers and statistical learning methods for robust proteome-wide localisation prediction [11]. These approaches have been utilised to gain information about the sub-cellular location of proteins in several biological examples, such as Arabidopsis [7, 12–16], Drosophila [17], yeast [18], human cell lines [19, 20], mouse [8, 21] and chicken [22], using a number of algorithms, such as, SVMs [23], k-NN [15], random forest [24], naive Bayes [14], neural networks [25], and partial-least squares discriminant analysis [7, 17, 22].
Although the application of computational tools to spatial proteomics is a recent development, the determination of protein localisation using in silico data such as amino acid sequence features (e.g. [26–40]), functional domains (e.g. [41, 42]), protein-protein interactions (e.g. [43–45]) and the Gene Ontology (GO) [4] (e.g. [46–49]) is well-established (reviewed in [50–52]). One may question the biological relevance and ultimate utility to cell biology of such predictors as protein sequences and their annotation do not change according to cellular condition or cell type, whereas protein localisation can change in response to cellular perturbation. Notwithstanding the inherent limitations of using in silico data to predict dynamic cell- and condition-specific protein location, transfer learning [6, 47–49, 53] may allow the transfer of complementary information available from these data to classify proteins in experimental proteomics datasets.
Here, we present a new transfer learning framework for the integration of heterogeneous data sources, and apply it to the task of sub-cellular localisation prediction from experimental and condition-specific MS-based quantitative proteomics data. Using the k-NN and SVM algorithms in a transfer learning framework we find that when given data from a high quality MS experiment, integrating data from a second less information rich but more plentiful auxiliary data source directly in to classifier training and classifier creation results in the assignment of proteins to organelles with high generalisation accuracy. Five experimental MS LOPIT datasets, from four different species, were employed in testing the classifiers. We show the flexibility of the pipeline through testing four auxiliary data sources; (1) Gene Ontology terms, (2) immunocytochemistry data [9], (3) sequence and annotation features, and (4) protein-protein interaction data [54]. The results obtained demonstrate that this transfer learning method outperforms a single classifier trained on each single data source alone and on a class-by-class basis, highlighting that the primary data is not diluted by the auxiliary data. This methodology forms part of the open-source open-development Bioconductor [55] pRoloc [56] suite of computational methods available for organelle proteomics data analysis.
Here, we have adapted a classic application of inductive transfer learning (TL) [6] using experimental quantitative proteomics data as the primary source and Gene Ontology Cellular Compartment (GO CC) terms as the auxiliary source. Using this TL approach, we have exploited auxiliary data to improve upon the protein localisation prediction from quantitative MS-based spatial proteomics experiments using (1) a class-weighted k-NN classifier, and (2) an SVM classifier in a TL framework. We also show the flexibility of the framework by using data from the Human Protein Atlas [57] and input sequence and annotation features from the YLoc [58, 59] web server, and protein-protein interaction data from the STRING database [54] as auxiliary data sources.
To assess classifier performance we employed the classic machine learning schema of partitioning our labelled data into training and testing sets, and used the testing sets to assess the strength of our classifiers. Parameter optimisation was conducted on the labelled training data using 100 rounds of stratified 80/20 partitioning, in conjunction with 5-fold cross-validation in order to estimate the free parameters via a grid search, as implemented in the pRoloc package [56] (and described in the methods below). Here, for the k-NN TL algorithm these parameters are the weights assigned to each class for each data source, and for the SVM TL algorithm these are C, γP and γA for the two kernels, as described in the materials and methods. The testing set is then used to assess the generalisation accuracy of the classifier. By applying the best parameters found in the training phase on test data, observed and expected classification results can be compared, and then used to assess how well a given model works by getting an estimate of the classifier’s ability to achieve a good generalisation, that is given an unknown example predict its class label with high accuracy. This schema was repeated for all 5 datasets, and for the SVM and k-NN classifiers, trained on (i) LOPIT alone, and (ii) GO CC alone, for comparison with the TL algorithms.
For simplicity, throughout this manuscript we refer to the mouse pluripotent embryonic stem cell dataset as the ‘mouse dataset’, the human embryonic kidney fibroblast dataset as the ‘human dataset’, the Drosophila embryos dataset as the ‘fly dataset’, the Arabidopsis thaliana callus dataset as the ‘callus dataset’ and finally the second Arabidopsis thaliana roots dataset, as the ‘roots dataset’.
The median macro-F1 scores for the mouse, human, callus, roots and fly datasets were 0.879, 0.853, 0.863, 0.979, 0.965, respectively, for the combined k-NN transfer learning approach. A two sample t-test showed that over 100 test partitions, the mean estimated generalisation performance for the k-NN transfer learning approach was significantly higher than on profiles trained solely from only primary or auxiliary alone for the mouse (p = 2e−21 for primary alone and p = 7e−78 for auxiliary alone), human (p = 1e−7 for primary alone and p = 8e−32 for auxiliary alone), plant roots (p = 4e−17 for primary alone and p = 4e−22 for auxiliary alone), and fly (p = 3e−5 for primary alone, p = 1e−112 for auxiliary alone) data (Fig 1). We found that the plant callus dataset did not significantly benefit (nor detrimentally affected) by the incorporation of auxiliary data. This was unsurprising as this dataset is extremely well-resolved in LOPIT (Fig A in S1 File, top right) and the median macro F1-score over 100 rounds of training and testing with a baseline k-NN classifier resulted in a median macro F1-score of 0.985 (the combined approach yielded a macro F1-score of 0.973).
The k-NN transfer learning classifier uses optimised class weights to control the proportion of primary to auxiliary neighbours to use in classification. One advantage of this approach is the ability for the user to set class weights manually, allowing complete control over the amount of auxiliary data to incorporate. As previously described, the class weights can be set through prior optimisation on the labelled training data. Fig 2 shows the detailed results for the mouse dataset and the distribution of the 100 best weights selected over 100 rounds of optimisation are shown on the top left. We found the distribution of weights for each dataset reflected closely the sub-cellular resolution in each experiment. For example, for the experiment on the mouse dataset the distribution of best weights identified for the endoplasmic reticulum (ER), mitochondria and chromatin niches are heavily skewed towards 1 indicating that the proportion of neighbours to use in classification should be predominantly primary. Note, as described in the methods if the class weight is assigned to 1, then strictly only neighbours in primary data are used in classification and similarly, if the class weight is 0 then all weight is given to the auxiliary data. If the weight falls between these two limits the neighbours in both the primary and auxiliary data sources is considered. From examining the principal components analysis (PCA) plot (Fig 2, top right) we indeed found that these organelles are well separated in the LOPIT experiment. Conversely, we found that the 40S ribosome overlaps somewhat with the nucleus (non-chromatin) cluster (Fig 2, top right) which is reflected in the best choice of class weights for these two niches; they were both assigned best weights of 1/3 and their weight distributions are skewed towards 0 indicating that more auxiliary data should be used to classify these sub-cellular classes. If we further examine the class-F1 scores for these two sub-cellular niches (Fig 2, bottom) we indeed find that including the auxiliary data in classification yields a significant improvement in generalisation accuracy (p = 1e−16 for 40S ribosome (red) and p = 1e−10 for the nucleus (non-chromatin) (pink)). We also found this to be the case for the proteasome, which is overlapping with the cytosol. We found LOPIT alone did not distinguish between these two sub-cellular niches in this particular experiment, however, the addition of auxiliary data from the Gene Ontology resulted in a significant increase in classifier prediction (p = 2e−16) as shown by the class-specific box plot in Fig 2, bottom (black).
Many experiments are specifically targeted towards resolving a particular organelle of interest (e.g. the TGN in the roots dataset) which requires careful optimisation of the LOPIT gradient. In such a setup sub-cellular niches other than the one of interest may not be well-resolved which may simply be due to the fact that the gradient was not optimised for maximal separation of all sub-cellular niches, but only one or a few particular organelles. Such experiments in particular may benefit from the incorporation of auxiliary data. We found that for the roots dataset all sub-cellular classes, except the TGN sub-compartment, benefitted from including auxiliary data (Fig C in S1 File, bottom), highlighting the advantage of using more than one source of information for sub-cellular protein classification. The best weight for the TGN was found to be 1 (Fig C in S1 File, top left), as expected and indicating high resolution in LOPIT for this class. In this framework we are able to resolve different niches in the data according to different data sources, further highlighted in the class-specific boxplots in Figs A-D in S1 File.
Adapting Wu and Dietterich’s classic application of transfer learning [6] we have implemented an SVM transfer learning classifier that allows the incorporation of a second auxiliary data source to improve upon the classification of experimental and condition-specific sub-cellular localisation predictions. The method employs the use of two separate kernels, one for each data source. As previously described, to assess the generalisation accuracy of our classifier we employed the classic machine learning schema of partitioning our labelled data into training and testing sets, and used the testing sets to assess the strength of our classifiers. This was repeated on 100 independent partitions for (i) the SVM TL method, (ii) a standard SVM trained on LOPIT alone, and (iii) a standard SVM trained on GO CC alone.
For the SVM TL experiments the resultant median macro-F1 scores for the mouse, human, callus, roots and fly datasets were 0.902, 0.868, 0.956, 0.875, 0.961, respectively, over the 100 partitions. As per the k-NN TL, we found the macro-F1 scores for the SVM TL Fig A in S2 File) were significantly higher than on profiles trained solely from only primary or auxiliary alone; mouse (p = 5e−56 for primary alone and p = 6e−37 for auxiliary alone), human (p = 7e−3 for primary alone and p = 1e−21 for auxiliary alone), callus (p = 4e−3 for primary alone and p = 1e−92 for auxiliary alone), roots (p = 2e−45 for primary alone and p = 7e−25 for auxiliary alone), and fly (p = 3e−3 for primary alone and p = 4e−105 for auxiliary alone) data. This was also evident on the organellar level as seen in the supporting figures in the S2 File.
One of the advantages of the transfer learning framework is the flexibility to use different types of information for both the primary and auxiliary data source. We demonstrate the flexibility of this framework by testing other complementary sources of information as an auxiliary data source.
We applied the two transfer learning classifiers to a real-life scenario, using the E14TG2a mouse stem cell dataset as our use-case to (i) demonstrate algorithm application, and (ii) highlight the applicability of the method for predicting protein localisation in MS-based spatial proteomics data over other single-source classifiers.
We compared the macro- and class-F1 scores from all experiments on all 5 datasets used to assess the classifier performance of the k-NN TL and SVM TL methods. We found that no single method systematically outperformed the other, as described in S5 File.
When applying the SVM TL and k-NN TL classifiers to the unlabelled proteins (see biological validation) an analysis of the final assignments (as classified based on a FDR of 5%) showed that the predicted protein localisations were in high agreement. Although there were no protein-organelle assignment mismatches between TL methods we did find a few cases where one TL method would assign a protein to one of the sub-cellular classes but the other TL method did not result in any organelle assignment, due to low classification scores (see Table C in S4 File). Overall, we did not find any contradicting sub-cellular class assignments.
We also compared Wu’s original k-NN algorithm against our TL methods. Wu’s method was significantly outperformed by our k-NN TL method for the mouse (p = 4e−4) and roots dataset (p = 4e−3) and by our SVM TL algorithm for the mouse (p = 7e−13), roots (p = 7e−8), and human (p = 0.004) datasets (see Figs F and G in S5 File).
In this study we have presented a flexible transfer learning framework for the integration of heterogeneous data sources for robust supervised machine learning classification. We have demonstrated the biological usage of the framework by applying these methods to the task of protein localisation prediction from MS-based experiments. We further show the flexibility of the framework by applying these methods to the five different spatial proteomics datasets, from four different species, in conjunction with four different auxiliary data sources to classify proteins to multiple sub-cellular compartments. We find the two different classifiers—the k-NN TL and SVM TL—perform equally well and importantly both of these methods outperform a single classifier trained on each single data source alone. We further applied the algorithm to a real-life use case, to classify a set of previously unknown proteins in a spatial proteomics experiment on mouse embryonic stem cells, which was validated using the most high resolution map of the mouse E14TG2a stem cell proteome produced to date [61]. We find integrating data from a second data source directly into classifier training and classifier creation results in the assignment of proteins to organelles with high generalisation accuracy. Finally, we find that using freely available data from repositories we can improve upon the classification of experimental and condition-specific protein-organelle predictions in an organelle-specific manner.
To our knowledge, no other method has been developed to date that allows the incorporation of an auxiliary data source for the primary task of predicting sub-cellular localisation in spatial proteomics experiments. In this study we have developed methods that not only allow the inclusion of an auxiliary data source in localisation prediction, but we have created a flexible framework allowing the use of many different types of auxiliary information, and furthermore allowing the user complete control over the weighting between data sources and between specific classes. This is extremely important for the analysis of biological data in general, and spatial proteomics data in particular, as many experiments are targeted towards resolving specific biologically relevant aspects (sub-cellular niches in spatial proteomics) and thus users may wish to control the impact of auxiliary information for aspects that have been specially targeted for analysis by the primary experimental method. In this context the setting of weights manually in the k-NN transfer learning classifier allows users complete power to explicitly choose whether to call upon an auxiliary data source or simply use data from their own experiment, on an organelle-by-organelle basis.
The effectiveness of using databases as an auxiliary data source will depend greatly on abundance and quality of annotation available for the species under investigation. For example, human is a well-studied species and there is a large amount of information available in the Gene Ontology and Human Protein Atlas. Furthermore, some organelles are easier to enrich for and thus there exists much more information available to utilise as an auxiliary source on a organelle by organelle basis. The transfer learning methods we present here allow the inclusion of any type of auxiliary data, provided of course there is information available for the proteins under investigation.
The integration of auxiliary data sources is a double-edged sword. On the one hand, it can shed light on (i) the primary classification task by reinforcing weak patterns or (ii) complement the signal in the primary data. On the other hand however it is easy to dilute valuable signals in an expensive experiment by shadowing the uniqueness, and hence biological relevance of the experimental primary data when integration is not performed with care, a phenomenon coined negative transfer (see Fig H in S5 File). Thus one needs to be cautious with data integration in general and not overlook the biological relevance of the primary data. Here, we provide a solution to this issue by using transfer learning: the k-NN transfer learning classifier uses optimised class-specific weights so as not to penalise any strong signals in the primary, if no signal is found in the auxiliary; similarly, the SVM transfer learning method uses optimised data-specific gamma parameters for each data-specific kernel.
The transfer learning framework forms part of the open-source open-development Bioconductor pRoloc suite of computational methods available for organelle proteomics data analysis. Moreover, as the pipeline utilises the formal Bioconductor classes, different data types, for example from gene expression technologies among others, can be easily used in this framework. The integration of different data sources is one of the major challenges in the data intensive world of computational biology, and here we offer a flexible and powerful solution to unify data obtained from different but complimentary techniques.
Five datasets, from studies on Arabidopsis thaliana [7, 15], Drosophila embryos [17], human embryonic kidney fibroblast cells [20], and mouse pluripotent embryonic stem cells (E14TG2a) [56] were collected using the standard LOPIT approach as described by Sadowski et al. [12]. In the LOPIT protocol, organelles and large protein complexes are separated by iodixanol density gradient ultracentrifugation. Proteins from a set of enriched sub-cellular fractions are then digested and labelled separately with iTRAQ or TMT reagents, pooled, and the relative abundance of the peptides in the different fractions is measured by tandem MS. The number of measurements obtained per gradient occupancy profile (which comprises of a set of isotope abundance measurements) is thus dependent on the reagents and LOPIT methodology used.
The first Arabidopsis thaliana dataset [7] on callus cultures employed dual use of four isotopes across eight fractions and thus yielded 8 values per protein profile. The aim of this experiment was to resolve Golgi membrane proteins from other organelles. Gradient-based separation was used to facilitate this, including separating and discarding as much nuclear material as possible during a pre-centrifugation step, and carbonate washing of membrane fractions to remove peripherally associated proteins, thereby maximising the likelihood of assaying less abundant integral membrane proteins from organelles involved in the secretory pathway.
The second Arabidopsis thaliana dataset on whole roots is one of the replicates published by Groen et al. [15], which was set up to identify new markers of the trans-Golgi network (TGN). The TGN is an important protein trafficking hub where proteins from the Golgi are transported to and from the plasma membrane and the vacuole. The dynamics of this organelle are therefore complex which makes it a challenge to identify true residents of this organelle. For each replicate, sucrose gradient fractions were subjected to a carbonate wash to enrich for membrane proteins and four fractions were iTRAQ labelled. Following MS the resultant iTRAQ reporter ion intensities for the four fractions were normalised to six ratios and then each protein’s abundance was further normalised across its six ratios by sum. In Groen’s original experiment the iTRAQ quantitation information for common proteins between the three different gradients were concatenated to increase the resolution of the TGN [23].
The aim of the Drosophila experiment [17] was to apply LOPIT to an organism with heterogeneous cell types. Tan et al. [17] were particularly interested in capturing the plasma membrane proteome (personal communication). There was a pre-centrifugation step to deplete nuclei, but no carbonate washing, thus peripheral and luminal proteins were not removed. In this experiment four isotopes across four distinct fractions were implemented and thus yielded four measurements (features) per protein profile.
The human dataset [20, 67] was a proof-of-concept for the use of LOPIT with an adherent mammalian cell culture. Human embryonic kidney fibroblast cells (HEK293T) were used and LOPIT was employed with 8-plex iTRAQ reagents, thus returning eight values per protein profile within a single labelling experiment. As in the LOPIT experiments in Arabidopsis and Drosophila, the aim was to resolve the multiple sub-cellular niches of post-nuclear membranes, and also the soluble cytosolic protein pool. Nuclei were discarded at an early stage in the fractionation scheme as previously described, and membranes were not carbonate washed in order to retain peripheral membrane and lumenal proteins for analysis.
The E14TG2a embryonic mouse dataset [56] also employed iTRAQ 8-plex labelling, with the aim of cataloguing protein localisation in pluripotent stem cells cultured under conditions favouring self-renewal. In order to achieve maximal coverage of sub-cellular compartments, fractions enriched in nuclei and cytosol were included in the iTRAQ labelling scheme, along with other organelles and large protein complexes as for the previously described datasets. No carbonate wash was performed.
For validation of the predicted localisations made using the transfer learning classifiers on the E14TG2a dataset above, a new high resolution mouse map was used as a gold standard [61]. This high resolution map was generated using hyperplexed LOPIT (hyperLOPIT), a novel technique for robust classification of protein localisation across the whole cell. The method uses an elaborate sub-cellular fractionation scheme, enabled by the use of Tandem Mass Tag (TMT) 10-plex and application of a recently introduced MS data acquisition technique termed synchronous precursor selection MS3 (SPS)-MS3 [68], for high accuracy and precision of TMT quantification. The study used state-of-the-art data analysis techniques [56, 67] combined with stringent manual curation of the data to provide a robust map of the mouse pluripotent embryonic stem cell proteome. The authors also provide a web interface to the data for exploration by the community through a dedicated online R shiny [69] application (https://lgatto.shinyapps.io/christoforou2015).
All datasets are freely distributed as part of the Bioconductor [55]pRolocdata data package [56].
Spatial proteomics relies extensively on reliable sub-cellular protein markers to infer proteome wide localisation. Markers are proteins that are defined as reliable residents and can be used as reference points to identify new members of that sub-cellular niche. Here, marker proteins are selected by domain experts through careful mining of the literature. Markers for each LOPIT experiment were specific to the system under study and conditions of interest and are distributed as part of the Bioconductor [55] pRoloc package [56].
The primary MS-based experimental datasets P consist of multivariate protein profiles. The auxiliary data A is a presence/absence binary matrix of Gene Ontology Cellular Compartment (GO CC) terms. Data are annotated to either (i) a single known organelle (labelled data), or (ii) have unknown localisation (unlabelled data). Thus we split P and A into labelled (L) and unlabelled (U) sections such that P = (LP, UP) and A = (LA, UA).
The labelled examples for P and A are represented by LP = {(xl, yl)|l = 1, …, |LP|} where x l ∈ R S, and LA = {(vl, yl)|l = 1, …, |LA|} where v l ∈ R T. Thus each lth protein is described by vectors of S and T features (generally, S << T, for P and A respectively. Each dataset shares a common set of proteins that is annotated to one of the same yl ∈ C = {1, …, |C|} sub-cellular classes, where | C | ∈ N is the total number of sub-cellular classes. Unlabelled data, UP and UA are represented by UP = {xu|u = 1, …, |UP|} where x u ∈ R S and UA = {vu|u = 1, …, |UA|} where v u ∈ R T, respectively.
The labelled data for the ith organelle class, with Ni indicating the number of proteins for the ith organelle class, is given for P by g i P = { ( x , y ) ∈ L P | y = i } and for A by g i A = { ( v , y ) ∈ L A | y = i }. The labelled dataset of all available proteins over the |C| different sub-cellular classes is given for P by L P = ∪ i = 1 | C | g i P and for A by L A = ∪ i = 1 | C | g i A.
We adapt Wu and Dietterich’s [6] classic application of inductive transfer using experimental quantitative proteomics data as the primary source (P) and GO CC terms as the auxiliary source (A). We aim to exploit auxiliary data to improve upon the sub-cellular classification of proteins found in MS-based LOPIT experiments in an organelle-specific way, using the baseline k-nearest neighbours (k-NN) algorithm in a transfer learning framework.
In k-NN classification, an unknown example is classified by a majority vote of its labelled neighbours, with the example being assigned to the class most common among its k nearest neighbours. Independent of the transfer learning classifier we compute the best k for each data source for values k ∈ {3, 5, 7, 9, 11, 13, 15} through an initial 100 rounds of 5-fold cross-validation using each set of labelled training data for P and then independently for A (as implemented in pRoloc). We denote by kP the best k for P, and by kA the best k for A.
Having obtained the best k for each data source, the transfer learning algorithm works as follows. For the uth protein (xu,vu) we wish to classify in U, we start by finding the kP and kA labelled nearest neighbours for xu and vu in LP and LA, respectively. Denote these sets N u P and N u A. We then define the vectors p u T = ( p 1 u , … , p | C | u ) and q u T = ( q 1 u , … , q | C | u ) to contain counts for each class in the sets of nearest neighbours; that is,
p i u = | { ( x , y ) ∈ N u P | y = i } | q i u = | { ( v , y ) ∈ N u A | y = i } | .
For each protein, let p ^ u = p u / k P and q ^ u = q u / k A be normalized vectors with elements summing to 1 and representing the distribution of classes among the sets of nearest neighbours for each protein. Finally, let NN P = { p ^ u | u = 1 , . . . , | U P | } and NN A = { q ^ u | u = 1 , . . . , | U P | }.
To include both the primary and auxiliary data in the set of potential neighbours we took a weighted combination of the votes in NNP and NNA for each sub-cellular class. Class weights are defined by the parameter vector θT = (θ1, …, θ|C|) with values θ i ∈ { 0 , 1 3 , 2 3 , 1 } chosen by optimisation through a prior 100 independent rounds of 5-fold cross-validation on a separate training partition of the labelled data. For the uth unknown protein (xu,vu) in U, the voting scores for each class i ∈ C are calculated as
V ( i ) = θ i p ^ i u + ( 1 - θ i ) q ^ i u (1)
and the protein is assigned to the class c ∈ C maximizing V(i)
c = arg max i V ( i ) .
The class weights θi in Eq 1 control the relative importance of the two types of neighbours for each class i ∈ C. This differs from Wu and Dietterich’s [6] original approach as they only weight the data sources and not the classes and the data sources. In this paper we select each class weight θi from the set { 0 , 1 3 , 2 3 , 1 }; however, the algorithm allows us to use any real-valued θi ∈ [0, 1]. If θi = 1, then all weight is given to the primary data in class i and only primary nearest neighbours in class i are considered. Similarly, if θi = 0, then all weight is given to the auxiliary data in class i and only auxiliary nearest neighbours in class i are considered. If 0 < θi < 1 then a combination of neighbours in the primary and auxiliary data sources is considered.
In order to evaluate the generalisation accuracy of each transfer learning classifier we employed the following schema in all experiments. A set of LOPIT profiles labelled with known markers, and their counterpart auxiliary GO CC profiles, were separated at random into training (80%) and test (20%) partitions. The split was stratified, such that the relative proportions of each class in each of the two sets matched that of the complete set of data. The test profiles were withheld from classifier training and employed to test the generalisation accuracy of the trained classifiers. On each 80% training partition 5-fold stratified cross-validation was conducted to test all free parameters via a grid search and select the best set of parameters for each classifier. In each experiment, for each dataset, this process of 80/20% stratified splitting, training with 5-fold stratified cross-validation on the 80% and testing on the 20% was repeated 100 times in order to produce 100 sets of macro F1 scores and class-specific F1 scores. The F1 score (He [83]) is a common measure used to assess classifier performance. It is the harmonic mean of precision and recall, where
precision = tp tp + fp , recall = tp tp + fn
and tp denotes the number of true positives, fp the number of false positives, and fn the number of false negatives. Thus
F1 = 2 × precision × recall precision + recall .
A high macro F1 score indicates that the marker proteins in the test data set are consistently correctly assigned by the algorithm.
To assess whether incorporating an auxiliary data source into classifier training and classifier creation was better than using primary or auxiliary data alone, we conducted three independent experiments for each data source and for each transfer learning method. We used the above schema to assess the generalisation accuracy of using (1) the transfer learning k-Nearest Neighbours (k-NN) classifier, (2) the primary LOPIT data alone, using a baseline k-NN, (3) the auxiliary GO CC data alone, using a baseline k-NN. We repeated this for the lpSVM transfer learning classifier and used a standard SVM with an RBF kernel for single data source experiments. Using these experiments we were able to compare using a simple k-NN versus the transfer learning k-NN, and also the use of a standard SVM versus the combined transfer learning lpSVM approach.
A two-sample two-tailed t-test, assuming unequal variance, was used to assess whether, over the 100 test partitions, the estimated generalisation performance using the optimised class-specific fusion approach was better than using either primary data alone, or auxiliary data alone. A threshold of 0.01 was used in all t-tests to determine significance.
|
10.1371/journal.ppat.1001225 | Dimeric 2G12 as a Potent Protection against HIV-1 | We previously showed that broadly neutralizing anti-HIV-1 antibody 2G12 (human IgG1) naturally forms dimers that are more potent than monomeric 2G12 in in vitro neutralization of various strains of HIV-1. In this study, we have investigated the protective effects of monomeric versus dimeric 2G12 against HIV-1 infection in vivo using a humanized mouse model. Our results showed that passively transferred, purified 2G12 dimer is more potent than 2G12 monomer at preventing CD4 T cell loss and suppressing the increase of viral load following HIV-1 infection of humanized mice. Using humanized mice bearing IgG “backpack” tumors that provided 2G12 antibodies continuously, we found that a sustained dimer concentration of 5–25 µg/ml during the course of infection provides effective protection against HIV-1. Importantly, 2G12 dimer at this concentration does not favor mutations of the HIV-1 envelope that would cause the virus to completely escape 2G12 neutralization. We have therefore identified dimeric 2G12 as a potent prophylactic reagent against HIV-1 in vivo, which could be used as part of an antibody cocktail to prevent HIV-1 infection.
| Most successful vaccines function by eliciting antibodies that bind to the surface of pathogens of interest from the host immunologic repertoire. This should also be the case for an HIV-1 vaccine, but broadly neutralizing anti-HIV-1 antibodies have proven hard to elicit with any reagent. Thus, methods to directly administer broadly neutralizing anti-HIV-1 antibodies, such as passive transfusion, become appealing. It is therefore important to find out which antibodies, or antibody cocktails, would provide effective protection against HIV-1 before administering them. Here, we show that the dimeric fraction of 2G12, a unique monoclonal anti-HIV-1 antibody that dimerizes naturally, provides better protection against HIV-1 than its monomeric fraction. As an added bonus, although HIV-1 can evolve to completely escape antibody control, the 2G12 dimer does not favor such evolution. Our study suggests that the 2G12 dimer may be a suitable reagent for direct administration to protect people from HIV-1 infection.
| Human efficacy trials of vaccine candidates designed to elicit antibody-based immunity against HIV-1 have mostly failed [1], [2], raising questions as to whether such an approach to HIV-1 vaccination is at all feasible. A recent human vaccine trial in Thailand [3], however, provided a promising signal of efficacy. While there is no direct evidence of which component of the vaccine was effective, it could be antibody-based immunity. In the trial, 98.6% of vaccinated individuals produced “binding antibodies” against HIV-1 envelope protein gp120 although no broadly neutralizing antibodies. The possibility that antibody-mediated protection was effective has reenergized the search for effective anti-HIV-1 antibodies.
Existing broadly neutralizing anti-HIV-1 antibodies are valuable starting points for generating protection against HIV-1. Several broadly neutralizing antibodies have been proposed as the basis for designing protective mechanisms against HIV-1 in recent years [4], [5]. Among them, 2G12 is unique, because it recognizes a constellation of carbohydrates on gp120 [6], [7], [8], [9] and has an unusual structure that involves a domain swap between the two heavy chains [8]. 2G12 is most effective at neutralizing clade B strains of HIV-1 [10].
A series of studies have described the in vivo protective effects of 2G12 against simian/HIV-1 in macaques [11], [12], [13] and against HIV-1 in humans [14], [15], [16], [17]. Interestingly, in the studies where 2G12 was combined with other broadly neutralizing antibodies such as 4E10 and 2F5 [16], [17], 2G12 provided the dominant protective effect against HIV-1. The relatively long in vivo half-life of 2G12 can partially explain this phenomenon [18]. However, albeit protective, 2G12 also selected HIV-1 escape mutants in vivo [16], [19]; therefore, it is important to identify a new reagent or method to minimize the rate of appearance of such escape mutants.
We have previously shown that 2G12 IgG1 can form natural dimers that are 50-80–fold more potent than monomeric 2G12 IgG1 in in vitro neutralization of various strains of HIV-1 [20]. 2G12 monomer, in common with typical IgGs, contains two antigen-binding Fabs and one Fc region, but the heavy chain regions of the Fabs are domain-swapped to create a single (Fab)2 unit [8]. 2G12 dimer contains four Fabs and two Fcs, which form a structure, presumably through inter-molecular domain swapping, that does not interconvert with 2G12 monomer [20]. The present study was designed to investigate the in vivo potency of dimeric 2G12 in controlling HIV-1 infection in a humanized mouse model. We show that dimeric 2G12 is effective at providing protection against HIV-1 without selecting viral mutants that would completely escape 2G12 neutralization, suggesting that the 2G12 dimer is a suitable prophylactic reagent for use against HIV-1.
The dimeric form of the monoclonal antibody 2G12 possesses increased in vitro neutralization potency compared to the monomeric form [20]. It is unknown, however, whether dimeric 2G12 would have a long enough half-life to be more effective than the 2G12 monomer at preventing HIV-1 infection in vivo. To address this question, we prepared separate stocks of purified 2G12 monomer and dimer and passively transferred 0.5 mg/mouse of 2G12 monomer or dimer into Rag2−/−γc−/− mice reconstituted with human immune cells (Supporting Figure S1A). We then challenged the mice intravenously (i.v.) with the CCR5-tropic strain of HIV-1, JR-CSF, at a dose of 400 ng of p24.
Using an ELISA targeting a Myc tag fused to the light chain of the purified antibodies, we found that the concentration of 2G12 monomer declined quickly in the mouse plasma whereas the 2G12 dimer was relatively stable (Figure 1A). The elimination (β phase) half-lives of the purified human IgGs in the humanized mouse plasma were estimated as 3.5±0.9 days for the 2G12 dimer and 0.9±0.2 days for the 2G12 monomer. The 2G12 dimer prevented CD4 T cell loss in the peripheral blood following HIV-1 infection, whereas the 2G12 monomer did not provide protection (Figure 1B). In addition, the 2G12 dimer moderately suppressed the increase of viral load in the blood (Figure 1C), causing an overall reduction of 97.5% in viral load compared to the control that lacked antibody (Figure 1D). The 2G12 monomer, on the other hand, did not suppress the increase of viral load following HIV-1 infection (Figure 1C). We also analyzed the percentages of T cells and the numbers of p24+ cells in the spleen, thymus, and mesenteric lymph node. As shown in Figure 1E, we found that without 2G12, HIV-1 almost completely depleted CD4+ cells in the spleen. The percentage of splenic CD8+ cells also decreased, presumably because they rely on CD4+ T helper cells for proliferation and survival [21]. Between the two forms of 2G12, the monomer had a minimal effect at preventing the loss of CD4+ and CD8+ splenocytes following HIV-1 infection, whereas the 2G12 dimer was able to rescue nearly half of the CD4+ cells and most of CD8+ cells in the spleen (Figure 1E). A similar effect was observed in the mesenteric lymph node (Figure S1B) but not in the thymus (Figure S1C), presumably because a CCR5-tropic virus was used and there are few CCR5+ T cells in the thymus [22]. Immunohistochemical analysis using an antibody against HIV-1 p24 confirmed that the 2G12 dimer was effective at limiting HIV-1 infection in both the spleen and the mesenteric lymph node (Figure 1F). HIV-1 p24+ cells were hardly found in the thymus (data not shown).
These results demonstrated increased protection against HIV-1 of purified 2G12 dimer compared to 2G12 monomer when the antibodies were administered to humanized mice prior to HIV-1 challenge.
To investigate whether the higher potency of 2G12 dimer compared to the monomer resulted only from its longer in vivo half-life, we modified the conventional humanized mice [23], [24] to carry antibody-expressing cells as backpacks [25] that produced antibodies continuously throughout the course of HIV-1 infection (Figure 2). This strategy avoided the dramatic fluctuation of antibody concentrations that usually occur when antibodies were administered through multiple administrations [16], [17]. The antibody-expressing cells were injected subcutaneously (s.c.) and formed localized backpacks whose size could be controlled by the administration of ganciclovir, a prodrug that killed backpacked cells co-expressing herpes simplex virus thymidine kinase (TK) along with the antibody [26]. Because the backpack size positively correlated with the concentration of 2G12 in the blood (Figure S2A), we could control the backpack size to limit the antibody concentration within a reasonably small range.
We made mice with backpacks that expressed wild-type 2G12 (named “2G12 BP”) and those with backpacks expressing D2, a mutant of 2G12 that is expressed with an increased dimer/monomer ratio [20] (named “D2 BP”). We previously reported that wild-type 2G12 cells produce 78% monomer and 22% dimer whereas the D2 clone produced 60% monomer and 40% dimer; and that the monomers and dimers produced by wild-type 2G12 or D2 2G12 exhibited no significant differences in biophysical and neutralization characteristics [20]. Since the 2G12 monomer and dimer share the same heavy and light chains, an ELISA would not distinguish between the two forms, making it difficult to directly measure the dimer:monomer ratios in the backpacked mice. Size exclusion chromatography, which could be normally used to determine relative levels of monomer and dimer, would require several milliliters of mouse blood for each sample collection, which was not feasible. Instead, we calculated the monomer:dimer ratios based on the production ratios of monomer versus dimer in the two cell lines (3.5: 1 for 2G12 BP and 1.5: 1 for D2 BP) and their individual half-lives in the humanized mice (see Materials and Methods for details). We then used the ratios to estimate the concentrations of 2G12 dimer and 2G12 monomer in the blood samples (Table 1; concentrations of total 2G12 and 2G12 dimer are shown; the concentration of 2G12 monomer can be obtained by subtracting the dimer concentration from the concentration of total 2G12).
The D2 BP provided an estimated 3-5-fold more dimer than the 2G12 BP during the first 3 weeks of HIV-1 infection (Table 1; p<0.02 for weeks 0, 1, 2 and 3). The concentrations of 2G12 monomer were not significantly different between the D2 BP and 2G12 BP groups at each time point (p>0.05) although the combined 2G12 concentrations were higher in the D2 BP group due to significantly greater dimer concentrations. Analysis of the peripheral blood lymphocytes showed that 2G12 BP barely had any protective effect on CD4 T cells compared to the control that lacked antibody (Figure 3A, weeks 1, 2 and 4). In contrast, D2 BP effectively protected CD4 T cells from being cleared by HIV-1 after one week of infection (Figure 3A, week 1; p<0.05). D2 BP also appeared to offer some protection for CD4 T cells 2 and 4 weeks after HIV-1 inoculation although the effect was not statistically significant. Analysis of HIV-1 copy numbers in the mouse plasma showed that D2 BP moderately suppressed the viral load at each time point (Figure 3B) and significantly suppressed the overall viral load (Figure S2B; p<0.01), suggesting that D2 BP is potent at preventing viral entry and/or eliminating HIV-1 from the circulation. The mice with D2 BP also had significantly lower numbers of p24+ cells in the mesenteric lymph node than mice carrying 2G12 BP (Figure 3C), although neither backpack significantly protected the spleen from HIV-1 infection (Figure 3D for the percentage of CD4 T cells and Figure S2C for the number of p24+ cells).
Since D2 BP did not completely prevent HIV-1 infection of humanized mice (i.e., HIV-1 viral load was still detectable in the mouse plasma), we asked if increasing the concentration of 2G12 to over 100 µg/ml [11], [12], [13] would provide better protection against HIV-1. Thus, we included a group of mice (named “BP”) that carried large wild-type 2G12 backpacks as a means to maintain both 2G12 monomer and 2G12 dimer at high concentrations in the peripheral blood (Table 1). Our results showed that the large backpacks prevented HIV-1-induced CD4 T cell loss in the peripheral blood (Figure 3A, weeks 1, 2, and 4), suppressed HIV-1 viral load in the mouse plasma (Figure 3B and Figure S2B), decreased the number of p24+ cells in the mesenteric lymph node (Figure 3C), and minimized the decrease of CD4 T cell percentage in the spleen (Figure 3D). However, the virus was still detectable in the periphery (Figure 3B). In fact, the overall viral load in BP mice was similar to that of D2 BP mice (Figure S2B), suggesting that the concentration of 2G12 dimer required to neutralize HIV-1 in vivo might be as low as 5–25 µg/ml (Table 1, dimer concentrations in the D2 BP group from week 0 to week 4), a level that led to over 70% neutralization of the virus (Figure 3B, comparing D2 BP to the control group lacking antibody). Providing 10-fold more of the 2G12 dimer could potentially prevent CD4 T cell loss in the peripheral blood for a longer period of time (Figure 3A), but it would not prevent HIV-1 entry or further decrease HIV-1 viral load in the plasma (Figure 3B and Figure S2B) or mesenteric lymph node (Figure 3C).
These results showed that a continuous supply of dimeric 2G12 at 5–25 µg/ml during the course of HIV-1 infection is effective at protecting humanized mice against HIV-1 infection.
Since 2G12 is known to induce HIV-1 escape mutants [16], [19], we extracted viral RNA from the week-4 plasma of 3 or 4 representative mice per experimental group, cloned the JR-CSF envelope gene from viral cDNA, and sequenced at least 10 clones per mouse sample. Some viral clones had spontaneous mutations at residue N339 regardless of the presence of 2G12 and might represent a background in the inoculum (Table 2 and Figure 4A). In addition, both 2G12 BP and D2 BP selected mutations at residue N386. Surprisingly, we observed an unusually high percentage of mutations at residue N295 when the 2G12 concentration was kept at 100 µg/ml or higher (Table 2 and Figure 4A; BP). This residue, along with N332 that was not significantly mutated in this study, have been suggested as the key anchors of glycans that form the 2G12 epitope [7]. To assess the sensitivity of mouse-derived viruses to 2G12 neutralization, we performed in vitro neutralization assays using pseudoviruses made with JR-CSF envelope genes that we obtained from mouse plasma samples. Both the input virus (the pseudovirus that shared the same JR-CSF envelope as the inoculum) and the virus with mouse-derived envelope that did not encounter any neutralizing antibody in vivo (HIV-1 only; No mutation) were effectively neutralized by 2G12 monomer and 2G12 dimer in vitro (Figure 4B and 4C); but the half maximal inhibitory concentration (IC50) of 2G12 dimer was 33-fold less than the IC50 of the monomer, suggesting that the 2G12 dimer was more potent at neutralizing the JR-CSF strain of HIV-1 than the 2G12 monomer. More importantly, we found that the viral envelope from a BP mouse with the mutation N295S caused the pseudovirus to completely escape the neutralization effect of both the 2G12 monomer (Figure 4B) and the 2G12 dimer (Figure 4C). In contrast, a virus variant with a mutation at residue 386 was partially neutralized by the 2G12 monomer and 2G12 dimer. This suggests that, unlike the >100 µg/ml condition (provided by BP), the presence of 2G12 dimer at 5–25 µg/ml (provided by D2 BP) did not select for complete HIV-1 escape mutants.
Therefore, our results showed that although constant administration of 2G12 at high concentrations was potent at protecting humanized mice from HIV-1 infection in vivo, it resulted in HIV-1 envelope mutations that could completely escape 2G12 neutralization. However, at least over the time-course of our experiments, a low level of 2G12 dimer did not specifically select the same mutations, providing an additional benefit to its high potency.
In this study, we used a humanized mouse model to investigate the in vivo potency of dimeric 2G12 in controlling HIV-1 infection. This mouse model supports human hematopoietic development, provides human CD4 T cells as natural targets of HIV-1 infection, and allows for possible selection of viral resistance [27]. Using these mice, we first examined the stability and protective effects of monomeric and dimeric forms of 2G12 in HIV-1-challenged humanized mice by passively transferring purified antibodies. We found that the 2G12 dimer had a longer in vivo half-life and was more potent than the 2G12 monomer at controlling HIV-1 infection in vivo. The elimination half-life of the 2G12 dimer was 3.5 days in humanized mice and comparable to the reported elimination half-life (3.2 days) of human IgG1 in mice [28]. This is shorter than the half-life of human IgG1 in humans [18] but correlates with the difference in body weight between mice and humans [29]. To investigate whether a continuous supply of the 2G12 monomer would overcome its poor in vivo efficacy, we next used a backpacking approach to provide the antibody continuously. Using wild-type 2G12 as the backpacked gene, we achieved a sustained level of 2G12 monomer and dimer in the mouse plasma. However, constant delivery of 2G12 monomer plus a small amount of 2G12 dimer at a low level (1–4 µg/ml dimer for the first 3 weeks and 16.6 µg/ml dimer after 4 weeks) did not protect the mice from HIV-1 infection. In contrast, backpacks containing the D2 mutant, which produced increased levels of 2G12 dimer (60% monomer, 40% dimer) provided effective protection against HIV-1 by maintaining a 2G12 dimer concentration of 5–25 µg/ml in the mouse plasma. Thus, our results suggest that, administered either through a single injection or continuously, dimeric 2G12 is a more potent prophylactic anti-HIV-1 antibody than 2G12 monomer.
Several in vivo studies have estimated that concentrations of 2G12 of 100 µg/ml or higher exert a protective effect against HIV-1 when the virus is given at a 50% tissue culture infective dose (TCID50) of 500—5,000 [11], [12], [13]. In order to establish a robust and consistent infection in humanized mice, we administered HIV-1 intravenously at a dosage of 400 ng p24, or a TCID50 of 400,000. Although sterilizing immunity was not achieved in this study, we found that, even with high-dose HIV-1 challenge, 2G12 monomer and dimer at combined concentrations of 100 µg/ml or higher could significantly reduce the severity of HIV-1 infection in the humanized mice (Figure 3). More importantly, the D2 BP that delivered 2G12 at a much lower concentration exerted a similar protective effect against HIV-1. In particular, D2 BP provided the 2G12 dimer at 5–25 µg/ml, which was sufficient to prevent peripheral blood CD4 T cell loss (Figure 3A) and suppress the increase of the viral load following HIV-1 infection (Figure 3B and Figure S2B). Therefore, 2G12 dimer represents a promising prophylactic reagent against HIV-1 in vivo because it neutralizes HIV-1 at a relatively low concentration.
Having a low effective concentration is not the only advantage of the 2G12 dimer as a protective reagent against HIV-1. 2G12 is known to select HIV-1 escape mutants both in vitro [30], [31] and in vivo [16], [30], [31], with in vivo escape mutants detectable as early as 4 weeks after HIV-1 inoculation [16], [30], [31]. Here we analyzed the diversity of HIV-1 viral RNA isolated from the mouse plasma, focusing on regions of the JR-CSF envelope gene where 2G12 epitope-containing carbohydrates would attach [7], [32]. We found that while low levels of 2G12 dimer induced mutations at residue N386, 2G12 at monomer plus dimer concentrations of >100 µg/ml specifically selected mutations at another residue (Table 2). This residue, N295, has been suggested to be one of the two central players in the interaction between 2G12 and its carbohydrate epitope [7]. A mutation at N295 would be more likely to allow HIV-1 to escape 2G12 neutralization than mutations at other sites such as N386 (Figure 4B and Figure 4C). Thus, at least over the time-course of our experiments, dimeric 2G12 provided protection against HIV-1 without selecting for complete HIV-1 escape mutants.
In summary, we found in the present study that dimeric 2G12, or the D2 mutant that increases the production of dimeric 2G12, might be potential prophylactic reagents against HIV-1. However, more research is necessary to characterize the tissue distribution of dimeric 2G12 and its in vivo antibody-dependent cellular cytotoxicity activity. It is also important to assess the immunogenicity of 2G12 in its dimeric form since it is twice the size of a typical IgG. In addition, the pharmacokinetics of dimeric 2G12 should be carefully established in human studies, as the half-life of the antibody in humans is likely to be different from that in humanized mice. Furthermore, because the neutralization spectrum of 2G12 is not particular good when tested against a large panel of HIV-1 isolates [10] and neutralizing antibodies have demonstrated synergy when combined together [33], the 2G12 dimer may be more beneficial when used as part of an antibody cocktail to protect people from HIV-1 infection.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee (IACUC) of California Institute of Technology (Animal Assurance Number: A3426-01). All animal experiments were conducted under IACUC protocols 1536-09G and 1547-08G.
The wild-type 2G12 heavy chain gene (IgG1) and a Myc-tagged 2G12 light chain gene were linked by an F2A sequence and subcloned into a lentivector. The vector is a third-generation, self-inactivating lentiviral vector backbone based on pHRST [34], [35]. Briefly, the StuI fragment of pHRST containing a complete viral genome was ligated into the pUC19 backbone to remove exogenous flanking genomic sequences. PCR-cloning was employed to introduce restriction sites flanking the promoter and transgene to facilitate subsequent cloning. Further modifications were made to pHAGE6 to remove extraneous viral sequences with no effect on virus function (A.B., to be published elsewhere). Lentiviruses were then generated by transient transfection of HEK-293T cells using the Trans-IT reagent (Mirus Bio; Madison, WI) and used to create a 293T stable cell line that produced 2G12. The 2G12-expressing, adherent stable cell line was adapted for growth in suspension for large-scale production of 2G12 at the Caltech Protein Expression Center. Cell culture supernatants were collected and passed over protein A resin (Pierce Biotechnology; Rockford, IL), and eluted using using pH 3.0 citrate buffer. Protein A eluates were immediately neutralized and then subjected to size exclusion chromatography in 20 mM Tris pH 8.0, 150 mM NaCl using a Superdex 200 16/60 (GE Healthcare). Fractions corresponding to monomer and dimer were collected and then separately passaged over a Superdex 200 10/30 column (GE Healthcare) to remove contaminating amounts of monomer or dimer from the separated purified species.
Frozen human cord blood CD34+ cells from single donors were purchased from AllCells (Emeryville, CA) or Lonza (Basel, Switzerland). One-day-old Rag2−/−γc−/− pups were irradiated and intrahepatically (i.h.) injected with 0.1-0.2×106 human cord blood CD34+ cells per pup. Mice were then screened for human CD45+ cells at 6 weeks of age and those with good reconstitution were chosen for the study (Figure S1A). For passive transfer experiments, one single dose of 0.5 mg/mouse of purified 2G12 dimer or 2G12 monomer was injected retro-orbitally (i.v.) into 4-month-old humanized mice 1 day before HIV-1 challenge. The HIV-1 JR-CSF plasmid was obtained from NIH AIDS Research and Reference Reagent Program and transiently transfected into 293T cells to produce infectious HIV-1 particles. The culture medium containing HIV-1 was then harvested and titered using the p24 ELISA kit from PerkinElmer (Waltham, MA). The virus was injected (i.v.) at 400 ng p24/mouse. For non-HIV-1 mice, conditioned medium was injected as the control. All mice involved in this study were age-matched since the CD4:CD8 ratio naturally increased with the age of these mice.
Wild-type 2G12 and D2 mutant genes were cloned into lentiviral vectors. Lentiviruses were then generated and used to create stable cell lines that produced wild-type 2G12 and D2, respectively. The parent cell line was a stable 293T cell line that expressed herpes simplex virus thymidine kinase (TK), so the progeny lines were named 293T/TK/2G12 and 293T/TK/D2 cell lines. When well-reconstituted humanized mice were 3-month-old, 1×106 of backpacked cells were injected (s.c.) on the back of the mice at the lower right side. Backpack size (length × width) was measured weekly and controlled by injection (i.p.) of 62.5 µg or 125 µg (depending on the backpack size) of ganciclovir (Sigma; St. Louis, MO) per mouse after HIV-1 challenge and when the backpack size reached 1.5 cm2.
Weekly blood samples were obtained retro-orbitally and the plasma was immediately separated from blood cells and stored for viral RNA extraction and Myc-specific ELISA (see below for details). The peripheral blood mononuclear cells after antibody staining were analyzed by the FACSCalibur (BD Biosciences; San Jose, CA). Mice were sacrificed 4 weeks after HIV-1 challenge. Blood, spleen, thymus, and mesenteric lymph node were collected for flow cytometry analysis or fixation in formalin. The fixed tissues were then send to University of California, Los Angeles for immunohistochemical analysis.
Mouse plasma was diluted 1∶10, 1∶100, and 1∶1000 in sample diluent and heat-inactivated at 55°C for 1 h. Myc-tagged 2G12 was captured by anti-human IgG-Fc (Bethyl Laboratories; Montgomery, TX) and detected by anti-Myc conjugated with horseradish peroxidase (Bethyl Laboratories; Montgomery, TX). The plates were read at 450 nm on a SpectroMax Reader (Molecular Devices, Sunnyvale, CA) after the addition of the TMB substrate and the stop solution. In passive transfer experiments, the half-life of the elimination phase (β phase), which took place after the redistribution phase, was determined using a one-phase exponential decay model using data points from week 0 (24 h after the injection of 2G12 monomer or dimer) to week 4. The half-lives were estimated as 3.5±0.9 days for the 2G12 dimer and 0.9±0.2 days for the 2G12 monomer. In backpacking experiments where both 2G12 monomer and dimer were present in the plasma, we determined their individual concentrations by calculating the monomer:dimer ratios as following: where P = protein (monomer or dimer), β = production rate; α = degradation rate. Assuming that at the time of HIV-1 challenge (4 weeks after backpack injection), the monomer and dimer had reached their individual steady state (i.e. ),
If the dimer had a production rate of β and a degradation rate of α, then the monomer should have a production rate of 3.5β (78% monomer versus 22% dimer produced from 2G12 backpacks) and a degradation rate of 3.9α (dimer:monomer ratio of half-lives 3.5/0.9 = 3.9) for 2G12 backpacks. Thus,
Therefore, the 2G12 monomer and dimer concentrations were calculated as:where Ptotal = total 2G12 concentration as measured by Myc-specific ELISA. For D2 backpacks, since the dimer's production rate was 1.5β (60% monomer versus 40% dimer produced from D2 backpacks) and the degradation rate stayed the same,
Viral RNA was extracted from mouse plasma using QIAamp Viral RNA Mini Kit from Qiagen (Valencia, CA). The RNA (200 ng) was reverse transcribed and quantified using the Taqman RNA-to-CT One-Step Kit (Applied Biosystems; Foster City, CA) and the Eppendorf Realplex real-time PCR system (Hauppauge, NY). The primers were designed to anneal to the pol region of the HIV-1 genome within the first intron, so that only unspliced viral RNA could be detected. The primer sequences were: forward primer, 5′-CAA TGG CAG CAA TTT CAC CA-3′; reversed primer, 5′-GAA TGC CAA ATT CCT GCT TGA-3′. The probe sequence was: 5′-/56-FAM/CCC ACC AAC AGG CGG CCT TAA CTG/36-TAMSp/-3′. HIV-1 RNA standard was generated using the Riboprobe T7/SP6 kit from Promega (Madison, WI) and the pGEM FL2 plasmid was provided by Dr. Dong Sung An at University of California, Los Angeles. The detection limit of the assay was 20,000 HIV-1 copies/ml mouse plasma.
Viral RNA was extracted from mouse plasma using QIAamp Viral RNA Mini Kit from Qiagen (Valencia, CA). The RNA (500 ng) was reverse transcribed and amplified using the SuperScript III One-Step RT-PCR System with Platinum Taq High Fidelity from Invitrogen (Carlsbad, CA). The primer sequences were: JR-CSF env forward primer, 5′-GGC AAT GAG AGT GAA GGG GAT CAG-3′; JR-CSF env reversed primer, 5′-CAT CTT ATA GCA AAG CCC TTT CCA AGC C-3′. The primers flanked the whole 2.5-kb envelope gene. The PCR product was then gel-purified and cloned into the TOPO vector using the TOPO XL PCR Cloning Kit from Invitrogen (Carlsbad, CA). More than 10 clones were picked for each RNA sample. The plasmids were then extracted and sent to sequencing at Laragen (Los Angeles, CA) or Sequetech (Mountain View, CA). The sequencing primer was 5′-GTC AGC ACA GTA CAA TGT ACA CAT GGA ATT AG -3′ and annealed upstream of the Asn residues that linked 2G12 epitope-containing carbohydrate chains [7]. Mutations at N295, N332, N339, N386, N392, N448 and adjacent Ser/Thr residues were then analyzed.
We used a previously described pseudovirus neutralization assay, which measures the reduction in luciferase reporter gene expression in the presence of 2G12 monomer or dimer following a single round of pseudovirus infection in TZM-bl cells [20]. Pseudoviruses were generated by cotransfection of 293T cells with an envelope expression plasmid and a replication-defective backbone plasmid. (For envelope expression, viral RNA was extracted from mouse plasma 4 weeks after HIV-1 challenge and reverse transcribed. The complete envelope gene was amplified from viral cDNA and the PCR product was then gel-purified and cloned into the pcDNA3 vector.) Each 2G12 protein was tested in triplicate with a 3-fold dilution series, and incubated with the pseudoviruses (250 infectious viral units per well) for 1 h at 37°C. After the incubation, 10,000 TZM-bl cells were added to each well, followed by incubation for 2 days. Cells were then lysed and assayed for luciferase expression by using Bright-Glo (Promega; Madison, WI) and a Victor3 luminometer (Perkin-Elmer; Waltham, MA).
|
10.1371/journal.pgen.1002555 | Interpreting Meta-Analyses of Genome-Wide Association Studies | Meta-analysis is an increasingly popular tool for combining multiple genome-wide association studies in a single analysis to identify associations with small effect sizes. The effect sizes between studies in a meta-analysis may differ and these differences, or heterogeneity, can be caused by many factors. If heterogeneity is observed in the results of a meta-analysis, interpreting the cause of heterogeneity is important because the correct interpretation can lead to a better understanding of the disease and a more effective design of a replication study. However, interpreting heterogeneous results is difficult. The standard approach of examining the association p-values of the studies does not effectively predict if the effect exists in each study. In this paper, we propose a framework facilitating the interpretation of the results of a meta-analysis. Our framework is based on a new statistic representing the posterior probability that the effect exists in each study, which is estimated utilizing cross-study information. Simulations and application to the real data show that our framework can effectively segregate the studies predicted to have an effect, the studies predicted to not have an effect, and the ambiguous studies that are underpowered. In addition to helping interpretation, the new framework also allows us to develop a new association testing procedure taking into account the existence of effect.
| Genome-wide association studies are an effective means of identifying genetic variants that are associated with diseases. Although many associated loci have been identified, those loci account for only a small fraction of the genetic contribution to the disease. The remaining contribution may be accounted by loci with very small effect sizes, so small that tens of thousands of samples are needed to identify them. Since it is costly to conduct a study collecting such a large sample, a practical alternative is to combine multiple independent studies in a single analysis called meta-analysis. However, many factors, such as genetic or environmental factors, can differ between the studies combined in a meta-analysis. These factors can cause the effect size of the causal variant to differ between the studies, a phenomenon called heterogeneity. If heterogeneity exists in the data of a meta-analysis, interpreting the meta-analysis results is an important but difficult task. In this paper, we propose a method that helps such interpretation, in addition to a new association testing procedure that is powerful when heterogeneity exists.
| Meta-analysis is a tool for aggregating information from multiple independent studies [1]–[3]. In genome-wide association studies (GWASs) [4], the use of meta-analysis is becoming more and more popular because one can virtually collect tens of thousands of individuals that will provide power to identify associated variants with small effect sizes [5]–[7]. Several large scale meta-analyses have been performed for diseases including type 1 diabetes [8], type 2 diabetes [9]–[11], bipolar disorder [12], Crohns disease [13], and rheumatoid arthritis [14], and have identified associations not revealed in the individual studies.
In meta-analyses, the effect size between studies may differ and this difference, or heterogeneity, can be caused by many factors [15]–[18]. If the populations are different between studies, the genetic factors can cause heterogeneity [19], [20]. If the subjects are from different regions, the environmental factors can cause heterogeneity [21]. Even if the true effect size is invariant, the design factors can also cause a phenomenon that looks like heterogeneity, what is often called the statistical heterogeneity [22]. If the linkage disequilibrium structures are different between studies, the collected marker can show heterogeneity [23]. If the studies use different genotyping platforms, different imputation accuracies and different genotyping errors can cause heterogeneity [24].
In current meta-analyses of genome-wide association studies, heterogeneity is often observed in the results [9]–[11], [13], [17]. Interpreting the cause of such heterogeneity is important. If the heterogeneity is caused by either genetic or environmental factors, understanding the cause of heterogeneity can help our understanding of the disease mechanism. If the heterogeneity is statistical heterogeneity caused by the design factors, understanding the cause of heterogeneity is crucial in designing a replication study so that we can eliminate the design factors that can hinder the revelation of the true effect in the replication study.
However, interpreting heterogeneous results is difficult. One standard approach is to examine the association p-values of the studies. The inherent limitation of this approach is that a non-significant p-value is not evidence of the absence of an effect. Thus, a p-value does not provide the full information necessary for the interpretation whether or not there is an effect in the study. Another standard approach is to plot observed effect sizes and their confidence intervals of all studies [17], [25], [26]. This plot can be overly complicated when the number of studies is large and does not provide a single estimate that represents the existence of an effect in each study. The limitation of both approaches is that they use classical estimates that are calculated using only the data of each single study. That is, they utilize only within-study information.
In this paper, we propose a framework facilitating the interpretation of the results of a meta-analysis. Our framework is based on a new statistic termed the m-value which is the posterior probability that the effect exists in each study. Plotting the new statistic together with p-values in a two-dimensional space helps us distinguish between the studies predicted to have an effect, the studies predicted to not have an effect, and the ambiguous studies that are underpowered. We name this plot a P-M plot. In this framework, the outlier studies showing distinct characteristics from the other studies are easily identified, as we demonstrate using data from type 2 diabetes and Crohns disease meta-analyses [10], [13]. Our new statistic is fundamentally different from traditional estimates based on the data of single studies. We use all studies simultaneously to calculate the new statistic based on the assumption that the effect sizes are similar if the effect exists. Thus, we utilize cross-study information as well as within-study information.
In addition to helping interpretation, the new framework allows us to develop a new association testing procedure which takes into account the presence or absence of the effect. The new method called the binary effects model is a weighted sum of z-scores method [5] assigning a greater weight to the studies predicted to have an effect and a smaller weight to the studies predicted to not have an effect. Application to the Crohns disease data [13] shows that the new method gives more significant p-values than previous methods at certain loci already identified as associated.
The new method is available at http://genetics.cs.ucla.edu/meta.
In our framework, we use a simplified model to describe heterogeneity among the studies which makes two assumptions. The first assumption is that effect is either present or absent in the studies. This assumption is different from the traditional assumption assuming normally distributed effect sizes [27]–[29]. Our assumption is inspired by the phenomenon that the effect sizes are sometimes observed to be much smaller in some studies than in the others. It is reported that different populations can cause such phenomenon [19], [20], [30], [31]. For example, the homozygosity for APOE 4 variant is known to confer fivefold smaller risk of Alzheimer disease in African Americans than in Asians [19], [30]. The HapK haplotype spanning the LTA4H gene is shown to confer threefold smaller risk of myocardial infraction in the populations of Europeans decent than in African Americans [31]. The HNF4A P2 promoter variants are shown to be associated with type 2 diabetes in Ashkenazi and the results have been replicated [20]. However, in the same study, the same variants did not show associations in four different cohorts of UK population suggesting a heterogeneous effect. Gene-environmental interactions can also cause such phenomenon. If a study lacks an environmental factor necessary for the interaction, the observed effect size can be much smaller in that study. It is generally agreed that the gene-environmental interactions exist in many diseases such as cardio vascular diseases [32], respiratory diseases [33], and mental disorders [34].
The second assumption is that if the effect exists, the effect sizes are similar between studies. We call these two assumptions together the binary effects assumption. While other types of heterogeneity structures are possible such as arbitrary effect sizes, for identifying which studies have an effect and which studies do not have an effect, we expect that this model will be appropriate.
We propose a statistic called the m-value which is the posterior probability that the effect exists in each study of a meta-analysis. Suppose that we analyze studies together in a meta-analysis. Let () be the observed effect size of study and let be the estimated variance of . It is a common practice to consider the true variance. In the current GWASs, the distribution of is well approximated by a normal distribution due to the large sample sizes. Let denote the observed data.
If there is no effect in study ,where is the probability density function of a normal distribution whose mean is and the variance is . If there is effect in study ,where is the unknown true effect size.
Since we want a posterior probability, the Bayesian framework is a good fit. We assume that the prior for the effect size isA possible choice for in GWASs is 0.2 for small effect and 0.4 for large effect [35], [36].
Let be a random variable which has a value 1 if study has an effect and a value 0 if study does not have an effect. Let be the prior probability that each study will have an effect such thatThen we assume a beta prior on Through this paper, we use the uniform distribution prior ( and ), but other priors can also be chosen.
Let be the vector indicating the existence of effect in all studies. can have different values. Let be the set of those values.
Our goal is to estimate the m-value , the posterior probability that the effect exists in study . By the Bayes' theorem,(1)where is a subset of whose elements' th value is 1. Thus, we only need to know for each the posterior probability of ,consisting of the probability of given and the prior probability of .
The prior probability of iswhere is the number of 1's in and is the beta function.
And the probability of given is(2)where is the indices of 0 in and is the indices of 1 in . We can analytically work on the integration to obtainwherewhere is the inverse variance or precision. The summations are all with respect to .
is a scaling factor such thatThe details of the derivation is in Text S1 in Supporting Information S1. As a result, we can calculate for every and therefore obtain for each study .
We propose plotting the studies' p-values and m-values together in two dimensions. This plot, which we call the P-M plot, can help interpreting the results of a meta-analysis. Figure 1 shows that how to interpret such a plot. The right-most (pink) region is where the studies are predicted to have an effect. Often, a study can be in this region even if the p-value is not very significant. The left-most (light-blue) region is where the studies are predicted to not have an effect. This suggests that the sample size is large but the observed effect size is close to zero, suggesting a possibility that there exists no effect in that study. The middle (green) region is where the prediction is ambiguous. A study can be in this region because the study is underpowered due to a small sample size. If the sample size increases, the study will be drawn to either the left or the right side.
If the binary effects assumption does not hold, a study can sit in an unexpected region and a careful interpretation is necessary. For example, if the effects are significant but the effect sizes are in opposite direction in some studies, the studies can sit in the unusual top left region. However, such case will be rare and may be a result of the strand errors.
We propose a new type of random effects model meta-analysis approach called the binary effects model. If the binary effects assumption holds, that is, if the effect is either present or absent in the studies, taking into account this pattern of heterogeneity in the association testing procedure can increase power compared to the general RE approach [23]. The binary effects model we propose is the weighted sum of z-scores method [5] where the m-values are incorporated into the weights. Intuitively, this is equivalent to assigning a greater weight to the studies predicted to have an effect and a smaller weight to the studies predicted to not have an effect.
Let be the z-score of study . The common form of the weighted sum of z-scores statistic for the fixed effects model isIn many cases, the weight approximates to the form where is the sample size and is the minor allele frequency [23]. When the minor allele frequency is similar between studies, the weight approximates to the popular form of [5].
The binary effects model statistic we propose isOur method is an empirical approach that uses estimated from the data as the prior weight for each study. Since the m-value is estimated using all studies, our approach can be thought of as gathering information from all studies and distributing back to each study in the form of weight. We choose this approach because of its simple formulation.
Since is not independent of , the statistic does not follow a normal distribution. Thus, the p-value is obtained using sampling which can be inefficient. We use two ideas to expedite the sampling. First, we propose an importance sampling procedure which is more efficient than the standard sampling. Second, we use an efficient approximation of m-value. See Text S2 and S3 in Supporting Information S1 for details.
In order to evaluate our methods, we use the following simulation approach. Assuming a minor allele frequency, a relative risk, and the number of individuals of cases and controls, a straightforward simulation approach is to sample alleles for cases and alleles for controls according to the expected minor allele frequencies in the cases and controls respectively [38]. However, since we perform extensive simulations assuming thousands of individuals, we use an approximation approach that samples the minor allele count from a normal distribution and rounds it to the nearest non-negative integer.
The URL for methods presented herein is as follows:
http://genetics.cs.ucla.edu/meta
We demonstrate a simple simulation example showing how m-value behaves depending on the presence and absence of the effect and the sample size. First, we make the following assumptions throughout all of the experiments in this paper. We assume that the minor allele frequency of the collected marker is 0.3. We assume that the equal number of cases and controls are collected and refers to the total number of individuals as sample size . We also assume a very small disease prevalence when we calculate the expected minor allele frequencies for cases and controls given a relative risk . For the details how the expected values are calculated, see Han and Eskin [23]. Note that these assumptions are not critical factors affecting our simulation results. In all experiments, the random effects model (RE) denotes the RE method of Han and Eskin [23]. We omit the results of the conventional RE method [15] because they are highly conservative [23]. Throughout this paper, we use the following priors for calculating m-values. We use for the prior of the effect size (). We use the uniform distribution prior, , for the prior of the existence of effect ().
In this simulation example, we assume four different types of studies. The first type is a large study having an effect ( and ). The second type is a small study having an effect ( and ). The third type is a large study not having an effect ( and ). The fourth type is a small study not having an effect ( and ). We generate two studies per each type, constructing a simulated meta-analysis set of total eight studies. We accept this simulation set only if none of eight studies' p-values exceeds the genome-wide threshold () but the meta-analysis p-value calculated by the RE approach exceeds the genome-wide threshold. Otherwise, we repeat. We construct 1,000 meta-analysis sets.
Given this simulated data, we plot the histogram of m-values for each type of studies separately in Figure 2. Figure 2A shows that almost all (99.9%) of large studies with an effect are concentrated on large m-values (), showing that the m-values effectively predict that the effect exists in the studies. Figure 2C shows that a large amount (78.6%) of large studies without an effect are concentrated on small m-values (). Figure 2B and 2D show that when the sample size is small, m-value tends to the mid-range regardless of the effect, suggesting that the studies are underpowered to determine the presence of an effect.
In this experiment, we compare the p-value, m-value, and BF by measuring how well they predict which studies have an effect and which studies do not have an effect. We assume a meta-analysis of 10 studies where the effect is either present () or not. We randomly pick the number of studies having an effect () from a uniform distribution ranging from 1 to 9, and randomly decide which studies have an effect. We randomly pick the sample size of each study from a uniform distribution between 500 and 2,000. Given the sample sizes and the effect sizes, we generate a meta-analysis study set. We accept the meta-analysis set only if none of the studies' p-values exceeds the genome-wide threshold () and the meta-analysis p-value exceeds the genome-wide threshold. We repeat until we construct 1,000 meta-analysis sets.
We examine each of 10,000 studies included in the simulated 1,000 meta-analysis sets. For each study, we calculate the p-value, m-value, and BF. We use the asymptotic BF of Wakefield [39] assuming the same prior distribution about the effect size as the m-value. Then we evaluate the performance of each statistic as follows. To evaluate the performance of m-value, we fix an arbitrary threshold so that we predict the studies having m-value to have an effect. Since we know the underlying truth if the effect exists or not in each study, we can measure what proportion of the studies actually having an effect is correctly predicted to have an effect (true prediction rate) and what proportion of the studies actually not having an effect is incorrectly predicted to have an effect (false prediction rate). Then we change the threshold to draw a curve between the true prediction rate and the false prediction rate, which is often called the receiver-operating-characteristic (ROC) curve. We do the same analysis for p-value and BF.
Figure 3A shows that m-value is superior to p-value and BF in predicting the studies having an effect. This is because m-value can utilize the cross-study information when the binary effects assumption holds. The performances of p-value and BF are almost identical.
Next, we evaluate the performance of the statistics in predicting studies not having an effect. The experiment is exactly the same as the previous experiment except that, given a threshold , we predict the studies having m-value to not have an effect. We similarly draw the ROC curves for the three statistics. True and false prediction rates are defined similarly for the objective of predicting the studies not having an effect.
Figure 3B shows that the m-value is even more superior to the other statistics in this experiment than in the previous experiment. The p-value shows the most inferior performance. This is expected because p-value is designed for detecting the presence of an effect but not for detecting the absence of an effect. That is, a non-significant p-value is not evidence of the absence of an effect but can be the result of a small sample size. On the other hand, the BF testing for the absence of an effect is just the reciprocal of the BF testing for the presence of an effect. Thus, the same BF can be used for both purposes. Although the BF performs better than the p-value, the m-value is even more superior. The relative performance gain of the m-value compared to the BF is due to the cross-study information utilized.
We apply our P-M plot framework to the real data of the meta-analysis of type 2 Diabetes (T2D) of Scott et al. [10]. The meta-analysis consists of three different GWAS investigations, the Finland-United States Investigation on NIDDM Genetics (FUSION) [10], the Diabetes Genetics Initiative (DGI) [11], and the WTCCC [9], [40].
In their analysis, two SNPs are shown to have a heterogeneous effect, rs8050136 and rs9300039. Ioannidis et al. [17] provide an insightful explanation about the heterogeneity at rs8050136. The WTCCC/UKT2D groups identified evidence for T2D and body mass index (BMI) associations with a set of SNPs including rs8050136 in the FTO region [40]. On the other hand, in the DGI study, the SNP rs8050136 was not significant. The explanation that Ioannidis et al. suggest is that the observed association at rs8050136 (FTO) may be mediated by its association with obesity. In fact, DGI is the only study where the BMI is matched between cases and controls, and the T2D association appears to be mediated through a primary effect on adiposity [11]. Thus, although the truth is unknown, the explanation of Ioannidis et al. is reasonable. Compared to rs8050136, the cause of heterogeneity at rs9300039 is less understood. It is suggested that the heterogeneity might reflect the different tag polymorphisms used in the studies [17].
To gain insights on these studies, we apply our P-M plot. Figure 4A shows the forest plot, the plot showing only the p-values, and the P-M plot for rs8050136. In the P-M plot, DGI appears to be well separated from the other two studies, even though its m-value () is not below the threshold (). Thus, the P-M plot visualizes that DGI can have a different characteristic from the others. Such a separation is not clear in the plot showing only the p-values. In the plot showing only the p-values, DGI is close to FUSION since FUSION is also not very significant (). However, the m-value of FUSION is much greater () than that of DGI. This suggests that the effect is much more likely to exist in the FUSION study than in the DGI study.
Figure 4B shows the plots for rs9300039. The P-M plot shows a different pattern from the P-M plot of rs8050136. In this P-M plot, every study has an m-value greater than 0.5. Thus, no study shows evidence of no effect. Comparing the plots of rs8050136 and rs9300039 gives an interesting observation. In the plot showing only the p-values, both SNPs show a specific pattern of p-values that a single study is considerably more significant than the other two. However, despite of this similarity in the pattern of p-values, the two SNPs' P-M plots look different enough that can lead us to different interpretations. This shows that our P-M plot can provide information that is not apparent in the analysis of only the p-values.
We apply our plotting framework to the data of the recent meta-analysis of Crohns disease of Franke et al. [13]. This meta-analysis consists of six different GWAS comprising 6,333 cases and 15,056 controls, and even more samples in the replication stage. In this study, 39 associated loci are newly identified increasing the number of associated loci to 71. We apply our framework to six loci where a high level of heterogeneity is observed. Han and Eskin [23] showed that at these six loci, RE gave more significant p-values than the fixed effects model (FE).
Figure 5 shows the P-M plots of two loci. See Figure S1 for the plots of all six loci. The names of the studies follow the names used in Franke et al. [13]. At these two loci, rs3024505 and rs17293632, the m-value of WTCCC is close to the threshold for predicting no effect. A possible explanation is that the different marker sets could have caused the statistical heterogeneity at these loci. WTCCC [40] used the Affymetrix platform while others used the Illumina platform. Although we do not further investigate this hypothesis, it is true that the P-M plots visualize an interesting outlier behavior of WTCCC at these loci. Such an observation is not clear in both the forest plot and the plot showing only p-values. In the plot showing only p-values, studies having non-significant p-values are all clustered and WTCCC is only one of them. In the forest plot, WTCCC is not the only study showing a small effect size at both loci. For example, at rs3024505, NIDDKNJ shows a smaller effect size than WTCCC. However, the m-value of WTCCC is much smaller than NIDDKNJ's because of the large sample size. Such an interaction between the sample size and the prediction can also be inferred from the forest plot since the forest plot includes the confidence interval. However, it is difficult to numerically quantify the effect of sample size on the prediction by visually examining the forest plot.
We estimate the false positive rate of the new binary effects model (BE). Assuming the null hypothesis of no association, we construct 5 studies of sample size 1,000 to build a meta-analysis set. We calculate the meta-analysis p-value of BE using our importance sampling procedure with 10,000 samples. We also calculate the meta-analysis p-values of FE and RE. We build 100 million sets of meta-analysis and estimate the false positive rate as the proportion of the simulated sets whose p-value exceeds a threshold. We vary the threshold levels from 0.05 to . Table 1 shows that all methods including BE control the false positive rates accurately, at all threshold levels examined. When we increase the number of studies from 5 to 10, the results are essentially the same and the false positive rates are controlled (Data not shown).
We compare the power of BE to the powers of FE and RE. Assuming a meta-analysis of five studies of an equal sample size 1,000, we construct 10,000 meta-analysis sets. The power of each method is estimated as the proportion of the meta-analysis sets whose meta-analysis p-value calculated by each method exceeds the genome-wide threshold ().
We measure power in two different situations. First, we assume a situation that the effect is either present or absent. We decrease the number of studies having an effect () from 5 to 2. We increase the relative risk as decreases, using for respectively, in order to show the relative performance between methods.
Figure 6 shows that except for the case that there is no heterogeneity (), BE is the most powerful among all methods. BE is more powerful than RE, even though both are a random effects model, possibly because it learns the fact that some studies do not have an effect from the data. When there is no heterogeneity (), FE achieves the highest power and BE achieves the lowest power.
Second, we assume a classical setting where the effect sizes follow a normal distribution. Assuming that the mean effect size of , we sample the log of effect size of each study from a normal distribution having the mean and the standard deviation where is the parameter we vary. As increases, the heterogeneity increases. We measure the power of each method varying from zero to one. Figure 7 shows that in this situation, BE is generally less powerful than RE. The power difference between BE and RE is the greatest when the heterogeneity is small. As the heterogeneity increases, BE shows a similar power to RE.
We apply BE to the real data of Crohns disease of Franke et al. [13]. Han and Eskin [23] showed that out of 69 associated loci analyzed, RE gave more significant p-values than FE at six loci where high level of heterogeneity is observed. We calculate the p-values at these loci using BE and compare to the p-values of FE and RE.
Table 2 shows that at all six loci where RE gave more significant p-values than FE, BE gives even more significant p-values. The reason why BE gives more significant p-values can be explained by examining the P-M plots of these loci in Figure 5 and Figure S1. The P-M plots show that at these loci, some studies show high m-values and some studies show low m-values, suggesting a bimodal distribution of effect size. Thus, the situation is very similar to the case that the effect is either present or not, in which case BE achieves higher power than RE as shown in Figure 6.
We measure how accurately the importance sampling procedure of BE estimates the p-value depending on the number of samples used. We calculate the BE p-value for the same dataset in 100 different runs to estimate the variance of the p-value estimate. Our criterion of interest is the ratio between the standard deviation of our estimate and the target p-value. For this, we use the 69 associated loci in the Crohns disease data of Franke et al. [13] that were previously analyzed in Han and Eskin [23]. We measure the ratio for each locus and average over all loci. We do this varying the number of samples from 1,000 to 1,000,000.
Table 3 shows that as the number of samples used for importance sampling increases, the accuracy increases. The pattern of accuracy increase is what we would usually expect in a sampling procedure; standard deviation is decreased approximately by the square root of the sample size increase. When the number of samples is 1,000, the ratio is roughly 0.5. A ratio of 0.5 is large, but can be enough for initial screening if we would apply an adaptive sampling that samples larger number of samples only for loci that are at least moderately significant (e.g. ).
We measure the computational efficiency of the importance sampling procedure of BE. In our software, we implemented an adaptive sampling procedure that samples smaller number first () and then larger number () for the loci that are at least moderately significant. In the machine equipped with Intel Xeon 1.68 GHz CPU, when we use 1,000 samples in the importance sampling, calculating BE p-values of 1,000 loci for the meta-analysis of 10 studies takes 100 seconds. Thus, to calculate BE p-values of one million loci assuming that 1,000 loci among them are moderately significant, it will take approximately 30 hours which is a feasible amount of time. If the number of samples is increased to achieve better accuracy, such as and , the procedure will still be efficient if one uses multiple computers or a cluster since the procedure is parallelizable.
We introduce a framework facilitating the interpretation of meta-analysis results based on a new statistic representing the posterior probability that the effect exists in each study. Our framework utilizes cross-study information and is shown to help interpretations in the simulations and the real data. The new statistic also allows us to develop a new association testing procedure called the binary effects model.
In the current meta-analyses of genome-wide association studies, heterogeneity is often observed and our framework will be a useful tool for interpreting such results. We expect that our framework will be even more useful in the future meta-analyses. As the number of studies in a meta-analysis grows, the chance of heterogeneity will increase [6]. Also, a meta-analytic approach can often be applied to a broader area such as to multiple diseases with similar etiology, in which case the heterogeneity is more likely to occur. Moreover, the majority of the current meta-analyses only use the fixed effects model (FE). The use of a random effects model (RE) approach [23] such as the binary effects model presented herein will increase the number of identified associations showing heterogeneity, since an RE approach is more powerful than FE for detecting associations with heterogeneity.
One limitation of our approach is that although the new statistic can predict the studies having an effect and the studies not having an effect, it does not distinguish the true heterogeneity and the statistical heterogeneity [22]. Discriminating between the two can be very difficult based on the observed data and might often be possible only by external data such as the replication studies. In that sense, our method can help discriminating them because one can come up with a hypothesis based on m-values that the heterogeneity is caused by specific design factors and then control the factors in the replication stage. The heterogeneity will disappear in the replication stage if it was due to the design factors.
Similarly to other Bayesian approaches [35], [36], the prior choice in our method can have a non-negligible effect on the predictions. For the prior of the effect size , it is important to set a reasonable value based on the prior information about the effect size. See Stephens and Balding [35] for the general guideline for this choice. For the prior of the probability that the effect exists , we used the uniform distribution () in this paper. However, different priors can also be used for different situations. If one expects that most of the studies have an effect, an asymmetric prior such as can be used. If one is certain that the studies having an effect and the studies not having an effect are mixed, a bell-shape prior such as can be used. See Figure S2 for the plots of the possible choices of priors.
|
10.1371/journal.pbio.0050261 | MSN2 and MSN4 Link Calorie Restriction and TOR to Sirtuin-Mediated Lifespan Extension in Saccharomyces cerevisiae | Calorie restriction (CR) robustly extends the lifespan of numerous species. In the yeast Saccharomyces cerevisiae, CR has been proposed to extend lifespan by boosting the activity of sirtuin deacetylases, thereby suppressing the formation of toxic repetitive ribosomal DNA (rDNA) circles. An alternative theory is that CR works by suppressing the TOR (target of rapamycin) signaling pathway, which extends lifespan via mechanisms that are unknown but thought to be independent of sirtuins. Here we show that TOR inhibition extends lifespan by the same mechanism as CR: by increasing Sir2p activity and stabilizing the rDNA locus. Further, we show that rDNA stabilization and lifespan extension by both CR and TOR signaling is due to the relocalization of the transcription factors Msn2p and Msn4p from the cytoplasm to the nucleus, where they increase expression of the nicotinamidase gene PNC1. These findings suggest that TOR and sirtuins may be part of the same longevity pathway in higher organisms, and that they may promote genomic stability during aging.
| There are only a few techniques that reliably promote longevity in multiple, distantly related species. Perhaps the best known, caloric restriction (CR), was first shown to promote lifespan in rodents in the 1930s and has since been shown to work in most species it has been tested on. We and others have previously proposed that CR extends lifespan in budding yeast by boosting the activity of sirtuin deacetylases, which work to extend lifespan by suppressing genomic instability. A competing theory is that CR works by suppressing the TOR (target of rapamycin) signaling pathway, which has recently been discovered to extend the lifespan of yeast and worms, but the downstream players are not yet known. We show that TOR inhibition and sirtuins are part of the same CR pathway that extends yeast lifespan by stabilizing the genome. CR and TOR inhibition promote longevity by relocalizing two transcription factors, Msn2p and Msn4p, from the cytoplasm to the nucleus, where they increase expression of the nicotinamidase gene PNC1, a regulator of sirtuin activity. We propose that TOR signaling and sirtuins may also be part of the same CR pathway in mammals.
| In the budding yeast Saccharomyces cerevisiae, replicative lifespan is measured by the number of divisions that a mother cell undergoes before senescing [1–3]. A primary cause of aging in this organism is homologous recombination between ribosomal DNA (rDNA) repeats, resulting in the formation of extrachromosomal rDNA circles (ERCs) that accumulate to toxic levels in mother cells [4]. Sir2p and a closely related homolog, Hst2p, belong to the sirtuin family of NAD+-dependent deacetylases [5] that can forestall aging by stabilizing the rDNA locus [4,6,7]. Although rDNA recombination is not known to play a role in the aging of metazoans, the function of Sir2p enzymes in lifespan determination appears to be conserved. In Caenorhabditis elegans and Drosophila melanogaster, additional copies of the SIR2 gene or pharmacological modulation of the Sir2p deacetylase also extend lifespan [8–11].
The diet known as calorie restriction (CR) prolongs the lifespan of numerous species, including fungi, invertebrates, and mammals [1,12,13]. Whether or not Sir2p enzymes play a role in CR-mediated lifespan extension is hotly debated. In support of their playing a role, additional copies of either SIR2 or HST2 suppress rDNA recombination and extend yeast replicative lifespan, whereas strains lacking SIR2 and HST2 fail to live longer when subjected to CR [14,15]. Similarly, CR diets or genetic mimics of CR fail to extend the lifespan of D. melanogaster and C. elegans lacking Sir2p [12,16]. However, other researchers favor a model in which Sir2p plays no role in CR-mediated lifespan extension, and instead the TOR (target of rapamycin) pathway is proposed to play the central role [17,18].
TOR is a nutrient-responsive phosphatidylinositol-kinase-related kinase that regulates protein synthesis and cell growth, and is inhibited by rapamycin, an immunosuppressive and anticancer drug that specifically inhibits TOR [19]. It has recently been discovered that lifespan can be extended in a variety of species by inhibition of TOR signaling, including S. cerevisiae, C. elegans, and D. melanogaster [17,20,21]. The mechanism by which inhibition of TOR signaling extends lifespan is unclear, but it has been proposed that it may act by altering ribosome assembly and translation [17,21–23]. Similar to CR, inhibition of TOR signaling can extend yeast lifespan in the absence of SIR2 [17], but whether SIR2 plays a role in lifespan extension during inhibition of TOR signaling is not known.
We have previously investigated the pathways by which CR operates in yeast. The enzymatic activity of Sir2p is regulated by endogenous levels of nicotinamide (NAM), a sirtuin inhibitor [24]. Yeast strains grown on standard 2% glucose medium have an intracellular concentration of ∼50 μM NAM, which is almost precisely the IC50 of Sir2p [24–28]. We and others have shown that CR and other mild stresses, including heat stress and osmotic shock, extend yeast lifespan by increasing expression of the PNC1 gene, which encodes a nicotinamidase [27,29]. Recent evidence indicates that mammalian Nampt/PBEF, a putative functional ortholog of PNC1 that is required for the conversion of NAM to NAD+, also regulates sirtuin activity [30,31].
PNC1 is an intriguing longevity gene because its expression is regulated by environmental stimuli that extend lifespan, such as heat, osmotic stress, low amino acids, and CR. Binding sites for the stress-responsive zinc-finger transcription factors Msn2p and Msn4p have been identified in the PNC1 promoter [32]. Msn2p and Msn4p have previously been shown to regulate chronological lifespan extension in response to deletion of the yeast Akt homolog SCH9 by controlling the expression of the superoxide dismutase SOD2 [33,34]. Msn2p/4p are therefore good candidates to regulate PNC1 expression in response to nutrient availability, but a previous study, which used a cdc25–10 mutant strain to mimic CR, concluded that MSN2 and MSN4 play no role in CR-mediated lifespan extension [14].
Here we show that Msn2p/4p are important regulators of yeast replicative lifespan and that they relocalize from the cytoplasm to the nucleus during CR, where they bind to and activate the PNC1 gene. Moreover, inhibition of TOR signaling acts via this same pathway to promote the expression of PNC1 and suppress rDNA recombination. These data provide evidence for a pathway from the cell's environment to an actual cause of aging, via which both CR and TOR signaling modulate lifespan.
We began by asking whether MSN2/4 are mediators of yeast lifespan extension by CR (0.5% glucose). In contrast to previous work using a genetic mimic of CR [14], we found that lifespan extension by CR was completely MSN2/4-dependent (Figure 1A). Single deletions of MSN2 or MSN4 did not block the ability of CR to extend lifespan, consistent with their known redundancy (Figure 1B). Interestingly, MSN2 and MSN4 were also absolutely necessary for lifespan extension resulting from TOR inhibition (Figure 1C).
The requirement of MSN2/4 for CR- and TOR-mediated lifespan extension indicated that CR and TOR might extend lifespan via the same mechanism. We and others have presented evidence that CR works by increasing Sir2p activity [8,14,29], as indicated by increases in telomeric silencing and a reduction in rDNA recombination [7,15,35]. We therefore examined whether rapamycin increased telomeric silencing and reduced rDNA recombination, and whether this was altered by the presence or absence of MSN2/4 or PNC1. We observed that rapamycin increased telomeric silencing in a wild-type strain, but not in strains lacking MSN2/4 or PNC1 (Figure S1). Treatment with rapamycin also suppressed rDNA recombination and, again, this effect required MSN2/4 and PNC1 (Figure 2A and 2B). The effect of rapamycin on silencing and rDNA recombination was also completely blocked by treating cells with NAM (Figure 2B), indicating that a sirtuin is likely required for this effect. Neither CR nor rapamycin increased protein levels of Sir2p (Figure 2C). Taken together, these data indicate that inhibition of TOR signaling increases the activity of Sir2p and/or another sirtuin.
In addition to Sir2p, S. cerevisiae contains four HST (homolog of sir two) genes, HST1–4. We and others have uncovered considerable redundancy in the sirtuin family, with Hst1p and Hst2p able to substitute for Sir2p during CR [15]. Recent studies in our laboratory have demonstrated that overexpression of any one of the sirtuin genes HST1–4 can suppress rDNA recombination (D. Lamming and M. Latorre-Esteves, unpublished data). Thus, we were not surprised to find that rapamycin could suppress rDNA recombination in a sir2Δ fob1Δ strain (Figure S2). This result supports a previously published report that inhibition of TOR signaling can extend lifespan in the absence of SIR2 [17]. Although TOR inhibition can suppress rDNA recombination in the absence of SIR2, deletion of both SIR2 and HST2 blocked the ability of rapamycin to suppress rDNA recombination (Figure S2). Consistent with these data, we found that rapamycin could extend the lifespan of a sir2Δ fob1Δ strain but had no effect on the lifespan of a sir2Δ hst2Δ fob1Δ strain (Figure S2). Thus, rather than working through a single sirtuin, we favor a model in which inhibition of TOR signaling suppresses rDNA recombination and promotes longevity by activating multiple sirtuins, including Sir2p and Hst2p.
In agreement with this model, PNC1 was required for lifespan extension by rapamycin (Figure 2D), and overexpression of PNC1 was sufficient to suppress rDNA recombination and extend lifespan in an msn2Δ/4Δ background (Figure 2E and 2F). Thus, within the framework of our model, TOR signaling is upstream of PNC1 and PNC1 is downstream of MSN2/4. The simplest mechanistic explanation is that inhibition of TOR activates MSN2 and MSN4, which then increase the expression of PNC1. To test this, we examined the effect of CR and rapamycin on Pnc1p levels in the presence and absence of MSN2/4. Both treatments induced Pnc1p-GFP (Figure 3A) as well as native Pnc1p (Figure 3B). In contrast, there were considerably lower levels of Pnc1p in the untreated msn2Δ/4Δ strain, and Pnc1p levels remained below those of the untreated wild-type strain, even in response to CR or rapamycin (Figure 3B).
Msn2p/4p are normally maintained in the cytoplasm by the activity of PKA. As cAMP levels fall, PKA activity decreases, and dephosphorylated Msn2p/4p relocalize to the nucleus [36]. A cdc25–10 mutant that has constitutively decreased cAMP/PKA signaling [37] expressed higher levels of Pnc1p than the wild-type strain, and this was MSN2/4-dependent (Figure 3C). Upregulation of PNC1 expression by other stresses was also MSN2/4-dependent (Figure 3D), and in agreement with the lifespan data in Figure 1B, single deletion of MSN2 or MSN4 did not greatly affect the ability of CR to induce PNC1 (Figure S3).
We have previously shown that heat shock extends lifespan in a PNC1-dependent manner [29], so we were curious whether this was an MSN2/4-dependent process. Heat shock induced the expression of PNC1 and extended lifespan even in the absence of MSN2/4 (Figure 3D and 3E). To explore the MSN2/4-independent mechanism by which heat shock induces PNC1, we analyzed the PNC1 promoter and found putative binding sites for the heat shock factor Hsf1p (positions −251 to −275 and −319 to −353, with respect to the start codon). Since HSF1 is an essential gene, we used an msn2Δ/4Δ strain containing doxycycline-repressible HSF1 to examine the role of Hsf1p in PNC1 regulation [38]. Repression of HSF1 largely blocked the ability of heat shock to upregulate PNC1 (Figure 3F), but it did not affect the ability of CR or rapamycin to induce PNC1 (data not shown).
Under standard yeast growth conditions (2% glucose, 30 °C), Msn2p/4p localize predominately to the cytoplasm [36], but in response to a variety of stresses, Msn2p/4p localize to the nucleus, where they activate stress-responsive genes [39]. We observed that CR induced the translocation of Msn2p into the nucleus, and the extent of the translocation was proportional to the degree of CR (Figure 4A and 4B). CR also induced the nuclear localization of Msn4p, but quantification of the nuclear localization indicated that Msn4p was less sensitive to glucose restriction than Msn2p (Figure S4). The reasons for this difference are unknown, but it may allow for differential gene regulation as nutrients are depleted. Nuclear localization of Msn2p also increased in both tor1Δ and cdc25–10 strains (Figure 4C and 4D), in agreement with previous reports [40,41].
Cells grown in moderate CR conditions (0.5% and 0.1% glucose) showed heterogeneous localization patterns for Msn2p/4p, indicating that Msn2p/4p might be oscillating between the nucleus and cytoplasm, as has recently been noted for cells exposed to light stress or osmotic shock [42]. To determine if this was the case, time-lapse photomicrographs of cells expressing Msn2p-GFP were taken at 30-s intervals during growth under various glucose concentrations. The vast majority of cells grown in 2% glucose showed a cytoplasmic localization of Msn2p (Figure 5A), while cells incubated in medium lacking glucose had exclusively nuclear localization of Msn2p (Figure 5B). These patterns did not change over time. However, in cells grown under intermediate levels of CR (0.1% or 0.5% glucose), nucleo-cytoplasmic oscillations of Msn2p-GFP were observed (Figures 5C and S5). The GFP signal eventually became bleached (data not shown), demonstrating that Msn2p was not simply being degraded in the nucleus and re-synthesized in the cytoplasm but was instead actively being transported in and out of the nucleus [43]. Msn4p-GFP showed a similar oscillatory behavior, and localization of Msn2p or Msn4p under 2% or 0.5% glucose was not altered by a lack of the other transcription factor (data not shown).
The PNC1 promoter contains four Msn2p/4p-binding sites known as stress response elements (STREs) [32]. To determine whether Msn2p/4p directly regulate PNC1 in response to CR, we tested several reporter constructs containing different regions of the PNC1 promoter (Figure 6A). The full-length promoter construct was greatly induced in response to CR (Figure 6B), whereas the reporters lacking one or more of the STREs were induced at significantly lower levels. The construct with no STREs showed no induction. The involvement of MSN2/4 was confirmed by the finding that CR completely failed to upregulate the reporter in an msn2Δ/4Δ strain (Figure 6C). The ability of CR to induce the PNC1 reporter was largely unaffected in single msn2Δ or msn4Δ mutants (Figure 6D and 6E), which is consistent with the ability of CR to extend the lifespan of single but not double msn mutants (see Figure 1A and 1B). Interestingly, deletion of MSN4 did not affect the expression of the STRE4 reporter construct under 2% glucose, but significantly reduced its response to CR (Figure 6E). In contrast, the STRE2 reporter construct was unaffected by the MSN4 deletion, indicating that the influence of Msn4p on PNC1 expression may be mediated by the distal two STRE elements.
We have previously shown that while single deletion of MSN2 impairs the ability of PNC1 expression to be induced by stress, single deletion of MSN4 has little effect (Figure S3). Furthermore, Msn2p localizes to the nucleus more readily in response to decreased glucose concentration (compare Figures 4 and S4), deletion of MSN2 results in approximately 2-fold less expression of the reporter construct than deletion of MSN4, and the STRE2 constructs were less responsive to CR in an msn2Δ strain than in an msn4Δ strain (Figure 6D and 6E). Together these data support the conclusion that Msn2p is more important than Msn4p for regulating PNC1 expression in response to CR.
Msn2p also appears to be more important for the response of PNC1 to heat stress (37 °C). Deletion of MSN2 resulted in an approximately 2-fold decrease in the expression of the STRE4 reporter construct in response to heat, relative to the wild-type strain (compare Figure 6B and 6D). In contrast, single deletion of MSN4 actually leads to increased expression of the STRE 4 reporter in response to heat (compare Figure 6B and 6E). This upregulation in the msn4Δ strain is likely to be due to MSN2, because deletion of MSN2 in the msn4Δ strain resulted in the poorest response of the STRE4 reporter to heat (Figure 6C). The data also indicated that elements present in the STRE4 reporter but not in the STRE2 reporter are responsible for the MSN2/4-independent induction of the promoter in response to heat stress. Consistent with this hypothesis, the putative binding sites for Hsf1p identified by our promoter analysis lie within this region.
Given the relative importance of Msn2p, we next asked if the nuclear localization of Msn2p was sufficient to induce PNC1. We expressed a constitutively nuclear mutant of Msn2p, Msn2p(S4A) [39], in an msn2Δ and msn4Δ mutant, and observed robust induction of Pnc1p (Figure 7A), demonstrating that nuclear localization of Msn2p is sufficient to induce PNC1 expression. We also noted that this strain grew poorly, in agreement with previous reports that constitutive nuclear localization of Msn2p/4p antagonizes growth [44].
The data thus far strongly supported a model in which CR promotes the binding of Msn2p directly to the PNC1 promoter, yet there remained the possibility that PNC1 was regulated by Msn2p indirectly. To distinguish between these two possibilities, we used chromatin immunoprecipitation to determine if Msn2p binds to the PNC1 promoter in response to CR (Figure 7B). We detected Msn2p at the PNC1 promoter, and the apparent abundance of Msn2p was proportional to the degree of CR (Figure 7C), which was consistent with the glucose-dependent nuclear localization of Msn2p that we previously observed.
In order to better understand the target specificity of Msn2p/4p in response to CR, we utilized previously published microarray data [45,46] to examine the expression of 82 previously identified STRE-containing genes [47]. We were surprised to find that the genes varied greatly in their responsiveness to conditions of low glucose. Ten genes, including PNC1, were highly responsive to small changes in glucose concentration, whereas other STRE-containing genes were responsive only to large changes in glucose concentration or were unresponsive (Table S1). In an effort to understand why some genes are more responsive to Msn2p/4p than others, we analyzed the promoters of these STRE-containing genes. The genes that were less responsive to low glucose averaged fewer STRE elements than the highly responsive genes and, on average, had binding sites approximately 50–60 base pairs further from the start of the coding sequence than genes that responded to low glucose (Table S2). We speculate that the placement of transcription factor binding sites at varying distances from the promoter may be a conserved mechanism for the differential regulation of stress-induced genes.
Because our work indicated that two major longevity pathways, namely CR and TOR, promote lifespan by inducing the expression of PNC1, we wondered whether other yeast longevity genes also modulate PNC1 expression. Deletion of the glycolysis pathway gene HXK2 extends lifespan and has been proposed as a genetic mimic of CR [15,17,37]. Consistent with this, an hxk2Δ strain had higher levels of PNC1 (Figure S6), placing it upstream of PNC1. Similarly, there are abundant data linking the Snf1p/AMPK pathway to yeast longevity. The yeast homolog of AMPK, Snf1p, phosphorylates Msn2p in response to glucose deprivation, is regulated by TOR, and influences lifespan by modulating ERC formation [48,49]. Deletion of SNF4, an activator of Snf1p, has been shown to extend lifespan, whereas deletion of SIP2, a repressor of Snf1p, shortens lifespan [49]. Surprisingly, we saw no evidence for involvement of the SNF pathway (SNF1, SNF4, or SIP2) in the regulation of PNC1 (Figure S6), indicating that the activity of Snf1p/AMPK regulates ERC formation independently of PNC1. Furthermore, CR induced PNC1 expression equally well in wild-type and snf1Δ mutant strains (Figure S6).
We also examined the potential role of ADR1, a transcription factor that is regulated by SNF1/AMPK and that we suspected from our promoter analysis of PNC1 might regulate PNC1 [50]. We found that ADR1 was not required for the induction of PNC1 by CR (Figure S6 and data not shown), and that overexpression of ADR1, or expression of a constitutively active form of Adr1p, did not induce expression of PNC1 (Figure S6). Together, these data show that while attenuation of TOR signaling, PKA activity (cdc25–10), or glucose metabolism (hxk2Δ) extends replicative lifespan, ostensibly by mimicking the effects of CR, the SNF pathway regulates lifespan via a PNC1-independent mechanism.
How CR delays aging and extends the lifespan of various species is poorly understood. In this study, we have connected two sections of the yeast CR pathway, namely the cytoplasmic components (the glucokinase/cAMP/PKA pathway) and the nuclear components (Pnc1p, Sir2p, and ERCs). Furthermore, we have shown that TOR signaling, which was previously thought to regulate lifespan independently of sirtuins and ERCs, actually governs the activity of the sirtuins and suppresses rDNA recombination (Figure 8). This provides additional support to the theory that CR extends replicative lifespan, at least in part, by activating sirtuins.
We also demonstrate that the induction of PNC1 in response to numerous stresses is largely controlled by the transcription factors Msn2p and Msn4p. Under conditions of high salt or sorbitol, PNC1 expression is increased in a manner that is completely dependent on MSN2/4 (Figure 3C). While we have linked the increase in Pnc1p levels during heat stress in an msn2Δ/4Δ strain to the transcription factor Hsf1p (Figure 3F), we did not observe a role for Hsf1p in the response to CR or low amino acids.
There must be additional factors that control the expression of PNC1, because an increase in Pnc1p levels still occurs in an msn2Δ/4Δ strain grown in 0.5% glucose or in medium with low amino acids. One possibility is that PNC1 is co-activated by Gcr1p, a transcriptional activator with potential binding sites ∼500 base pairs upstream of the PNC1 start codon. GCR1 regulates glycolytic enzyme genes, ribosomal gene synthesis, and trehalose/glycogen metabolism [51,52], making it an interesting candidate for future analysis, although we note that any such analysis is complicated by the severe growth defect of a gcr1Δ strain.
In contrast to our study, a previous study utilizing a cdc25–10 mutant as a mimic of CR found that replicative lifespan extension of a PSY316 strain can occur in the absence of MSN2/4 [14]. A recent study has shown that PSY316 may differ substantially from other yeast strains in terms of Sir2p-mediated lifespan extension [18], and our data may reflect that difference. We favor the notion that while the cdc25–10 mutation mirrors aspects of CR, such as lower PKA activity and increased lifespan, it does not fully replicate it.
A previous study has shown that inhibition of TOR signaling can extend lifespan, even in the absence of SIR2 [17]. In agreement with this data, we find that treatment with rapamycin can suppress rDNA recombination and extend lifespan in a sir2Δ fob1Δ strain (Figure S2). Yeast contain four additional sirtuin genes (HST1–4), some of which can compensate for the lack of Sir2p during CR [15]. Under the conditions and with the strain used in this study, we have observed that rapamycin no longer lowers rDNA recombination or promotes longevity if both SIR2 and HST2 are deleted (Figure S2), implying that these two genes are primarily responsible for the effect. However, a W303 sir2Δ hst2Δ fob1Δ strain has a high rate of rDNA recombination and a short lifespan, which may serve to obscure the role of additional sirtuins or other mediators in the response to TOR inhibition. In fact, overexpression of PNC1 in a wild-type strain lowers rDNA recombination more than in a strain lacking MSN2/4, which may indicate that genes downstream of MSN2/4 besides PNC1 also function to repress rDNA recombination. These alternative pathways may be especially important when glucose concentrations are extremely low [53] and may include pathways that directly regulate rDNA stability, such as RPD3-dependent loading of condensin onto the rDNA array in response to nutrient signaling [54]. TOR signaling also promotes the synthesis of ribosomal proteins, and downregulation of ribosomal biogenesis can extend the lifespan of both yeast [17] and C. elegans [22,23]. These data suggest that TOR signaling may act to promote lifespan via multiple pathways that act in parallel to promote longevity (Figure 8).
Our analysis of the responsiveness of STRE-containing genes found ten genes, including PNC1, that are upregulated more than 2-fold in response to a slight decrease in the glucose concentration (2% to 1.75%) (Table S1). In general, the genes in this category are highly sensitive to environmental stresses, including heat shock and osmotic stress [46]. We speculate that other genes in this category, which includes both metabolic and heat shock genes, may also play a role in lifespan extension. Heat shock proteins in particular have been shown to promote longevity in numerous organisms, and are upregulated during CR in rodents [55,56].
Interestingly, MSN2/4 have also been shown to be required for the extension of yeast chronological lifespan [57]. MSN2/4 are responsible for the activation of numerous stress-responsive genes, including the superoxide dismutase SOD2, a gene that promotes chronological lifespan [34]. Yet, overexpression of SOD2 shortens replicative lifespan, and it has been demonstrated that deletion of MSN2/4 can actually lead to increases in replicative lifespan [58]. Even though we saw no such effect (Figure 1A), perhaps because of a difference in strain background, there may be a reciprocal relationship between replicative and chronological lifespan. A recent study showed that deletion of SIR2 can extend chronological lifespan in several strains [59], and we have observed that overexpression of SIR2 or HST2 shortens chronological lifespan in W303 (unpublished data).
The identification of the stress response factors Msn2p/4p as key components of the CR pathway in yeast supports two theories about CR. The first is known as the hormesis hypothesis of CR, which states that CR is a mild biological stress that provides health benefits because it activates an organism's defenses against adversity [60,61]. The second hypothesis is that the promoter elements of key longevity genes are just as important as the longevity genes themselves [28,62]. These promoters serve as sensors of the organism's environment by accepting different and additive inputs from environmentally sensitive transcription factors. The existence of short DNA sequences that dictate longevity could explain how new lifespans evolve so rapidly in response to a new ecological niche. Theoretically, if a transcription factor binding site regulates a key longevity gene, then a single base change might be sufficient to alter how long a species lives in response to an environmental condition.
In contrast to previous suggestions, we find that TOR and sirtuin signaling are components of the same longevity pathway that extends yeast replicative lifespan by stabilizing the repetitive rDNA (Figure 8). Given the high degree of functional conservation of TOR and sirtuins between yeast and higher organisms, and the recent discovery of a role for mammalian sirtuins in DNA repair [63], the findings in this study raise the possibility that the mammalian TOR pathway influences sirtuin activity and that together they may promote the health and longevity of mammals.
W303AR MATa, W303AR MATa pnc1::kanr, W303AR PNC1-GFP::kanr, and W303AR SIR2-3xHA::kanr have been previously described [24,29,64]. Gene disruptions in W303AR MATa were achieved by replacing the wild-type genes with the kanr, hphr, or natr marker as described [65,66] and verified by PCR using oligonucleotides flanking the genes. PNC1 was overexpressed as previously described [29]. W303AR cdc25–10 was created by replacing the endogenous copy of CDC25 with a plasmid-borne copy of cdc25–10 (the kind gift of S. J. Lin) as previously described [37]. pPNC1-STRE.4, pPNC1-STRE.2, and pPNC1-STRE.0 constructs and the pAdh-Msn2p-GFP/HA constructs were kindly provided by M. Ghislain [33] and C. Schuller [36], respectively. Msn4p-GFP and Msn2p(S4A)-GFP constructs were the kind gift of M. Jacquet. Plasmids for expression of Adr1p were obtained from E. Young. BY4741 deletions in this background were from F. Winston and P. Silver (Harvard), BY4742 and BY4742 hxk2::kanr were from B. Kennedy [18], W303 msn2Δ/4Δ and W303 msn2Δ/4Δ tetO-HSF1 were from H. Nelson [38]. All primer sequences, strains, and plasmid maps are available upon request.
Yeast were grown in yeast peptone dextrose (YPD) medium supplemented with an additional 0.015% w/v adenine, histidine, leucine, tryptophan, and uridine, and containing 2% w/v glucose during normal growth and 0.5% glucose for CR unless otherwise stated. For growth in low amino acid medium, synthetic complete medium containing 0.03 % w/v essential amino acids and 2% glucose was used. Strains were pre-grown overnight at 30 °C. The following day, cells were inoculated at an O.D. 600 = 0.1 and grown until log phase of growth was attained during the various conditions mentioned (O.D. 600 = 0.7). For treatment with rapamycin or heat shock, cells were grown for 2 h untreated, at which point rapamycin was added to a final concentration of 1 nM, or cells were moved to 37 °C, and cells were then grown for an additional 2 h.
rDNA recombination rates were determined by determining the frequency of loss of ADE2 in the rDNA of strain W303AR as previously described [15,29]. For rapamycin recombination assays, cells were grown for 2 h without rapamycin, followed by growth with 1 nM rapamycin for 2 h. More than 6,000 colonies were examined for each strain. Results are average values and standard deviation of at least three experiments. For replicative lifespan analyses, strains were pre-grown overnight on YPD plates unless otherwise noted. All lifespan analyses were carried out by using micromanipulation as previously described [14], and all micromanipulation dissections, including for cells grown under heat stress (37 °C), were carried out at laboratory temperature. For cells treated with rapamycin, yeast that growth-arrested in the G1 phase of the cell cycle due to toxicity [67] within the first nine divisions were excluded from the datasets. Statistical analysis was carried out using the JMP-IN statistics package (SAS, http://www.sas.com/). Wilcoxon rank-sum test p-values were calculated for each pair of lifespans, as shown in Table S3.
For the observation of nuclear migration of Msn2p-GFP or Msn4p-GFP, yeast were grown in YPD for 30 min at 1%, 0.5%, 0.1%, and 0.05% glucose (w/v), after first pre-growing to log phase in 2% glucose. Nuclei in live cells were stained with Hoechst #33342 (Sigma-Aldrich, http://www.sigmaaldrich.com/). Time-lapse photomicrographs were captured every 30 s using 1-s exposures. Image analysis was performed using the imageJ software package (National Institutes of Health, http://www.nih.gov/) in order to calculate the ratios of average nuclear intensity versus average cytoplasmic intensity.
Whole cell extracts were used to assay β-galactosidase activity as described [33]. Enzymatic activity is expressed as nanomoles o-nitrophenol-β-D-galactopyranoside cleaved per minute per milligram total protein.
Rabbit anti-Pnc1p polyclonal antibodies were generated by immunization of rabbits (Covance, http://store.crpinc.com/) with recombinant protein, and fresh serum was used at a dilution of 1:5,000. Mouse monoclonal anti-β-tubulin antibody (MAB3408, clone KMX-1, Upstate, http://www.upstate.com/), mouse monoclonal anti-actin antibody (Upstate/Chemicon MAB1501) and polyclonal rabbit anti-HA antibody (Abcam, http://www.abcam.com/) were used at a dilution of 1:1,000. Anti-rabbit (Amersham, http://www.amersham.com/) and anti-mouse (Amersham) horse radish peroxidase–conjugated antibodies were used at dilutions of 1:7,000.
The chromatin immunoprecipitation procedure was a modification of the method described by Strahl-Bolsinger et al. [68]. Changes to the protocol are as follows: 100 ml of cells (2.0 × 107 cells/ml) was cross-linked with 2% formaldehyde for 15 min at room temperature. Glycine was added to a final concentration of 250 mM, and the incubation continued for an additional 5 min. The suspension was sonicated seven times for 10 s with the amplitude set at 30% using a Branson model 450 digital sonifier (Branson, http://www.bransonultrasonics.com/). The suspension was clarified by centrifugation for 5 min, maximum setting, at 4 °C in a microcentrifuge. Samples were incubated on ice for 2 min between pulses. Then 1 μl of RNase (10 μg/ul) was added to samples, and they incubated for 30 min at 37 °C. Afterwards, sheared chromatin was purified using QIAquick spin columns (Qiagen, http://www1.qiagen.com/). Then 250 μl of supernatant was incubated with 15 μl of rabbit anti-HA antibody (Abcam). For the PCR analysis, the actin control primers used were as follows: ACT1-Chip-Fwd, GCCTTCTACGTTTCCATCCA [69], and ACT1-Chip-Rev, GGCCAAATCGATTCTC AAAA [69]. The PNC1 promoter primers used were as follows: Pnc1p-Chip-Fwd, GATCAAGGTGGCACACAGGG, and Pnc1p-Chip-Rev, ATACATAGTGGGCCAAACGG. The PCR protocol used was one cycle with 2 min at 95 °C, 30 s annealing at 55 °C, and a 1-min extension at 72 °C, followed by 30 cycles with 30 s at 95 °C, 30 s annealing at 55 °C, and 1-min extension at 72 °C. A final extension was performed for 4 min at 72 °C. Specific binding of Msn2p-HA to the endogenous PNC1 promoter (PNC1p) was analyzed by calculating the ratio of the percent IP of PNC1p to the percent IP of ACT1, using the V4.2.2 Quantity One 1-D analysis package (Bio-Rad, http://www.bio-rad.com/).
|
10.1371/journal.ppat.1002659 | Proteolytic Processing of Nlrp1b Is Required for Inflammasome Activity | Nlrp1b is a NOD-like receptor that detects the catalytic activity of anthrax lethal toxin and subsequently co-oligomerizes into a pro-caspase-1 activation platform known as an inflammasome. Nlrp1b has two domains that promote oligomerization: a NACHT domain, which is a member of the AAA+ ATPase family, and a poorly characterized Function to Find Domain (FIIND). Here we demonstrate that proteolytic processing within the FIIND generates N-terminal and C-terminal cleavage products of Nlrp1b that remain associated in both the auto-inhibited state and in the activated state after cells have been treated with lethal toxin. Functional significance of cleavage was suggested by the finding that mutations that block processing of Nlrp1b also prevent the ability of Nlrp1b to activate pro-caspase-1. By using an uncleaved mutant of Nlrp1b, we established the importance of cleavage by inserting a heterologous TEV protease site into the FIIND and demonstrating that TEV protease processed this site and induced inflammasome activity. Proteolysis of Nlrp1b was shown to be required for the assembly of a functional inflammasome: a mutation within the FIIND that abolished cleavage had no effect on self-association of a FIIND-CARD fragment, but did reduce the recruitment of pro-caspase-1. Our work indicates that a post-translational modification enables Nlrp1b to function.
| Inflammasomes are multi-protein complexes that respond to signals derived from microbial pathogens or damaged tissue. The function of an inflammasome is to activate pro-caspase-1, a protease that contributes to the inflammatory response by generating the cytokines IL-1β and IL-18. A common feature of inflammasomes is their ability to cluster multiple copies of pro-caspase-1 in a manner that allows inter-molecular auto-proteolysis. The Nlrp1b inflammasome assembles in response to anthrax lethal toxin by using two oligomerization regions: the NACHT domain and the FIIND-CARD region. Here, we demonstrate that the FIIND is proteolytically cleaved, but that the two fragments of Nlrp1b generated from the cleavage remain associated with one another. Cleavage within the FIIND is functionally important, however, because mutants of Nlrp1b that are not cleaved are not able to activate pro-caspase-1. Furthermore, we were able to control cleavage by inserting a heterologous protease site into Nlrp1b, which allowed us to establish that processing of Nlrp1b is required for its activity. Finally, we provide evidence that processing of Nlrp1b facilitates the recruitment of pro-caspase-1. Our work identifies a novel mechanism by which the Nlrp1b inflammasome may be regulated.
| Inflammasomes are multi-protein complexes that facilitate the activation of pro-caspase-1 in response to pathogen associated molecular patterns (PAMPS) or endogenous danger associated molecular patterns (DAMPS). A common feature of inflammasomes is that they recruit multiple copies of pro-caspase-1, which allows auto-proteolytic processing to generate active caspase-1 that cleaves downstream targets including pro-IL-1β and pro-IL-18. There are distinct mechanisms by which inflammasomes are thought to cluster molecules of pro-caspase-1. AIM2 has a HIN domain that binds cytosolic DNA derived from viruses or intracellular bacterial pathogens; it is the binding of several molecules of AIM2 to the same fragment of cytosolic DNA that leads to the grouping of pro-caspase-1 via the ASC adaptor [1]–[3]. In contrast, NLRP3 and NLRC4 each have a NACHT domain that self-associates to assemble a pro-caspase-1 activation platform after their respective triggers release auto-inhibitory intra-molecular interactions. Formation of the NLRP3 inflammasome has been proposed to occur in response to lysosomal permeabilization or reactive oxygen species generation [4]–[6], although recent evidence suggests that reactive oxygen species may serve only to induce the expression of NLRP3 [7]. The NLRC4 inflammasome detects bacterial flagellin and secretion system components [8], [9]. Human NLRP1 and the murine homolog, Nlrp1b, have NACHT domains as well as a Function to Find Domain (FIIND) that facilitate self-association [10], [11]. Nlrp1b indirectly senses the proteolytic activity of anthrax lethal toxin (LeTx) and there is some evidence that NLRP1 may detect peptidoglycan [12]–[16].
LeTx is comprised of a protease, lethal factor (LF), and a second component, protective antigen (PA), which binds mammalian cells and translocates LF to the cytosol [17], [18]. The proteolysis of multiple members of the MAPKK family by LF results in downregulation of the ERK, p38 and JNK signaling pathways. Interference of these pathways explains many of the pathogenic effects of LeTx, such as the inhibition of cytokine expression and chemotaxis, but there is no clear link between inactivation of these targets and activation of the Nlrp1b inflammasome [19]. Activation of Nlrp1b by LeTx causes macrophages to undergo a form of caspase-1 dependent cell death known as pyroptosis [20]. Interestingly, induction of pyroptosis in macrophages promotes the survival of mice infected with Bacillus anthracis by initiating IL-1β signaling and neutrophil recruitment [21], [22].
There are five alleles of Nlrp1b. Murine macrophages that express either allele 1 or 5 are susceptible to LeTx-mediated pyroptosis, but those that express allele 2, 3, or 4 are resistant [13]. The alleles are highly polymorphic: amino acid differences are found within the NACHT domain, the leucine rich repeat (LRR) domain, the FIIND, the caspase recruitment domain (CARD), as well as inter-domain regions (Figure 1). Although there has been little work done to examine the differences between the alleles, we previously developed a reconstituted cell-based system and confirmed that LeTx activated Nlrp1b allele 1 (Nlrp1b1, AAZ40509.1), but not Nlrp1b allele 3 (Nlrp1b3, AAZ40520.1) – there is a functional difference(s) between these proteins because they were expressed at similar levels [11]. We also noted in immunoblots that expression of Nlrp1b1 yielded two products, which could be interpreted as a cleavage event occurring within the protein, whereas Nlrp1b3 yielded only one product [11]. Recently, the human FIIND has been demonstrated to undergo autoproteolysis [23]. D'Osualdo and colleagues detected sequence similarity between the FIIND and the ZU5-UPA domain found in the autoproteolytic protein PIDD, which is activated by DNA damage. In both NLRP1 and PIDD, it is thought that a catalytic serine attacks a strained segment of the polypeptide backbone to generate two protein fragments that remain associated. Autoproteolysis of PIDD affects downstream signaling [24], but the consequence of NLRP1 autoproteolysis is not known.
Here, we have studied the FIIND of Nlrp1b. We made a series of truncation mutants to identify the minimal region of the FIIND that can drive oligomerization of the CARD domain and activation of pro-caspase-1. A multiple alanine substitution mutant of this Nlrp1b1 fragment impaired self-association and pro-caspase-1 activation. The same mutation abolished function of full-length Nlrp1b1 and prevented cleavage within the FIIND, so we hypothesized that cleavage of Nlrp1b is required for its function. Comparison of the Nrlp1b1 and Nlrp1b3 sequences allowed us to identify a single amino acid that determines the difference in their susceptibilities to cleavage. We demonstrated that cleavage is required for activity by introducing a heterologous TEV protease site into a cleavage-deficient mutant of Nlrp1b1 and showing that activation required co-expression of TEV protease. Finally, we provide evidence that cleavage of Nlrp1b1 does not facilitate its self-association, but does enhance the recruitment of pro-caspase-1.
Previous work demonstrated that Nlrp1b11086–1233 activates pro-caspase-1 in the absence of LeTx [11]. To identify the smallest constitutively active fragment of Nlrp1b1, we tested various Nlrp1b1 truncation mutants for their potential to activate pro-caspase-1 by reconstituting the Nlrp1b1 inflammasome in HT1080 fibroblasts. Various Nlrp1b1 truncation mutants were transfected into HT1080 cells along with pro-caspase-1 and pro-IL-1β and then the supernatants were probed for HA-tagged IL-1β by immunoblotting. In agreement with previous findings, IL-1β was detected in supernatants of cells expressing Nlrp1b11086–1233, which contains part of the FIIND plus the CARD domain, but not the CARD domain alone, Nlrp1b11142–1233 (Figure 2A). Nlrp1b11100–1233 induced IL-1β secretion, whereas Nlrp1b11102–1233 exhibited little activity. These results indicate, therefore, that Nlrp1b11100–1233 is the smallest fragment of Nlrp1b1 that can optimally activate pro-caspase-1.
We next performed co-immunoprecipitation experiments to test self-association of the Nlrp1b1 truncation mutants. HT1080 cells were transfected with His6-Nlrp1b1-HA and His6-Nlrp1b1-T7 vectors containing the truncation mutants Nlrp1b11086–1233, Nlrp1b11100–1233, Nlrp1b11102–1233 and Nlrp1b11142–1233. HA-tagged proteins were immunoprecipitated and associated T7-tagged proteins were detected by immunoblotting (Figure 2B). Nlrp1b11086–1233 and Nlrp1b11100–1233 self-associated, but Nlrp1b11102–1233 and Nlrp1b11142–1233 did not (Figure 2B). These data support an induced-proximity model in which the self-association of Nlrp1b1 is required to activate pro-caspase-1.
We next substituted residues 1100EIKLQIK1106 to alanine in Nlrp1b11100–1233 and found that this mutant, Nlrp1b11100–1233-7A, was impaired in its ability to self-associate and activate pro-caspase-1 (Figure 3A, B). The mutation did not, however, affect binding of catalytically inactive pro-caspase-1-C284A, which demonstrates that Nlrp1b11100–1233-7A was not grossly misfolded (Figure 3C). Notably, Nlrp1b11100–1233 and Nlrp1b11100–1233-7A bound pro-caspase-1-C284A, whereas the CARD domain alone (Nlrp1b11142-123) did not, indicating that a region N-terminal to the CARD domain facilitates the recruitment of pro-caspase-1. The multiple alanine substitution mutation was also introduced into full-length Nlrp1b1; Nlrp1b1-7A was not able to active pro-caspase-1 in response to LeTx (Figure 3D). Interestingly, Nlrp1b1-7A appeared as a single band in an immunoblot in contrast to the two bands that are observed upon expression of wild-type Nlrp1b1 (Figure 3E). These results suggest a relationship between self-association of the FIIND, cleavage of the FIIND, and activation of pro-caspase-1.
We sought to determine if endogenous Nlrp1b1 is proteolytically processed. First, we knocked down Nlrp1b1 in the murine macrophage cell line J774A.1 and assayed for cell death in response to LeTx. J774A.1 cells that were stably expressing Nlrp1b1 shRNA showed considerable protection from LeTx-induced cell death compared to cells expressing control shRNA, indicating that the knockdown was functionally effective (Figure 4A). We next probed these lysates with an antibody raised against the N-terminus of Nlrp1b1. Although the Nlrp1b1 antibody detects several proteins non-specifically, comparing lysates made from J774A.1 cells stably expressing either control shRNA or Nlrp1b shRNA demonstrated that the knockdown of Nlrp1b1 led to the disappearance of two bands (Figure 4B). These bands had slightly lower molecular weights than the bands corresponding to transfected TAP-tagged Nlrp1b1, which is predicted to be 8 kDa larger than endogenous Nlrp1b1. That the difference in molecular weight is similar between the two endogenous bands and the two TAP-tagged bands suggests that endogenous Nlrp1b1 is cleaved at the same location as transfected Nlrp1b1.
Because the TAP tag antibody and the Nlrp1b antibody both detect epitopes at the N-terminus of Nlrp1b, the two bands observed on the immunoblots likely correspond to full-length Nlrp1b1 and the N-terminal fragment of cleaved Nlrp1b1. To detect the C-terminal fragment, we generated an Nlrp1b1 construct with an additional TAP tag at the C-terminus. As predicted, expression of this doubly tagged protein yielded three bands: full-length Nlrp1b1, the N-terminal fragment, and the C-terminal fragment (Figure 4C). Judging by the molecular weight of the C-terminal fragment (estimated to be ∼30 kDa without the TAP tag), cleavage of Nlrp1b1 occurs in the FIIND.
Based on the size of the C-terminal fragment and on the observation that Nlrp1b1 is cleaved whereas Nlrp1b3 is not cleaved, we speculated that an amino acid difference(s) between Nlrp1b1 and Nlrp1b3 in the region near amino acids 970–1120 accounts for the lack of cleavage of Nlrp1b3. Six amino acid differences were found near this region (Figure 5A), so we generated 6 single swap mutants of Nlrp1b1. Of the 6 mutations, only the V988D mutation resulted in a loss of cleavage of Nlrp1b1 (Figure 5B), suggesting that this difference between the two proteins is responsible for the lack of cleavage of Nlrp1b3. Notably, this mutation also resulted in a complete loss of inflammasome activity in response to LeTx, as assessed by the loss of mature IL-1β in the cell supernatant (Figure 5C). Several other mutants (most markedly A996D and N1026S) resulted in decreased inflammasome activity without affecting cleavage of Nlrp1b1. These data suggest that the non-responsiveness of Nlrp1b3 to LeTx is not only a consequence of it not being cleaved.
To determine if a single mutation could lead to the cleavage of Nlrp1b3, a mutant containing the reverse of the V988D mutation, Nlrp1b3-D927V, was made (the difference is numbering is due to an insertion in Nrlp1b1). Nlrp1b3-D927V was cleaved, but it was not able to activate pro-caspase-1 in response to LeTx (Figure 5D, E). In an attempt to generate a functional version of Nlrp1b3, we introduced swap mutations based on the mutations that impaired function of Nrlp1b1 (A996D and N1026S). Nlrpb3-D935A/S965N did not exhibit any activity, but Nlrp1b3-D927V/D935A/S965N was cleaved and displayed activity (Figure 5F, G). This triple mutant was constitutively active, presumably because the mutations also interfere with auto-inhibition of the protein.
We next sought to determine whether proteolytic processing of Nlrp1b1 is important for its activity by introducing a heterologous protease site whose cleavage could be controlled. A cleavage site for the tobacco etch virus (TEV) NIa protease was inserted between amino acids 981 and 982 of Nlrp1b1-V988D. Importantly, TEV protease has high sequence specificity and is well tolerated by mammalian cells [25]. Nlrp1b1, Nlrp1b1-V988D, and Nlrp1b1-V988D-TEV were each co-expressed with pro-caspase-1 and pro-IL-1β. As expected, LeTx caused the release of IL-1β by cells that expressed Nlrp1b1 regardless of whether TEV protease was expressed (Figure 6A). TEV protease did not cause the release of IL-1β in the absence of LeTx. Nlrp1b1-V988D was not functional in any of the conditions tested, whereas Nlrp1b1-V988D-TEV activated pro-caspase-1, but only in the presence of TEV protease. That TEV protease was able to promote pro-caspase-1 activation by Nlrp1b1-V988D-TEV in the absence of LeTx indicated that TEV cleavage overrides auto-inhibition of Nlrp1b1.
Cleavage of Nlrp1b1-V988D-TEV was not observed under conditions similar to those used to test activity (with the exception that the catalytically inactive mutant pro-caspase-1-C284A was used in order to prevent release of inflammasome components from the cell), indicating that only a small fraction of Nlrp1b1-V988D-TEV may have been cleaved (Figure 6B). To confirm that direct cleavage of Nlrp1b-V988D-TEV by TEV protease was responsible for the observed activity, rather than an indirect effect of the protease, a TEV-site mutant was generated. Nlrp1b1-V988D-TEVmut has a Q to H mutation in the consensus sequence ENLYFQS, which has been shown to prevent cleavage by TEV protease [26]. Because cleavage by TEV protease could not be visualized under the conditions used to test inflammasome activation (Figure 6B), 5 times as much plasmid encoding TEV protease was transfected with the Nlrp1b1 constructs. Under these conditions, a small amount of cleaved product of Nlrp1b1-V988D-TEV was observed, whereas no cleavage of Nlrp1b1-V988D-TEVmut was detected (Figure 6C). Nlrp1b1-V988D-TEV exhibited activity when co-expressed with TEV protease, but Nlrp1b1-V988D-TEVmut did not (Figure 6D). Cumulatively, these results provide strong evidence that cleavage of Nlrp1b1 is important for its function.
We next wanted to determine whether the N-terminal and C-terminal fragments of Nlrp1b1 are associated in the auto-inhibited state and if activation of Nlrp1b1 by LeTx causes the C-terminal fragment to be released as an activated moiety. Nlrp1b1 was expressed with an N-terminal TAP tag and a C-terminal HA tag. When Nlrp1b1 was precipitated using the TAP tag, the C-terminal fragment co-precipitated (Figure 7). Treatment of cells with LeTx for 1–2 h did not reduce the amount of C-terminal fragment precipitated. Treatment of cells with LeTx for 3 h slightly reduced the amount of the C-terminal fragment precipitated, although this was because less was present in the lysate possibly as a result of inflammasome secretion. Thus, LeTx does not cause cleaved Nlrp1b1 to dissociate.
We next sought to determine how cleavage within the FIIND promotes inflammasome assembly. To ascertain whether processing within the FIIND facilitates the ability of the FIIND-CARD fragment to self-associate or to recruit pro-caspase-1, we used the constructs Nlrp1b1720–1233 and Nlrp1b1720–1233-V988D, which contain the entire FIIND-CARD region. Nlrp1b1720–1233 displayed constitutive activity as we have shown previously [11], but the V988D mutant had minimal activity (Figure 8A). An experiment was then performed in which T7-tagged Nlrp1b1 constructs were immunoprecipitated and the co-precipitation of HA-tagged Nlrp1b1 constructs and FLAG-tagged pro-caspase-1-C284A was tested. Nlrp1b1720–1233 self-associated in the presence or absence of pro-caspase-1-C284A (Figure 8B). Nlrp1b1720–1233 also complexed with pro-caspase-1-C284A. Nlrp1b1720–1233-V988D exhibited a slightly higher level of self-association compared to the wild-type construct, but bound less pro-caspase-1-C284A. These experiments suggest that cleavage within the FIIND facilitates the formation of an optimally assembled platform that can recruit pro-caspase-1.
NLRP1 and Nlrp1b appear to be unique among the NOD-like receptors in that they possess two regions that facilitate oligomerization: a NACHT domain and a FIIND-CARD region. The NACHT domain is a member of the well-characterized AAA+ ATPase family [27], while there is only limited information available on the FIIND. In this study, we discovered that cleavage of the Nlrp1b FIIND is functionally important for inflammasome activity and that this processing of the FIIND facilitates the recruitment of pro-caspase-1.
Previously, we noted that immunblotting for Nlrp1b1 using an antibody against the N-terminal TAP tag detected two bands [11]. We speculated that the upper band represented the full-length protein and that the lower band resulted from a proteolytic event. By generating a construct with tags at both termini of Nlrp1b1, we were able to detect the C-terminal fragment and determine that it remained associated with the N-terminal fragment (Figure 7). That the processing of Nlrp1b1 might be important for function was hinted at by the finding that the LeTx-non-responsive Nrlp1b3 was not cleaved. Furthermore, a substitution mutation that we designed to impair self-association of the FIIND unexpectedly prevented Nlrp1b1 cleavage. The functional importance of FIIND cleavage was established by an experiment in which the TEV protease site was introduced at a position within Nlrp1b1 that we estimated to be near the natural cleavage site – TEV protease cleaved this site and activated the inflammasome.
During the preparation of this manuscript, D'Osualdo and colleagues published that the FIINDs of human NLRP1 and CARD8 undergo autoproteolytic processing [23]. These researchers used computational approaches to detect similarity between the FIIND and the ZU5-UPA domains found in the PIDD autoprotease. PIDD forms a complex called the PIDDosome that either activates pro-casapase-2 or NF-κb in response to DNA damage. PIDD that is autoproteolytically cleaved at a single site initiates activation of NF-κb; an additional autoproteolytic event facilitates activation of pro-caspase-2 [24]. Autoproteolysis at each site occurs before a serine residue – the hydroxyl group of the serine is believed to attack a strained backbone to initiate the breakage of the peptide bond.
Cleavage of the CARD8 FIIND was demonstrated to occur after a phenylalanine residue in a conserved SFS motif [23], which corresponds to 982SFS984 in Nlrp1b1. Notably, we found that the nearby V988D mutation in Nlrp1b1 abolished cleavage and the corresponding swap mutation in Nlrp1b3 caused this protein to be cleaved. Furthermore, the TEV site that allowed activation of Nlrp1b1-V988D-TEV by TEV protease was inserted between amino acids 981 and 982. Collectively, these findings suggest that the V988D mutation disrupts the protein conformation required for autoproteolysis between amino acids 983 and 984, but that activity can be rescued by cleavage at a nearby location. We believe that Nlrp1b1 undergoes autoproteolysis by the same mechanism as CARD8 because Nlrp1b1-S984A, in which the predicted catalytic serine is mutated, is not cleaved and is not functionally active (Figure S1). We further note that although lethal factor is itself a protease, the toxin does not affect Nlrp1b1 cleavage and, therefore, we do not believe that Nlrp1b1 is a direct target.
Only approximately half of overexpressed Nlrp1b1 and endogenous Nlrp1b1 is cleaved. We did not observe an increased amount of cleavage over time (data not shown), which suggests that only a fraction of Nlrp1b1 is capable of undergoing autoproteolysis. We speculate, therefore, that auto-inhibited Nlrp1b1 exists as a dimer in which one of the monomers is cleaved and the other is not cleaved. Dimerization might induce autoproteolysis in one of the monomers and not in the other, which could explain why mutation of amino acids 1100–1106 not only prevents self-association of Nlrp1b11100–1233, but also prevents cleavage of Nlrp1b1.
To address the mechanistic consequence of FIIND processing, we compared Nlrp1b1720–1233 to Nlrp1b1720–1233-V988D, which contain the complete FIIND and CARD domains. Nlrp1b1720–1233 is cleaved and is constitutively active; the V988D mutation abolishes both cleavage and activity. We found that the V988D mutation does not impair self-association – and in fact seems to increase self-association to a small extent – but the mutation does impair the recruitment of pro-caspase-1. We interpret these data to mean that cleavage of the FIIND is required for the proper assembly of a platform that is capable of recruiting molecules of pro-caspase-1 in orientations that allow cross-proteolysis. Incorrect assembly of uncleaved FIIND-CARD may partially occlude the pro-caspase-1 binding site because self-association of FIIND-CARD does not appear to be required for binding pro-caspase-1, as demonstrated by binding of pro-caspase-1 to oligomerization-defective Nlrp1b11100–1233-7A. Because Nlrp1b11100–1233 is active even though it lacks the region that undergoes proteolysis, it is presumably the uncleaved region N-terminal to amino acids 1100–1233 that exerts a negative restraint that prevents correct assembly.
HT1080 cells and J774A.1 cells (ATCC) were cultured in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum and 1% penicillin-streptomycin. PA and LF were purified as described previously [28].
Rabbit antibody was raised against an N-terminal epitope of Nlrp1b1, MEESPPKQKSNTKVAQHE. Membranes were probed with the following antibodies: anti-Nlrp1b polyclonal antibody (1∶5000), anti-caspase-1 p10 (M20) polyclonal antibody (1∶500, Santa Cruz Biotechnology sc-514), anti-HA polyclonal antibody (1∶1000, Santa Cruz Biotechnology sc-805), anti-T7-tag monoclonal antibody (1∶1000, Novagen 69522), anti-β-actin monoclonal antibody (1∶10000, Sigma Aldrich A-5441), and anti-calmodulin binding peptide (CBP) antibody (1∶1000, Upstate 07-482). For immunoprecipitations, anti-T7-tag monoclonal antibody (Novagen 69522), and anti-HA monoclonal antibody (Sigma H9658) were used. Anti-green fluorescent protein (GFP, Covance MMS-118R) was used as a control.
Construction of plasmids pcDNA3-T7, pcDNA3-His6-HA, pcDNA3-His6-T7, pN-TAPA Nlrp1b1, pcDNA3-pro-caspase-1-T7, pcDNA3-IL-1β-HA were described previously [11]. All substitution mutations were made using Quikchange mutagenesis (Stratagene).
To construct pcDNA3-pro-caspase-1-FLAG, pcDNA3-pro-caspase-1-T7 was cut with Apa1 and Nhe1 to remove the T7 tag. FLAG-tag oligonucleotide was constructed by annealing forward oligonucleotide 5′-CAT GGA CTA CAA GGA CGA CGA TGA CAA GG-3′ and reverse oligonucleotide 5′-CTA GCC TTG TCA TCG TCG TCC TTG TAG TCC ATG GGC C-3′. The resulting annealed oligonucleotide was ligated at Apa1 and NheI restriction sites of the restriction digested pcDNA3-pro-caspase-1-T7 vector.
pN-TAPB-T7 plasmid was constructed by annealing T7-tag forward oligonucleotide 5′- TCG AGA TGG CTA GCA TGA CTG GTG GAC AGC AAA TGG GTT AGG GGC C-3′ and reverse oligonucleotide 5′-CCT AAC CCA TTT GCT GTC CAC CAG TCA TGC TAG CCA TC-3′. The resulting annealed T7 oligonucleotide was ligated at Xho1 and Apa1 restriction sites of pN-TAPB.
Nlrp1b truncation plasmids were constructed by amplifying fragments from pN-TAPA -Nlrp1b1. The PCR products were digested with the restriction enzymes BamHI and XhoI, and the resulting products were ligated into vectors pN-TAPB-T7, pcDNA3-His6-HA and pcDNA3-His6-T7.
pC-TAPA-Nlrp1b1 was constructed by amplifying Nlrp1b1 with the forward primer 5′-GCG GGA TCC GCC GCC ACC ATG GAA GAA TCC CCA CCC AAG-3′ and the reverse primer 5′-GCG CTC GAG TGA TCC CAA AGA GAC CCC AC-3′. The PCR product was digested with BamHI and XhoI and ligated into pC-TAPA.
pN-TAPA-Nlrp1b1-TAP was constructed by amplifying the C-terminus of Nlrp1b1 and the TAP tag from the pC-TAPA-Nlrp1b1 plasmid using the forward primer 5′-CGC ACC CAA GCT TCT CCC CAA TGG-3′ and the reverse primer 5′-CGC CTC GAG CTA AAG TGC CCC GGA GGA TG-3′. The PCR product and pN-TAPA-Nlrp1b1 were digested with HindIII and XhoI. The vector backbone of pN-TAPA including the N-terminus of Nlrp1b1 were isolated by gel extraction (Qiagen) and ligated with the PCR product.
The NIa protease of the tobacco etch virus (TEV protease) was amplified using the forward primer 5′-CGC GGT ACC GCC GCC ACC ATG GGA TCC AGC TTG TTT AAG GGA C-3′ and the reverse primer 5′-CGC TCT AGA GTC ACG ATG AAT TCC GGG CGA G-3′. The PCR product was digested with KpnI and XbaI and ligated into pcDNA3-T7. In order to reduce the self-cleavage and increase the catalytic efficiency of TEV protease a S219V mutation [29] was introduced using the forward primer 5′-GGG GCC ATA AAG TTT TCA TGG TCA AAC CTG AAG AGC CTT TTC-3′ and the reverse primer 5′-GAA AAG GCT CTT CAG GTT TGA CCA TGA AAA CTT TAT GGC CCC-3′.
The pN-TAPA-Nlrp1b1-V988D-TEV construct was cloned in two steps. The N-terminus of Nlrp1b1 was amplified with the forward primer 5′-CGC GGA TCC TAT GGA AGA ATC CCC ACC CAA G-3′ and the reverse primer (which includes coding for the TEV-site) 5′-CGC ATC GTC GAC TGG AAG TAG AGA TTC TCT GGG TTT TTC AGT ACT GTG TAT CC-3′. The C-terminus of Nlrp1b1 was amplified from pNTAP-Nlrp1b1-V988D with the forward primer 5′-CGC ATC GTC GAC GAG CTT CTC CCC AAT GGG AGA TG-3′ and the reverse primer 5′-CGC CTC GAG TCA TGA TCC CAA AGA GAC CCC ACC TG-3′. The PCR fragments were digested with SalI and ligated. After gel extraction, the ligated PCR fragments were digested with BamHI and XhoI and ligated into pN-TAPA.
To construct pcDNA3-His6-Nlrp1b1-HA, Nlrp1b1 was amplified using the forward primer 5′-CGC GGA TCC ATG GAA GAA TCC CCA CCC AAG-3′ and the reverse primer 5′- CGC CTC GAG TGA TCC CAA AGA GAC CCC AC-3′. The PCR product was digested using BamHI and XhoI followed by ligation into pcDNA3-His6-HA.
The pN-TAPA-Nlrp1b1-HA construct was created by amplifying Nlrp1b1 including the C-terminal HA tag from pcDNA3- His6-Nlrp1b-HA. The forward primer used was 5′-CGC GGA TCC TAT GGA AGA ATC CCC ACC CAA G-3′ and the reverse primer was 5′-CGC ATC GTG GTC GAC TCA CAA GCT AGC GTA ATC TGG-3′. The PCR product was digested with BamHI and SalI and ligated into pN-TAPA.
pN-TAPB-Nlrp1b1720–1233-T7 was constructed by amplifying Nlrp1b1 using the forward primer 5′-GCG GGA TCC GAC CTG TCC TCT CTC AGT G-3′ and the reverse primer 5′-CGC CTC GAG TGA TCC CAA AGA GAC CCC AC-3′. The PCR product was digested with BamHI and XhoI and ligated into pN-TAPB-T7.
The pN-TAPB-Nlrp1b1720–1233-HA construct was created by amplifying the FIIND and CARD domains of Nlrp1b1 including the C-terminal HA tag from pcDNA3- His6-Nlrp1b-HA. The forward primer used was 5′-GCG GGA TCC GAC CTG TCC TCT CTC AGT GCC-3′ and the reverse primer was 5′-CGC ATC GTG GTC GAC TCA CAA GCT AGC GTA ATC TGG-3′. The PCR product was digested with BamHI and SalI and ligated into pN-TAPB.
Design and infection of shRNA were performed according to pLKO.1 protocol from Addgene. Nlrp1b1 shRNA was targeted against the 3′ UTR of Nlrp1b1. The Nlrp1b1 shRNA forward sequence was 5′-CCG GGG TTG TCT TTG TCT CTG TTG ACT CGA GTC AAC AGA GAC AAA GAC AAC C-3′ and the reverse sequence was 5′-AAT TCA AAA AGG TTG TCT TTG TCT CTG TTG ACT CGA GTC AAC AGA GAC AAA GAC AAC C-3′. The scrambled shRNA sequences were generated from the Nlrp1b1 shRNA sequence and the forward sequence was 5′-CCG GGG TCT TGT ATT CGG TTT CTG TCT CGA GAC AGA AAC CGA ATA CAA GAC CTT TTT G-3′ while the reverse sequence was 5′-AAT TCA AAA AGG TCT TGT ATT CGG TTT CTG TCT CGA GAC AGA AAC CGA ATA CAA GAC C-3′. Forward and reverse oligonucleotides were annealed and ligated into the pLKO.1 vector at the EcoRI and AgeI sites. pLKO.1 vector containing the shRNA was transfected with polyethylenimine, pH 7.2 into HEK-293T cells (ATCC). Approximately 48 h following transfection, cell medium was collected and filtered with a 0.45 µm filter and 2 mL of the lentiviral particle solution was added to a 10 cm dish containing J774A.1 cells. Approximately 24 h later, media was removed and media containing 5 µg/ml puromycin (Sigma) was added to J774A.1 cells to select cells that had stably integrated the shRNA expressing plasmid.
J774A.1 cells were seeded onto a 24-well plate at 1.4×105 cells per well. Approximately 24 h later cells were treated with LeTx (10−8 M PA and 10−9 M LF) for 4 h. Release of cytoplasmic LDH into the cell medium was measured using the CytoTox 96 nonradioactive cytotoxicity assay (Promega G-1780) in accordance with the manufacturer's instructions. The percent of LDH release was calculated as 100×(Experimental LDH−Spontaneous LDH)/(Maximum LDH−Spontaneous LDH). J774A.1 WT −LeTx and +LeTx values were set to 0% and 100% respectively, and scrambled and Nlrp1b shRNA J774A.1 results were normalized to those values. Results from three independent experiments were averaged.
Approximately one million HT1080 cells were seeded on a 10-cm dish the day before transfection. On the day of transfection, 1 µg each of pN-TAP-Nlrp1b1, pcDNA3-pro-caspase-1-FLAG (or T7), and pcDNA3-pro-IL-1β-HA was transfected using 9 µl of 1 mg/ml polyethylenimine, pH 7.2. Note that for experiments involving TEV protease an additional 1 µg of either pcDNA3-TEV-protease-T7 or pcDNA3 empty vector was transfected as indicated. Approximately 24 h after transfection, cells were treated with LF (10−8 M) and PA (10−8 M) for 3 h. The cell supernatant was mixed with 1 µl of anti- HA antibody (Sigma-Aldrich H9658) overnight, followed by the addition of 100 µl of BSA-blocked protein A Sepharose beads (GE Healthcare) and a 2 h incubation. Proteins were eluted from the protein A Sepharose beads with sodium dodecyl sulfate (SDS) loading dye and subjected to immunoblotting using a polyclonal HA antibody (Santa Cruz sc805). Cells from each 10-cm plate were scraped into 300 µl of EBC lysis buffer (50 mM Tris, pH 8, 150 mM NaCl, 0.5% [vol/vol] NP-40, 1 mM phenylmethylsulfonyl fluoride). Each sample was lysed by rotation at 4°C for 1 h or by sonication 3 times, each for 10 s followed by a 10 s incubation on ice. Lysates were clarified by centrifugation, and protein concentrations were determined using the Bradford assay. Equivalent amounts of cell lysate protein (∼40 µg) were subjected to SDS-polyacrylamide gel electrophoresis and immunoblotted with anti-HA, anti-T7, and anti-β-actin antibodies.
One to five 10-cm dishes of HT1080 cells were transfected with 1–4 µg of the indicated constructs. Approximately 24 h after transfection, cells were harvested and cell pellets from each plate were lysed with approximately 300 µl EBC buffer by sonication. Cell lysates were clarified by centrifugation. Equivalent amount of cell lysate protein was incubated with 25 or 50 µL streptavidin agarose resin (Thermo Scientific 20349) for ∼2 h or overnight. Beads were washed three times with 1 mL EBC buffer. Proteins were eluted with SDS and analyzed by immunoblotting with anti-Nlrp1b or anti-CBP antibody. CBP is an epitope found within the TAP tag.
To test self-association of Nlrp1b1 truncation mutants, two plates of HT1080 cells were transfected with pcDNA3-His6-Nlrp1b1-HA and pcDNA3-His6-Nlrp1b1-T7 vectors containing trunction mutants. Approximately 24 h following transfection, cells were lysed in 300 µL EBC buffer by sonication, and lysates were clarified by centrifugation. Equal amount of lysate protein was incubated overnight with 1 µL of anti-HA antibody (Sigma-Aldrich H9658), followed by the addition of 50 µL of BSA-blocked protein A Sepharose beads (GE Healthcare) and a 2 h incubation at 4°C. Complexes were resolved by SDS-polyacrylamide gel electrophoresis and immunoblotted using an anti-T7 antibody.
A similar protocol as above was followed to test for Nlrp1b1720–1233 self-association and pro-caspase-1 binding, with the exception that the vectors indicated in Figure 8 were transfected to a total of 4 µg. Lysates were incubated with 1 µl of anti-T7 antibody or 1 µL of control anti-GFP antibody overnight. Membranes were immunoblotted with anti-caspase-1 p10, anti-HA, anti-T7, and anti-β-actin antibodies.
A similar protocol as above was followed to test for the association of the Nlrp1b1 cleaved products with the exception that the vectors indicated in Figure 7 were transfected to a total of 4 µg. Lysates were incubated overnight with 50 µL streptavidin agarose resin slurry, followed by washing three times with 1 mL EBC buffer. Proteins were eluted from the beads with SDS and analyzed by immunoblotting with anti-HA and anti-β-actin antibodies.
For the experiment shown in Figure 3C, two 10-cm dishes of HT1080 cells were transfected with pcDNA3-pro-caspase-1-T7-C284A and either pcDNA3-His6-Nlrp1b11100–1233-HA, pcDNA3-His6-Nlrp1b11100–1233-7A-HA or pcDNA3-His6-Nlrp1b11100–1233-HA. Cells were lysed in 600 µL EBC buffer by sonication, and the lysates were clarified by centrifugation. Lysates were incubated with 1 µL of anti-T7 antibody or 1 µL of control anti-GFP antibody overnight, followed by the addition of 50 µL of protein A Sepharose for 2 h. Complexes were resolved by SDS-polyacrylamide gel electrophoresis and immunoblotted using an anti-HA antibody and anti-caspase-1 p10 antibody.
|
10.1371/journal.pmed.1002710 | Metabolic syndrome in pregnancy and risk for adverse pregnancy outcomes: A prospective cohort of nulliparous women | Obesity increases the risk for developing gestational diabetes mellitus (GDM) and preeclampsia (PE), which both associate with increased risk for type 2 diabetes mellitus (T2DM) and cardiovascular disease (CVD) in women in later life. In the general population, metabolic syndrome (MetS) associates with T2DM and CVD. The impact of maternal MetS on pregnancy outcomes, in nulliparous pregnant women, has not been investigated.
Low-risk, nulliparous women were recruited to the multi-centre, international prospective Screening for Pregnancy Endpoints (SCOPE) cohort between 11 November 2004 and 28 February 2011. Women were assessed for a range of demographic, lifestyle, and metabolic health variables at 15 ± 1 weeks’ gestation. MetS was defined according to International Diabetes Federation (IDF) criteria for adults: waist circumference ≥80 cm, along with any 2 of the following: raised trigycerides (≥1.70 mmol/l [≥150 mg/dl]), reduced high-density lipoprotein cholesterol (<1.29 mmol/l [<50 mg/dl]), raised blood pressure (BP) (i.e., systolic BP ≥130 mm Hg or diastolic BP ≥85 mm Hg), or raised plasma glucose (≥5.6 mmol/l). Log-binomial regression analyses were used to examine the risk for each pregnancy outcome (GDM, PE, large for gestational age [LGA], small for gestational age [SGA], and spontaneous preterm birth [sPTB]) with each of the 5 individual components for MetS and as a composite measure (i.e., MetS, as defined by the IDF). The relative risks, adjusted for maternal BMI, age, study centre, ethnicity, socioeconomic index, physical activity, smoking status, depression status, and fetal sex, are reported. A total of 5,530 women were included, and 12.3% (n = 684) had MetS. Women with MetS were at an increased risk for PE by a factor of 1.63 (95% CI 1.23 to 2.15) and for GDM by 3.71 (95% CI 2.42 to 5.67). In absolute terms, for PE, women with MetS had an adjusted excess risk of 2.52% (95% CI 1.51% to 4.11%) and, for GDM, had an adjusted excess risk of 8.66% (95% CI 5.38% to 13.94%). Diagnosis of MetS was not associated with increased risk for LGA, SGA, or sPTB. Increasing BMI in combination with MetS increased the estimated probability for GDM and decreased the probability of an uncomplicated pregnancy. Limitations of this study are that there are several different definitions for MetS in the adult population, and as there are none for pregnancy, we cannot be sure that the IDF criteria are the most appropriate definition for pregnancy. Furthermore, MetS was assessed in the first trimester and may not reflect pre-pregnancy metabolic health status.
We did not compare the impact of individual metabolic components with that of MetS as a composite, and therefore cannot conclude that MetS is better at identifying women at risk. However, more than half of the women who had MetS in early pregnancy developed a pregnancy complication compared with just over a third of women who did not have MetS. Furthermore, while increasing BMI increases the probability of GDM, the addition of MetS exacerbates this probability. Further studies are required to determine if individual MetS components act synergistically or independently.
| Obesity increases the risk for developing gestational diabetes mellitus (GDM) and preeclampsia (PE), which both associate with increased risk for type 2 diabetes mellitus (T2DM) and cardiovascular disease (CVD) in women in later life.
In the general population, metabolic syndrome (MetS) associates with T2DM and CVD.
The impact of maternal MetS on pregnancy outcomes in nulliparous pregnant women has not been thoroughly investigated.
We assessed the association between MetS, measured at 15 ± 1 weeks’ gestation, and a range of pregnancy complications in low-risk, nulliparous women recruited to the multi-centre, international prospective Screening for Pregnancy Endpoints (SCOPE) study.
Women with MetS in early pregnancy had an increased risk for GDM and PE, after adjustment for a range of demographic and lifestyle variables.
Diagnosis of MetS may be useful to broadly identify a group of women at risk for pregnancy complications, which may additionally associate with future CVD risk.
Future studies are required to substantiate our findings and to determine whether MetS is a useful discriminator of pregnancy complications in lean and obese women. Future studies are also encouraged to define clear metabolic health phenotypes that will produce the optimal probability threshold for each outcome.
| Obesity is an established risk factor for pregnancy complications, increasing risk for gestational diabetes mellitus (GDM), preeclampsia (PE), and delivering large for gestational age (LGA) infants, by 2- to 3-fold [1–4], and small for gestational age (SGA) infants, by 24% [5]. Such adverse outcomes place women and their infants at greater risk for cardiovascular and metabolic diseases both perinatally and in later life [6,7]. However, not all obese women experience a pregnancy complication [8], nor do they all develop chronic disease [9].
Metabolic syndrome (MetS) is a cluster of risk factors that encompasses metabolic, vascular, and inflammatory indicators. Although there are several definitions and cut-points used to describe and characterise MetS [10], the metabolic disturbances underpinning MetS are consistent and include atherogenic dyslipidemia, raised blood pressure (BP), insulin resistance, obesity, and pro-thrombotic and pro-inflammatory states. While some expert definitions deem obesity an essential criterion [11,12], other definitions largely focus on insulin resistance [11,13,14]. To date, there are no obligatory components to define MetS [11] but rather a constellation of risk factors that, irrespective of components and cut-offs, have consistently been demonstrated to increase risk for cardiovascular disease (CVD) [15], some cancers [16], type 2 diabetes mellitus (T2DM) [17], and chronic kidney disease [18] in the adult population.
Normal pregnancy is a pro-inflammatory, pro-thrombotic, highly insulin resistant [19], and hyperlipidemic state [20]. However, there are no recognised healthy metabolic variable cut-points in pregnancy. Studies that have assessed metabolic components in pregnancy have generally used accepted definitions for the adult population [21,22] or used a range of population-specific criteria [23–25]. Appropriate definitions that associate metabolic health parameters and pregnancy complications have yet to be defined, and the utility of using previously defined variables from the non-pregnant adult population is unclear.
Given the established links between MetS and chronic diseases in adulthood [15,17], as well as between pregnancy complications such as PE and GDM and later life T2DM and CVD [26,27], pregnancy may offer a window of opportunity to identify women with MetS and elevated risk of adverse pregnancy outcomes, as well as later life chronic disease. Most studies to date, however, have only assessed individual metabolic components in pregnancy: raised triglycerides (TGs) and low-density lipoprotein cholesterol (LDL-C) and reduced high-density lipoprotein cholesterol (HDL-C) are associated with increased risk for GDM, PE, LGA, and spontaneous preterm birth (sPTB) [24,25,28]. In a small study, multiparous Greek women who had MetS, determined in early pregnancy, were at a 3-fold increased risk for preterm birth [29].
Assessment of MetS and a broader range of pregnancy complications in nulliparous women has not been done. We hypothesized that women with MetS are at greater risk for pregnancy complications than women who do not have MetS. The aims of this study were to (i) determine the prevalence of MetS in women participating in the Screening for Pregnancy Endpoints (SCOPE) study and (ii) determine whether MetS is associated with maternal and neonatal outcomes including GDM, LGA, PE, SGA, and sPTB.
Participants recruited to SCOPE were nulliparous pregnant women recruited from Adelaide (Australia), Auckland (New Zealand), Cork (Ireland), Leeds (UK), London (UK), and Manchester (UK) (n = 5,628). SCOPE is a multi-centre prospective cohort study with the primary aim of developing screening tests for prediction of PE, sPTB, and SGA babies. Recruitment was between 11 November 2004 and 28 February 2011, and data collection occurred until 30 September 2011, when the last babies were born. Included women were nulliparous with singleton pregnancies. Women were excluded if they were considered to be at high risk for PE, SGA, or sPTB due to underlying medical conditions (e.g., chronic hypertension requiring antihypertensive medication or diabetes); if they had previous cervical knife cone biopsy; if they had 3 terminations or 3 miscarriages; if their pregnancy was complicated by a known major fetal anomaly or abnormal karyotype; or if they received interventions that may modify pregnancy outcome (e.g., aspirin or cervical suture). Women were also excluded if they were taking supplements of calcium (>1 g/d), eicosopentanoic acid (≥2.7 g/d), vitamin C (>1,000 mg/d), or vitamin E (>400 IU/d) or if they had diabetes (type 1 or type 2). An uncomplicated pregnancy was defined as a normotensive pregnancy, delivered at ≥37 weeks, resulting in a liveborn, non-SGA, and non-LGA baby. Pregnancies with other complications—such as placenta praevia, placental abruption, cholestasis of pregnancy, or other significant pregnancy complication—were not included in the uncomplicated pregnancy group.
Study data were obtained by a research midwife at 15 ± 1 weeks’ gestation, including demographics; smoking; family, medical, and gynaecological history; diet and supplement use; systolic and diastolic BP; height and weight (measured at 15 ± 1 weeks’ gestation, to determine BMI), and waist circumference (WC). The socioeconomic index (SEI) is a measure of the individual’s socioeconomic status and is derived from the specific occupation of the woman, producing a score between 10 and 90, with a lower score reflecting greater disadvantage [30]. The Edinburgh Postnatal Depression Scale was evaluated at 15 weeks’ gestation, and women were categorised as unlikely to experience depression (score <5), at increased risk of depression in the next year (score of 5–9), or likely depressed (score >9). Physical activity at 15 weeks’ gestation was categorised as none (no moderate or vigorous exercise, with ≤1 instance/day of recreational walking), light (no moderate or vigorous exercise, with >1 instance/day of recreational walking), or moderate or vigorous (some moderate or vigorous exercise, with or without recreational walking). Ethnicity was binary, coded as white or other. Smoking status was binary, coded as yes or no for any cigarette smoking at 15 ± 1 weeks’ gestation.
A non-fasting blood sample was taken for measurement of HDL-C and TGs at 15 ± 1 weeks’ gestation. Details of the immunoassay methodology for measuring lipids can be found in previous publications [31,32]. Plasma blood glucose was measured as a random blood sample by glucometer at 15 ± 1 weeks’ gestation. Additional information on data collection has been provided in detail previously [33]. Ethical approval was obtained from local ethics committees (New Zealand AKX/02/00/364; Australia REC 1712/5/2008; London, Leeds, and Manchester 06/MRE01/98; and Cork ECM5 [10] 05/02/08), and all women provided written informed consent. The study is reported as per STROBE guidelines (S1 STROBE Checklist).
We defined MetS using the International Diabetes Federation (IDF) MetS criteria [12], assessed at 15 ± 1 weeks’ gestation. A WC ≥80 cm is a prerequisite risk factor, along with any 2 of the 4 following variables: raised TGs (≥1.70 mmol/l [≥150 mg/dl]), reduced HDL-C (<1.29 mmol/l [<50 mg/dl]), raised BP (i.e., systolic BP ≥130 mm Hg or diastolic BP ≥85 mm Hg), or raised fasting plasma glucose (≥5.6 mmol/l). As this was a pregnancy cohort, a random plasma glucose was measured. The utility of random plasma glucose has recently been demonstrated in a sample of 25,543 women in the UK, showing that in 69% of these women, random plasma glucose was able to predict GDM (area under the curve [AUC] 0.8) better than maternal BMI (AUC 0.65) and maternal age (AUC 0.60) [34].
PE was defined as systolic BP ≥140 mm Hg or diastolic BP ≥90 mm Hg, or both, on at least 2 occasions at least 4 hours apart after 20 weeks’ gestation but before the onset of labour, or postpartum, with either proteinuria (24-hour urinary protein ≥300 mg or spot urine protein:creatinine ratio ≥30 mg/mmol creatinine or urine dipstick protein ≥ ++) or any multisystem complication of PE [33,35]. GDM was defined using the new World Health Organization classification (fasting glucose of ≥5.1 mmol/l or, following an oral glucose tolerance test, a 2-hour level of ≥8.5 mmol/l) [36]. Universal screening was not employed for GDM in Ireland and the UK; women were only screened if they were identified as being at risk based on factors such as family history and BMI. Therefore, GDM analysis was confined only to women from Australia and New Zealand (n = 3,126). sPTB was defined as spontaneous delivery before 37 + 0 weeks’ gestation. Infant measurements (head circumference, length, and birth weight) were recorded by research midwives within 72 hours of birth. Customised birth weight centiles were calculated correcting for gestational age at birth, maternal ethnicity, maternal weight and height in early pregnancy, parity, and infant sex [37]. SGA and LGA were defined as a birth weight <10th or >90th customised centile, respectively.
A total of 5,530 women were included in the current analysis. S1 Fig shows the participant flow. Overall, 12.3% (n = 684) had MetS. Table 1 describes maternal and infant characteristics according to whether women did or did not have MetS. Women with MetS had an approximately 5-kg/m2 higher average BMI, had a lower average SEI, and were more likely to smoke, and the mean gestational age at birth was slightly lower.
Box plots of measured values for each metabolic component in women with and without MetS are shown in S2 Fig. S2 Table shows the number and percentage of women with each metabolic abnormality, in women with and without MetS. In women with MetS, 88.3% of women had TGs above the IDF MetS cut-point, and 83.9% had raised glucose. All women, by definition, had a high WC. In comparison, in women who did not have MetS, 55.2% had high WC, 28.4% had raised glucose, and 20.8% had raised TGs.
S3 Fig shows the UpSet plot with the frequency for each combination of MetS components, and S4 Fig and S5 Fig show the UpSet plots for women with BMI <30 kg/m2 and ≥30 kg/m2, respectively. WC ≥80 cm was the most common single criterion, followed by the combination of WC and glucose. S3 Table shows the number and percentage of women with each pregnancy complication according to MetS.
For each pregnancy complication, the relative risk, the number of women with MetS or its individual components (N), and the number of women with the outcome (n) are shown in Fig 1. Individual MetS components that increased risk for PE were raised BP (RR 1.73, 95% CI 1.11 to 2.68), raised TGs (RR 1.73, 95% CI 1.37 to 2.18), and high WC (1.43, 95% CI 1.04 to 1.96), after adjusting for maternal BMI, age, study centre, ethnicity, SEI, physical activity, smoking status, depression status, and fetal sex. Of the total cohort, the number of women with raised BP at 15 weeks’ gestation was small (n = 158); 20 (12.7%) of these women developed PE. In all, 3,346 women had a high WC, and 213 (6.4%) developed PE. Raised glucose and reduced HDL-C were not significantly associated with risk for PE. Women with MetS were at an increased risk for PE, with a relative risk of 1.63 (95% CI 1.23 to 2.15), and 66 (9.7%) of these women developed PE. In absolute terms, women with MetS had an adjusted excess risk of 2.52% (95% CI 1.51% to 4.11%), with a number needed to harm of 40.
Individual MetS components that increased risk for GDM were raised TGs (RR 3.38, 95% CI 2.24 to 5.10), raised BP (RR 2.69, 95% CI 1.43 to 5.05), and raised glucose (RR 2.46, 95% CI 1.63 to 3.72) (Fig 1). Ninety-eight women had raised BP, of whom 12 (12.2%) developed GDM; 796 women had raised TGs, of whom 65 (8.2%) developed GDM; and 1,091 women had raised glucose, of whom 67 (6.1%) developed GDM. High WC and reduced HDL-C were not associated with risk for GDM. Women with MetS were at the greatest increased risk for GDM, with a relative risk of 3.71 (95% CI 2.42 to 5.67). In absolute terms, women with MetS had an adjusted excess risk of 8.66% (95% CI 5.38% to 13.94%), with a number needed to harm of 12.
Risk for SGA was increased with raised BP (RR 1.89, 95% CI 1.33 to 2.68), but not with any of the other metabolic components, nor with MetS. For sPTB, risk was increased with raised TGs (RR 1.35, 95% CI 1.02 to 1.77), but not with any of the other metabolic components, nor with MetS. Raised glucose modestly increased risk for LGA (RR 1.23, 95% CI 1.04 to 1.45), but none of the other individual metabolic components, nor MetS, increased risk for LGA. MetS was associated with reduced probability of having an uncomplicated pregnancy (RR 0.82, 95% CI 0.73 to 0.93). In absolute terms, for having any pregnancy complication, women with MetS had an adjusted excess risk of 9.35% (95% CI 10.30% to 7.98%), with a number needed to harm of 11. Obesity (BMI ≥30 kg/m2) increased risk for GDM (RR 3.28, 95% CI 2.22 to 4.84), PE (RR 1.91, 95% CI 1.47 to 2.47), and SGA (RR 1.27, 95% CI 1.04 to 1.56).
To determine the adjusted probability of pregnancy complications with increasing maternal BMI (non-linear) according to MetS, we used penalised spline (cubic spline with 4 knots) analysis in GAMs (Fig 2). As maternal BMI increases, the presence of MetS increases the estimated probability for GDM and SGA and also decreases the probability of an uncomplicated pregnancy.
To our knowledge, this is the first large, population-based, multi-centre prospective cohort study in low-risk nulliparous pregnant women to assess the association between MetS—rather than just its individual components—at 15 weeks’ gestation and pregnancy outcomes. We report that 12.3% (n = 684) of women had MetS, and these women were at an increased risk for PE and GDM, after adjustment for a range of demographic and lifestyle variables. Increasing BMI in combination with MetS also increased the estimated probability for GDM, and decreased the probability of an uncomplicated pregnancy. The findings build on the literature assessing the effects of individual metabolic health components or obesity on pregnancy complications, by using established clustering of abnormalities that are evident in non-obese pregnant women.
There is an abundance of literature demonstrating that maternal obesity increases risk for pregnancy complications including GDM and PE [40–42] and also for babies born too small or large [43]. Some studies reporting on individual metabolic components, such as raised TGs or LDL-C or reduced HDL-C, and pregnancy complications show that these are associated with increased risk for GDM, PE, LGA, and sPTB [28]. In addition, raised maternal glucose, insulin resistance, and raised BP are associated with GDM [44,45], and pre-pregnancy WC is associated with GDM and LGA [46]. Examination of a composite of components (MetS) and its association with pregnancy complications extends these earlier studies assessing individual components.
Currently, all obese pregnant women are considered to be at the same increased risk for GDM. However, only 15%–30% of these women will develop the disorder [47]. Recently, in a sample of 1,303 obese pregnant women, a model based on clinical and anthropometric variables (e.g., age and systolic BP) could modestly predict GDM (AUC 0.71), with the AUC increased to 0.77 with the addition of candidate biomarkers such as random glucose and TGs [48]. However, while obese women are at higher risk for GDM than their lean or overweight counterparts, 7%–15% of lean women may still develop GDM [49,50]. Our results suggest that assessment of MetS in all pregnant women may provide information on risk for GDM that extends beyond BMI. Although MetS conferred a slightly greater risk for GDM than glucose, TGs, and WC assessed individually, the impact of MetS as a function of a combination of components or exclusion of certain components requires investigation in other pregnancy cohorts. Additionally, whether MetS improves prediction of risk of GDM above that of raised glucose alone needs to be examined, and should be assessed in women who are obese or not. This will help determine any applicability of diagnosing MetS in the antenatal setting, in addition to standard BMI and glucose measurements.
Recent figures estimate that PE affects 2%–7% of pregnancies and occurs more frequently in nulliparous women [51]. Women who have experienced PE are at a more than 2-fold increased risk for future CVDs such as hypertension and ischemic heart disease [52]. Early identification of women at risk for PE remains one of the major focuses of antenatal care. We found that MetS, a risk factor for future CVD, increased the risk for PE, but in combination with increasing BMI did not demonstrate greater risk. We found that a smaller number of women had raised BP (n = 198) than had MetS (n = 681), but a similar proportion of each went on to develop PE (~10%). Future work comparing whether MetS provides risk information beyond that of raised BP would be useful.
Traditionally, antenatal care pathways are based on individual needs, history, and specific risk factors such as obesity. Despite the best care, a number of women still go on to develop pregnancy complications. Not all available screening tests are offered to all women, and some are only offered to those who are deemed at risk. Testing for impaired glucose tolerance in early pregnancy is rarely provided but may be undertaken before 24 weeks’ gestation if the woman has risk factors. Some guidelines recommend that women have their risk for GDM evaluated by assessment of maternal risk factors followed by an oral glucose tolerance test when required. Furthermore, in nulliparous women, risk assessment for other pregnancy complications such as PE is difficult as there is no maternal history and no widely adopted screening test.
Our findings suggest that MetS diagnosed in early pregnancy may be used to broadly identify women at increased risk for pregnancy complications. The utility of MetS in predicting pregnancy complications, as a complementary component to standard routine antenatal care and compared to standard risk factors, requires exploration. Future studies are also encouraged to define clear metabolic health phenotypes that will produce the optimal probability threshold for each outcome.
The major strength of this study is that it used a population-based prospective cohort design that included a large number of nulliparous women across 6 centres in 4 countries. Participants comprised a clearly defined population of nulliparous low-risk women with no pre-existing disease, which should be considered in study design for other populations. Risk estimates may potentially be underestimated for the general population as the participants in this study are at low risk for pregnancy complications compared to the general population of pregnant women. The study captured 5,530 pregnant women for whom complete information on plasma and other maternal metabolic variables and pregnancy and neonatal outcomes was available. These were ascertained using rigorous assessment methods standardised across recruiting centres, resulting in negligible information bias. A key strength is the assessment of individual and composite MetS compoments and pregnancy complications, which have not been thoroughly investigated to date.
There are limitations in this study. Currently, there are several different definitions for MetS in the adult population as different numbers and types of metabolic variables are used, as well as different cut-off values, between studies [10,11]. However, there are no definitions at all specifically for pregnancy, and, despite pregnancy being a hyperlipidemic state, there is no agreed threshold for pathological hyperlipidemia in pregnancy. We used the IDF definition for MetS for adults, as previous studies have demonstrated no remarkable increase in lipids in first trimester of pregnancy [53], and it is unlikely that WC would have significantly increased at the time of recruitment at 15 weeks’ gestation. There may be inherent limitations with using non-fasting compared to fasting metabolic components, particularly for TGs and glucose; however, an increased risk for PE and GDM has been demonstrated using fasting [25,54,55] and non-fasting [56–61] lipids. Universal screening for GDM in the UK and Ireland was not available at the time of recruitment, and screening was only undertaken for women who were deemed at risk; thus, we only included data for the centres with universal screening (n = 3126) for this outcome. A recent study has highlighted the importance of early screening for GDM in obese women, demonstrating differences in metabolic profile between obese women who do and do not develop GDM [45]. Finally, assessment of MetS was made in the first trimester, and it is unclear whether metabolic components may have altered since conception.
Around 90% of the study participants were white; thus our results are likely generalisable to populations of pregnant women who are white and considered to be at low risk for disease. The results may have limited applicability to other ethnic groups with potentially different patterns of metabolic health and known differences in risk for pregnancy complications. It is also unclear whether these results are generalisable to women of all parities. However, women who have had a complication in a previous pregnancy may modify their behaviours, and clinicians would monitor them more closely in subsequent pregnancies.
We did not compare the impact of individual metabolic components with that of MetS as a composite, and therefore cannot conclude that MetS is better at identifying women at risk. However, more than half of the women who had MetS in early pregnancy developed a pregnancy complication compared with just over a third of women who did not have MetS. Furthermore, while increasing BMI increases the probability of GDM, the addition of MetS exacerbates this probability. Further studies are required to determine if individual MetS components act synergistically or independently. It is currently unclear whether women in this study had an adverse metabolic profile prior to pregnancy, and, therefore, studies to assess and monitor changes in metabolic profile pre-conception and throughout pregnancy may be helpful. Young women identified as having poor metabolic health in pregnancy are at increased risk for pregnancy complications, and these women will likely need metabolic follow-up in the years after their pregnancy.
|
10.1371/journal.pcbi.1004004 | Vesicular Stomatitis Virus Polymerase's Strong Affinity to Its Template Suggests Exotic Transcription Models | Vesicular stomatitis virus (VSV) is the prototype for negative sense non segmented (NNS) RNA viruses which include potent human and animal pathogens such as Rabies, Ebola and measles. The polymerases of NNS RNA viruses only initiate transcription at or near the 3′ end of their genome template. We measured the dissociation constant of VSV polymerases from their whole genome template to be 20 pM. Given this low dissociation constant, initiation and sustainability of transcription becomes nontrivial. To explore possible mechanisms, we simulated the first hour of transcription using Monte Carlo methods and show that a one-time initial dissociation of all polymerases during entry is not sufficient to sustain transcription. We further show that efficient transcription requires a sliding mechanism for non-transcribing polymerases and can be realized with different polymerase-polymerase interactions and distinct template topologies. In conclusion, we highlight a model in which collisions between transcribing and sliding non-transcribing polymerases result in release of the non-transcribing polymerases allowing for redistribution of polymerases between separate templates during transcription and suggest specific experiments to further test these mechanisms.
| RNA dependent RNA Polymerases tight association with their template creates an almost infinite dilution transcription machinery. Polymerases are delivered to the host cytoplasm associated with the genome template, however, they initiate transcription only at or near the 3′ end of the genome template. How these polymerases initiate and sustain transcription is completely unknown. Given the efficiency of these polymerases and their nontrivial template interactions, understanding their mechanism has both medical and nano-technological applications. Here we show that efficient transcription requires a sliding mechanism for non-transcribing polymerases and can be realized with different polymerase-polymerase interactions.
| Transcription is the process of polymerase driven synthesis of mRNA from the genome template. In eukaryotic cells, polymerases engage their promoters through 3D diffusion [1], [2] and have a dissociation constant from their promoters in the range of 40–60 nM [3], [4], [5]. In many viral infections, transcription is the first step for efficient replication. Many viruses, however, do not rely on cellular polymerases for transcription. Specifically non segmented negative strand (NNS) RNA viruses which include potent human pathogens e.g. Rabies, Ebola and measles, deliver special RNA dependent RNA polymerases to transcribe and replicate their genome template [6], [7], [8]. Transcription initiates only at or near the 3′ end of the genome template [9], [10], [11] which immediately poses the question of initiation and sustainability of transcription during early stages of infection.
Here we focus on vesicular stomatitis virus (VSV) which is a prototype of NNS RNA viruses. VSV genomic RNA encodes the 5 viral proteins in the order 3′-N-P-M-G-L-5′ and is fully encapsidated by 1,250 copies of nucleoproteins (N) to form the genome template (N-RNA) [12], [13]. Transcription and following genome replication of VSV is very efficient. Within the first hour of VSV infection, a single template is estimated to yield at least one round of transcription all the way through the L gene with substantially higher transcripts released for the upstream genes [14], [15]. Within 10 hours post infection, a single cell is estimated to produce ∼10,000 progeny VSV virions [16].
The catalytic unit of the VSV polymerase is the L protein, 50 copies of which are packaged within the virion along a single genome template [17]. While the L protein can initiate transcription from its naked genomic RNA [18], it is unable to bind and process the genome template. To transcribe the genome template, L requires association with phosphoprotein P to form a functioning polymerase unit [6], [19] reviewed in [8]. The exact copies of P associated with L during each stage of transcription and replication remains unclear, however in total 400 copies of P are incorporated within each virus [17].
Understanding the position of polymerases along the genome template may provide important clues as to the mechanism of transcription initiation. Polymerases remain associated with the N-RNA templates released from detergent-disrupted virions. When these purified templates are examined with immunogold electron microscopy, the polymerases are found to be almost randomly distributed along the genome template [20]. Within intact virions, Cryo-EM measurements have shown that the 5′ end of the genome template resides at the blunt end of the bullet shaped VSV virion [13] and more recently high resolution fluorescence microscopy also localized polymerases at the blunt end of the VSV virions [21]. Therefore, in combination these measurements localize the polymerases to the 5′ end of the genome template. The mechanism by which polymerases can redistribute along the genome template when extracted from detergent-disrupted virions in the absence of transcription is not well understood.
The polymerases initiate transcription either at the 3′ end of the genome template [10] which results in synthesizing a ∼50 nt leader RNA before synthesizing the N gene or start the synthesis directly at the start of the N gene [9], [11]. Both of these transcription initiation sites (TIS) are located at or near the 3′ end of the genome template and in our manuscript we will not distinguish between initiation at either of these two positions and therefore refer to them collectively as TIS. After initiation, polymerases transcribe the genes coded on the RNA genome starting from the 3′ end in a sequential form [22]. There is a drop of ∼30% in transcription at each gene junction which yields the highest transcription for the N gene and lowest transcription for L [15] (reviewed in [23]). Given that within the first hour of transcription on average at least one L mRNA is synthesized and there is a 30% reduction of transcription at each gene junction [15], a lower bound estimate of the number of N mRNA can be calculated as where n is the number of genes on the genome. In case of VSV, n = 5 and therefore the number of N mRNA synthesized at the first hour would be ∼120. Although the attenuation rates are known, the mechanistic details of how polymerases reach the 3′ end and initiate transcription are not clear.
One plausible mechanism to initiate transcription is for a few polymerases to dissociate from the genome template during its delivery to the cytoplasm. The genome template is delivered to the cytoplasm after G protein facilitated fusion of viral envelope with endocytic membranes [24], [25]. During entry, virion interior acidifies, facilitating release of the matrix protein M from the genome template [26]. If some polymerases dissociate during the entry process, they would be able to bind at or near the 3′ end followed by transcription of the genome template [8]. Sustained transcription is dependent on supplying polymerases to the TIS at 3′ end of the genome template, which in principle can happen through dissociation of a fraction of transcribing polymerases at the gene junctions and their subsequent binding to the TIS though 3D diffusion. Although this simple mechanism is attractive, it is difficult to know a priori if such mechanism would be capable to initiate and sustain transcription. The exact topology of the genome is also not clear, genome templates of VSV are assumed to be linear although some circular topologies have been observed through electron microscopy [27]. Since there are only 50 polymerases associated with a single template and rates of transcription of the polymerases are known [15], plausibility of early transcription models with different mechanisms and genome topologies can be tested based on predicted N mRNA levels at the end of one hour of transcription in Monte-Carlo simulations.
Here we constructed Monte Carlo simulations to gauge the fitness of various early transcription models based on their N-mRNA production levels after one hour of transcription. In total 20,000 simulations of early transcription were performed testing a range of polymerase dissociation rates, sliding rates, TIS binding strengths, linear and circular topologies of the genome template and two different polymerase–polymerase collision conditions. Our simulations show that the initial distribution of polymerases is not critical since the transcription machinery rapidly reaches a steady state. We identify the following mechanisms that can sustain transcription and are consistent with our measurements of the dissociation constant of polymerases from their genome template: i) Transcription through sliding facilitated polymerase release, in which polymerases are released from the template through their collision with a transcribing polymerase and engage the TIS through 3D diffusion; ii) Transcription through sliding facilitated initiation on circular templates, in which polymerases find the TIS through 1D sliding on a circular genome template; iii) Transcription through sliding and non-colliding polymerases (in which transcribing and non-transcribing polymerases can pass each other). At the end we propose that sliding of non-transcribing polymerases plays a critical role in VSV transcription. Specifically we highlight the sliding facilitated polymerase release model and discuss how the predictions of this model are consistent with low dissociation constant of polymerases under no transcription and the previously observed redistribution of polymerases between templates during transcription [28]. We conclude that the transcription machinery of NNS RNA viruses is capable of functioning at almost infinite dilutions. We further suggest specific experiments to narrow possible mechanisms.
We measured the dissociation constant of L from the whole genome template by detergent disrupting the VSV virions and separating the N-RNA template bound with P and L proteins from the rest of the viral proteins using a glycerol cushion spin as detailed in the (S1 Text). The L protein hosts the catalytic site which performs the RNA dependent RNA polymerization and binds the genome template through the phosphoprotein P [8]. Western blot and protein gel analysis was used to detect a small fraction of free L, from which we calculated the dissociation constant to be 20 pM (S1 Text, S1 Figure and S2 Figure). It is important to note that this rate is calculated between the L and the full genome template bound with P and L extracted from detergent-disrupted virions. Since L binds the genome template through its interactions with P, the presence of excess P bound to the N-RNA template produces many L binding sites along the N-RNA templates. We will present an estimate of the dissociation from a single binding site during the discussion. Regardless, the low dissociation constant of the L from the full template guarantees that active polymerases remain bound to the template even when a single template is delivered to the host cytoplasm.
We simulated transcription from a set of 50 polymerases on the linear genome template using the Monte Carlo rules explained in the methods. In brief, 50 polymerases were followed on the template. Polymerases would only initiate transcription from the TIS at the 3′ end. During transcription, polymerases had a 30% probability of falling off the template at the end of each gene. Non-transcribing polymerases would slide along the template with the 1D diffusion rate Dsl and would fall off spontaneously from the genome template with the rate Koff. In the event of a collision between a transcribing and a non-transcribing polymerase, the non-transcribing polymerase would be forced off the template. Free polymerases in solution would find the template through a diffusion limited reaction and would have a binding affinity R to the TIS relative to a random binding site on the template. Neither Koff, Dsl nor R values are known experimentally, therefore we performed 4,000 simulations spanning a large range of these parameters (Dsl: 10 nm2/s to 105 nm2/s; Koff: 10−5/s to 10−1/s; R: 1 to 500). We report the measured N mRNA production for each condition as shown in Fig. 1. Initially, polymerases were positioned at the templates 5′ end. The asymmetric initial distribution was compared with random initial distribution which resulted in similar N mRNA productions as shown in S3 Figure. The most critical parameter in this model is the TIS binding strength R; we could not achieve sustained transcription with R = 1 as shown in Fig. 1A. When R was increased to 500, there were two distinct conditions that allowed sustained transcription.
The one possible caveat for the single track linear genome model presented in Fig. 1 is the relatively high TIS binding strength (R = 500) required for sustaining transcription. The TIS binding strength in the NNS RNA viruses has not been measured independently. Given the unique structure of the 3′ end of the genome template, it is possible that a high TIS binding strength can exist, but it can also be argued that the direct TIS binding strength may not be so high since the promoter is buried under a layer of N proteins. To create a model that does not require high TIS strength, we made an assumption that the N-RNA template can be circularized. Such circular genome templates have been isolated from VSV infected Hela cells [27], however the biological implications and the mechanism of circularization have remained unclear. By assuming that polymerases which reached the 5′ end could move forward and start at the 3′ end of the template we practically inserted the circular genome template into the MonteCarlo simulations as described in previous sections and methods. Similar to the model presented for the linear genome, any polymerase reaching the TIS could initiate transcription.
The collision between transcribing and non-transcribing polymerases is not trivial. It is possible that the transcribing polymerases may be capable of bypassing a non-transcribing polymerase without dissociating the non-transcribing polymerase from the template. Although the RNA is buried under N protein and cannot form secondary structures, it is possible that the N-RNA genome template can form transient kissing loops which would facilitate polymerase transfer from one section of the N-RNA to the next effectively bypassing a transcribing polymerase.
To test these conditions, we changed the collision rules so polymerases can bypass each other and simulated this model by exploring the sliding, dissociation and TIS binding strengths (Dsl: 10 nm2/s to 105 nm2/s; Koff: 10−5/s to 10−1/s; R: 1 to 500) as shown in Fig. 3. Contrary to the single track model proposed in Fig. 1, the double track model was capable of sustaining transcription both under high R = 500 as well as low R = 1 relative TIS binding strengths as shown in Fig. 3.
As expected under high dissociation rates (Koff>10−2/s) we could observe sustained transcription independent of the sliding diffusion coefficient Dsl similar to results shown in Fig. 1 for colliding polymerases.
It can be speculated that transcription can be sustained by a one-time dissociation of polymerases from the template during cellular entry in the absence of sliding. Our simulations however show that such a model is incapable of sustained transcription and polymerases eventually stop transcribing as shown in Fig. 4A. This Fig. also includes the time trajectory of N mRNA production for various other models including transcription through high dissociation rates of polymerases with no sliding as shown in Fig. 4B, Sliding facilitated polymerase release as shown in Fig. 4C and transcription through sliding and non-colliding polymerases as shown in Fig. 4D, all of which except 4A can sustain transcription.
To address the issues of initial conditions further, we have measured the N mRNA production with initial conditions: i) randomly distributed polymerases and ii) asymmetrically distributed polymerases at the 5′ end. These measurements were performed on the same range of Koff (from 10−5/s to 10−1/s) and Dsl (from 10 nm2/s to 105 nm2/s) as shown in S3 Figure produced minimal effects on the overall transcription rates.
The Monte Carlo simulations presented here were used to quantify the transcription efficiency and sustainability of various models. Simple mechanisms like jump starting the transcription with initial dissociation of polymerases during entry did not produce sustainable transcription. The most critical issue for sustaining transcription is to have efficient mechanisms for delivery of polymerases to TIS at the 3′ end. This can be achieved either through 3D diffusion or 1D sliding on the genome template. Below we will discuss specific models presented from our results that can sustain transcription.
The most simple model that can deliver the polymerases to the 3′ initiation site through 3D diffusion is based on dissociation of non-transcribing polymerases from the genome template. In order for this model to sustain transcription, two conditions need to be satisfied. Like any other model that relies on 3D diffusion, this model requires a high relative affinity for the TIS (R>500) and a high dissociation rate of (Koff>10−2/s). The interesting aspect of this model is that it is insensitive to sliding rates, genome topology and/or collision conditions as shown in Fig. 1C, 2C and 3C right panels. While the TIS binding affinity has not been experimentally measured, the more substantial problem with this model is the large fraction of free polymerases predicted in the absence of transcription in all conditions (Fig. 1C, 2C and 3C right panels). This is inconsistent with our measurements of dissociation constant of polymerases from the genome template as presented in S1 Text and also inconsistent with live imaging visualization of the fluorescently modified P proteins during the replication cycle of VSV in Hela cells, in which most of the P is localized with puncta assumed to be genome templates ([29] and our unpublished observations). The existence of a mechanism for high dissociation rates in vivo cannot be completely dismissed. The live imaging data cited above do not show conclusively that the observed P molecules are in association with genome templates and not less substantial aggregates of N proteins. There also exist substantial differences between transcription initiation in vivo and in vitro [9], [11] and therefore there is a finite possibility that the dissociation constant of the polymerases can be sufficiently increased within the cytoplasm to make this model work. However there is no current data indicating such a possibility and further experiments are required to measure the concentration of free polymerases in the cytoplasm. Therefore we will not highlight this model as the most likely model for transcription.
When collision conditions between transcribing polymerases and non-transcribing polymerases on a linear template are such that the non-transcribing polymerases get ejected from the template due to this collision, these ejected polymerases can reach the TIS at the 3′ end of the genome given high relative TIS binding affinity (R = 500). This creates an interesting model in which immediately upon halting transcription, all the polymerases would bind back to the genome template as shown in Fig. 1C left panel. In this model, during transcription 43% of polymerases are free in solution and this percentage drops to <1% immediately upon halting transcription. This model is both consistent with our measurements of polymerase dissociation constants under non-transcribing conditions as well as the observed re-distribution of polymerases from UV irradiated templates [28]. In the experiments with UV irradiated templates, genome templates along with associated polymerases were purified from UV irradiated VSV virions. These damaged templates showed a significant reduction of transcription in vitro. In subsequent experiments, genome templates minus their polymerases purified from WT undamaged VSV virions were added to the UV irradiate genome templates bound with polymerases. This mixing resulted in a boost of total transcription indicative of transfer of polymerases from the damaged template to the undamaged template. This experiment is indeed in agreement with the sliding facilitated polymerase release model since this model predicts a significant pool of polymerases are released from the templates due to collision of transcribing and non-transcribing polymerases. The important point being that although the genome templates are UV irradiated, their polymerases are still capable of initiating transcription. This initiation leads to either premature dissociation or creation of a roadblock that facilitates dissociation of sliding non-transcribing polymerases due to collisions. To demonstrate this effect we simulated the transcription from the UV damaged templates as shown in S1 Text S4 Figure and demonstrate that polymerases would fall off the UV irradiated templates. Based on these experiments we highlight the sliding facilitated transcription model as a most likely model of transcription.
This is the first model we tested in which polymerases would find the initiation site through 1D sliding and are capable of sustaining transcription. The advantage of this model is that it is operational even under low relative binding strength of the TIS (R = 1) as shown in Fig. 2C left panel. Although in this model, similar to the sliding facilitated polymerase release model, >50% of polymerases are free in solution during transcription, careful analysis of the polymerase trajectories in the simulations shows that 99.5% of the polymerases that initiate transcription, find the initiation site through 1D sliding. This model is also consistent with the observed low dissociation rate of the polymerases since in non-transcribing conditions 99% of polymerases remain bound to the templates as shown in Fig. 2C right panel. The major problem with this model is that aside from early electron microscopy data [27], the presence of circular templates during transcription has not been verified experimentally, however this model is consistent with the observed redistribution of polymerases from UV irradiated templates [28] and therefore would be the most likely model of transcription if measurements show a low TIS binding strength.
If we assume that transcribing polymerases can bypass non-transcribing polymerases on the template and that non-transcribing polymerases can slide on the genome template we arrive at a model that does not require circular templates, sustains a high rate of transcription and can operate with almost no free polymerases in solution either during or under non-transcribing conditions as shown in Fig. 3C left panels. This is an attractive mechanism since it will create an infinite dilution, transcription machinery with no free polymerases even under transcribing conditions. However, the mechanics of polymerase collisions are nontrivial and it is not clear how a non-transcribing polymerase can retain its affiliation to the template under heavy reorganization of the template during passage of a transcribing polymerase [8], [19]. More substantially, since this model predicts almost no dissociation of polymerases during transcription, it fails to explain the previously observed redistribution of polymerases between UV irradiated and undamaged genome templates [28].
All the mechanisms represented here and tested in the MonteCarlo simulations are inspired by previously observed mechanisms in DNA enzymes and molecular motors. It is however possible that redistribution of polymerases in NNS RNA viruses has a unique and different mechanism. One such mechanism, can be transferring of the polymerases from one section of the template to the other through collisions between template sections, such a model can be termed effectively the “Spider-Man” model. In such a model, sliding of polymerases would not be essential since polymerases can redistribute along the template through strand transfer collisions. Theoretical prediction of the spider-man model is complicated, since it depends on the three dimensional dynamics of the genome template, which is currently unknown. The spider-man model however is predicted to be heavily dependent on template self collisions and therefore it predicts a severe decrease in polymerase redistributions under stretched template geometries which can be achieved by mechanical manipulations of the template.
It becomes clear when comparing the panels in Figs. 1, 2 and 3C that one of the main indicators of the various models are the amount of free polymerases during active transcription. Therefore, the essential next step should be to directly measure the amount of free polymerases during transcription. Since transcription can be reconstituted from purified VSV genome templates, these experiments are highly feasible. With the advent of new quantitative fluorescent measurements like Fluorescence Correlation Spectroscopy [30], [31], [32], the minimal amount of free polymerases can also be measured during transcription in live cells, given the proper tagging of fluorescent polymerase components.
Failure to detect free polymerases during transcription would make the sliding non-colliding polymerase mechanism proposed in Fig. 3 likely. Our simulations indicate that sliding of non-transcribing polymerases plays a critical role in all proposed mechanisms. Measurements of this sliding rate in vitro using available single molecule techniques, although possible, is very challenging due to lack of tools for immobilizing the genome template. These measurements however, would be crucial to validate the proposed sliding mechanism in NNS RNA virus transcriptions.
Given the 20 pM dissociation constant measured for L with its full genome template, how can a sliding model be justified? Indeed, the dissociation constant of molecular motors from their tracks is in the low nanomolar range and the dissociation constant of the T7 RNA polymerase to its class III TIS is in the range of 40–60 nM [3], [4], [5]. Based on our approximate division of the template to 10 nm regions (the size of the polymerase), we conclude that around 300 polymerases could fit onto a single N RNA template. If we assume that all these interactions are linear, we can estimate that a single binding site would have a dissociation constant of 6 nM, which would be much closer to the observed dissociation constants from other procesive motors. The sliding movement of the polymerase is likely realized through transient binding and dissociation of the P protein to the N-RNA, structural evidence of such transient interactions have been previously demonstrated in paramyxoviruses [33].
Cellular DNA based RNA polymerases have been shown to bind directly to their promoters to initiate transcription [1], [2]. Although we incorporate sliding of polymerases along the N-RNA template, our model of transcription in the linear genome also relies on binding of the polymerases through 3D diffusion to a TIS ((R>500) Fig. 1). It is only under the low TIS strength and circular templates that our model predicts polymerases will find the TIS through sliding (Fig. 2). Combinations of sliding and 3D diffusion has been proposed previously for lac repressor facilitated diffusion to reach its target sequences up to 100 fold faster than the diffusion limit [34], [35]. DNA repair enzymes were also shown to utilize long range 1D sliding on template [36], [37]. Sliding along the genome template therefore can be a plausible mechanism.
In NNS RNA viruses, aside from initiation at the 3′ end [10], transcription can be efficiently initiated at the start of the N gene as shown in vivo [9] and in vitro [11]. These observations can be reconciled if one assumes that polymerases can scan the template, therefore always entering at the 3′ end but sometimes scanning their way to the start of the N gene without synthesizing the leader RNA [38], [39]. Scanning is not the only proposed mechanism, a direct binding model based on structural rearrangement of the N proteins around the TIS can also explain the start of initiation directly at the N gene junction [40]. The proposed scanning as discussed above is likely different from the sliding mechanism we propose here since the scanning takes place with a polymerase already engaged with the RNA, while our proposed sliding can happen on the surface of the N-RNA without the need to rearrange the template. The asymmetric localization of polymerases at the 5′ end of the genome within intact virions [41] in conjunction with their random distribution along the genome template when extracted from detergent solubilized virions [20] supports a sliding mechanism by which polymerases can redistribute on the genome template after release from the virion.
In our model we investigated only the first one hour of transcription. The kinetics of transcription during the replication cycle beyond the first hour have been previously modeled for viral infection cycles including: kinetic modeling of viral growth in VSV [42], attenuated VSV [43], Influenza transcription replication kinetics [44] and HCV replication kinetics [45]. Our goal was to find a set of mechanisms that would allow transcription while maintaining the polymerase association with the template as shown experimentally. We conclude that sliding on the N-RNA template would be essential for efficient transcription of NNS RNA viruses.
Understanding how polymerases initiate and sustain transcription is attractive both as a potential antiviral target and also because these polymerases represent an almost infinite dilution transcription machinery. Understanding the mechanics of this machine therefore may open doors to nano-technological applications.
Negative stain electron microscopy of single VSV polymerases shows them as 10 nm ring-like structures which can adopt different conformations with appendages [7], [19]. At our low-resolution approximation, we divided the genome template into 10 nm segments and at any given time, polymerases could occupy any of these segments, this approximation resulted in nTotal = 300 segments along the template (n = 1 to 300). There are 50 polymerases that occupy the genome template Li (i = 1 to 50). No two polymerases can occupy the same segment simultaneously. To create a Monte Carlo simulation, we also digitized the time domain. The essential time scale for the simulation τ is the time required for transcribing one segment of the N-RNA by a polymerase. Given an average rate of transcription of 3.7 nt/s [15], we calculate τ = 7.8 Sec. Based on this calculation, we set up our Monte Carlo simulation in which 50 polymerases (Li) were tracked along the digitized genome template with incremental time τ. In each time step all the polymerases are forced to choose a reaction with the following rules:
1) Transcription is initiated only if the polymerase is on the TIS; 2) Transcription follows a linear trajectory during each time step τ, with 30% fall off rate at the end of each gene junction; 3) Non-transcribing polymerases move along the genome randomly, with a step size calculated based on the mean square displacement associated with a 1D random walker with a diffusion coefficient Dsl; 4) Collision of non-transcribing polymerases results in reflection of trajectories without dissociation from the template; 5) Reflective boundary conditions are set up at the ends of the template for non-transcribing polymerases; 6) Non-transcribing polymerases fall off the template with a rate Koff and can immediately rebind back with a rate Khopp (S1 Text) or can stay in solution till next time step; 7) Free polymerase in solution would bind back to the template with a rate Kon, once binding back, they have a higher probability of binding to the TIS compared to a regular template section with a factor R.
Kon is not determined experimentally, however an upper limit is calculated based on the diffusion-limited reaction. The upper limit of a diffusion limited reaction is . For a single polymerase and a single template in the volume of cytosol, the maximum reaction rate would be:
To incorporate the stochastic binding within the Monte Carlo simulations, we used a Gillespie Algorithm (SSA) [46] to determine a stochastic binding time tbinding based on the Kon in the simulation. r is a pseudorandom number drawn from the uniform distribution on the open interval (0,1).
Each polymerase that falls off the template will be assigned its own tbinding time, and will be kept in solution until the simulation time exceeds tbinding time. The polymerase then gets reassigned to an available site on the template. The parameter R, which is the ratio of binding strength between TIS and unspecific sites, determines the probability of polymerase binding to the TIS PTIS = (R/(nTotal+R)) or the probability of randomly binding on the template Pnonspecific = (nTotal/(nTotal +R)) in which nTotal is the total number of available binding sites.
In the described simulations with no sliding, we could not reconcile a healthy transcription rate, which produces ∼100 N mRNA in 1 hr [15] with the 20 pM dissociation constant of the polymerases from the template (S1 Text).
|
10.1371/journal.pcbi.1003115 | Understanding the Molecular Determinants Driving the Immunological Specificity of the Protective Pilus 2a Backbone Protein of Group B Streptococcus | The pilus 2a backbone protein (BP-2a) is one of the most structurally and functionally characterized components of a potential vaccine formulation against Group B Streptococcus. It is characterized by six main immunologically distinct allelic variants, each inducing variant-specific protection. To investigate the molecular determinants driving the variant immunogenic specificity of BP-2a, in terms of single residue contributions, we generated six monoclonal antibodies against a specific protein variant based on their capability to recognize the polymerized pili structure on the bacterial surface. Three mAbs were also able to induce complement-dependent opsonophagocytosis killing of live GBS and target the same linear epitope present in the structurally defined and immunodominant domain D3 of the protein. Molecular docking between the modelled scFv antibody sequences and the BP-2a crystal structure revealed the potential role at the binding interface of some non-conserved antigen residues. Mutagenesis analysis confirmed the necessity of a perfect balance between charges, size and polarity at the binding interface to obtain specific binding of mAbs to the protein antigen for a neutralizing response.
| Group B Streptococcus (GBS) is the leading cause of neonatal invasive diseases and pili, as long filamentous fibers protruding from the bacterial surface, have been discovered as important virulence factors and potential vaccine candidates. The bacterial surface is the main interface between host and pathogen, and the ability of the host to identify molecular determinants that are unique to pathogens has a crucial role for microbial clearance. Here, we describe a strategy to investigate the immunological and structural proprieties of a protective pilus protein, by elucidating the molecular mechanisms, in terms of single residue contributions, by which functional epitopes guide bacterial clearance. We generated neutralizing monoclonal antibodies raised against the protein and identified the epitope region in the antigen. Then, we performed computational docking analysis of the antibodies in complex with the target antigen and identified specific residues on the target protein that mediate hydrophobic interactions at the binding interface. Our results suggest that a perfect balance of shape and charges at the binding interface in antibody/antigen interactions is crucial for the antibody/antigen complex in driving a successful neutralizing response. Knowing the native molecular architecture of protective determinants might be useful to selectively engineer the antigens for effective vaccine formulations.
| The bacterial surface is the foremost interface between host and pathogen, and recognition of the specific epitopes by the immune system provides the host a key signature to initiate microbial clearance. Identification and characterization of antigenic epitopes is a rapidly expanding field of research with potential contributions to the tailored design of improved, safe and effective vaccines [1], [2]. A number of approaches are currently being used that require atomic-level information in understanding the rules governing antibody/antigen interaction, in particular it is the degree of complementarity between surfaces on epitope and paratope that determines the affinity and specificity of this interaction [3]–[5]. To date, the concept of complementariness is directly related to the conservation of the amino acid sequence on a specific neutralizing epitope. A single amino acid change resulted crucial to alter the surface antigenic properties of a specific epitope of Neuraminidase (NA) in Influenza virus [6].
Streptococcus agalactiae (also known as Group B Streptococcus or GBS) is a Gram-positive pathogen causing severe diseases in newborn and young infants worldwide [7]. Pilin proteins, structural components of cell surface-exposed appendages, have been discovered in GBS as important virulence factors as well as promising vaccine candidates [8]. These high molecular weight structures are made by a major shaft subunit (named backbone protein, BP), a major ancillary protein (named AP1), and a minor ancillary protein (named AP2). BP is distributed regularly along the pilus structure and is fundamental for pilus assembly whereas the two ancillary proteins are dispensable [9]. AP1 is thought to be located at the tip of the assembled pilus structure, while AP2 is involved in pilus attachment to the cell wall [10]–[12]. In GBS three pathogenicity islands, named Pilus Island-1 (PI-1), Pilus Island-2a (PI-2a) and Pilus Island-2b (PI-2b), each encoding pilin protective subunits, have been identified [9]. Among them, the backbone protein of Pilus Island 2a (BP-2a) is a key component of a promising pilus-based vaccine formulation against Group B Streptococcus infections [13], [14]. However, this protein showed the highest level of gene variability among all pilin antigens, characterized by six non cross-protective allelic variants [13]. Each variant identified was able to induce protective immunity in mouse models and opsonophagocytosis killing of live bacteria, but only against GBS strains expressing the homologous variant [13], [14]. A recent Structural Vaccinology approach applied to the BP-2a protein led to the identification of the minimal protein domain carrying the protective epitopes [14]. By using in vitro opsonophagocytosis assays and in vivo animal infectious models, we demonstrated that, within each variant, the domain D3 is responsible for eliciting neutralizing antibodies against pathogen homologous infections [14]. This structure-based approach combined with immunological assays succeeded in the generation of an easy-to-produce chimeric antigen capable to elicit protection against the majority of circulating GBS serotypes [14]. However, the specific mechanisms by which antibodies raised against each variant can mediate a neutralizing response only against GBS strains expressing the homologous variant are not completely understood.
The aim of this work is to contribute a deeper understanding of the molecular basis driving the immunogenic specificity of single BP-2a variants, explaining the mechanism by which amino acid variability on the antigen surface may allow the bacteria to adapt to the host environment and/or escape its immune system. To investigate the variant-specific immunogenicity of BP-2a at the molecular level, we generated neutralizing monoclonal antibodies (mAbs) raised against a specific allelic variant of the protein, the 515 allele [13], [14]. The produced mAbs were functionally screened according to their ability to recognize the polymerized pilus structure on the bacterial surface and to mediate opsonophagocytosis GBS killing in vitro. By Surface Plasmon Resonance (SPR) technology we determined their binding affinity. An approach based on partial digestion, immunocapture and mass spectrometry was used to identify the epitope region in the antigen. Finally, docking and molecular simulation prediction studies were used to elucidate the intrinsic and functional affinity between mAbs and antigen at a residue level.
This work describes a potential strategy to investigate the immunological and structural properties of a surface virulence factor, by elucidating at the molecular level the chemical-physical properties directly related to an effective neutralizing response against pathogen infection.
Recent advances in monoclonal antibodies (mAbs) technology suggested us to use them as tool to investigate the principles governing functional antibody/antigen interactions [6]. So, to understand the immunological differences among the six different variants of the highly immunogenic GBS protein BP-2a through the identification of neutralizing epitope(s), mouse monoclonal antibodies (mAbs) against the 515 allelic variant were generated following standard procedures (see Materials and Methods). Since surface accessibility of bacterial proteins is a fundamental pre-requisite of antibodies for mediating an effective humoral response against bacterial infections, selection criteria for mAb identification were based on variant-specificity and on bacterial surface staining, which was investigated by Flow Cytometry (FACS) analysis. The screening procedure resulted in the identification of six different monoclonal antibodies (named 4H11/B7, 17C4/A3, 27F2H2/H9, 14F6/A1, 25B7/D7, 28E7/E4) able to recognize only the polymeric pilus structure on the bacterial surface of their homologous strain 515 (Figure 1A). In fact, they were not able to stain the surface of GBS strains expressing a different BP-2a variant (Figure 1A).
To assess if the six selected mAbs could also mediate a functional immunogenic response against GBS we performed an in vitro opsonophagocytosis assay, using as effector cells differentiated HL60 cells, as described in the Materials and Methods section, and GBS strain 515. We analyzed each monoclonal antibody at three different dilutions in presence of baby rabbit complement. As shown in Figure 1B, only three out of six mAbs (4H11/B7, 17C4/A3 and 27F2H2/H9) were able to mediate an effective complement-dependent opsonization and killing of GBS bacteria, meaning that these mAbs can recognize and bind neutralizing epitopes exposed in the pilin protein on the bacterial surface.
Classes and subclasses of the six monoclonal antibodies were determined as described in the Materials and Methods section. Clone 4H11/B7 secreted the IgG2b subclass; clone 17C4/A3 and 27F2H2/H9 secreted the IgG2a subclass whereas clones 14F6/A1, 25B7/D7 and 28E7/E4 had the IgG1 isotype.
To evaluate the interaction between BP-2a 515 variant and the selected mAbs, we conducted SPR (Surface Plasmon Resonance) analyses. A convenient strategy to study this interaction is to capture the antibody on a surface containing Fc-receptors in order to place the antibody in a well-defined orientation for binding analysis. Two CM5 biosensors, one coated with Protein A and the second with Protein G, were prepared in order to steadily capture the different isotypes of the mAbs and study their interaction with BP-2a 515 variant in terms of association (ka) and dissociation (kd) rate constants, and binding affinity (KD = kd/ka).
The monoclonal antibodies 17C4/A3 and 27F2/H2/H9 were captured on both Protein A and Protein G biosensors while 4H11/B7 mAb was stably captured only in presence of Protein G. Two out of the three IgG1 mAbs, 25B7/D7 and 28E7/E4, were captured by Protein G, increasing the RU (refractive unit) of capture according to the concentration (5 or 15 nM), while 14F6/A1 was not captured by protein G up to the concentration of 15 nM. After the capture, the mAbs (4H11/B7, 17C4/A3 and 27F2H2/H9) that were the same antibodies that were able to mediate opsonophagocytic killing of GBS cells could bind BP-2a 515 variant, while the two IgG1 mAbs captured by the Protein G biosensor did not bind to the protein in the range of 0.5 to 2.5 µM. For the mAbs which were able to bind the BP-2a 515 variant single cycle kinetics were performed on both biosensors when possible. The average of three independent runs is reported in Table 1. Data showed that the association phase (ka) resulted comparable for all the tested mAbs, within a range of ∼2.5-fold (ka max/ka min). Larger differences were measured in the kinetic of dissociation kinetics (kd), with the most stable binding observed for 4H11/B7, kd approximately 10-fold slower than for 17C4/A3 and 5.5-fold slower than for 27F2/H2/H9 mAb. Nevertheless, the corresponding thermodynamic dissociation constants (KD) did not differ among the three mAbs, which showed to strongly bind the BP-2a 515 variant.
To identify the neutralizing epitope on BP-2a 515 variant, an epitope mapping with the three functionally active monoclonal antibodies (27F2/H2/H9, 17C4/A3 and 4H11/B7) was performed. Two different MS-based approaches were used, one of them allowing the identification of conformational epitopes (see Materials and Methods). With either approach, the experiments were performed six times, using the proteases trypsin, LysC and GluC, and using the mutant form of the entire protein lacking the three isopeptide bonds (BP-2a-515K199A/K355A/K463A) previously generated [14]. It is well-known that the presence of internal isopeptide bonds is important for the stability and resistance to proteolysis of single structurally independent domains in which the protein is organized. The results indicate that the three monoclonal antibodies recognize the same region of BP-2a-515 in domain D3, with sequence 411-TYRVIERVSGYAPEYVSFVNGVVTIK-436. The domain D3 was the same protein portion previously characterized as the domain carrying most of the epitopes inducing protective antibody responses [14]. Figure 2A shows the mass spectrum of the total LysC digestion of BP-2a-515 (upper panel) and that of the peptides immunocaptured with 4H11/B7 (lower panel). The two labeled signals in the lower panel correspond to the fragments of BP-2a-515 sharing the 411–436 sequence (Figure 2B). To confirm the sequence of these peptides, MS/MS spectra were obtained for the peak with a m/z = 2946.530 Da. For the peak with a m/z = 4156.461Da no MS/MS was recorded due to the low intensity of the signal. When using trypsin and GluC no immunocaptured peptide fraction could be detected. Since the peptides retained by the antibodies after LysC digestion contained R and E residues, potential cleavage sites for trypsin and GluC, respectively, these results suggest that the residues R and E and their immediate neighbors may play a role either in the interaction with mAbs or in the structural arrangement of the epitope, since cleavage at either site prevents binding. In the BP-2a-515 structure (PDB code: 2XTL) [14] it can be observed that the residues that are exposed at the surface are comprised between Glu424 and Lys 436 (Figure 2C). The fact that the two approaches tested (see Materials and Methods) lead to the same result indicates that epitope recognition is primarily based on sequence.
To investigate the mode of action of neutralizing mAbs on the antigen BP-2a 515, we performed mAbs-protein docking. Monoclonal antibodies sequences were obtained by isolation of total RNA from each hybridoma cell line and reverse-transcription. Then, using the generated cDNA as template, the heavy (VH) and light (VL) chains were amplified using specific PCR primers. Sequence comparison of the three neutralizing mAbs is showed in Figure 3 and all light chains were of the κ-type.
To elucidate the residue-specific interaction between antigen and antibody at the binding interface, after mAbs sequencing, a structural model of the Fv domain of two of them (17C4/A3 and 4H11/B7) was developed using Modeler 9v8 [15]. Template crystal structures for mAb 17C4/A3 were selected from PDB showing 80% sequence identity for an antibody variable heavy chain (PDB entry 1H3P) and 79% sequence identity for an antibody variable light chain (PDB entry 2ROW). The same procedure led to the selection of two template crystal structures for mAb 4H11/B7 sharing 90% sequence identity for the light chain (PDB entry 1I9J) and 76% sequence identity for the heavy chain (PDB entry 3O6M). In both cases, light and heavy chains were packed together and energy minimized before proceeding with docking studies.
To investigate at the amino-acid level the molecular interactions between neutralizing mAbs and BP-2a antigen, the modeled structures of antibodies were docked against the partial crystal structure of the antigen [14] using ATTRACT [16]. Knowing that mAbs bind to the D3 domain of the antigen in the region T411-K436, we used this information to screen the most accurate complexes within a range of 15000–20000 complexes generated by the docking program.
To validate the stability and reliability of the best selected complexes, we performed explicit solvent molecular dynamics simulation using GROMACS 4.0.5 simulation package [17].
Molecular dynamics results of best docked solutions of mAbs/BP-2a antigen revealed different binding orientations of the neutralizing mAbs against the target protein which showed the importance of specific residues both in the epitope and in the paratope. Molecular simulation results indicated that a shorter portion of the previous identified epitope might be necessary at binding interface: P423-K436 in 17C4/A3-complex (Figure 4A) and V426-K436 in 4H11/B7-complex (Figure 4B). In particular, during the course of the simulation, the distance between those residues and the CDR remained at a contact distance of around 4 Angstrom, indicating their importance for complex interaction and stability. Although two different mAb binding orientations were identified as sterically possible, two amino acid residues located on the target antigen were identified as fundamental at the binding interface in both cases: Val429 and Asn430. Remarkably, residue 429 had been identified as being under selective pressure, which is consistent with a direct role in the interaction with the antibody. Both mAb binding interfaces form a deep cleft filled by a loop region of domain D3. In the docking models residues Val429 and Asn430 fill the deeper cavity of both clefts in a water-free environment (Figure 5). Val429 is involved in hydrophobic interaction with Val206, Arg208 and Ser204 in the 17C4A3-antigen complex (Figure 6A), whereas it interacts with Val207, Leu317, Asn320 and Tyr321 in the 4H11/B7-antigen complex (Figure 6B). Asn430 is predicted to interact with residue Glu269 and Arg208 through H-bonds and makes polar interactions with Ala252, Ser254 and Ile318 in the 17C4A3-complex (Figure 6A), while it establishes through polar interactions with Tyr321 and Tyr277 residue in the 4H11/B7-complex (Figure 6B).
The level of conservation of the epitope was analyzed in a set of 144 BP-2a sequences from different GBS isolates. An alignment of the protein sequences revealed six main variants, as described in previous work [13]. In the full protein alignment, inter-variant variability is high (p distance = 0.31) with relatively few conserved positions (25%), while intra-variant variability is low, suggesting a mosaic structure (see full alignment in Supplementary Material). Comparative sequence analysis of the identified epitope region among the BP-2a protein sequences divided the isolates into the same variants (Figure 7A) with the same variability pattern of the full protein (p distance = 0.37 and 26% conserved sites). The alignment also showed that the epitope region has a more conserved first half, residues 411 to 423 (sequence 515), and a more divergent second half, residues 424 to 436 (Figure 7B). In addition, the unique BP-2a nucleotide sequences were aligned to study the genetic events causing the observed variability. A recombination analysis using GARD [18] identified two statistically significant break points located at codon positions 198 and 636 (sequence 515). To distinguish the effect of mosaicism and point mutations in epitope variability, results from the GARD algorithm were taken into account for the estimation of positive selection. Thus, the REL algorithm implemented in HyPhy [19] identified 16 sites under selective pressure in BP-2a (Supplementary Material), one of them (Val429, sequence 515) located in the epitope region (Figure 7A). These results suggest that both recombination and selection of advantageous mutations have acted to generate the six BP-2a variants for both the full protein and the epitope region. In particular, residue 429 could be changing in response to the pressure of the immune system and be thus a key element for immunological specificity.
To further confirm single amino acid contributions to mAb-epitope binding and elucidate molecular determinants of the immunological specificity of each BP-2a allelic variant, a peptide dot blot immunoassay was performed. Combining epitope mapping, docking and molecular dynamics results with alignment data and detected signatures of positive selection (Figure 7A), four mutated peptides were synthesized (Figure 7B) and used to evaluate the contribution of single point mutations at the mAb-antigen binding interface. In particular, major consideration was reserved on those residues found to reach the deeper cavity of the mAbs cleft: Val429 and Asn430. The designed mutations were aimed at testing the effect of size and charge of the side chains of these two residues on the binding to the antibodies. Mutating both the Asn430 and Val429 to an alanine residue resulted in a conserved capability of binding to the two mAbs, even with increased affinity in the case of Val429 substitution. On the other hand, mutating either residue to lysine, that carries a bulkier and positively charged side chain, substantially reduced or completely abolished the binding of the mAbs to the spotted peptides (Figure 7B). The results from lysine substitution support the presence of residues 429 and 430 at the interaction interface (highly perturbed by a lysine) while those from alanine substitution suggest that these two residues are not responsible for specificity but their nature is probably limited by their fit in the cavity.
The ability of the host to identify microbial molecular determinants that are unique to pathogens has a crucial role in host defense. The recognition by the immune system of the host of surface exposed components, such as proteins and polysaccharides represents the start signal for microbial clearance. Characterization studies of vaccine formulations require a deep knowledge of the interactions between pathogen and host immune system and vaccine components should include the molecular determinants able to stimulate an effective immune response against a specific pathogen. The data reported in this work show a significant correlation between the molecular interactions of monoclonal antibody/target protein with successful neutralizing response against bacterial infection.
The protection against GBS has been associated with the production of high levels of neutralizing antibodies which specifically recognize the antigens exposed on bacterial surface [20], [21]. However, the specific mechanisms, in terms of single molecular determinants, by which antibodies neutralize GBS infections, are not completely understood. Moreover, it is well-known that GBS as well as many other bacteria have evolved a wide range of mechanisms to escape the immune system of their hosts or to adapt to environmental variation, for instance, adopting the strategy of gene variability and/or differential gene expression. These strategies play a crucial role in the capacity of pathogens to trigger disease and also explain why it is so difficult to develop vaccines against these microorganisms. Thus, positive selection and recombination have played an important role in adaptation of the core-genome of different Streptococcus species to different hosts [22]. In this context, the identification of different allelic variants of a key vaccine candidate, such as the pilin protein BP-2a, able to induce variant-specific protection [13] clearly reflects a typical strategy of the bacterium to escape the immune system of the host and, at the same time, represent an additional confirmation of the important role of this protein in GBS virulence. Recent data showed that the majority of protective epitopes of the different BP-2a alleles are located in a single structurally independent domain, called D3 domain [14]. A synthetic chimeric protein constituted by the protective D3 domain of the six BP-2a variants was able to protect mice against the challenge with all of the type 2a pilus-carrying strains [14]. In this work we have investigated the contribution of single amino acid residues within the immunodominant domain D3 of BP-2a, able to drive the neutralizing host humoral response. As a tool to investigate the principles governing functional antibody/antigen interactions at the amino acid level, we successfully selected functionally active monoclonal antibodies targeting the 515 allelic variant of the pilin BP-2a. The selected mAbs were able both to recognize the polymerized pilus structure on bacterial surface and to mediate complement-dependent opsonophagocitic killing of live bacteria.
Epitope mapping analysis of two of the neutralizing mAbs identified showed that the mAbs bind to the same region of BP-2a-515 in domain D3, the same domain previously identified as the immunodominat protein region carrying protective epitopes [14]. Although these results confirmed the importance of D3 domain for immunogenicity and protection capacity of BP-2a, to elucidate the specific affinity of the antibodies versus their protein target a structural analysis of the mAbs alone and in complex with the target antigen has been performed. Molecular docking and MD simulation studies indicated that only two specific residues on the target protein, Val429 and Asn430, were able to reach the deepest cavity formed by the antibody binding interface, mediating specific hydrophobic and polar/H-bond interactions, respectively.
Previous studies support the importance of single amino-acid residues at the binding interface in mAb/antigen interactions, responding to a strict balance of shape and energetics. In the case of Influenza virus (H3N2 vaccine strains 1968–2007), modeling and antigen/antibody docking analyses revealed the molecular basis of the interactions between Hemagglutinin (HA) protein, the primary target of the human immune system, and monoclonal antibodies [6]. Specific mutations both in the neutralizing epitopes and in their vicinity altered the protein surface and the surface electrostatics of the virus, leading to the loss of recognition by the antibody [6]. Though the epitopes responsible for immunity were very similar in successive variants of HA, the simulations could explain the antigenic drift of pathogen surface determinants that has been responsible for the loss of immunity against Influenza infection even in vaccinated population [6]. It has been also shown that single-residue mutants of an antigen may prevent docking by increasing the free energy barrier to conformational rearrangements required for binding to the antibody [23]. In light of this data, the variability of the epitope region identified in BP-2a has been analysed, including the detection of events of recombination and positive selection. Both factors have been found to be significant players in epitope variability. Thus, recombination is likely to be at the basis of the six allelic variants known today and, in addition, residue 429 is predicted to be under positive selection. Interestingly, this residue is located in a part of the epitope region that has low conservation, strengthening the hypothesis of a GBS antigenic drift to escape the immune response and to adapt to the host.
To further characterize the epitope recognized by 4H11/B7 and 17C4/A3 mAbs, a functional dot blot assay using mutated peptides was performed. To identify functional residues within the neutralizing epitope, we mutated those residues predicted to be fundamental at the binding interface (Val429 and Asn430) to lysine and alanine. Mutating Val429 and Asn430 into alanine did not drastically affect mAb surface electrostatics and did not generate steric interferences that could inhibit the binding of antibody. This indicates that a perfect surface antigen-antibody complementarity in this region is not necessary for binding. Conversely, changing the same residues in lysine resulted in a decreased, in the case of Val429, and in a complete abolishment, in the case of Ans430, of mAb binding. The substitution of valine or asparagine by lysine introduces a drastic perturbation of the shape and electrostatics of the antigen surface at this region. The fact that this perturbation inhibits binding supports the presence of these two residues at the interface.
Overall, this study provides new insights into mAbs-BP-2a 515 variant interactions and highlights the molecular correlation between BP-2a variability and its immunological specificity. Moreover, the identification of a neutralizing epitope of a highly immunogenic antigen could be useful for a knowledge-based design of effective vaccines, avoiding the side effects of unfavorable epitope(s) and stringently targeting the immune response only on those one(s), belonging either to the same or to different antigenic protein(s), responsible of pathogenic clearance. Knowing the native molecular architecture of protective determinants could be possible to selectively engineer the antigens for including them in a more effective vaccine formulation.
Animal treatments were performed in compliance with the Italian laws and approved by the institutional review board (Animal Ethical Committee) of Novartis Vaccines and Diagnostics, Siena, Italy.
GBS strains used in this work are 515 (serotype Ia, expressing BP-2a-515 allele); CJB111 (serotype V, expressing BP-2a-CJB111); H36B (serotype II, expressing BP-2a-H36B); 3050 (type II, expressing BP-2a-2603); CDC84 (serotype II, expressing BP-2a-DK21); and strain CDC89 (serotype Ia, expressing BP-2a-CJB110). Bacteria were grown at 37°C in Todd Hewitt Broth (Difco Laboratories) or in trypticase soy agar supplemented with 5% sheep blood.
The full-length BP-2a 515 variant and the mutated form of BP-2a-515 (BP-2a-515K199A/K355A/K463A) were produced as previously reported [14]. Recombinant proteins were expressed in E. coli BL21 (DE3) (Novagen) cells as His-tagged fusion proteins and purified by affinity chromatography and gel filtration.
Mouse monoclonal antibodies (mAbs) were generated by Areta International (Varese, Italy) using standard protocols. Briefly, B-cell hybridoma clones were isolated from spleen cells of immunized CD1 mice with the purified recombinant BP-2a-515 protein. Positive clones were first selected by ELISA and then culture supernatants were screened for binding to the surface of GBS 515 strain by flow cytometry. Positive primary hybridoma clones were subjected to single cell cloning and sub-cloning by limiting dilution. Monoclonality of a clone was accepted only when all the wells of a microtitre plate with growing cells gave positive reaction in indirect ELISA after repeated sub-cloning. The selected mAbs were finally purified by protein G affinity chromatography. Classes and subclasses of the monoclonal antibodies were determined by IsoQuick Mouse Monoclonal Isotyping Kit (Sigma).
Flow Cytometry Analysis (FACS) analysis was performed as described elsewhere [9]. Briefly, mid-exponential phase bacterial cells were fixed in 0.08% (wt/vol) paraformaldehyde and incubated for 1 hour at 37°C. Fixed bacteria were then washed once with PBS, resuspended in Newborn Calf Serum (Sigma) and incubated for 20 min. at 25°C. The cells were then incubated for 1 hour at 4°C in presence of mAbs diluted 1∶200 in dilution buffer (PBS, 20% Newborn Calf Serum, 0.1% BSA). Cells were washed in PBS-0.1% BSA and incubated for a further 1 h at 4°C with a 1∶100 dilution of R-Phycoerythrin conjugated F(ab)2 goat anti-mouse IgG (Jackson ImmunoResearch Laboratories; Inc.). After washing, cells were resuspended in PBS and analyzed with a FACS CANTO II apparatus (Becton Dickinson, Franklin Lakes, NJ) using FlowJo Software (Tree Star, Ashland, OR).
The assay was performed using differentiated HL-60 as phagocytic cells and live bacteria as target cells. GBS strain 515 was grown to mid-exponential growth phase (A650 nm = 0.3), harvested by centrifugation, and, after washing in cold saline solution, was resuspended in HBSS buffer (Invitrogen). Promyelocytic HL-60 cells (ATCC, CCL-240) were expanded in RPMI 1640 (Invitrogen) containing 10% Fetal clone I (HyClone) at 37°C with 5% CO2 and differentiated into granulocyte-like cells to a density of 4×105 cells/ml by the addition of 100 mM N,N dimethylformamide (DMF, Sigma) to the growth medium. After 4 days, cells were harvested by centrifugation and resuspended in HBSS buffer. Briefly, the reactions took place in a total volume of 125 µl containing ≈3×106 differentiated HL-60, ≈1,5×105 CFU of GBS cells, 10% baby rabbit complement (Cedarlane), and different dilutions of purified mAbs. Immediately before and after 1 h of incubation at 37°C with shaking at 350 rpm, a 25-µl aliquot was diluted in sterile distilled water and plated onto trypticase soy agar plates with 5% sheep blood. A set of negative controls consisted of reactions without phagocytic cells or with heat-inactivated complement. The amount of opsonophagocytic killing (log kill) was determined by subtracting the log of the number of colonies surviving the 1 h assay from the log of the number of CFU at the zero time point.
Surface plasmon resonance (SPR) analyses were performed using a Biacore X100 instrument (GE Healthcare). Protein A and Protein G (Sigma) were immobilized on CM5 biosensors (Biacore) using standard primary amine coupling (Amine Coupling Kit, GE Healthcare) in which the carboxymethylated CM5 dextran layers were activated by mixing equal volumes of 0.4 M N-ethyl-N′-(3-dimethylaminopropyl)carbodiimide (EDC) and 0.1 M N-hydroxysuccinimide (NHS) at a flow rate of 10 µL/min for 7 min injection. Protein A (250 µg/mL) and Protein G (150 µg/mL) in 10 mM sodium acetate pH 4.5, were immobilized on the activated biosensors using a contact time of 9 min at 10 µL/min flow rate. Unreacted NHS-esters were blocked with three injections (4 min each) of 1.0 M ethanolamine hydrochloride, pH 8.5. The immobilization procedure allowed obtaining a Protein A and a Protein G coated biosensors of ∼3500 RU and ∼1000 RU respectively. Untreated flow cell 1 was used as reference. PBS buffer pH 7.2 with 0.005% (v/v) Tween 20 was used as running buffer for protein immobilization and binding experiments.
To perform single cycle kinetics (SCK), monoclonal antibodies (ranging from 5 to 15 nM in running buffer) were captured onto the Protein A and Protein G surfaces, according to their isotype, at a flow rate of 2 µL/min for 4 min injection.
The analyte BP-2a 515 variant in 2-fold serial dilutions in running buffer (starting from 500 or 250 nM, five concentrations in total) was injected over the captured antibody for 2 min at 45 µL/min followed by a 5 or 10 min dissociation. Biosensor regeneration was performed after each cycle and achieved using urea 8 M, pH 10.5 (4 min, 10 µL/min). This treatment did not damage the biosensor surface as shown by equivalent signals of capturing ligand on different runs. Each kinetic experiment was preceded by an identical binding-regeneration cycle of buffer as analyte after mAb capture. This cycle was used as blank and subtracted from all the active curves to correct background effects.
The association, dissociation and affinity constants (ka e kd and KD respectively) were determined by a simultaneous fitting of the kinetic curves with a model of equimolar stoichiometry (1∶1) using the BIAevaluation X100 software version 1.0 (GE Healthcare).
The epitope-mapping approach was based on the method described by Peter and Tomer [24], which we adapted to the following two protocols [25], [26]:
Approximately 5×106 monoclonal antibody-secreting hybridoma cells were collected. Poly(A)+ RNA was isolated using RNeasy Mini Kit according to the manufacturer's instructions (QIAGEN). cDNA was produced via reverse transcription using ∼2 ug of poly(A)+ RNA template and oligo-(dT)12–18 primer using First Strand cDNA Synthesis kit (Novagen). The resulting cDNA was used as a template for PCR amplification using PfuUltra High-Fidelity DNA Polymerase (Stratagene) and degenerated primers specific for FvH and FvL gene fragments (Mouse Ig-Primer Set, Novagen). PCR timing has been set according to the manufacturer's instruction (Mouse Ig-Primer Set, Novagen). Positive PCR products have been purified using a QIAprep Spin Miniprep Kit (QIAGEN) and sequenced.
A total of 144 S. agalactiae BP-2a sequences were examined both from GenBank, including previously published sequences [13], and from complete genome sequences. Codon alignments and phylogenies were constructed using the coding region for each of the 22 unique BP-2a gene sequences with MEGA5 [27]. To detect codons that show signs of adaptive evolution the program HyPhy [19] was used, as implemented in Datamonkey [28]. The codon-based maximum likelihood method REL (Random Effects Likelihood) [29] was used to estimate the dN/dS ratio at every codon in the alignment. The REL method can also take recombination into account, provided that prior to the selection analysis a screening of the sequences for recombination breakpoints is performed. The recombination analysis was performed with GARD [18], using the HKY85 substitution model.
The corresponding amino acid sequences of monoclonal antibodies were used to search the Protein Data Bank (PDB) in order to retrieve suitable templates for modeling. Ten models for each antibody have been obtained with Moleder 9v8 [30]. The best models were selected according to the objective function scoring. The quality of the refined structures obtained was checked with verify Profile-3D module of Discovery Studio 3.0 (Accelrys).
Molecular docking has been performed with ATTRACT docking program [31]. The docking protocol of ATTRACT has already been described in previous publications [16], [31]. Briefly, the antibodies and the target protein (BP-2a 515 variant) coordinates are translated into a reduced protein presentation made up to three pseudo atoms per amino acid residue: the protein backbone is represented by one pseudo atom, small aminoacid side chains (Ala, Asp, Asn, Cys, Ile, Leu, Pro, Ser, Thr, Val) are represented by one pseudo atom and larger and more flexible side chains are represented by two pseudo atoms, to better describe the shape and dual chemical character of side chains. The contacts between pseudo atoms are described by different interaction: Lennard–Jones (LJ)-type potentials (A/r8-B/r6-potential), repulsive and attractive LJ-parameters describing approximately the size and physico-chemical character of the side chain chemical groups [31]. For systematic docking studies, the antibody, called the ligand protein, was used as probe and placed at various positions and various orientations on the surface of the domain D3 of BP-2a 515 crystal structure. We also took into account the experimental data from Mass analysis, setting a weight of 1.5 on exposed aminoacid residues identified and on the surface area (4A) around them. Best docked complexes were selected according to energy scoring function and were finally energy-minimized using the Sander program from the Amber8 package.18. During energy minimization, a Generalized Born (GB) model was employed to implicitly account for solvation effects as implemented in Amber8.
All molecular dynamics (MD) simulations were performed with the GROMACS 4.0.5 simulation package [17] using the AMBER99SB-ILDN force field [32] with explicit water (TIP3P) [33]. The selected energy minimized complexes served as starting structure for MD simulations. After stepwise heating of the systems to 310 K production runs were performed for up to 20 ns with a time step of 2 fs in the NPT ensemble at 310 K and 1 bar. Temperature and pressure were controlled by Nosé-Hoover [34], [35] (coupling constant tt = 2.5) and Parrinello-Rahman [36], [37] (tp = 5.0 ps) schemes, respectively. Figures of the molecular structures were generated with VMD [38] and Discovery Studio 3.0 (Accelrys).
Amounts of 5 – 2 – 0.2 µg of purified peptides (Thermo scientific) were spotted on nitrocellulose membrane (0.45 µm pore size, Biorad) and left to dry for at least 30 minutes at room temperature. The spotted membranes were washed three times with PBST (0.05% Tween 20 in phosphate-buffered saline or PBS pH 7.4) applying a constant vacuum flow using SNAP i.d. Protein Detection System (Millipore) and blocked for 1 h at room temperature in PBST buffer containing 10% of non-fat-dry milk (Biorad). The membranes were then probed 1 h at room temperature with specific anti-BP2a mAb (diluted ∼4.5 µg/mL in PBST/1% non-fat-dry milk) and washed 5 minutes (3X) with PBST and further incubated in PBST/1% non-fat-dry milk for 1 h containing a dilution of 1∶1000 goat anti-mouse horseradish peroxidase-conjugated secondary antibody (Dako, Glostrup, Denmark). Subsequently, the filters were washed 15 minutes (2X) with PBST and developed by enhanced chemiluminescence (ECL) detection assay (Pierce ECL Western blotting substrate, Thermo Fisher Scientific Inc.) following manufacturer's protocols.
|
10.1371/journal.pbio.2006409 | A single pair of leucokinin neurons are modulated by feeding state and regulate sleep–metabolism interactions | Dysregulation of sleep and feeding has widespread health consequences. Despite extensive epidemiological evidence for interactions between sleep and metabolic function, little is known about the neural or molecular basis underlying the integration of these processes. D. melanogaster potently suppress sleep in response to starvation, and powerful genetic tools allow for mechanistic investigation of sleep–metabolism interactions. We have previously identified neurons expressing the neuropeptide leucokinin (Lk) as being required for starvation-mediated changes in sleep. Here, we demonstrate an essential role for Lk neuropeptide in metabolic regulation of sleep. The activity of Lk neurons is modulated by feeding, with reduced activity in response to glucose and increased activity under starvation conditions. Both genetic silencing and laser-mediated microablation localize Lk-dependent sleep regulation to a single pair of Lk neurons within the Lateral Horn (LHLK neurons). A targeted screen identified a role for 5′ adenosine monophosphate-activated protein kinase (AMPK) in starvation-modulated changes in sleep. Knockdown of AMPK in Lk neurons suppresses sleep and increases LHLK neuron activity in fed flies, phenocopying the starvation state. Further, we find a requirement for the Lk receptor in the insulin-producing cells (IPCs), suggesting LHLK–IPC connectivity is critical for sleep regulation under starved conditions. Taken together, these findings localize feeding-state–dependent regulation of sleep to a single pair of neurons within the fruit fly brain and provide a system for investigating the cellular basis of sleep–metabolism interactions.
| Neural regulation of sleep and feeding are interconnected and are critical for survival. Many animals reduce their sleep in response to starvation, presumably to forage for food. Here, we find that in the fruit fly Drosophila melanogaster, the neuropeptide leucokinin is required for the modulation of starvation-dependent changes in sleep. Leucokinin is expressed in numerous populations of neurons within the two compartments of the central nervous system: the brain and the ventral nerve cord. Both genetic manipulation and laser-mediated microablation experiments identify a single pair of neurons expressing this neuropeptide in the brain as being required for metabolic regulation of sleep. These neurons become active during periods of starvation and modulate the function of insulin-producing cells that are critical modulators of both sleep and feeding. Supporting this notion, knockdown of the leucokinin receptor within the insulin-producing cells also disrupts metabolic regulation of sleep. Taken together, these findings identify a critical role for leucokinin signaling in the integration of sleep and feeding states.
| Dysregulation of sleep and feeding has widespread health consequences, and reciprocal interactions between these processes underlie a number of pathologies [1–4]. Sleep loss correlates with increased appetite and insulin insensitivity, while short-sleeping individuals are more likely to develop obesity, metabolic syndrome, type 2 diabetes, and cardiovascular disease [1,3,4]. Although the neural basis for sleep regulation has been studied in detail, little is known about how feeding state and changes in metabolic function modulate sleep [5,6]. Understanding how sleep and feeding states are integrated may provide novel insights into the comorbidity of disorders linked to sleep and metabolic regulation.
Animals balance nutritional state and energy expenditure in order to achieve metabolic homeostasis [6,7]. In both flies and mammals, diet potently affects sleep regulation, supporting the notion that sleep and metabolic state interact [5,6,8]. Starvation leads to sleep loss or disrupted sleep architecture, presumably to induce foraging behavior, while high-calorie diets have complex effects on sleep depending on macronutrient content [9–12]. Behavioral and physiological responses to changes in feeding state are modulated both by cell-autonomous nutrient centers in the brain that detect changes in circulating nutrients and through communication between brain and peripheral tissues [13], yet the neural basis for the integration of sleep and feeding state remain poorly understood.
The fruit fly, D. melanogaster, provides a powerful model for investigating sleep regulation. Flies display all the behavioral hallmarks of sleep, including extended periods of behavioral quiescence, rebound following deprivation, increased arousal threshold, and species-specific changes in posture [14,15]. Many genetic mechanisms regulating sleep are conserved from flies to mammals. In addition, high-throughput systems for sleep analysis in Drosophila have led to the identification of both novel and highly conserved sleep genes [16,17]. Further, stimulants including caffeine, amphetamine, and cocaine have been shown to suppress sleep in flies [15,18,19]. Thus, at the molecular, pharmacological, and behavioral levels, flies provide a model for studying genetic regulation of mammalian sleep.
A number of genes and neurons that are required for the integration of sleep and feeding states have been identified, including core-circadian clock genes, metabolic hormones, and sensory neurons [9,20–22]. While many identified regulators of sleep–metabolism interactions broadly impact these processes [6], a mutation of the DNA/RNA binding protein translin (trsn) disrupts starvation-induced sleep suppression without affecting sleep or metabolic regulation under fed conditions. Targeted knockdown in approximately 30 leucokinin (Lk) neurons phenocopies trsn mutants, raising the possibility that these neurons are required for the integration of sleep and metabolic state [23].
Here, we identify a single pair of Lk neurons in the lateral horn of the fly brain that are required for the integration of sleep and metabolic state. These neurons project near the insulin-producing cells (IPCs), which are critical modulators of sleep and metabolic regulation [24–26]. Lateral Horn leucokinin (LHLK) neurons are dispensable for sleep under fed conditions but are required for starvation-induced sleep suppression. Functional imaging reveals that LHLK neurons have reduced activity in response to glucose application and increased activity under starved conditions. The identification of single neurons that integrate sleep and metabolic state provide a model for investigating the cellular mechanisms regulating the integration of sleep and metabolic state.
Leucokinin (Lk) neuropeptide has been implicated in regulation of feeding, sleep, and circadian activity [27,28]. To determine whether Lk is required for metabolic regulation of sleep, we measured sleep under fed and starved conditions in mated female flies with disrupted Lk expression. In agreement with previous findings, driving Lk-RNAi (RNA interference) under control of Lk-galactose-responsive transcription factor (GAL4) (yeast transcription factor, Lk-GAL4>upstream activation sequence [UAS]-dicer-2 [dcr2],Lk-RNAi) significantly reduced Lk expression (Fig 1A and 1B and S1B Fig). Control flies harboring Lk-RNAi or Lk-GAL4 transgenes alone significantly suppressed sleep during starvation, while no significant differences were detected between the fed and starved states in Lk>Lk-RNAi knockdown flies (Fig 1C and 1D). To confirm that these phenotypes are not due to off-target effects, we measured starvation-induced sleep suppression in two Lk mutants. Lkc275 is a hypomorphic allele containing a piggyBac element upstream of the Lk gene transcription start site (S1A Fig), with approximately 30% reduction in Lk levels (Fig 1E and 1F and S1B Fig) [28]. Flies homozygous for Lkc275 failed to suppress sleep when starved (Fig 1G and 1H). We also used Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR/Cas9) gene-editing to generate a recombinant transgenic line (Lk−/−(GAL4)) by replacing bases 1 to 7 downstream of the Lk ATG start codon with a GAL4-element containing cassette (S1A Fig). Lk protein was not detected in the brains of Lk−/−(GAL4) mutants, confirming that this genetic modification resulted in a robust reduction of Lk function (Fig 1I and 1J and S1B Fig). Sleep while fed did not differ between Lk−/−(GAL4) lines and w1118 controls, suggesting that Lk is not required for sleep regulation under fed conditions (Fig 1K and 1L). Conversely, control w1118 flies, but not Lk−/−(GAL4) flies, robustly suppressed sleep under starved conditions during the day and night, indicating that Lk is required for starvation-induced sleep suppression (Fig 1K and 1L). In agreement with previous findings, starvation induces hyperactivity in control flies (S1C and S1D Fig) [20,29]. Starvation did not alter waking activity in Lkc275 and Lk−/−(GAL4) flies, indicating that Lk is required for both starvation-induced changes in sleep regulation and hyperactivity (S1C and S1D Fig).
To determine whether the sleep phenotype was caused by loss of Lk, we restored Lk in the background of each mutant and measured sleep. Pan-neuronal rescue (Lkc275; embryonic lethal abnormal vision [elav]-GAL4/UAS-Lk) restored starvation-induced sleep suppression (S1E Fig) and starvation-induced hyperactivity (S1F Fig). Rescue flies did not differ from heterozygous controls (Lkc275; elav-GAL4/+ or UAS-Lk;Lkc275/+). The GAL4 insertion in Lk−/−(GAL4) drives expression to a pattern similar to Lk antibody, providing the opportunity to restore Lk to its endogenous expression pattern (S1G Fig). Similar to Lkc275, rescue in Lk neurons (Lk−/−(GAL4); UAS-Lk) restored starvation-induced sleep suppression (S1H Fig) and starvation-induced hyperactivity (S1I Fig), confirming that Lk is required for the metabolic regulation of sleep. Flies heterozygous for Lk−/−(GAL4) failed to suppress sleep in response to starvation (Fig 1K) and displayed reduced starvation-induced hyperactivity compared to control (S1D Fig), raising the possibility that haploinsufficiency impacts sleep regulation. This result is consistent with the notion that the moderate reduction in Lk levels in Lkc275 mutants can affect diverse behavioral phenotypes [27]. Expression of a rescue transgene in Lk neurons of flies heterozygous for Lk (Lk+/−(GAL4);UAS-Lk) also restored starvation-induced sleep suppression (S1J Fig), confirming the specificity of phenotype in Lk−/−(GAL4) flies. Taken together, three independent genetic manipulations that perturb Lk expression inhibit starvation-induced changes in sleep and activity.
Leucokinin antibody labels the LHLK neurons and the subesophoageal ganglion Lk (SELK) neurons, as well as a number of abdominal Lk (ABLK) neurons in the ventral nerve cord (Fig 2A). To localize the population of neurons that regulate starvation-induced sleep suppression, we restricted GAL4 expression primarily to the brain through the expression of GAL80, a GAL4 repressor, in the ventral nerve cord using teashirt-GAL80 (tsh-GAL80) [30]. Expression of CD8::GFP (Lk-GAL4>CD8:green fluorescent protein [GFP];tsh-GAL80) revealed tsh-GAL80 blocks expression in all but two ventral nerve cord neurons without affecting expression in the brain (Fig 2B). Silencing the remaining Lk neurons with light-chain tetanus toxin (TNT) (tsh-GAL80;Lk-GAL4>UAS-TNT) abolished starvation-induced sleep suppression, phenocopying the effects of silencing all Lk neurons (Lk-GAL4>UAS-TNT) (Fig 2C and 2D) [31]. Further, no differences in sleep were detected between groups in fed flies, and there was no effect of expressing an inactive variant of TNT light chain (impTNT) in Lk neurons (Fig 2C and 2D). These findings suggest Lk neurons within the brain are required for sleep–metabolism interaction.
It has previously been reported that Apterous-GAL4 (Apt-GAL4) drives expression in the LHLK neurons, as well as neurons in the optic lobe and antennal mechanosensory and motor centers (AMMCs), and a small population of mushroom-body neurons (S2A Fig) [27,32]. Immunostaining with Lk antibody in Apt-GAL4>UAS-mCD8::GFP flies confirmed colocalization exclusively within the LHLK neurons (S2A and S2B Fig). To functionally assess the role of LHLK neurons, we genetically silenced LHLK neurons, as well as other non-Lk cells labeled by Apt-GAL4. Silencing neurons labeled by Apt-GAL4 (Apt-GAL4>UAS-TNT) inhibited starvation-induced sleep suppression and promoted sleep while fed, while no effects were observed in flies expressing impTNT (Apt-GAL4>UAS-impTNT) (Fig 2E). While these findings suggest a role for Apt-GAL4–labeled neurons in sleep regulation, it is possible that the phenotype is independent of Lk function.
To verify that the sleep phenotype was due to blocking Lk release from LHLK neurons, we sought to disrupt Lk function selectively in Apt-GAL4–labeled neurons. Expression of Lk-RNAi in Apt-Gal4 neurons (Apt-GAL4>UAS-dcr2, Lk-RNAi) disrupted starvation-induced sleep suppression (Fig 2F). As a complementary approach, we restored Lk to Apt-GAL4 neurons in the Lkc275 mutant background and measured sleep. Rescue in Apt-GAL4 neurons (Lkc275; Apt-GAL4>Lkc275; UAS-Lk) restored starvation-induced sleep suppression to heterozygote control levels (Apt-GAL4; Lkc275/+ and UAS-Lk; Lkc275/+), whereas flies harboring either GAL4 or UAS in the Lkc275 mutant background failed to suppress sleep (Apt-GAL4; Lkc275 and UAS-Lk; Lkc275; Fig 2G). These data support a role for the LHLK neurons in starvation-induced sleep suppression but do not rule out the possibility that other neurons labeled by Apt-GAL4 also contribute to this phenotype.
To complement genetic silencing experiments, we sought to precisely ablate the LHLK neurons and measure their role in starvation-induced sleep suppression (Fig 3A). Multiphoton microscopy has been used in diverse genetic models for targeted ablation of neuronal cell types [33–35]. All adult Lk neurons are present in third-instar larvae and labeled by the Lk-GAL4. The SELK and anterior Lk (ALK) neurons are easily visualized, providing the opportunity to independently ablate these subtypes and measure the effect on adult behavior in an intact animal (Fig 3B). We selectively induced bilateral ablations of LHLK neurons or unilateral ablation of two control ALK neurons in immobilized third-instar larvae with a titanium sapphire multiphoton laser. Ablation of individual neurons could be visualized in larvae as a disruption of the GFP-labeled neuronal cell body (Fig 3C). Following ablation, larvae were transferred back into food vials, and 5- to 7-day–old adult flies were tested for sleep under fed and starved conditions. After behavioral testing, brains were dissected and imaged to verify selective bilateral ablation of the LHLK neurons or unilateral ablation of two ALK neurons (Fig 3F). Flies with ablated ALK neurons suppressed sleep during starvation similarly to controls (Fig 3D and 3E). Conversely, bilateral ablation of the LHLK neurons abolished starvation-induced sleep suppression without affecting sleep while fed, revealing an essential role for the LHLK neurons in the integration of sleep and metabolic state (Fig 3D, 3E and 3F).
The finding that LHLK neurons are required for starvation-induced sleep suppression raises the possibility that the activity of Lk neurons is modulated by nutritional state. We selectively expressed a GFP-calmodulin and MP13 peptide sequence (GCaMP6m)-mCherry (UAS-Gerry) fusion protein that allows for ratiometric detection of Ca2+ activity [36,37] in Lk neurons and measured the response to nutrients. The brains of fed or 24-hr–starved flies were imaged for GCaMP and mCherry signal ex vivo (Fig 4A). Flies expressing Gerry in Lk neurons suppressed sleep similarly to control flies harboring Lk-GAL4 alone, indicating that expression of the Ca2+ sensor does not affect starvation-induced regulation of sleep (S3A Fig). The GCaMP/mCherry ratio was elevated in the LHLK neurons of starved flies compared to fed controls, suggesting these neurons are more active during starvation (Fig 4B). Conversely, no difference in the GCaMP/mCherry ratio between the fed and starved states was detected in the SELK neurons (Fig 4C).
In mammals, the activity of some sleep- and wake-promoting neurons are directly modulated by glucose and other circulating nutrients [38,39]. It is possible that the activity of Lk neurons is modulated in accordance with feeding state by sensory detection of tastants or indirectly result upon detection of changes in circulating nutrients. To differentiate between these possibilities, the brains of fed flies were removed and treated with either glucose or the competitive inhibitor of glycolysis, 2 deoxy-glucose (2DG) [40–42]. Application of glucose reduced Ca2+ activity in LHLK neurons compared to controls treated with Drosophila artificial hemolymph alone, suggesting these neurons are sensitive to circulating glucose (Fig 4D). To verify that application of glucose was specific to LHLK neurons, we applied glucose or artificial hemolymph while measuring Ca2+ response in SELK neurons. No significant differences in Ca2+ activity were found between controls and glucose application (S3B Fig). Further, the combined application of 2DG and glucose increased Ca2+ activity to levels greater than hemolymph alone, mimicking the starved state (Fig 4D). Taken together, these findings indicate that the activity of Lk neurons are modulated in accordance with nutrient availability and support the notion that the LHLK neurons are more active during starvation, thereby suppressing sleep.
To examine whether the activity in Lk neurons is modulated by feeding state in an intact animal, we performed in vivo recordings in tethered flies (Fig 4E). Briefly, a portion of the head capsule was removed so that the LHLK neurons were accessible. The activity of LHLK neurons was then recorded in flies that had been previously fed or starved for 24 hr. In agreement with ex vivo findings, the GCaMP/mCherry ratio was elevated in the LHLK neurons of starved flies, fortifying the notion that Lk neurons are more active during starvation (Fig 4F). Refeeding flies with standard food reduced the GCaMP/mCherry ratio 3 hr after refeeding (Fig 4F). Taken together, these findings suggest the activity of LHLK neurons are modulated by circulating nutrient levels.
The identification of nutrient-dependent changes in activity of LHLK neurons raises the possibility that cell-autonomous nutrient sensors or signaling pathways function within Lk neurons to modulate sleep. To identify regulators of sleep that modulate the activity of Lk neurons, we expressed RNAi targeted to 28 RNAi lines encoding putative nutrient sensors or signaling pathways using Lk-GAL4 and measured starvation-induced changes in sleep (S4A Fig). RNAi knockdown of AMP-activated protein kinase alpha (AMPKα) in Lk neurons (Lk-GAL4>UAS- AMPKα) alone abolished starvation-induced sleep suppression compared to GAL4 controls crossed to the isogenic host strain for the RNAi library (Fig 5A and 5B and S4A Fig) [43]. Feeding did not differ in flies expressing AMPKα-RNAi in Lk neurons, suggesting the sleep phenotype is not due to generalized changes in hunger (S4B Fig). Targeting AMPKα-RNAi with a second, independently derived RNAi line also abolished starvation-induced sleep suppression (Lk-GAL4>UAS-dcr2, UAS-AMPKα-RNAi #2; S4C Fig). Genetically restricting AMPK knockdown in flies harboring tsh-GAL80 (tsh-GAL80;Lk-GAL4>AMPKα-RNAi) also impaired starvation-induced sleep suppression (Fig 5C and 5D), suggesting AMPKα functions in the Lk neurons within the brain to regulate sleep.
To determine whether inhibition of AMPK signaling changes the activity of Lk neurons to resemble a starved-like state, we genetically expressed AMPKα-RNAi under control of Lk-GAL4 and measured neuronal activity in vivo using UAS-Gerry (Fig 5E and 5F; Lk-GAL4>UAS-AMPKα-RNAi; UAS-Gerry). Genetic disruption of AMPKα increased Ca2+ activity in LHLK neurons of fed flies compared to flies expressing UAS-Gerry alone (Fig 5E). The increase Ca2+ activity phenocopies changes found in starved control flies, suggesting that the loss of AMPK increases the activity of Lk neurons, thereby suppressing sleep (Fig 5F). Together, these findings suggest AMPK is active within LHLK neurons during the fed state, and reduced AMPK signaling during starvation increases LHLK activity.
Lk signals through a single Lk receptor (Lkr) that is highly expressed in the IPCs and the dorsal fan-shaped body (dFSB), both of which have been implicated in sleep regulation [24,44–46]. To determine the role of Lkr in sleep regulation, we used CRISPR/Cas9 gene-editing to generate a recombinant transgenic line (Lkr−/−(GAL4)) with a GAL4 element inserted 106 to 111 base pairs preceding the ATG translational start site, disrupting its function. Consistent with previous reports of Lkr expression, transgene expression in Lkr−/−(GAL4)>UAS-mCD8::GFP flies labeled the IPCs, dFSB, and a number of other brain regions (Fig 6A). To determine the role of Lkr in sleep, we tested flies for sleep under fed and starved conditions. Lkr−/−(GAL4) flies failed to suppress sleep when starved, phenocopying loss of Lk function (Fig 6B). Restoring Lkr to Lkr mutant flies (Lkr−/−(GAL4)>UAS-Lkr) rescued starvation-induced sleep suppression (Fig 6B). The promoter-fusion R65C07-GAL4 predominantly labels the dFSB, while R67D01-GAL4 primarily labels the IPCs ([47], Fig 6C and 6F). Silencing R65C07-GAL4–labeled neurons with TNT reduced sleep in fed flies, consistent with a sleep-promoting role for the dFSB ([45]; Fig 6D). However, there was no effect of Lkr knockdown in these cells (R65C07>UAS-dcr2, Lkr-RNAi), suggesting that Lk does not signal through the dFSB to modulate sleep. Conversely, genetic silencing or expression of Lkr-RNAi in R67D01-GAL4 neurons that label the IPCs abolished starvation-induced sleep suppression (Fig 6G and 6H). Together, these findings suggest Lkr function in the IPCs is required for starvation-induced sleep suppression. To validate a role for the IPCs, we selectively targeted Lkr in neurons labeled by Drosophila insulin-like peptide 2 (Dilp2) (Fig 6I). Selectively knocking down Lkr in Dilp2 neurons (Dilp2-GAL4>UAS-dcr2, Lkr-RNAi) prevented starvation-induced sleep loss, indicating that Lkr is required in Dilp2 neurons for starvation-induced sleep suppression (Fig 6J and 6K).
Our findings reveal that Lk signals through a single Lkr to integrate sleep and metabolic state. Previous studies have identified a number of genes required for starvation-induced changes in sleep or locomotor activity, yet many of these genes have pleiotropic functions on behavior or metabolic function [20,29,48,49]. For example, the glucagon-like adipokinetic hormone (AKH) is responsible for energy mobilization, and genetic disruption of AKH induces obesity and abolishes starvation-induced hyperactivity [29,50,51]. Similarly, the circadian transcription factors clock and cycle are required for starvation-dependent regulation of behavior, and loss of function affects sleep both in fed and starved conditions [20]. Conversely, neuropeptide F functions within a subpopulation of circadian neurons and is selectively required for metabolic regulation of sleep [22]. Our findings that genetic manipulations that inhibit Lk signaling selectively disrupt starvation-induced modulation of sleep suggest that different neural mechanisms regulate sleep under basal conditions and in response to environmental perturbation.
The failure of Lk mutants to suppress sleep under starved conditions phenocopies mutation of the RNA/DNA binding protein trsn. Loss of trsn does not impact feeding behavior but impairs starvation-induced sleep suppression, suggesting that trsn is not generally required for hunger-induced behavior [23]. While trsn is broadly expressed in the fly nervous system [52,53], we previously found that selective knockdown of trsn in Lk neurons disrupted starvation-induced sleep suppression [54]. These findings raise the possibility that trsn functions to regulate changes in the physiology of Lk neurons to modulate sleep under starved conditions.
A central question is how Lk neurons modulate numerous complex behaviors and physiological processes. In adults, Lk is expressed in four pairs of neurons in the brain and 11 pairs in the ventral nerve cord, which regulate diverse behaviors and physiological processes [55–57]. Recent work suggests the ABLK neurons in the thoracic ganglion are critical regulators of water consumption and contribute to the altered stress resistance and water content in Lk mutant flies [46]. The SELK neurons connect the gustatory receptors to the subesophageal ganglia and ventral nerve cord. Although a specific function has not been identified to SELK neurons, silencing of all Lk neurons disrupts gustatory behavior, and a mutation in the Lk locus affects meal size [28,58], raising the possibility that these behaviors are regulated by SELK neurons. Lastly, the LHLK neurons project to the superior lateral protocerebrum, medial protocerebrum, and peduncle and axonal stalk of the mushroom bodies [56]. The LHLK neurons we identify here as regulating sleep–metabolism interactions receive inputs from clock neurons and are thought to modulate locomotor activity and sleep [27]. Supporting the notion that LHLK neurons are outputs of the clock, previous work has found that silencing Lk within these neurons attenuates circadian rhythms [27]. Conversely, we find that silencing of these neurons has little effect on sleep yet abolishes starvation-induced sleep suppression. Together, these findings support a role for LHLK neurons in sleep regulation, yet discrepancies remain about the conditions under which these neurons regulate sleep.
Lk neurons have also been implicated in the modulation of sleep following the ingestion of a meal. Postprandial sleep is enhanced in Lk-deficient flies and reduced in flies with thermogenetically activated Lk neurons sleep, whereas Lk knockdown increases the probability of falling asleep after a meal [59]. Similarly, we found that knockdown of Lk in Lk-expressing neurons results in increased sleep during starvation. In contrast, it was previously reported that Lk signals the fan-shaped body neurons to regulate postprandial sleep [59]. These results raise the possibility that distinct circuitry regulating starvation-induced sleep suppression and postprandial sleep, which could be influenced by distinct time-scales at which both behaviors are being executed. In addition, it is possible that a distinct subpopulation of Lk neurons is responsive to individual nutrients in comparison to starvation itself, which is sensed by LHLK neurons. Taken together, these findings suggest different neural mechanisms underlie Lk-dependent regulation of postprandial feeding and circadian rhythms.
The Drosophila genome encodes for a single Lk target, the Lkr, that is expressed in the lateral horn, the ventral nerve cord, the IPCs, and the sleep-promoting fan-shaped body [27,28,59]. The fan-shaped body, a subregion of the Drosophila central complex, is a primary sleep-promoting region [45,60], while the IPCs are proposed integrators of sleep and feeding state [61]. Previous studies suggest that Lkr function within the fan-shaped body is required for proper regulation of circadian rhythms and postprandial sleep [27,59]. Here, we find that targeted knockdown of Lkr within the IPCs phenocopies LHLK ablation, suggesting the LHLK neurons signal to the IPCs to modulate starvation-induced regulation of sleep. Supporting these findings, recent work has shown that Lk and Lkr mutants display increased levels of Dilp2 and Dilp3 immunoreactivity in the IPCs [46]. However, anterograde trans-synaptic labeling revealed no direct synaptic inputs between Lk neurons and the IPCs, raising the possibility that that Lk inputs to the IPCs could occur via paracrine signaling [46]. Taken together with previous studies, these findings suggest different Lk targets regulate postprandial feeding, circadian rhythms, and sleep or that Lk functions through paracrine signaling to modulate different targets.
In Drosophila, a number of circulating nutrients including fructose, trehalose, and glucose have been found to affect central brain physiology and behavior [42,62,63]. While nutrients may be detected by gustatory receptors expressed in the periphery to regulate sleep [9,64], sugar receptors and transporters are also expressed in the brain [65]. The identification of LHLK neurons as being active under starvation conditions and suppressed by glucose provide a system to investigate feeding-state–dependent changes in neural activity. A number of neurons in the fly brain are acutely regulated by feeding state, including the starvation-active Taotie neurons that inhibit IPCs of the pars intercerebralis to regulate insulin-like peptide release under nutrient deprivation conditions [66–68]. Conversely, the IPCs function as cell-autonomous nutrient sensors that are activated by glucose through the inhibition of KATP channels [69]. Further, the LHLK nutrient phenotype is similar to the neurons within the ellipsoid body labeled by the sodium/glucose co-transporter SLC5A11 that are active during starvation and promote feeding [41,70]. SLC5A11 and its cognate neurons are required for a variety of hunger-induced feeding behaviors, but the effect on sleep has not been identified [70]. Our screen found that knockdown of SLC5A11 in Lk neurons did not affect starvation-induced sleep suppression, suggesting alternative regulators of sleep. The identification of LHLK neurons as starvation-active neurons provides a system for identification of additional nutrient sensors that regulate sleep.
AMPK functions as a cell-autonomous regulator of energy allocation and induces physiological changes associated with starvation [71,72]. AMPK consists of a heterotrimeric complex that is activated by AMP and modulates diverse intercellular signaling pathways, including mammalian target of rapamycin (mTOR), forkhead box (FoxO), and sirtuin-1 (SIRT1) [73]. Canonically, AMPK is activated during starvation and increases neuronal activity, though this effect varies by neuronal subtype [74,75]. For example, in Caenorhabditis elegans, starvation-induced AMPK activation leads to inhibition of neurons that modulate local search behavior in response to food deprivation while promoting activity in neurons that trigger dispersal behavior [76]. Here, we find that knockdown in LHLK neurons using multiple independently derived RNAi lines results in flies that reduce sleep under the fed state and increases the activity of LHLK neurons, similar to neural activity seen during the starved state. Ubiquitous disruption of AMPK in Drosophila induces hypersensitivity of the locomotor response to starvation and reduces starvation resistance [74]. Conversely, we find that selectively disrupting AMPK function in Lk neurons promotes starvation-induced hyperactivity and sleep loss during the fed state, suggesting a neural-circuit–specific function for AMPK. While our findings suggest that AMPK functions as an important modulator of LHLK neuronal activity and state-dependent changes of activity within LHLK neurons, it is also possible that AMPK generally modulates the activity of Lk neurons, resulting in sleep loss. The findings that the activity of SELK neurons is not elevated during starvation raises the possibility of neuron-specific AMPK function.
Taken together, the identification of LHLK neurons as critical modulators of sleep–metabolism interactions provides a system for identifying novel nutrient-sensing and signaling mechanisms that modulate sleep. These findings illustrate the need to determine how Lk neurons modulate different aspects of sleep regulation, including their reported role as circadian output neurons, in regulation of sleep–metabolism interactions, and in postprandial sleep regulation [23,27,59]. Further investigation of feeding-state–dependent changes in Lk signaling and the identification of neuronal inputs and targets of LHLK neurons will provide mechanistic insight into how animals integrate sleep with changes in their internal and external environments.
Flies were grown and maintained on standard food (Bloomington Recipe, Genesee Scientific, San Diego, CA, USA). Flies were kept in incubators (Dros52; Powers Scientific, Warminster, PA, USA) at 25 °C on a 12:12 LD cycle with humidity set to 55%–65%. The background control line used in this study is the w1118 fly strain, and all experimental flies were outcrossed 6–8 generations into this background. All the experiments performed in this manuscript used mated female flies. The following fly strains were ordered from Bloomington Stock Center: w1118 (5905; [77]), Lkc275 (16324; [28]), elav-GAL4 (8765; [78]), Apt-GAL4 (3041; [79]), UAS-TNT (28996; [31]), UAS-impTNT (28840; [31]), UAS-mCD8::GFP (32186; [80]), UAS-dcr2 (Chr II;24650; [43]), UAS-dcr2 (Chr III;24651; [43]), AMPKα-RNAi#2 (35137; [81]), UAS-Lkr-RNAi (65934; [81]), UAS-luciferase (31603; [81]), Lkr-GAL4 (39344; [47]), and Lkr-GAL4 (39412; [47]). The following lines were generated in this study: Lk−/−(GAL4), Lkr−/−GAL4), and UAS-Lk. UAS-Gerry was a kind gift from Greg Macleod, Lk-GAL4 and Dilp2-GAL4 from Young-Joon Kim, and UAS-Lkr from Bader Al Anzi [28]. tsh-GAL80 [30] was provided by Julie Simpson. Drosophila lines used in the RNAi screen and UAS-Lk-RNAi (14091) originate from the VDRC library [43] and are described in Table 1.
Lk−/−(GAL4) and Lkr−/−(GAL4) were generated by Wellgenetics (Taipei City, Taiwan) using the CRISPR/Cas9 system to induced homology-dependent repair (HDR) using one guide RNA (gRNA) (Lk−/−(GAL4): GATCTTTGCCATCTTCTCCAG and Lkr−/−(GAL4): GTAGTGCAATACATCTTCAG). At the gRNA target site, a donor plasmid was inserted containing a GAL4::VP16 and floxed 3xP3-RFP cassette [82]. For Lk−/−(GAL4), following the translational ATG start site, bases 1 to 7 were replaced by the knock-in cassette. For Lkr−/−(GAL4), preceding the ATG start site, bases 111 to 106 were replaced by the knock-in cassette. All lines were generated in the w1118 background [77]. Proper insertion loci for both mutations were validated by genomic PCR.
The full-length open reading frame of Lk was amplified from the Lk-pOT2 plasmid (Drosophila Genomics Resource Center [DGRC], #1378621) using specific primers (forward primer: GCCTTTGGCCGTCAAGTCTA and reverse primer: CTCCAAGTACCGCAGGTTCA) generated by Integrated DNA Technologies (Coralville, IA, USA). Amplified sequence was inserted into the pENTER vector (Invitrogen) via TOPO cloning and subsequently recombined into pTW destination vector (DGRC, #1129) using standard gateway cloning protocol as per manufacturer’s instructions (Invitrogen, Carlsbad, CA, USA). The plasmids were verified by sequencing (Genewiz, Morrisville, NC, USA). Transgenic lines were established via phiC31-mediated integration at the attp40 landing site [83] on the second chromosome (BestGene, Chino Hills, CA, USA).
The DAMS detects activity by monitoring infrared beam crossings for each animal [84]. These data were used to calculate sleep information by extracting immobility bouts of 5 minutes using the Drosophila Counting Macro [85,86]. For experiments examining the effects of starvation on sleep, flies were kept on a 12:12 LD cycle. Mated female flies were briefly anesthetized with CO2 and placed into plastic tubes containing standard food. All flies were given 24 hr to recover after being anesthetized. Activity was recorded for 24 hr in food tubes prior to transferring flies into tubes containing 1% agar diluted in dH2O (Fisher Scientific) at Zeitgeber time (ZT) 0. Activity was monitored for an additional 24 hr on agar. For the screen, percent change in sleep during starvation was calculated as the sleep duration on agar minus the sleep duration in food tubes, divided by the sleep duration in food tubes for each fly assayed multiplied by a hundred [11,54].
The brains of 5- to 7-day–old female flies were dissected between ZT 4–ZT 9 in ice-cold phosphate-buffered saline (PBS) and fixed in 4% paraformaldehyde, PBS, 0.5% Triton-X for 30 minutes as previously described [87]. Brains were then rinsed 3× with PBS, 0.5% Triton-X (PBST) for 10 minutes and overnight. In the following day, brains were incubated for 24 hr in primary antibody (1:1,000 rabbit anti-Lk [88] and mouse 1:20 nc82; Iowa Hybridoma Bank, University of Iowa, Iowa City, IA, USA) diluted in PBST at 4 °C. Brains were rinsed in PBST 3× for 10 minutes and placed in secondary antibody (1:400 donkey anti-rabbit Alexa 555 and 1:200 donkey anti-mouse Alexa 647; Thermo Fisher Scientific, Waltham, MA, USA), diluted in PBST for 90 minutes at room temperature. Finally, all samples were washed in PBST for a total of 120 minutes and mounted in Vectashield (VectorLabs, Burlingame, CA, USA). Samples were imaged in 2-μm sections with a Nikon A1R confocal microscope (Nikon, Tokyo, Japan) using a 20× or 60× oil immersion objective. Images were then processed with NIS Elements 4.40 (Nikon).
Briefly, flies were maintained on standard fly food. At ZT 0, flies were transferred to vials containing 1% agar, 5% sucrose, and 2.5% blue dye (FD&C Blue Dye No. 1). Following 30 minutes of feeding, flies were flash frozen on dry ice, and four flies were homogenized in 400 μL PBS (pH 7.4, Thermo Fisher Scientific) per sample. Color spectrophotometry was then used to measure absorbance at 655 nm in a 96-well plate reader (iMark; Millipore Sigma, Burlington, MA, USA). Baseline absorbance was determined by subtracting the absorbance measured in non-dye–fed flies from each experimental sample.
Five- to seven-day–old female flies were collected and placed in vials containing fresh food (fed) or a wet KimWipe paper (starved) for 24 hr. All experiments were done between ZT 4–ZT 7 to account for rhythmic excitability of Lk neurons [27]. For imaging brain explants, previously established methods for calcium imaging were used with modifications [42,65]. Brains of fed or 24-hr–starved flies were dissected and placed in glass wells (Pyrex, Corning, Corning, NY, USA) containing artificial hemolymph (140 mM NaCL, 2 mM KCl, 4.5 mM MgCl2, 1.5 mM CaC2, and 5 mM HEPES-NaOH with pH 7) and allowed a 5-minute recovery period before being recorded. For 2DG experiments, fed brains were dissected and placed in 400 mM 2DG (Sigma Aldrich) in artificial hemolymph, 200 mM glucose (Sigma Aldrich) in artificial hemolymph, or artificial hemolymph alone for a total of 70 minutes. Every 20 minutes, solutions were exchanged. Coverslips were treated with poly-L-lysine (Sigma Aldrich) to ensure that brains were in the same position during imaging and placed onto chamber (RC-21BBDW; Warner Instruments, Hamden, CT, USA). Fly brains were bathed in artificial hemolymph solution and imaged using a 20× air objective lens on an inverted confocal microscope (Nikon A1R on a Ti-E inverted microscope). The pinhole was opened to 244.43 μm to allow a thicker optical section to be monitored. UAS-GCaMP-R (GCaMP and mCherry) was expressed in Lk neurons and simultaneously excited with wavelengths of 488 nm (FITC) and 561 nm (TRITC). Recording were taken for 120 seconds, capturing 1 frame/5 seconds with 512 × 512 resolution. For analysis, regions of interest (ROIs) were drawn manually, capturing the same area between experimental and control. The mean fluorescence intensity was subtracted from background mean fluorescence intensity for FITC and TRITC per frame. Then, the ratio of GCaMP6.0 to mCherry was calculated and plotted as an average of the total time recorded per brain imaged.
In vivo imaging was performed using a previously described protocol with some modifications [89,90]. Briefly, fed, 24-hr–starved, or 3-hr re-fed (standard Bloomington Recipe) flies were anesthetized on ice and secured in a 200-μL pipette tip with head and proboscis accessible. The pipette tip was placed in a small chamber at an angle of 140°, allowing the dorsal and posterior surface of the brain to be imaged. A small hole was cut in the tin foil and fixed to the stage and fly head, leaving a window of cuticle exposed, then sealed using dental glue (Tetric EvoFlow; Ivoclar Vivadent, Schaan, Lichtenstein). The proboscis was extended, and dental glue was used to secure it in place, ensuring the same position throughout the experiment.
A 21-gauge 1 1/4 needle (PrecisionGlide; Becton Dickinson, Franklin Lakes, NJ, USA) was used to cut a window in the fly cuticle. A drop of artificial hemolymph was placed on the cuticle, and the connective tissue surrounding the brain was dissected. Flies were allowed to recover from the procedure for 30–45 minutes in a humidified box. Mounted flies were placed under a confocal microscope (Nikon A1R on an upright microscope) and imaged using a 20× water-dipping objective lens. The pinhole was opened to 244 μm to allow a thicker optical section to be monitored. The settings and data analysis were performed as described above.
Female third-instar larvae expressing GFP in Lk neurons were selected and anesthetized in ethyl ether (Thermo Fisher Scientific, E134-1) for 2–5 minutes. Larvae were placed dorsally on a microscope slide, and a coverslip was placed on the larvae. Ringer’s solution was applied onto the larvae below the coverslip. Larvae were imaged using a 25× water-dipping objective lens on a multiphoton microscope (Nikon A1R) containing a Chameleon Vision II Ti:Sapphire tunable laser. Excitation laser light of 870 nm was used. Images were acquired at 1 frame per second with a resolution of 512 × 512 pixels. For each neural ablation, a total of four frames were acquired. Two frames were captured prior to ablation for a duration of approximately 2 seconds, followed by ROI stimulation of 2–4 seconds and two frames after ablation. Following ablations, larvae were placed in vials containing food and allowed to grow. Sleep in food tubes and on agar was measured 5–7 days posteclosion in the DAMS. In order to verify which neurons were ablated after behavioral assay, flies were anesthetized on ice, and the central nervous system (CNS) was dissected. Fly CNS was fixed in 4% paraformaldehyde, 0.5% PBST for 30 minutes. Following fixation, samples were imaged in 2-μm sections with a Nikon A1R confocal microscope (Nikon) using a 20× oil immersion objective. Ablations that resulted in the formation of supernumerary neurons or deletions of two different subpopulations of Lk neurons were removed from analysis.
The experimental data are presented as means ± SEM. Unless otherwise noted, a one-way or two-way analysis of variance (ANOVA) followed by Tukey’s post hoc test was used for comparisons between two or more genotypes and one treatment and two or more genotypes and two treatments. Unpaired t test was used for comparisons between two genotypes. All statistical analyses were performed using InStat software (GraphPad Software 6.0) with a 95% confidence limit (p < 0.05).
|
10.1371/journal.pcbi.1005129 | Optimized Treatment Schedules for Chronic Myeloid Leukemia | Over the past decade, several targeted therapies (e.g. imatinib, dasatinib, nilotinib) have been developed to treat Chronic Myeloid Leukemia (CML). Despite an initial response to therapy, drug resistance remains a problem for some CML patients. Recent studies have shown that resistance mutations that preexist treatment can be detected in a substantial number of patients, and that this may be associated with eventual treatment failure. One proposed method to extend treatment efficacy is to use a combination of multiple targeted therapies. However, the design of such combination therapies (timing, sequence, etc.) remains an open challenge. In this work we mathematically model the dynamics of CML response to combination therapy and analyze the impact of combination treatment schedules on treatment efficacy in patients with preexisting resistance. We then propose an optimization problem to find the best schedule of multiple therapies based on the evolution of CML according to our ordinary differential equation model. This resulting optimization problem is nontrivial due to the presence of ordinary different equation constraints and integer variables. Our model also incorporates drug toxicity constraints by tracking the dynamics of patient neutrophil counts in response to therapy. We determine optimal combination strategies that maximize time until treatment failure on hypothetical patients, using parameters estimated from clinical data in the literature.
| Targeted therapy using imatinib, nilotinib or dasatinib has become standard treatment for chronicle myeloid leukemia. A minority of patients, however, fail to respond to treatment or relapse due to drug resistance. One primary driving factor of drug resistance are point mutations within the driving oncogene. Laboratory studies have shown that different leukemic mutants respond differently to different drugs, so a promising way to improve treatment efficacy is to combine multiple targeted therapies. We build a mathematical model to predict the dynamics of different leukemic mutants with imatinib, nilotinib and dasatinib, and employ optimization techniques to find the best treatment schedule of combining the three drugs sequentially. Our study shows that the optimally designed combination therapy is more effective at controlling the leukemic cell burden than any monotherapy under a wide range of scenarios. The structure of the optimal schedule depends heavily on the mutant types present, growth kinetics of leukemic cells and drug toxicity parameters. Our methodology is an important step towards the design of personalized optimal therapeutic schedules for chronicle myeloid leukemia.
| Chronic Myeloid Leukemia (CML) is an acquired hematopoietic stem cell disorder leading to the over-proliferation of myeloid cells and an increase in cellular output from the bone marrow that is often associated with splenomegaly. The most common driving mutation in CML is a translocation between chromosomes 9 and 22 that produces a fusion gene known as BCR-ABL. The BCR-ABL protein promotes proliferation and inhibits apoptosis of myeloid progenitor cells and thereby drives expansion of this cell population. By targeting the BCR-ABL oncoprotein, imatinib (brand name Gleevec) is able to induce a complete cytogenetic remission in the majority of chronic phase CML patients. A minority of patients, however, either fail to respond or eventually develop resistance to treatment with imatinib [1]. It is thought that a primary driver of this resistance to imatinib is point mutations within the BCR-ABL gene. A recent study utilizing sensitive detection methods demonstrated that a small subset of these mutations may exist before the initiation of therapy in a significant fraction of patients, and that this status is correlated with eventual treatment failure [2]. Second generation agents such as dasatinib and nilotinib have been developed and each has shown efficacy against various common mutant forms of BCR-ABL. This leads to the observation that the various mutant forms of BCR-ABL result in CML that have unique dynamics under therapy, and that combinations of these inhibitors may be necessary to effectively control a rapidly evolving CML population. Patients with CML often die due to transformation of the disease into an acute form of leukemia known as blast crisis. It has been shown that blast crisis is due to the accumulation of additional mutations in CML progenitor cells [3].
The goal of this work is to leverage the differential responses of CML mutant strains to design novel sequential combination treatment schedules using dasatinib, imatinib and nilotinib that optimally control leukemic burden and delay treatment failure due to preexisting resistance. We develop and parametrize a mathematical model for the evolution of both wild-type (WT) CML and mutated (resistant) CML cells in the presence of each therapy. Then we formulate the problem as a discrete optimization problem in which a sequence of monthly treatment decisions is optimized to identify the temporal sequence of imatinib, dasatinib and nilotinib administration that minimizes the total CML cell population over a long time horizon.
There has been a significant amount of work done in the past to mathematically model CML. For example, in [4] the authors developed a system of ordinary differential equations (ODEs) that model both the normal progression from stem cell to mature blood cells and abnormal progression of CML. A hierarchical system of differential equations was used to model the response of CML cells to imatinib therapy in [5]; this model fit the biphasic nature of decline in BCR-ABL positive cells during imatinib treatment. In [6] the authors investigated the number of different resistant strains present in a newly diagnosed chronic phase CML patient. An optimal control approach was utilized to optimize imatinib scheduling in [7]. Particularly relevant to our work is [8, 9] where the authors investigated simultaneous continuous administration of dasatinib, nilotinib and imatinib; in particular, they explored the minimal number of drugs necessary to prevent drug resistance. In the current work, we focus on understanding the optimal administration schedule of multiple therapies to prevent resistance, and studying the impact of toxicity constraints on optimal scheduling. Since several of the available tyrosine kinase inhibitors (TKI) share similar toxicities (in particular neutropenia, see e.g., [10–12]) combining them together can lead to elevated risk of adverse events. Thus we consider sequential combination therapies in which only one TKI may be administered at a time. Moreover, it has been shown that the risk of treatment failure and blast crisis are highest within the first 2 years from diagnosis [1]. Therefore it is possible that optimized, sequential single agent therapy may be sufficient to minimize the risk of treatment failure. Allowing only one treatment at a time leads to a complex, time-dependent discrete optimization problem.
Another line of research closely related to the current work is the use of optimal control techniques in the design of optimal temporally continuous drug concentration profiles (see, e.g., review articles [13, 14] and the textbook [15]). In this field the tools of optimal control such as the Pontryagin principle and the Euler-Lagrange equations are used to find drug concentration profiles that result in minimal tumor cell populations under toxicity constraints. Particularly relevant to the current work is [16] where the authors searched for optimal anti-HIV treatment strategies. They dealt with the similar problem of treating heterogeneous populations with multiple drugs. One major drawback of these works is the fact that it is nearly impossible to achieve a specific optimal continuous drug-concentration profile in patients, since drug concentration over time is a combined result of a treatment schedule (e.g. sequence of discrete oral administrations) and pharmacokinetic processes in the body including metabolism, elimination, etc. Thus the clinical utility of an optimal continuous drug concentration profile is limited. In contrast to these previous works, here we model the optimization problem as a more clinically realistic sequence of monthly treatment decisions. Imposition of this fixed discrete set of decision times leads to a challenging optimization problem. Such dynamical systems are referred to as ‘switched nonlinear systems’ in the control community [17], and our problem additionally imposes fixed switching times. In this work we will leverage the system structure and tools from mixed-integer linear optimization [18] to solve this problem numerically, resulting in optimal therapy schedules that are easy to implement in practice.
We consider an ODE model of the differentiation hierarchy of hematopoietic cells, adapted from [5, 19, 20]. Stem cells (SC) on top of the hierarchy give rise to progenitor cells (PC), which produce differentiated cells (DC), which in turn produce terminally differentiated cells (TC). This differentiation hierarchy applies to both normal and leukemic cells [21]. We consider in our model leukemic WT cells as well as preexisting BCR-ABL mutant cell types. We use type 1, type 2, and type i (3 ≤ i ≤ n) cells to denote normal, leukemic WT, and (n − 2) leukemic mutant cells; layer 1, 2, 3, 4 cells to denote SC, PC, DC, and TC; and drug 0, 1, 2, 3 to denote a drug holiday, nilotinib, dasatinib, and imatinib, respectively. Let xl,i(t) denote the abundance of type i cell at layer l and time t, and x(t) = (xl,i(t)) be the vector of all cell abundance at time t. If drug j ∈ {0, 1, 2, 3} is taken from month m to month m + 1, then the cell dynamics are modeled by the following set of ODEs.
x ˙ ( t ) = f j ( x ( t ) ) , t ∈ [ m Δ t , ( m + 1 ) Δ t ] , x ( m Δ t ) = x m ,
for some function fj, where Δt = 30 days and xm is the cell abundance at the beginning of month m. The concrete form of function fj under drug j is described as follows.
SC level x ˙ 1 , i = ( b 1 , i j ϕ i - d 1 , i j ) x 1 , i , i = 1 , … , n PC level x ˙ 2 , i = b 2 , i j x 1 , i - d 2 , i j x 2 , i , i = 1 , … , n DC level x ˙ 3 , i = b 3 , i j x 2 , i - d 3 , i j x 3 , i , i = 1 , … , n TC level x ˙ 4 , i = b 4 , i j x 3 , i - d 4 , i j x 4 , i , i = 1 , … , n . (1)
See Fig 1 for an illustration of the differentiation hierarchy of hematopoietic cells, including neutrophils as part of the TC.
Here we describe the function of each parameter of this model. For a detailed discussion of how these parameters were estimated from biological data, please see Section A of S1 Text. Type i stem cells divide at rate b 1 , i j per day under drug j. The production rates of type i progenitors, differentiated cells and terminally differentiated cells under drug j are b l , i j per day for l = 2, 3, 4, respectively. The type i cell at layer l dies at rate d l , i j per day under drug j, for each i, l and j. The competition among normal and leukemic stem cells is modeled by the density dependence functions ϕi(t), where ϕ i ( t ) = 1 / ( 1 + p i ∑ i = 1 n x 1 , i ( t ) ) for each i; these functions ensure that the normal and leukemic stem cell abundances remain the same once the system reaches a steady state. The parameter p1 (resp. p2) is computed from the equilibrium abundance of normal (resp. leukemic WT) stem cells assuming only normal (resp. leukemic WT) cells are present, and we set pi = p2 for each i ≥ 3. In particular, p 1 = ( b 1 , 1 0 / d 1 , 1 0 - 1 ) / K 1 and p 2 = ( b 1 , 2 0 / d 1 , 2 0 - 1 ) / K 2, with K1 (resp. K2) being the equilibrium abundance of normal (resp. leukemic WT) stem cells assuming only normal (resp. leukemic WT) cells are present.
Assume the initial population of each cell type is known. Our goal is to select a treatment plan to minimize the tumor size at the end of the planning horizon. We call this the Optimal Treatment Plan problem (OTP). Each treatment plan is completely characterized by a temporal sequence of monthly treatment decisions over a long time horizon. Between each monthly treatment decision, the dosing regimen is identical from day to day. The standard regimens for each drug, which we will utilize throughout the work, are 300mg twice daily for nilotinib, 100mg once daily for dasatinib, and 400mg once daily for imatinib [22]. For example, let 1 denote nilotinib, 2 denote dasatinib, 3 denote imatinib, and 0 denote drug holiday. Then the sequence {1, 1, …, 1} represents that the patient takes the standard nilotinib regimen, 300mg twice daily, every day, every month. The sequence {2, 0, 2, 0, …} represents that the patient alternates between the standard dasatinib regimen, 100mg once daily, and a drug holiday on alternate months.
We introduce the binary decision variables zm,j to indicate whether drug j is taken in month m or not, for each j = 0, 1, 2, 3 and m = 0, 1, …, M − 1, in an M-month treatment plan. An assignment of values to all zm,j variables that satisfy all constraints in the optimization model gives a feasible treatment plan.
Note the total leukemic cell abundance at day t is given by ∑l≥1∑i≥2 xl,i(t). The OTP can be formulated as the following mixed-integer optimization problem with ODE constraints.
min ∑ l ≥ 1 ∑ i ≥ 2 x l , i ( M Δ t ) (2) s . t . x ˙ ( t ) = ∑ j = 0 3 z m , j f j ( x ( t ) ) , t ∈ [ m Δ t , ( m + 1 ) Δ t ] , m = 0 , 1 , … , M - 1 , (3) ∑ j = 0 3 z m , j = 1 , m = 0 , 1 , … , M - 1 , (4) z m , j ∈ { 0 , 1 } , j = 0 , 1 , 2 , 3 , m = 0 , 1 , … , M - 1 , (5) x ( 0 ) = x 0 . (6)
To summarize the previous display, in eq (2) we state that our objective is to minimize the leukemic cell population at the end of the treatment horizon. In eq (3) we stipulate that the cell dynamics are governed by the system of differential equations given by eq (1). Together eqs (4) and (5) stipulate that during each time period we administer either one drug or no drug.
The OTP problem is a mixed-integer nonlinear optimization problem, in which some constraints are specified by the solution to a nonlinear system of ODEs eq (3). This optimization problem is beyond the ability of state-of-the-art optimization software. However, if we assume the TKI therapies do not affect the stem cell compartment, then it is possible to handle the ODE constraints numerically. This is because the non-linearities in the ODE model are only present in the stem cell compartment, and the remaining compartments are modeled by linear differential equations. Thus we are able to build a refined linear approximation to the ODE constraints (see Section C of S1 Text), and recast the problem as a mixed-integer linear optimization problem, which can be solved efficiently for the size of our problem (see Section B of S1 Text).
It should be noted that our model does not explicitly consider the phenomena of TKI resistance acquired during therapy. Our model suggests a method for designing optimized treatment plans, beyond monotherapies, at the beginning of treatment. In addition, this optimization procedure can be re-run and modified during the course of treatment, with updated inputs from each patient’s response to the treatment—including the presence of acquired mutations.
Below we summarize our notation for the ease of the reader.
In this work we consider the dynamics of CML response to single-agent and combination schedules utilizing the standard therapies imatinib, dasatinib and nilotinib.
We first utilize the model to demonstrate the dynamics of CML populations with preexisting BCR-ABL mutations under monotherapy with the standard therapies imatinib, dasatinib and nilotinib. Recall that the standard dosing regimens are 300mg twice daily for nilotinib, 100mg once daily for dasatinib, and 400mg once daily for imatinib [22]. The birth rate parameters for each leukemic mutant type under different TKIs (b l , i j for j = 1, 2, 3, l = 1, 2, 3, 4, and i ≥ 3) in the model are estimated using in vitro IC50 values reported in [23] for each drug. The initial cell populations at the start of therapy are derived by running the model starting from clonal expansion of a single leukemic cell in a healthy hematopoietic system at equilibrium [19] until CML detection (when the total leukemic burden reaches approximately 1012 cells [24]). At this point the total cell burden is 2–3 times the normal cell burden in a healthy individual and thus the total leukemic cells make up approximately 77% of the total cell population; this is consistent with clinical reports [25]. Details on deriving the initial cell abundances at diagnosis are provided in Section A of S1 Text.
In the first example we consider a patient harboring a low level of the BCR-ABL mutant F317L before the initiation of TKI therapy. According to the in vitro IC50 value reported in [23], F317 is resistant to dasatinib, and moderately resistant to nilotinib and imatinib. The initial population conditions are given in Table 1 with the leukemic WT and F317L cells taking up 95% and 5% of the leukemic cells, respectively.
We plot in Fig 2 the cell dynamics over 120 months for four treatment plans: (1) nilotinib monotherapy (2) dasatinib monotherapy, (3) imatinib monotherapy, (4) no therapy—control. We observe that as predicted, the disease burden responds well to imatinib and nilotinib; the percentage of cancerous cells after a 24 month treatment drops to 0.19% with nilotinib and 0.26% with imatinib, respectively. However, the F317L mutant population is fairly resistant to dasatinib; we observe that the percentage of cancerous cells after 24 months is 58.1% with dasatinib and 95.4% with no treatment. Over the 120 month period dasatinib treatment provides only modest improvement over the ‘no drug’ option in controlling the F317L population; however, dasatinib remains quite effective in controlling the WT leukemic population. It is interesting to note that overall, nilotinib is the most effective in controlling both the WT and F317L leukemic populations. However, nilotinib also negatively impacts the healthy cell population more severely than imatinib, which is slightly less effective in controlling the leukemic populations. This suggests that some trade-offs between these drugs exist, and these trade-offs may be exploited in designing combination therapies.
In the next example we consider a patient with BCR-ABL mutant type M351T preexisting therapy. In contrast to the previous example, this commonly occurring mutant has been found to be partially sensitive in varying degrees to all three therapies. The initial conditions are given in Table 2. Once again we have assumed that WT and M351T cells take up 95% and 5% of total leukemic cells, respectively.
In Fig 3 the cell dynamics over 120 months for the four standard treatment plans are plotted: (1) nilotinib monotherapy (2) dasatinib monotherapy, (3) imatinib monotherapy, (4) no therapy—control. Since the M351T mutant is responsive to each drug in contrast to the previous example, the percentage of cancerous cells after a 24 month treatment drops to 0.18% with nilotinib, 0.18% with dasatinib, and 0.25% with imatinib, respectively. Without treatment, the percentage of cancerous cells after 24 months is 95.4%. Here, we observe that although nilotinib is more effective than dasatinib in controlling the total mutant M351T burden, the effect is reversed in the progenitor population. Higher levels of stem and progenitor populations will lead to faster rebound during treatment breaks, suggesting another trade-off to consider in the combination setting.
We next solve the discrete optimization problem to identify sequential combination therapies utilizing imatinib, dasatinib and nilotinib to optimally treat CML patients with preexisting BCR-ABL mutations. We consider schedules in which a monthly treatment decision is made between one of four choices: imatinib, dasatinib, nilotinib, and drug holiday. During months in which one of the three drugs is administered, the dosing regimen is fixed at 300mg twice daily for nilotinib, 100mg once daily for dasatinib, and 400mg once daily for imatinib. In the following we optimize over feasible treatment decision sequences that result in a minimal leukemic cell burden after 3 years. Each treatment plan is completely characterized by a temporal sequence of drugs over a long time horizon.
The extent to which TKIs affect leukemic stem cells is currently unknown. Several studies have demonstrated that these cells may in fact be resistant to TKIs, see e.g. [27] and references therein. Based on these results we have assumed throughout this work that TKIs do not affect the leukemic stem cell population. Given that there is uncertainty regarding this issue, in this subsection we investigate how our optimal schedules will perform if we assume TKIs impact the leukemic stem cells. In Tables 9 and 10, we look at the leukemic cell burden after a 36-month treatment with our optimal therapy and the best monotherapy for preexisting M351T and F317L, respectively, assuming that each TKI reduces the production rate of leukemic stem cells by a certain level. In both tables we observe that the optimized schedule outperforms the monotherapy even if leukemic stem cells are susceptible to TKI. These result imply that the optimal treatment schedules we derive here outperform a traditional monotherapy regardless of the impact of TKIs on leukemic stem cells.
In this work we have considered the problem of finding optimal treatment schedules for the administration of a variety of TKIs for treating chronic phase CML. We modeled the evolution of wild-type and mutant leukemic cell populations with a system of ordinary differential equations. We then formulated an optimization problem to find the sequence of TKIs that lead to a minimal cancerous cell population at the end of a fixed time horizon of 36 months. The 36-month therapeutic horizon is clinically meaningful since it appears that the risk of therapeutic failure and disease progression to blast crisis is highest within the first two years from diagnosis [1].
At first glance the optimization problem studied in this work (OTP) is quite challenging. It is a mixed-integer nonlinear optimization problem, where the nonlinear constraints are specified by the solution to a nonlinear system of differential equations. However, one factor mitigating the complexity of the problem is the assumption that the TKIs do not effect the stem cell compartment. This has the effect of making the evolution of the stem cell compartment independent of the TKI schedule chosen. In addition, the remaining layers in the cellular hierarchy are modeled by linear differential equations. We can thus numerically solve the differential equation governing the stem cell layer, and treat this function as an inhomogeneous forcing term in the linear differential equation governing the progenitor cells. This allows us to approximate the nonlinear constraints specified by the differential equations by linear constraints with high accuracy. Then the OTP problem can be approximated by a mixed-integer linear optimization problem, which we are able to solve with state-of-the-art optimization software CPLEX [28] within one hour.
Importance of minimizing progenitor cell population. We first aimed to minimize leukemic cell burden at 36 months after initiation of therapy, starting with an initial leukemic population of wild-type CML cells and either M351T (sensitive to all three therapies) or F317L (resistant to dasatinib) mutant leukemic cells. For both starting mutant populations, we observed that the optimal schedule involves initiating therapy with dasatinib and later switching to nilotinib, although the timing of the switch differed. To further understand this result, we noted that within this parameter regime, dasatinib is the most effective of the three TKIs at controlling leukemic progenitor cells, while nilotinib is the most effective at controlling the differentiated cells, which comprise most of the total leukemic burden. Thus, we note that controlling the leukemic progenitor cell population is important in long-term treatment outcome. This is further supported by the observation that blast crisis emerges due the acquisition of additional mutations in CML progenitor cells (not stem cells) [3]. Our approach suggests that using combination TKI therapies may be a viable method of controlling this population. Our modeling suggests that it is best to reduce the progenitor cells early and then reduce the differentiated cells towards the end of the treatment planning horizon. An early reduction in progenitor cells pays off in later stages of the treatment planning horizon, since a small progenitor cell population results in a lower growth rate for differentiated cells which leads to a greater response to subsequent TKI therapy.
Effects of toxicity constraints. We also imposed a toxicity constraint on therapy optimization procedure by mandating that patient ANC levels stay above a given threshold that reduces the risk of infections. We observed that incorporating this toxicity constraint does impact the structure of the optimal schedules significantly, resulting in mandated treatment breaks as well as switching some months to imatinib therapy, which has a lower toxicity effect. We also noted that the choice of treatment breaks occurring also in one-month intervals may result in dangerous rebound of leukemic burden to levels close to pre-treatment, suggesting that shorter breaks to combat toxicity may be recommended. Although the model we have used for describing the dynamics of the ANC levels is simple, our findings demonstrate that incorporating a mechanistically modeled toxicity constraint into optimization of therapy scheduling is both feasible and important in determining optimal scheduling.
Multiple preexisting mutant types. While some previous studies have suggested that the majority of CML patients are diagnosed with 0 or 1 preexisting BCR-ABL mutations, some patients do harbor multiple mutants at the start of therapy [2, 6]. Thus we also investigated the impact of having 2 mutant types present (M351T and F317, or E255K and F317L) at the start of therapy, on optimal schedules. We observed the number and specific combination of preexisting mutants present can significantly impact the optimization results. This points to the importance of determining which BCR-ABL mutations preexist in patients at diagnosis, before treatment planning is done.
Throughout this work we have observed that the structure of the optimal therapy depends heavily on model parameters, e.g., cellular growth rates and ANC decay rates. It is likely that each individual patients will have unique model parameters, and therefore a unique best schedule. An exciting application of this work would be the development of personalized optimal therapeutic schedules. Determination of (i) the mutant types (if any) present in a patient’s leukemic cell population, (ii) growth kinetics of their leukemic cell populations, and (iii) patient ANC level responses under various TKIs, would enable our optimization framework to build treatment schedules in a patient-specific setting. At the start of treatment for acute lymphoblastic leukemia (ALL), Quantitative RT-PCR or similar techniques are sometimes used to perform mutational analysis to identify the preexisting mutant types. Indeed, studies have demonstrated that BCR-ABL mutants are present at the time of diagnosis in many ALL patients, and as sequencing technologies improve, smaller and smaller subclones with resistant phenotypes will likely be discovered [29]. In addition, the ANC level can be tested weekly by a routine complete blood count and in principle, the growth kinetics of each leukemic cell type could be analyzed in the laboratory using standard techniques such as flow cytometry or quantitative imaging. However, very few laboratories currently have the ability to analyze the growth kinetics of each type of leukemic cell population in a reproducible way. We believe that the testing procedure needs to be standardized before it becomes helpful for the treatment of CML. If such a standardized growth kinetic analysis can be realized, during treatment a patient’s response to various TKIs could be monitored so that the impact of TKI on growth kinetics of leukemic cell population and ANC levels could be quantified. Then the optimization procedure could be re-run on the fly with updated patient parameters, providing dynamic feedback into each patient’s optimal therapy schedule.
|
10.1371/journal.pntd.0002244 | International External Quality Assessment of Molecular Detection of Rift Valley Fever Virus | Rift Valley fever (RVF) is a viral zoonosis that primarily affects animals resulting in considerable economic losses due to death and abortions among infected livestock. RVF also affects humans with clinical symptoms ranging from an influenza-like illness to a hemorrhagic fever. Over the past years, RVF virus (RVFV) has caused severe outbreaks in livestock and humans throughout Africa and regions of the world previously regarded as free of the virus. This situation prompts the need to evaluate the diagnostic capacity and performance of laboratories worldwide. Diagnostic methods for RVFV detection include virus isolation, antigen and antibody detection methods, and nucleic acid amplification techniques. Molecular methods such as reverse-transcriptase polymerase chain reaction and other newly developed techniques allow for a rapid and accurate detection of RVFV. This study aims to assess the efficiency and accurateness of RVFV molecular diagnostic methods used by expert laboratories worldwide. Thirty expert laboratories from 16 countries received a panel of 14 samples which included RVFV preparations representing several genetic lineages, a specificity control and negative controls. In this study we present the results of the first international external quality assessment (EQA) for the molecular diagnosis of RVF. Optimal results were reported by 64% of the analyses, 21% of the analyses achieved acceptable results and 15% of the results revealed that there is need for improvement. Evenly good performances were achieved by specific protocols which can therefore be recommended as an accurate molecular protocol for the diagnosis of RVF. Other protocols showed uneven performances revealing the need for improved optimization and standardization of these protocols.
| Rift Valley fever (RVF) is a zoonotic viral disease posing an increasing threat to animals and humans worldwide. Recent severe outbreaks of the disease in animal and human populations in endemic regions and outside the disease's traditional geographic boundaries necessitate the need for evaluating the diagnostic performance of RVF expert laboratories. Molecular methods are increasingly used for a rapid and accurate detection of viral nucleic acid. In this study we present the results of the first international external quality assessment (EQA) for the molecular diagnosis of RVF. Such EQA studies allow participating laboratories to monitor the quality and identify possible weaknesses of current diagnostic methods. Participants to this RVF EQA were 30 expert laboratories from 16 different countries worldwide. The study demonstrated that optimal results could be achieved by the majority of laboratories. Specific protocols showed evenly good performances and can therefore be recommended to all expert laboratories. However, other methods showed uneven performances suggesting the need for improved optimization and standardization of these protocols.
| Rift Valley fever (RVF) is a mosquito-borne viral zoonosis that primarily affects animals but also has the capacity to infect humans. An epizootic of RVF is usually first indicated by a wave of unexplained abortions as infected pregnant livestock abort virtually 100% of fetuses. The disease is less fatal to humans as most human infections are asymptomatic and when clinical symptoms appear they are in majority influenza-like. Nevertheless, some cases may develop a severe RVF disease with variable clinical signs. More severe cases occur in 2% of the RVF cases and fall into three categories: liver necrosis with hemorrhaging, retinitis with visual impairment and meningoencephalitis [1], [2].
The causative agent of RVF, the RVF virus (RVFV), is a negative-stranded RNA virus, a member of the genus Phlebovirus of the Bunyaviridae family. The number of identified viral lineages of RVFV has increased from 3 in an early analysis [3] to 7 in a 2007 study [4], and in the most recent report 15 distinct genetic groups were reported [5]. Phylogenetic analysis shows that the virus emerged in the mid-19th century, but it was first identified in 1930 during an outbreak of abortions and deaths among sheep in the Rift Valley region of Kenya. In 1977–78, several millions of people were infected and more than 600 died during a severe epidemic in Egypt [6]. Since then, the geographical distribution of the virus has widely spread and now includes most countries of the African continent as well as Madagascar and the Arabian Peninsula. During the past five years, outbreaks have been reported in Kenya [7], Somalia, Tanzania [8], Sudan [9], Mayotte [10], Madagascar [11], Swaziland, South Africa and Mauritania [12], [13] Another important concern is the increasing number of human fatalities during the most recent outbreaks [14].
The emergence or re-emergence of RVFV activity is periodic and associated with exceptionally heavy rainfalls which allow massive breeding of flood-water Aedes mosquitoes with the capacity for transovarial transmission [15] and other competent vectors such as Anopheles and Culex species [9]. These mosquitoes may initiate outbreaks among livestock, particularly breeds of cattle and sheep. The virus can be transmitted to humans by mosquito bite or by contact with infected tissues of domestic and wildlife ruminants. The sudden onset of large numbers of abortions and fatalities in RVFV affected livestock, resulting in the virus spread to humans can greatly strain public health and veterinary infrastructures.
Unavailability of effective antiviral drugs and commercial vaccines for human or animal use outside endemic countries, including the US and Europe, and the recent spread of RVFV beyond its usual boundaries has resulted in increased international demand for qualified diagnostic tools for a rapid and accurate diagnosis of RVF.
Diagnostic methods for RVFV detection include virus isolation [16], antigen [17], [18] and antibody detection methods [19]–[21] and nucleic acid amplification techniques. Isolation procedures are expensive, time-consuming and require high biocontainment facilities. Serological methods such as antigen or antibody-detection enzyme immunoassays (EIA) require several samples and often lack sensitivity. Therefore, considerable efforts have been made to develop molecular methods which allow a rapid, accessible and accurate detection of RVFV. The use of direct diagnostic methods such as molecular methods, can detect the disease during the acute phase of the infection thus allowing efficient patient management, avoiding nosocomial cases and providing rapid outbreak response. Highly sensitive nucleic acid detection methods have been developed including polymerase chain reaction (PCR) assays such as reverse-transcriptase PCR (RT-PCR) [22], real-time RT-PCR (qRT-PCR) [23]–[25] and more recently real-time reverse-transcription loop-mediated isothermal amplification (RT-LAMP) [26] and recombinase polymerase amplification assays (RPA) [27].
The performance of the different techniques applied for molecular diagnosis of RVFV may vary between laboratories. External quality assessment (EQA) studies to assess the quality of RVFV molecular diagnostics have not been performed until now. The EQA study allows the participating laboratories to monitor the quality of current diagnosis, identify possible weaknesses of particular diagnostic methods and evaluate their capacity for surveillance activities. Therefore the first EQA study for the molecular diagnosis of RVFV was organized by the European Network for Diagnostics of ‘Imported’ Viral Diseases (ENIVD) (http://www.enivd.org) in 2012. Using the results of this study, the ENIVD can also provide support and advice to all laboratories performing RVFV molecular diagnosis.
A total of 33 laboratories involved in diagnostics of RVF infections were invited to participate in this study. Invitees were selected from the register of ENIVD members, national/regional reference laboratories for RVF or vector-borne diseases as well as on the basis of their contributions to the literature relevant to this topic. The participation to the study was open and free of charge and included publication of the results in a comparative and anonymous manner. This EQA was coordinated by the ENIVD following comparable procedures used during previous studies performed by the network [28], [29].
A proficiency test panel of 14 samples was prepared which included inactivated and stable RVFV preparations generated from Vero E6 cell culture supernatants of different RVFV genetic lineages and origin. Viral cell supernatants were inactivated by heating for 1 h at 60°C and gamma irradiation (25 kilogray) to assure their non-infectivity. A serum sample spiked with Toscana virus, another phlebovirus, was included as a specificity control as well as two negative controls. The RVFV positive samples selected for this EQA panel are detailed in Table 1. Two dilutions of sample Tambul/Egypt/1994 and 5 dilutions of sample F057/Kenya/2007 were obtained by serial 10-fold dilutions and included in the panel for sensitivity testing.
All virus material used for the preparation of the EQA panel was obtained from cell culture and not from clinical samples of infected patients. Therefore, there is no requirement for any ethical statement in this study.
All samples were diluted with fresh thawed human plasma previously confirmed as negative for RVFV. Aliquots of 100 µl were number-coded, freeze dried for 24 h (Christ, AlphaI-5, Hanau, Germany) and stored at 4°C until dispatched.
Before dispatching the panels, 3 different sets of EQA samples were tested and validated by 2 expert laboratories. For validation, the samples were resuspended in 100 µl of water and the RNA extracted using the QIAamp viral RNA minikit (Qiagen, Hilden, Germany). The number of RVFV genome copies present in these samples was determined by qRT-PCR.
Panel samples were shipped by regular post at ambient temperature. We requested participant laboratories to resuspend the samples in 100 µl of water and to analyze the material as serum samples for nucleic acid detection of RVFV following their routine protocols. The EQA panels were distributed to participants with documentation including full instructions and an evaluation form to fill in their results. Participants were also asked to report information on the adopted protocol, the type of RVFV strain and the number of genome copies in each sample when possible as well as any problems encountered concerning the shipment or the packaging of the samples.
To guarantee anonymous participation, an individual numerical identification code was assigned to the results reported by each laboratory. This number was followed by a letter (a, b, c) when distinguishable data sets of results based on different methods were sent.
The results were scored in reflection of analytical sensitivity and specificity as in previous EQA studies performed by the ENIVD [29], [30]. We assigned one point for correct positive or negative result whereas false-negative/-positive results were not scored. Equivocal or borderline results were not counted as molecular diagnostic methods should always provide a clear positive or negative result.
Results were classified as:
We obtained from the invitees a response rate of 91% representing a total of 30 participating laboratories from 16 different countries (10 European, 2 African, 3 Middle-Eastern/Asian countries and one American country):
CODA-CERVA, Department of Virology, Epizootic Diseases Section, Uccle, Belgium; ANSES, Virology Unit, Laboratory of Lyon, France; CIRAD, Department BIOS «Control of exotic and emerging diseases», Montpellier, France; IRBA-IMTSSA, Virology Unit, Le Pharo, Marseille, France; BNI, National Reference Centre for Tropical Infectious Diseases, Hamburg, Germany; Bundeswehr Institute of Microbiology, Munich, Germany; Institute for Novel and Emerging Infectious Diseases Friedrich-Loeffler-Institut, Germany; Robert Koch Institute, Berlin, Germany; Institute of Virology, Georg-August University, Gottingen, Germany; Central Virology Laboratory, Ministry of Health, Public Health Laboratories Sheba Medical Center, Israel; Army Medical and Veterinary Research Center, Rome, Italy; Department of Infectious, Parasitic and Immune-Mediated Diseases, Istituto Superiore di Sanità, Rome, Italy; Padiglione Baglivi National Institute for Infectious Diseases “L. Spallanzani”, Rome, Italy; Department of Histology, Microbiology and Medical Biotechnologies, University of Padova, Italy; Center for Vectors and Infectious Diseases Research, National Institute of Health, Aguas de Moura, Portugal; King Fahd Medical Research Center, King Abdulaziz University, Saudi Arabia; Arboviruses and viral hemorrhagic fever Unit, Institut Pasteur de Dakar, Senegal; Defense Medical & Environmental Research Institute, DSO National Laboratories, Singapore; Institute of Microbiology and Immunology, Faculty of Medicine, University of Ljubljana, Slovenia; Onderstepoort Veterinary Institute, South Africa; Deltamune (Pty) Ltd, Centurion, Gauteng, South Africa; Special Viral Pathogens Laboratory, National Institute for Communicable Diseases, South Africa; Laboratory of Arboviruses and Imported Viral Diseases, National Center for Microbiology, Instituto de Salud Carlos III, Spain; National Institute for Agricultural Research and Experimentation (INIA), Madrid, Spain; Viral Diseases Unit, CReSA, Barcelona, Spain; Swedish Institute for Infectious Disease Control, Sweden; Virology group, Spiez Laboratory, Switzerland; Laboratory of Virology, University Hospitals of Geneva, Switzerland; WHO Collaborative Centre for Virus Reference and Research (Arboviruses & VHFs), Health Protection Agency, United Kingdom; Viral Special Pathogens Branch, Infectious Diseases, CDC, Atlanta, United States of America.
A total of 39 datasets were received including 5 double sets from laboratories using 2 methods (lab #6, 7, 21, 27 and 28) and 2 triple sets from lab #5 and #14. Methods used by the same laboratory could differ from the type of technique, the protocol used for a specific technique or the type of instrument used for a specific protocol.
Performances varied among laboratories and scores ranged from 7 to the maximum value of 14. Optimal results were reported by 64% (n = 25) of the analyses; 21% (n = 8) of the analyses achieved acceptable results due to the inability to detect one positive sample, and 15% (n = 6) revealed several false negative and/or one or more false positive results indicating that there is still need for improvement (Table 2 and 3).
RVF reference laboratories responded keenly to this EQA study (91% response rate), including laboratories situated in RVFV endemic countries such as South Africa and Saudi Arabia. Nonetheless, there is still a need to encourage more laboratories situated in RVF-endemic areas to participate in quality assurance programs. In fact, the increasing amplitude of this disease in Africa necessitates the rapid recognition of RVF outbreaks and implementation of effective control measures in order to prevent uncontrolled and wider spread of the virus.
Most of the laboratories (93%, 28 out of 30) reported the use of qRT-PCR techniques allowing a rapid detection as well as quantification of the virus genome. This confirms that the use of qRT-PCR has remarkably expanded although it requires expensive equipment. All datasets obtained by qRT-PCR only were scored with 13 or 14 points indicating an evenly high performance of all qRT-PCR procedures performed by the different laboratories.
Protocols from Drosten et al, 2002 [24], Bird et al. 2007 [25], Weidmann et al 2008 [31] as well as all in-house qRT-PCR protocols (dataset #6b, #10 and #24) have demonstrated the capacity of providing optimal performances indicating a good specificity and sensitivity for these techniques. The sets of results obtained by applying the qRT-PCR protocols of Mweango et al. 2012 [35], Garcia et al. 2001 [23] and Busquets et al. 2010 [32] did not achieve optimal performances (scores 13, 11 and 13 respectively) but these techniques are not sufficiently represented to conclude on their overall performances.
Information on the viral load of RVFV in human samples can be very useful to monitor the progress of clinical manifestations and to study the pathogenesis of RVFV. Interestingly, not all laboratories employing qRT-PCR techniques have reported quantified results and most of them (64%) reported the results as cycle thereshold (Ct) values and not the number of genome copies. This indicates that most laboratories do not resort to RVFV standards while performing qRT-PCR although such standards would allow them to quantify viral genome in each sample without performing any additional assay. Accordingly to the results of this EQA as well as previous EQA studies, there is still room for improvement concerning viral load determination [29], [30].
The most widely used technique after qRT-PCR was nested RT-PCR with 5 laboratories which referred to 2 different protocols [22], [34]. Nested RT-PCR performances varied greatly compared to qRT-PCR with scores ranging from 7 to 13 thus never reaching optimal performances. The dataset #14c obtained a score of 13 with the protocol of Sall et al. 2002 [22] because it could not detect the highest dilution of the RVFV-Egypt/1994 strain indicating a slightly low sensitivity just as observed for some of the qRT-PCR methods. Nevertheless other datasets referring to nested RT-PCR (#9 and #22) also reported false positive results indicating a lack of specificity of these procedures with both nested RT-PCR protocols [22], .
It is interesting to notice the appearance of newly developed techniques which are suitable for rapid field diagnostics such as RT-LAMP developed in 2009 [26] and RPA technology developed in 2012 [27]. No general conclusion can be achieved concerning the performances of these two techniques as they both have been performed by only one laboratory. However RPA has shown optimal results for this EQA demonstrating equivalent sensitivity and specificity to the qRT-PCR techniques (dataset #27b).
On the other hand, RT-LAMP results indicated difficulties in detecting RVFV genome in the less concentrated samples of the panel (sample #4, #13 and #14). These results suggest some limitations in test sensitivity. However, very high test sensitivity is not essential for field diagnostics in an outbreak situation where most diagnosed patients are in the acute phase of the disease and are expected to present a high viremia.
Three laboratories have provided different sets of results which referred to the same technique and protocol but using different instruments (datasets #5b/c, #14a/b and #28 a/b). These datasets provided all optimal results by using two different instruments except for dataset #14 which reported a slightly lower sensitivity using the SmartCycler System from Cepheid (#14b, 13 points) compared to the 7500 Real-Time PCR System from Applied Biosystems (#14a, 14 points). However, this difference cannot be attributed with certainty to the use of a different instrument as result variability can also arise from a lack of repeatability of the procedure.
Only a few participants provided complete or partial information regarding strain typing (13%, 4 out of 30). However, correct results without strain or genetic lineage specification are satisfactory in the context of laboratory diagnosis. Nonetheless, RVFV strain typing is relevant for surveillance activities in order to monitor which strains are circulating in RVFV-endemic areas and what type of clinical manifestations are associated with these strains.
Comparing the results of this EQA panel to previous EQA studies [29], [30], [36], we observe a higher concordance in terms of performance within laboratories using the same type of diagnostic method. In fact, all qRT-PCR techniques demonstrated an overall good performance with scores ranging from 13 to 14. On the other hand, nested RT-PCR methods have shown a common need for improvement in terms of test sensitivity and/or specificity.
Nevertheless, variations in performance between laboratories using the same method were noted. The reason for such variations is difficult to establish but can be minimized by standardizing procedures, including controls and testing conditions.
In order to ensure optimal performances for RVFV molecular diagnosis in expert laboratories, we recommend conducting EQA studies on a regular basis. Future EQA studies should include a wide range of RVFV isolates with limiting concentrations to assess as precisely as possible the diagnostic performances of various molecular protocols in different reference laboratories.
|
10.1371/journal.pgen.1004096 | New MicroRNAs in Drosophila—Birth, Death and Cycles of Adaptive Evolution | The origin and evolution of new microRNAs (miRNAs) is important because they can impact the transcriptome broadly. As miRNAs can potentially emerge constantly and rapidly, their rates of birth and evolution have been extensively debated. However, most new miRNAs identified appear not to be biologically significant. After an extensive search, we identified 12 new miRNAs that emerged de novo in Drosophila melanogaster in the last 4 million years (Myrs) and have been evolving adaptively. Unexpectedly, even though they are adaptively evolving at birth, more than 94% of such new miRNAs disappear over time. They provide selective advantages, but only for a transient evolutionary period. After 30 Myrs, all surviving miRNAs make the transition from the adaptive phase of rapid evolution to the conservative phase of slow evolution, apparently becoming integrated into the transcriptional network. During this transition, the expression shifts from being tissue-specific, predominantly in testes and larval brain/gonads/imaginal discs, to a broader distribution in many other tissues. Interestingly, a measurable fraction (20–30%) of these conservatively evolving miRNAs experience “evolutionary rejuvenation” and begin to evolve rapidly again. These rejuvenated miRNAs then start another cycle of adaptive – conservative evolution. In conclusion, the selective advantages driving evolution of miRNAs are themselves evolving, and sometimes changing direction, which highlights the regulatory roles of miRNAs.
| During Metazoan evolution, the architecture of the genome changed dramatically in size, gene number and regulatory elements. Genomic architecture is often assumed to be correlated with morphological complexity. However, it is still not known whether the gene repertoire, both for protein coding and non-coding genes, is continually increasing. In the last decade, a large family of small non-coding RNAs, or microRNAs (miRNAs), has been shown to play an important role in diverse developmental processes. The genes controlled by miRNAs often evolve rapidly, potentially contributing to functional novelty, diversity and speciation. Here we estimated the birth and death rate of new adaptive miRNAs in Drosophila melanogaster. We found most new adaptive miRNAs disappear over long periods of time; hence, the miRNA repertoire stays close to that of a steady state. This steady state is commensurate with the morphological constancy of the genus of Drosophila.
| MicroRNAs (miRNAs) are a class of small, endogenous RNAs that regulate gene expression post-transcriptionally [1], [2]. Each miRNA gene is first transcribed as a stem-loop (hairpin) RNA structure, 70–90 nt in length in animals, and then processed in several steps into the ∼22-nt mature product, referred to as miR [3]. In animals, miR binds to the 3′ untranslated region (UTR) of target mRNAs through perfect base-pairing of the seed region (position 2–8 of a miR), inducing translation repression or mRNA degradation [4]. As the seed is only 7 nt long, each miRNA may potentially regulate hundreds of transcripts while each transcript may in turn be regulated by more than one miRNA [5].
The emergence of new miRNAs is of special interest in evolutionary biology for two reasons. First, they buffer gene expression noises and thus have been hypothesized to be a key player in canalization [6], [7]. As proposed by C. H. Waddington [8], [9], canalization contributes to developmental stability and, in a recent interpretation, it may also contribute to evolvability via hidden genetic variations [10], [11]. Second, due to their small size, miR-producing hairpins can form readily and de novo emergence of miRNAs from non-miRNA transcripts is a frequent phenomenon [12], [13]. There are hundreds of thousands of potential miRNA structures in each Drosophila genome [12] and millions in a mammalian genome [14]. Given such a propensity for new miRNAs to emerge, the birth, death and adaptation of new miRNAs are a significant part of understanding the evolution of transcriptional regulation [12]. In contrast, protein-coding genes require long open reading frames to yield functional peptides. Hence, local duplication or retrotransposition [15], rather than de novo origination, is the common mode for the formation of coding genes.
In Drosophila, the birth and death rates of miRNAs have been estimated to be about 12 and 11.7 genes per Myr, respectively, with a net gain of about 0.3 per Myr [12]. It is generally agreed that the net gain is low, ranging between 0.3 and 1 new gene per Myr [16], [17]. Despite this, the total repertoire of miRNAs should still be increasing dramatically over long periods. While the net gain (birth – death) is not in dispute, there is disagreement over the estimated birth and death rates of new miRNAs [12], [16], [17]. Because numerous putative miRNAs are found in the transcriptome, these lowly expressed, evolutionarily neutral, and short-lived miRNAs account for the bulk of the estimated births and deaths. The debate is about which ones should be counted as new miRNAs.
To resolve the issue, we propose to define new miRNAs in an evolutionary context by a set of stringent criteria, requiring a signature of initial adaptive evolution soon after their birth. Numerous small RNAs that emerge and vanish with the dynamics of neutral sequences are excluded from the evolutionary analysis. Given this definition, only a small fraction of miRNA-like sequences in any species would qualify as new miRNAs. We collected extensive small RNA-seq data available for four Drosophila species (D. melanogaster, D. simulans, D. pseudoobscura and D. virilis) [12], [16], [18]–[23] and three mosquitoes (Aedes albopictus, Aedes aegypti and Culex quinquefasiatus) [24], [25]. We further generated small RNA-seq data for sex organs and imaginal discs in D. simulans and D. pseudoobscura. The extensive dataset permits systematic identification of new miRNAs and in-depth analyses of their long-term fates.
Our first objective is to understand the origin and early evolution of new miRNAs in the species D. melanogaster. The second objective is to track the long-term evolutionary trajectory of new miRNAs, which may be in any of the following four modes after their initial adaptive evolution:
From the D. melanogaster miRNA repository (miRBase Release 19.0, Ref. [26]), 238 miRNA genes, including 204 canonical miRNAs and 34 mirtrons, were evaluated for their expression levels by examining small RNA sequencing data from different tissues and developmental stages (Ref. [12], [16], [18]–[21], [23], see Table S1 and Materials and Methods). The phylogenetic distributions of the 238 miRNA genes in Drosophila (D. melanogaster, D. simulans, D. pseudoobscura and D. virilis), with mosquitoes (Aedes albopictus, A. aegypti and Culex quinquefasiatus) as the outgroup, were determined from the available small RNA libraries (Ref. [12], [16], [22], [24], [25], see Table S1 and Materials and Methods). In addition, we sequenced five additional libraries from D. simulans and D.pseudoobscura to ensure that all Drosophila species in this survey included samples from testes and ovaries. Genes represented by more than 200 reads per million (RPM) in at least one library were designated “highly expressed” (Table S2). The rest were denoted as “lowly expressed” miRNAs.
The 204 canonical miRNAs and 34 mirtrons have very different patterns in age and expression level. Table 1 shows the emergence time of each miRNA, which falls in the interval of 0–4, 4–30, 30–60, 60–250, and >250 Myrs before present as depicted in Fig. 1. More than half of the highly expressed, canonical miRNAs (71 out of 136) came from the oldest age group (>250 Myrs) but none of the mirtrons were from that group (Table 1), suggesting mirtrons contribute very little to miRNA repertoire over long periods of time. The result is consistent with previous findings that mirtrons have different evolutionary trajectories from canonical miRNAs [16]. The majority of the lowly expressed genes, both canonical miRNAs (60/68) and mirtrons (21/25), came from the young age group of 0–4 Myrs (Table 1), corroborating that lowly expressed miRNA genes are likely to be evolutionarily transient [12].
In this study, we will focus on the 136 highly expressed canonical miRNAs because, with respect to long-term evolution, they are the most significant class among the four categories of Table 1.
Starting with the youngest genes, we first analyzed the 22 new miRNA genes that emerged in the last 30 Myrs, since D. melanogaster diverged from D. pseudoobscura (Fig. 1). Among them, 21 originated de novo; only miR-983-2 in D. melanogaster (dme-miR-983-2) was duplicated from another miRNA (dme-miR-983-1). More than half of the 22 new miRNAs are found in clusters – five in the miR-972 cluster (abridged as miR-972s), two in miR-310s and five in miR-982s. Members in a cluster have significantly higher expression levels than the orphan miRNAs (Mann-Whitney U test, p<0.05). The miR-982 cluster consists only of members emerging in the last 30 Myrs, whereas both miR-310s and miR-972s are mixtures of old and new miRNAs (Table S3). Thus, the former is most informative about the birth and early evolution of new miRNAs.
The miR-982s is X-linked, comprising five distinct miRNA families: miR-982, -2582, -303, -983 and -984. With the exception of the recently duplicated dme-miR-983-1/-2, miRNAs in this cluster do not share a seed sequence (Fig. 2A &B). Against the 12 Drosophila species [27], copies of this cluster can be found in D.simulans, D. sechellia, D. yakuba and D. erecta but are absent in all other more distantly related species. The expression of miR-982s members was confirmed by RT-PCR (Fig. S1). The evolution of this cluster in the D. melanogaster subgroup is depicted in detail in Fig. 2A.
As shown in Fig. 2A, each member of miR-982s appears to emerge in situ from local non-miRNA sequences. Due to their small sizes, unstructured genomic sequences evolving into miRNA-like transcripts have often been suggested [28] but have not been convincingly proven. The cluster of miR-982/2582/303/983/984 appears to be a good example of de novo origin (see below) with point mutations improving miRNA processing step by step (Fig. 2B and Fig. S2 and S3). For example, the secondary structure of miR-982 in D. erecta can only form a poor hairpin (−18.20 kcal/mol). Many nucleotide substitutions, accumulated subsequently in the stem and loop regions, have greatly improved the thermodynamic stability of the hairpin in D.melanogaster (−24.00 kcal/mol) and in the three paralogs of D. simulans (−21.52 to −27.50 kcal/mol; Fig. 3A and Fig. S2 and S3).
After each miRNA emerges from the unstructured sequence, gene duplication appears common [29], [30]. miR-2582 and miR-982 were expanded by whole-gene (Fig. 2A, Duplication 1, 2 and 3) or segment duplication (Duplication 4) in D. melanogaster and D. simulans, followed by gene conversion in D. sechellia (Fig. 2A). Moreover, miR-983 was duplicated in D. melanogaster (Duplication 5). In this species alone, miR-984 emerged de novo next to miR-983 (Fig. 2A).
These duplicates soon accumulated many nucleotide substitutions (Fig. 2B). Meanwhile, seed shifting and arm switching occurred in the miR-982/2582/303/983 families (Fig. 2B). These modifications presumably lead to new targets, resulting in neo-functionalization after gene duplication [28].
After new miRNAs emerged de novo, the question is whether the subsequent evolution is driven by natural selection. A greater level of divergence in miRNA genes than in flanking regions might suggest positive selection (Ref. [31]; Fig. S4A). A proper analysis would require the comparison of between-species divergence (D) and within-species polymorphism (P) using a modified McDonald-Kreitman (MK) test [32].
In this study, we generated DNA sequences from 42 D. melanogaster (∼7.5 kb from each line) and 25 D. simulans lines (∼8.1 kb) (Table S4). The D/P ratios for each precursor miRNA from miR-982s, as well as the 1 kb upstream flanking regions, were compared [33]. As shown in Table 2, all the miRNA genes from the miR-982, miR-303 and miR-983 families have a significantly higher D/P ratio than the flanking regions in both D. melanogaster and D. simulans (p<0.05), suggesting positive selection. Members of the miR-2582 family show significantly higher D/P ratios in D. melanogaster, but not in D. simulans (Table 2, also see next section).
Because each individual miRNA gene, being small, would yield a significant result in the MK test only when the selection is extremely strong, we also performed the test on new miRNAs collectively, relative to the genome-wide 4-fold degenerate sites (from Drosophila Population Genomics Project (DPGP); see Materials and Methods). Table 3 shows that the new miRNAs emerging in the last 30 Myrs have a higher D/P ratio than in the genome-wide 4-fold degenerate sites. In fact, more than 79% of the observed divergence in the precursors and more than 89% in the mature regions is estimated to have been fixed adaptively (see Materials and Methods and Table 3). A higher D/P ratio could also be attributed to an increase in selective constraint, rather than positive selection [34]. However, we excluded such possibility in Text S1. Due to the large number of adaptive sites, every new miRNA is likely to carry one or more of them. As expected, signatures of positive selection are much weaker for the lowly expressed miRNAs and mirtrons (Table S5).
Other lines of evidence for recent adaptive evolution include the pattern of polymorphism within species and the differentiation between populations. The miR-982 cluster was examined further by the sliding window analysis of Fay and Wu's H (θH), an estimator of nucleotide diversity sensitive to positive selection [35], [36]. The profile of θH peaks near miR-983/984 and miR-303 in both species, a common footprint of hitchhiking with positive selection [35]. The signature is stronger in D. simulans for miR-982 than in D. melanogaster (Fig. S4B and S4C). In addition, we analyzed the M and Z populations of D. melanogaster [37]–[39] using the Fst statistic [40]. For dme-miR-984 and dme-miR-303, the precursor sequences are strongly differentiated between M and Z lines (Fst = 0.318 for dme-miR-984 and Fst = 0.252 for dme-miR-303) compared to all SNPs within the miR-982s region (Mann-Whitney U test, p = 0.057 for dme-miR-984 and p = 0.068 for dme-miR-303, Table S6) or the 238 D. melanogaster miRNAs (Mann-Whitney U test, p = 0.046 for dme-miR-984 and p = 0.008 for dme-miR-303; data were obtained from DPGP2 [41], see Materials and Methods). The analyses collectively suggest that the rapid evolution of new miRNAs is driven by natural selection.
After the initial adaptive evolution, one might reasonably expect these new adaptive miRNAs to be integrated into the transcriptional network and begin evolving at a slower rate. Surprisingly, the most likely fate of these new miRNAs was death, rather than integration. This can be seen in the number of observable new miRNAs from two different time periods – 22 surviving miRNAs from the last 30 Myrs but only 9 from the preceding 30 Myrs (30–60 Myrs before present).
By assuming a constant birth rate, we can estimate the number of newborn miRNAs in each time interval, which can then be compared with the surviving miRNAs from that time period. Using the estimated rate of 3 newborn miRNAs per Myr (12 in the last 4 Myrs), Figure 4A shows that 87% of new miRNAs disappeared in 4–30 Myrs (68 out of 78). The proportion of death in older miRNAs increased only marginally, to 90%, for the period of 30–60 Myrs (81 out of 90). Therefore, most miRNAs seem to die quickly at an early stage of evolution, soon after the initial adaptive evolution. Only 6.0% of new miRNAs (34 out of 570) survived after 60 Myrs. It is unexpected that new adaptive miRNAs favored by natural selection should suffer such quick and massive death, albeit at a somewhat lower rate than neutrally evolving new miRNAs [12]. The former has a survival rate of 6.0% while the latter has a lower rate, at 2.5% [12]. Apparently, the initial adaptation is evolutionarily transient and the continual adaptation toward integration is not a common fate. We should note that alternative explanations have been considered. A most obvious one concerns the possibility of a bust of adaptive new miRNAs in D. melanogaster since its split from D. simulans. These explanations are compared in Discussion.
Interestingly, miRNA death may sometimes be an adaptive process. The miR-2582-like gene in D. melanogaster is shown to be evolving adaptively in Table 2, but its evolution is toward degeneracy. Three lineage-specific mutations that disrupt the duplex structure are shown in Fig. 3B, probably associated with the degeneration of dme-miR-2582. Presumably, conditions changed causing the adaptive function initially performed by the new miRNA to become deleterious at a later time.
Upon survival, new miRNAs eventually became integrated into the transcriptional network and evolved conservatively. There is a transitional phase after the adaptive phase, but before either integration or death. During the transition, these miRNAs often appeared to have a neutral evolutionary rate. Figure 4A shows that all the surviving miRNAs began to evolve either neutrally or conservatively (three transitional and six conservative miRNAs, respectively) within 30–60 Myrs (See Materials and Methods). The miR-2582 gene in D. simulans appears to be in such a transition (Table 2). It is interesting that miR-2582 orthologs in sibling species may be at different stages of evolution.
Over long periods of time, new miRNAs will have died or have been integrated into the transcriptional network and are now conservatively evolving. miRNAs born 60–250 Myrs ago have largely vanished (94% have disappeared, see Fig. 4A). However, some of the cohort of the 34 surviving miRNAs are not behaving as expected. In fact, only 26 of them are evolving conservatively. Nearly a quarter of them (8 out of 34) are evolving either neutrally or adaptively (Fig. 4A) and most of these (7 out of 8) come from miR-972s or miR-310s (Table 4). At this rate of evolution, none of them should have been recognizable as homologs between D. melanogaster and D. virilis.
We suggest that the 8 unusual miRNAs may have been conservatively evolving for most of their evolutionary history. Four of them have been adaptively evolving once again and the remaining four appear to be in transition, away from the previous selective constraints. If the hypothesis is correct, we expect to see stronger evolutionary conservation in more distant comparisons than in recent ones. We use KmiR/KS, where KmiR denotes the divergence in the precursor region of the miRNA, to measure conservation. Table 4 shows their KmiR/KS values for the last 4 Myrs and for the distant past (60 Myrs after the split between D. melanogaster and D. virilis). The evolutionary conservation has indeed been relaxed substantially in the last 4 Myrs with the average value increasing from 0.337 to 0.825, a 2.5-fold difference. Such fold-changes of KmiR/KS were significantly high in the eight miRNAs, compared with the whole repertoire of 238 miRNAs (Mann-Whitney U test, p = 0.00014). The rate increase appears to be true in both D. melanogaster and D. simulans lineages when the homologous sequences from D. yakuba and D. erecta were used as outgroups to calculate the rate in each lineage separately. Among the eight genes, two and six are evolving slightly faster in D. melanogaster and D. simulans, respectively (see Table S7). It is interesting that some old miRNAs go through the reverse transition (or rejuvenation) from conservative to adaptive evolution, the latter being the hallmark of young miRNAs.
Rejuvenation can also lead to the death of old miRNAs. The miR-972s may be such an example. Some members of this cluster emerged 60–250 Myrs ago and should have been integrated into the ancestral genome by the time D. pseudoobscura split from D. melanogaster. However, the entire miR-972s region was lost in D. pseudoobscura since the split.
Taken together, new miRNAs (such as miR-310s and miR-972s) may go through cycles of adaptation, integration (if escaping death) and rejuvenation, which would start another cycle of adaptation and integration (Fig. 4B).
To study the evolution of new miRNAs sequences, we characterized their expression patterns. We did so by using the global small RNA profiling datasets (see Table S1 and Materials and Methods). Figure 5 shows young miRNAs (<30 Myrs) are lowly expressed in specific tissues, generally in the testes and larval brain/gonads/imaginal discs. Middle-aged miRNAs (30–60 Myrs) broadened their expressions to include ovaries and embryos. The older miRNAs (60–250 Myrs) showed moderate and even broader expressions, which then evolved to become highly abundant in all tissues and developmental stages as seen in the oldest miRNAs (>250 Myrs). The simplest explanation is that new miRNAs increase the expression level and expand the breadth as they get older. The change in expression parallels that in sequence evolution (Fig. 4A and Table S8). There are other explanations that may also account for the different expression patterns between new and old miRNAs (see Text S2). Detailed descriptions of the evolution in expression patterns are given in Text S3.
During Metazoan evolution, the miRNA repertoire expanded dramatically from a few genes to several hundreds [28], [42]. By limiting the analysis to new miRNAs that evolve adaptively soon after their birth, we avoided the large number of lowly expressed miRNA-like sequences. These sequences may or may not be considered miRNAs and are generally thought to be evolutionarily ephemeral and adaptively insignificant [43], [44]. The inclusion of only new miRNAs that evolve adaptively at emergence reveals an unexpected pattern of an excess of such miRNAs in the last 4 million years of the D. melanogaster lineage. The possible explanations are therefore either a burst of birth since D. melanogaster split from D. simulans, or a decline in the survivorship of adaptive new miRNAs as they age.
We consider the latter explanation as more plausible for several reasons. First, the birth rate of miRNA-like sequences indeed appears constant because different Drosophila species have comparable numbers of such new transcripts [16]. Given the ease in forming precursor-like hairpins, the constant rate is hardly surprising. Second, as a result, the birth rate of adaptive new miRNAs may not deviate much from a constant value either. Indeed, the burst of adaptive new miRNAs is observable in D. simulans as well as the common ancestor of D. melanogaster and D. simulans, as is evident in the miR-982 cluster (Fig. 2A). Third, the proportion of adaptive miRNAs born in the period of 4–30 Myrs is also higher than that in the 30–60 Myrs period. Overall, an excess of new adaptive miRNAs appears to be a decreasing function of time, rather than of particular lineages; hence, their death over time is a simpler explanation.
Because only a small number of new adaptive miRNAs remain active after cycles of evolution through phases of adaptation and degeneration, the repertoire of miRNAs in the D. melanogaster genome has been nearly static in 40 Myrs of evolution, with only 0.18 miRNA integrations per Myrs. We should note that this low rate may still be an over-estimate because not all death has been accounted for. This (near) steady state echoes the view of a correlation between morphological complexity and the size of miRNA repertoire [45], as the Drosophila genus has been relatively invariant in form since its diversification.
Despite the low integration rate, many new miRNAs continue to emerge and some briefly evolve adaptively before their demise. This “transient utility” is puzzling as gene functions are lost usually through environmental changes (such as vision genes in caves [46]) or redundancies [47]. A possible explanation may be the suggested role of miRNAs in evolutionary canalization [7]. In such a role, the regulators and their targets need not be stringently wired as long as the system remains properly buffered. By this scheme, new miRNAs may emerge to fill in the transiently vacated role created by the shifting interactions between established miRNAs and their targets [7]. They disappear when the role is no longer needed.
A small number of new miRNAs that become integrated into the transcriptional network begin this process in the testis, in parallel with new protein coding genes [48]–[54]. Since sexual selection driving male reproduction is a very potent force of evolution, this expression pattern may not be all that surprising [50], [55]–[58]. In the example of miR-982s, the predicted targets are indeed enriched in genes of male courtship behavior and other male sexual traits (Table S9). Once a new miRNA is established, its expression is often broadened to other tissues. Testis may be the beachhead that permits the new miRNA to gradually modulate its expression and interactions with potential targets. In addition, new miRNAs with distinct seeds often emerge in clusters, which presumably facilitate their co-expression [29], [30], [59].
Unlike protein coding genes, miRNAs can easily emerge de novo, thanks to their small size, but can often be derived from existing genes as well [60]. The simple structure of miRNAs may permit general inferences on features and dynamics of genic evolution. A previous example is the rate of evolution as a correlate of expression level [61]. It would be interesting to see if the inferred cycles of evolution experienced by new miRNAs are a general process.
Total RNA was extracted from D. simulans (NC48S) and from D. pseudoobscura using TRIzol (Ambion). Ovaries and testes from 3 to 5-day adults were dissected and collected for both NC48S and D. pseudoobscura. Imaginal discs including central nerve system (CNS) were dissected from wandering third-instar larva of D. pseudoobscura. Small RNA libraries were generated from each RNA sample using Illumina Small RNA Sample Preparation kit, and sequenced with the Illumina HiSeq 2000 at the Beijing Genomics Institute (Shenzhen). The data were deposit at Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/) under the accession numbers GSM1165052-GSM1165056.
The publicly available small RNA sequencing reads from four Drosophila species (D. melanogaster, D. simulans, D. pseudoobscura and D. virilis) were downloaded from GEO database (http://www.ncbi.nlm.nih.gov/geo/, Table S1). The miRNA sequences of three Culicinae species (Aedes albopictus, A. aegypti and Culex quinquefasiatus) were adopted from two previous small RNA sequencing studies [24], [25]. Drosophila genome sequences were retrieved from UCSC (http://genome.ucsc.edu); the Whole Genome Alignment (WGA) and CDS alignment were obtained from 12 Drosophila Assembly/Alignment/Annotation (http://rana.lbl.gov/drosophila). The genome versions used were: D. melanogaster, dm3; D. simulans, droSim1; D. sechellia, droSec1; D. yakuba, droYak2; D. erecta, droEre2; D. ananassae, droAna3; D. pseudoobscura, dp4; D. persimilis, droPer1; D. willistoni, droWil1; D. mojavensis, droMoj3; D. virilis, droVir3; D. grimshawi, droGri2. The genome coordinates and sequences of miRNA genes were retrieved from miRBase Release 19 (http://www.mirbase.org). The genome coordinates and sequences of intron, rRNA, tRNA, snRNA and transposon elements were obtained from FlyBase (r5.41, http://flybase.org,)
We defined canonical miRNAs and mirtrons according to Ruby et al. [62]. Mirtrons were defined as pre-miRNAs with both 5′ and 3′ ends matching the splicing sites of host introns. The rest of the miRNAs were then classified as canonical miRNAs. When more than three miRNAs were located within a 20 kb region, these miRNAs were considered as a cluster.
Small RNA reads (18–30 nt) were extracted from sequencing data. Firstly, we excluded reads mapped to transposon elements and structural RNAs (rRNA, tRNA and snRNA) using bowtie [63], allowing no mismatch. Next, we annotated novel miRNAs by miRDeep2 [64] with default parameters. Finally, miRNAs with no read matching miR* were removed following previous practice [23]. We combined novel miRNAs sequences and known miRNA sequences for expression analysis.
For each species, small RNA reads (18–30 nt) were mapped to miRNA precursor sequences using bowtie [63], allowing no mismatch. Each read count was divided by the number of matches to miRNA precursors. The miRNA expression was normalized by total miRNA counts and scaled to reads per million (RPM), as previous described [18].
We examined phylogenetic distributions of the D. melanogaster miRNAs in three other Drosophila species (D. simulans, D. pseudoobscura and D. virilis) and three Culicinae species (Aedes albopictus, A. aegypti and Culex quinquefasiatus), where small RNAs have been profiled via deep sequencing [12], [16], [22], [24], [25]. Based on the comprehensive dataset, miRNA homologs were determined by homology search using either the whole genome alignment (WGA) within the Drosophila group or BLAST (threshold E<10−5) between Drosophila species and mosquitoes, and cross-checked with small RNA reads in the species in query (at least one read matching mature and miR*).
The homologous sequences of the D. melanogaster miRNA precursors in D. simulans (droSim1), D. pseudoobscura (dp4) and D. virilis (droVir3) were extracted from UCSC pairwise WGAs using LiftOver (http://hgdownload.cse.ucsc.edu/, minMatch = 0.6). The precursors failing to obtain hits in the genomes were subjected to BLASTN search against NCBI trace archives (http://www.ncbi.nlm.nih.gov/Traces/home/). Matched sequences with E-values <10−5 were also considered as miRNA homologs and recovered for the analysis below. The WGA output was compared with miRNA annotation by miRDeep2 [64]; miRNA orthologs confirmed by miRDeep2 were retained.
The miRNA precursor sequences in Aedes albopictus, Culex quinquefasiatus and A. aegypti were adopted from the studies of Li et al. [24] and Skalsky et al. [25]. These sequences were combined and subjected to BLASTN search against miRNA precursors in D. melanogaster. The best reciprocal hits with E-values <10−5 were retained as the corresponding miRNA homologs in the Culicinae lineage.
According to the phylogenetic distribution, maximum parsimony method was used to infer the origination of each miRNA along the main trunk of the phylogenetic tree of D.melanogaster, D. simulans, D. pseudoobscura, D. virilis and Culicinae. An miRNA is assumed to emerge in the most recent common ancestor of all the species bearing the authentic homologs. The branch lengths of the phylogenetic tree (in Myrs) were adopted from previous estimations [27], [65], [66]. The 238 miRNAs were classified into five age groups, corresponding to the time intervals of 0–4 Myrs, 4–30 Myrs, 30–60 Myrs, 60–250 Myrs and >250 Myrs.
The genomic coordinates and precursor sequences of dme-miR-982/303/983-1/983-2/984 and dsi-miR-982c/2582b/982b/2582a/982a/303/983 were retrieved from miRBase (Release v19). Based on the WGA of 12 Drosophila genomes [27], genomic sequence of the whole miR-982s cluster (∼9 Kb) in D. melanogaster (dm3) was extracted and used as a query to search against the other 11 Drosophila genomes using BLAT [67] with an E-value threshold of 0.001. We only detected hits in D. simulans, D. sechellia, D. yakuba and D. erecta, indicating that miR-982s is specific to the melanogaster subgroup. Homologous sequences of the miR-982s cluster from the five species were aligned using MUSCLE [68]. Homologs of miR-982s members in each species were identified using BLAST with the query of known precursor sequences (miRBase Release v19) and an E-value threshold of 0.001. The hits were further inspected in the alignment of the whole miR-982s cluster. The phylogenetic tree of each family of miR-982, miR-2582, miR-303, and miR-983 was reconstructed using the maximum likelihood method as implemented in MEGA 5.0 [69].
To validate the existence of miR-982s members in D. yakuba and D. erecta, we first predicted the secondary structure and thermo-stability of each miRNA homolog using RNAfold (http://rna.tbi.univie.ac.at/) with the default parameters [70]. A good hairpin with minimum free energy (MFE) >15 kcal/mol was considered as a potential miRNA candidate. There were four such candidates: dya-miR-2582-anc, dya-miR-303-anc, der-miR-982-anc, and der-miR-983-anc, where “anc” indicates ancestor. Then, we validated the expression of each candidate by amplifying the potential miRNA precursor from cDNA because the mature miRNA is hard to define. Total RNAs were extracted from testes of D. yakuba and D. erecta using TRIzol (Ambion) and treated with TURBO DNase Kit (Ambion). 0.5 ug RNA was reverse transcribed (RT) in a 20 ul reaction volume using PrimeScript II 1st Strand cDNA Synthesis Kit (TaKaRa). 1 ul RT products were used for PCR with Ex Taq DNA Polymerase (TaKaRa). PCR primers used are listed in Table S10.
A total of 25 D.simulans lines and 42 D.melanogaster lines, including 29 M lines and 13 Z lines, were used for population sequencing of the miR-982s cluster. The fly strains used were listed in Table S4. The genomic sequences of D.simulans (droSim1) and D.melanogaster (dm3) were used to design primer pairs that amplify a ∼8 Kb region spanning the whole miR-982s cluster and ∼1.5 Kb each of the upstream and downstream flanking regions. The PCR product of each primer set was designed to be about 2 Kb in length and overlapped with each other by at least 300 bp. The primers used are listed in Table S10 and their genomic coordinates are displayed in Fig. S5. PCR was carried out using LA Taq DNA Polymerase (TaKaRa). PCR products were subject to direct sequencing or clone sequencing on an ABI 3730xl DNA Analyzer (Applied Biosystems). DNA sequences were assembled using SeqMan software (DNASTAR Inc., USA) and aligned using MUSCLE [68] with manual inspection. Haplotypes were inferred with the PHASE program when heterozygous sites were present [71]. The sequences obtained in this study have been deposited in GenBank under the accession numbers JX648211-JX648278.
Using the population sequencing data, several methods were used to detect positive selection of miR-982s in D. melanogaster and D.simulans, respectively. First, MK tests were applied on each member of miR-982s based on the divergence between D. melanogaster and D.simulans consensus sequences and polymorphism within either species. Each miRNA precursor was tested against a 1 kb region about 1.5 kb upstream of the 5′ end of miR-982s. Second, sliding window analysis of divergence and polymorphism was applied to the whole miR-982s cluster and its flanking region. The divergence was calculated using Kimura's 2-parameter model [72] based on the genomic sequences of D. simulans (droSim1) and D. melanogaster (dm3). The polymorphism within either species was estimated using the method described previously [35], [36], [73], [74]. D. simulans (droSim1) and D. melanogaster (dm3) were used as the outgroup for each other reciprocally, in order to polarize the derived alleles. The window size is 100 bp and the step width is 25 bp. Finally, based on our miR-982s population data or DPGP2 data (see below) [41], the pattern of population differentiation (Fst) between Z and M lines was estimated for each miRNA precursor using Weir's method [40].
We used the McDonald-Kreitman test (MK test) [32] framework to detect positive selection in miRNAs from each age group based on the polymorphisms within D. melanogaster and the divergence between D. melanogaster and D. simulans. Precursor or mature sequences of each miRNA group were combined and treated as the functional category, while the 4-fold degenerate sites in the whole genome were used as the neutral control. The divergence is calculated by counting the number of changed nucleotide sites between D. melanogaster (dm3) and D. simulans (droSim1) based on the UCSC whole genome alignment. Polymorphism data was retrieved from Drosophila Population Genomics Project (DPGP, http://www.dpgp.org/, release 1.0). SNPs that were detected on more than thirty individuals and exhibited a derived allele frequency (DAF) >5% were used for the MK test.
The proportion of adaptively fixed mutations (α) was estimated as previously described [75]. To estimate the evolutionary fate of each miRNA, we first screened for adaptive miRNAs among the 238 candidates by using each miRNA's precursor together with the 50 bp flanking sequences on both sides as the functional sites. The p-values of multiple MK tests were adjusted by the Benjamini-Hochberg method [76] and the adaptive significance of each candidate is re-validated by using the precursor alone in the MK test. We then identified the conservative miRNAs by comparing the number of substitutions in the miRNA precursors (KmiR) with the number of substitutions in the synonymous sites (KS) between D.melanogaster and D.simulans. miRNAs with KmiR/KS<0.5 were considered to be conservatively evolving. Kimura's 2-parameter model [72] and the Nei-Gojobori model [77] were used to calculate KmiR and KS, respectively. Finally, excluding the adaptive and conservative miRNAs, the remaining were considered to be in transition between adaptive to conservative/death.
Data processing of small RNA deep sequencing libraries from different development stages and tissues of D. melanogaster [12], [16], [18]–[21], [23] was conducted as described above. The read counts of each miRNAs were normalized to Reads Per Million (RPM), which is the read number of each miRNA per million mapped reads in each library. The normalized counts were log2 transformed and subject to hierarchical clustering using R package heatmap2.
miR-982s targets were predicted by seed match using TargetScan (v5.0 http://www.targetscan.org/fly_12/) [5]. Taking all the miRNA members together, 1,002 targets were obtained in D. melanogaster and 3,563 in D. simulans, of which 454 were shared by both species. We used DAVID to perform a Gene Ontology (GO) enrichment test for the predicted targets in the two species (DAVID v6.7, http://david.abcc.ncifcrf.gov/) [78]. Only the GO terms for biological processes were used for the enrichment test.
|
10.1371/journal.pbio.1001064 | Cracking the Code of Oscillatory Activity | Neural oscillations are ubiquitous measurements of cognitive processes and
dynamic routing and gating of information. The fundamental and so far unresolved
problem for neuroscience remains to understand how oscillatory activity in the
brain codes information for human cognition. In a biologically relevant
cognitive task, we instructed six human observers to categorize facial
expressions of emotion while we measured the observers' EEG. We combined
state-of-the-art stimulus control with statistical information theory analysis
to quantify how the three parameters of oscillations (i.e., power, phase, and
frequency) code the visual information relevant for behavior in a cognitive
task. We make three points: First, we demonstrate that phase codes considerably
more information (2.4 times) relating to the cognitive task than power. Second,
we show that the conjunction of power and phase coding reflects detailed visual
features relevant for behavioral response—that is, features of facial
expressions predicted by behavior. Third, we demonstrate, in analogy to
communication technology, that oscillatory frequencies in the brain multiplex
the coding of visual features, increasing coding capacity. Together, our
findings about the fundamental coding properties of neural oscillations will
redirect the research agenda in neuroscience by establishing the differential
role of frequency, phase, and amplitude in coding behaviorally relevant
information in the brain.
| To recognize visual information rapidly, the brain must continuously code
complex, high-dimensional information impinging on the retina, not all of which
is relevant, because a low-dimensional code can be sufficient for both
recognition and behavior (e.g. a fearful expression can be correctly recognized
only from the wide-opened eyes). The oscillatory networks of the brain
dynamically reduce the high-dimensional information into a low dimensional code,
but it remains unclear which aspects of these oscillations produce the low
dimensional code. Here, we measured the EEG of human observers while we
presented them with samples of visual information from expressive faces (happy,
sad, fear, etc.). Using statistical information theory, we extracted the
low-dimensional code that is most informative for correct recognition of each
expression (e.g. the opened mouth for “happy,” the wide opened eyes
for “fear”). Next, we measured how the three parameters of brain
oscillations (frequency, power and phase) code for low-dimensional features.
Surprisingly, we find that phase codes 2.4 times more task information than
power. We also show that the conjunction of power and phase sufficiently codes
the low-dimensional facial features across brain oscillations. These findings
offer a new way of thinking about the differential role of frequency, phase and
amplitude in coding behaviorally relevant information in the brain.
| Invasive and noninvasive studies in humans under physiological and pathological
conditions converged on the suggestion that the amplitude and phase of neural
oscillations implement cognitive processes such as sensory representations,
attentional selection, and dynamical routing/gating of information [1]–[4]. Surprisingly,
most studies have ignored how the temporal dynamics of phase code the sensory
stimulus, focusing instead on amplitude envelopes (but see [5]), relations between amplitude and
frequency [6],
or coupling between frequencies ([7]–[10]; see [11] for a review). But there is compelling evidence that
phase dynamics of neural oscillations are functionally relevant [12]–[16].
Furthermore, computational arguments suggest that if brain circuits performed
efficient amplitude-to-phase conversion [17],[18], temporal phase coding could
be advantageous in fundamental operations such as object representation and
categorization by implementing efficient winner-takes-all algorithms [17], by providing
robust sensory representations in unreliable environments, and by lending themselves
to multiplexing, an efficient mechanism to increase coding capacity [18],[19]. To crack the
code of oscillatory activity in human cognition, we must tease apart the relative
contribution of frequency, amplitude, and phase to the coding of behaviorally
relevant information.
We instructed six observers to categorize faces according to six basic expressions of
emotion (“happy,” “fear,” “surprise,”
“disgust,” “anger,” “sad,” plus
“neutral”). We controlled visual information, by presenting on each
trial a random sample of face information—smoothly sampled from the image
using Gaussian apertures at different spatial frequency bands. The Gaussian
apertures randomly sampled face parts simultaneously across the two dimensions of
the image and the third dimension of spatial frequency bands (Figure S1
illustrates the sampling process for one illustrative trial; [20],[21]). We recorded the
observers' categorization and EEG responses to these samples (see Materials and Methods, Procedure).
To quantify the relative coding properties of power, phase, and frequency, we used
state-of-the-art information theoretic methods (Mutual Information,
MI, which measures the mutual dependence between two variables;
[22]) and
computed three different MI measurements: between sampled pixel
information and behavioral responses to each emotion category (correct versus
incorrect), between EEG responses (for power, phase, and the conjunction of phase
and power) and behavior, and finally between sampled pixel information and EEG
response (see Figure
S2 for the mutual information analysis framework and Computation: Mutual
Information).
First, to characterize the information that the brain processes in the cognitive
task, for each observer and category, we computed MI(Pixel;
Behavior), the MI between the distribution of grey-level values of
each image pixel (arising from the summed Gaussian masks across spatial frequency
bands, down-sampled from a 380×240 pixels image to a 38 to 24 image and
gathered across trials) and equal numbers of correct versus incorrect categorization
responses. Figure 1,
MI(Pixel; Behavior) illustrates MI on a scale
from 0 to 0.05 bits. High values indicate the face pixels (e.g., forming the mouth
in “happy”) representing the visual information that the brain must
process to correctly categorize the stimuli (see Figure S3 for a
detailed example of the computation).
We now compare how the parameters of oscillatory frequency, power, and phase code
this information in the brain. For each observer, expression, electrode of the
standard 10–20 position system, and trial, we performed a Time ×
Frequency decomposition of the signal sampled at 1,024 Hz, with a Morlet wavelet of
size 5, between −500 and 500 ms around stimulus onset and every 2 Hz between 4
and 96 Hz. We make three points:
(a) The conjunction of phase and power (phase&power) codes more
information about complex categorization tasks than phase and power on their
own. In Figure 2,
MI(EEG response; Behavior) measures the reduction of
uncertainty of the brain response, when the behavioral variable correct versus
incorrect categorization is known. We provide the measure for each electrode of the
standard 10–20 position system over the Time × Frequency space. Pz, Oz,
P8, and P7 had highest MI values of all electrodes, irrespective of
whether the brain response considered was power (blue box), phase (green box), or
the phase&power (red box). The adjacent MI scales reveal that
phase&power was 1.25 times more informative of behavior than phase, itself 2.4
times more informative than power. Phase&power was 3 times more informative than
power alone. Henceforth, the analyses focus on these four electrodes and on
phase&power, the most informative brain measurement for the cognitive task.
(b) Phase&power codes detailed categorization-relevant features of
sensory stimuli. MI(Pixel; Behavior) revealed that the two eyes and the
mouth are prominent features of expression discrimination (see Figure 1). As explained, with Gaussian masks we
sampled pixels from the face on each trial. Consequently, for all correct trials of
an expression category (e.g., “happy”), we can measure at each pixel
location the mutual information between the distribution of grey-level values of the
Gaussian masks across trials and each cell of the Time × Frequency brain
response. Figure 3 reports
MI(Pixel; Phase&Power), focusing on Pz, Oz, P8, and P7. The
red box represents, at 4 Hz and 156 ms, following stimulus onset (a time point
chosen for its prominence in face coding [21]), the color-coded
MI value of each face pixel—overlayed on a neutral face
background for ease of feature interpretation (the yellow box presents mutual
information at 12 Hz and 156 ms). The scale is the adjacent rainbow colors ranging
from 0 to 0.03 bits. Electrodes P7 (over left occipito-temporal cortex) and P8 (over
right occipital-temporal cortex) reveal the highest MI to the
contra-lateral eye (i.e., left eye for P8; right eye for P7). At the same time on Pz
and Oz, the highest MI is to both eyes and to the mouth.
To generalize across Time × Frequency, for ease of presentation, we computed
three masks extracting pixel locations from the left eye, right eye, and mouth. We
averaged MI values within each mask, independently for each Time
× Frequency cell. We then color-coded MI for each feature in
RGB color space—red for “right eye,” green for
“mouth,” and blue for “left eye”; see schematic colored
faces adjacent to the Time × Frequency plot for complete color coding. The
broad red (versus blue) cloud on electrode P7 (versus P8) denotes highest
MI to the right (versus left) eye in this Time ×
Frequency region, whereas Pz and Oz demonstrate sensitivity to the two eyes (in
purple) and to the mouth (in green). To conclude, phase&power codes detailed
categorization-relevant features of the sensory input.
(c) Phase&power coding is multiplexed across oscillatory
frequencies. Theta (4 Hz) and low beta (12 Hz) on both Oz and Pz
demonstrate the remarkable multiplexing property of phase&power coding: the idea
that the brain codes different information in different oscillatory bands. In Figure 3, Oz and Pz reveal that
beta encodes two eyes (see the purple RGB code and the yellow framed faces) when
theta encodes the mouth (see the green RGB code and the red framed faces).
Multiplexing is also present to a lesser degree on P8 and P7. MI
values critically depend on the joint distribution of variables (see Figure S3), and
so we turn to Figure 4 to
understand how the variables of phase and power jointly contribute to the coding of
facial features. Figure 4
develops the red and yellow framed faces of Figure 3, for electrode Pz. At 156 ms, at 4 and
12 Hz, we discretized the distribution of power and phase neural responses in
3×3 bins—represented in Cartesian coordinates as
. In each bin, we averaged the pixel values leading to this
range of imaginary numbers. At 12 Hz, what emerges is a phase&power coding of
the two eyes (in red, between 45 and 90 deg of phase) and an encoding of the mouth
(in red, between 270 and 315 deg of phase). At 4 Hz, the encoding of mostly the
mouth and the two eyes (in red) occurs between 90 and 135 deg of phase. The 4 and 12
Hz colored boxes in Figure 4
therefore illustrate the prominence of phase coding for facial features.
Here, using the concept of mutual information from Information Theory, we compared
how the three parameters of neural oscillations (power, phase, and frequency)
contribute to the coding of information in the biologically relevant cognitive task
of categorizing facial expressions of emotion. We demonstrated that phase codes 2.4
times more information about the task than power. The conjunction of power and phase
(itself 3 times more informative than power) codes specific expressive features
across different oscillatory bands, a multiplexing that increases coding capacity in
the brain.
In general, the relationship between our results on the frequency, power, and phase
coding of neural oscillations cannot straightforwardly be related to the coding
properties of more standard measures of the EEG such as event related potentials
(ERP). However, an identical experimental protocol was run on the N170
face-sensitive potential [21],[23], but using reverse correlation analyses, not MI. Sensor
analyses revealed that the N170 ERP initially coded the eye contra-lateral to the
sensor considered, for all expressions, followed at the N170 peak by a coding of the
behaviorally relevant information [21], together with a more detailed coding of features (i.e.,
with their Higher Spatial Frequencies) at the peak [23]. Interestingly, distance
of behaviorally relevant information (e.g., the wide-opened eyes in
“fearful” versus the mouth in “happy”) to the initially
coded eye determined the latency of the N170 peak (with the ERP to a
“happy” face peaking later than to a “fearful” face). ERPs
confer the advantage of precise timing, leading to precise time course of coding in
the brain, including phase differences across visual categories. However, we do not
know whether this coding occurs over one or multiple sources of a network that might
oscillate at different temporal frequencies (as suggested here between theta and
beta), for example to code features at different spatial resolutions (as suggested
in [19] and [24]). In sum, the
complex relations between EEG/MEG data, the underlying cortical networks of sources,
their oscillatory behaviors, and the coding of behaviorally relevant features at
different spatial resolutions open a new range of fundamental questions. Resolving
these questions will require integration of existing methods, as none of them is
singly sufficient.
In these endeavors, the phase and frequency multiplexing coding properties of neural
oscillations cannot be ignored.
Six observers from Glasgow University, UK, were paid to take part in the
experiment. All had normal vision and gave informed consent prior to
involvement. Glasgow University Faculty of Information and Mathematical Sciences
Ethics Committee provided ethical approval.
Original face stimuli were gray-scale images of five females and five males taken
under standardized illumination, each displaying seven facial expressions. All
70 stimuli (normalized for the location of the nose and mouth) complied with the
Facial Action Coding System (FACS, [25]) and form part of the
California Facial Expressions (CAFE) database [26]. As facial information is
represented at multiple spatial scales, on each trial we exposed the visual
system to a random subset of Spatial Frequency (SF) information contained within
the original face image. To this end, we first decomposed the original image
into five non-overlapping SF bands of one octave each (120–60,
60–30, 30–15, 15–7.5, and 7.5–3.8 cycles/face, see Figure S1).
To each SF band, we then applied a mask punctured with Gaussian apertures to
sample SF face information with “bubbles.” These were positioned in
random locations trial by trial, approximating a uniform sampling of all face
regions across trials. The size of the apertures was adjusted for each SF band,
so as to reveal six cycles per face. In addition, the probability of a bubble in
each SF band was adjusted so as to maintain constant the total area of face
revealed (standard deviations of the bubbles were 0.36, 0.7, 1.4, 2.9, and 5.1
cycles/degree of visual angle from the fine to the coarse SF band). Calibration
of the sampling density (i.e., the number of bubbles) was performed online on a
trial-by-trial basis to maintain observer's performance at 75%
correct categorization independently for each expression. The stimulus presented
on each trial comprised the randomly sampled information from each SF band
summed together [27].
Prior to testing, observers learned to categorize the 70 original images into the
seven expression categories. Upon achieving a 95% correct classification
criterion of the original images, observers performed a total of 15 sessions of
1,400 trials (for a total of 21,000 trials) of the facial expressions
categorization task (i.e., 3,000 trials per expression, happy, sad, fearful,
angry, surprised, disgusted, and neutral faces, randomly distributed across
sessions). Short breaks were permitted every 100 trials of the experiment.
In each trial a 500 ms fixation cross (spanning 0.4° of visual angle) was
immediately followed by the sampled face information, as described before (see
Figure
S1). Stimuli were presented on a light gray background in the centre
of a monitor; a chin-rest maintained a fixed viewing distance of 1 m (visual
angle 5.36°×3.7° forehead to base of chin). Stimuli remained on
screen until response. Observers were asked to respond as quickly and accurately
as possible by pressing expression-specific response keys (seven in total) on a
computer keyboard.
We recorded scalp electrical activity of the observers while they performed the
task. We used sintered Ag/AgCl electrodes mounted in a 62-electrode cap
(Easy-Cap) at scalp positions including the standard 10–20 system
positions along with intermediate positions and an additional row of low
occipital electrodes. Linked mastoids served as initial common reference and
electrode AFz as the ground. Vertical electro-oculogram (vEOG) was bipolarly
registered above and below the dominant eye and the horizontal electro-oculogram
(hEOG) at the outer canthi of both eyes. Electrode impedance was maintained
below 10 kΩ throughout recording. Electrical activity was continuously
sampled at 1,024 Hz. Analysis epochs were generated off-line, beginning 500 ms
prior to stimulus onset and lasting for 1,500 ms in total. We rejected EEG and
EOG artefacts using a [−30 µV; +30 µV]
deviation threshold over 200 ms intervals on all electrodes. The EOG rejection
procedure rejected rotations of the eyeball from 0.9 deg inward to 1.5 deg
downward of visual angle—the stimulus spanned 5.36°×3.7° of
visual angle on the screen. Artifact-free trials were sorted using EEProbe (ANT)
software, narrow-band notch filtered at 49–51 Hz, and re-referenced to
average reference.
In Information Theory [28],[29], Mutual Information
MI(X;Y ) between random
variables X and Y measures their mutual
dependence. When logarithms to the base 2 are used in Equation 1, the unit of
mutual information is expressed in bits.(1)
The critical term is p(x,y),
the joint probabilities between X and Y. When the variables are
independent, the logarithm term in Equation 1 becomes 0 and
MI(X;Y
) = 0. In contrast, when X and
Y are dependent
MI(X;Y ) returns a value
in bits that quantifies the mutual dependence between X and
Y. Derived from the measure of uncertainty of a random
variable X expressed in Equation 2 and the conditional
uncertainty of two random variables X and Y
(Equation 3),(2)(3)
Mutual Information measures how much bits of information X and
Y share. It quantifies the reduction of uncertainty about
one variable that our knowledge of the other variable induces (Equation
4),(4)
Here, we use Mutual Information to measure the mutual dependence between the
sampling of input visual information from faces and the oscillatory brain
responses to these samples and between the same input information and behavior
(see Figure
S2 for an overall illustration of our framework; see Figure S3
for a detailed development of the computations between face pixels and correct
versus incorrect behavioral responses). For all measures of MI, we used the
direct method with quadratic extrapolation for bias correction [22]. We
quantized data into four equi-populated bins, a distribution that maximizes
response entropy [22]. Results were qualitatively similar for a larger
number of bins (tested in the range of 4 to 16). Below, we provide details for
the computation of mutual information with behavioural and EEG responses,
including number of trials taken into consideration for the MI computations and
the determination of statistical thresholds of mutual information.
On each of the 21,000 trials of a categorization task, the randomly located
Gaussian apertures make up a three-dimensional mask that reveals a sparse face.
Observers will tend to be correct when this sampled SF information is diagnostic
for the categorization of the considered expression. To identify the face
features used for each facial expression categorization, we computed mutual
information, per observer, between the grey levels of each face pixels and a
random sample of correct matching the number of incorrect trials (i.e., on
average 5,250 correct trials and 5,250 incorrect trials). For each expression,
we then averaged mutual information values across all six observers,
independently for each pixel. To establish statistical thresholds, we repeated
the computations 500 times for each pixel, after randomly shuffling the order of
response—to disrupt the association between pixel values and
categorization responses. For each of the 500 computations, we selected the
maximum mutual information value across all pixels. We then chose as statistical
threshold the 99th percentile of the distribution of maxima. This maximum
statistic implements a correction for multiple comparisons because the
permutation provides the null distribution of the maximum statistical value
across all considered dimensions [30]. Behavioral mutual
information is reported as the top row of faces in Figure 1.
Here, we examined two different measures: MI(EEG Response;
Behavior) and MI(Pixel; EEG Response). MI(EEG
Response; Behavior) computed, for each electrode, subject, and expression, the
mutual information between correct and incorrect trials and the power, phase,
and phase&power of the Time × Frequency EEG signal. For this
computation, we used the same number of trials as for Behavior MI (i.e., on
average 5,250 correct trials and 5,250 incorrect trials). As with behavior, for
each electrode and type of EEG measurement, we averaged the mutual information
values across subjects and expression. To establish statistical thresholds, we
repeated the computations 500 times, permuting the trial order of the EEG Time
× Frequency values and identified the 500 maxima each time across the
entire Time × Frequency space. We identified the statistical threshold as
the 99th percentile of the distribution of maxima (see Figure 2).
MI(Pixel; Phase&Power) computed, for each subject,
expression, and face pixel (down-sampled to 38×24 pixel maps), the mutual
information between the distribution of each face pixel grey-level value and the
most informative of the brain responses, phase&power Time × Frequency
responses, for correct trials only. That is, an average of 15,750 trials per
subject. To establish statistical thresholds, given the magnitude of the
computation, we computed z scores using the pre-stimulus
presentation baseline (from −500 to 0 ms) to estimate mean and standard
deviation. In Figure 3, .01
bits of mutual information correspond to a z score of 55.97, so
all mutual information values this number of bits (see the level marked with an
asterisk in Figure 3) are
well above an uncorrected threshold of .0000001 (itself associated with a
z score of 5).
Figure 2 indicated two
clusters of maximal MI in all three measures (Power, Phase, and
Phase&Power) at a latency of 140–250 ms in two frequency bands (4 Hz
and 12–14 Hz). We averaged the MI measures, for each
cluster, electrode, and subject, and subjected these MI
averages to a two-way ANOVA with factors electrode (P7, P8, Pz, and Oz) and
measure (Power, Phase, and Phase&Power). Both clusters revealed a
significant main effect of electrode (F(1,
3) = 8.38, p<0.001 for 4 Hz and
F(1, 3) = 79.34,
p<0.001 for 12–14 Hz) and measure
(F(1, 2) = 44.24,
p<0.001 for 4 Hz and F(1,
2) = 104.77, p<0.001 for 12–14
Hz). Post hoc t test confirmed that
MI(Phase&Power) is significantly higher than
MI(Phase)
(p = 0.013), which itself is significantly
higher than MI(Power)
(p = 0.003).
|
10.1371/journal.pbio.1002115 | Ih Channels Control Feedback Regulation from Amacrine Cells to Photoreceptors | In both vertebrates and invertebrates, photoreceptors’ output is regulated by feedback signals from interneurons that contribute to several important visual functions. Although synaptic feedback regulation of photoreceptors is known to occur in Drosophila, many questions about the underlying molecular mechanisms and physiological implementation remain unclear. Here, we systematically investigated these questions using a broad range of experimental methods. We isolated two Ih mutant fly lines that exhibit rhythmic photoreceptor depolarization without light stimulation. We discovered that Ih channels regulate glutamate release from amacrine cells by modulating calcium channel activity. Moreover, we showed that the eye-enriched kainate receptor (EKAR) is expressed in photoreceptors and receives the glutamate signal released from amacrine cells. Finally, we presented evidence that amacrine cell feedback regulation helps maintain light sensitivity in ambient light. Our findings suggest plausible molecular underpinnings and physiological effects of feedback regulation from amacrine cells to photoreceptors. These results provide new mechanistic insight into how synaptic feedback regulation can participate in network processing by modulating neural information transfer and circuit excitability.
| Feedback regulation is a common feature of neural circuits during the process of acquiring information. Therefore, it is important to understand how this phenomenon occurs. Using the primary visual system of the fruit fly Drosophila melanogaster as a model, we systematically investigated the molecular mechanisms and the physiological implementation of feedback regulation from amacrine cells (second order neurons that are present in the lamina) to photoreceptors. We isolated two fly lines with mutations in the gene that encodes for the ion channel known as Ih, whose photoreceptors exhibited rhythmic depolarizations in the absence of light stimulation. We demonstrated that Ih channels function in amacrine cells to regulate the release of the neurotransmitter glutamate by modulating the activity of the voltage-gated calcium channel, Cac. We further found that the glutamate signal released by amacrine cells is sensed and transduced by glutamate receptors expressed by the photoreceptors. Finally, we showed that this feedback regulation is critical for maintaining light sensitivity in the presence of ambient light. Our results suggest that regulation of synaptic feedback in a neuronal network modulates information transfer and circuit excitability.
| Feedback regulation is common in neural circuit information processing. In both vertebrate and invertebrate visual systems, photoreceptor output is feedback-regulated by interneurons, which is an important mechanism for shaping the transmission of light information [1,2]. In the vertebrate retina, bipolar cells receive synaptic input from rod and cone photoreceptors and transfer information to ganglion cells. Meanwhile, the laterally distributed horizontal cells provide a feedback signal to photoreceptor axon terminals, controlling their output gain [3,4]. The structure, function, and development of the vertebrate and insect visual systems possess many evolutionary parallels [5]. In the Drosophila lamina, 12 neuron classes have been identified, and specific interneurons may serve similar functions [6,7]. Serial electron-micrograph (EM) studies have revealed that outer photoreceptor (R1–R6) axons project their outputs to L1–L3 monopolar cells and amacrine cells (AC) and receive synaptic inputs from L2, L4, AC, Lawf, and C3 cells [6–8]. Because connectivity in the Drosophila lamina has been elucidated to the level of individual synapses, this system provides a good model to study how the feedback neural circuit works and facilitates network information processing [6,7,9,10].
Upon light stimulation, Drosophila photoreceptors undergo depolarization via activating the phototransduction cascade, which opens transient receptor potential (TRP) channels [11]. In turn, depolarized photoreceptors release the inhibitory neurotransmitter histamine [12] and hyperpolarize postsynaptic L1–L3 neurons and ACs by opening their histamine-gated chloride channel, HisCl 2 [13]. Intracellular recordings from L1–L3 neurons and R1–R6 photoreceptors imply that L2 and AC receive inhibitory input from R1–R6 axons and subsequently depolarize the photoreceptors through synaptic excitation (glutamate, acetylcholine, or both) [14]. Microinjection, immunolabeling, and genetic reporter line experiments suggest that AC and L2 are either glutamatergic or cholinergic neurons, while L4 is either cholinergic or gabaminergic [12,15–18]. However, the physiological roles of these feedback regulations and the underlying molecular mechanisms remain unclear. In addition, the types of excitatory neurotransmitter receptors in R1–R6 photoreceptors are still unknown.
Ih channels, also called hyperpolarization-activated cyclic nucleotide-gated (HCN) channels, are low-threshold, voltage-gated ion channels that are normally activated at negative potentials below −50 mV [19,20]. As Ih channels are permeable to Na+ and K+ and form an inward current at rest, they may depolarize the neuronal resting membrane potential (RMP) and influence excitatory postsynaptic potential kinetics and integration [21–23]. A recent study demonstrated that HCN1 colocalizes with low-threshold voltage-gated T-type Ca2+ channels (Cav3.2) in presynaptic terminals and inhibits glutamate release by suppressing Cav3.2 activity [24]. In the present study, we examined two Ih mutant fly lines that exhibit rhythmic depolarization in photoreceptors without light stimulation. Our results demonstrate that Ih channels are expressed in ACs and suggest that Ih channels regulate synaptic glutamate release by modulating the activity of Cacophony (Cac) channels. We further showed that the eye-enriched kainate receptor (EKAR) receives the retrograde glutamate signal in photoreceptor terminals. Finally, we investigated how feedback regulation from ACs affects photoreceptor output and fly optomotor behavior. Our studies elucidate the molecular mechanism and physiological roles of feedback regulation from ACs to photoreceptors in the Drosophila visual system.
To identify additional genes involved in fly photoreceptor functions, we performed an electroretinogram (ERG)-based genetic screen in the mutant lines from Exelexis collections [25,26] and gene disruption project (GDP) collections [27]. In this screen, we identified two Ih mutant lines, PBac{XP}Ihf01485 and PBac{XP}Ihf03355, that exhibited normal light responses and distinctive ERG baseline oscillations (Fig. 1A–C). Although the amplitudes and frequencies of oscillations were variable within individual flies (S1A Fig.), this ERG phenotype was easily detectable and distinct from that of wild-type flies, which never exhibited ERG baseline oscillations (Fig. 1B,C). Strikingly, ERG baseline oscillations in Ih mutant flies were sustained for more than 15 min, although their amplitude and frequency were attenuated (S1B Fig.). Similar results were observed in the recombinants with two deficiency lines Df(2R)Exel7131 and Df(2R)BSC700, in which the entire Ih gene (Gene ID: 36589) was deleted (Fig. 1B,C) [26,28]. Splicing of the Ih gene creates several transcriptional variants that encode Ih channels with long or short N-termini and different lengths of the inter-loop regions between the membrane-spanning domains S3–S4 and S4–S5 (Fig. 1A) [29]. The PBac{XP}Ihf03355 mutant contains a piggyBac inserted into the extron of all transcriptional variants of the Ih gene, whereas PBac{XP}Ihf01485 mutant has a piggyBac inserted into the intron of most transcriptional variants of the Ih gene (Fig. 1A) [25]. Reverse transcription polymerase chain reaction (RT-PCR) analysis revealed that the piggyBac insertion completely abolished the mRNA transcription of Ih gene in these two mutant lines (Fig. 1D). Using antibodies against the intracellular C terminal domain that exists in all Ih channel variants, we revealed four major Ih channel variants (170, 125, 73, and 71 kDa) that were expressed in wild-type flies but absent in Ih mutant flies (Fig. 1E). In addition, two low intensity bands (74 and 52 kDa) exist in Ih mutants (Fig. 1E), which might be nonspecific bands. To further confirm that piggyBac insertion actually disrupts the Ih gene and leads to ERG baseline oscillations in PBac{XP}Ihf03355 mutants, we performed piggyBac precise excision from PBac{XP}Ihf03355 mutants. ERG recording showed that this completely abolished the ERG baseline oscillations phenotype (Fig. 1B,C). Taken together, these results demonstrate that loss of Ih channels results in an abnormal ERG baseline oscillation phenotype.
Because ERG is an extracellular recording technique that measures the light-induced mass response of the eye, we next conducted intracellular recordings in photoreceptors to investigate whether the ERG baseline oscillation phenotype was due to a photoreceptor abnormality. Interestingly, Ih mutant photoreceptors, but not wild-type photoreceptors, showed rhythmic depolarization without light stimulation (Fig. 2A,B). In the first several minutes of recording, the pacemaker traces in Ih mutant flies showed a uniform amplitude (3.10 ± 0.36 mV) and frequency (0.24 ± 0.05 Hz). Each depolarization had a rise time of 0.82 ± 0.12 s, lasted 1.57 ± 0.13 s, and had a decay time of 0.60 ± 0.15 s (Fig. 2A). The frequency of depolarization in the intracellular recordings was much lower than that in ERG recordings, likely due to the desynchronized depolarization of multiple photoreceptors in the extracellular recordings. In addition, Ih mutant photoreceptors showed a reduced amplitude (10.3 ± 0.7 mV versus 11.6 ± 0.8 mV, p < 0.01) as well as a prolonged decay time (0.23 ± 0.03 s versus 0.06 ± 0.01 s, p < 0.001) of light-evoked depolarization (Fig. 2C). These alterations might reflect abnormal photoreceptors and/or cell communication in the visual system.
To explore whether loss of Ih channels causes a photoreceptor abnormality, we conducted EM studies and biochemical analyses of photoreceptors. EM images showed normal rhabdomeral structures in Ih mutant flies (Fig. 3A), and the expression and localization of phototransduction components in Ih mutant flies were also normal (Fig. 3B,C), indicating that the rhythmic depolarization noted in Ih mutant photoreceptors was not due to abnormalities in rhabdomere structure or phototransduction cascades. We also excluded the possibility that the rhythmic depolarization in Ih mutant photoreceptors was dependent on phototransduction cascades activation by genetically blocking phototransduction cascade activation through introducing the mutant of norpA, which encodes the key phototransduction component phospholipase C [30]. ERG recordings showed that norpA mutation did not suppress the rhythmical depolarization in Ih mutant photoreceptors (Fig. 3D).
To elucidate how the loss of Ih channels leads to rhythmic depolarization in photoreceptors, we examined the expression pattern of endogenous Ih channels. In wild-type flies, Ih channel staining was observed in the lamina and medulla, whereas photoreceptor cell bodies were either weakly or not at all labeled (Fig. 4A). Although the axons of outer photoreceptors project into the lamina and form synaptic connections with multiple lamina neurons, Ih channels were undetectable in photoreceptor axons (Fig. 4B). Strong Ih labeling was observed in the somata of L1 and L2 neurons, whose membranes had been labeled with mCD8-GFP markers (Fig. 4C,D). Ih channels were also found in the somata and processes of ACs, which was identified by expression of mCD8-GFP markers under the control of AC-specific split GAL4 (Lai-GAL4) (Fig. 4E) [31].
To determine whether Ih channels in the lamina neurons or glia might contribute to rhythmic depolarization in photoreceptors, we specifically depleted Ih channels in various cell types using an RNAi knockdown approach. The RNAi line THU02084.N recognizes all Ih transcriptional variants (Fig. 1A) [32], and western blotting analysis demonstrated that RNAi knockdown of Ih channels using pan-neural elav-GAL4 but not glia-specific repo-GAL4 successfully repressed the expression of all Ih variants (Fig. 5A). Moreover, RNAi against all Ih transcriptional variants using elav-GAL4 but not repo-GAL4 phenocopied the abnormal ERG baseline oscillations observed in Ih mutant flies (Fig. 5B,C), indicating that the loss of Ih channels in neurons, but not glia, results in rhythmic depolarization in photoreceptors. Consistent with no or low Ih channel expression in photoreceptors, the Rh1-GAL4 driver did not recapitulate the ERG deficits observed in Ih mutant flies (Fig. 5B,C). These results suggest that rhythmical depolarization in Ih mutant photoreceptors is caused by abnormal communication between photoreceptors and other neurons.
Photoreceptor terminals receive synaptic inputs from L2, L4, and Lawf neurons and ACs [6–8]. Given that Ih channels are expressed in L1 andL2 neurons and ACs, we next asked whether the loss of Ih channels in these neurons causes rhythmic depolarization in photoreceptors. Although RNAi knockdown of Ih channels in L1 and L2 neurons using L1L2-GAL4 [33,34] significantly reduced the levels of Ih channels (Fig. 5A), these RNAi knockdown flies did not exhibit obvious ERG baseline oscillations (Fig. 5B,C), suggesting that the loss of Ih channels in L1 and L2 neurons is not sufficient to trigger rhythmic depolarization in photoreceptors. Our immunostaining data from Lai-GAL4;UAS-mCD8-GFP flies have shown that the number of ACs is small (Fig. 4E), and anti-Ih antibody staining revealed that Ih expression in ACs is much lower than that in L1/L2 neurons (Fig. 4E). Therefore, depletion of Ih channel from ACs may not lead to a detectable reduction in total Ih channels in the whole fly head (Fig. 5A). To validate that RNAi knockdown in ACs works well, we performed immnuostaining analysis and showed that RNAi knockdown of Ih channels using Lai-GAL4 successfully did deplete Ih channels in ACs (Fig. 5D). Interestingly, RNAi knockdown of Ih channels using Lai-GAL4 recapitulated the ERG abnormalities observed in Ih mutant flies (Fig. 5B,C), indicating that the loss of Ih channels in ACs leads to rhythmic depolarization in photoreceptors. Although knockdown of Ih channels in L1 and L2 neurons using L1L2-GAL4 caused occasional ERG baseline oscillations, the recombinant of Lai-GAL4 and L1L2-GAL4 resulted in greater ERG baseline oscillations compared with those caused by Lai-GAL4 alone (Fig. 5B,C). These results suggest that the loss of Ih channels in L1 and L2 neurons might also contribute to rhythmic depolarization in photoreceptors.
We next generated p[UAS-Ih] transgenic flies and performed rescue experiments. To choose an appropriate variant for transgene generation, we first determined which variant is expressed in the retina and lamina. The RNAi line P{KK100190}VIE-260B only recognizes the long isoforms of Ih gene (Fig. 1A) [35]. RNAi against long Ih transcriptional variants using eye-specific eyeless-GAL4 showed normal Ih protein levels and ERG response (S2A,B Fig.). Furthermore, pan-neural elav-GAL4 failed to recapture the ERG baseline oscillations (S2B,C Fig.), although it caused a significant reduction in the long isoforms of Ih proteins (170 and 125 kDa, S2A Fig.). The above observations indicated that only short isoforms of Ih play essential roles in suppressing the ERG baseline oscillation phenotype. Western blotting also revealed that 71- and 73-kDa variants but not 170- and 125-kDa variants were highly expressed in the isolated retina and lamina (Fig. 6A). RT-PCR analysis further showed that the annotated transcript form Ih-RK, which encodes a 618aa isoform Ih-PK protein, was abundant in the retina and lamina (Fig. 6B). Thus, we amplified the cDNA of Ih-RK and generated p[UAS-IhK] transgenic flies to confirm that this transgene can be successfully expressed in neurons (Fig. 6C).
Expression of Ih channels in ACs but not in L1 and L2 neurons restored normal ERG activity in Ih mutant flies (Fig. 7A–C). Conversely, expression of Ih channels in photoreceptors (Rh1-GAL4) or glia (repo-GAL4) had no inhibitory effects on ERG baseline oscillations (Fig. 7A,B). Intracellular recordings further validated that the expression of Ih channels in ACs suppressed rhythmic depolarization in photoreceptors (Fig. 7D). These observations provide evidence that AC-derived Ih channels are critical to inhibit rhythmic depolarization in photoreceptors.
As photoreceptor terminals receive synaptic inputs directly from ACs, we first investigated whether rhythmic depolarization in Ih mutant photoreceptors was due to abnormal cartridge structure or connections between photoreceptors and ACs. However, EM images showed no obvious morphological differences between Ih mutant and wild-type cartridges (Fig. 8A). Therefore, we suspected that rhythmic depolarization in photoreceptors might be due to abnormal synaptic output from ACs. To test this hypothesis, we blocked neurotransmitter release from ACs by expressing tetanus toxin light chain (TeTxLC) [36] or silencing ACs by expressing a mutant form of open rectifier potassium channel (dORKΔC) [37]. TeTxLC and dORKΔC expression was suppressed during development by the ubiquitous expression of temperature-sensitive Gal80ts but selectively induced during adulthood by exposure to 30°C for 4 h [38]. Interestingly, intracellular recordings revealed that TeTxLC expression in ACs abolished rhythmic depolarization in Ih mutant photoreceptors and led to a prolonged decay time of light-induced depolarization (Fig. 8B and S3A Fig.). By contrast, TeTxLC expression in L1 and L2 neurons failed to suppress rhythmic depolarization in Ih mutant photoreceptors (Fig. 8B and S3A Fig.). Consistently, ectopic expression of dORKΔC in ACs but not in L1L2 neurons also suppressed rhythmic depolarization in Ih mutant photoreceptors and caused a prolonged decay time of light-induced depolarization (Fig. 8C). These observations indicate that rhythmic depolarization in Ih mutant photoreceptors is due to abnormal synaptic output from ACs.
Given that ACs are likely glutamatergic neurons [16], we further tested whether rhythmic depolarization in photoreceptors is caused by uncontrolled glutamate release from ACs. Vesicular glutamate transporter (vGluT) functions in loading glutamate into synaptic vesicles and is therefore critical for synaptic glutamate output [16]. Consistently, knockdown of vGluT expression in ACs but not in L1 and L2 neurons suppressed rhythmic depolarization in Ih mutant photoreceptors (Fig. 8D and S3B Fig.). Taken together, the above observations demonstrate that rhythmic depolarization in Ih mutant photoreceptors is due to uncontrolled synaptic glutamate release from ACs.
Previous studies suggest that synaptic release may depend on Ca2+ entry via voltage-gated Ca2+ channels (VGCCs) [39–41]. Thus, we attempted to identify which VGCC contributes to the mediation of retrograde glutamate release from ACs. The Drosophila genome contains three putative homologs of vertebrate VGCCs (Ca-α1D, Gene ID: 34950; Cac, Gene ID: 32158; and Ca-α1T, Gene ID: 31550) [20]. Interestingly, introducing the cac mutant but not the Ca-α1T mutation suppressed ERG baseline oscillations in Ih mutant flies (Fig. 9A). In addition, RNAi against cac (elva-GAL4/+;Ih;UAS-cac-RNAi/+) but not Ca-α1D (elva-GAL4/+;Ih;UAS-Ca-α1D-RNAi/+) suppressed ERG baseline oscillations in Ih mutant flies (Fig. 9B). Furthermore, RNAi against cac in ACs was also able to suppress ERG baseline oscillations in Ih mutant flies (Ih; Lai-GAL4/UAS-cac-RNAi) (Fig. 9B). These results suggest that Cac channels mediate synaptic glutamate release from ACs.
Cac channels can produce high voltage-activated Ca2+ currents above −30 mV and low voltage-activated Ca2+ currents between −70 and −60 mV in vivo [42]. Ih channels are normally open at potentials more negative than −50 mV, leading to depolarization of the RMP. A recent study demonstrated that HCN1 colocalizes with Cav3.2 in presynaptic terminals and inhibits glutamate release by suppressing Cav3.2 activity [24]. Thus, it is possible that the loss of Ih channel activity hyperpolarizes the RMP of ACs to a small window (−70mV to −60 mV) in which Cac channels are active. If this is true, then we should be able to suppress Cac activity and subsequent rhythmic depolarization in photoreceptors by changing AC RMP. Previous recordings in larval muscle fibers showed that expression of the sodium channel NaChBac evoked robust voltage-gated inward currents that began to activate at approximately −60 mV and peak at approximately −30 mV [43]. Therefore, we genetically depolarized the RMP of ACs by expressing NaChBac [43]. Interestingly, its expression in ACs but not in L1L2 neurons suppressed rhythmic depolarization in Ih mutant photoreceptors (Fig. 9C). However, both expression of dORKΔC and NaChBac in ACs failed to trigger rhythmic depolarization in wild-type flies (Fig. 8C and Fig. 9C). Taken together, these findings suggest that rhythmic depolarization in Ih mutant photoreceptors is due to changes in the RMP that relieve Cac inactivation in ACs.
To identify the glutamate receptor that mediates retrograde glutamate signaling from ACs to photoreceptors, we screened 15 known ionotropic glutamate receptor (iGluR) subunits, including three conserved classes (kainate, AMPA, and NMDA types) of cation iGluRs and one chloride channel (GluClα) (Table 1) [20]. Interestingly, knockdown of ekar (CG9935, Gene ID: 43806) in photoreceptors by two individual RNAi lines, THU3080 and THU4260, suppressed ERG baseline oscillations in Ih mutant flies (Fig. 10). However, knockdown of other glutamate receptors in photoreceptors did not show any inhibition effects (Fig. 10).
We also obtained an ekar mutant allele, Mi{ET1}CG9935MB00001 [44], that does not produce ekar mRNA (Fig. 11A,B). Intracellular recordings revealed that light-induced depolarization of photoreceptors was significantly reduced in ekar mutant flies compared with wild-type flies (11.6±0.8 mV versus 4.3±0.8 mV, p < 0.001, Fig. 11C). However, EM images revealed normal rhabdomere structures in ekar mutants (S4A Fig.), and western blotting showed that ekar mutants express normal protein levels of phototransduction components (S4B Fig.). These observations suggest that EKAR might contribute to the light-evoked depolarization of photoreceptors. To further validate the role of EKAR in mediating the retrograde glutamate signal, we generated Ih;;ekar double mutant flies. Intracellular recordings revealed that ekar mutation suppressed rhythmic depolarization in Ih mutant photoreceptors (Fig. 11D,E). Given that rhythmic depolarization in Ih mutant photoreceptors was independent of phototransduction cascades activation (Fig. 3D), the suppression of rhythmic depolarization in Ih;;ekar double mutant flies indicates that the kainate receptor EKAR receives synaptic glutamate output from ACs.
To explore the potential role of this feedback regulation, we first examined whether feedback regulation is required for photoreceptor survival. EM images revealed normal rhabdomere structure in 14-day-old Ih and ekar mutant flies raised under regular light cycles (12 h light/12 h dark) or in constant darkness (S5 Fig.). These observations suggest that feedback regulation from ACs to photoreceptor terminals is not essential for photoreceptor survival.
We next performed intracellular recordings to examine photoreceptors’ excitability in response to various light intensity stimulations. The results showed that photoreceptors underwent light-induced depolarization in a light intensity-dependent manner (Fig. 12A,B). 10 Lux light stimulation evoked a 2.91 ± 0.19 mV depolarization in wild-type photoreceptors and a significantly reduced depolarization in Ih mutant photoreceptors (2.01 ± 0.11 mV versus 2.91 ± 0.19 mV; p < 0.05, Fig. 12A,B), which was indistinguishable with rhythmic depolarization without light stimulation in Ih mutant photoreceptors (Fig. 12A). In contrast, 10 Lux light stimulation triggered a significantly reduced depolarization in ekar mutant photoreceptors (0.63 ± 0.10 mV versus 2.91 ± 0.19 mV; p < 0.001) and in Lai-GAL4/UAS-TeTxLC photoreceptors (1.04 ± 0.10 mV versus 2.91 ± 0.19 mV; p < 0.001) (Fig. 12A,B). These observations indicate that feedback regulation from ACs to photoreceptor terminals facilitates photoreceptor excitability and helps maintain light sensitivity in presence of ambient light.
To investigate the potential role of this feedback regulation in visual behavior, we assessed the flies’ optomotor responses under various light conditions. We placed single flies on a circular platform and examined their ability to track moving light patterns (Fig. 12C,D). With high-intensity moving light patterns (85 and 800 Lux), Ih mutant flies but not ekar mutant or Lai-GAL4/UAS-TeTxLC flies exhibited a reduced ability to track moving patterns (Fig. 12E). However, with low-intensity moving light patterns (1 and 10 Lux), Ih mutant flies, ekar mutant, and Lai-GAL4/UAS-TeTxLC flies were less able to track moving patterns (Fig. 12E). These findings demonstrate that feedback regulation from ACs to photoreceptor terminals enhances the flies’ optomotor response in dim light conditions, whereas uncontrolled feedback regulation also disturbs motion detection in ambient light conditions (Fig. 12E).
Based on these results, we propose a feedback regulation model from ACs to photoreceptors (Fig. 12F). Photoreceptors synthesize the inhibitory neurotransmitter histamine, which is released upon light stimulation [12,45]. Thus histamine hyperpolarizes ACs by opening HisCl 2 channels [13,46,47]. Hyperpolarization of ACs activates Ih channels, which depolarizes AC RMP and limits Cac channel activity. Without Ih channels in ACs, Cac channels are activated, resulting in Ca2+ influx and subsequent glutamate release from ACs. EKAR is expressed in photoreceptor terminals and depolarizes photoreceptors upon receiving the retrograde glutamate released from ACs.
It has been shown Ih channels are expressed in several classes of interneurons that exhibit spontaneous firing activity and provide tonic inhibition to principal neurons, thus contributing to the regulation of firing frequency and excitability [48]. In this study, we revealed that loss of Ih channels in ACs results in rhythmic depolarization in photoreceptors. This phenotype was suppressed by either blocking neurotransmitter release or impairing synaptic glutamate output from ACs. Our studies provide solid evidence that feedback regulation from ACs to photoreceptors is regulated by Ih channels. Although expression of Ih channels in L1 and L2 neurons failed to suppress rhythmic depolarization in Ih mutant photoreceptors, L1/ L2 neuron-expressed Ih channels may also contribute to feedback regulation, as knockdown of Ih channels using recombinant Lai-GAL4 and L1L2-GAL4 led to enhanced ERG baseline oscillations compared with Lai-GAL4 alone. These findings are consistent with previous morphological studies showing that outer photoreceptor terminals also directly receive feedback inputs from L2 neurons [6–8].
Low-threshold Ca2+ channels are expressed in a variety of tissues such as the brain, heart, smooth muscle, kidney, and various endocrine glands [49]. These channels play important roles in controlling intracellular Ca2+ levels, modulating neuronal excitability, and regulating hormone and neurotransmitters secretion [50]. Here, we show that cac mutant flies suppressed glutamate release and subsequent rhythmic depolarization in Ih mutant photoreceptors. We further showed that rhythmic depolarization in Ih mutant photoreceptors were suppressed by changing the RMP of ACs. Our results suggest that HCN channels depolarize the RMP, thereby restricting Ca2+ entry via Cac channels and preventing glutamate release. Therefore, glutamate release is enhanced in Ih mutants, which causes rhythmic depolarization of photoreceptors. In addition, we showed that the released glutamate may induce long-lasting depolarization by opening of EKAR, which may contribute to the slow repolarization at the end of light stimulation in Ih mutant photoreceptors. A recent study reports that HCN1 channels localize in the active zone of mature asymmetric synaptic terminals and inhibit synaptic glutamate release by suppressing the activity of low-threshold voltage-gated T-type Ca2+ channels [24]. Thus, this form of regulation might be a common mechanism by which these channels modulate neuronal excitability.
Intracellular recordings from wild-type and shibireTS mutant flies reveal that cessation of all synaptic feedback to photoreceptors results in a 10–15 mV hyperpolarization shift upon light stimulation [14], suggesting that feedback regulation depolarizes photoreceptors upon light stimulation. However, the receptors that receive the excitatory neurotransmitter are still unknown. Our findings indicate that the EKAR receptor receives the glutamate signal in photoreceptor terminals: both knockdown of EKAR in photoreceptors and mutation of ekar suppressed rhythmic depolarization in Ih mutant photoreceptors. This result is supported by the previous microarray data, which showed that EKAR is highly expressed in the eye [51]. Our intracellular recording result also revealed that light-induced depolarization of photoreceptors was significantly reduced in ekar mutant flies compared with wild-type flies. The recordings further showed that Ih mutant photoreceptors undergo rhythmic depolarization with slow rise and decay times, which is consistent with the physiological properties of kainate receptors that mediate postsynaptic depolarization with slow rise and decay time [52]. Taken together, these results indicate that EKAR is expressed in photoreceptor terminals and depolarizes photoreceptors upon light stimulation.
In vertebrates, the synaptic input from horizontal cells to cone cells contributes to many visual functions including the formation of center-surround receptive fields, retinal synchronization, and light adaptation [53–55]. Fly ACs are structurally equivalent to horizontal cell in vertebrates. A recent study showed that silencing ACs reduces optomotor responses to regressive rotation stimulation, whereas activation of ACs leads to slightly decreased responses to high and low contrast stimulations [31]. Although ACs project most of their synapses to epithelial glia, they also form direct feedback synapses to photoreceptor axons [6]. In this study, we show that feedback regulation from ACs to photoreceptors improves ambient light-induced visual signal transmission and motion detection under dim light conditions, which might be important for fly activity at dawn and dusk. Conversely, uncontrolled feedback regulation in Ih mutant flies impairs visual signal transmission and motion detection in ambient light conditions, suggesting that feedback regulation is strictly modulated. In this study, we showed that Ih mutants exhibit a significantly reduced ability in tracking the moving patterns with high light intensity. Since L1 and L2 neuron play essential roles in normal motion vision [31,56], loss of Ih channels in L1 and L2 neurons may contribute to this reduced ability in motion detection.
In Drosophila photoreceptors, the rapid termination of photoresponse in Drosophila is thought to be achieved by fast deactivation of rhodopsin and calcium-mediated intrinsic feedbacks [57,58]. In addition, a slow termination of photoresponse has been reported in mutants with blocked photoreceptor transmission [59], suggesting a retrograde regulation is likely to contribute to the termination speed. In this study, Our ERG and intracellular recordings showed that blocking neurotransmitter release from ACs resulted in slow repolarization at the end of light stimulation. Similar phenotypes are observed in flies with reduced glutamate signal output from ACs and cac mutant flies [60]. Given that most of mutants with slow termination phenotype did not undergo rhythmic depolarization [57–59,61–64], the slower depolarization phenotype is not sufficient to trigger the rhythmic depolarization. These observations suggest that feedback regulation from ACs is essential for the rapid repolarization of photoreceptors at the end of light stimulation. However, repolarization speed is normal in ekar mutant flies. Because ACs form most of feedback synapses to epithelial glia [6]. Rapid repolarization of photoreceptors might be regulated by glia cells. The mechanism that facilitates rapid repolarization at the end of light stimulation need to be further investigated.
In summary, our studies reveal the molecular mechanism and physiological roles of feedback regulation from ACs to photoreceptors. This might represent a general mechanism by which feedback regulation modulates synaptic transmission and facilitates neural circuit excitability and network information processing.
Transposon piggyBac insertion flies Ihf03355 and Ihf01485 [25,26] were obtained from Harvard Medical School. The mutant alleles used for other genes in this work are CacH18 [65], PBac{WHr}Ca-α1Tdel [42], and Mi{ET1}CG9935MB00001 [44,66]. UAS-RNAi lines were ordered from Vienna Drosophila RNAi Center and Tsinghua RNAi Stock Centre. L1-GAL4, L2-GAL4, and L1/L2-GAL4 lines [33,34] were obtained from Dr. Jens Rister, and the split Lai-GAL4 line (R92A10AD attP40; R17D06DBD attp2) [31] was provided by Dr. Aljoscha Nern. Other lines used in this work were obtained from the Bloomington Stock Center. The wild-type flies used in this study were w1118. All flies were maintained in standard medium at 25°C, with 60%–80% relative humidity. Less than 3-day-old flies were used for ERG recording, and 3-day-old flies were used in optomotor response assays. In all experiments, an equal number of male and female flies were used. Full genotypes of samples shown in main figure panels are provided in S1 Table.
Anti-Ih antibodies were generated in rabbits against a purified glutathione S-transferase fusion fragment (aa332–618) of Ih-PK protein and generated by GenScript (Nanjing, China). An affinity column, generated by coupling a MBP-Ih fragment (aa332–618) to Sepharose 4B, was used to purify the antibody. The sources of other antibodies were rabbit anti-TRP [67], rabbit anti-Arr2 [61], rabbit anti-INAD [63], anti-PLC antibodies [68], anti-GFP (Abcam), and mouse anti-24B10 [69] and anti-Rh1 (4C5) (DSHB).
Electroretinogram (ERG) recordings were conducted at 25°C as previously described [70]. Less than 3-day-old flies were collected, immobilized with strips of tape and put in darkness for 5 min for adaptation. Two glass microelectrodes were filled with Ringer’s solution and placed on the thorax and compound eye. Flies were stimulated with 5-s light pulses (4000 Lux) every 25 s using a Newport light projector. For each fly, ERG recording lasted for more than 100 s. The signal was amplified and recorded using a Warner IE210 Intracellular Electrometer. In ERG and intracellular recording without light stimulation, more than five continuous depolarizations with amplitude >1 mV was defined as an ERG baseline oscillation phenotype. The fraction of flies that exhibited ERG oscillation was quantified and presented in the figures.
Intracellular recordings were performed as described previously [63]. Briefly, flies were fixed with strips of tape, and a small opening was made on surface of the eye using fine tweezers. A low resistance (>30 MΩ) glass microelectrode filled with 2 M KCl was gradually inserted into the opening until light-induced membrane depolarization was observed. A reference electrode was filled with Ringer’s solution and placed inside the eye at the retina layer. The signal was amplified and recorded using a Warner IE210 Intracellular Electrometer. The fractions of photoreceptors exhibiting rhythmic depolarization in each genotype were calculated. To quantify the amplitudes of light responses, 10 photoreceptors from 10 flies were measured for each genotype, and the mean±SEM was calculated and showed in figures.
Western blotting was carried out as previously described [71]. Fly heads were homogenized in SDS-sample buffer. The proteins were fractionated by SDS-PAGE and transferred to PVDF membranes (Pall) in Tris-glycine buffer. After blocking, the blots were probed with anti-Rh1 antibody (1:3,000 dilution), rabbit anti-Arr2 antibody (1:1,000 dilution), rabbit anti-INAD antibody (1:1,000 dilution), anti-Ih antibody (1:200), rabbit anti-TRP antibody (1:1,000), anti-PLC antibody (1:1,000) at RT for 2 h. After three washes with PBS, the blots were subsequently probed with either anti-rabbit or mouse IgG-peroxidase conjugate (GE Healthcare) at RT for 1 h, and the signals were detected using ECL reagents (Amersham Biosciences)
To separate the retina and lamina from the brain for western blotting analysis, fly heads were cut and immersed in 100% ethanol for 2 h before the retina and lamina were carefully dissected from the brain. Separated tissues (retina and lamina, the head without retina and lamina) were homogenized in SDS-sample buffer. To isolate the retina and lamina for RT-PCR analysis, they were dissected carefully, and the mRNA was extracted from the separated tissues using TRIzol reagent (Invitrogen).
Section staining was performed as previously described [61]. Fly heads were fixed with 4% paraformaldehyde. After three washes, the fly heads were dehydrated with acetone and embedded in LR White resin, and 1-μm cross-sections were cut across the top half of the eye. The sections were incubated in diluted primary antiserum (Rh1, 1:200; INAD, 1:400; TRP, 1:400) at room temperature for 1 h. After three times of washing in PBS, sections were incubated with diluted secondary antibodies at room temperature for 1 h. The stained sections were examined under a ZEISS Axio observer A1 microscope.
Whole-head staining was performed to locate endogenous Ih channel in wild-type adult flies. Fly heads were dissected in PBS buffer and fixed with 4% paraformaldehyde in PBS buffer. After fixation, the heads were double labeled with diluted primary antiserum (anti-Ih antibodie 1:50 and 24B10 1:100 or anti-GFP 1:200). After three washes in PBS buffer, fly heads were incubated with diluted secondary antibodies at room temperature for 1 h [71]. After three additional PBS washes, the stained heads were examined under an LSM 700 confocal microscope.
EM was carried out as previously described [72]. Fly heads were fixed at 4°C for 12 h in 2.5% gluteraldehyde, 0.1 M sodium cacodylate (pH 7.2). After three washes with 0.1 M sodiumcacodylate, the heads were stained with 1% osmium tetroxide at room temperature for 1 h. After a standard ethanol dehydration series, the heads were immersed in propylene oxide for two 10-min washes before they were embedded with standard procedures. Thin sections (100 nm) were cut at the top 2/3 of retina to show ommatidia whereas cut at half the lamina to display cartridges. Sections were collected on Cu support grids and stained with uranyl acetate for 8 min, followed by 5 min with lead citrate. Micrographs were taken at 80 kV on a Hitachi-7650 transmission EM.
Fly optomotor responses were tested as previously described [73]. Briefly, 3-day-old flies were collected and their wings were cut off. After recovering for more than 24 h in a 12-h light/12-h dark cycle, a single fly was placed on a circular platform for the optomotor response test. The platform is surrounded with a water-filled moat to prevent the fly from escaping, and the moat was surrounded with a panoramic LED display that controlled by LED Studio software (Shenzhen Sinorad Medical Electronics). Bright and dark stripes were used to generate a clockwise motion light for 90 s followed by an anti- clockwise motion light for another 90 s (180°/s corresponding to a temporal frequency of 4 Hz). The walking traces of flies were recorded by a camera (WV-BP330, Panasonic System Networks), and the data were analyzed in MATLAB (Mathworks). The flies’ optomotor responses were quantified by the performance index of tracking time (PITT). The PITT is defined as (Ptracking time—Pun-tracking time)/ (Ptracking time + Pun-tracking time). The probabilities of fly movement in the platform in accordance with LED rotating direction or not are defined as tracking or un-tracking, respectively. The male and female flies were used alternately.
Statistical analysis was done by using MS Excel. The numerical data used in all figures are included in S1 Data. For quantitative data of ERG and intracellular recordings, fraction of flies or photoreceptors that exhibit oscillations and standard error of rate are shown. Fisher’s exact probability tests were used to compare genotypes. For statistical analysis of depolarization amplitudes and decay time of depolarization in Figs. 2C, 8B, 8D, 11C, and 12B, data are presented as mean ± SEM. Two-tailed Student’s t tests were used to compare genotypes. Significance was classified as follows: *, p ≤ 0.05; **, p < 0.01; ***, p < 0.001; n.s. p > 0.05.
|
10.1371/journal.pgen.1006961 | Positional cloning of quantitative trait nucleotides for blood pressure and cardiac QT-interval by targeted CRISPR/Cas9 editing of a novel long non-coding RNA | Multiple GWAS studies have reported strong association of cardiac QT-interval to a region on HSA17. Interestingly, a rat locus homologous to this region is also linked to QT-intervals. The high resolution positional mapping study located the rat QT-interval locus to a <42.5kb region on RNO10. This region contained no variants in protein-coding sequences, but a prominent contiguous 19bp indel polymorphism was noted within a novel predicted long non-coding RNA (lncRNA), which we named as Rffl-lnc1. To assess the candidacy of this novel lncRNA on QT-interval, targeted CRISPR/Cas9 based genome-engineering approaches were applied on the rat strains used to map this locus. Targeted disruption of the rat Rffl-lnc1 locus caused aberrant, short QT-intervals and elevated blood pressure. Further, to specifically examine the significance of the 19bp polymorphism within the Rffl-lnc1 locus, a CRISPR/Cas9 based targeted knock-in rescue model was constructed by inserting the 19bp into the strain which contained the deletion polymorphism. The knock-in alleles successfully rescued the aberrant QT-interval and blood pressure phenotypes. Further studies revealed that the 19bp polymorphism was necessary and sufficient to recapitulate the phenotypic effect of the previously mapped <42.5kb rat locus. To our knowledge, this study is the first demonstration of a combination of both CRISPR/Cas9 based targeted disruption as well as CRISPR/Cas9 based targeted knock-in rescue approaches applied for a mammalian positional cloning study, which defines the quantitative trait nucleotides (QTNs) within a rat long non-coding RNA as being important for the pleiotropic regulation of both cardiac QT-intervals and blood pressure.
| Diseases of the cardiovascular system such as essential hypertension do not have a clear cause, but are known to run in families. The inheritance patterns of essential hypertension and other cardiac diseases suggest that they are not due to a single defective gene but instead are caused by multiple genetic defects that are inherited together in a patient. This complex inheritance makes it difficult to pinpoint the underlying defects. Here, we describe a panel of genetically-engineered rats, using which we have discovered a novel gene, which does not code for any protein, as a gene required for maintenance of normal blood pressure. Structural defects within this non-coding RNA cause hypertension and cardiac short-QT interval. Further, by performing genome surgery to correct the gene defect, we demonstrate the precise error in nucleotides that was inherited and caused hypertension and cardiac short-QT interval syndrome.
| It is estimated that hypertension affects nearly 75 million Americans (about 1 in every 3 U.S. adults) [1]. Essential hypertension is the most common type of hypertension and remains a major risk factor for cardiovascular diseases, such as cardiomyopathy [2], coronary artery diseases [3] and peripheral vascular diseases [4]. However, essential hypertension is of unknown origin and is characterized as a multifactorial disease involving genetic and environmental factors [5]. Familial and twin studies show that 30%-50% of the phenotypic variation of blood pressure (BP) is attributable to genetic heritability [6]. This implies that the contributions of genetic determinants to the development of hypertension are significant and elements on our genome may predispose some people to develop hypertension [7].
Over the past 50 years, several animal models of essential hypertension, predominantly in the rat, have been developed as valuable tools to study the genetic factors associated with hypertension [8]. One such tool generated from our laboratory is the inbred Dahl salt-sensitive (S) rat. The S rat develops hypertension even on a low-salt diet but develops more severe hypertension when fed with a high-salt diet. Using this rat strain, we and others have applied classic genetic approaches of linkage followed by substitution mapping to locate regions of its genome as quantitative trait loci (QTLs) that are inherited causes of hypertension [9–19]. As relevant to the current study, we have previously mapped one such BP QTL on rat chromosome 10 by linkage followed by the construction and characterization of a custom series of congenic strains which contained introgressing genomic segments of normotensive Lewis rat (LEW) on the genetic background of the S rat. The mapped locus was within a <42.5kb region and reported as a quantitative trait locus (QTL) for BP as well as cardiac QT-interval [20–25] (Fig 1A). LEW alleles within the <42.5kb region significantly shortened QT-interval and increased blood pressure of the hypertensive S rat [20]. Interestingly, a large meta-analysis of three genome-wide association studies (GWAS) using 13,685 individuals reported that the region homologous to the rat 42.5kb region in humans, which lies on human chromosome 17, has multiple minor alleles that are reportedly associated with shorter QT-intervals [26] (Fig 1B). Of notable interest, nearly 30% of individuals in the GWAS study also had hypertension [26]. A second GWAS further confirmed the association of this locus to QT-interval [27]. Collectively, these observations suggest that the critical region in focus for the current report is of significance in the cardiovascular health of two mammalian species, the rat and human.
The <42.5kb critical region in rats contains a single protein coding gene, Rffl, which is without any exonic variants. The region also contains a novel long non-coding RNA (lncRNA), named Rffl-lnc1, located within Rffl 5’-UTR intronic region. Rffl-lnc1 harbors a 19bp indel polymorphism between the S (+19bp) and the S.LEW congenic strain (-19bp), which were the two strains used to map this locus (Fig 1C). Based on this observation, we hypothesized that Rffl-lnc1 is a genetic determinant of QT-interval and blood pressure.
To test this hypothesis, using the CRISPR/Cas9 technology, a panel of Rffl-lnc1 disruption models was developed on the genomic background of the Dahl S rat. These models harbored varied disruptions around the critical 19bp region. The disruption of Rffl-lnc1 significantly shortened QT-interval and increased blood pressure of the S rat, suggesting an important role of Rffl-lnc1 in regulating cardiovascular function. To further evaluate the specific effect of the 19bp indel polymorphism within the Rffl-lnc1 locus, a 19bp knock-in rescue model was developed on the genomic background of the S.LEW congenic strain using the CRISPR/Cas9 technology. The 19bp insertion successfully corrected the aberrant short QT-interval phenotype and lowered blood pressure of the S.LEW congenic strain, demonstrating that the 19bp indel polymorphism within Rffl-lnc1 is an inherited genetic variation responsible for regulating cardiovascular disease in the rat. Further, our study has demonstrated that among all the variants located within the <42.5kb QTL region, the 19bp polymorphism was sufficient to regulate both QT-intervals and blood pressure. Overall, this study is the first to precisely define the quantitative trait nucleotides within a long non-coding RNA as a genetic determinant of cardiovascular function and is also the first to apply both gene-disruption and knock-in strategies using the CRISPR/Cas9 based genome editing approaches for delineating a complex cardiovascular trait locus in a mammalian model.
To disrupt Rffl-lnc1 on the genomic background of the Dahl S rat, a custom gRNA, rRffl.g4, was designed to target the 19bp containing genomic segment. In vitro validation of rRffl.g4 using mismatch detection assay confirmed its target efficiency (S1 Fig). Microinjection of gRNA and Cas9 mRNA into single cell embryos of the S rat followed by implantation into 6 pseudo-pregnant females resulted in a total of 67 pups. Genotyping and sequencing data showed that 21 out of these 67 pups were mutants within the Rffl-lnc1 locus with disruptions both within and outside of the 19bp critical region. We used 4 founders with different deletions occurring within the 19bp locus for subsequent phenotypic studies (Fig 2A).
All 4 Rffl-lnc1 disruption models demonstrated elevated systolic, diastolic and mean arterial pressures compared to wild-type hypertensive S rats (Fig 2B–2M). Interestingly, these disruption models exhibited different levels of BP increasing effects (Fig 2B–2M). The heart/body weight ratios were also higher in Rffl-lnc1 disruption models (S2 Fig), suggesting BP associated cardiac hypertrophy and potential dysfunction. To further assess cardiac function, we focused on QT-interval because shorter QT-intervals were reported to be associated with alleles within the homologous segment in humans as well as observed in our previous high resolution positional mapping study in rats [20, 26]. As shown in Fig 3A and 3B, the QT-intervals of the gene-edited Rffl-lnc1 model were significantly shorter than that of the S rat. Collectively, these results demonstrate that Rffl-lnc1 is a potential genetic determinant of blood pressure and QT-interval.
Since the annotation for the novel Rffl-lnc1 was limited to a few base-pairs, we performed RACE experiments to ascertain its full sequence. 5’RACE amplifications using the primer P1 (Fig 4A) and Universal Primer A Mix (UPM) for the initial PCRs followed by nested PCR amplification using the primer P2 (Fig 4A) and Universal Primer Short (UPS) resulted in four 5’RACE products, labeled as a, b, c and d, in Fig 4B. Unlike 5’RACE, Fig 4C shows the unique 3’RACE product in lane 5, which was obtained using the primer P3 (Fig 4A) and UPM for the initial PCR followed by nested PCR amplification using P4 (Fig 4A) and UPS. Further characterization of these PCR products by sequencing confirmed the existence of four different isoforms of Rffl-lnc1, each with a different 5’ end. Each isoform contained a single-exon of more than 3000bp (Fig 4D). The secondary structures of these isoforms of Rffl-lnc1 were predicted using RNAfold Webserver [28] (Fig 5). The 4 isoforms of Rffl-lnc1 showed different secondary structures in the wild-type S rat (Fig 5A–5D). Interestingly, it was observed that sequence deletions around the critical 19bp region of Rffl-lnc1 caused a range of perturbations of the secondary structures, such as double helices, internal loops and stem loops, of all the four isoforms in each of the gene-disruption models (Fig 5E–5Q). The most deleterious perturbation was that in Rffl-lnc1 disruption model 4, wherein, due to the large deletion of the 5’ end of Rffl-lnc1, the secondary structural integrity was lost in all the isoforms except one (Fig 5Q). Secondary structures of Rffl-lnc1 in all the disruption models appeared to correlate well with physiological impact. For instance, the secondary structure of Rffl-lnc1 transcript 1 was drastically altered in Rffl-lnc1 disruption model 2 compared to other models and the S rat (Fig 5A, 5E, 5I, 5M and 5Q), which correlated with a dramatic BP increasing effect observed in Rffl-lnc1 disruption model 2 compared to other models (Fig 2). The correlation indicates a potential link between lncRNA structure and its physiological impact.
The above evidence obtained with gene-disruption models, albeit strong, does not directly test causality for the 19bp as the naturally occurring quantitative trait nucleotides within Rffl-lnc1 affecting cardiovascular function. To directly evaluate the contribution of the 19bp indel polymorphism on cardiovascular function, we further used the CRISPR/Cas9 system to generate a targeted knock-in rescue model by precisely inserting the 19bp into the Rffl-lnc1 locus of the S.LEW congenic strain. A total of 73 pups were born after the microinjection of rRffl.g4, Cas9 mRNA and donor oligonucleotide into single cell embryos of the S.LEW congenic strain, followed by implantation into 7 pseudo-pregnant females. Genotyping and sequencing validations identified 3 successful 19bp knock-in founders (Fig 6A–6C).
Knock-in rescue rats exhibited significantly lower systolic, diastolic and mean arterial pressures compared to wild-type S.LEW congenic rats (Fig 6D). Echocardiac evaluation demonstrated that knock-in rescue rats tended to have lower relative wall thickness and exhibited significantly improved cardiac function and contractility as evidenced by significantly lower MPI and increased FS/MPI, respectively (S1 Table). Moreover, short QT-intervals were also improved in targeted knock-in rescue rats compared to wild-type congenic rats (Fig 6E). Fig 7 catalogs the structural differences of the four Rffl-lnc1 transcripts between targeted knock-in rescue rats and wild-type S.LEW congenic rats, which reflect the direct rescue status of the transcripts. These results demonstrate that the 19bp indel polymorphism is specifically responsible for functioning as quantitative trait nucleotides within four isoforms of a long non-coding RNA that are involved in cardiovascular regulation.
The critical <42.5kb QTL region contains 171 variants including the continuous 19bp variation [20]. To evaluate whether the 19bp are sufficient to regulate cardiac function, we further compared the phenotypes between S and targeted knock-in rescue rats. Interestingly, knock-in rescue rats demonstrated no differences in blood pressure and QT-interval compared to S rats (Fig 8). No significant differences in echocardiographic parameters were seen (S2 Table). These results demonstrate that the other variants within the QTL region are not contributing to the QTL effect and importantly, that the 19bp indel polymorphism within the previously resolved <42.5kb QTL region is necessary and sufficient to demonstrate the full effect of the QTL region independent to the other allelic variations within the QTL segment.
The search for inherited factors responsible for blood pressure and cardiovascular traits has been dominated by genome wide-association studies in humans [29–37] and positional mapping approaches in model organisms [38, 39]. The former has resulted in the identification of hundreds of genomic variants and the latter has been successful in mapping genomic segments of mammalian model organisms, predominantly the rat. Data from both of these approaches is consistently pointing to variants in non-coding elements as being the most prevalent signals for associations to blood pressure. Beyond such associations, there is a wide gap in our understanding of the precise identities of the allelic variants of non-coding elements and their functional link to cardiovascular health and disease.
The present work, which is backed by persistent and systematic mapping [20–25], for the first time, identifies a 19bp indel polymorphism as the precise variation within a novel lncRNA, Rffl-lnc1, which is necessary for BP regulation. This is also the first study wherein the state-of-the-art genome engineering using the CRISPR/Cas9 based genome targeted rescue strategy is applied beyond a mapping approach to identify quantitative trait nucleotides responsible for cardiovascular effects.
The impetus for focusing on non-coding elements within our previously mapped <42.5kb genomic region, was the observation that the region contained only a single protein-coding gene, Rffl (Fig 1A), which did not contain any exonic variants. More interestingly, this short <42.5kb genomic segment was also involved in the regulation of QT-intervals [20], which further translationally supported a large meta-analysis of human GWAS on its homologous genomic region on human chromosome 17 [26] (Fig 1B).
There were a total of 171 variants to consider within the QTL region [20], among which a long stretch of contiguous 19bp indel polymorphism caught our attention. At this point, we were also developing a first-generation catalog of rat lncRNAs involved in BP regulation [40], through which we identified a novel lncRNA, named Rffl-lnc1. As a happenstance, the contiguous 19bp indel polymorphism overlapped the available annotation for Rffl-lnc1 (Fig 1C). Therefore, we chose to prioritize the 19bp indel polymorphism and focused on testing the hypothesis that this particular variation is responsible for the QTL effect. In obtaining positive results that prove our hypothesis, our present study has: (1) Further improved the high resolution mapping from the <42.5kb QTL to 19bp; (2) Eliminated the remainder of the QTL region as not independently contributing to the QTL effect; (3) Defined the QTL as being due to the quantitative nucleotide variation within a novel, functional long non-coding Rffl-lnc1; (4) Determined the function of Rffl-lnc1 as being important for two cardiovascular complex traits, i.e., blood pressure and cardiac QT-interval; and (5) Identified the involvement of at least 4 isoforms of Rffl-lnc1 with disparate secondary structures to be considered as being important for further functional cardiovascular regulatory studies.
Considering that data from the targeted rescue model is sufficient evidence to assign causality for the 19bp indel polymorphism, it may appear that the gene-disruption models were not that important. On the contrary, the disruption models were quite informative. The targeted disruption models of Rffl-lnc1 served as important evidence for suspecting the role of Rffl-lnc1 in regulating cardiovascular functions. We also predicted the secondary structures in the disruption models and surprisingly, large structural changes were observed in Rffl-lnc1 transcripts of different models due to different sequence deletions occurring around the 19bp locus (Fig 5). For example, Rffl-lnc1 transcript 1 in Rffl-lnc1 disruption model 2 is drastically different compared to that in other three disruption models and the S rat (Fig 5A, 5E, 5I, 5M and 5Q), which corresponded to much higher BP increasing effects in the model 2 compared to the other 3 models (Fig 2). Due to the large deletion occurring in the 5’ end of Rffl-lnc1 in Rffl-lnc1 disruption model 4, only one isoform appears to remain intact in this model (Fig 5Q). Interestingly, the insertion of only 19bp in the targeted rescue model caused discernible structural modifications, such as double helices, internal loops and stem loops, in all the transcripts (Fig 7). The correlation between molecular lncRNA isoform structures and their physiological effects further provides evidence to the point that different lncRNA isoforms may have varied physiological impact. Whether all or only some of these 4 Rffl-lnc1 transcripts are required for the observed physiological effects remains to be determined. It is also unknown whether additional, yet undetected, Rffl-lnc1 transcripts exist. Nevertheless, the role of the identified 19bp indel polymorphism is clearly established as being functional at least through one (or more) of the isoforms of Rffl-lnc1. Also, these first-generation genome-engineered targeted disruption models for a lncRNA serve as excellent tools for further studies.
The most rigorous test for assigning the ‘quantitative trait nucleotide’ status to an allelic variant on the genome is to demonstrate a direct cause-effect relationship between the allelic variation and an alteration in a physiological trait using a targeted rescue approach in a model organism [38, 41]. To apply this level of rigor was impossible in rat models until the CRISPR/Cas9 technology became applicable to the rat model. Therefore, we have taken advantage of this technological advancement to further test whether the 19bp served as quantitative trait nucleotides. Our results demonstrate that by restoring Rffl-lnc1 with the 19bp insertion, the rescue model lowered hypertension and corrected the short QT-interval phenotype. Thus, the 19bp indel polymorphism is hitherto defined as quantitative trait nucleotides for cardiovascular regulation of blood pressure and cardiac QT-intervals.
The next ambiguity pertained to the contribution of the remainder of the QTL region, because the experimental design of comparing the targeted rescue model, which was developed on the genomic background of the S.LEW congenic strain, with the congenic strain as the control strain, was not informative for the contributions, if any, of the remainder of the variants within the QTL region. To overcome this ambiguity, we used the experimental design of comparing the targeted rescue model with the S rat, whereby, the precise contribution of the remainder of the QTL region (without the 19bp variation) could be assessed. The result from this study wherein there were no phenotypic differences between the targeted rescue model and the S rat (Fig 8), suggests two possibilities. The first possibility is that other variants are not important and the 19bp play an exclusive role in cardiovascular regulation. The second possibility is that other variants may be in epistasis, whereby they may not be able to exert their effects independent of the 19bp. Either way, this provides evidence to indicate that the 19bp polymorphism is critical for the QTL effect and that the remainder of the QTL region is relatively insignificant in independently contributing to the QTL effect.
Although our study identifies QTNs for arterial blood pressure, it does not specifically address whether the QT interval shortening in the Dahl S and the rescue with the transgenic 19bp insertion is a direct consequence of changes in the cardiac conduction system, or due to a primary modification of neural autonomic pathways that alter electrogenic cardiac functions, or due to the result of cardiac hypertrophy as a secondary consequence to the chronic elevation of BP (which is well known to alter electrical conduction intervals [42, 43]). Teasing apart these coupled features requires combinatorial experimental designs of using the genetic models developed through our study and pharmacological approaches.
Our previous study provided evidence for this <42.5kb QTL region as also being important for the regulation of tumorigenesis [44], therefore our future study will further investigate the roles of Rffl-lnc1 and the 19bp within it in the process of tumorigenesis using our disruption and rescue models. As we mentioned earlier, a homologous locus on human chromosome 17 of this <42.5kb critical region on rat chromosome 10 was studied in a large meta-analysis of human GWAS, showing multiple alleles near human RFFL gene were associated with the short QT-interval syndrome [26] (Fig 1B) as confirmed by our congenic and transgenic rat models. Although hypertension was not pointed out to be associated with this locus in humans, it is reported that nearly 30% of individuals in this GWAS study had hypertension [26]. Therefore, our study serves as the basis to consider hypertension as an important co-segregating phenotype along with QT-interval in humans. Further the delineation of a lncRNA as the quantitative trait gene within the rat locus prompts the consideration of a similar lncRNA within the homologous human region with reported association for QT-interval. To this point, it is interesting to note that a human lncRNA does exist within the 5’-UTR intronic region of the human RFFL gene (Fig 1B). Given the rat data, our study may serve as a translational foundation for considering this human lncRNA as a candidate regulator for cardiovascular diseases.
This study has contributed to the advancement of QTL mapping in the rat for cardiovascular phenotypes in general by pinpointing the quantitative trait nucleotides underlying a QT-interval and a BP QTL. It is also the first report of a polymorphism detected within a long non-coding RNA as a candidate gene verified within a mammalian QTL. The translational significance of the study is that it provides additional confirmatory evidence for the homologous region in humans detected to be associated with QT-interval. Due to the lack of sequence conservation between rats and humans, the precise polymorphisms within the homologous Rffl-lnc1 locus may or may not exist in humans. The relevance of this positional cloning study is therefore that the results obtained by mapping a locus in the rat provide a functional basis to assess similar effects of variants within the lncRNA as candidates within the homologous region in humans reported by GWAS for QT-interval.
All animal procedures and protocols described in this study were approved by the University of Toledo Institutional Animal Care and Use Committee. Animal experiments were performed in accordance with the Guide for the Care and Use of Laboratory Animals. The inbred Dahl salt-sensitive (SS/Jr or S) rat strain was from stocks maintained in our animal facility at our institution. Rats were weaned at 28–30 days of age and fed with a low-salt diet (0.3% NaCl, TD 7034, Harlan Teklad). High-salt diet (2% NaCl, TD 94217, Harlan Teklad) was used for experiments involving a high-salt regimen. Only male rats were used for the current study, in order to match the blood pressure QTL inference drawn from the previous study [20] conducted using male rats. In each phenotypic study, any different experimental rat groups were concomitantly bred and co-housed to minimize environmental effects.
Guide RNAs (gRNAs) were designed to target the 19bp locus within Rffl-lnc1 (Genome Engineering and iPSC Center, Washington University, St. Louis, MO). Bioinformatics analysis was performed to detect potential off-target sites of all gRNA candidates on the rat genome. The gRNA, rRffl.g4 (AAGCCATGGAGTTAGGCCATNGG), which had minimum off-target potential based on homology, was further validated in rat C6 cells. The gRNA, rRffl.g4, was chosen for the transgenesis in generating both disruption and knock-in rescue models. Additionally, a donor oligonucleotide (CACCACCCCAGCAGCTCCTGTTGAGCACTGCAGCGGCCTCATCCATGTGACAGGCCTGACGCCCTCACGCAGGTCTGGCCTATGGCCTAACTCCATGGCTTTCCAAGTGCTGGAAGTTCCCCAGGCGACATTCAGTGTC), which contains the 19bp sequence, was designed for the knock-in rescue model.
Oocyte microinjection was conducted at the University of Michigan Transgenic Animal Model Core (Ann Arbor, MI). For the disruption model, a mixture of rRffl.g4 (2.5 ng/μl) and Cas9 mRNA (5 ng/μl) was injected into one-cell stage Dahl salt-sensitive (S) rat embryos. Microinjected embryos were implanted into 6 pseudo-pregnant Sprague-Dawley female rats and a total of 67 pups were born. For the knock-in rescue model, a mixture of rRffl.g4 (2.5 ng/μl), Cas9 mRNA (5 ng/μl) and the donor oligonucleotide (10 ng/μl) was injected into one-cell stage S.LEW congenic strain embryos. Microinjected embryos were implanted into 7 pseudo-pregnant Sprague-Dawley female rats and a total of 73 pups were born. At 14 days of age, tail tip biopsies were collected from transgenic pups for extracting genomic DNA. Three different primer sets were used for initial genotyping. The forward (F) and reverse (R) sequences of these three primer sets (1, 2, 3) are: 1-F: AGCAGCTCCTGTTGAGCACT; 1-R: GAACTTCCAGCACTTGGAAAGC; 2-F: ACTGCCCTGAACCAAACCTG; 2-R: ACTTGGAAAGCCATGGAGTTAG; 3-F: ATGCAGACGATTTCTGACAGC; 3-R: ATCCCTGAGGGCTTTTCTACA. Due to large deletions in Rffl-lnc1 disruption model 4, the forward (ATGCAGACGATTTCTGACAGC) and reverse (GGTCTTCACTCTCCAGAATATG) primers were used for further genotyping. After breeding all the potential founders to homozygotes, the PCR products of the genotyping from the homozygotes were sent for sequencing validation (Eurofins MWG Operon, https://www.eurofinsgenomics.com/en/home.aspx) and sequencing data was analyzed using Sequencher 4.10.1. The homozygotes of disruption and knock-in models were used for subsequent phenotypic studies.
Blood pressure was recorded and analyzed using radiotelemetry transmitters (HD-S10 or previously C40), receivers and software from Data Sciences International, as described previously [21]. The specific details on the age of the rat and type of diet used in each study are provided in the legend to each figure.
ECG data was collected and analyzed using CTA-F40 transmitters, receivers and software from Data Sciences International. Briefly, the transmitters were surgically implanted into the peritoneal cavity of rats under anesthesia and transmitter electrodes were arranged in Lead II configuration. ECG data was collected at 5-minute intervals and analyzed using Ponemah v.5.2 (Data Sciences International). Bazett’s formula was used as the standard correction method for normalizing QT-intervals specifically for rats. The specific details on the age of the rat and type of diet used in each study are provided in the legend to each figure.
Total RNA was extracted from heart tissues of the Dahl S rat using the TRIzol reagent (Life Technologies) according to the manufacturer’s instructions. The integrity and concentration of the RNA was assessed by gel electrophoresis and NanoDrop 2000 (Thermo Scientific). 5’RACE and 3’RACE procedures were performed according to the SMARTer RACE 5’/3’ Kit (Clontech) protocol. Briefly, about 5μg RNA was used for making 5’RACE and 3’RACE cDNA, respectively. For 5’RACE amplification, P1 (GATTACGCCAAGCTTACCCCAGCAGCTCCTGTTGAGCACT) and Universal Primer A Mix (UPM) were used for initial PCR amplification according to Program 1 (touchdown PCR) in the protocol. Using the diluted (50X) PCR product from the previous step, P2 (GATTACGCCAAGCTTTGGGCACAATAGCTTGGCTTTTATGGAC) and Universal Primer Short (UPS) were used for the nested PCR according to Program 2 in the protocol to obtain the 5’RACE products for the following characterization. For 3’RACE amplification, P3 (GATTACGCCAAGCTTAACCATTCAGGAAGCCACAGGCCTTCC) and UPM were used for initial PCR amplification according to Program 1 (touchdown PCR) in the protocol. Using the diluted (50X) PCR product from the previous step, P4 (GATTACGCCAAGCTTGTCCCGCCTTCCTATTTTCCAGATGAGG) and UPS were used for the nested PCR according to Program 2 in the protocol to obtain the 3’RACE product for the following characterization. The 5’RACE and 3’RACE products were further characterized following the steps of gel extraction and in-fusion cloning in the protocol. The cloned inserts were PCR amplified and sent for sequencing (Eurofins MWG Operon, https://www.eurofinsgenomics.com/en/home.aspx) and sequencing data was analyzed using Sequencher 4.10.1.
Left ventricular function and geometry of Dahl S rats, S.LEW congenic rats and 19bp targeted knock-in rescue model were evaluated by echocardiography, as described previously [45, 46]. The specific details on the age of the rat and type of diet used in each study are provided in S1 and S2 Tables.
Two-tailed Student’s t-test was used for statistical analyses. Data are presented as mean ± SEM. A p-value of <0.05 was considered to be statistically significant.
|
10.1371/journal.pcbi.1003052 | Long-range Order in Canary Song | Bird songs range in form from the simple notes of a Chipping Sparrow to the rich performance of the nightingale. Non-adjacent correlations can be found in the syntax of some birdsongs, indicating that the choice of what to sing next is determined not only by the current syllable, but also by previous syllables sung. Here we examine the song of the domesticated canary, a complex singer whose song consists of syllables, grouped into phrases that are arranged in flexible sequences. Phrases are defined by a fundamental time-scale that is independent of the underlying syllable duration. We show that the ordering of phrases is governed by long-range rules: the choice of what phrase to sing next in a given context depends on the history of the song, and for some syllables, highly specific rules produce correlations in song over timescales of up to ten seconds. The neural basis of these long-range correlations may provide insight into how complex behaviors are assembled from more elementary, stereotyped modules.
| Bird songs range in form from the simple notes of a Chipping Sparrow to the complex repertoire of the nightingale. Recent studies suggest that bird songs may contain non-adjacent dependencies where the choice of what to sing next depends on the history of what has already been produced. However, the complexity of these rules has not been examined statistically for the most elaborate avian singers. Here we show that one complex singer—the domesticated canary—produces a song that is strongly influenced by long-range rules. The choice of how long to repeat a given note or which note to choose next depends on the history of the song, and these dependencies span intervals of time much longer than previously assumed for birdsong. Like most forms of human music, the songs of canaries contain patterns expressed over long timescales, governed by rules that apply to multiple levels of a temporal hierarchy. This vocal complexity provides a valuable model to examine how ordered behaviors are assembled from more elementary neural components in a relatively simple neural circuit.
| Brains build complex behaviors from simple modules [1], [2].The ultimate example is speech where sequences of phonemes form words that in turn are rearranged to form sentences. So too, the complex performances of a musician or swordfighter are composed of discrete motor gestures that may be composed of more elementary motor modules or muscle synergies [3]. Songbirds, in their own ways, build complex vocal forms from elementary units known as syllables. Among the 4500+ species of songbirds, simple and complex songs can be found, and a rich history of detailed song descriptions can be found across a wide variety of literature [4]–[12]. However, quantitative information about the statistical complexity of song is available only for a few species [4], [8]–[12]. Birdsong has often been described in terms of first-order transition statistics, e.g. between adjacent syllables [13] in the zebra finch or syllable chunks [9], [14] in the nightingale and Bengalese finch. However, analysis of the Bengalese finch song also reveals non-adjacent dependencies where transition probabilities between syllables depend not only on the current active syllable, but also one or more prior syllables sung [15]–[17]. Formally, this implies that song syntax must be modeled with a second-order or higher order Markov chain. Higher-order Markov chains can also be represented through first-order statistics in a hidden Markov model (HMM). In the latter case, statistically complex sequences will require a large number of hidden states, relative to the number of observed syllables.
In addition to the detailed quantitative studies of syntax in Bengalese finches in laboratory settings, many field studies have described an array of influences on the delivery of song. For example, some species such as the swamp sparrow engage in antiphonal song type-matching—selecting a song that best matches what the neighbor just sang [18]. In this case, an auditory stimulus is involved in the selection of elements from a vocal repertoire, and the choices are not simply determined by the current motor state [19]. Other examples of complex vocal behavior can be found for species that sing many song types. In some cases, e.g. with Western Meadowlarks and American Redstarts, the probability of producing a given song type decreases after the first time it is delivered in a bout of singing, and as a result, the full repertoire of songs is expressed more frequently than expected if the selection of songs was random [4], [12]. In a related example of song performance memory, nightingales can pause for a few seconds, and then resume singing where they left off in a ordered set of songs [20]. Taken together, numerous threads suggest that songbirds can maintain a memory trace for songs recently heard or sung for at least a few seconds. If syllables sung or heard can introduce a memory trace that lasts for seconds, and if this memory trace can impact future decisions about what to sing, then a substrate exists that could introduce long-range correlations between decision points in song. How deep is the memory for past choices in song among the most elaborate singers?
One of the most complex singers that can be easily reared in a laboratory setting is the domesticated canary. Here we investigated the long, complex songs of the Belgian Waterslager strain. We show that their songs are governed by long-range correlations in syntax at multiple hierarchical levels. The time a bird spends repeating a syllable or the choice of what to sing next depends on the history of the song and correlations between the past and present can extend over durations up to 10 seconds, encompassing 4 or more phrases consisting of dozens of syllable repeats. Canary song, like most popular music, contains structure in a range of time-scales through which sequence flexibility is balanced by long-range order [21]. The neural basis of these long-range correlations may provide insight into how complex behaviors are assembled from more elementary, stereotyped behavioral modules.
The smallest indivisible unit of the domesticated canary's song is the syllable, a stereotyped sound that typically ranges from 20 to 200 ms in duration. The adult repertoire usually contains between 25–35 distinct syllable types whose acoustic forms are learned by an interplay between innate programming and flexible imitation [22]. Syllables are repeated multiple times to form a phrase, which can range from 500 ms to 3 s, and phrases are flexibly chained together to form songs (Fig. 1), typically 5–15 s long [23]. In the present context, the term “phrase type” refers to the syllable type repeated in a given phrase.
These two fundamental units of the canary song—syllables and phrases—form distinct time-scales. Syllable durations range from 28 to 480 milliseconds, while phrase time-scales are on the order of a second (1.3375–1.3589 95% bootstrap confidence interval of the median), and the full song is an order of magnitude longer (Fig. 1 and Fig. 2). Fig. 2b shows that there is no general correlation between syllable duration and phrase duration (r = .013, p = .90); canaries persist on a single syllable for a duration that is roughly one second, whether the syllable is short or long [24]. To produce a phrase of the characteristic duration, the shortest syllables are repeated 20–30 times, while the longest syllables are repeated only 3–4 times. We discuss later the implications of a phrase time-scale that is not simply related to syllable time-scales.
As a group, there is no general correlation between phrase length and syllable length (Fig. 2b). However, particular syllable types do have their own characteristic phrase lengths, and the duration of specific phrases can vary depending on the context in which they occur. Specifically, we found highly significant mutual dependence between the length of specific phrases and the phrase type sung before or after the phrase (Fig. 1a and Fig. S1) (81 phrase types examined in 6 birds, p<.001 Fisher-Freeman-Halton test, see Materials and Methods). For a given phrase type, the length of a phrase can depend on the recent history of the song. Moreover, the choice of what to sing at a branch-point can depend not only on the currently active syllable, but also on the amount of time elapsed since the onset of the last phrase.
Renditions of what appeared to be the same syllable in short and long phrases might show subtle distinctions acoustically. If they are distinct, then the apparent long-range correlations could be more simply described by nearest neighbor rules in the syllable sub-types [16]. Here we sought to provide a methodology for direct visual representation of syllable variability. To proceed, we first isolated all renditions of a chosen syllable based on an automated template matching procedure (see Materials and Methods). We confirmed by visual inspection of sonograms that all examples of the chosen syllable were extracted, with no errors. We then separated the syllables into three groups depending on whether they came from short, medium or long phrases. (Phrase durations were discretized into three bins of uniform time-span, an arbitrary choice.) We then generated spectral density images (see Materials and Methods) for each group. The spectral density image provides a quantitative representation of syllable form and its variability across multiple renditions. In the spectral density image, color scale indicates the probability of finding a time-frequency contour [25] at a given point in the time-frequency plane. For a few syllables, visual inspection of the spectral density image revealed no variation for syllables drawn from different phrase duration bins (Fig. 3a).
This qualitative observation can be quantified using a similarity score based on the spectral density images (see Materials and Methods). Specifically, we computed the all-to-all overlap of binary contour images by computing the inner product between all pairs. Then, the all-to-all scores were sorted by their corresponding phrase groups and the distributions were compared using a Kolmogorov-Smirnov (KS) test and a d′ measure (see Materials and Methods). By the KS test, the distributions were distinct (p<.01), however the scale of the acoustic differences was very small by our measure. In all cases d′<0.2 (see Materials and Methods), indicating that the average differences between syllable shapes in different groups were smaller than the variations within a given group (Fig. 3b, and Table S2). Corroborating measurements were found using scores computed from acoustic features defined in the Sound Analysis Pro for MATLAB package [26]. For these scores, d′<0.1 (see Text S1).
In this analysis, we chose the most stable syllables. In other syllable types (particularly the fastest syllables), a systematic shift in acoustic form may occur over the course of a phrase. Also, for many phrase types in canaries, the first syllable of a new phrase type shows a structure that matches neither the preceding nor the succeeding phrase. If a switch in syllable forms is made in the central motor control loops, the syringeal or respiratory pattern may require a finite time to reconfigure. Ongoing phonation during this period of reconfiguration may produce syllable forms that differ from the steady state syllable forms [27]. Fig. S10 reveals that these specific context dependent effects can be acoustically significant. Excluding the special case of the first transitional syllables in a new phrase, in the syllables analyzed here, changes in form in different contexts were too small to allow a single instance of a syllable to be reliably assigned to short, medium, or long phrases using scores based on either SAP features or spectral density images.
Canary song is organized around a mesoscopic structure, the phrase. Is the larger sequence of song explained by a first-order Markov process in phrases, or is phrase sequencing more complex than a first-order Markov process? To examine this possibility, we first observed that the succession of phrases is quite constrained—each phrase is typically followed by just a few downstream possibilities (Fig. 4a). The top three or four transitions account for most of the variations that follow a given phrase. We next examined the entropy of phrase sequences of various lengths, and compared this entropy with random sequences that preserve only first-order transition statistics. We found that the entropy of phrase sequences is almost as high as a first-order model would imply (Fig. 4b). However, the match is not perfect, and for sequences 4–6 phrases long, it is clear that the set of song sequences is smaller than the set of possible sequences in a first-order random model. Song is thus more ordered than a first-order Markov process acting on phrase types.
To examine further the constraints placed on canary phrase sequences, we first tested all phrase types for statistically significant second-order structure. Specifically, for a sequence of three phrases XYZ, we asked whether the phrase type X impacted the phrase type probabilities for Z, for a given phrase type Y. The test reveals that this mutual dependence exists for 70% of all examined phrase types (p<.001, Fisher-Freeman-Halton test, see Fig. S1). Here too, a high-resolution analysis of selected examples using spectral density images confirmed that the acoustic form of the syllable in position Y can remain relatively constant even when flanked by diverse phrase types in position X or Z (Fig. S5a). As for the earlier analysis of syllable forms in phrases of different lengths, this observation was quantified both with spectral density similarity scores (d′<0.35) and through the use of SAP scores (d′<0.1), indicating that variations of syllable form within each group are larger than the separation between syllable forms in different syntactic contexts. (As before, there are detectable differences between the distributions, p<.01 two-sample Kolmogorov-Smirnov test).
It is not known whether peripheral motor variables such as air pressure or muscular tone change over the course of a long canary song. Time-dependent changes in the periphery could impact the acoustic details of song [27]–[29]. We examined how syllable form changed when syllables occurred early or late in song, for a fixed phrase context defined by the immediate preceding syllable. Pairwise similarity analysis was performed for all syllables examined in the phrase-context analysis described in the preceding paragraph. Grouping syllables into renditions that occur before or after the median song duration for a given bird, detectable differences in acoustic form could be found between groups for all syllable types analyzed (p<.01 two-sample Kolmogorov-Smirnov test, n = 3 syllable types). The d′ value of the group differences is comparable to the changes reported in the previous paragraph for phrase context. For song position effects, spectral density based similarity scores reveal d′<0.2 and for SAP similarity scores d′<0.1. To summarize the analysis of syllable stability: a memory for past phrase choices impacts future phrase choices or phrase durations, and this memory may have a very limited impact on the acoustic form of some syllables (see Fig. S5b and Table S3). It is possible that the minor acoustic changes in syllables can be largely explained by small-scale drift in peripheral control variables.
The previous analysis indicated that second-order correlations introduce a statistically detectable shift in phrase transition probabilities for most phrase types, but these second-order effects could be weak. Still, weak higher-order correlations could in principle explain the gap between a first-order random model and canary song. However, the next stage of analysis revealed that while higher-order correlations are weak for many phrase types, for some phrase types, strong long-range rules apply to the delivery of song.
To examine how long-range correlations varied by phrase type, we constructed a prediction suffix tree (PST) [30] to represent each bird's song. A PST provides a visual representation of how past information in a sequence impacts transition probabilities. Formally, the tree is built from a collection of Markov chains, one for each phrase type. Each chain is initialized as a zero-order Markov chain, and the order is increased only if the information gained by looking further back in time justifies the added complexity (see Materials and Methods). In Fig. 5, the PST is displayed radially, with each syllable arranged around the inner circle. For a given syllable, the number of nodes between the tree trunk and the outer branches indicates the order of the Markov chain for that syllable. For a syllable impacted by high-order correlations, that syllable on the trunk of the PST tree will be connected to long, multi-branched limbs. Similar methods were recently used in the analysis of Bengalese finch syntax [15].
As a control, we used a 10-fold cross-validation procedure. Prediction suffix trees were computed for the training data, and then the average negative log-likelihood of the test data was computed for each tree (Fig. 6). The PST that leads to the minimum in average negative log-likelihood on the test set is considered the best fit. The depth of the best fit ranged from 4 to 7 phrases in the six birds examined here (Fig. S7). This corresponds to a propagation of song information over a time-scale of approximately 5 to 10 seconds. Many syllables in this analysis showed no significant structure beyond first-order; just a few syllables are governed by long-range rules. (The prevalence of second-order structure revealed in Fig. S1 suggests that the PSTs provided a conservative estimate of statistical depth for many syllables.) The structure of example songs with long time-scale correlations is illustrated in Fig. 7 where the syllable identity of the phrase at the top of a chain impacts transition probabilities many phrases later. (Fig. S6 contains similar song barcodes showing the full song context for the examples in Fig. 7.) In the examples given in Fig. 7, the history of previous syllable selections can impact future syllable transitions over 4–5 intervening phrases, spanning a time-scale of up to ten seconds.
The apparent statistical depth of the phrase structure could be reduced when sequences of phrases occur as a unit [15], [16]. In sequence DABN for example, the choice of syllable D impacts choices after N, implying a fourth-order correlation. However, the PST transition probabilities for nodes A, B and N all have equivalent or near-equivalent values. The state space can be reduced to D (ABN)—a second-order model in phrase sequence “chunks.” In the process of constructing a PST, the nodes that can be collapsed into a single chunk without changing the predictive power of the model are “internal nodes.” These nodes do not impact the transition probability at the end of a chain, but just provide a connection from the leaves of the tree to the trunk. In Fig. 5 and Fig. S8 internal nodes are labeled with blue text labels. (Here internal nodes are defined as those nodes that would not have been added to the PST on their own strength, but are simply added to show the connections from the outer branches to the core of the graph. As such, the definition of internal node depends on the parameters used in the PST fit.) After collapsing internal nodes in these figures, the maximum depth of the suffix trees reduce from a range of 4–7 to a range of 2–3 in the 6 birds analyzed here, whereas the log-likelihood of the model changed, on average, by less than 1 percent, indicating that sequence chunks could be regarded as monolithic states without impacting the quality of the model.
The PST provides a particularly compact representation of long-range dependencies in song. The compactness of the PST representation is emphasized by comparison of the PST graph with its corresponding probabilistic finite automaton (PFA) [30]–a first-order transition model that can be more easily related to first-order dynamical models of neural activity. Fig. 8 illustrates the graph structure of the PFA for one bird, whose PST is given in the top of Fig. 5. To render the PFA visually interpretable, 363 edges were deleted from this figure that occur with less than 20 percent probability (more complete PFAs are shown in Fig. S11). We emphasize that in spite of the complexity of Fig. 8 and Fig. S11, the PFA is only a statistical model for phrase transitions. The model does not account for phrase durations, or the fact that phrase durations depend on the recent history of the song–a point documented earlier–these features should be addressed in a more complete statistical model.
We next examined whether the long-range correlations followed a simple adaptation rule. Long-range correlations could appear if the probability of a given phrase transition decays with the frequency of its use. That is, as each phrase transition occurs, its subsequent probability of occurrence is decreased. In the cases examined in Fig. S9, no simple rule was observed—the most frequent syllable transition from a given phrase can increase or decrease in likelihood as a function of the number of times the phrase is produced in a given song. Over the transitions that we could analyze for this property, 33% strictly increased while 50% decreased, and 13% both increased and decreased over the course of a song (17/52, 26/52 and 7/52, respectively).
We then checked to see if the statistical depth of canary song could be explained by limitations on song duration. The concern here is that some branches in the path of song might lead to unusually long songs that could be prohibited for physiological reasons. We analyzed all examples of context-dependencies and found no evidence for this effect (see Text S1). This search for simple rules explaining the apparent depth of canary song was not exhaustive, and a more extensive analysis involving additional samples from other canaries is needed.
Canary song is built from elementary units, the syllables, repeated in groups to form a mesoscopic structure, the phrase. Phrases are flexibly sequenced to form songs. Correlations among phrase choices can extend over time-scales of 7–10 seconds. Over this time-scale, 4–5 phrases may be produced consisting of typically dozens of syllables. This observation significantly extends the time-scales that must be considered in dynamical models for song generation. We first discuss the time-scale of a single phrase. Dynamical theories for the central control of song are, in various forms, attractor models [31]–[36]. If each phrase type is a separate attractor (or closed-neural chain) in canaries, then the phrase transition could be produced by a “kick” that recurs every second or so, inducing a hop from one attractor to another. Statistically, the phrase durations of canary song could also be described by a POMMA model [16] if each syllable has a self-return probability that decreases with each repeat, as long as the adaptation rate scales inversely with syllable duration. More simply, phrase time-scales could be introduced into first order models like POMMA by introducing an adaptation that changes as a function of time rather than syllable repeats. Experiments are needed to determine whether canary phrase time-scales are defined by a fine-tuning of syllabic adaptation rates or a separate phrase transition process with its own intrinsic time-scale. Whatever the mechanism of defining phrase durations, it apparently does not need to be informed by any auditory experience with natural canary song, since birds reared in acoustic isolation also develop canary-typical phrase structure [22].
The observed structure of canary song significantly extends the time-scale of long-range correlations documented in bird song. For example, the sequence CDABNE from the top of Fig. 5 has an average duration of eight seconds and the sequence YFHX from the bottom of Fig. 5 has an average duration of ten seconds. Long-range rules in canary song can be compactly described, by 4th–7th order Markov processes acting on phrases, or a 2nd–3rd order Markov process acting on larger units that include blocks of multiple phrases.
A suffix tree of depth 7 for 30 phrase types could in principle have 30∧7 nodes, and even a third order process in 30 phrase types could have 30∧3 nodes. In contrast to this large state space, the PSTs included in Fig. 5 are quite sparse. The first-order generative models for the canary songs represented by the PFA diagrams contain only 47–56 states representing syllables or syllable strings. This is not greatly larger than the number of observable syllables (17–26). For canary song, the PST provides a particularly compact representation of syntax dependencies. The compactness of this representation is clear in comparisons between the PST tree and its corresponding first-order models (Fig. S11). The speed and convergence properties of the PST algorithm make it possible to quickly cross-validate the structure of the PST over large ensembles of trials, defining the point at which over-fitting occurs. Taken together, these properties suggest that the PST analysis will be generally useful for characterizing the structure of birdsong syntax.
Statistically complex phenomena that are best described by higher-order Markov processes such as the PST can be generated by simple physical processes. For example, if circulating neuromodulators in song nuclei depend on the syllables that are sung, and if syllable transitions themselves depend on the hypothetical neuromodulator, then stochastic variations in the beginning of a song could impact future transition probabilities, generating apparent “long-range rules”. In this scenario, additional information is needed to explain why some syllables show strong long-range rules, and others behave in a simple first-order manner.
On a fine-grained scale, neural dynamics should be captured by a first-order statistical process. In statistical terms, long-range correlations in syntax imply that multiple “hidden states” can give rise to the observable syllable. This duplication of states statistically does not imply that the motor program for a syllable is duplicated–the smallest change in a syllable program satisfies the duplication of “hidden states.” Recent studies in Bengalese finches have observed that the stereotyped neural program for a syllable depends on its context–in particular, changes were observed in Basal Ganglia projecting neurons in nucleus HVC (used as a proper name) [37], and subtle acoustic changes in syllable form were observed for syllable in different contexts [16]. Whatever the mechanism of the long-range rules in canaries, the neural variables that carry the memory for past song choices can exert a powerful effect on transition statistics without significantly altering the acoustic form of syllables. This observation is supported through the high-resolution “spectral density” images introduced here to characterize syllable variability.
Canary phrase structure and canary syllable form appear to be encoded by separable processes. This distinction is supported by the observation that phrase time-scales are not simply predicted by syllable time-scales. Another line of evidence arises in studies of song learning in juvenile canaries. As juveniles, canaries can learn to imitate artificial songs that lack normal phrase structure [22]. Rising testosterone levels that occur with the onset of the breeding season cause a rearrangement of song–the imitated syllables are reorganized into phrased repetitions. In this artificial tutoring paradigm, what is most dramatically reprogrammed in the transition to adulthood is not syllable acoustic structure, but the sequential organization of the syllables.
Many questions remain about the neural basis of phrase structure in canary song, but we may also wonder about the relevance of the long-range rules for the natural behavior of the species. Is the statistical depth of song an epiphenomenon of little ethological relevance, or do canaries show preferences for songs with long-range order? Can a canary fine-tune the long-range rules to match a tutor song, or can a bird be trained to alter rule-sets in different behavioral contexts? These questions are addressable since canaries readily imitate artificial songs designed to pose specific questions about their vocal learning processes [22].
Lashley emphasized that the control of serial order in behavior is one of the most important and least understood aspects of neuroscience over 60 years ago [38]. Songbirds have provided an opportunity for examining sensory-motor learning of stereotyped neural sequences; dynamical models for song sequence generation have generally focused on stringing together bursts of neural activity in a long chain of elementary states [34], [39]. This representation is remarkably similar to observed neural dynamics in zebra finches and Bengalese finches, and can be related to simple first-order statistical models for song production [16], [17]. However, the long-range rules that govern canary song extend to time-scales of 10 seconds, and persist while a bird vocalizes five or six intervening phrases, consisting of dozens of syllables each. How is information in the song circuit transferred over these time-scales? Answers to this question may provide general principles of how complex behaviors with long-range correlations are assembled from simple modules.
The care of animals in this study was carried out in accordance with Boston University IACUC protocol number 09-007.
Canaries (Belgian Waterslager strain) used in this study were a gift from Fernando Nottebohm. Birds were isolated at least two weeks before recording in soundproof boxes and kept on a light-dark cycle matched to the external annual light cycle in Boston (Boston University IACUC protocol number 09-007). All birds were at least one year old before isolation. Song was recorded between the months of March and April.
Spectrograms of the song were calculated in MATLAB (Mathworks, Natick, MA), and the beginning, end, and syllable identity of each phrase was marked on the image by visual inspection. For all data described here, two independent observers annotated the songs. Observer 1 and observer 2 annotated 33,469 and 36,447 phrases, respectively, between 6 birds (see Table S1 for the number of phrases analyzed for individual birds). The annotated sonograms were then scanned and converted into strings, and statistical analysis of the strings was performed using custom MATLAB scripts. Zero to two syllables were excluded per bird because the syllable form consisted of subtypes that could not be labeled consistently.
For each bird (6 total), we first examined the mutual dependence between a given phrase's duration and the path into or out of the phrase. That is, in a given phrase sequence XYZ, we examined the mutual dependence between the length of phrase Y and identity of the syllable type in phrase X or the identity of the syllable type in phrase Z, which we call MD(dur,pathin) and MD(dur,pathout), respectively. We discretized phrase durations by terciles and then used a variation of the Fisher test for contingency tables suitable for arbitrarily large, sparse contingency tables referred to as the Fisher-Freeman-Halton test [40]. Significance values for the test were then computed using a standard Monte Carlo procedure where the contingency tables were randomized while preserving the marginal values (i.e. the row and column totals), and to derive a p-value we calculated 1,000,000 randomizations for each test [41]. Using the same method, we also examined the mutual dependence between the syllable identity of phrase X and the syllable identity of phrase Z, for each Y, which we label MD(pathin,pathout).
Custom MATLAB (Mathworks, Natick, MA) scripts were used for automated syllable clustering. After choosing a template, spectral features from the template and the rest of the data were computed using a sparse time-frequency representation [42] (see Text S1). As a final step, candidate sounds were plotted in two dimensions and a decision boundary was drawn by the user.
For a collection of syllables, we first generate a sparse time-frequency representation of each syllable using auditory contours [25]. Auditory contours provide a high-resolution binary image consisting of sparse, continuous lines that follow the features of the sound with high precision. Summing over contours for all renditions of a syllable produces a two-dimensional probability density in time and frequency, which we call a spectral-density image, . Specifically, a single contour image calculated at a given resolution is a sparse, binary time-frequency image or spectrogram , which we compute for each syllable , denoted . The spectral density image is then defined as , where N is the number of binary images. By definition, is the probability of finding a contour in pixel in a single sample, and in the images shown here, the value is represented by color scale. These images provide a direct representation of the variability of syllable form for every point in time and frequency.
Since the starting matrices are sparse, and high precision, the spread of the acoustic energy revealed in may be narrower than the resolution of standard sonograms, for sparse stereotyped song elements. When interpreting these images, it must be understood that variations in frequency or timing of auditory contours both lead to a spread in spectral density; separating these sources of variability is not generally possible without additional analysis such as time-warping [22].
We define a simple similarity measure between two syllable contour images and to be ; With this definition, the average similarity between two groups A and B is just , the inner product of their spectral density images. The average similarity between a one specific syllable and an ensemble of syllables B is . To quantify the separability of two distributions we use the d′ measure, , or the difference in mean similarity scores in units of pooled standard deviation.
To test for the existence of structure beyond second-order (as in the MD(pathin,pathout) test), we constructed prediction suffix trees (PSTs) for the sequence of canary phrase types using a previously published algorithm [30], [43]. Further details can be found in Text S1.
To compute block entropy we used standard methods based on Shannon entropy [44]. The maximum entropy line in Fig. 4 assumes symbols are emitted with uniform probability.
|
10.1371/journal.ppat.1003058 | Infection-Induced Interaction between the Mosquito Circulatory and Immune Systems | Insects counter infection with innate immune responses that rely on cells called hemocytes. Hemocytes exist in association with the insect's open circulatory system and this mode of existence has likely influenced the organization and control of anti-pathogen immune responses. Previous studies reported that pathogens in the mosquito body cavity (hemocoel) accumulate on the surface of the heart. Using novel cell staining, microdissection and intravital imaging techniques, we investigated the mechanism of pathogen accumulation in the pericardium of the malaria mosquito, Anopheles gambiae, and discovered a novel insect immune tissue, herein named periostial hemocytes, that sequesters pathogens as they flow with the hemolymph. Specifically, we show that there are two types of endocytic cells that flank the heart: periostial hemocytes and pericardial cells. Resident periostial hemocytes engage in the rapid phagocytosis of pathogens, and during the course of a bacterial or Plasmodium infection, circulating hemocytes migrate to the periostial regions where they bind the cardiac musculature and each other, and continue the phagocytosis of invaders. Periostial hemocyte aggregation occurs in a time- and infection dose-dependent manner, and once this immune process is triggered, the number of periostial hemocytes remains elevated for the lifetime of the mosquito. Finally, the soluble immune elicitors peptidoglycan and β-1,3-glucan also induce periostial hemocyte aggregation, indicating that this is a generalized and basal immune response that is induced by diverse immune stimuli. These data describe a novel insect cellular immune response that fundamentally relies on the physiological interaction between the insect circulatory and immune systems.
| Mosquitoes transmit diseases such as malaria, dengue fever, West Nile virus and lymphatic filariasis. A mosquito initially acquires a pathogen when she ingests a blood meal from an infected person or animal. Then, after a period of development and/or replication in the mosquito gut, the pathogen enters the hemocoel (body cavity) and undergoes an obligate migration to the salivary glands (the destination for viruses and protozoans) or the mouthparts (the destination for larger worms). During this migration, pathogens are subject to two potentially antagonistic mosquito forces: immune responses and circulatory currents. In this study, we examined the physiological interactions between the mosquito immune and circulatory systems. We show that when mosquitoes are infected with bacteria or malaria parasites, mosquito immune cells (hemocytes) migrate to the areas surrounding the valves of the heart. At these areas of rapid and dynamic hemolymph (mosquito blood) flow, hemocytes swiftly phagocytose and kill pathogens. These experiments describe a novel and basal insect immune response that fundamentally relies on the physiological interaction between the mosquito circulatory and immune system. Furthermore, because traversal of the hemocoel is required for pathogen transmission, this new knowledge could be used in the development of novel disease control strategies.
| Pathogens transmitted by mosquitoes must traverse the insect's open body cavity (hemocoel) during their journey from the midgut to the salivary glands, and this obligate migration places them in direct contact with the insect's circulatory and immune systems. The insect circulatory system consists of hemolymph (blood), the hemocoel, and pulsatile organs, of which the dorsal vessel is the most important [1]. The dorsal vessel extends along the dorsal midline of the insect and is anatomically divided into a thoracic aorta and an abdominal heart. In adult mosquitoes, the heart drives hemolymph propulsion by sequentially contracting in a wave-like manner, with the contractile waves periodically alternating between propagating in the anterograde (toward the head) and retrograde (toward the posterior abdomen) directions (Figure 1) [2], [3]. When the heart contracts in the anterograde direction, hemolymph enters the lumen of the vessel through six pairs of incurrent ostia (valves) located in the anterior portion of abdominal segments 2 through 7 and exits through an excurrent opening located in the head region [2], [3]. When the heart contracts in the retrograde direction, hemolymph enters the vessel through a single ostial pair located at the thoraco-abdominal junction and exits through an excurrent opening located in the terminal abdominal segment. Variants of this arrangement are seen in all insects and similar systems are present in all arthropods [4], conclusively supporting its ancient origin.
While insect circulatory processes have been primarily studied for their role in transporting nutrients, wastes and signaling molecules, one aspect that has been overlooked is the relationship between hemolymph circulation and immune responses. The insect immune system relies on innate reactions to fight pathogens and involves both cellular and humoral components [5]–[7]. To date, studies on insect immunity have focused primarily on dissecting the molecular bases of immunity and on understanding the cellular biology of hemocytes (immune cells). These experiments have largely assessed immune responses at single points in time, and have used methods that require insect death during the extraction of hemocytes or other tissues. In the hemocoel, hemocytes, humoral immune factors and pathogens exist in contiguous association with the insect's circulatory organs and have likely shared such an existence throughout the course of arthropod evolution. Because this association occurs in a fluid and dynamic space, we hypothesized that hemolymph currents influence the temporal and spatial control of anti-pathogen responses. Furthermore, given that the closed circulatory and lymphatic systems of vertebrate animals are integrally associated with immune surveillance [8], [9], we hypothesized that coordinated interactions between the insect's open circulatory system and immune system are essential for effective insect immune responses. Whether this interaction occurs remains unexplored in any insect, and mosquitoes are an exceptional model for its investigation because: (1) physiological interactions between mosquitoes and a taxonomically diverse array of pathogens have been explored [10]–[13], (2) the mosquito circulatory system has been well characterized [2], [3], (3) the phylogenetic distance between mosquitoes and Drosophila melanogaster provides a unique perspective on the evolution of the insect immune and circulatory systems, and (4) it has been proposed that mosquito immune responses could be harnessed for the control of mosquito borne disease [14]–[16].
We have reported that during the course of an infection in the malaria mosquito, Anopheles gambiae, pathogens accumulate in discrete foci along the surface of the mosquito heart [11]. Specifically, pathogens accumulate as flanking lines on the anterior portion of each abdominal segment in areas that match the location of the heart's ostia [3]. Here, we used intravital imaging and microdissection techniques to show that during the course of bacterial and malarial infections, mosquito hemocytes migrate to the areas surrounding the heart's ostia, defined here as the periostial regions (Figure 1), where they bind the musculature and each other, and engage in the rapid phagocytosis of pathogens. This invertebrate cellular immune response integrally involves hemolymph circulation and the heart, and relies on the physiological interaction between the mosquito circulatory and immune systems, supporting their co-adaptation to counter pathogens. Furthermore, this discovery suggests that many studies into mosquito immune responses have underestimated the role hemocytes play in controlling infection.
The initial goal of this study was to investigate whether pathogen accumulation around the dorsal vessel is due to a physical barrier or due to an active immune response (Figure 2A–B) [11]. Hemolymph in the abdominal cavity flows dorsally toward the ostia [2], [3], and as part of this flow it is possible that pathogens could become stuck in the fenestrated dorsal diaphragm or at the narrow openings of the ostia. Alternately, pathogens could be sequestered as part of an active immune response, and an earlier study suggested that this would most likely be a PC-mediated immune response [11]. To distinguish between these two scenarios we developed two novel methods for the labeling of PCs in vivo. The first employs the injection of AlexaFluor-conjugated IgG and relies on the pinocytic nature of insect PCs, while the second employs LysoTracker Red and relies on the ability of this stain to label cells with high acidic, and presumably lysosomal, content.
Co-staining of PCs and the abdominal musculature showed that PCs are binucleated cells that flank the mosquito heart (Figure 2C). PCs occur along the length of the heart, with large gaps present between the diamond shaped alary muscles that tether the heart to the cuticular epidermis, and small gaps present slightly posterior of the anterior-posterior midline of the alary muscles. Based on flow experiments (movie S1) as well as our previously published work on the structural mechanics of the heart [3], we conclude that the smaller gaps flank the heart's ostia, and thus, in this study we refer to these areas as the periostial regions (Figure 1).
Staining of PCs after infection with GFP-expressing Escherichia coli showed that pathogens are not phagocytosed by PCs. Instead, pathogens accumulate at the periostial regions between the four PCs that flank the ostia (Figure 2D). Further examination of LysoTracker Red-stained dorsal abdomens revealed that the accumulated bacteria are inside cells with nuclear and cellular diameters that are considerably smaller than that of PCs, and that these phagocytic cells are similar in size to the phagocytic subpopulation of circulating hemocytes (granulocytes) (Figure 2E). Thus, these data show that pathogen accumulation in the periostial regions is due to an immune response that is not mediated by PCs, and suggest that the phagocytic cells are likely hemocytes that bind the alary muscles and the outer surface of the heart at the location of the ostia.
To determine whether the phagocytic cells identified in the periostial regions are hemocytes, we developed a novel in vivo hemocyte-staining method using the dye chloromethylbenzamido-1,1′-dioctadecyl-3,3,3′,3′-tetramethylindocarbocyanine-perchlorate (CM-DiI). Intrathoracic injection of CM-DiI into live mosquitoes, followed by hemolymph perfusion and ex vivo analysis of cell staining efficiencies showed that CM-DiI stains greater than 97% of the hemocytes collected from naïve (99%), injured (sterile injection with LB broth; 99%) and E. coli infected (97%) mosquitoes (Figure 3A–D). Quantitative co-localization of CM-DiI staining and the phagocytosis of GFP-E. coli by circulating hemocytes yielded a Mander's overlap of 70%, illustrating the high hemocyte-staining efficacy of CM-DiI. Further qualitative analyses suggested that the slightly lower hemocyte-staining efficiency seen in E. coli infected mosquitoes is due to lower incorporation of CM-DiI in hemocytes that have phagocytosed extremely large numbers of bacteria.
In addition to hemocytes, hemolymph perfusion results in the collection of small amounts of fat body as well as free nuclei from lysed cells [17], [18]. Visual analysis of all cellular components collected by perfusion revealed that CM-DiI only stains hemocytes (Figure 3C–D), and analysis of carcasses and other tissues from CM-DiI injected mosquitoes conclusively showed that cells other than hemocytes do not incorporate CM-DiI.
With the knowledge that a major cellular immune response occurs in the periostial regions, we then aimed to determine whether this immune response is mediated by hemocytes. Complementary hemocyte, PC, and muscle staining confirmed that this indeed is the case. Co-labeling of hemocytes and muscle showed that hemocytes laterally flank the heart at the periostial regions in both naïve and E. coli-infected mosquitoes (Figure 4A). Infection with GFP-E. coli, followed by CM-DiI injection 24 h later, revealed strong spatial overlap between GFP-fluorescence and hemocytes (Figure 4B), and suggested that hundreds of hemocytes phagocytose pathogens in the periostial regions during an active infection. Then, examination of abdomens from mosquitoes infected with non-fluorescent E. coli confirmed that PCs and hemocytes are two distinct cell populations, and that hemocytes occur in aggregates positioned between the 4 PCs that flank each ostial pair (Figure 4C). Finally, comparisons of cellular and nuclear diameters confirmed that hemocytes are identical in size to (1) the phagocytic cells previously detected using LysoTracker Red staining (Figure 2E), and (2) the circulating population of granulocytes (Figure 3A–C).
Because immune processes within the mosquito hemocoel occur in three-dimensional space, we performed a series of dual labeling experiments to elucidate the three-dimensional relationship between bacterial aggregation and the periostial hemocytes, and the relationship between hemocytes, PCs and the heart. Analysis of deconvolved and volume rendered Z-stacks revealed that aggregated periostial hemocytes and E. coli overlap in all three dimensions, supporting the phagocytic activity of heart-associated hemocytes (Figure 4D–E). Then, similar experiments where heart muscle was labeled along with PCs, phagocytosis foci or hemocytes showed that PCs intertwine with the alary muscles while phagocytosis foci and periostial hemocytes are located dorsal of the alary muscles and in the vicinity of the ostia (Figure 4F).
Finally, because sessile hemocytes occur elsewhere in the mosquito, we examined whether hemocytes or bacteria preferentially aggregate elsewhere in the abdomen. We found that, throughout the abdomen, the periostial regions are the major location of hemocyte aggregation following infection (Figure 4G–H), as well as the major location of E. coli sequestration (Figure 4I–L). Quantitative analysis of the dorsal abdomen showed that between 6 and 24 h post-infection, ≥62% of E. coli aggregation, as measured by GFP fluorescence, is confined to the periostial regions (Figure 4J–L). When the analysis was repeated for the entire abdomen (dorsal+ventral), ≥42% of the aggregated E. coli was confined to the periostial regions. Taken altogether, these data describe a novel mosquito immune response, where hemocytes in the periostial regions of the heart phagocytose and degrade pathogens.
Because periostial hemocytes appeared to represent a substantial proportion of the total number of hemocytes present in mosquitoes [17], [19]–[23], and because their positioning suggested a strong interaction between the mosquito circulatory and immune systems, we then investigated whether hemocytes are recruited to the periostial regions during the course of an infection. Intravital video imaging of periostial hemocytes during the first 15 min post-infection revealed that resident sessile periostial hemocytes (basal population of hemocytes always present at the periostial regions) phagocytose E. coli within seconds of their injection into the hemocoel (Figure 5A–C; Movie S1). Pearson's correlation coefficient measurements of Movie S1 quantitatively proved this, as the fluorescence overlap of periostial hemocytes and E. coli increased within the first few frames of the video and began to plateau at a correlation near 60% by 3 min post-infection (Figure 5D). Longer video recordings confirmed the high phagocytic activity of resident periostial hemocytes, and also showed that the number of hemocytes in the periostial regions increases during the course of an infection (Figure 5E–G; Movie S2). Specifically, during the first hour of infection with E. coli (OD600 = 4), the number of hemocytes in the periostial regions roughly doubles. This increase is due to the movement of hemocytes into the periostial regions and not hemocyte replication at the pericardium. As seen in Movie S2, some hemocytes flowing into the pericardium adhere to the heart-associated musculature. These hemocytes then slowly glide into the periostial regions, where they settle and phagocytose pathogens. While hemocytes exist both in circulation and attached to tissues (sessile), the relatively low number of sessile hemocytes observed in non-periostial areas of the dorsal cuticular epithelium, together with the relatively rapid arrival of hemocytes into the pericardium suggests that the recruited cells originate from the circulating hemocyte population. The molecular trigger for their arrival is unclear, but we hypothesize that hemolymph flow brings circulating hemocytes to the pericardium, where they bind the alary muscles and then slowly undergo a directed migration into the periostial regions. This migration is facilitated by hemolymph flow but is not exclusively driven by it; the median velocity of the hemocytes tracked in Movie S2 was 1.2 µm/sec (Figure 5G), which is several orders of magnitude slower than the 200–1,000 µm/sec and ∼8,000 µm/sec hemolymph flow velocities in the periostial regions and heart lumen, respectively [3]. Overall, while the vast majority of hemocytes migrate to the periostial regions as individual cells, a minor proportion of hemocytes in a minority of mosquitoes arrives at the pericardium in small cellular aggregates. Many single or aggregated migrating hemocytes contain phagocytosed E. coli prior to entering the pericardial space, indicating that they have been immune activated elsewhere in the mosquito.
To further elucidate the process of periostial hemocyte aggregation, naïve mosquitoes, mosquitoes that had been injured 24 h earlier, and mosquitoes that had been infected with various doses of GFP-E. coli for 24 h were injected with CM-DiI and the number of periostial hemocytes were counted. On average, 5-day-old naïve mosquitoes contain 43 periostial hemocytes, distributed among the 6 periostial regions of the abdomen (Figure 6A, C). Injury does not result in an increase in the number of periostial hemocytes, but infection with large numbers of E. coli for 24 h leads to a >4-fold increase in the number of periostial hemocytes (Figure 6A). Hemocyte aggregation in the periostial regions is induced in an infection dose-dependent manner (Figure 6A), with the increases in mean number of periostial hemocytes per mosquito being nearly linear for E. coli infection intensities of OD600 = 1 through OD600 = 5 (R2 = 0.92; means of 69, 89, 106, 138, and 188, respectively). Quantitative analysis of periostial hemocyte numbers during the course of an infection revealed that hemocyte numbers in the periostial regions approximately double within the first hour, and plateau at 4 h post infection (Figure 6B). Depending on the bacterial dose and the time following infection, periostial hemocyte aggregates vary from small and dispersed groups of hemocytes to expansive and contiguous aggregates of hemocytes (Figure 6C–E). Infected mosquitoes at times contain more than 300 periostial hemocytes (Figure 6E), even when counting is done using our conservative protocol, which favors the exclusion of a small number of hemocytes rather than the inclusion of non-hemocytes. Finally, quantitative analysis of E. coli fluorescence in the periostial regions showed that the level of viable (fluorescent) bacteria in the periostial regions steadily increases until 12 h post-infection and then begins to decline (Figure 6F). This loss of GFP-fluorescence in the periostial regions strongly suggests that bacterial killing and degradation is dynamically occurring in these regions.
While infection induces the migration of hemocytes to the periostial regions, their numbers decrease once the acute stage of infection has passed. Infection of mosquitoes with large numbers of E. coli (OD600 = 4) revealed that hemocytes are recruited to the periostial regions within the first 24 h post-infection, but that the number of periostial hemocytes decreases by day 3 post-infection (not shown). Infection of mosquitoes with fewer E. coli (OD600 = 2) revealed a similar trend: the number of periostial hemocytes decreases significantly by 3 days post-infection, but remains elevated for the lifetime of the mosquito, relative to similarly aged naïve controls (Figure 7A). All mosquitoes assayed at 12 days post challenge had live E. coli in their hemocoels (and periostial regions; Figure 7B), and the amount of melanin deposition in the periostial regions also increased as mosquitoes aged. Thus, the maintenance of elevated numbers of periostial hemocytes may be due to the continued need of cellular antimicrobial activity, as mosquitoes appear to be incapable of completely clearing an E. coli bacterial infection [24], [25]. Finally, based on the number of circulating hemocytes [17], [19]–[23], the periostial hemocyte population represents between 10% and 25% of the total hemocyte population post-infection, which highlights the importance of pathogen sequestration in these areas of high hemolymph flow.
Phagocytosis of 1 µm diameter microspheres triggers the recruitment of hemocytes to the periostial regions (Figure 2A–B), suggesting that immune activation via phagocytosis pathways induces hemocyte migration to the pericardial space. We tested whether several soluble immune elicitors also induce hemocyte aggregation in the periostial regions. Being solubilized, these microbial components are outside of the size range that induces phagocytosis. Examination of dorsal abdomens 24 h after injection with peptidoglycan and β-1,3-glucan revealed that soluble immune elicitors also induce the aggregation of hemocytes in the periostial regions (Figure 8A–D). The aggregation response following treatment with PGN and β-1,3-glucan was lower than following infection with live E. coli (Figure 6A), but this reduced response could be related to the doses used (see Figure 6A) and may not be an indication that soluble immune elicitors are less (or more) capable of inducing periostial hemocyte aggregation. Melanization in the periostial regions was prevalent following injection with peptidoglycan and β-1,3-glucan (Figure 8B–D), and this melanization response was considerably stronger than what was observed following infection with E. coli. Finally, while periostial hemocyte aggregation remained elevated several days following injection of soluble elicitors, hemocytes at these later time-points could not be accurately counted because of the extensive melanization response. Taken altogether, these data suggest that periostial hemocyte aggregation is a basal immune response that is induced by a broad range of immune stimuli.
Mosquitoes are vectors of disease-causing pathogens. Among the most important mosquito-borne pathogens are Plasmodium parasites, which are the etiological agents of malaria. Plasmodium infection represents a complex and long-term immune stimulus for mosquitoes [11], [26], [27], and we have previously observed Plasmodium sporozoites near the heart's ostia [11]. For these reasons, and because Plasmodium infection occurs without breaching the outer cuticle, we analyzed whether Plasmodium infection induces periostial hemocyte aggregation. Examination of the periostial regions of mosquitoes that had received a normal blood meal and mosquitoes that had received a Plasmodium-infected blood meal revealed that the process of sporozoite migration to the salivary glands induces the aggregation of hemocytes in the periostial regions: mean numbers of periostial hemocytes increased from 72 in the non-infected group to 106 in the infected group (Figure 8E). Moreover, periostial hemocytes phagocytosed Plasmodium sporozoites, and a minority of these sporozoites were also melanized (Figure 8F–H). In some cases, multiple Plasmodium nuclei were observed within an individual hemocyte (Figure 8G). Some of these sporozoites were fragmented and exhibited varying levels of fluorescence intensity, which is indicative of death or dying (Figure 8G–H). Overall, these data show that Plasmodium infection induces the recruitment of hemocytes to the periostial regions, and suggest that the cellular immune response against Plasmodium may be stronger than what was reported in an earlier study [11] that assayed circulating hemocytes alone.
Based largely on practical constraints, the insect immune and circulatory systems have been conceptually divided into discrete elements, and the immune system further dissected into cellular and humoral components [28]. However, these entities are physiologically interrelated and have apparently evolved in integral association since the beginning of animal evolution [28]–[32]. The cellular immune response remains only partially understood in mosquitoes as well as in other adult insects [5], [7]. Likewise, interactions between major circulatory elements and immune cells are virtually unknown and have received little attention. Using methods we previously developed for the study of hemolymph circulation [2], [3], along with novel techniques for the in vivo investigation of hemocyte biology, we analyzed the cellular immune response in the pericardial region of the malaria mosquito, Anopheles gambiae. We discovered that during an active infection, hemocytes migrate to the periostial regions, where they form a major component of the cellular immune response. Exemplifying the interrelationship of cellular immunity and circulatory processes, periostial hemocytes form phagocytic foci in regions of high hemolymph flow, which are also in the direct vicinity of the mosquito's major nephrocytes, the pericardial cells. Together, the data presented herein describe the formation of a novel immune tissue in mosquitoes. Because previous mosquito studies did not recognize sessile hemocyte aggregations as a major player in immunity, they failed to examine a large proportion of the cellular immune response, and thus, underestimated the relative contribution of hemocytes in anti-pathogen responses.
Using a correlative imaging approach, we scrutinized what appeared to be major phagocytic foci forming in the periostial regions of the mosquito heart. We previously reported that these foci form in response to infection, but their cellular composition and their functional role remained unknown [11]. Similar foci form in Drosophila, although they have never been directly studied [10], [33]. Here, we dispel the notion that phagocytosis on the surface of the heart is due to the activity of PCs. This finding was not entirely surprising, as in two different insect orders the PCs have been shown to be surrounded by a basement membrane [32], [34], [35], and the presence of this physical barrier should impede the direct phagocytosis of invading pathogens. Instead, we report that immune foci on the surface of the heart are composed of periostial hemocytes that rapidly and efficiently phagocytose pathogens. Rapid phagocytosis is believed to be an essential immune process, which culminates in pathogen death and the production of humoral immune components [36], [37]. The large number of periostial hemocytes present in infected mosquitoes, when compared to the total number of circulating hemocytes [17], [19]–[23], suggests that this response involves between 10% and 25% of the hemocytes present in mosquitoes, thus highlighting the importance of pathogen sequestration in areas of high hemolymph flow. Finally, because nodulation and cellular encapsulation do not occur in mosquitoes [5], [7], periostial foci formation is the primary hemocyte aggregation immune response in the culicid lineage.
Periostial immune aggregates are composed of a mixture of resident hemocytes and circulating hemocytes that settle in the pericardial regions in response to infection. Given that bacterial infection induces an increase in the number of circulating hemocytes in An. gambiae [21], and that Aedes aegypti hemocytes can replicate in response to various stimuli [19], [20], we hypothesize that some of the migrating hemocytes seen in this study are the products of circulating hemocyte replication in response to infection. While the origin of periostial hemocytes seems clear, the molecular trigger that induces hemocyte aggregation in the periostial regions is unknown. The finding that hemocyte recruitment is induced by bacteria, Plasmodium, carboxylate modified latex microspheres and soluble immune elicitors suggests that multiple pathways of immune activation can induce periostial hemocyte aggregation. Studies in other insects have identified several molecular components involved in sessile hemocyte aggregation and release. For example, Noduler mediates hemocyte aggregation in Lepidoptera [38], and multiple pathways mediate hemocyte proliferation and adhesion in Drosophila [39]. However, a great amount of genomic divergence is seen in insect immune genes [40], [41], and thus, alternate pathways may be involved in mosquitoes.
What is perhaps most interesting about periostial hemocyte aggregates is their location. In dipterans, the cardiac ostia are the major incurrent valves for hemolymph entry into the heart [2], [3], [42]. Thus, hemocyte aggregation in these regions greatly increases the likelihood that hemocytes encounter circulating pathogens, and that toxic products produced during pathogen breakdown are either immediately diluted as they are swept into the rapidly flowing hemolymph or are captured by the nephrocytic PCs that flank the heart. PCs filter proteins and colloids from the hemolymph [32], and thus, it is likely that the proximity of the PCs to the periostial hemocytes is essential for the quick absorption of pathogen breakdown products. We speculate that, during the course of arthropod evolution, periostial hemocytes and PCs adapted to their current locations because of their proximity to each other in an area of high hemolymph flow.
Data from the first hour post-infection showed that circulating hemocytes bind the alary or cardiac musculature within 100 µm of the ostia and then glide into the periostial regions at velocities that are orders of magnitude slower than that of the surrounding hemolymph flow. The molecular mechanism for this process was not a focus of this study, but the movement of insect hemocytes is known to be controlled by a number of different molecular pathways, and to be governed by processes such as adhesive capture and chemotaxis [43]–[47]. The process observed here is likely a variant of adhesive capture, a process described in Drosophila where injury induces the capture of hemocytes at epidermal wound sites [43]. In agreement with the process of adhesive capture is our observation that mosquito hemocytes that reach the periostial regions originate from the circulating pool of hemocytes; video analyses show the binding of hemocytes to the alary and cardiac musculature and not the gliding of sessile hemocytes across extended distances. However, in contrast with adhesive capture in Drosophila, where individual hemocytes arrive at injury sites and do not disperse or migrate [43], mosquito hemocytes bind within the general vicinity of the ostia and then move into their final point of attachment at velocities considerably slower than that of hemolymph flow. This spatially directed gliding process is likely mediated by shear-flow dynamics, a process that has been shown to drive cell migration toward areas of high flow in phylogenetically diverse organisms [48], [49]. This process is also reminiscent of the early stages of vertebrate leukocyte extravasation in response to inflammation, where activated endothelial cells produce factors that capture circulating leukocytes, at which point the leukocytes roll in the direction of flow, and then undergo diapedesis [50], [51]. Finally, on rare occasions hemocytes arrive into the pericardial regions as small aggregates. Thus, there remains the possibility that hemocyte aggregation in circulation, or increases in hemocyte size post-infection [25], [37], [52], are physical mechanisms for an evolutionarily pragmatic response to infection.
In a manner comparable to what we describe in mosquitoes, the macrophages of many lower vertebrates aggregate in areas of high blood flow in response to infection [53], [54]. These aggregates assemble in the spleen and liver, where they concentrate the destruction and recycling of exogenous and endogenous material. Likewise, the tissue macrophages of higher vertebrates (e.g., stellate cells and Kupffer cells) are believed to have originated from the same cellular ancestors as the phagocytes found in all animals, and are also commonly located in areas of high blood flow [55]. It seems parsimonious to speculate that such phylogenetically disparate phagocyte aggregation responses are entirely based on functional analogy. However, mounting evidence suggests that these aggregation responses rely on conserved molecular and physiological components that were present in an ancient bilaterian [29]–[32], [55], [56]. From a physiological perspective, we believe the data presented here exemplifies the taxonomically widespread importance of evolutionary constraints imposed on the cellular branch of the immune system as a consequence of its long history of evolution in close association with the circulatory system.
Although it is known that most Plasmodium sporozoites rapidly die during their migration through the mosquito hemocoel [11], the specific interactions between sporozoites and hemocytes remain largely unknown. Earlier reports showed that phagocytosis of Plasmodium by hemocytes occasionally occurs [11], [37]. Our data show the in vivo interaction between Plasmodium and hemocytes, along with the first evidence of a systemic cellular immune response to late-stage malaria infection (increase in periostial hemocyte numbers). While the large number of sporozoites released by each oocyst makes it unlikely that phagocytosis is the primary component of the anti-Plasmodium response in the hemocoel, increases in melanization and periostial hemocyte aggregation suggest that hemocyte activation leads to the production of humoral factors that target Plasmodium via lytic and melanization pathways. Evidence from others supports this idea, as Plasmodium development in mosquitoes induces the transcriptional regulation of immune genes in hemocytes [57], [58], and our data on melanin deposition near hemocyte-Plasmodium interactions are in agreement with studies on the anti-sporozoite response [37], [59].
In most insects, one problem foiling hemocyte research is that no effective means of specifically staining hemocytes in vivo exists [17], [58]. As part of this investigation, we developed a CM-DiI based method that fluorescently labels hemocytes in vivo. CM-DiI is a lipophilic probe with high affinity to plasma membranes. While this probe stains all cells grown in culture, it only stains hemocytes when injected into mosquitoes. Several lines of evidence support the specificity and efficiency of CM-DiI hemocyte labeling. First, CM-DiI stains virtually all circulating hemocytes and also stains cells attached to tissues in a random pattern (sessile hemocytes). Second, CM-DiI stains cells that fit the morphological description of hemocytes given by previous authors [17], [18], [37]. Finally, the vast majority of cells that stain with CM-DiI exhibit the characteristic phagocytic signature of mosquito granulocytes. Although the mechanism by which CM-DiI specifically stains hemocytes remains unknown, our data suggest that CM-DiI is unable to cross the basal lamina surrounding internal tissues, and that it initially stains hemocytes by binding their membranes and then becoming subsequently phagocytosed. Evidence supporting this mechanism of labeling includes: (1) co-injection of formaldehyde along with CM-DiI eliminates hemocyte specificity and all tissues become labeled; (2) incubation of CM-DiI injected mosquitoes in a solution containing a detergent releases the hemocyte-captured CM-DiI and leads to the staining of other tissues; (3) injection of carbon particles prior to CM-DiI treatment blocks hemocyte staining; (4) CM-DiI staining appears most brightly as puncta within the hemocytes but over time spreads over the entire cell membrane; and (5) within minutes of mixture with PBS, CM-DiI precipitates out of solution and loses its hemocyte staining efficacy. Taken altogether, this suggests that CM-DiI could be categorized as a functional marker that stains only hemocoelic phagocytes in vivo. Techniques based on similar principles have been used for the study of macrophage biology in mammals [60]. Given the technical and practical difficulties associated with the creation of transgenic mosquito strains [61], as well as the fact that some of the more common Drosophila hemocyte markers are not encoded in the mosquito genome (e.g., hemese), the CM-DiI approach described here for the first time allows the study of mosquito hemocyte cell biology in vivo and in real time. We expect that this procedure could be adapted for the study of hemocyte biology in a broad range of insects.
In conclusion, there remain deficits in our current knowledge of hemocyte biology in adult insects, as well as in our understanding of the direct interactions between the insect circulatory and immune systems. Here, we developed new methods for the in vivo study of mosquito hemocytes and pericardial cells (nephrocytes), and applied these methods to discover a novel mosquito immune response. Namely, we uncovered periostial hemocyte aggregates, an immune tissue that is located on the surface of the mosquito heart and represents a basal component of the cellular immune response against bacteria and malaria parasites.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health, U.S.A. The protocol was approved by Vanderbilt University's Institutional Animal Care and Use Committee (IACUC; VU animal use protocols M/10/381 and M/08/041). Animals were maintained in a certified animal room and were cared for by trained personnel and veterinarians.
Anopheles gambiae (G3 strain) were reared and maintained in an environmental chamber as described [3]. Briefly, larvae were hatched in plastic containers and fed a mixture of koi food and yeast. Pupae were separated by size, allowed to develop into adults, and maintained on a 10% sucrose solution at 27°C, 75% relative humidity and a 12 h light/12 h dark photoperiod. Unless stated otherwise, all experiments were carried out on female mosquitoes at 5 days post-eclosion.
For injections, mosquitoes were cold anesthetized and a finely pulled glass needle was inserted through the thoracic anepisternal cleft. A volume of 0.2 µl was slowly injected into the hemocoel and mosquitoes were then placed back in an environmental chamber until assayed.
For bacterial infections, tetracycline-resistant GFP-expressing E. coli (modified DH5α) were grown overnight in a shaking incubator at 37°C in Luria-Bertani's rich nutrient medium (LB broth) and, unless otherwise stated, cultures were normalized to OD600 = 2 or OD600 = 4 using a BioPhotometer plus spectrophotometer (Eppendorf AG, Hamburg, Germany) prior to injection. To determine the absolute infection dose, dilutions of each OD600 = 2 and OD600 = 4 E. coli culture were plated on LB agar with tetracycline, incubated at 37°C, and the resultant colony forming units were counted 18 h later. On average, OD600 = 2 and OD600 = 4 represented infection doses of 38,000 and 131,000 E. coli per mosquito, respectively. As a non-living phagocytosis elicitor [11], 1 µm diameter FluoSpheres carboxylate modified microspheres (Molecular Probes; Eugene, OR) were also injected into mosquitoes. Microspheres were mixed with phosphate buffered saline (PBS; pH 7.0) to a final concentration of 0.08% solids per volume prior to injection.
Three soluble immune elicitors were used in this study: peptidoglycan (PGN) purified from Gram(−) E. coli, PGN purified from Gram(+) Bacillus thuringiensis, and β-1,3-glucan. For PGN purification, E. coli and B. thuringiensis were grown overnight in LB broth at 37°C. A volume of 2 ml of each bacterial culture was independently centrifuged for 1 min at 10,000 rcf, and the resulting pellets were suspended in 1 ml of PBS. Bacteria were then lysed, while on ice, by sonication for 1 min using a Branson Sonifier 450 (Branson Ultrasonics; Danbury, CT) equipped with a 3 mm tip that was set to 20% power and 30% duty cycle. Trichloroacetic acid (TCA) was added to the lysed bacteria to a final concentration of 10% v/v, the resultant solutions were incubated for 10 min at 90°C, and the PGNs were extracted by centrifugation for 1 min at 12,000 rcf [62]. The PGN pellets were then washed 3 times by resuspending in 1 ml 75% ethanol and centrifuging for 1 min at 12,000 rcf. After the final centrifugation the PGNs were resuspended in 1 ml PBS, and the purified PGNs were dissociated into soluble fragments while on ice by sonicating for 30 min at 20% power and 30% duty, with 50% rest periods every 40 sec. Following dissociation, all non-soluble material was removed by centrifugation per standard protocol [62], and PGN solutions were normalized to OD600 = 1 prior to injection.
β-1,3-glucan (microparticulate curdlan), a fungal immune elicitor, was prepared by sonicating 1% w/v curdlan (Sigma-Aldrich; St. Louis MO) in PBS for 5 min at 20% power and 20% duty, while on ice. Particulates remaining in the solution were then allowed to settle for 30 min while on ice, and the clear top phase was removed and used for injections [63]. Finally, although lipopolysaccharide (LPS) is commonly used as an immune elicitor, it was excluded from this study because it has been shown that LPS has little or no immunostimulatory effect in non-mammalian animals [64], [65]. Regardless, preliminary experiments using LPS yielded results that were qualitatively similar to the PGN experiments, but these observations could be due to the residual PGN found in LPS preparations.
Five-day-old female adult mosquitoes were starved for 6 h and then allowed to feed on a P. berghei-infected mouse with approximately 10% blood-stage parasitemia and a 1–2% gametocytemia. A control group of mosquitoes originating from the same cohort was allowed to feed, concurrently, on an uninfected mouse. Both groups were then housed in a humidified chamber at 20.5°C for 20 days prior to the assessment of the periostial cellular immune response (see below). The PbGFPCON P. berghei strain [66] was used for experiments where periostial hemocyte numbers were counted. The RedStar strain [67] was used to observe the interaction between hemocytes and sporozoites, as this strain retains a higher level of fluorescence following aldehyde fixation. The infection status of each mosquito was determined by visualizing parasites in the midgut and the salivary glands.
To stain hemocytes inside live mosquitoes, 0.2 µl of a solution consisting of 75 µM CM-DiI (Vybrant CM-DiI Cell-Labeling Solution, Invitrogen) and 0.75 mM Hoechst 33342 (Invitrogen) in PBS was injected into mosquitoes. It was crucial that this solution be injected within minutes of its preparation, as once the CM-DiI is placed in an aqueous environment its hemocyte-staining effectiveness rapidly decreases, approaching 0% after 10–15 min of mixing. After CM-DiI injection, mosquitoes were immediately returned to 27°C and 75% relative humidity for an incubation period of 20 min.
Circulating hemocytes were collected by perfusing the hemolymph onto the center of 1 cm diameter etched rings on Rite-On (Gold Seal; Portsmouth, N.H.) glass slides [25]. Cells were allowed to adhere to slides for 20 min at room temperature, fixed for 20 min with 4% formaldehyde in PBS, washed 3 times for 5 min with PBS, and coverslips were mounted with Aqua Poly/Mount (Polysciences; Warrington, PA). Visual examination of adherent perfused hemocytes was conducted using a Nikon 90i compound microscope (Nikon; Tokyo, Japan) equipped with a Nikon Intensilight C-HGFI fluorescence illumination unit and a CoolSNAP HQ2 digital camera (Roper Scientific; Ottobrunn, Germany). Cells were counted at 1000× magnification by scanning the slides from the far left to the far right until 50 hemocytes from each individual mosquito were visualized, also keeping track of the number of fat body cells observed. Intact cells were first identified as either hemocytes or fat body by confirming the presence of a nucleus using fluorescence microscopy (Hoechst 33342) and comparing cell morphology by differential interference contrast (DIC) microscopy to the descriptions of previous authors [18]. Specifically, hemocytes are significantly smaller than fat body cells and do not contain large lipid droplets. Hemocytes were then examined for the presence of phagocytosed GFP-expressing E. coli, and all cell types were examined for their incorporation of CM-DiI. Three treatments were performed (naïve, LB injected and E. coli injected), and for each treatment, hemocytes from 15 individual mosquitoes that originated from 5 independent but paired cohorts were examined (i.e., for each treatment, 3 mosquitoes per cohort).
The dorsal portion of mosquito abdomens were analyzed after the labeling of PCs, hemocytes and heart muscle. These tissues were labeled in the presence or absence of an immune challenge, and were labeled in the following combinations: (1) PCs and heart muscle, (2) PCs and hemocytes, and (3) hemocytes and heart muscle. In all experiments, Hoechst 33342 was used as a nuclear stain.
Depending on the experiment, PCs were stained using one of two novel methods. In the first, more permanent method, 0.2 µl of 0.2 mg/ml Alexa Fluor conjugated IgG (either 488 or 568 nm; Molecular Probes) in PBS was intrathoracically injected and the mosquitoes were placed at 27°C for 1 h to allow the labeled proteins to be pinocytosed by the PCs. The mosquitoes were then placed in PBS and their abdomens were bisected along a coronal plane. The dorsal half of each abdomen, sans any internal organs, was isolated, rinsed, fixed in 4% formaldehyde in PBS for 10 min, washed 3×5 min in 0.1% Tween 20 in PBS (PBST), and mounted on a glass slide using Aqua Poly/Mount. For the second PC staining method, 0.2 µl of a 0.1 mM mixture of LysoTracker Red (Molecular Probes) in PBS was intrathoracically injected and allowed to incubate for 10 min. Abdomens were then bisected along a coronal plane, rinsed in PBS, and mounted using Aqua Poly/Mount. Because LysoTracker Red stains any region that contains high lysosomal activity, it also serves as a marker for hemocyte-mediated phagocytosis of bacteria. However, LysoTracker Red cannot be aldehyde-fixed, so this method stains the PCs briefly before the dye diffuses out, and is not useful in combination with most other staining techniques.
To examine and quantify hemocytes that were adhered to tissues, 0.2 µl of a solution consisting of 75 µM CM-DiI and 0.75 mM Hoechst 33342 in PBS was injected into mosquitoes. After allowing this solution to incubate in live mosquitoes for 20 min at 27°C, mosquitoes were injected with 0.2 µl of 16% formaldehyde. Tissues were then allowed to fix for 5 min, the mosquitoes were bisected along a coronal plane, and the dorsal halves were mounted on glass slides using Aqua Poly/Mount. Immediately following mounting, CM-DiI stained hemocytes in the periostial regions were counted through the 90i's oculars at 400× or 1000× total magnification. Hemocytes were only counted if their presence was supported by both CM-DiI and Hoechst 33342 staining, so a small number of cells (<5%) might have been excluded from our counts using this conservative method. Cells were counted as periostial hemocytes only if they were attached to the dorsal vessel at the ostia, or formed part of a contiguous mass of hemocytes that were attached to this region. Numbers of periostial hemocytes were recorded on a per mosquito basis, and for each treatment cell counts were conducted on at least 12 mosquitoes that originated from no fewer than 4 independently-reared cohorts. In these experiments, trials were excluded when background staining interfered with the cellular boundaries of hemocyte aggregates in any of the treatment groups.
For the co-staining of PCs and heart muscle, 1 h after the injection of Alexa Fluor-conjugated IgG (568 nm) muscle was stained by injecting a formaldehyde-phalloidin-Hoechst-Triton×100 mixture as described [3]. Abdomens were then washed by perfusion with PBST 3 times for 5 min each, fixed a second time using 4% formaldehyde, and bisected and mounted as above. For the co-staining of PCs and hemocytes, Alexa Fluor-conjugated IgG (488 nm) was injected to stain the PCs, and after 1 h the hemocytes were stained with CM-DiI as described above. For the co-staining of hemocytes and heart muscles, CM-DiI was injected and allowed to incubate for 20 min at 27°C, and 0.2 µl of 16% formaldehyde was then intrathoracically injected to kill and preserve the mosquito. Abdomens were then bisected, suspended for 15 min in a Phalloidin-Hoechst-Triton×100 mixture [3], washed, and mounted on glass slides using Aqua Poly/Mount.
Perfused hemocytes and abdominal whole mounts were imaged between 200× and 1000× magnification depending on which phenomenon was being captured. Specimens were viewed under differential interference contrast microscopy (DIC) and/or fluorescence illumination using the Nikon 90i microscope ensemble described above, and Z-stack images were captured using a linear encoded Z-motor and Nikon's Advanced Research NIS-Elements software. To display 2 dimensional images, all images within a stack were combined to form a focused image using the Extended Depth of Focus (EDF) module of NIS Elements. For 3 dimensional rendering, Z-stacks were quantitatively deconvolved using the AQ 3D Blind Deconvolution module of NIS Elements and rendered using the volume view feature.
To quantify relative levels of E. coli fluorescence in the abdomen following infection, a series of images were captured and analyzed using the fixed thresholding feature in NIS-Elements. All images were acquired under non-saturating conditions using identical settings, and were identically thresholded to isolate only the portions containing GFP-E. coli fluorescence. For each specimen, fluorescence intensity data was collected for the dorsal abdomen, the entire abdomen (dorsal+ventral), and the periostial regions.
For intravital time-lapse video recording of the interactions between hemocytes, pathogens and the mosquito heart, hemocytes were labeled in vivo using the CM-DiI method described above. Mosquitoes were then restrained on glass slides using small strips of Parafilm “M” (Pechiney Plastic Packaging, Chicago) to gently adhere the proboscis, legs and wings (extended) to glass slides, and spheres of Parafilm were placed on either side of the abdomen to restrict side-to-side movement. Once mosquitoes were restrained, GFP-expressing E. coli were injected and fluorescence-based intravital video recording of the heart region was initiated within 10 sec of treatment. Time-lapse image sequences were captured for green (GFP-E. coli) and red (CM-DiI-stained hemocytes) channels at 5 sec intervals using the ND-experiment capture module of NIS Elements. Hemocyte tracking was then done using the manual feature of the Object Tracker module of NIS-Elements, and hemocyte velocity was calculated by dividing the path length by the total time of tracking. To reduce the potential for damage to the mosquito, light shutters were only open while each image was being acquired, and light intensity was greatly reduced by using a neutral density filter of 16.
Cell count data was separated by treatment group and tested for normality using the Kolmogorov-Smirnov test. After confirming normality, datasets with one variable and two groups were analyzed by the t-test. Datasets with one variable and more than two groups were analyzed by one-way ANOVA, and multiple comparisons were done using Tukey's post hoc test. Differences were deemed significant at P<0.05.
For statistical evaluation of the co-localization of E. coli fluorescence (GFP) and hemocyte fluorescence (CM-DiI) within perfused hemocyte samples or whole mount abdomens, images were analyzed using the Mander's overlap or Pearson's correlation coefficient (PCC) features in NIS-Elements. These measures are similar in that they describe the amount of spatial overlap between two fluorescence channels, with the main difference being that Mander's overlap corrects for differences in signal intensity while PCC does not [68].
|
10.1371/journal.pcbi.1000270 | Coordinated Concentration Changes of Transcripts and Metabolites in Saccharomyces cerevisiae | Metabolite concentrations can regulate gene expression, which can in turn regulate metabolic activity. The extent to which functionally related transcripts and metabolites show similar patterns of concentration changes, however, remains unestablished. We measure and analyze the metabolomic and transcriptional responses of Saccharomyces cerevisiae to carbon and nitrogen starvation. Our analysis demonstrates that transcripts and metabolites show coordinated response dynamics. Furthermore, metabolites and gene products whose concentration profiles are alike tend to participate in related biological processes. To identify specific, functionally related genes and metabolites, we develop an approach based on Bayesian integration of the joint metabolomic and transcriptomic data. This algorithm finds interactions by evaluating transcript–metabolite correlations in light of the experimental context in which they occur and the class of metabolite involved. It effectively predicts known enzymatic and regulatory relationships, including a gene–metabolite interaction central to the glycolytic–gluconeogenetic switch. This work provides quantitative evidence that functionally related metabolites and transcripts show coherent patterns of behavior on the genome scale and lays the groundwork for building gene–metabolite interaction networks directly from systems-level data.
| Metabolism is the process of converting nutrients into usable energy and the building blocks of cellular structures. Although the biochemical reactions of metabolism are well characterized, the ways in which metabolism is regulated and regulates other biological processes remain incompletely understood. In particular, the extent to which metabolite concentrations are related to the production of gene products is an open question. To address this question, we have measured the dynamics of both metabolites and gene products in yeast in response to two different environmental stresses. We find a strong coordination of the responses of metabolites and functionally related gene products. The nature of this correlation (e.g., whether it is direct or inverse) depends on the type of metabolite (e.g., amino acid versus glycolytic compound) and the kind of stress to which the cells were subjected. We have used our observations of these dependencies to design a Bayesian algorithm that predicts functional relationships between metabolites and genes directly from experimental data. This approach lays the groundwork for a systems-level understanding of metabolism and its regulation by (and of) gene product levels. Such an understanding would be valuable for metabolic engineering and for understanding and treating metabolic diseases.
| Cellular metabolism—the process by which nutrients are converted into energy, macromolecular building blocks, and other small organic compounds—depends upon the expression of genes encoding enzymes and their regulators. Well-characterized transcriptional regulatory circuits such as the lac and trp operons in E. coli and the galactose utilization system in S. cerevisiae illustrate how the concentration of metabolites such as tryptophan or galactose can modulate gene expression. In addition, changes in gene expression can lead to increases or decreases in the concentrations of enzymes and regulatory proteins, thereby affecting concentrations of intracellular metabolites. While individual cases of mutual regulation by metabolites and gene products have been and continue to be described, identifying the full scope of these interactions is important for improving rational control of metabolism to meet therapeutic and bioengineering objectives. Clinical scientists, for instance, may be interested in developing novel treatments that control blood glucose levels in diabetic patients, or that fight cancer by disrupting metabolism in tumor cells. This line of inquiry is also relevant to bioengineers seeking to increase the production of small molecules (such as biofuels or flavor molecules) by knocking out or overexpressing individual genes.
The simultaneous measurement of metabolite and transcript concentrations is one method that has begun to show promise for identifying gene products and small molecules involved in the same biological processes [1]. A number of studies [2]–[6] have followed the behavior of specific secondary metabolites of interest such as volatile signaling molecules [4] or compounds with pharmaceutical properties [3], as well as transcripts, in response to genetic or biochemical perturbations. The further refinement of high-throughput experimental technologies such as mass spectrometry has enabled recent studies to measure many functional classes of metabolites together with a large proportion of the transcriptome [7]–[14]. For example, one recent ground-breaking study collected extensive data on metabolite, protein, and transcript levels in E. coli following the disruption of genes in primary carbon metabolism or changes in growth rate, and concluded that metabolite concentrations tended to be stable with respect to these perturbations [15]. Another study [12] compared transcript and metabolite concentrations in S. cerevisiae under two different growth conditions, and using a novel computational method in which known metabolic pathways were divided into smaller pathways termed “reporter reactions,” the authors observed that when two different growth conditions were compared, the majority of the reporter reactions showed changes in transcript concentrations, with fewer revealing significant alterations in metabolite levels. Such methods, which make inferences based on comprehensive reconstructions of biochemical pathways in an organism, represent valuable tools for analyzing metabolomic and transcriptional data together. However, there is still a need for approaches that are designed to answer the problem of identifying novel interactions between specific gene products and metabolites that include both enzymatic and regulatory relationships.
Of prime importance to the problem of finding gene–metabolite relationships from data is the question of whether functionally-related metabolites and transcripts do indeed show coherent patterns of concentration changes that can be used to make valid predictions. Studies aimed at addressing this question have relied on computing correlation coefficients between profiles of transcript and metabolite concentrations, which can then be ranked [7] or used to co-cluster the metabolomic and transcriptomic data [8]. However, it is possible that other types of regulation, such as post-translational protein modifications and feedback inhibition, could be more predominant in the aggregate than transcriptional regulation [11]. Accordingly, a major limitation with these computational techniques is that the extent to which transcripts and metabolites are co-regulated is not known. The proportion of strong gene–metabolite correlations that are due to chance or indirect effects, as opposed to enzymatic or regulatory relationships, has also not been determined by previous investigations.
In part due to these concerns, previous work has come to contradictory conclusions about the extent of coordination between metabolite and transcript concentrations. Some qualitative evidence has been provided for the claim that transcripts and metabolites are substantially co-regulated [8],[9],[16], including the comparison of clustering patterns in each data set [8], and examples of coherent correlations between biosynthetic enzymes and their products [9]. In contrast, other studies contend that transcript and metabolite profiles tend to behave differently [10], and some have argued that correlative approaches are not specific enough to draw conclusions about which genes and metabolites are functionally related (such that the expression of a gene product controls the concentration of a metabolite, or vice versa) [11],[17].
Indeed, observed correlations within metabolic networks often confound straightforward interpretations. Metabolic networks, unlike transcriptional or protein-interaction networks, consist of molecular species which chemically interconvert. As a result, metabolites that are only distantly related in terms of the underlying pathways can show high levels of correlation [18]. This is especially true in the case of global perturbations (e.g., nutrient starvation, diurnal cycles) which affect many different branches of metabolism at once [19]. It is therefore likely that the interpretation of correlations between transcript and metabolite concentrations will depend on contextual factors, such as the branch of metabolism being studied or the experimental perturbation under which the correlations were observed.
In order to examine these questions further, we conducted a systems-level investigation of the metabolome and transcriptome of S. cerevisiae, in which we measure the dynamic responses of metabolites and transcripts to two nutrient deprivations. We examine whether transcripts and metabolites are co-regulated in general, and demonstrate the existence of a strong trend for correlated genes and metabolites to participate in related biological processes. We also demonstrate that the correlations observed for related gene–metabolite pairs are dramatically different depending on the type of metabolite and the perturbation to which the cells are subjected, and we develop a Bayesian algorithm capable of accounting for these dependencies. When applied to our experimental data, this algorithm makes gene–metabolite interaction predictions that are significantly more precise and complete than those made by correlation alone.
Transcript levels (Dataset S1, GEO accession number GSE11754) were measured via microarray following the induction of carbon starvation (glucose removal) or nitrogen starvation (ammonium removal) at 0, 10, 30, 60, 120, 240, and 480 minutes post-induction. These data complement a previously-published study that measured metabolites in Saccharomyces cerevisiae using liquid chromatography–tandem mass-spectrometry (LC-MS/MS), under the same experimental conditions [20]. Both metabolite and transcript samples were collected utilizing a filter-culture approach, which allows the rapid modification of the extracellular environment and fast quenching of intracellular metabolism and transcription 20,21.
The extent to which transcripts and metabolites show coordinated behavior in response to environmental perturbations remains an open question. It has been observed that metabolite data and transcript data cluster in similar ways [8], yet other studies have noted marked differences in the temporal dynamics of the metabolic and transcriptional responses [10]. Previous systems-level analyses have not presented quantitative evidence either for or against the similarity of the transcriptional and metabolic responses as a whole. To investigate this question, we used singular value decomposition to mathematically extract the signals in the transcriptional and in the metabolic data, and then tested how well these signals were correlated to each other.
Singular value decomposition (SVD) of the gene expression data and of the metabolite data shows that the dominant metabolite abundance patterns are closely aligned with the corresponding transcript abundance patterns (Figure 1). The first eigenvector for both genes and metabolites corresponds to a roughly monotonic starvation response that is similar across carbon and nitrogen deprivation. The second eigenvector is consistent with a nutrient-specific response, and exhibits opposite directionality between carbon and nitrogen deprivation. The third eigenvector represents a difference in dynamics between carbon and nitrogen starvation. Since neither the eigenvectors found by SVD nor the correlation analysis is sensitive to the absolute scale of each pattern in the transcript and metabolite data, the similarities described above are due to the response dynamics, and do not imply similar magnitudes of the responses. However, the magnitudes of the responses that we observed in the transcriptional and the metabolomic data appear to be comparable: the root-mean-squared fold change from the zero timepoint was 3.3-fold for metabolites and 3.1-fold for transcripts. This analysis supports the conclusion that metabolite and transcript concentrations change in quantitatively similar manners following nutrient starvation. Thus, although metabolism and transcription operate on different time scales, in the present study these two processes can be directly compared without explicitly accounting for such a temporal difference.
This result enabled us to ask whether the transcripts and metabolites that show similar dynamics tend to be biologically related: although instances of relationships between the concentrations of metabolites and related biosynthetic enzymes have been described [9],[16], other systems-level studies have noted that the majority of individual gene–metabolite correlations that they observed had no direct interpretation [11]. In order to investigate whether a trend in fact exists for metabolites involved in a certain biological process to show coordinated response patterns with related genes, we conducted a statistical enrichment analysis covering multiple metabolite classes (Materials and Methods). In this analysis, the metabolites that we measured were divided into four broad classes according to their functional role: (a) glycolysis and pentose-phosphate pathway compounds, (b) TCA cycle compounds, (c) amino acids, and (d) biosynthetic intermediates. For each of these classes, a list of associated genes was assembled, such that if a gene was significantly correlated to a metabolite belonging to a particular class, then that gene was considered to be associated with that metabolic class. Significance of correlation was assessed empirically via permutation test and corrected for multiple hypotheses, setting the false discovery rate at 0.01. To find which functions were statistically over-represented in these lists of associated genes, we then performed Gene Ontology term enrichment analysis, using the hypergeometric distribution to obtain p-values which were then Bonferroni-corrected. We selected the Gene Ontology (GO) to perform this enrichment since it has annotations for S. cerevisiae that encompass not only enzymes but also regulatory proteins, and since the ontology extends beyond metabolism to cover a wide range of other biological processes such as protein translation and the cell cycle. The full list of enriched biological processes is shown in Table 1.
Despite the complexity of the interplay between metabolism and transcription and complicating factors such as post-translational regulation, we found a strikingly logical and biologically relevant relationship between classes of metabolites and the types of gene products to which they were highly correlated. For example, the single significant enrichment result for TCA cycle compounds is the biological process “tricarboxylic acid cycle intermediate metabolism” (). Additionally, the gene products correlated to the amino acid metabolite category are enriched for “amino acid metabolism” () and “tRNA aminoacylation” (). Transcripts correlating with biosynthetic intermediates are enriched for “biosynthesis” (), among other processes, and the glycolysis and pentose-phosphate pathway compounds are enriched for “protein amino acid N-linked glycosylation” (). Not all terms show a direct relationship to the metabolite class for which they are enriched: except for the TCA cycle compounds, the profiles of metabolites in every class appear to be correlated to transcripts involved in lipid, ergosterol, and steroid metabolism, a result whose functional relevance has yet to be determined. Additionally, the profiles of the glycolysis and pentose-phosphate pathway compounds also tend to be highly correlated to the expression of genes involved in mitosis and the cell cycle. This enrichment may relate to the fact that, while yeast cells deprived of nitrogen continue to proliferate and divide over the course of an eight-hour experiment, presumably by catabolizing intracellular nitrogen sources, yeast cells starved for glucose arrest and enter stationary phase almost immediately [20].
While the above approach is adequate to reveal an overall trend for co-regulation of functionally related genes and metabolites, the nature of the co-regulation could vary depending on the experimental condition and the functional role of the metabolite involved. Furthermore, correlations between genes and metabolites can be of varying strengths, ranging from no correlation to a perfectly linear relationship between transcript concentration and metabolite concentration. These different strengths of correlation can be more or less informative about a gene–metabolite relationship, depending on the circumstance under which they are observed. For example, since amino acids and the enzymes involved in their biosynthesis and catabolism are both likely to be strongly affected by a lack of ammonium, it could be the case that instances of co-regulation between genes and amino acids under nitrogen starvation would be more meaningful than correlations of the same strength observed under carbon starvation.
In addition, correlation can be either positive (as the concentration of the gene rises, the concentration of the metabolite also rises) or negative (“inverse”—as the concentration of one rises, the other falls). The levels of related genes and metabolites could exhibit a positive correlation under one condition while having an inverse relationship or no relationship under another, due to condition-specific differences in regulation. For example, 3-phosphoglycerate (3PG) and phosphoenolpyruvate (PEP) are important in both ATP production and biosynthesis (in which they provide carbon skeletons). 3PG and PEP are known to accumulate during carbon starvation via an allosteric regulatory mechanism that prepares the cell for gluconeogenesis and the metabolism of alternate carbon sources; conversely, their abundances fall under nitrogen starvation [20]. However, many of the enzymes that use the metabolites of lower glycolysis as biosynthetic precursors are repressed under both starvation conditions, perhaps to avoid wasting limited resources. These enzymes include ILV2 (acetolactate synthase, which catalyzes the first step in isoleucine and valine biosynthesis from pyruvate) and ARO3 (which catalyzes the first step in aromatic amino acid biosynthesis from PEP and erythrose-4-phosphate). Calculating the correlations of 3PG or phosphoenolpyruvate with genes like ILV2 or ARO3 over both experimental conditions would, in effect, average two opposite behaviors: anti-correlation in carbon starvation and positive correlation in nitrogen starvation. There would be no overall correlation, although the behavior could well be consistent with a functional gene–metabolite relationship.
Condition-specific behavior is indeed observed for these gene–metabolite pairs, as well as for the pairs “ALD6 to phosphoenolpyruvate,” “GLK1 to hexose phosphate,” and “PGM2 to hexose phosphate” (Figure 2A–E, in which the concentrations of metabolites belonging to the “glycolysis and pentose-phosphate pathway” class and the concentrations of functionally related gene products are plotted against each other). Ald6p oxidizes acetaldehyde to acetate, and in addition to its key role in redox metabolism [22],[23], is involved in the production of acetyl-CoA from glycolytic end products [24]–[27]. The enzyme Glk1p phosphorylates glucose to glucose-6-phosphate in the first irreversible step of glycolysis [28], and Pgm2p catalyzes the conversion of glucose-1-phosphate to glucose-6-phosphate during the metabolism of alternative carbon sources such as galactose [29]. The metabolite “hexose phosphate” refers to glucose-6-phosphate as well as its isomers (e.g., fructose-6-phosphate, with which glucose-6-phosphate is interconverted), which were not distinguishable in the present LC-MS/MS analysis.
Overall, these glycolytic and pentose-phosphate pathway metabolites show positive correlations () with a number of related genes under nitrogen starvation but negative correlations () under carbon starvation (Figure 2A–E; representations of the nitrogen starvation data and best-fit lines using expanded x-axes can be found in Figure S2). Computing correlation across both conditions would lead to the erroneous conclusion that no relationship exists between these genes and metabolites (). In contrast, for metabolites belonging to the “amino acids” category (Figure 2F–H), related metabolites and genes tend to show strong positive correlations under both conditions: histidine and HTS1 (the histidine tRNA synthetase), methionine and MET6 (methionine synthase), and threonine and THR4 (threonine synthase) exemplify this behavior (). We have therefore developed a Bayesian algorithm capable of automatically learning and exploiting the way in which different signs and strengths of correlation can be suggestive of a functional relationship, depending on the experimental condition and the metabolite class.
Bayesian networks [30]–[32] are a general class of graphical probabilistic models. Because they allow the specification of dependencies between quantities of interest, such as relationships observed between genes and metabolites under different conditions, Bayesian networks are well-suited for leveraging such dependencies in order to make specific predictions. In these networks, variables, or “nodes,” are connected by arrows, or “edges,” indicating which variables depend on which others. Each node is parametrized by a conditional probability distribution (CPD), which describes the probability of observing the variable in a certain state, given the states of the variables on which it is dependent (for example, in Figure 3B, “gene–metabolite correlation observed under carbon starvation” is dependent on both “metabolite class” and whether a “gene–metabolite functional relationship” exists).
Our objective in constructing this Bayesian network was to formalize the concept that the strength and direction of correlation observed between a certain gene and metabolite is particular to the experimental perturbation, and depends on the functional class to which the metabolite in question belongs. We also expect to observe different correlations for metabolites and genes that are truly related than we would observe for random, unrelated gene–metabolite pairs. The Bayesian network that we constructed (Figure 3B) therefore consists of four nodes. Two of these nodes correspond to observed correlations calculated from LC-MS/MS and microarray data (“gene–metabolite correlation observed under nitrogen starvation” and “gene–metabolite correlation observed under carbon starvation”); each of these nodes can take one of five different values, depending on the strength and sign of correlation. The other two nodes (“gene–metabolite functional relationship,” which can be yes or no, and “metabolite class,” which can be any of the four metabolite classes enumerated above) correspond to intrinsic attributes of the gene–metabolite pair. To represent the dependencies described above, edges have been drawn from the node representing “functional relationship” and from the node representing “metabolite class” to both of the nodes representing gene–metabolite correlations observed under a specific experimental condition.
Given a set of positive and negative examples, the conditional probability distributions that constitute the parameters of our model can be automatically learned. These distributions are given by and , where refers to the correlation of gene and metabolite under carbon starvation, to correlation under nitrogen starvation, to whether or not a functional relationship exists between gene and metabolite , and to the class of metabolite . By Bayes' theorem, these class-specific conditional probability distributions (CPDs) are equivalent to the probability that a pair is functionally related given a certain observed level of correlation, normalized by 1) whether that level of correlation is rare or common overall and by 2) whether functional relationships are rare or common overall (i.e., and ). To learn these parameters, we calculated how often different correlations were observed for a set of gene–metabolite pairs known to be either functionally related or unrelated. Positive examples were drawn from genes and metabolites belonging to the same pathway in the Kyoto Encyclopedia of Genes and Genomes (KEGG [33]); negative examples were random gene–metabolite pairs that were not in the positive example set (see Materials and Methods for details).
A key advantage of Bayesian networks, compared to other machine-learning techniques, is that since the parameters are probability distributions, they have a direct meaning which can be informative about the system being modeled. With this in mind, the parameters and are shown in Figure 3C, for two of the metabolite classes (“glycolysis and pentose-phosphate pathway” and “amino acids”) and all possible values of , , and . Intuitively, these probabilities capture how likely an observed gene–metabolite correlation would be if the gene–metabolite pair were either related (dark grey) or unrelated (light grey). For example, in the plots on the right-hand side of Figure 3C (nitrogen deprivation data), the distribution for functionally-related pairs is shifted substantially to the right: this indicates that functionally related gene–metabolite pairs tend to be positively correlated under nitrogen starvation.
Another visualization of these conditional probability distributions is shown in Figure 3D. Here, the CPDs are collapsed into a single bar chart for each metabolite class and environmental condition by taking the log-ratio of the CPDs represented by the light and dark lines in Figure 3C. These log-odds scores are given mathematically by and . This visualization is particularly useful because it clarifies whether a particular level of correlation is more likely to be observed for a related gene–metabolite pair (above zero) or for an unrelated pair (below zero). For instance, this figure shows that for amino acids (second row), negative correlations under either condition are more likely to be observed for unrelated gene–metabolite pairs than for pairs where a functional relationship exists. The magnitude of each bar corresponds to how much more probable a particular correlation is for either related or unrelated pairs. For example, in the case of the amino acids, while a positive correlation under either experimental condition suggests a functional gene–metabolite relationship, positive correlation is more informative under nitrogen starvation than it is under carbon starvation.
The values that the network learned for these parameters indicate that the magnitude and direction of correlation between a given gene and metabolite do in fact depend strongly on that metabolite's class, as suggested by Figure 2. For instance, the amino acid methionine and the biosynthetic gene MET6, which converts homocysteine to methionine, have a clear functional relationship. Consistent with the parameters learned, methionine and MET6 exhibit a strong positive correlation under both conditions, especially nitrogen starvation (Figure 2G). In contrast, for glycolysis and pentose-phosphate pathway compounds, while related gene–metabolite pairs do exhibit positive correlations under nitrogen starvation, interacting pairs actually tend to be inversely correlated under carbon starvation. This relationship is typified by GLK1 and hexose-phosphate (Figure 2D). Additionally, when hexose-phosphate concentrations are plotted against GLK1 transcript concentrations, it is readily apparent that because hexose-phosphate and GLK1 are positively correlated under nitrogen starvation but inversely correlated under carbon starvation, they exhibit a very weak relationship when Pearson correlation is computed across both conditions ().
This pattern of positive correlation under nitrogen starvation and inverse correlation under carbon starvation is also observed for a number of other gene–metabolite pairs in our standard of examples (Figure 2A–E), including phosphoenolpyruvate (PEP) and ALD6 (Figure 2C). In terms of chemical steps, PEP is linked to ALD6 indirectly (being first converted to pyruvate by CDC19 and then to acetaldehyde via pyruvate decarboxylase, the major isozyme of which is PDC1). However, PEP, like ALD6, is predominantly cytoplasmic, whereas the intermediate species pyruvate and acetaldehyde exist in both cytoplasmic and mitochondrial pools, which could be regulated differently. This suggests that the total cellular concentrations of PEP might be more strongly related to ALD6 concentrations than would those of the other intermediate species, and furthermore that gene–metabolite pairs that are not directly linked by a single biochemical reaction may still have important functional relationships.
This type of Bayesian integration does not attempt to infer causality between changes in gene and metabolite levels. In certain cases, however, we do have a prior expectation that can explain some of the learned parameters. For example, lack of ammonium under nitrogen starvation likely leads directly to falling amino acid concentrations. Nitrogen starvation also leads to decreased activity of the transcription factor GCN4 and thus reduced expression of amino acid biosynthetic genes. Although the mechanism is not fully understood, there is evidence that the TOR pathway, which is believed to sense intracellular concentrations of glutamine [34], is responsible for causing reduced translation of GCN4 via the protein Eap1p [35]. Under carbon starvation, many transcripts may be induced or repressed by a combination of extracellular pathways for the sensing of glucose (via Ras/PKA and Snf3p) and intracellular sensing of hexose-phosphate (potentially mediated by HXK2) [36]. While these pathways are elaborate and involve many layers of regulation, it has been observed that during growth without glucose, repression involving HXK2 and MIG1 is relieved [37]. In the absence of glucose, we would expect glucose-6-phosphate, fructose-6-phosphate, and FBP levels to drop: since HXK1 and GLK1 have been shown to be under the control of HXK2-dependent glucose repression [38], this would explain the inverse correlation observed between, for example, GLK1 and hexose-phosphate.
Following parameter learning, we performed inference using the Bayesian network, which assigned to each gene–metabolite pair a confidence score. This score is equal to the posterior probability of a functional relationship, given the metabolite class and the correlations observed in the data (i.e., ). Since this value is continuous between 0 and 1, different cutoffs can be chosen depending on whether a certain application requires more precision (the fraction of pairs above the cutoff that are true positives) or more recall (fraction of total true positives with a score above the cutoff). One way to assess performance that takes this trade-off into account is to plot precision against recall for every possible cutoff, yielding a precision-recall curve (PRC). The same type of PRC can also be generated using the Pearson correlation between metabolite and gene concentrations instead of the gene–metabolite confidence score. We have employed these PRCs to compare the performance of our method relative to simply computing correlation across both experiments (Figure 4).
Given the differences between the parameters learned for distinct perturbations and metabolic classes, we expected that many physiologically relevant, specific gene–metabolite interactions that can be discovered by this Bayesian analysis would be missed by looking only at overall correlation. In agreement with this expectation, when evaluated against our set of known gene–metabolite interactions (using three-fold cross validation to avoid overfitting) and compared to Pearson correlation, Bayesian integration performs significantly better (Figure 4). It is more precise than correlation overall, and reaches twice the precision at the most stringent cut-off (the leftmost end of the curve), which corresponds to the most confidently-predicted gene–metabolite interactions.
To investigate the potential of the Bayesian network to find biologically relevant interactions beyond the set of examples, we searched for support in the scientific literature for the most confident predictions of our network (764 predicted gene–metabolite interactions, excluding those belonging to the example set derived from KEGG), as well as for 250 random gene–metabolite pairs. While many true predictions could be novel and thus unsupported in the literature, we still expect that accurate predictions would be enriched for pairs supported by existing published evidence. Each gene–metabolite pair was scored on four specific criteria (see Materials and Methods). The evaluation was performed blind to whether gene–metabolite pairs were predicted or randomly picked. Of the random pairs, only 1.2% received literature support. In contrast, 9.4% of the highly-predicted pairs were supported by at least one piece of literature evidence, an enrichment of 7.8-fold ( by Fisher's exact test; for contingency table, see Table 2). Whereas no random pair satisfied all four evidence criteria, three predicted pairs did: methionine-MET3, methionine-MET22, and methionine-MET10. These three pairs were not in our gold standard because they participate in the assimilation of sulfur into homocysteine, and although homocysteine is converted in one step to L-methionine, in the KEGG database “sulfur metabolism” does not contain the molecular species “methionine” and is a separate pathway from “methionine metabolism.” Nevertheless, MET3, MET10, and MET22 are essential for methionine biosynthesis and the knockouts are methionine auxotrophs. We also found a variety of other genes and metabolites for which there was substantial evidence: e.g., valine-PDC5 (PDC5 is involved in the catabolism of valine to isobutyl alcohol [39]), and methionine-MIS1 (MIS1 is required for the formylation of the mitochondrial initiator [40]). The full results can be found in Dataset S3. These results suggest that, despite the limited scale of the present work, our approach is capable of generalizing from our training set to find other biologically relevant gene–metabolite interactions.
A further example of the potential utility of the Bayesian approach is illustrated in Figure 5, in which we describe an interaction identified by Bayesian integration between a metabolite and a protein that regulates enzyme concentrations. This regulatory protein functions as an important part of the system that S. cerevisiae has evolved to face a fundamental metabolic challenge: namely, the diauxic shift, during which the cell changes from fermentative to respirative growth. In the first phase of growth on fermentable sugars, S. cerevisiae cultures initially grow quickly, metabolizing all the available glucose to ethanol (high ethanol concentrations are toxic to many other microbes, giving S. cerevisiae a competitive advantage). This fermentative phase is followed by a second phase of growth in which yeast cells use ethanol as a substrate, and perform oxidative respiration. The switch between these two states involves extensive metabolic and transcriptional remodeling [41]. Chief among the changes induced by the diauxic shift is the shift from using glucose to generate ATP (glycolysis), to using ethanol and ATP to make glucose and the carbon skeletons necessary for biosynthesis (gluconeogenesis). Many of the steps in both glycolysis and gluconeogenesis are readily reversible and are therefore catalyzed by the same enzymes. For this reason, it is imperative that the cell be able to commit to one pathway or the other by controlling the enzymes that are unique to each pathway, as otherwise the cell would waste energy through futile cycles. Accordingly, S. cerevisiae has evolved extensive regulatory machinery at the metabolic, transcriptional, and post-transcriptional levels that allows it to successfully negotiate this transition.
One of the key steps of glycolysis is the irreversible conversion of fructose-6-phosphate (F6P) to fructose-1,6-bisphosphate (FBP), catalyzed by phosphofructokinase (the genes PFK1 and PFK2); in gluconeogenesis, the opposite reaction is catalyzed by a separate enzyme, fructose-1,6-bisphosphatase (Fbp1p). A schematic of this pathway is given in Figure 5A. One of the top predictions made by the Bayesian network for the metabolite fructose-1,6-bisphosphate (FBP) was the gene VID24, which was not in our gold standard of examples. However, VID24 is known to play an important regulatory role in governing the gluconeogenetic enzyme Fbp1p: during the switch from gluconeogenesis to glycolysis, Fbp1p is specifically targeted to and degraded in the vacuole in a way that is dependent on VID24 [42]. This example highlights the promise of Bayesian integration to find relationships that correlation alone would miss. Pearson correlation calculated between VID24 and FBP across both conditions yields equal to just 0.03. However, as shown in Figure 4B, VID24 and FBP exhibit an inverse correlation under carbon starvation () and a strong positive correlation under nitrogen starvation (). According to the parameters learned by the Bayesian network for the “glycolysis and pentose-phosphate pathway” metabolite class, this behavior is indicative of a gene–metabolite functional relationship with a high likelihood. It is important to note that this interaction was found despite the fact that our study did not explicitly target the diauxic shift, suggesting the capacity of this method to recover diverse functional signals in the data. Moreover, this example shows that interactions can be found not only between genes encoding enzymes and the metabolites they act on, but also between metabolites and proteins that play roles in metabolic regulation.
We have generated paired transcriptional and metabolomic data that capture the dynamic responses to two perturbations over time, and find substantial evidence for the co-regulation of transcripts and metabolites. At a general level, singular value decomposition reveals that the dominant dynamic patterns exhibited by transcript and metabolite concentrations are closely aligned. Functional enrichment demonstrates that metabolites tend to show significant correlations to genes that play roles in related biological processes. Finally, using a Bayesian framework, we are able to find patterns of co-regulation between genes and metabolites that take into account both the experimental context where the correlations are observed, and the functional classification of the metabolite in question.
By analyzing metabolite and transcript data within this framework, we can identify new interactions and detect both direct and indirect regulatory relationships between a broad range of genes and small molecules. For example, we identified a regulatory link between FBP and VID24, although they exhibit almost no net correlation across both of the environmental perturbations tested (Figure 5). Additionally, our top predictions also include gene–metabolite relationships that connect metabolism to other key biological processes: for instance, methionine is known to play a unique and important role in the initiation of translation, and indeed two of our top predictions link methionine to FUN12 and GCN3, which are both involved in the formation of the 80S initiation complex that includes [43]–[45].
This type of Bayesian integration has been shown to outperform conventional correlation-based analyses (Figure 4), and the literature study suggests that we are able to find true gene–metabolite interactions outside of the gold standard. Furthermore, our ability to verify via literature search a substantial fraction of the predicted gene–metabolite pairs (Table 2) implies that Figure 4 markedly underestimates the precision of the Bayesian approach: many of the apparent “false positives” reflect real interactions that were not included in the limited set of positive examples selected from KEGG. These include both real interactions that are already known in the literature, and novel interactions to be verified in follow-up efforts. Identifying such novel gene–metabolite relationships could be used not only to drive further experimentation, but also to contribute to other modeling approaches that rely on extensive knowledge about cellular metabolic networks and their connectivity.
Despite this progress, there is undoubtedly still room for advances to be made in the accuracy of the predicted gene–metabolite interactions. For instance, advances in analytical techniques continue to allow the measurement of larger numbers of known compounds. Although the metabolite classes that we describe in the present work are broadly applicable and cover the majority of primary metabolism, they could also be extended to cover biomolecules that were not measured in the current study, such as lipids or secondary metabolites. Additionally, an increase in measured compounds could allow broader classes such as “biosynthetic intermediates” to be divided into smaller groups like amino acid intermediates or nucleotides, allowing more specific predictions to be made without the risk of overfitting based on a small number of examples. Using a larger number of classes could also help to avoid situations in which a small number of metabolites in a particular class exhibit different behavior from the majority, potentially leading to incorrect predictions for those outlier metabolites. Another area for future development is the gold standard itself, which, although certainly sufficient to make valid predictions, is still incomplete, as shown by the literature study. The gold standard could productively be combined with an extensive curation of the yeast metabolism literature, so that known regulatory as well as enzymatic interactions between genes and metabolites would then be included.
It should also be noted that the current predictions were made on the basis of only two experimental conditions. As interest in the measurement of multiple biomolecule types grows, more paired gene–metabolite data of the type presented here will continue to be published, and we imagine that these data will prove a valuable resource for integration efforts like the present work. Selected data sets that could prove particularly illuminating include metabolome and transcriptome sampling under other elemental starvations, such as phosphate and sulfur. Additionally, since prototrophic yeasts are capable of growth on a variety of carbon and nitrogen sources, monitoring gene and metabolite concentrations under these conditions could be illuminating with respect to both general (e.g., preferred vs. non-preferred nutrient sources, such as ammonium vs. proline) and specific gene–metabolite interactions (e.g., repression or activation of the GAL pathway by galactose).
As compounds from more branches of metabolism can be measured, and as data sets that track multiple biomolecule types in response to perturbations become available for more experimental conditions, analyses that are sensitive to biochemical context are likely to become increasingly critical. This work represents proof-of-concept of the potential of context-sensitive approaches for building networks relating metabolic activity and gene expression directly from experimental data.
Cultures of FY4 (a prototrophic, Mata derivative of S288C [46], Princeton strain DBY11069) were grown overnight in liquid minimal media (YNB, see below). After these overnight cultures were set back, 10 mL of early exponential phase culture (Klett 60, 1.5×106 cells/mL) was filtered onto pore-size nitrocellulose filters (82 mm in diameter). The cells (1.5×107 cells with diameter ) covered 5% of the filter surface. The filters were then placed on minimal media-agarose plates and allowed to grow for 3 h at , or approximately one doubling on the filter. To initiate the starvation time-course, the filters were transferred from the minimal plates to plates made with media lacking either ammonium (YNB-N, nitrogen deprivation) or D-glucose (YNB-C, carbon deprivation). The filter-culture approach, which allows for both rapid modification of the extracellular environment and rapid quenching of metabolism, is described in detail in previous work [20],[21].
The transcriptome and metabolome were sampled during exponential growth (before switching) and at 10, 30, 60, 120, 240, and 480 minutes following the switch to nitrogen-free or carbon-free media. Measurements of both metabolites and transcripts were collected in parallel. The metabolite measurements and extraction procedures have been previously published [20]. The observed quantitative metabolite concentration changes were verified by an independent experiment that included isotopically-labeled standards of 34 metabolites during the measurement process. This validation demonstrates that the metabolite measurements are robust to potential ion suppression artifacts and experimental noise (see Figure S1 and Brauer et al. [20]).
Experimental controls also demonstrate that the presented metabolomic and transcriptomic data are dominated by biological signal and not by noise. Raw LC-MS/MS data ( transformed ion counts) for two independent replicates of exponentially growing yeast are plotted in Figure S3. The agreement between the two samples was found to be high (y = 1.03x. ). The Lin's concordance coefficient [47], a normalized measure of the distance from the 45° line representing , where 0 is non-reproducible and 1 is perfectly reproducible, was 0.98, indicating very high reproducibility.
For the transcriptomic data, two additional negative control replicates were collected by extracting RNA from filter cultures moved to plates containing both a carbon and a nitrogen source (i.e., the same nutrient conditions as before the switch). The median standard deviation in transcript measurements collected from these replicates was found to be 0.099 ( units). In contrast, the median standard deviation for the carbon starvation timecourse was 0.45, and for the nitrogen starvation timecourse 0.46. This demonstrates that the primary source of variability in the presented data is not due to technical or biological noise.
In order to take timepoints of transcription, the filter cultures were submerged in liquid nitrogen and stored at . Yeast cells were washed from filters with 10 mL of lysis buffer, and RNA was subsequently extracted using a Qiagen RNEasy kit (QIAGEN, Valencia, CA). Oligo(dT) resin from an Oligotex midi kit (QIAGEN, Valencia, CA) was used to purify the poly(A)+ fraction from the total extracted RNA. cRNA labeled with cyanine (Cy)5- (experimental) or Cy3-dCTP (reference) was then synthesized from 1 to of the poly(A)+ RNA. The transcriptional profiles of yeast cultures at the time of harvest were measured by hybridization of the Cy3- and Cy5-labeled cRNA to an Agilent Yeast Oligo Microarray (V2). The reference sample was the zero-timepoint (taken during exponential growth prior to media switching).
Metabolite levels were normalized by cell dry weight to account for cell growth and division during the time course. As metabolites were roughly evenly-distributed between increasing and decreasing in response to nutrient starvation, no normalization for total metabolome size or total LC-MS/MS signal was required. For transcript levels, cell growth and division was accounted for by loading an approximately equivalent amount of reference and experimental RNA onto each array. This loading also normalized for decreases in total RNA pool size induced by nutrient starvation. Such normalization is useful to enable the identification of specific transcriptional regulatory events, as opposed to changes in the overall level of transcription. To correct for biases in hybridization efficiency between the Cy3 and Cy5-labeled RNA, microarray chip scans were normalized so that the total intensities across all probes in the red and in the green channels were equal.
For the purposes of our analyses, both metabolite and transcript levels were expressed as log base 2 ratios of the zero timepoint. Missing values for the transcriptional data were imputed using KNNimpute [48] with 10 neighbors, discarding any genes having more than 30% missing values; metabolites with missing values were discarded.
Singular value decomposition (SVD) is a process used to elucidate predominant patterns in large data matrices; its applications include image compression and noise reduction. SVD transforms a single data matrix into three matrices: these correspond to (i) the characteristic patterns, or “eigenvectors”; (ii) the amount of information each pattern contributes to the original data set as a whole; and (iii) the weight of each pattern for individual variables. Alter et al. [49] contains a more detailed treatment. Singular value decomposition was performed in MATLAB using the svd command.
To determine the extent of coordination between metabolism and transcription under the conditions tested, we computed the Pearson correlation between the most informative gene patterns (top eigenvectors) and corresponding metabolite patterns. We found that each of the first three gene patterns correlated significantly with the corresponding metabolite patterns, suggesting that similar overall trends were exhibited in both types of data. Significance was established via t-test ().
The root-mean-squared fold change () for each of the data types was computed according to the following formula:where is the number of timepoints (12 for either data type), is the number of molecular species measured (5373 for transcripts and 61 for metabolites), and is a particular abundance level observed at timepoint for gene or metabolite expressed as a log2 ratio to time 0. The root-mean-squared fold change was 3.1 for transcripts and 3.3 for metabolites.
For each metabolite and gene measured, we calculated the Pearson correlation between them:where and correspond to the sample variance of and , and is the sample size (i.e., total number of observations) of (or ).
We then conducted a permutation test, rearranging the columns (i.e., experimental conditions) of the metabolite data matrix 104 times to get bootstrapped p-values for these correlation values, which were then corrected using a false discovery rate of according to the procedure described by Benjamini and Hochberg [50]. The significantly-correlated genes for each metabolite were assembled into lists. We combined all the gene lists for every metabolite in a particular class (TCA cycle, glycolysis and pentose-phosphate pathway, amino acids, or biosynthetic intermediates), yielding four larger gene lists, one for each metabolite class.
The Gene Ontology (GO) [51] is a hierarchical categorization scheme for genes in several organisms, including S. cerevisiae. There are three top-level nodes, or “terms,” namely, “molecular function,” “cellular component,” and “biological process”; the majority of gene products in yeast are annotated to more specific (i.e., descendant) terms. We calculated the enrichment of these per-metabolite-class gene lists for all possible GO “biological process” terms using the hypergeometric distribution. Let be the number of class-associated genes, the number of genes in the genome, the number of genes in a GO term, and the number of class-associated genes that are also in the GO term. The p-value is then given by:where iterates from 0 to . This equation therefore yields one minus the probability of observing or fewer class-associated genes belonging to a given GO term, or equivalently, the probability of observing or greater class-associated genes belonging to that GO term.
These enrichment p-values were then Bonferroni corrected (i.e., , where is the number of tests). Since only terms containing at least one of the significantly-correlated genes were tested for enrichment, the number of hypotheses tested was 138 for the “TCA cycle” class, 611 for the “amino acids” class, 468 for the “glycolysis and pentose-phosphate pathway” class, and 620 for the “biosynthetic intermediates” class. All significant () enrichments are given in Table 1.
We assembled a “gold standard,” or a set of positive and negative examples of gene–metabolite interactions, from the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database [33]. In order to find positive examples for the metabolite classes “amino acids” and “biosynthetic intermediates,” for each distinct pathway (e.g., “arginine and proline metabolism”) as defined by KEGG, the set of reactions in a pathway was collected and then matched to the enzymes that catalyzed these reactions. To generate gene–metabolite pairs, every measured metabolite that appeared in that pathway was then paired with this set of enzymes. For example, in the pathway “arginine and proline metabolism,” arginine, ornithine, and proline are all paired with all of the enzymes involved in the catabolism and biosynthesis of arginine and proline, including arginase (CAR1), ornithine-oxo-acid transaminase (CAR2), and proline oxidase (PUT1).
In the case of the “TCA cycle” and “glycolysis and pentose-phosphate pathway” metabolite classes, a similar procedure was used. However, compounds from both of these classes are used as carbon skeletons for a wide variety of metabolites. Therefore, to improve the specificity of these positive examples, positive examples for the “TCA cycle” class were drawn only from the list of reactions in the “TCA cycle” pathway, and positive examples for the “glycolysis and pentose-phosphate pathway” class were drawn only from “glycolysis and gluconeogenesis” and “pentose-phosphate pathway.” Additionally, to properly capture the structure of glycolysis and the pentose-phosphate pathway, each of these KEGG pathways was divided into two separate subpathways: these subpathways were upper and lower glycolysis (genes and metabolites upstream and downstream of fructose-1,6-bisphosphate, respectively, with FBP itself belonging to upper glycolysis), and the oxidative and non-oxidative branches of the pentose-phosphate pathway. Matching of metabolites within a pathway to reactions and to enzymes was performed in the same way as above (because of the structure of KEGG, this included certain enzymes outside the pathway that directly acted on one of the metabolites in these pathways, such as ILV6).
Certain “distributor” metabolites (2-oxoglutarate, acetyl-CoA, ADP, AMP, ATP, L-glutamate, L-aspartate, L-glutamine, NAD+, and NADP+) were excluded from the gold standard because they are common reactants or products in a very large number of reactions. For each metabolite class, 50 times as many random gene–metabolite pairs (drawn from outside the positive example set for all metabolite classes) were picked as negative examples, so that the final gold standard was 1.96% positives and 98.04% negatives (Dataset S2).
In order to perform Bayesian integration, we first calculate the Pearson correlation of every metabolite and gene separately for each experimental perturbation. In order to ensure that these correlations are comparable between conditions, we enforce normality on the observed correlations by applying a Fisher transform:The resulting distribution is then centered by the mean and divided by the standard deviation (). This process transforms the correlation distributions observed under nitrogen and under carbon starvation to be approximately equal to a normal distribution centered around zero, with a standard deviation of one. The Z-scores are then discretized into five bins; bin edges were , so that the “strong inverse” bin contained Z-scores more than 1.5 standard deviations below the mean, the “weak inverse” bin contained Z-scores from 0.5 to 1.5 standard deviations below the mean, the “no relationship” bin contained Z-scores 0.5 standard deviations above or below the mean, and so forth. These discretized data become the input for the Bayesian networks described below.
The algorithm for finding gene–metabolite interactions is based on the Bayesian network shown in Figure 3. This network, whose structure is depicted in Figure 3B, relates the correlations observed between a gene and metabolite under each condition to (1) whether the gene and metabolite are related and (2) the class of the metabolite. More rigorously, this network specifies that, for a given gene and metabolite , the discretized correlations observed under nitrogen starvation () and under carbon starvation () are dependent on the class () of the metabolite and whether or not the gene and metabolite are functionally related (). This network is therefore parametrized by the conditional probability distributions and , along with the prior probability of a gene–metabolite relationship , which simply reflects the proportion of positive and negative examples in our gold standard for each metabolite class (see above). The conditional probability distributions and were calculated from the data using maximum likelihood [52]. In each of our examples, the value of every node is known, so this calculation reduced to counting the examples falling into each bin of correlation under nitrogen or carbon starvation for each possible value of and . These counts were then divided by the total number of observations satisfying those values of and to yield probability distributions summing to one for and .
After learning the parameters for this Bayesian network (shown in Figure 3C and 3D), we calculated the probability that a gene and metabolite were actually related given the observed correlations and the metabolite class, or . In our network, exact inference can be used to calculate :The numerator can be calculated directly from the learned parameters, and the denominator can be obtained by marginalization over .
We assessed this algorithm by generating a precision-recall curve, employing three-fold cross-validation to ensure unbiased evaluations. The gold standard was divided into random thirds; the network was then trained on two-thirds of the examples and evaluated on the remainder. This training was repeated three times, each time holding out a different third of the gold standard. Histograms of the confidence scores received by the positive and negative examples in the Bayesian integration process reveal that the positive examples from our gold standard indeed have significantly higher scores ( by Kolmogorov-Smirnov test), and can be found throughout the top predictions (Figure S1).
Our Bayesian network was trained and evaluated using the Bayes Net Toolbox for MATLAB [53].
The top 788 gene–metabolite pairs were predicted to be related by the Bayesian network with equal confidence. These top predictions were compiled and the pairs in the gold standard were removed; this yielded 764 predicted pairs. We then added 250 random gene–metabolite pairs, and analyzed the random and predicted sets together. This analysis was performed blind to whether pairs were predicted by the algorithm or randomly selected. The predictions were evaluated based on four categories:
Since relatively few genes and metabolites have been studied for interactions, we expect that the gene–metabolite pairs scored according to this evaluation will contain many false negatives, or gene–metabolite pairs for which there is no evidence simply because the relationship between those particular genes and metabolites have not yet been studied, despite the presence of a functional interaction.
“YNB” minimal media consisted of 6.7 g yeast nitrogen base without amino acids and 20 g D-glucose per 1 L. “YNB-C” carbon starvation media consisted of 6.7 g yeast nitrogen base without amino acids per 1 L, with no glucose. “YNB-N” minimal media consisted of 6.7 g yeast nitrogen base without amino acids and without ammonium sulfate and 20 g D-glucose per 1 L. 30 g of three-times-washed ultrapure agarose was added per 1 L to make agarose plates.
|
10.1371/journal.pcbi.1003707 | Signaling Domain of Sonic Hedgehog as Cannibalistic Calcium-Regulated Zinc-Peptidase | Sonic Hedgehog (Shh) is a representative of the evolutionary closely related class of Hedgehog proteins that have essential signaling functions in animal development. The N-terminal domain (ShhN) is also assigned to the group of LAS proteins (LAS = Lysostaphin type enzymes, D-Ala-D-Ala metalloproteases, Sonic Hedgehog), of which all members harbor a structurally well-defined center; however, it is remarkable that ShhN so far is the only LAS member without proven peptidase activity. Another unique feature of ShhN in the LAS group is a double- center close to the zinc. We have studied the effect of these calcium ions on ShhN structure, dynamics, and interactions. We find that the presence of calcium has a marked impact on ShhN properties, with the two calcium ions having different effects. The more strongly bound calcium ion significantly stabilizes the overall structure. Surprisingly, the binding of the second calcium ion switches the putative catalytic center from a state similar to LAS enzymes to a state that probably is catalytically inactive. We describe in detail the mechanics of the switch, including the effect on substrate co-ordinating residues and on the putative catalytic water molecule. The properties of the putative substrate binding site suggest that ShhN could degrade other ShhN molecules, e.g. by cleavage at highly conserved glycines in ShhN. To test experimentally the stability of ShhN against autodegradation, we compare two ShhN mutants in vitro: (1) a ShhN mutant unable to bind calcium but with putative catalytic center intact, and thus, according to our hypothesis, a constitutively active peptidase, and (2) a mutant carrying additionally mutation E177A, i.e., with the putative catalytically active residue knocked out. The in vitro results are consistent with ShhN being a cannibalistic zinc-peptidase. These experiments also reveal that the peptidase activity depends on .
| Hedgehog proteins are important “morphogens” that steer embryonic development in concentration-dependent ways. Despite many years of intensive research, the mechanism of morphogen action is still under debate. We have studied properties of ShhN, the actual signaling part of Sonic Hedgehog, by a comprehensive set of computational methods and based on experimentally determined molecular structures. Our work suggests surprising new features of ShhN, namely that ShhN is an enzyme that acts as a cannibalistic peptidase, and that this enzymatic function is switched off by the binding of calcium ions to ShhN. Our computational approach reveals the details of this novel switching mechanism. To test the predicted autodegradation of ShhN, we study in vitro the stability of specific mutants, and we find that these experiments are in agreement with predictions.
| Hedgehogs (Hhs) are a conserved family of secreted growth factors essential for development in bilateral animals [1]. Hh proteins realize the so-called morphogen principle [2]: they are secreted by specific cells and form extracellular concentration gradients. These are sensed by receiving cells and translated into specific cellular responses in a concentration dependent manner [3]. The basic physical process shaping the concentration gradient is the diffusion of Hh through the extracellular matrix, e.g. in the form of large oligomers [4] or other agglomerates [5]. As altered morphogen concentration may severely affect the cellular responses, the morphogen gradient has to be tightly controlled. Mathematical modeling [6], [7] has pointed to self-enhanced degradation of the morphogen as one possible mechanism to establish the predicted concentration gradients. Molecular feedback loops that lead to self-enhanced removal of Hh, e.g. by receptor–ligand internalization in the target cells, have indeed been found [8]. Another simple molecular implementation of Hh removal would be self-digestion, requiring, of course, that Hh is a peptidase.
In fact, the first crystal structure of the N-terminal domain of Sonic hedgehog (ShhN, the carrier of the morphogen function) revealed that ShhN harbors a center with a striking similarity to that of the zinc peptidase Thermolysin [9]. Later it was found that the geometry of the ShhN center closely matches a whole class of zinc peptidases [10]: the zinc is tetrahedrally coordinated by and of two histidines, respectively, an aspartate , and a water molecule. Moreover, key zinc coordinating residues are attached to a -sheet with the same topology in ShhN and all these zinc peptidases. Such a zinc center defines the LAS group of proteins (LAS = Lysostaphin-type peptidases, D-Ala-D-Ala metallopeptidases, Sonic hedgehog) [10]. For some LAS enzymes even the global fold is similar to ShhN; e.g. the VanX amino-dipeptidase has a root-mean-square deviation of 2.35 Å over 92 atoms, and its catalytic center maps onto the zinc center of ShhN including the zinc co-ordinating residues and the catalytically active glutamic acid [11]. Yet, although having the same zinc center, ShhN has been the only member of the LAS group without confirmed enzymatic activity.
One confirmed function of the ShhN zinc center is the recognition of cell-surface receptors and their antagonists, though without cleaving these binding partners [12]–[14]. For one of these ligands, Hhip, the binding mode to ShhN was found to be similar to the binding mode of the zinc metalloprotease stromelysin-1 (MMP-3) to its inhibitor TIMP-1 [12], [15] in that the inhibitor completes the tetrahedral co-ordination shell of the zinc ion.
Fuse et al. demonstrated that a crucial element of ShhN function is the direct binding to its receptor Ptc [16], but they could not find evidence for peptidase activity in assays with a variety of potential substrates, including some containing D-amino acids (often substrates of LAS enzymes). Given these tests, it is unlikely that ShhN is a broad specificity peptidase such as Thermolysin [17], [18].
A unique feature of ShhN in the LAS group is a second metal ion center with two calcium ions in the vicinity of the zinc center [19]. This double- binding site is evolutionary highly conserved and important for the specific recognition of several binding partners of ShhN [12], [13], [19]. In many proteases, binding activates the enzymatic function. Examples of such activated proteases with a catalytic include matrix-metalloproteases [20], Thermolysin [18], or Helicobacter pylori metalloprotease [21]. The relevance of binding in ShhN and other hedgehog proteins is underlined by the strong conservation of this binding site. Consequently, mutations at this site are associated with fatal developmental defects [22]–[24].
In this work we explore possible functions of the metal ion centers in ShhN. One of the hypotheses that are suggested by what we know from other calcium dependent zinc proteases is that ShhN with its LAS peptidase-like center could be a activated LAS peptidase.
To study this hypothesis and to characterize other possible roles of in ShhN, we have investigated by computational means and based on the available ShhN sequences and X-ray structures the effect of metal ions, in particular the effect of on the structure, dynamics, and electrostatics of ShhN. From these analyses we conclude that is a regulator of ShhN peptidase, though by an unexpected molecular mechanism. We present in vitro experiments testing for the predicted peptidase activity and autodegradation of ShhN by monitoring the stability of specific mutants. These experiments support the hypothesis that ShhN acts as self-degrading peptidase.
The first question that we addressed was whether the calcium ions stabilize the structure of ShhN (for the zinc ion see Figure 8 in Text S1). We investigated this question by molecular dynamics simulations of ShhN structures with 0, 1, or 2 calcium ions. Figure 1 shows that the root mean square fluctuation (RMSF, see Methods) of the protein backbone strongly depends on the number of calcium ions. While the states Ca1 (1 ) and Ca2 (2 ) can barely be distinguished, state Ca0 (no ) has overall a higher RMSF, especially in the two binding loops (residues 88–94) and (residues 128–139), and in the neighboring (residues 66–72). The flexibility pattern observed in the molecular dynamics simulations is in good agreement with that derived from crystallographic B-factors (Figure 1 in Text S1) and the experimental observation that , the loop with the highest RMSF values, is disordered in Ca0 [12]. The residue averaged RMSF values are 0.12 nm for Ca0, 0.09 nm for Ca1, , and (the latter two based on ShhN complexes with Hhip [12] and Ihog [19]). We applied Wilcoxon tests to all pairs of RMSF distributions, i.e. Ca0 vs. Ca1, Ca0 vs. , etc., and found that only Ca0 was highly significantly different from Ca1 and Ca2 (, see Table 1 in Text S1), while RMSF distributions of Ca1, , and were not clearly different at significance level . Thus, with respect to flexibility, the binding of the first switches ShhN from a more flexible state Ca0 to a more rigid state Ca1, while the binding of the second has little impact on flexibility.
The transition from Ca2 to Ca0 not only increases the flexibility of ShhN, but induces a different group of structures: the removal of the calcium ions breaks up the binding site and pushes the loops , , , which are located close to the calcium ions, away from each other (Figure 2). This is not surprising as the binding sites abound with negatively charged side chains that repel each other.
To further test the stabilizing effect of the calcium ions we simulated the molecular dynamics of states Ca0, Ca1, Ca2 based on X-ray structures of the corresponding states, and measured the root-mean-square deviation (RMSD, see Methods) between initial structures and structures sampled in simulations, expecting to see a higher RMSD for Ca0 and a lower RMSD for the calcium stabilized Ca1 and Ca2. Additionally, we simulated the molecular dynamics of artificial states Ca0, Ca1, Ca2, generated by removing calcium ions from X-ray structures of Ca2 or by introducing calcium ions into the binding pockets in the Ca0 X-ray structure. We found that the RMSDs are consistently higher (usually between 0.15 nm and 0.2 nm) for all Ca0 simulations, and consistently lower (usually 0.15 nm or lower) for Ca2 simulations, irrespective of whether the state was the original one in the X-ray structure, or artificially generated (Figure 2 in Text S1).
Since ShhN in state Ca0 was found to be more flexible than ShhN in states Ca1 and Ca2, while the latter two had similar flexibility (Figure 1), it seems that the two calcium ions are not equally important for structural stabilization. A closer inspection of the structure shows that, while one calcium is more exposed to the solvent (in Figure 2 close to D132), the other is almost completely engulfed by its ligands (in Figure 2 below E90). The latter calcium ion is the only one present in PDB entry 3n1r that was used for the simulation of the Ca1 state (blue trace in Figure 1). In the following we call this the Ca1 calcium, while the more solvent exposed calcium ion that completes the Ca2 state is called Ca2 calcium.
The observation of structure 3n1r with a single calcium at Ca1 position and a missing Ca2 ion suggests that Ca1 is the ion with the higher affinity. This is consistent with the smaller solvent accessible surface of the Ca1 ion, as there is a negative correlation of calcium affinity and solvent accessible surface [25]. Based on this empirical correlation, we obtain a rough estimate of for the difference between the affinities of Ca2 and Ca1 (see Materials and methods).
This ranking of affinities is also in agreement with the concentrations applied in crystallization and the correspondent Ca1 and Ca2 occupancies: For the crystallization of Ca2 structures 3d1m [19], 3mxw [14], and 2wfx [12], of , , and , were applied respectively, while the only Ca1 structure 3n1r [26] had a of . McLellan et al. [19] attempted to measure Ca1 affinity and estimated that it should lie above . These experiment based values narrow the range of Ca1 affinity to about to (or to . Together with our estimate of the affinity difference between Ca1 and Ca2 of this makes clear that Ca2 is very weakly bound.
Naturally, the question arises whether binding of the Ca2 calcium ion has specific effects on the structure and function of ShhN. A candidate mechanism could be the activation of the Thermolysin-like or LAS zinc peptidase function by binding of the Ca2 calcium ion.
In the following we therefore focus on the putative catalytic zinc center and investigate the effect of binding on its structure. Figure 3 shows zinc centers of several peptidases known from earlier studies [9], [10] for their high similarity to the zinc environment of ShhN. In all structures, the zinc ion is co-ordinated by nitrogens of two imidazole rings contributed by histidines, and a carboxylate group from a glutamate or aspartate residue. Both imidazole rings are fixated by hydrogen bonds between their N–H groups and neighboring hydrogen bond acceptors. In all four cases, one of these acceptors is a carbonyl oxygen, the other a carboxylate group.
We first hypothesized that addition of the Ca2 calcium makes the zinc center of ShhN more similar to the zinc center of other LAS peptidases. To test this hypothesis, we simulated the molecular dynamics of ShhN in states Ca0, Ca1, and Ca2 and compared the conformations of the zinc environments sampled in this way with the X-ray structures of LAS peptidases with PDB entries 1lbu, 2vo9, 1u10, and 1r44 (Figure 4A). Contrary to our hypothesis, the median of the RMSDs between the sampled zinc center conformations and the LAS peptidase zinc centers were significantly higher in the Ca2 state. In other words: the binding of the second calcium makes the ShhN zinc center less similar to a LAS enzyme zinc center.
The differences shown in Figure 4A between Ca0 and Ca2, and between Ca1 and Ca2 are significant (significance level of , see Table 2 in Text S1). The differences between Ca0 and Ca1 are not significant (). This means that while the Ca1 calcium ion is responsible for stabilizing the overall structure as described in the previous sections, the binding of the Ca2 calcium ion switches the zinc center from a LAS enzyme conformation to a significantly different conformation.
Although this pattern was consistent across the tested LAS enzyme structures, it could theoretically be an artifact related to the comparison of LAS enzyme X-ray structures and ShhN molecular dynamics simulations. To exclude this, we simulated the two LAS enzyme structures (PDB entries 1lbu, 2vo9) that according to their X-ray structures had zinc centers geometrically most similar to the ShhN zinc center. We then compared zinc center structures sampled by molecular dynamics simulations for both the LAS enzymes and the ShhN states (Figure 4B). We found that ShhN zinc centers in Ca0 and Ca1 were significantly more similar to LAS zinc centers than ShhN zinc centers in Ca2 (; p-values in Tables 6 and 7 in Text S1). Ca0 and Ca1 did not differ significantly in this respect, neither did different versions , . This result agrees with the previous one, again suggesting that binding of Ca2 calcium switches the zinc center of ShhN from a LAS enzyme conformation to a significantly different conformation.
By which mechanism does the binding of the Ca2 calcium switch the zinc center, which lies 1 nm away? The authors of the first X-ray structure [9] had proposed for ShhN a peptidase reaction mechanism closely related to that of Thermolysin, involving seven residues and the zinc ion. They were unaware of the fact that ShhN binds in two well-defined binding pockets. Figure 5A shows the residues of the putative catalytic center in states Ca0, Ca1, Ca2 as obtained by X-ray crystallography, and, for orientation, the positions of the calcium ions. According to the postulated reaction mechanism, E177 abstracts a proton from the catalytic water at the fourth tetrahedral co-ordination site of the zinc ion, followed by a nucleophilic attack of the on the substrate carbon.
Amongst the residues in the putative catalytic center, E127 is directly affected by binding as it co-ordinates both the Ca1 and the Ca2 calcium ions. The X-ray structures (Figure 5A) show that, as the Ca2 calcium ion binds, the carboxylate of E127 is markedly drawn towards this calcium ion. Further, there is a hydrogen bond between E127 and H135 observed in the X-ray structures and MD simulations of all three calcium binding states. Hence, as E127 is dragged towards the calcium center, it pulls H135 with it. According to Ref. [9], the second N–H of the H135 imidazole could stabilize the peptidase substrate in a conformation amenable to hydrolysis by forming a hydrogen bond with the carbonyl-O of the substrate. In Ca2, with H135 pulled away, the substrate stabilization in this critical conformation will be affected.
Opposite to H135, at the other side of the substrate, lie the zinc bound water molecule and E177, the two actually catalytically active components according to the enzymatic model in Ref. [9]. As the Ca2 calcium ion is introduced, the gap between H135 and E177 widens by about 1 Å according to the X-ray structures (Figure 5B). This observation is in agreement with the previous one that the binding of the Ca2 calcium ion is accompanied by a significant perturbation of the putative catalytic center, possibly affecting substrate stabilization.
This pattern of conformational change due to the binding of the Ca2 calcium ion is corroborated by the analysis of the MD trajectories of the different states. Figure 5C compares MD simulations of Ca0 and Ca2. It shows that the hydrogen bond between E127 and H135 is conserved between the binding states, despite the shift of E127 from Ca0 to Ca2 due to its attraction to the calcium center. Conversely, from Ca0 to Ca2 the distance widens between H135 and E177, thus opening the tight clamp previously gripping putative substrate and catalytic water. Moreover, the variance of this distance increases with the binding of the Ca2 calcium. Numerical details characterizing the distance distributions are reported in Tables 8 and 9 in Text S1.
The pulling of E127 percolates towards the zinc center also along another route. In the description of the zinc center we have mentioned that the zinc co-ordinating histidines are stabilized by hydrogen bonds with carbonyl and carboxylate groups (Figure 3). One of these carbonyl groups comes from the backbone of G128, the sequence neighbor of the co-ordinating E127. In Ca0 and Ca1, the G128 carbonyl forms a hydrogen bond with the zinc co-ordinating H141. In Ca2, as E127 is pulled towards the Ca2 calcium ion, the neighboring G128 is twisted away from the H141, and the bond between G128 and H141 is broken, the distance and distance variance is increased, and thus the zinc environment destabilized further (Figure 3, right panel, in Text S1).
H183, the other zinc co-ordinating histidine, lies distal to the binding sites and is stabilized by a hydrogen bond with the carboxylate of E54, also distal to the binding sites. Consequently, the geometry of this interaction is barely affected by calcium ion binding (Figure 3, left panel, in Text S1).
According to the enzymatic mechanism postulated by Hall et al. [9], there is another essential component of the catalytic zinc center: a water molecule occupying the fourth corner of the tetrahedron formed by the zinc ligands, as reported in the first X-ray structure [9]. This is another commonality with active LAS enzymes as noted by [10] (see Table 13 in Text S1). The accessibility of this co-ordination site to water is a necessary condition for an active enzyme with the same reaction mechanism, and it is fulfilled in ShhN.
If ShhN is a peptidase that employs the postulated mechanism, including the involvement of the water molecule at the zinc ion, and if further, this peptidase function is switched off by the Ca2 calcium ion, then we could expect that states Ca0 and Ca1 favor a water molecule at that position compared to Ca2. We have therefore analyzed our MD simulations for the behavior of water molecules close to the position of the putative catalytic water in terms of distance to the zinc ion and angles with other zinc ligands, e.g. the angle between of H148, , and water oxygen. There was in all simulations, irrespective of state (Ca0, Ca1, Ca2) almost always a water molecule at a distance of 0.2 nm to the zinc ion, i.e. at the distance of the putative catalytic water in X-ray structure 1vhh (Figure 5 in Text S1). This is not too surprising as the oxygen atom of water and the zinc have opposite partial charges, and therefore, given the water accessible zinc co-ordination site, there will always be a water molecule realizing this interaction. The angle showed a more interesting pattern. We found for states Ca0 and Ca1 that a water molecule constantly has an value very close to the crystallographically determined (Figure 6). In these states, this water molecule does not exchange with other water molecules during the simulations. The picture looks qualitatively different for state Ca2. There, the water molecule closest to the position of the putative catalytic water assumes clearly different positions, and also flips between distinct values (Figure 6). In Ca2, we could also observe exchanges of water molecules at the zinc ion. To completely define the position of the water molecule we have determined another angle (Figure 6 in Text S1), but the message remains the same, i.e. in Ca0 and Ca1 states, the water is fixed close to the experimentally observed position, while the water in Ca2 shows much more dynamics and in general is off the Ca0 crystal position. This loss of a well-defined water co-ordination also affects the flexibility and position of catalytic E177, linked to this water by a H-bond. This contributes to the widening and loosening of the H135-E177 clamp introduced above.
This behavior of the putative catalytic water molecule is consistent with the conformational switching of the zinc environment between more LAS-like in Ca0 and Ca1 to less LAS-like in Ca2. The conservation of a water molecule at this well-defined position in states Ca0 and Ca1 makes also sense in the light of a hydrolase function in these states, and the loss of this water co-ordination is in agreement with a loss of hydrolase function in state Ca2.
The observed switch from the conservation of the zinc co-ordinating water molecule in states Ca0 and Ca1, to the more liquid-like behavior of water around the zinc ion in Ca2 could be solely due to conformational changes induced by the binding of the Ca2 calcium ion (opening of clamp), or there could be also contributions by direct electrostatic interactions between the calcium ion and the zinc co-ordinating water molecule. To test whether there is a noticeable direct electrostatic interaction, we solved the Poisson-Boltzmann equation for different calcium states of ShhN, (a) in complex with the zinc co-ordinating water, and (b) with that water released into the bulk water. The corresponding affinity differences of the zinc co-ordinating water to ShhN were computed for different ShhN structures. Figure 7 shows that for each added calcium ion the electrostatic energy of the water molecule increases, i.e. the affinity of the water molecule is diminished; this is true for all ShhN structures. Thus the behavior of the water molecule as a function of calcium state observed in the molecular dynamics simulation can be partially attributed to direct electrostatic interactions between the ions and the dipole of the zinc co-ordinating water.
While the evidence presented so far is compatible with an enzymatic function of ShhN, it is unclear whether the structure of the zinc center is specific for enzymes, or whether other non-enzymes show a similar structure. To test this, we have searched with EpitopeMatch [27] the complete PDB (last access Jan 30, 2013; see Methods) for an arrangement similar to the one in the putative active site of ShhN [9], including the following eight groups: zinc ion, E127, H135, H141, D148, E177, H181, H183 (PDB entry 1vhh). Not surprisingly, all 20 X-ray structures of Hedgehogs were found with complete coverage of all eight groups and low RMSD (0 - 0.1 nm). Further, we found two matching proteins covering seven of the groups, a phosphodiesterase (1bf6) and a putative metalloprotease (3iuu). When we allowed for conservative exchanges of amino acids, another group of proteins were found also with full coverage of eight residues, all of them enzymes, including eight LAS enzymes (PDB entries 1lbu, 1qwy, 1r44, 1u10, 2b13, 2b44, 2vo9, 4f78), a hydrolase (3csq), and Thermolysin (8tln). No other proteins with a similar center were found. Thus in the set of available protein structures, the zinc center in ShhN was characteristic of enzymes.
Having an enzymatic function in an organism or not having it seems to be a fundamental property of a protein. Accordingly, one intuitively expects that the putative catalytic center of ShhN is evolutionary conserved. However, it had been noted from early on that in Drosophila the zinc binding site is not conserved [9], [28]. To assess the degree of evolutionary conservation we therefore computed a phylogenetic tree of Hedgehog proteins (see Methods) and checked where on the tree the putative catalytic center is present (Figure 8). The gross topology of the tree agrees with that of earlier trees [29] based on fewer sequences, and shows a clear branching between Drosophila hedgehog proteins and vertebrate hedgehog proteins, including Sonic, Indian, Desert, and Tiggy-Winkle hedgehog. We found that 16 out of 17 vertebrate hedgehog sequences contain the full putative EHHDEHH catalytic motif corresponding to mouse ShhN E127, H135, H141, D148, E177, H181, H183, while all of the Drosophila sequences carried instead the motif EHHTVHY, with the D148/H183 co-ordinating zinc in ShhN replaced by a threonine/tyrosine in Drosophila hedgehog. The only sequence in the vertebrate branch that has not the full active site motif is ShhN of rat. In this sequence, all zinc co-ordinating residues are present, but H181 is replaced by an arginine. H181 is not co-ordinating the zinc ion but may fixate the catalytically active E177 in a position appropriate for interaction with the catalytic water. To test this, we returned to the structural comparison of the ShhN zinc center with the zinc centers of the LAS enzymes, and we found that there is at least one confirmed LAS enzyme, L-alanoyl-D-glutamate endopeptidase of a bacteriophage (PDB entry 2vo9; [30]) that has a proline at the structural position corresponding to H181. This means that a histidine at this position is not strictly required for peptidase function, and hence rat ShhN could still be a peptidase. These results are compatible with a conservation of ShhN enzyme function in vertebrates. In the insect class the picture is less clear: while Drosophila hedgehog has no zinc center, Mosquito hedgehog has all seven residues of the putative ShhN active site [9].
Surprisingly, the carboxylate carrying residues (in mouse ShhN E90, E91, D96, E127, D130, D132) that co-ordinate the calcium ions are almost completely conserved in all 30 sequences, including Drosophila and vertebrates. The only exception is Drosophila ananassae Hedgehog protein in which E90 is conservatively replaced by an aspartate.
If ShhN is an enzyme, there should be a substrate. So far, no substrates of ShhN have been described, despite efforts to identify such substances [9], [16]. We tried to narrow the set of possible substrates by using the available ShhN structures. In our search we restricted ourselves to peptide substrates, as most of the proteins with the highest structural similarity to the putative ShhN catalytic center are peptidases or peptidoglycan amidases [10].
Hedgehog Interacting Protein (Hhip) is a functionally important binding partner at the zinc centers of ShhN and the corresponding DhhN domain in the related Desert Hedgehog protein [12]. In X-ray structures of Hhip – ShhN complexes, Hhip binds with a negatively charged patch to the zinc center. Hhip affinity to ShhN and DhhN is influenced by the calcium concentration: for high , the for binding of Hhip to ShhN and DhhN is on the order of 10 nM, for low of the order of about 100 nM [12]. This means that under the low conditions hypothesized to trigger the peptidase function, the affinity of Hhip to ShhN is lower. Moreover, Hhip replaces the putative catalytic water molecule by the carboxylate group of a glutamate side chain, similar to the binding of inhibitor TIMP-1 to its cognate zinc protease MMP-3 [13], [15]. Thus, Hhip is probably not a good substrate model.
The authors of the free ShhN X-ray structure [9] (PDB entry 1vhh) noted that in the crystal, one ShhN molecule binds to the C-terminal peptide of its crystal lattice neighbor, so that this peptide comes close to the zinc ion but without replacing the putative catalytic water. The C-terminal carboxylate oxygen forms an isosceles triangle of side length 3.4 Å with the of H135 and the zinc ion, but is only 2.5 Å away from the oxygen of the catalytic water. It is well imaginable that this situation represents the state of the peptidase before the dissociation of the product of the cleavage reaction. Except for the last amino acid, the C-terminal peptide SVAAK in 1vhh is not particularly polar. The amino group of the terminal lysine side chain is not engaged in contacts with the protein, but points away from it. This means that the putative active site binds non-polar peptides, and thus that the peptide substrate could be non-polar. If this is true, we should see for the putative enzymatically active states Ca0 and Ca1 a low electrostatic potential close to that binding site, while the potential should flip to higher values in Ca2 to allow for binding of Hhip. To test this we have taken a high resolution X-ray structure of Ca2 state (PDB: 3d1m) and computed the electrostatic potential around ShhN in Ca0, Ca1, Ca2 states, with the former two generated by removing calcium ions from the structure. Figure 9 shows that the zinc center region indeed changes electrostatic potential from slightly negative (Ca0), over slightly positive (Ca1), to high (Ca2). This means that in the putative enzymatically active states Ca0 and Ca1 the zinc center region has an electrostatic potential close to zero, suggesting non-polar substrates, similar to the C-terminal peptide of ShhN discussed above.
In functional vertebrate ShhN, the C-terminal peptide is not SVAAK as in the X-ray structure 1vhh, but SVAAK(S/T)GG, which is still compatible with a rather non-polar peptide substrate. The only known exception to this rule in vertebrates is again rat where the first G (G197) is replaced by D197. Recall that in rat we also have mutation H181R. Hence, two neutral residues, H181 and G197, are replaced by two oppositely charged residues, R and D. This could be a compensatory pair of mutations if both residues were close in space. In fact, 1vhh shows that, if we would extend the C-terminus by (S/T)(G/D)G, the G/D197 in the C-terminal peptide of one ShhN protein could be very close to H/R181 of a neighboring ShhN protein (Figure 9D). Such a compensatory mutation makes sense, if the proximity between zinc center and C-terminal peptide observed in Ca0 (1vhh) is functionally relevant, e.g. if the C-terminal peptide is a substrate of a ShhN peptidase.
As mentioned above, ShhN proteins form large oligomers [4], a situation in which many ShhN molecules are close neighbors, and thus may expose their C-termini or other hydrophobic parts to cannibalysis [18] by neighboring ShhN molecules.
Amongst the LAS enzymes, the lysostaphins have zinc centers that are most similar to ShhN. Lysostaphins are known to have slight elastase activity [31], and most of lysostaphin-like enzymes are believed to cleave glycyl-glycine or glycyl-alanine peptide bonds [10]. Hence, conserved glycines in the lysostaphin-like ShhN could be potential cleavage sites for an autodegrading ShhN peptidase.
There is only one glycyl-glycine motif in ShhN, right at the C-terminus of ShhN. This GG motif is conserved in vertebrates, i.e. in organisms with an intact zinc center, but not in Drosophila, where the zinc center is missing. Cleavage of the peptide bond between these glycines would be functionally relevant, as the C-terminal G198 carries an important cholesterol modification [32]–[34], and cleavage of the C-terminal glycine would remove this cholesterol. At the molecular level, removal of cholesterol may e.g. dissolve ShhN oligomers [35] and thus expose ShhN to other peptidases, or it could increase the mobility of ShhN [36]. Very recently, it has been reported that a fraction of ShhN may act in an unlipidated form, though it remained unclear how this form is generated [37]. The removal mechanism of the cholesterol by ShhN proposed here could contribute to this fraction.
A GG-cholesterol motif is also a noteworthy substrate candidate because of its hydrophobic nature. Remarkably, there is a sizable, shallow hydrophobic pocket close to the zinc (black ellipse in Figure 9D), extending from L140 over W173 to F48. This pocket has an electrostatic potential close to zero, independently of the binding state (Figure 9), which means that the binding of the cholesterol group could be decoupled from the peptidase activity.
To test whether the shallow hydrophobic pocket is suitable for binding cholesterol, we have applied molecular docking to the complete surface of ShhN as receptor and cholesterol as ligand. As the proposed binding pocket lies in the vicinity of the zinc ion, and the zinc is presumed to be co-ordinated in states Ca0 and Ca1 by a well-defined water molecule, we carried out two docking runs, one with this water and one without it. In both runs the hydrophobic pocket close to the zinc was identified as the dominant binding site, accommodating the cholesterol molecule in various orientations and poses (Figure 4 in Text S1). There was only one further binding site in a distance of more than 2 nm to that pocket, but this alternative binding site turned up at rank 14 or worse amongst the 20 best poses (Tables 10 and 11 in Text S1).
The binding of the cholesterol moïety in this hydrophobic pocket is compatible with the observed pair of assumed compensatory mutations between mouse (H181, G197) and rat (R181, D197): If we place G/D197 close to H/R181, the C-terminal G198 will hover just above the zinc ion, pointing with its cholesterylated C-terminus to the hydrophobic pocket (Figure 9D).
This extended hydrophobic pocket also indicates a possible mechanism for the extraction of cholesterol from the membrane. The pocket is large, shallow, and well-accessible at the surface of ShhN. Thus, ShhN may dip with this part into the membrane and recruit a cholesterol C-terminally attached to a neighboring ShhN. Similar mechanisms have been proposed for other lipid binding proteins with extended hydrophobic regions [38], [39].
The other substrate pattern suggested by comparison with lysostaphins is glycyl-alanine. There are two such GA motifs conserved in ShhN. The first lies at G58, and this GA motif is conserved in vertebrates, but not in Drosophila. The other lies at G94 and is almost conserved in all Hh proteins (including Drosophila), except for Tiggy-Winkle Hh in zebrafish where this G is replaced by N. The GA motif at G58 seems to be a better candidate as the glycine is well-accessible, and it is part of a predominantly hydrophobic peptide TLGASG that is strictly conserved in all vertebrate Hh, but not in Drosophila.
The set of potential cleavage sites discussed above could be extended to a total of 13 conserved glycines, several of which are well exposed at the surface and have hydrophobic neighbor residues.
So far, the presented arguments for the predicted calcium-regulated autodegrading ShhN peptidase have been based mainly on simulation results and on comparisons with enzyme structures, but no direct experimental evidence for this hypothesis has been available. In early tests, no ShhN peptidase activity could be detected with peptide substrates [9], [16], including peptide stretches from ShhN. It had been shown for other peptidases that the structural presentation of the peptidase substrate can be relevant [40]. Moreover, earlier tests were based on bacterial expression systems that lack the capability of adding functionally important modifications to ShhN, such as the N-terminal palmitoyl and the C-terminal cholesterol. We therefore asked whether a potential autodegrading peptidase function of ShhN may primarily target natively folded and properly modified ShhN.
For a first test, we wanted to establish a mutant ShhN () that by construction is in state Ca0 and therefore should be a constitutively active peptidase. To this end we introduced mutations E90A, E91A, E127A, expected to weaken the affinities in both calcium binding pockets (Figure 2B). In MD simulations, did not show stable calcium binding. In the putative catalytic center, the structures sampled by the MD simulations of had an RMSD that was significantly lower to the Ca0 structure 1vhh than to the Ca2 structure 3d1m ().
Further, we constructed a four-fold mutant (), carrying the three mutations of , but additionally mutation E177A, which eliminates the residue assumed to play the central catalytic role. Thus, should not be a peptidase.
If the mutations in did not abolish the cleavage site, we expect that a self-degrading peptidase is unstable in an autodegradation assay while should be stable. We tested this in vitro (data in Table 14 in ). We found that at both ShhN mutants were stable for many hours (Table 14 and Figure 9 in Text S1). At , both mutants were degraded (Figure 10). However, the decay of was much faster leading to significantly lower protein content at ( in Wilcoxon test) and (). At the latter time point, had almost completely vanished, while there was still about half of the . A least squares fit to an exponential decay (straight lines in Figure 10) yields a half-life of to for and of to for with the boundaries computed from fitted slopes standard errors.
The fact that not only but also decayed at can be attributed to other proteases in the medium that cannot be completely eliminated in this genetic experiment and that are active at this . However, the highly significant difference between and can be most easily explained as difference between an autodegrading peptidase and a knock-out of the catalytic function in by point mutation E177A.
Alternatively, we could explain the result if we assume that neither nor is a peptidase and that E177A abolishes a recognition site for another peptidase that is present in the medium. Although we currently cannot rule out this more complex hypothesis, inspection of the available structures shows that E177 lies protected in the putative substrate binding crevice of the ShhN peptidase, and it therefore seems to be not plausible that this is a recognition site for another protease.
Peptidases often have a range of optimal activity [41]. Thus, it is not surprising that the stability of ShhN in the above in vitro experiments had a strong dependence on , with a decay of ShhN at and stable ShhN at . A simple interpretation of this observation is that the ShhN peptidase is more active at than at . We therefore estimated values for all protonatable groups in ShhN based on structure 1vhh [9].
We found that the protonatable groups in the calcium binding site have values of 6 and lower (Table 12 in Text S1). This means that at neutral the binding site should be fully deprotonated. However, there are in the calcium binding site three carboxylate groups (E91, D96, E127) with between 5 and 6, indicating a tendency for protonation at and thus a weaker affinity to . Thus, ShhN peptidase could be switched on by shifting the from neutral to . For the observed difference between and , this additional switch is irrelevant since both mutants do not bind . There is only one further group with a between 5 and 6, namely H181, with a of 5.5 in Ca0 (1vhh). In the first X-ray structure [9], H181 is described as forming a charged hydrogen bond with the carboxylate of putative catalytic E177, and thus stabilizing the latter. According to our estimate, the shifting of from 5 to 6 would make protonation of H181 less likely and thus destabilize E177. This could in part explain the observed dependency of stability in our experiments (Figure 9 in Text S1).
The presence of a calcium-regulated zinc peptidase function of ShhN would explain the following observations:
1. The close similarity of the ShhN zinc center to the the centers of LAS enzymes. Apart from ShhN, all proteins with such a center are enzymes, mostly peptidases.
2. According to MD simulations and X-ray structures, the loss of calcium ions makes the ShhN zinc center significantly more similar to the zinc center of LAS enzymes.
3. In states Ca0 and Ca1, a water molecule stably co-ordinates the zinc ion, in position and pose matching catalytic waters in related enzymes. This water position and pose is disfavored in state Ca2.
4. The outcome of our comparative genetic experiments with mutants and : If our hypothesis is correct that calcium loss activates the peptidase function and that E177 is the key catalytic residue, should be active, inactive, which is compatible with our observation.
While other explanations for these observations are imaginable, the presence of a calcium regulated zinc peptidase function in ShhN seems to be the most parsimonious one. Early attempts to find peptidase function relied on bacterially expressed ShhN that is lacking functionally relevant modifications, as outlined above. These modifications are important for the formation of ShhN oligomers [5], and we have argued above that oligomerization and autodegradation could well be coupled.
In summary, we find that our computational and experimental results are in agreement with the hypothesis of ShhN as a calcium regulated autodegrading zinc peptidase. According to our model, the peptidase function should be switched off if both calcium binding pockets are occupied. The switching mechanism involves dragging of glutamic acid residue 127 towards the calcium center, a movement that is transmitted to the zinc center through several residues coupled to E127, notably histidine 135 which is bound to E127 by a hydrogen bond. The dragging of E127 pulls this H135 away from the putative substrate. Another notable effect of the binding of the second calcium is the destabilization of a catalytic water molecule that co-ordinates the zinc ion.
These conformational changes typically amount to distance changes of about to Å. Yet, such small conformational changes can leverage hundred to thousandfold changes of enzymatic activity (Ref. [42], p. 166). The propagation of functional conformational changes of this size through proteins has been observed before [43].
While our in vitro experiments are consistent with the hypothesis that ShhN is a calcium regulated autodegrading peptidase, the target motifs of this peptidase have still to be determined. By comparison with Lysostaphins we have proposed conserved GG or GA motifs as targets, e.g. the C-terminal GG-Cholesterol motif, or the GA motif at G58 that both are conserved in vertebrates, but not in Drosophila where no zinc peptidase activity is expected. The importance of these motifs for autodegradation could be tested by site-directed mutagenesis and mass spectrometry experiments.
Our discussion of the two mutations in rat ShhN compared to mouse ShhN has led us to conclude that we have a pair of compensatory mutations: , and , with the latter a direct neighbor of the C-terminal G198. Accordingly, we predict that chimeric ShhN with combinations H181/D197 or R181/G197 should be partially compromised in their peptidase function.
The predictions above are experimentally testable. It is also clear that autodegradation could be an elegant mechanism to tune morphogen gradients [6], [7]. It is currently less clear where and when in vivo the calcium regulation of peptidase function comes into play, though phylogenetic conservation of all parts of its mechanism supports its relevance.
in blood is usually tightly controlled at a relatively high level, and hence this does not seem a likely place for enzymatic activation. More promising are interstitial or intracellular compartments. In various interstitial fluids values between [44] and [45] have been reported, so that there the hypothesized enzymatic function could be switched on and off. In neural tissue, concentrations can be transiently depleted [46], [47]. ShhN could also be internalized into cells and exposed to lower there. Finally, it is conceivable that ShhN competes for calcium ions with the negatively charged proteoglycans of the extracellular matrix, and in the course of this competition temporarily loses calcium ions and gains peptidase function. If this is true, we should find that autodegradation depends on the proteoglycan composition in the extracellular space.
The in vitro experiments have pointed to a strong dependence of protease activity with the peptidase being inactive at and active at . This raises again the question where and when such acidic conditions may be fulfilled. There are acidic intracellular compartments (lysosome or late endosome), but a more interesting candidate in view of the mode of action of Hhs is again the extracellular matrix which abounds with proteoglycans rich in acidic groups that lower local . In fact, it is well-known that proteoglycan - Hh interactions are important for Hh function [48], [49].
The autocleavage of a terminal peptide at low is a feature that ShhN could share with some zymogens such as pepsinogen, thus ShhN could be an autoactivating and autodegrading zymogen tuning its own concentration gradient. This does not imply that this is the only mechanism that shapes the ShhN concentration gradient. For instance, sheddases have been shown to contribute to ShhN function [35].
Although the work presented here is consistent with the hypothesis that ShhN is an autodegrading peptidase that is switched off by the binding of the Ca2 calcium ion, this does not preclude other functions of this calcium ion. It had been shown experimentally that CDO, a mammalian receptor of ShhN, binds ShhN preferentially in Ca2 state [19]. CDO does not bind at the putative substrate binding site, but closer to the binding pocket. The electrostatics of this region depends heavily on the presence of calcium, as we could show by computing the electrostatic potential of ShhN structures in states Ca0, Ca1, and Ca2 (Figure 7 in Text S1). On the other hand, it is remarkable that in Drosophila, one of the few known cases where Hedgehog has no zinc center, the CDO analog Ihog binds to a different region of Hedgehog, independently of calcium [19]. This could indicate that at the level of the molecular interaction network, both functions, CDO/Ihog binding and ShhN peptidase activity, may be linked.
All molecular structures were retrieved from the Protein Data Bank PDB [50]. Three different states – here called Ca0, Ca1, Ca2 – of murine ShhN with respect to binding were studied: ShhN without calcium (Ca0) based on PDB entry 1vhh [9], ShhN with one (Ca1) based on PDB entry 3n1r [26], and ShhN with two (Ca2) based on PDB entries 2wfx [12] and 3d1m [19]. In the latter two structures, ShhN is bound to Hhip and Ihog, respectively, and if necessary we therefore refer to these structures as and . For consistent comparison, all structures of Ca0-Ca2 were considered between residues L40 and E189, a range common to all structures. As the few clipped residues were disordered and far from the metal binding sites, this manipulation was not considered critical.
ShhN structures were compared to structures of five representative LAS enzymes: Streptomyces albus G D-Ala-A-Ala Carboxypeptidase (PDB entry 1lbu; [51]), L-alanoyl-D-glutamate endopeptidase (PDB entry 2vo9; [30]), VanX amino-peptidase (PDB entry 1r44; [11]), and peptidoglycan amidase MepA (PDB entry 1u10; [52]).
The molecular dynamics of ShhN, ShhN mutants (, ), and LAS enzymes was simulated with the GROMACS 4.6.5 package [53]. Nonbonded interactions were calculated on a GPU (GeForce GTX 780), and bonded interactions and PME summation on four CPUs. The GROMOS96 43a1 force field was used with optimized Lennard-Jones parameters and for interactions taken from Ref. [54]. For water we used the SPC/E model, as recommended for simulations with GROMOS96 force fields in the Gromacs documentation.
Each system was simulated in duplicate using the following protocol. Initial protein structures were solvated in a rhombic dodecahedron box of SPC/E water with a minimum of 1.0 nm distance between protein and faces of box. Residues were assumed to be protonated according to their normal states at , with the exception of histidines. The protons were assigned to histidines after inspection of H-bond patterns in ShhN X-ray structures to the following nitrogens: for H134, H181, H183; for H141, and both for H135. and ions were added to neutralize the system at an ionic strength of . The Particle Mesh Ewald method was used to compute electrostatic interactions under periodic boundary conditions. Structures were energy minimized and equilibrated by Molecular Dynamics simulation for 5 ns. Production simulations were run for 100 ns with a time step of 2 fs. NPT conditions were stabilized at 300 K by V-rescale thermostat, and at 1 atm by Parinello-Rahman barostat. Bonds were constrained using the LINCS algorithm. Snapshots of the trajectories were saved every 100 ps. The last 90 ns of all trajectories did not show increasing root-mean-square-deviations (RMSDs) with respect to the starting structures, but merely fluctuations. We analyzed the trajectories with the toolbox provided by GROMACS, especially g_rms for RMSDs and g_rmsf for root-mean-square-fluctuations (RMSFs). Representative structures for the different calcium binding states in Figure 2 were extracted from trajectories with g_cluster based on mutual RMSDs; the structures shown are close to the cluster centers and represent about 90% of the trajectories. Structures in trajectories were matched with other structures using EpitopeMatch [27] (see next section). For all statistical analyses we used R version 3.01 [55].
Similarity of molecular structures was quantified with EpitopeMatch [27], version 2013.09.20 (see also http://www.EpitopeMatch.org). For two molecular structures or structures of discontinuous fragments given by their geometries (e.g. atomic coordinates) and physico-chemical properties (e.g. atom types), EpitopeMatch aligns the two structures so that the Euclidean distances between physico-chemically similar parts are minimized. The quality of the match between the two structures can be described in terms of root-mean-square-deviation (RMSD), number of matched components, or a similarity index that combines several quantities. Here we have used a matching mode of EpitopeMatch in which the two structures are aligned so that a maximum number of chemically identical atoms from both structures is superimposed, and between these the match with the lowest RMSD is selected.
EpitopeMatch was used here for three purposes: (1) to characterize similarity between LAS enzyme zinc centers taken from X-ray structures and ShhN structures taken from MD trajectories; (2) to characterize the similarity between LAS enzyme zinc centers based on their MD trajectories; (3) to identify in the complete PDB (about 88.741 structures) arrangements similar to the zinc center of ShhN.
We have used an empirical model to compute rough estimates of affinities to the two binding sites [25]. Basically, the model correlates Gibb's free enthalpy of binding with the size of the solvent accessible surface of the : the more exposed the ion, the weaker the affinity. Quantitatively, we estimate(1)with the accessible surface for a probe sphere radius of 0.5 Å. This formula was applied to both calcium ions in the three ShhN X-ray structures in state Ca2 with PDB entries 3d1m, 3mxw, 2wfx. The first approximations of the affinities thus obtained are for Ca1 and for Ca2 (errors are standard deviations from calculations on the different X-ray structures). However, these absolute values are overestimating the affinities as the empirical model does not consider interactions between the closely neighboring calcium ions, and the model does also not account for the loss of entropy due to conformational condensation, especially on binding of the first calcium ion (Figure 1). We therefore use only the difference of the above estimates, , as an rough estimate of , indicating that Ca1 is bound more tightly than Ca2.
Electrostatic potentials were computed by solving the non-linear Poisson-Boltzmann Equation with APBS [56] using a grid with Å spacing, dielectric constants of 79 and 2 outside and inside ShhN, respectively, and a water probe of radius Å. Temperature was set to , ionic strength to 0.1 mol/l NaCl outside ShhN and 0 mol/l inside ShhN. Charges and radii were assigned with PDB2PQR Version 1.7 [57] using the AMBER99 [58] force field parameters.
We retrieved ShhN structures 1vhh, 2wfq, 2wfx, 2wg4, 3d1m, 3ho5, 3m1n, 3mxw and 3n1r from the PDB. After structural alignment of the zinc ligands, we used the catalytic water coordinates from the PDB entry 1vhh and the calcium coordinates from PDB entry 2wfx for the setup of the three states Ca0, Ca1, and Ca2. Protonations of histidines were consistent with those used for the molecular dynamics simulations.
For each structure and its three states we calculated the electrostatic component of the binding free energy between catalytic water and protein using a standard thermodynamic cycle. The differences of electrostatic binding free energies between the states (Figure 7) was then calculated as:(2)(3)
PROPKA version 3.1 [59] was used to estimate the values of side chains in ShhN. Default parameters were used to perform the calculations. Coupling effects between protonatable groups were included.
From UniProtKB (last access Feb 30, 2013) we retrieved all 30 reviewed amino acid sequences of full length Hedgehog proteins. The set contained sequences of Desert Hedgehog, Indian Hedgehog, Sonic Hedgehog, and Tiggy-Winkle Hedgehog from Drosophila species and several vertebrates. The sequences were aligned with t-coffee [60], Version 8.99, with default settings. Based on this alignment we computed a distance based tree with BioNJ [61]. Branching confidence was assessed by 100 bootstraps.
The cholesterol molecule was taken from ZINC database [62] (ID: 3869467). The mol2 file was translated to PDB format using Chimera v.1.6.2 [63]. The protein structure used was PDB entry 1vhh. The docking was prepared with AutoDockTools/MGLTools v.1.5.6 (The Scripps Research Institute) by adding polar hydrogens and assigning charges to all atoms. AutoDock Vina 1.1.2 [64] was used for all docking runs. General docking parameters for both calculations were kept at their default values, except for the exhaustiveness value, which was increased to 12. The docking grid had a size of 58 Å 52 Å 52 Å and covered the entire ShhN protein.
Shh and Hh acyltransferase (Hhat) coding sequences were generated from murine cDNA (NM_009170) and human cDNA (NM_018194) by PCR. PCR products (Shh nucleotides 1-1314, corresponding to amino acids 1-438 of ShhN and Hhat nucleotides 1-1481, corresponding to amino acids 1-493) were ligated into pDrive (Qiagen, Hilden, Germany) and sequenced. mutant lacking -coordinating amino acids E90, E91 and E127 (mouse Shh nomenclature) was generated by site directed mutagenesis (Stratagene, La Jolla, USA). mutant, additionally lacking putative catalytic residue E177, was also generated by site directed mutagenesis. For primer sequences see Table 15 in Text S1. Both cDNAs, together with Hhat were cloned into pIRES (ClonTech, Mountain View, USA) for bicistronic ShhNp/Hhat co-expression in Bosc23 cells, a HEK293 derivate. This resulted in the secretion and proteolytic processing of N-palmitoylated and C-cholesterylated ShhNp termini, respectively.
Bosc23 cells were cultured in DMEM (PAA, Cölbe, Germany) with 10% fetal calf serum (FCS) and penicillin/streptomycin and transfected using PolyFect (Qiagen). After cells had been cultured for 24 hours, media were removed, cells washed with -free PBS, and proteins secreted into serum-free/-free EpiLife media (Gibco, USA) for 16 h. Media were harvested, ultracentrifuged, the adjusted to and , and incubated at for 3 h, 6 h, and 18 h before Trichloroacetic acid (TCA)-precipitation. All proteins were analyzed by 15% SDS-PAGE, followed by Western blotting using PVDF-membranes. Ponceau-S staining of PVDF-membranes after blotting or Coomassie-staining of gels was conducted as loading controls. Blotted proteins were detected by polyclonal -Shh antibodies (goat IgG; R&D Systems, Minneapolis, USA). Incubation with peroxidase-conjugated donkey--goat IgG (Dianova, Hamburg, Germany) was followed by chemiluminescent detection (Pierce). Signals on blots were quantified with ImageJ.
|
10.1371/journal.ppat.1000408 | The Defective Prophage Pool of Escherichia coli O157: Prophage–Prophage Interactions Potentiate Horizontal Transfer of Virulence Determinants | Bacteriophages are major genetic factors promoting horizontal gene transfer (HGT) between bacteria. Their roles in dynamic bacterial genome evolution have been increasingly highlighted by the fact that many sequenced bacterial genomes contain multiple prophages carrying a wide range of genes. Enterohemorrhagic Escherichia coli O157 is the most striking case. A sequenced strain (O157 Sakai) possesses 18 prophages (Sp1–Sp18) that encode numerous genes related to O157 virulence, including those for two potent cytotoxins, Shiga toxins (Stx) 1 and 2. However, most of these prophages appeared to contain multiple genetic defects. To understand whether these defective prophages have the potential to act as mobile genetic elements to spread virulence determinants, we looked closely at the Sp1–Sp18 sequences, defined the genetic defects of each Sp, and then systematically analyzed all Sps for their biological activities. We show that many of the defective prophages, including the Stx1 phage, are inducible and released from O157 cells as particulate DNA. In fact, some prophages can even be transferred to other E. coli strains. We also show that new Stx1 phages are generated by recombination between the Stx1 and Stx2 phage genomes. The results indicate that these defective prophages are not simply genetic remnants generated in the course of O157 evolution, but rather genetic elements with a high potential for disseminating virulence-related genes and other genetic traits to other bacteria. We speculate that recombination and various other types of inter-prophage interactions in the O157 prophage pool potentiate such activities. Our data provide new insights into the potential activities of the defective prophages embedded in bacterial genomes and lead to the formulation of a novel concept of inter-prophage interactions in defective prophage communities.
| Bacterial viruses, known as bacteriophages or phages, are major factors promoting horizontal gene transfer (HGT) between bacteria, and this activity has sparked new interest in light of the discovery that many sequenced bacterial genomes harbor multiple prophages carrying a wide range of genes, including those related to virulence. However, prophages identified from genome sequences often contain various genetic defects, and they have therefore been regarded as merely genetic vestiges, with no attention paid to their potential activities as mobile genetic elements. Enterohemorraghic Escherichia coli O157, which harbors as many as 18 prophages, is the most striking such example. The O157 prophages carry numerous genes related to O157 virulence, but most possess multiple genetic defects. In this study, we analyze the functionalities of O157 prophages and report that many of the apparently defective prophages are inducible and released from the O157 cells as particulate DNA and that some can be transferred to other E. coli strains. We should therefore regard these prophages as having high potential to disseminate virulence determinants. Our results further suggest that their activities as mobile genetic elements are potentiated by various types of interactions among the prophages, formulating a novel concept of inter-prophage interactions in defective prophage communities.
| Horizontal gene transfer (HGT) is a major mechanism involved in bacterial evolution. In HGT between bacteria, viruses known as bacteriophages (or phages) play particularly important roles as gene transfer vehicles [1],[2]. Incoming temperate bacteriophages parasitize their hosts by integrating their genomes into the host genetic material. The additional genetic information that they provide to the host bacterium encodes various novel abilities, such as niche adaptation and the production of new virulence factors [2],[3]. Although phage-mediated HGT was first described in the 1950s in the conversion of Corynebacterium diphtheriae strains that did not produce a toxin to strains that did [4], studies in recent decades have identified a number of virulence determinants carried by phages [1]–[3], [5]–[7]. Furthermore, because numerous bacterial genomes have been sequenced, it has become increasingly clear that many bacterial genomes contain multiple prophages carrying a variety of genes [8]. However, the prophages identified from the genome sequences often contain genetic defects, such as deletions or disruptions of genes required for phage induction and propagation. Thus, such prophages are regarded simply as genetic remnants, and investigators tend to ignore the possibility that they might function as mobile genetic elements or participate in HGT.
Enterohemorrhagic Escherichia coli (EHEC) comprise a distinct class of E. coli strains that cause diarrhea, hemorrhagic colitis, and hemolytic uremic syndromes [9]. Among the various EHEC strains, the most dominant are the strains of serotype O157:H7 [10]. The genome of EHEC O157:H7 strain Sakai (referred to as O157 Sakai) contains 18 prophages (Sp1 to Sp18) and 6 prophage-like elements (SpLE1 to SpLE6), amounting to 16% of the total genome [11],[12]. These Sps and SpLEs have carried many virulence-related genes into the O157 Sakai genome, including the Shiga toxin genes (stx1 and stx2), a set of genes for a type III secretion system (T3SS), numerous T3SS effector proteins, and transcriptional regulators for T3SS gene expression [11],[13]. A recent genomic comparison of O157 strains has further revealed that variation in prophage regions is a major factor generating the genomic diversity among O157 strains [14],[15].
An initial analysis indicated that, among the 18 Sps, 11 (Sp3–Sp6, Sp8–Sp12, Sp14 and Sp15) retain features of lambdoid phages, one (Sp13) has features similar to those of P2, one (Sp1) contains P4 features, and one (Sp18) retains Mu features. The other four Sps (Sp1 Sp7, Sp16, and Sp17) were unable to be assigned to particular phage families due to their chimeric or highly disrupted genomic backbones. Most of the lambdoid prophages resemble one another and contain various genetic defects ranging from frame-shift mutations to deletions and insertions of so-called insertion sequence (IS) elements [12]. Thus, a functional analysis of the prophage pool of O157 Sakai could reveal whether defective prophages have any biological activity and, perhaps more importantly, whether they have the potential to disseminate virulence factors among bacteria.
In the present study, we used bioinformatic analyses to re-evaluate the genomic structures of each Sp and to define their genetic defects by comparing them with their respective well-characterized prototype phages. We then systematically analyzed each Sp for its ability to excise itself from the host genome, replicate, and package its phage DNA. Our results indicate that many of the apparently defective prophages can excise themselves, replicate, and be released from O157 cells as particulate DNA. Furthermore, using Sp derivatives carrying a chloramphenicol-resistance (CmR) gene, we demonstrated the transferability of apparently defective prophages to other E. coli strains. Our data indicate that defective prophages in the O157 prophage pool are not simply genetic remnants but have significant potential to act as mobile genetic elements that can mediate the spread of virulence-related genes from O157 to other bacteria. The results further suggest that various inter-prophage interactions in the prophage pool potentiate the biological activities of the defective prophages.
The results of a genomic comparison of 18 Sps with their corresponding prototype phages are summarized below (Figure 1 and Table 1; see Figure S1 for more detail).
i) Lambdoid prophages: Among the 11 prophages with well-conserved lambdoid features, Sp3 and Sp8 lack repressor and anti-repressor functions (CI and Cro), which are at the center of the regulation of lysogenization and induction of lambdoid phages [16]–[19]. In the other lambdoid prophages, the repressors of Sp11 and Sp12 have been disrupted and those of Sp4, Sp6, Sp9, and Sp10 lack a peptidase domain that is required for SOS-induction (Figure 1A and Figure S2).
Integrase (Int), which mediates a bidirectional process of phage genome integration and excision [20], has been disrupted in Sp3, Sp11, and Sp12 (Figure 1A). Although Sp4 and Sp14 appear to encode intact integrases, their putative excisionase (Xis) proteins that regulate the directionality of the Int function [21] lack DNA binding motifs (data not shown) and are probably non-functional.
Replication initiator protein O and elongation protein P [22]–[24] are apparently functional in all lambdoid prophages except for Sp3 (Figure 1A). Whereas six (Sp6, Sp8, Sp9, Sp11, Sp14 and Sp15) have λ-type helicase loaders, three (Sp4, Sp10 and Sp12) have DnaC-type helicase loaders and Sp5 has an elongation protein almost identical to that of phage HK022, which belongs to the DnaB family (Figure S3).
Three Sps (Sp3, Sp5, and Sp15) have a general recombination system similar to that of phage λ, consisting of Exo, Bet, and Gam proteins (Figure 1A), but those of Sp3 and Sp15 have been disrupted. Seven Sps (Sp4, Sp6, Sp9–Sp12, and Sp14) possess a different type of recombination system that contains enterobacterial exodeoxyribonuclease VIII (Exo VIII)-type proteins instead of the λ-type exonuclease, but this system is intact only in Sp10.
All lambdoid Sps, except for Sp4, encode intact terminase, the key enzyme for DNA packaging [25],[26]. In Sp4, the nu1 gene for the terminase small subunit protein has been disrupted by an IS insertion (Figure 1A).
The morphogenesis regions of Sp3, Sp4, Sp8, Sp11, Sp14, and Sp15 follow the gene organization of λ [18],[26], but Sp6, Sp9, Sp10, and Sp12 exhibit slightly different gene organizations in the head formation region (Figure 1A). Of these 10 Sps, all genes for morphogenetic function are conserved only in three (Sp8, Sp10, and Sp11). Sp15 (Stx1 prophage) also contains multiple defects in the morphogenic functions. Most of the putative morphogenic genes of Sp5 (Stx2 prophage) differ from those of λ and remain uncharacterized. However, another O157 Stx2-converting phage, called 933W, contains a set of genes nearly identical to that of Sp5 and has been shown to be fully active [27].
All 11 lambdoid Sps encode the Q protein, a regulator of late transcription. A full set of genes for cell lysis is also present in all lambdoid Sps, although some variation in gene organization is observed (Figure 1A). In Sp5, an IS-insertion has occurred upstream of the lysis region, but it has not disrupted any protein-coding genes; indeed, an Sp5 derivative has been shown to transfer from O157 Sakai to K-12 [28].
Based on our in silico analysis of the 11 lambdoid Sps, we predicted that (1) three (Sp3, Sp11, and Sp12) would no longer be able to excise themselves from the host chromosome and (2) the other eight prophages would be excisable, but seven (Sp4, Sp6, Sp8–Sp10, Sp14, and Sp15) would have some defects in morphogenesis or other functions, and only Sp5 would be capable of the full complement of viral functions (Table 1).
ii) P2 and P4-like prophages: Phage P2 functions as a helper for the satellite phage P4 [29]–[33]. The P2–P4 couple in O157 Sakai is atypical because homologues of the P2 ogr gene and the P4 ε gene [30],[31] are present in Sp2 (P4-like phage) and Sp13 (P2-like phage), respectively (Figure 1B and 1C). In addition, Sp13 lacks most phage functions, including most of the morphogenetic genes (Figure 1B and Table 1); thus, it may no longer propagate by itself or work as a helper for Sp2.
iii) Mu-like prophage: Sp18 is predicted to be intact and spontaneously inducible like the prototype Mu phage because most features for Mu-like phages [34],[35] are conserved (Figure 1D). Sp18 also contains an invertible host-specificity region, but the encoded genes are distinct from those of Mu [35].
iv) Others: The other four prophage genomes have been severely degraded, but all may have been derived from lambdoid phages because many of their residual genes are homologous to λ genes (Figure 1E). Reassessment of the Sp17 region revealed that two prophages (referred to as Sp17a and Sp17b) have been integrated in tandem in this region. Sp7 shows interesting chimeric features of lambdoid and P4-like phages. Sp7 encodes a P4-like Int, as well as a P4-like Xis (Vis-homologue) [36]. In addition, Sp7 contains a gene similar to the P4 α gene, but the gene has been disrupted by multiple frame-shift mutations.
To experimentally evaluate the inducibility of each Sp, we first examined the amplification of prophage DNA upon mitomycin C (MMC) treatment of O157 Sakai cells using an oligo DNA microarray. Cell lysis started 2 to 3 hr after the addition of MMC (1 µg/ml) to the early log-phase culture, and the optical density (OD) returned to basal levels within 6 to 8 hr as a result of cell lysis (Figure S4). We isolated total cellular DNA from aliquots of cultures at 1-hr intervals from 0 hr to 4 hr after the addition of MMC. We then analyzed the total DNA using the microarray (Figure 2A).
We observed selective amplification of Sp5, Sp13 and Sp15 regions, although amplification of the Sp13 regions was delayed relative to that of the other two regions. Interestingly, the Sp5-flanking regions exhibited significant amplification (R1 and R2 in Figure 2A). The Sp15-flanking regions also showed substantial amplification (R3 and R4 in Figure 2A), but the amplification of the R3 and R4 regions was asymmetric and smaller than that for the Sp5-flanking regions. A similar phenomenon in phage λ is known as the “regional replication” of prophage-flanking regions. In “regional replication,” upon induction, the chromosomal regions flanking the λ prophage genome replicate together with the λ genome that remains to be excised from the chromosome [37]. Similar phenomena have also been reported for other lambdoid phages [38]. Importantly, the amplified prophage-flanking regions of Sp5 and Sp15 included prophages Sp4 and Sp14, respectively (Figure 2A). Our preliminary microarray analysis of a spontaneous Sp5 deletion mutant confirmed that the amplification of the Sp5-flanking region in response to MMC treatment requires the presence of Sp5 (data not shown).
We also analyzed transcriptional changes of the prophage genes upon MMC treatment using the microarray (Figure 2B). A large number of chromosomal genes were up- or down-regulated by MMC treatment (data not shown), as described in K-12 [39],[40]. Among the prophage genes, we observed a marked increase in the transcript levels of the Sp5, Sp13, and Sp15 genes, especially in their early regions, which is in agreement with their selective DNA amplification upon MMC treatment (Figure 2A). No significant transcriptional changes in response to MMC treatment were detected for most genes of other prophages, except for those of Sp1. Although this highly degraded lambdoid prophage showed no DNA amplification upon MMC treatment (Figure 2A), many of its residual early genes exhibited clear induction. The biological significance of this phenomenon is unknown.
To determine whether the Sps amplified by MMC treatment are excised from the host chromosome into a circular form and whether other Sps are excisable but to a much lesser extent, we looked for the presence of circular forms of all Sps using PCR amplification of the attachment site (attP)-flanking regions that are generated by excision and circularization (Figure 3). In MMC-treated O157 Sakai cells, we detected circularized DNA not only from Sp5, Sp13, and Sp15 but also from Sp6, Sp7, Sp9, and Sp10 (Figure 3B). In addition, the circular forms of Sp4 and Sp14 genomes were detected, although the amount of circularized Sp14 was significantly lower than those of the other prophages. Furthermore, we also detected the circular form of DNA for all of these prophages, except for Sp4 and Sp14, in O157 Sakai cells that had not been treated with MMC. These results indicate that the nine prophages can be excised into a circular form by MMC-mediated or spontaneous induction and that the Sp4 and Sp14 genomes, which were amplified by regional replication, can also be excised and cyclized. Circularized Sp18 DNA was not detected, which is consistent with previous results for the prototype Mu phage [34],[35].
By sequencing the PCR products obtained in this analysis, we confirmed that the circularized prophage genomes arose by site-specific recombination between the left and right phage attachment sites (attL and attR). This analysis also allowed us to precisely determine the core attachment sequences of these nine prophages (Figure S5). The results largely agreed with our previous predictions, except for the results with Sp9 [12]. The attL and attR sites of Sp9 are each located 121 bp upstream of the predicted positions, and we identified the true core sequence of 28 bp.
We also used the same strategy to analyze six prophage-like elements (SpLEs) of O157 Sakai. However, we detected no excised and circularized DNA from any of these elements, either in untreated cells or in cells treated with MMC (data not shown). This suggests that these elements have no (or very low) mobility or require other types of stimuli to be mobilized.
Because we found that many Sps are excised into circular forms, we quantified the circularization and/or replication of these Sps using quantitative PCR (qPCR) (Figure 3). In untreated cells, we detected similar amounts of circularized DNA of Sp5, Sp6, Sp7, Sp13, and Sp15 and a slightly lower level of circularized Sp9 DNA. In cells treated with MMC, the relative amounts of circularized DNA of Sp5, Sp13, and Sp15 increased to approximately 300, 30, and 40 times higher than their levels in untreated cells, respectively (Figure 3C). In contrast, the levels of Sp6, Sp9 and Sp10 were lower than in the untreated cells. It is noteworthy that, in complete agreement with the results of the qualitative PCR analysis (Figure 3B), considerable amounts of circularized Sp4 and Sp14 were generated in the MMC-treated cells, whereas hardly any was detected in the untreated cells. Sp18 also proved to be non-inducible by MMC, as described for phage Mu [35].
DNA microarrays were used to monitor prophage induction [38],[41]. Microarray analysis can provide a gross image of the replication pattern of prophages augmented by MMC treatment, as seen in Figure 2, but it cannot detect spontaneous induction of prophages. Thus, as our present data show (Figure 3, see also Figure S6), qPCR analysis is required to obtain the true picture of prophage induction. This may also be true in transcriptome analysis of prophage genes (Figure 2B).
In general, the stability of a prophage is tightly coupled with the physiology of the host cell. Under conditions that generate DNA injury—in the present study through MMC treatment—prophages are de-repressed by a RecA-mediated mechanism (the SOS response) to enter the lytic pathway [42]. The RecA protein stimulates self-cleavage of the repressor protein, which leads to the expression of genes required for the lytic pathway [43],[44]. The non-inducible nature of Sp6, Sp9 and Sp10 by MMC treatment is consistent with the fact that the peptidase motif is missing in the repressors of these lambdoid Sps (Figure S2). Phage P2 is insensitive to the SOS response and is thus non-inducible by MMC treatment, because its repressor intrinsically lacks the peptidase motif [30]. The repressor of Sp13, a P2-like prophage, also lacks the peptidase motif. Thus, the MMC-mediated induction of Sp13 observed in this analysis is remarkable (Figures 2, Figure 3, and Figure S6). Although the mechanism is yet to be elucidated, a P4 ε-like gene encoded on Sp13 (Figure 1 and Figure S2) may be involved in this unique behavior because the P4 ε gene product de-represses the P2 genome by binding to the P2 repressor [30],[31].
To investigate whether the Sps that were circularized and replicated by spontaneous or MMC-mediated induction could be packaged into phage particles, we first attempted field inversion gel electrophoresis (FIGE) analysis of DNA isolated from phage particles. Particles were taken from supernatants of bacterial cultures that were either treated with MMC or left untreated (Figure 4A). In the untreated sample, we detected packaged DNA of Sp5, Sp10, and Sp18. However, upon MMC treatment, a large amount of Sp5 DNA accumulated and generated extensive smearing that prevented the visualization of minor species of packaged phage DNA.
We therefore quantified particulate DNA of each Sp by qPCR using the same set of PCR primers used for the quantification of intracellular phage DNA (Figure 4B). In the untreated sample, we detected DNase-resistant forms of DNA for at least five prophages (Sp5, Sp7, Sp9, Sp10, and Sp15). They included two Sps (Sp9 and Sp15) that contain genetic defects in head formation (Figure 1 and Table 1). This result suggests that the defects of these two prophages were complemented, probably by other prophages that provided all or some of the gene products required for head formation. The amount of Sp6 DNA was marginal compared with the control chromosomal DNA (CB1 and CB2 in Figure 4B). The Sp13 DNA appears to be inefficiently packaged or unstable. As expected from the data on phage Mu, the Sp18 DNA was efficiently packaged.
In the MMC-treated sample, a large amount of packaged Sp5 DNA was detected, reaching levels of ≥1010 molecules per milliliter of culture. Although at much lower levels, we detected considerable amounts of packaged DNA from at least six other Sps (Sp4, Sp7, Sp10, Sp13, Sp15, and Sp18). Interestingly, this group includes Sp13, which also lacks the genes for head formation and DNA packaging (Figure 1 and Table 1). Of the two Sps (Sp4 and Sp14) whose genomes are amplified only by the regional replication of Sp5 and Sp15, respectively, we detected packaged Sp4 genomic DNA, although it also contains defects in head formation and DNA packaging. Thus, these defects of Sp13 and Sp4 must have also been complemented by other prophages.
To examine the transferability of packaged Sp genomes, we marked the eight Sps (Sp4–Sp7, Sp9, Sp10, Sp13, and Sp15) by replacing “moron” genes of each phage genome, which are not required for phage propagation, with a CmR gene cassette (Figures 5A and 5B). The stx2 gene in Sp5 and stx1 gene in Sp15 were replaced with the cassette. Incorporation of the CmR cassette into the prophage genomes did not affect DNA packaging because the levels of particulate phage DNA detected for each CmR-derivative were similar to those observed for the wild-type O157 Sakai (data not shown).
To examine whether the CmR marker can be transferred to two K-12 derivatives (strains MG1655 and MC1061), we analyzed the culture supernatants prepared from O157 Sakai containing each CmR-marked Sp derivative with or without MMC treatment. We found that the CmR gene on four Sps (Sp5, Sp6, Sp10, and Sp15) is transferable to K-12 and stably maintained, although the efficiency of transfer was low in all cases except that of Sp5 (Table 2). Among the four Sps, three (Sp5, Sp6, and Sp10) were integrated at the same chromosomal loci in K-12, as in O157 Sakai (Figure 5C). The integration site of Sp10 was already occupied by the Rac prophage in K-12, but the Sp10 derivative was integrated in tandem with Rac using the attR sequence of Rac as the attB site (Figure 5D). In contrast, the Sp15 derivative was not integrated into the yehV locus in K-12, the chromosomal locus where Sp15 is present in O157 Sakai (Figure 5C). This suggests that recombination occurred between Sp15 derivatives and other Sps that allowed the transfer of the CmR gene (by which the stx1 gene was replaced) to K-12 (see the next section).
Among the three Sps that were successfully transferred to K-12, the Sp5 derivative produced infective phage particles in K-12 (Table 2). In contrast, we could not detect the production of infective Sp6 or Sp10 derivatives in K-12. This result suggests that these two Sps may require the support of other Sps, available only in the O157 cell, to produce infective phage particles efficiently.
To analyze the CmR-marked Sp15 (Sp15Δstx1::CmR) transductants, we first performed PCR scanning analysis of the stx1-flanking region of an Sp15Δstx1::CmR-transductant of K12 MG1655 (Figure 6A, see Figure S7 for more details). The results indicated that the transductant contains an Sp15Δstx1::CmR-derived DNA segment covering the stx1 region, but also that some recombination had occurred between the P and stx1 (replaced by the CmR cassette) genes and between the stx1 and nu1 genes in the Sp15Δstx1::CmR genome.
By DNA sequence homology analysis between Sp15 and other Sps, we found that, although many lambdoid Sps contain one or more genomic segments that are highly homologous to the stx1-flanking region of Sp15, only Sp5 contains both segments homologous to the upstream and downstream regions of the stx1 gene (Figure S7). These Sp5 segments are also present in the upstream and downstream regions of the stx2 gene. This suggested that the CmR Sp15 derivative transferred to K-12 may have been generated by recombination between Sp15Δstx1::CmR and Sp5. We therefore analyzed the genome of the CmR Sp15 derivative by PCR using two primer pairs: those specific to the CmR cassette and the Sp5 P gene and those specific to the CmR cassette and the Sp5 Nu1 gene (Figure 6B and Figure S7). The two primer pairs yielded 4.9-kb and 6.4 kb amplicons, respectively, both of which were absent in the donor O157 Sakai derivative containing Sp15Δstx1::CmR. Furthermore, we confirmed that the Sp15Δstx1::CmR transductant contains an Sp5-like phage in the wrbA locus, the integration site of Sp5 (Figure 6C and Figure S7). All these data indicated that the CmR cassette-carrying phage is a chimeric phage that was generated by replacing the stx2 regions of Sp5 with the stx1 region of Sp15Δstx1::CmR.
We analyzed 21 additional Sp15Δstx1::CmR transductants using the same methods. The results indicated that all of the transductants contain chimeric phages of Sp15Δstx1::CmR and Sp5 (data not shown). However, the data from our preliminary sequence analysis of the PCR products covering recombination points suggested that several (at least four) types of chimeric phages had been generated (more details of these chimeric phages will be described elsewhere). It may also be worth noting that this phenomenon was observed only in MMC-treated O157 cells (Table 2).
Finally, we performed an electron microscopic examination of phage particles that were present in the culture supernatants of MMC-treated and untreated O157 Sakai cells (Figure 7). The MMC-treated sample contained numerous phage particles with a short tail attached to a head approximately 56 nm in diameter (Figure 7A). The dominant induction of Sp5 by MMC treatment (Figure 4) suggests that these phage particles originated from Sp5. In fact, K-12 strains lysogenized by CmR-marked Sp5 produced phage particles with the identical morphology. The morphology of Sp5 (Figure 7A) is highly similar to that of the previously reported Stx2 phage of O157 EDL933 [27]. We were unable to detect other phage types in the MMC-treated sample. However, in addition to Sp5, at least two other types of phage particles were detected in the untreated sample. The second type had a head with a hexagonal outline approximately 49 nm in diameter, which was connected by a neck to a contractile and non-flexible tail (the uncontracted sheath is approximately 100 nm long and the contracted one is 55 nm) (Figure 7B). The similarity to the morphology of Mu phage particles indicates that this second type most likely originated from Sp18. The third type had a head with an elongated hexagonal outline (44 wide and 95 nm long) and a 147 nm long flexible tail (Figure 7C). This phage probably derived from some of the lambdoid prophages, but its origin is difficult to pinpoint because lambdoid prophages other than Sp5 (including Sp10 and Sp6) contain very similar morphogenetic genes (Figure 1).
We also examined the culture supernatants of K-12 strains lysogenized with CmR derivatives of Sp6 and Sp10 for the production of phage particles, but no phage particle was detected in the culture supernatants of Sp6 and Sp10 lysogens (data not shown).
The results of our in silico analysis of the potential activities of 18 prophages on the O157 Sakai genome indicate that all but Sp5 contain one or more genetic defects (Figure 1, Figure S1, and Table 1). This suggests that the present-day O157 prophage pool may have low potential activity as mobile genetic elements to spread virulence genes, although their mobility in evolutionary history has played an essential role in the emergence of this highly virulent E. coli lineage. Nevertheless, our systematic experimental evaluation of the Sps revealed that many have unexpectedly high potential activity to function as mobile genetic elements. First, nine Sps could excise themselves from the chromosome and replicate in the O157 cells in response to spontaneous or MMC-mediated induction (Figure 3, Figure 4, and Table 2). They can be divided into three groups according to their induction patterns: (i) spontaneously inducible (Sp6, Sp7, Sp9, Sp10, and Sp18), (ii) spontaneously inducible and further enhanced by MMC-mediated induction (Sp5, Sp13, and Sp15), and (iii) inducible only by the regional replication of other prophages (Sp4 by Sp5 and Sp14 by Sp15). Second, most of these Sps, except Sp6 and Sp14, were packaged into phage particles (Figure 4), although half of them (Sp4, Sp9, Sp13, and Sp15) contain defects in head formation or DNA packaging. Third, we found that the CmR gene cassette on four Sps is transferable to other E. coli strains, although we used only two K-12 derivatives as recipients (Table 2). Three (Sp5, Sp6, and Sp10) were transferred to K-12 and stably lysogenized in the chromosome (Figure 5). This result indicates that Sp6, which is also defective in head and tail formation, can be packaged, although this was not clear from the particulate DNA quantification by qPCR (Figure 4). It is also important that these three phages carry several important virulence determinants, including the stx2 genes and multiple non-LEE effector genes (Figure 1 and Figure S1). Fourth, the CmR gene cassette inserted into the Sp15 genome by replacing the stx1 gene was transferred to K-12, and this transfer was achieved through the generation of a chimera between Sp15 and Sp5 (Figure 6 and Figure S7). In addition, three types of phage with distinct morphologies were detected in the culture supernatant of O157 Sakai (Figure 7). Two derive from Sp5 and Sp18, respectively, but the origin of the third remains undetermined. These results indicate that many apparently defective prophages of O157 Sakai should not be regarded as simple phage remnants, but rather as active genetic elements that can potentially mediate or assist HGT of various virulence determinants encoded on the O157 genome. The results further suggest that various types of inter-prophage interactions occur in the O157 prophage pool, and these interactions induce the biological activities of the defective prophages.
Inter-prophage interactions that most likely occurred in the O157 prophage pool complemented various defects in morphogenetic functions by providing the proteins for phage particle formation. Although the details of this complementation remain to be elucidated, the lambdoid phages on the O157 genome share nearly identical morphogenetic genes [11],[12] in various combinations (Figure 1). They therefore appear capable of supplying virion proteins compatible with those of other lambdoid phages. In fact, we identified one type of phage particle that differs from that of Sp5 but has lambdoid features in the culture supernatant of O157 Sakai (Figure 7C). In some cases, whole virion proteins may be provided by other prophages; this may be the situation with Sp7 and Sp13. Both have severe defects in morphogenetic functions, but are nevertheless packaged. Furthermore, because Sp13 is the only member of the P2-like phage family in the O157 prophage pool, this type of inter-prophage interaction may occur between very different types of bacteriophages.
In the case of Sp7, another type of interaction may complement its defect in replication function. This highly degraded prophage lacks most morphogenesis genes, as well as repressor and antirepressor genes. Furthermore, the replication gene, which resembles the P4 α gene, has been disrupted into three fragments (Figure 1E and Figure S1). Nevertheless, Sp7 is spontaneously inducible and a significant amount of circularized DNA was observed to accumulate in the O157 cells (Figure 4). Thus, the replication of Sp7 may be mediated by the replication proteins of Sp13 (P2-like phage) or Sp2 (P4-like phage), although we cannot exclude the possibility that some (or all) of the fragmented polypeptides of Sp7 may still contain some replication initiation activity.
Replication or amplification of the Sp4 and Sp14 genomes is another type of inter-prophage interaction. Their genomic DNA can be amplified only by the regional replication of Sp5 and Sp15, respectively. Although both lack the genes for excisionase, integrases alone appear capable of mediating their excision from the chromosome. Thus, the two prophage genomes amplified by regional replication are excised into a circularized form (Figure 3 and Figure 4). More interestingly, although Sp4 does not encode an intact packaging enzyme (terminase) by itself (Figure 1 and Figure S1), its amplified genome was found to be packaged (Figure 4). Most likely, this packaging was carried out by the terminase of Sp14, because the putative cos sequence of Sp4, which needs to be digested by terminase for packaging, is nearly identical to that of Sp14 (Figure S8).
Finally, the recombination between Sp15 and Sp5 can also be regarded as an inter-prophage interaction because it occurred in the O157 prophage pool and generated new Stx1-tranducing phages (Figure 5 and Figure S7). This type of inter-prophage interaction can occur between other lambdoid prophages of O157 as well, because they share nearly identical sequences, which can therefore recombine [11],[12]. Similar recombination may also occur between the resident prophages and newly incoming phages. In this way, high levels of excision and replication of defective prophage genomes in O157 cells may provide significant opportunities for such recombination. This may explain why a surprisingly high level of structural variation is observed in the prophage regions among O157 genomes [14],[45], which, in turn, supports the hypothesis that O157 cells function as “phage factories” that produce a wide variety of bacteriophages in nature [12].
In conclusion, many of the prophages of O157 Sakai that contain a wide range of genetic defects show unexpectedly high potential activity as mobile genetic elements, and this mobility is probably achieved through various types of inter-prophage interactions that occur in the O157 prophage pool. Thus, these apparently defective prophages are not simply remnants generated in the course of O157 evolution, but instead should be regarded as genetic elements that are potentially capable of spreading virulence determinants and other genetic traits to other bacterial strains. Similarly to E. coli, many other bacteria contain multiple prophages with genetic defects, and the potential of these sequence elements to function as mobile elements has been largely ignored. Our findings suggest that more attention should be paid to their potential roles in HGT between bacteria and in the evolution of bacterial pathogens.
O157 Sakai (RIMD 050995) was isolated in a large outbreak that occurred in Sakai city, Japan, in 1996 [46], and the complete genome sequence has been determined [11]. Cells were grown overnight to stationary phase at 37°C in Luria-Bertani (LB) medium. For prophage induction with MMC, cells were grown to early log phase (OD600 0.2–0.4), and MMC was added to the culture to a final concentration of 1 µg/ml. At 1-hr intervals, we isolated aliquots of the culture and collected the cells by centrifugation at 4°C. Total cellular DNA was isolated from the cells using the Genomic-tip 100/G and the Genomic DNA buffer set (Qiagen, CA, USA) according to the manufacturer's instructions.
Phage particles were isolated from the culture supernatants 10 hr after the addition of MMC. The culture was first treated with chloroform, and bacterial cell debris was removed by centrifugation. The supernatant was filtered through 0.22-µm pore-size filters (Millipore Corp., MA, USA) and incubated with 200 U/ml DNase I (Invitrogen, CA, USA) at 37°C for 1 hour. After incubation, 0.25 volumes of a solution containing 20% polyethylene glycol 8000 (PEG) and 10% NaCl was added to the sample. The mixture was kept at 4°C overnight and then centrifuged at 12,000×g for 1 hour to precipitate phage particles. The phage particles were suspended in SM buffer (0.58% NaCl, 0.2% MgSO4⋅7H2O, 1 M Tris-Cl (pH 7.5), 0.01% gelatin) and incubated with DNase I (final concentration, 1000 U/ml) and RNase A (50 µg/ml) (Stratagene, CA, USA) at 37°C for 1 hour. After DNase and RNase treatment, the sample was treated with proteinase K (100 µg/ml; Wako, Osaka, Japan) at 50°C for 1 hour, and phage DNA was isolated using the Genomic-tip 20/G. Total cellar DNA and phage DNA from untreated samples were prepared by the same protocol, except that no MMC was added to the culture.
Comparative analysis of O157 Sakai prophage genomes was performed using the “in silico MolecularCloning(R) (IMC)” software (Genomics edition, version 1.4.71, In Silico Biology, Inc., Kanagawa, Japan). Homology searches were performed using BLAST2 [47], and functional motifs were searched using InterProScan (http://www.ebi.ac.uk/Tools/InterProScan/). Multiple protein sequence alignment was carried out using CLUSTALW [48] and analyzed by the multiple alignment editor Jalview (http://www.ebi.ac.uk/clustalw/). A neighbor-joining tree for replication elongation proteins was generated using MEGA3 [49].
An overnight culture in LB medium was diluted to an OD600 of 0.2 and grown for 1 hour. At this point (0 min), MMC was added to the culture at a concentration of 1 µg/ml. Samples were collected at 0-min, 45-min and 90-min intervals from the cultures treated with MMC or those left untreated. The RNAprotect Bacteria Reagent (QIAGEN, Valencia, CA) was immediately added to the samples, and total RNA was isolated using the RNeasy Plus Mini kit (QIAGEN, Valencia, CA) according to the manufacturer's instructions. RNA quality was assessed by spectrophotometry using the NanoDrop instrument (NanoDrop Technologies, Inc., USA) and by agarose gel electrophoresis.
Probes (60 mer) were designed for the 5,447 protein-coding genes of the O157 Sakai genome. The O157 Sakai genome contains many multi-copy genes that are derived from IS elements and lambdoid prophages sharing nearly identical sequences. Thus, to avoid effects due to cross hybridization, all data for the probes that showed >80% DNA sequence identity to any other genomic regions of the O157 Sakai genome were removed from the data set. Finally, we used 4,507 probes representing 4,507 genes. Among the 4,507 probes, 452 were for the genes on 18 prophage regions. Arrays were produced by Agilent Technologies (Palo Alto, CA, USA) by the in situ oligonucleotide synthesis method.
For DNA microarray analyses, test and reference DNA (250 ng) were chemically labeled with ULS-Cy3 and ULS-Cy5, respectively, using the Agilent Oligo CGH Microarray Kit (Agilent Technologies). The fluorescently labeled DNA was purified by the Agilent KREApure column. The Cy5-labeled and Cy3-labeled DNA were mixed and used for hybridization. For RNA analysis, total RNA (10 µg) was reverse transcribed and labeled with amino-allyl dUTP using MMLV-RT (Agilent Technologies) and random hexamers (Invitrogen). The cDNA from test and control samples was labeled with Cy3 and Cy5 dye, respectively (Monofunctional NHS-ester Dye, Amersham). The Cy3-labeled and Cy5-labeled cDNAs were purified, combined, and used for hybridization. The arrays were scanned using an Agilent scanner (Agilent Technologies), and data extraction, filtering and normalization were conducted using Feature Extraction software (Agilent Technologies) according to the manufacturer's instructions. Each sample was examined twice using the labeled DNAs and cDNAs independently prepared for each hybridization. Data analysis and visualization were done using Microsoft Excel and MultiExperiment Viewer (The Institute for Genome Research) [50].
PCR was carried out using an Ex-taq PCR amplification kit (Takara Bio, Kyoto, Japan). The PCR cycling program consisted of 29 cycles of 45 sec at 95°C, 45 sec at 60°C, and 1 min at 72°C, with an additional step of 2 min at 72°C. Sequencing of the PCR products was carried out using an ABI PRISM 3100 automated sequencer (PE Biosystems, CA, USA). Primers designed for PCR amplification were used as sequencing primers. Sequencher™ software (version 4.2.2, Gene Codes Corporation, MI, USA) was used for sequence data analyses. All primers used are listed in Table S1.
TaqMan probes and PCR primers for real-time qPCR (Table S1) were designed using Primer Express software (Primer Express™, PE BioSystems) according to the manufacturer's instructions. All analyses were performed using the ABI PRISM 7000 Sequence Detection System (PE BioSystems). To analyze the intracellular prophage DNA, 10 ng of total cellular DNA was used as template DNA in a 50 µl reaction volume. To analyze the phage particle DNA, phage DNA isolated from 1 ml (MMC-treated samples) or 50 ml (untreated samples) of culture supernatants was used as the template. Primers and TaqMan probes were used at a concentration of 400 nM and 250 nM, respectively. FAM (5′) and TAMRA (3′) were used as reporter and quencher dyes for TaqMan probes, respectively. The PCR cycling program consisted of 45 cycles of 15 sec at 95°C and 1 min at 60°C. For quantification of DNA, the standard curve method was employed. Standard curves were constructed over the range from 103 copies/µl to 107 copies/µl for each amplicon. Quantities of two chromosomal backbone regions (CB1 and CB2 in Figure 3 and Figure 4) were monitored as controls in qPCR analyses. The concentrations of chromosomal backbone DNA measured in the total cellular DNA preparations were comparable with those estimated from the numbers of cells used for DNA preparation (Figure 3 and Figure 4), indicating the validity of the real-time qPCR assays.
FIGE analysis of packaged phage DNA was performed using phage particles collected from culture supernatants by PEG/NaCl precipitation. Precipitated phage particles were embedded in plugs of 1% Certified Low Melt Agarose (Bio-Rad Laboratories, Inc., CA, USA) and treated at 37°C for 2 hr with DNase I (1000 U/ml) and RNase A (50 µg/ml) in DNase I buffer (10 mM Tris (pH 7.5), 2.5 mM MgCl2, and 0.5 mM CaCl2), followed by overnight incubation with proteinase K (100 µg/ml) at 50°C in TE containing 1% sodium dodecyl sulfate. After the plugs were washed three times with TE buffer at 15-minute intervals, the plugs were sliced into appropriate sizes and subjected to FIGE analysis. FIGE was performed using a CHEF MAPPER (Bio-Rad Laboratories) with a 1% agarose gel, and initial and final switch times of 0.11 and 0.92 sec, respectively. Total run time was 20.3 h with 9.0 V/cm (forward) and 6.0 V/cm (reverse) voltage at a constant temperature of 14°C. The gel was stained with ethidium bromide to visualize DNA bands.
The CmR cassette was inserted into eight Sps (Figure 5A) by replacing moron genes on each prophage genome using the method described by Datsenko and Wanner [51]. Primer sequences utilized for gene replacement and confirmation of replacement are listed in Table S2. In Sp5 (Stx2 phage) and Sp15 (Stx1 phage), the entire stx genes were replaced by the CmR cassette. In other Sps, genes for T3SS effectors or other morons were selected as targets for gene replacement. Therefore, the O157 Sakai derivatives generated in this study are predicted to have reduced potential virulence.
CmR-O157 Sakai derivatives (Table S3) were grown at 37°C overnight in LB containing 40 µg/ml of Cm. The cells were subcultured to an OD600 of 0.2–0.4 in LB without antibiotic and then cultivated at 37°C for 10 hrs with vigorous shaking in the presence or absence of MMC (10 µg/ml). Phage particles in the culture supernatants were recovered by PEG/NaCl precipitation and suspended in SM buffer. The precipitated phage preparations were gently treated with 0.1 volumes of chloroform. After brief centrifugation, aqueous phases of each sample were collected and incubated at 37°C for 15 min to remove residual chloroform. Phage solution prepared from a 20-ml culture supernatant of each O157 Sakai derivative was incubated with 108 recipient cells suspended in 100 µl of the SM buffer at 28°C for 1 hour. Finally, the recipient cells were plated on LB agar plates containing Cm (50 µg/ml) and incubated overnight at 37°C. Randomly selected colonies were checked by the agglutination test using anti-O157 serum to ensure that all were derived from K-12 strains; colonies were subsequently used in further analyses. The CmR transductants obtained were verified for the lysogeny of CmR-marked prophages using PCR (Figure 5). Primers used are listed in Table S4.
To analyze the chimeric phages of Sp15Δstx1::CmR and Sp5, PCR scanning analysis of the stx1-flanking region was performed using eight primer pairs (Figure 6A, Figure S7, and Table S5). To confirm the chimeric structure of the recombinant phage, we performed two types of PCR analyses, one using the forward (P_F) and reverse (4R) primers specific to the P gene of Sp5 and the CmR gene cassette, respectively, and another using the forward (6F) and reverse (T_R) primers specific to the CmR gene cassette and the Sp5 nu1 gene, respectively. These primer sequences are also listed in Table S6. Integration of an Sp5-like phage into the wrbA locus in Sp15Δstx1::CmR-transductants was confirmed by PCR using two primer pairs (Table S4) to amplify the left (primers LbF and LbR) and right (RbF and RbR) attachment sites of Sp5 (Figure 6C and Figure S7).
Phage particles were collected by PEG/NaCl precipitation from the culture supernatants of MMC-treated or untreated O157 Sakai and suspended in SM buffer. A 10-µl drop of the suspension was placed on copper grids with carbon-coated Formvar films and negatively contrasted with 2% uranyl acetate dihydrate. Samples were examined using a transmission electron microscope (1200EX, JEOL, Tokyo) operated at 80 kV. Average dimensions were measured with the image processing and analysis software ImageJ (http://rsb.info.nih.gov/ij/).
All sequence information for prophages and genes of the O157 Sakai genome is available at the DDBJ/EMBL/NCBI database (accession no. BA000007)
|
10.1371/journal.pcbi.1002457 | Automatic Filtering and Substantiation of Drug Safety Signals | Drug safety issues pose serious health threats to the population and constitute a major cause of mortality worldwide. Due to the prominent implications to both public health and the pharmaceutical industry, it is of great importance to unravel the molecular mechanisms by which an adverse drug reaction can be potentially elicited. These mechanisms can be investigated by placing the pharmaco-epidemiologically detected adverse drug reaction in an information-rich context and by exploiting all currently available biomedical knowledge to substantiate it. We present a computational framework for the biological annotation of potential adverse drug reactions. First, the proposed framework investigates previous evidences on the drug-event association in the context of biomedical literature (signal filtering). Then, it seeks to provide a biological explanation (signal substantiation) by exploring mechanistic connections that might explain why a drug produces a specific adverse reaction. The mechanistic connections include the activity of the drug, related compounds and drug metabolites on protein targets, the association of protein targets to clinical events, and the annotation of proteins (both protein targets and proteins associated with clinical events) to biological pathways. Hence, the workflows for signal filtering and substantiation integrate modules for literature and database mining, in silico drug-target profiling, and analyses based on gene-disease networks and biological pathways. Application examples of these workflows carried out on selected cases of drug safety signals are discussed. The methodology and workflows presented offer a novel approach to explore the molecular mechanisms underlying adverse drug reactions.
| Adverse drug reactions (ADRs) constitute a major cause of morbidity and mortality worldwide. Due to the relevance of ADRs for both public health and pharmaceutical industry, it is important to develop efficient ways to monitor ADRs in the population. In addition, it is also essential to comprehend why a drug produces an adverse effect. To unravel the molecular mechanisms of ADRs, it is necessary to consider the ADR in the context of current biomedical knowledge that might explain it. Nowadays there are plenty of information sources that can be exploited in order to accomplish this goal. Nevertheless, the fragmentation of information and, more importantly, the diverse knowledge domains that need to be traversed, pose challenges to the task of exploring the molecular mechanisms of ADRs. We present a novel computational framework to aid in the collection and exploration of evidences that support the causal inference of ADRs detected by mining clinical records. This framework was implemented as publicly available tools integrating state-of-the-art bioinformatics methods for the analysis of drugs, targets, biological processes and clinical events. The availability of such tools for in silico experiments will facilitate research on the mechanisms that underlie ADR, contributing to the development of safer drugs.
| Drug safety issues can arise during pre-clinical screening, clinical trials and, more importantly, after the drug is marketed and tested for the first time on the population [1]. Although relatively rare once a drug is marketed, drug safety issues constitute a major cause of morbidity and mortality worldwide.
In 1998, Lazarou et al estimated that yearly about 2 million patients in the US are affected by a serious adverse drug reactions (ADRs) resulting in approximately 100 000 fatalities, ranking ADRs between the fourth and sixth cause of death in the US, not far behind cancer and heart diseases [2]. Similar figures were estimated more recently for other western countries [3], [4], [5]. Serious ADRs resulting from the treatment with thalidomide prompted modern drug legislation more than 40 years ago [6]. Over the past 10 years, 19 broadly used marketed drugs were withdrawn after presenting unexpected side effects [1], [3]. The current and future challenges of drug development and drug utilization, and a number of recent high-impact drug safety issues (e.g. rofecoxib) highlight the need of an improvement of safety monitoring systems [5]. In this regard, initiatives such as the EC-funded EU-ADR project seek to develop methodologies to improve the way drug safety signals are detected and analyzed [7], [8].
Due to the important implications of an ADR in both public health and the pharmaceutical industry, unraveling the molecular mechanisms by which the ADR is elicited is of great relevance. Understanding the molecular mechanisms of ADRs can be achieved by placing the drug adverse reaction in the context of current biomedical knowledge that might explain it. Due to the huge amounts of data generated by the “omics” experiments, and the ever-increasing volume of data and knowledge stored in databases related with ADRs, the application of bioinformatics analysis tools is essential in order to study and analyze the molecular and biological basis of ADRs.
Although the factors that determine the susceptibility to ADRs are not completely well understood, accumulating evidence over the years indicate an important role of genetic factors [9]. ADRs can be mechanistically related to drug metabolism phenomena, leading for instance to an unusual drug accumulation in the body [9]. They can be associated with inter-individual genetic variants, most notably single nucleotide polymorphisms (SNPs), in genes encoding drug metabolizing enzymes and drug target genes [9]. One of the first ADRs explained by a genetic factor was the inherited deficiency of the enzyme glucose-6-phosphate dehydrogenase causing severe anemia in patients treated with the antimalarial drug primaquine [10]. Alternatively, an ADR can be caused by the interaction of the drug with a target different from the originally intended target (also known as off-targets) [11]. A well-known example of an off-target ADR is provided by aspirin, whose anti-inflammatory effect, exerted by inhibition of prostaglandin production by COX-2, comes at the expense of irritation of the stomach mucosa by its unintended inhibition of COX-1 [12], [13]. Furthermore, in addition to mechanisms related to off-target pharmacology, it is becoming evident that ADRs may often be caused by the combined action of multiple genes [9]. The anticoagulant warfarin, which shows a varying degree of anticoagulant effects, is often associated with hemorrhages, and leads the list of drugs with serious ADR in the US and Europe [9]. A 50% of the variable effects of warfarin are explained by polymorphisms in the genes CYP2C9 and VKORC1 [14], [15]. A recent study furthermore identified a third gene, CYP4F2 explaining about 1.5% of dose variance [16]. However, the genes accounting for the remaining variability in the response to warfarin are still unknown.
Other cases of ADRs may arise as a consequence of drug-drug interactions, or the interplay between the effect of the drug and environmental factors [9], [15]. Indeed, the interaction between genotype and environment observed in several aspects of health and disease also extend to drug response and safety. For example, alcohol consumption and smoking are both associated with changes in the expression of the metabolic enzyme CYP2E1, therefore affecting the pharmacokinetics of certain drugs [17].
From the above paragraphs, it is clear that the study of the molecular and biological mechanisms underlying ADRs requires achieving a synthesis of information across multiple disciplines. In particular, it requires the integration of information from a variety of knowledge domains, ranging from the chemical to the biological up to the clinical. Different resources cover information about these different knowledge domains, and many of them are freely available on the web, such as biological and chemical databases and the biomedical literature. On the other side, new data is produced continuously, and the list of resources and published papers that a researcher interested in ADRs needs to cope with is turning more into a problem than into a solution. It has been recognized that the adequate management of knowledge is becoming a key factor for biomedical research, especially in the areas that require traversing different disciplines and/or the integration of diverse and heterogeneous pieces of information [18]. A key aspect is the integration of heterogeneous data types, and several authors have discussed the challenges of data integration in the life sciences [19], [20], which are rooted in the inherent complexity of the biological domain, its high degree of fragmentation, the data deluge problem, and the widespread ambiguity in the naming of entities [21]. In addition to the complexity of extracting, storing and integrating heterogeneous data from multiple domains one needs to consider the lack of completeness of the data available [22], an aspect that has a direct impact on the scope and conclusions of any analysis performed on the integrated data.
On the other hand, approaching current biomedical research questions by computational analysis requires a combination of different methods. An attractive approach that emerged in the last years is the combination of different bioinformatics analysis modules by means of pipelines or workflows [23]. This technology allows the integration of a variety of computational techniques into a processing pipeline in which the input and outputs are standardized. This kind of integration has been greatly facilitated by the use of public APIs and web services allowing programmatic access to data repositories and analysis tools. The open source software Taverna is one of such approaches that allow integration of different analysis modules, shared as web services, into a scientific workflow to perform in silico experiments [24]. Similar approaches are also used for the processing of free-text documents (http://uima.apache.org/) or for combining data mining methods (http://www.knime.org/).
In this article we present a general framework developed in the context of the EU-ADR project for a systematic analysis of adverse drug reactions. The entry point of the system is a potential drug safety signal, which is composed of the drug and its associated adverse reaction. In the process of signal filtering, we search for previous reports of the potential signal in specialized databases and in the biomedical literature. In the process of signal substantiation, we seek to provide a plausible biological explanation to the potential signal. This framework was implemented by means of software modules accessible through web services and integrated into workflows ready to be used for automatic filtering and substantiation of drug-event associations. Finally, we present a detailed analysis of antipsychotic drugs and their association with the prolongation of the QT interval, as well as a large scale analysis of drug-side effect pairs from SIDER [25] emphasizing the usefulness of our signal filtering and substantiation workflows.
The here presented framework for the filtering and substantiation of drug safety signals consists of placing the potential signal in the context of current knowledge of biological mechanisms that might explain it. Essentially, we are searching for evidence that supports causal inference of the signal, i.e. feasible paths that connect the drug with the clinical event of the adverse reaction. The signal filtering analysis looks for evidence reporting the drug-event association in the biomedical literature and biomedical databases. The signal substantiation process considers two scenarios able to provide a causal inference of the signal (see Figure 1). First, we look for connections between the drug and the event through their associated protein profiles. Here, a connection is established if there are proteins in common between the drug-target and the event-protein profile (Figure 1A). Many ADRs are caused by altered drug metabolism for which genetic variants in metabolizing enzymes are often responsible. Consequently, we also consider drug metabolism phenomena as an underlying mechanism of the observed ADR by assessing if the drug metabolites are targeting proteins that are known to be associated with the clinical event. Second, the association between the drug and the clinical event can involve proteins that are not directly associated with the drug and the clinical event, but indirectly in the context of biological networks. The final consequence of the drug action is the observed clinical event. Thus, the proteins in the drug-target profile and event-protein profile are mapped onto biological pathways to evaluate if the drug and the event can be connected through biological pathways (Figure 1B).
Our approaches for signal filtering and signal substantiation were implemented using dedicated bioinformatics methods that are accessed through web services and integrated into processing pipelines by means of Taverna workflows. The substantiation workflow results can be visualized and analyzed by means of other bioinformatics tools such as Cytoscape [26], a software for network visualization and analysis. For the signal filtering process, we have implemented two Taverna workflows (ADR-FM and ADR-FD) that access data mined from databases such as DrugBank [27], DailyMed (http://dailymed.nlm.nih.gov/) and Medline®. A third Taverna workflow, (ADR-S), performs the signal substantiation process and was implemented by combining in silico target profiling, text mining and pathway analysis, among other bioinformatics approaches. More details about the implementation of web services and workflows can be found in the Methods section.
In the following section we describe the results of the analysis of potential drug safety signals as a proof of concept of the here proposed framework and tools.
In the 1990s, the occurrence of several cases of serious, life-threatening ventricular arrhythmias and sudden cardiac deaths, secondary to the use of non-cardiac drugs raised concerns with regulators [28]. In 1998, several drugs received a black-box warning in the US due to concerns regarding prolongation of the QT interval. Nowadays, it is known that many seemingly unrelated drugs can cause the prolongation of QT interval and Torsade de Pointes, which eventually may lead to fatal arrhythmias. For instance, cisapride, a drug for gastrointestinal protection, was withdrawn from the market in 2000 due to increased risk for QT prolongation. The first report of sudden cardiac death with an antipsychotic drug appeared in 1963 [29]. Since then, several studies found an increased risk for ventricular arrhythmias, cardiac arrest and sudden death associated with the use of antipsychotics [30], which can partly be explained by the prolongation of QT intervals observed with several antipsychotic drugs. It has been suggested that the mechanisms by which antipsychotics can cause prolongation of QT interval involve the potassium channel encoded by the KCNH2 gene that regulates myocyte action potential [31], [32]. Drugs blocking this potassium channel can slow down repolarization, which in turn may lead to the prolongation of the QT interval, eventually resulting in sudden cardiac death. We selected six antipsychotic drugs according to their risk of producing cardiac arrhythmias from [33] and from the QTdrugs database (http://www.qtdrugs.org) (Tables 1 and 2) and analyzed their association with the prolongation of the QT interval as defined in the EU-ADR project (referred to as QTPROL) using our signal filtering and substantiation workflows. The results of the filtering analysis (shown in Table 1) indicate that all drug-event associations are discussed in the literature or recorded in specialized databases, with the only exception of DrugBank that does not contain any information on the association of the selected drugs with QTPROL. When comparing both Medline-based filtering workflows, ADR-FM/MeSH and ADR-FD/Medline, the latter appears to be more sensitive as the number of abstracts found is generally higher (compare columns ADR-FM/MeSH and ADR-FD/Medline in Table 1). This difference might be explained by the different methods used by the two approaches. The MeSH®-based approach uses the MeSH terms assigned to each citation and the ADR-FD approach uses Natural Language Processing on title and abstracts to identify drug-event associations. Both Medline-based approaches can be compared with a PubMed query (“(QT or QTc) prolongation <one of the six antipsychotic drugs>”), which resulted in 2–3 times more abstracts being returned than by ADR-FD/Medline. This does not come as a surprise since PubMed searches for keyword co-occurrences at the abstract level. The workflows are more specific since they search at the sentence level (ADR-FD/Medline) or use additional information provided by the MeSH subheadings and the use of the pharmacological action (ADR-FM/MeSH). It should be noted that Medline is only one source of information to filter known signals; DrugBank and DailyMed are other, potentially complementary, sources. In the case of pimozide, no results are obtained from DailyMed®, since QT prolongation is not mentioned in the adverse reactions section but in the contraindications and warnings sections.
We furthermore explored the mechanisms underlying the association between QTPROL and the selected antipsychotics using the substantiation workflow. The results are summarized in Table 2 (see Table 3 for a quick reference guide to gene and protein names discussed throughout the example) and Figure 2, which shows a detail of the Cytoscape graph representing the drug-protein-event network resulting from analyzing haloperidol and its association with QTPROL. For all the antipsychotic drugs, with the exception of sulpiride, connections are established through proteins associated with both, drug and event. All the connections between the drug and the event include the protein HERG encoded by the KCNH2 gene. All of the found connections are statistically significant except for ziprasidone (see Table 3). The high-risk antipsychotics haloperidol, ziprasidone and pimozide are potent potassium channel blockers (IC50 or Ki in the 0.1 µM range, Table 2). In the case of ziprasidone, it is worth to mention that one of the metabolites of the drug is predicted to bind to the protein HERG. Contrasting, olanzapine shows a lower activity on the protein HERG, while sulpiride has no activity on this protein. In addition to HERG, for the high-risk antipsychotics pimozide and haloperidol the drug and the event can be connected through the proteins encoded by the genes KCNH1 and CANCNA1C. In the case of KCNH1, which encodes the protein hEAG1, the ADR-S workflow provides evidence indicating that mutations in an animal model showed an association with prolonged QT interval and cardiac arrhythmia [34]. The mutations in the CACNA1C gene, which encodes the depolarizing long-lasting calcium current channel, are associated with Timothy syndrome, characterized by severe prolongation of the QT interval.
Interestingly, our analysis also indicates that the antipsychotics in our study have an important activity on adrenergic receptors (Figure 2 B).
Moreover, haloperidol shows activity on the drug transporter encoded by the gene ABCB1 (Ki 0.2 µM, Figure 1B). Similar activities are found for pimozide, whereas ziprasidone, olanzapine, sulpiride and quetiapine do not show activity on the transporter.
Regarding the substantiation through pathways, for haloperidol and pimozide we found several Reactome pathways (Integration of energy metabolism, Axon guidance, Synaptic transmission, Signaling by GPCRs and Diabetes pathways), which connect the drug and the event, and where the involved proteins are expressed in cardiac tissues. It is likely that the effect of a drug on its target proteins will affect proteins in their direct neighborhood in the biological pathway. Hence, we computed the average shortest path length between pairs of drug and event associated proteins in the Reactome pathways and compared them to the average shortest path length between randomly selected drug and event proteins. Interestingly, for all five antipsychotic drugs, the drug and event proteins are in close proximity in the Reactome pathways with average shortest path lengths between 2 and 3, which are significantly shorter than the average shortest path length of 5 of randomly selected drug and event proteins (p-value< = 0.05).
In summary, the ADR-S workflow provides different hypotheses explaining the antipsychotics-induced QTPROL, including the direct action of the drug on proteins associated with the clinical event (e.g. HERG), the cross-talk between different biological processes (adrenergic signaling and cardiac action potential), and the differential distribution of drugs among tissues (due to inhibition of transporters exerted by the drug). Moreover, it also highlights several interesting evidences that might explain the differences between low and high-risk antipsychotics.
In addition to the example case presented above, the ADR-S workflow was evaluated on a large-scale data set. The SIDER database was used to extract drug-event pairs (see Methods). Here, an event refers to a known side effect of a drug compiled from package inserts of the drugs from several public sources [25]. For a total of 28251 drug-event pairs, 6108 (4265 with p-value< = 0.01) pairs can be directly linked through at least one protein connecting the drug with the side effect. Interestingly, 2692 (44%) of the 60108 drug-event pairs are connected by means of the drug metabolites. Moreover, the substantiation through pathways module finds connections between 21526 pairs (10789 with p-value< = 0.01). This quantitative analysis should be followed by a thorough qualitative study on selected drug-event pairs of interest in order to explore the found connections and derive mechanistic hypothesis. Hence, we make the results of the analysis available as Supplementary Material (Dataset S1 and S2).
Recent studies highlight the use of disparate data sets in the study of ADRs, enabled by bioinformatics methodologies. Combining the study of protein–drug interactions on a structural proteome-wide scale with protein functional site similarity search, small molecule screening, and protein–ligand binding affinity profile analysis, Xie and colleagues [35] have elucidated a possible molecular mechanism for the previously observed, but molecularly uncharacterized, side effect of selective estrogen receptor modulators (SERMs). In another study, the side effect information from prescription drug labels was exploited to identify novel molecular activities of existing drugs [25]. The Unified Medical Language System (UMLS) Metathesaurus® [36] was used as a vocabulary for the side effects, and a weighting scheme to account for the rareness and interdependence of side effects was developed. Since similarity in side effects correlated with shared targets between drugs, side effect similarity was used to predict novel targets between any two “unexpected” drug pair [25]. In another study, Berger and colleagues used a computational systems biology approach to analyze drug-induced long QT syndrome, and showed that the analysis of a human protein interaction network associated with congenital long QT syndrome can be used to predict new gene variants for long QT syndrome, to explain the complexity of the adverse drug reaction, and to predict the susceptibility of new drugs to cause long QT syndrome [32].
All these examples illustrate how computational approaches are paving the way toward elucidating the molecular mechanisms of ADRs. The here presented framework follows this direction, by traversing and integrating information from the chemical domain, through genes and proteins, molecular and cellular networks, and finally to the clinical domain. The filtering workflows interrogate specialized databases and literature repositories in order to determine the novelty of a drug-event association. On the other hand, the substantiation framework seeks to find hypotheses that might explain drug-induced clinical events by looking for evidences supporting causative connections between the drug, its targets, and their direct or indirect (through biological pathways) association to the clinical event. The signal substantiation process can be framed as a closed knowledge discovery process, analogous to the Swanson model based on hidden literature relationships [37], which we extend by considering not only relationships found in the literature, but also relationships discovered by mining other data sources or found by applying different bioinformatics methods (vide infra). For a drug-event association, we collect information about the drug-targets by querying publicly available databases and by applying in silico drug-target profiling methods [38]. In parallel, we retrieve information about the genes and proteins associated with the clinical event from a database covering knowledge about the genetic basis of diseases [39]. Then, we combine these two pieces of information under the following assumption: if the disease phenotype elicited by the drug is similar to the phenotype observed in a genetic disease, then the drug acts on the same molecular processes that are altered in the disease. This can be regarded as phenocopy, a term originally coined by Goldschmidt in 1935 [40] to describe an individual whose phenotype, under a particular environmental condition, is identical to the one of another individual whose phenotype is determined by the genotype. In other words, in the phenocopy the environmental condition mimics the phenotype produced by a gene. In the case of ADRs, the environmental condition is represented by the exposure to the drug, whose effect mimics the phenotype (disease) produced by a gene in an individual. In this way, we can capitalize on all the knowledge about the genetic basis of diseases to explore mechanisms underlying ADRs.
We illustrate our approach by analyzing a clinically relevant drug safety signal: prolongation of the QT interval (QTPROL) leading to cardiac arrhythmias produced by a set of antipsychotic drugs. The results of the filtering workflows show that the association of QTPROL with the antipsychotic drugs has been extensively discussed in the literature and is documented in specialized databases. On the other hand, the substantiation workflow provides different hypotheses explaining the antipsychotics-induced QTPROL. First, we were able to confirm the widely accepted mechanism proposed for drug-induced QTPROL, in which the drug blocks the potassium channel HERG (encoded by the KCNH2 gene) and this blockade leads to a prolongation of the QT interval [41], [42]. The known association of congenital long QT syndrome being associated with mutations in the KCNH2 gene furthermore supports this concept [38], [39]. Interestingly, our analysis reveals that high-risk antipsychotics show higher activities on the potassium channel than low-risk antipsychotics (see Table 2), suggesting that the strength of binding might explain the different risks of observing the side effect for different antipsychotics. For all except one antipsychotic (ziprasidone), the associations between the drugs and QTPROL are statistically significant (p-value< = 0.01). We want to point out, that even for ziprasidone with a higher p-value, the evidences provided by the workflow give enough confidence to establish the hypothesis of the blockage of HERG being related with QTPROL. We believe that each drug-event pair and the evidences provided by the workflows have to be studied carefully in order to generate hypotheses valid to be tested. We furthermore find a connection of high-risk antipsychotics and QTPROL through other proteins different from HERG, suggesting that the prolongation of the QT interval might result from the effect of the drugs on other channel proteins regulating the action potential. In addition to the direct blockade of channels creating ion currents involved in the action potential, other factors can be considered for the mechanism of antipsychotics-induced QTPROL. Adrenergic activation due to stress can precipitate cardiac arrhythmias [35]; in fact, the main treatment for patients with congenital long QT syndrome is beta-adrenergic blocking [41]. Alpha and beta-receptors agonists produce an inhibition of the potassium channel leading to the prolongation of QT [34]. Interestingly, our results indicate that the antipsychotics in our study have an important activity on adrenergic receptors. Haloperidol has been reported to act as partial agonist in cerebral alpha-adrenergic receptors [43]. Hence, our results suggest that the modulation of adrenergic signaling by haloperidol might be an additional factor resulting in the inhibition of the potassium repolarizing current. Thus, in the case of haloperidol, direct inhibition by the drug combined with an indirect mechanism involving the activation of beta adrenergic signaling might lead to HERG blockade. These findings are in line with evidences supporting the notion that ADRs may often be caused by the combined action of multiple genes [9].
We furthermore found that activities of haloperidol and pimozide on the drug transporter encoded by the gene ABCB1 (Ki 0.2 µM, Figure 1B), while ziprasidone, olanzapine, sulpiride and quetiapine do not show activity on this transporter. Titier and colleagues studied the myocardium to plasma concentration ratio of several antipsychotic drugs, reporting ratios of 2.7 for olanzapine and 6.4 for haloperidol [43]. Therefore, the different distributions of the antipsychotics between plasma and the heart could be another factor influencing the varying risk of different antipsychotic drugs to induce QTPROL.
Regarding the analysis through biological pathways, our workflow does not provide novel hypotheses that might explain drug-induced QTPROL in addition to the above presented hypotheses. Nevertheless, it is interesting that the drug target proteins and event-associated proteins are closely located in the Reactome pathways. All in all, a detailed analysis of the generated paths might add valuable information about the mechanism underlying the drug adverse reaction. Ultimately, the usefulness of the pathway module strongly depends on the drug-safety signal of interest. For example, the cholesterol-lowering drug cerivastatin was withdrawn from the market in 2001 due to its fatal risk to induce rhabdomyolysis leading to kidney failure [44]. While the ADR-S workflow connects cerivastatin and rhabdomyolisis through proteins and pathways, it only finds a meaningful connection between the drug and acute renal failure through the pathway module. Hence, in this example the pathway module adds valuable information to the analysis. We also want to mention some limitations of the pathways module. The publicly available information on pathways is not complete, and the level of detail differs between the pathways. Moreover, the Reactome pathways used are at a very high level in the Reactome hierarchy and can be very general; hence the substantiation results need to be carefully analyzed in order to determine if the connection found between the drug and the event represents a plausible explanation of the ADR.
In summary, using antipsychotics and their risk to induce QTPROL, we showed that the filtering workflows are able to extract relevant information from the literature and dedicated databases. We also showed that the substantiation workflow provides different hypotheses explaining the antipsychotics-induced QTPROL. These hypotheses include the direct action of the drug on proteins associated with the clinical event (e.g. HERG), the cross-talk between different biological processes (adrenergic signaling and cardiac action potential), and the differential distribution of drugs among tissues (due to inhibition of transporters exerted by the drug). Moreover, the analysis also highlights several interesting evidences that might explain the differences between low and high-risk antipsychotics. In addition, we provide the results of a large-scale analysis of drug-side effect pairs from SIDER and show that about 22% of the known side effects of drugs might involve direct effects of drugs on proteins being associated with the events. This relatively small number is not surprising because not all drug side effects can be attributed to the direct action of the drug onto its targets, such as on-target and off-target pharmacological effects. Other mechanisms of drug toxicity have been discussed. For example, metabolites can react with nucleophiles including DNA, which can trigger regulatory processes leading to inflammation, apoptosis and necrosis [45]. Moreover, the workflow uses public data sources on drug-target and event-protein associations, which are not complete. Interestingly, almost half (44%) of the direct connections through proteins involve metabolites of the drugs. This finding is in good agreement with current opinion on the relevance of drug metabolism for drug adverse reactions [9]. The pathway module connects many more drug-side effect pairs. Although, the results of our workflow for each drug-side effect pair have to be carefully analyzed in detail, this finding suggests that the indirect connection of drug and event in the context of biological networks plays an important role. We want to stress that the substantiation workflow provides a variety of evidences, such as the binding strength of the drug to its targets, as well as the provided literature sources supporting the associations of proteins to the events. All pieces of evidence need to be carefully considered to generate hypotheses of mechanisms that are valid to be further tested.
Both filtering and substantiation workflows are available to the community and allow a systematic and automatic analysis of drug safety signals detected by mining clinical records, providing a user-friendly framework for the analysis of drug-event combinations. We believe that with the availability of such tools for in silico experimentation, research on the mechanism that underlies drug-induced adverse reactions will be facilitated, which will have great impact in the development of safer drugs.
The signal filtering and substantiation framework has been implemented by means of software modules that perform specific tasks of the processes. To allow access and integration of the modules in high-level analysis pipelines, the modules were implemented as web services and combined into data processing workflows to achieve the aforementioned signal filtering and signal substantiation. To standardize data exchanges between the different web services, we have developed two complementary schemas using XSD to define a common XML interoperability structure. The first one describes general data types (http://bioinformatics.ua.pt/euadr/common_types.xsd) and the second one defines the specific types needed for signal filtering and substantiation in the context of the EU-ADR project (http://bioinformatics.ua.pt/euadr/euadr_types.xsd). Both schemas allow a smooth integration of the different modules in Taverna workflows, by enabling content and structure validation for the workflow input and output XML files. Moreover, the use of schemas facilitates further data transformations, for example, by applying XSL transformation to XML files of the signal substantiation workflow to create XGMML file graphs that can be visualized with Cytoscape. The workflows and web services are described in the following sections. All workflows have been implemented and tested using Taverna Workflow Management system version 2.2.
We have implemented two workflows for signal filtering. The ADR-FM workflow is a MeSH®-based approach to find drug-event pairs in Medline® citations. The ADR-FD workflow uses text-mining to find the drug-event pairs in Medline® abstracts, databases such as DrugBank and drug labels available at DailyMed®.
A dataset of drug-side effects was downloaded from SIDER (December 2011) [25]. We restricted the SIDER dataset of total 61102 drug-event associations to 28251 associations between 492 drugs and 974 side effects by (i) mapping the used drug and event identifiers to the vocabularies used in our framework (ATC codes for drugs and UMLS concept identifiers for adverse events), and (ii) restricting to drugs and events for which protein annotations were available. P-values were computed using Fisher exact test and FDR was used to correct for multiple hypothesis testing.
We used the protein-protein interaction representation of the Reactome pathways (http://www.reactome.org/download/current/homo_sapiens.interactions.txt.gz, January 2012) to calculate the shortest path between any pair of antipsychotic drug and QTPROL associated proteins. For this purpose, we used the implementation of the Dijkstra algorithm in the Perl package Graph (http://search.cpan.org/~jhi/Graph-0.94/lib/Graph.pod). We then computed the average shortest path length for randomly chosen combinations of drug and event proteins and used a one-sided t-test to assess if the shortest path between the drug and event proteins as observed in our analysis was shorter than compared to random.
The EU-ADR project focuses on a selection of adverse drug reactions that are monitored in electronic health records and further analyzed by the filtering and substantiation workflows [7], [8]. These events were defined in terms of UMLS Metathesaurus® concept identifiers as described in [51], [52]. The event codes and names as defined in the EU-ADR project are listed in Table 5. The mapping of events codes or strings to UMLS Metathesaurus® concept identifiers and other vocabularies such MeSH® and OMIM is implemented within the web services. The ADR-S workflow accepts events as defined in the EU-ADR project or any other clinical event defined by UMLS concept identifier. The UMLS concept identifiers are processed to map them to MeSH® and OMIM identifiers using the UMLS Metathesaurus®.
The availability of web services and workflows presented in this work is detailed in Table 4.
|
10.1371/journal.pgen.1007129 | Acute Smc5/6 depletion reveals its primary role in rDNA replication by restraining recombination at fork pausing sites | Smc5/6, a member of the conserved SMC family of complexes, is essential for growth in most organisms. Its exact functions in a mitotic cell cycle are controversial, as chronic Smc5/6 loss-of-function alleles produce varying phenotypes. To circumvent this issue, we acutely depleted Smc5/6 in budding yeast and determined the first cell cycle consequences of Smc5/6 removal. We found a striking primary defect in replication of the ribosomal DNA (rDNA) array. Each rDNA repeat contains a programmed replication fork barrier (RFB) established by the Fob1 protein. Fob1 removal improves rDNA replication in Smc5/6 depleted cells, implicating Smc5/6 in the management of programmed fork pausing. A similar improvement is achieved by removing the DNA helicase Mph1 whose recombinogenic activity can be inhibited by Smc5/6 under DNA damage conditions. DNA 2D gel analyses further show that Smc5/6 loss increases recombination structures at RFB regions; moreover, mph1∆ and fob1∆ similarly reduce this accumulation. These findings point to an important mitotic role for Smc5/6 in restraining recombination events when protein barriers in rDNA stall replication forks. As rDNA maintenance influences multiple essential cellular processes, Smc5/6 likely links rDNA stability to overall mitotic growth.
| Smc5/6 belongs to the SMC (Structural Maintenance of Chromosomes) family of protein complexes, all of which are highly conserved and critical for genome maintenance. To address the roles of Smc5/6 during growth, we rapidly depleted its subunits in yeast and found the main acute effect to be defective ribosomal DNA (rDNA) duplication. The rDNA contains hundreds of sites that can pause replication forks; these must be carefully managed for cells to finish replication. We found that reducing fork pausing improved rDNA replication in cells without Smc5/6. Further analysis suggested that Smc5/6 prevents the DNA helicase Mph1 from turning paused forks into recombination structures, which cannot be processed without Smc5/6. Our findings thus revealed a key role for Smc5/6 in managing endogenous replication fork pausing. As rDNA and its associated nucleolar structure are critical for overall genome maintenance and other cellular processes, rDNA regulation by Smc5/6 would be expected to have multilayered effects on cell physiology and growth.
| The conserved Smc5/6 complex (or Smc5/6) is required during normal growth and for coping with genotoxins [1–4]. Due to the essential nature of the complex, studies thus far have examined partial loss of function mutants of the complex in various organisms. As its chronically sick alleles give varied phenotypes, a coherent view of Smc5/6 function during growth has yet to be established. In budding yeast, studies using temperature sensitive alleles suggest that Smc5/6 is required during S phase, as shifting to non-permissive temperatures during S, but not G2-M phase, leads to defects [5]. However, two views about the S phase functions of Smc5/6 have been proposed based on distinct mutant phenotypes. One smc6 allele (smc6-56) impairs replication of longer chromosomes while another (smc6-9) only diminishes the duplication of chromosome XII (Chr XII), which harbors the entire ribosomal DNA (rDNA) array [5–7]. The former defect was interpreted as reflecting Smc5/6 roles in replication fork rotation [6], while the mechanism for the latter defect was unclear [5,7]. More recently, another study proposed that Smc5/6 is essential in G2, but not S phase, as fusion of Smc5/6 with a G2-cyclin, but not S-cyclin, cassette sustained cell growth [8]. A number of factors likely contribute to the diverse phenotypes observed for this collection of chronic alleles. For example, cells containing chronically altered alleles can accumulate different levels or types of stress over generations, engendering a phenotype compounded from primary defects and various secondary changes. Indeed, smc5/6 alleles show alterations in diverse processes, including cell cycle checkpoint responses, cohesin function, repair pathway usage, and centromeric regulation [9–12]. As such, it is a challenge to deconvolute chronic mutant phenotypes and derive primary roles for Smc5/6.
The lack of a cohesive understanding of primary Smc5/6 defects in normal cell growth hinders advances in the field and is an important issue to address. An effective way to overcome the drawbacks of chronic allele usage is to induce acute and potent Smc5/6 depletion, which enables identification of the immediate consequences of Smc5/6 loss. However, robust loss of Smc5/6 within one cell cycle is difficult to achieve. For example, the budding yeast Smc5/6 subunits appear to be stable and even low levels of the complex can be tolerated for multiple cell divisions [13]. Previous strategies using conditional promoters or DHFR-degron systems reduced Smc5/6 protein levels and cell growth, but failed to cause the null phenotype of lethality [14,15]. To improve the robustness of Smc5/6 depletion, we turned to an auxin-inducible degron (AID) system [16]. We found that while targeting a single subunit of the Smc5/6 complex with AID did not cause lethality, a null phenotype could be achieved by combining AID fusions of two subunits. Thus, we used this Smc5/6 ‘double degron’ system to examine the immediate consequences of robust Smc5/6 degradation in the first cell cycle after loss. Our findings using this system demonstrate that the primary effect of Smc5/6 loss is defective rDNA replication in yeast. We present further genetic and DNA analysis data to derive a mechanism by which Smc5/6 promotes rDNA replication. Our data suggest that Smc5/6 is involved in managing programmed replication pausing at rDNA and restrains Mph1-mediated recombinogenic events to enable proper duplication of this at-risk locus.
To circumvent the limitations of smc5/6 hypomorphs used in previous studies, we employed an inducible, plant hormone-based, AID degron system to achieve acute depletion of Smc5/6 complex subunits. This system exploits the ability of a diffusible plant hormone auxin (IAA) to bridge the interaction between a plant adapter protein (TIR1)-bound endogenous ubiquitin ligase complex (SCF) and a TIR1-binding cassette (AID) fused to a target protein [16]. IAA addition recruits AID-fusion proteins to the TIR1-SCF ubiquitin ligase complex, which polyubiquitinates them for proteasome-mediated degradation (S1A Fig). Neither TIR1 nor functional concentrations of IAA are toxic to cells or interfere with other cellular processes, and rapidly inducible degradation has been reported for many yeast proteins using this system [17].
Each of the eight subunits of the Smc5/6 complex (Smc5, Smc6, Mms21, and Nse1, 3–6) was tagged with a C-terminal AID at its endogenous locus. Without IAA or TIR1, these strains gave rise to wild-type sized colonies, except for the Nse1-AID allele, which showed slow growth and was thus excluded from further analyses (Fig 1A). IAA addition elicited a slow growth phenotype in strains containing AID-tagged Smc5/6 alleles only if TIR1 was also present (Fig 1A). Immunoblotting for targeted proteins confirmed significantly reduced levels within 90 minutes after IAA addition (Fig 1B). These results indicate that single Smc5/6 AID degron alleles cause growth defects.
As Smc5/6 null alleles are lethal [2], the slow growth seen with single AID degrons of Smc5/6 subunits suggested that cells tolerate low levels of Smc5/6. Similar conclusions were reached in previous studies titrating cellular levels of Smc5/6, or another SMC complex, cohesin [14,15,18]. To enhance the robustness of Smc5/6 depletion, we constructed combined degrons based on the following rationale. We observed that loss of one subunit of the yeast Smc5/6 complex generally did not affect the levels of the other obligate members of the complex (S1B–S1D Fig). Thus, reducing the level of a second subunit should further decrease the probability of forming intact complexes. Indeed, we found that combining AID degrons of Nse5 and Smc6 proved highly effective for eliminating colony formation upon IAA treatment (Fig 1C). Thus, Nse5-AID Smc6-AID strains containing TIR1, referred to as “Nse5-Smc6 double degron” or “double degron” for simplicity, were used to investigate the primary defects caused by Smc5/6 complex depletion.
To evaluate whether the Nse5-Smc6 double degron strain suffers chronic defects prior to induced protein degradation, we compared it to three frequently used hypomorphic alleles: smc6-56, smc6–P4, and mms21-11 [2,6,19]. These three mutants grow more slowly than wild-type cells at permissive temperatures and are lethal at 37°C [2,6,19]. Even at permissive temperatures, they showed higher levels of dNTPs, a condition associated with increased DNA stress and altered replication profiles (S2A Fig) [20,21]. In contrast, cells harboring the Nse5-Smc6 double degron exhibited dNTP levels similar to wild-type or strains containing TIR1 alone, in both G1 and asynchronous cultures (S2B Fig). These data suggest that Nse5-Smc6 double degron cells lack chronic genome stress prior to induced degradation with IAA. The double degron strain’s normal growth and wild-type dNTP levels under un-induced conditions offer a more optimal baseline for detecting primary defects after Smc5/6 depletion.
We also examined the extent of Nse5 and Smc6 protein loss upon IAA addition and effects on bulk DNA replication. To this end, G1-arrested double degron strains were treated with IAA for 90 min to induce protein degradation, and then released into cycling with IAA (Fig 1D, +IAA). As a control, the same strains were examined in parallel without IAA (Fig 1D,–IAA). Protein levels were monitored at four time points after release from G1 (Fig 1E). Relative to controls, double degron cells in IAA lost ~95% of Nse5 and 70% of Smc6 proteins in the first cell cycle (20–80 min) (Fig 1E). It is reasonable to infer that levels of the intact complex are likely lower than 5% of those in wild-type cells, as depletion of each subunit independently affects the complex.
Next, we assessed bulk genome replication by FACS analysis. Throughout the time course, the double degron cells showed similar cell cycle progression in the presence or absence of IAA (Fig 1D). On a molecular level, Clb2 kinetics, an indicator of cell cycle progression [22], were also comparable (Fig 1E). IAA-treated double degron cells also showed no increased phosphorylation of the Rad53 checkpoint kinase or gross differences in levels of γH2A, a marker for DNA replication or breaks (Fig 1E) [23,24]. Our data suggest that there are no detectable changes in cell cycle progression, DNA break marker increase, or checkpoint activity in the first cell cycle upon Smc5/6 subunit removal. Moreover, cells appear to undergo normal bulk genome replication even with robust Smc5/6 depletion.
To achieve more sensitive detection of chromosome synthesis in the first cell cycle of Smc5/6 loss than that afforded by FACS, we used BrdU incorporation coupled to pulsed field gel electrophoresis (PFGE). PFGE can separate replicated chromosomes, which enter the gel, from the branched forms still undergoing replication, which remain trapped in the wells [25]. BrdU labels the newly synthesized DNA in each replicated chromosome and can be detected by immunoblotting [26]. To apply these techniques, we used a similar procedure as described above and monitored cells for 180 min after G1 release (Fig 2A). To capture new DNA synthesis, we added BrdU immediately after G1 release. We also used nocodazole to prevent additional rounds of cycling in order to focus on effects of Smc5/6 depletion in the first cell cycle (Fig 2A).
Nse5-Smc6 double degron cells were first compared to cells containing the TIR1 adaptor but without degron alleles. As before (Fig 1D), FACS data showed that both strains were synchronized in G1, progressed through S phase, and achieved G2/M arrest (Fig 2B). For control cells, BrdU signals for all chromosomes increased from 20’ to 40’ and peaked at 60’ (Fig 2C). This is consistent with the FACS profile, which shows bulk replication having largely completed by 60’ (late S phase), with cells remaining in G2/M for subsequent time points. Strikingly, both BrdU blotting and DNA staining of PFGE gels showed little to no replicated Chr XII signal in Nse5-Smc6 double degron cells throughout the duration of the time course, despite wild-type-like cell cycle progression (Fig 2C; S3A Fig). Quantification of BrdU signals for other chromosomes found no significant differences between IAA-treated double degron and control cells (S3B Fig).
To verify that the lack of fully replicated Chr XII signal was due to the loss of AID-targeted proteins, we repeated the assay with double degron strains in the presence or absence of IAA. Identical first cell cycle progression was seen for both conditions (Fig 2D). Once again, only IAA-treated degron cells showed low Chr XII signal, as detected by both BrdU incorporation and DNA staining (Fig 2E; S3C Fig). Based on these results, we concluded that acute Smc5/6 depletion leads to a Chr XII-specific replication defect in the first cell cycle.
Chr XII is unique among yeast chromosomes in that it harbors the entire rDNA array. This large array (~1.4 Mb) represents 10% of the yeast genome and half of Chr XII, and is intrinsically difficult to replicate. Uniquely, each of the 100–200 rDNA repeats in the array contains a programmed replication fork barrier, or RFB, located between the 5S and 35S rRNA genes (Fig 3A) [27]. When bound by Fob1, the RFB sequence can block replication fork progression in the direction of 35S rRNA transcription [28]. This mechanism helps avert collisions between the replication and transcription machineries, but also requires careful management to enable replication completion and avoid repeat instability [29].
Knowing that rDNA harbors these specific sites of replication blockade, we asked whether their removal could ameliorate the observed Chr XII replication defect of Smc5/6-depleted cells. To this end, we removed Fob1 in the Nse5-Smc6 double degron strains, and repeated the BrdU and PFGE tests described above (Fig 2A). Both degron and degron fob1∆ strains showed identical first cell cycle FACS profiles (Fig 3B). Importantly, the fully replicated Chr XII BrdU signal was stronger in degron fob1∆ strains than in degron strains at 120’ and 180’ after release from G1 (Fig 3C). This finding was further confirmed by Southern blotting with an rDNA-specific probe (Fig 3D). Quantification showed ~3-fold increased Chr XII signals in degron fob1∆ cells over degron alone (Fig 3D). As expected, these fob1 effects were specific to Chr XII, as signals for other chromosomes such as Chr III were similar for the two strains (Fig 3D).
To determine if the observed fob1∆ effect on Chr XII replication in double degron cells reflects an improvement of rDNA replication per se, we directly examined the rDNA array, which can be released from Chr XII by the restriction enzyme XhoI. The rDNA array is flanked by XhoI recognition sites but contains no internal ones, so XhoI cleavage releases the entire array from its chromosomal context [30]. The array’s large size enables its resolution from other smaller chromosome fragments by PFGE and can be subsequently detected by hybridization to an rDNA-specific probe. As shown in Fig 3E, ~90% of the rDNA signal failed to enter the gel in degron cells even by 180 min, consistent with our data for intact Chr XII (Fig 3D). This confirmed that the rDNA array itself suffers from incomplete replication when Smc5/6 is depleted. The rDNA of degron fob1Δ cells entered the gel from 60 to 180 minutes, and in-gel levels of rDNA were 4–6 fold greater than those of degron cells (Fig 3E). The ability of fob1Δ to improve replication of the rDNA array more than that of Chr XII (Fig 3D and 3E) confirms that rDNA is responsible for the beneficial effect exerted by fob1Δ on Chr XII replication in double degron cells.
We also asked whether replication fork blockade by Fob1-RFB, outside the context of rDNA, were sufficient to create a requirement for Smc5/6. It is known that two RFB sites inserted on Chr III can pause replication forks upon Fob1 overexpression, and that such pauses are resolved by the recruitment of the Rrm3 helicase [31]. We found that Smc5/6 loss in this system impaired Chr XII replication as expected, but did not affect Chr III replication (S4 Fig). Thus, additional properties specific to the rDNA locus not recapitulated by these RFB sites contribute to the importance of Smc5/6 for rDNA replication (see Discussion).
We then investigated which other protein factors might play a role in Smc5/6-dependent effects at rDNA. Previous studies showed that Smc5/6 inhibits the pro-recombinogenic activity of the Mph1 DNA helicase at stalled forks under DNA damage conditions [19,32–34]. Although Mph1 has not been implicated in rDNA and Fob1-RFB-mediated replication pausing, replication forks stalled by Fob1-RFB, like those stalled by template lesions, require management to ensure replication completion. Thus, we tested whether Smc5/6 inhibition of Mph1 may also be relevant at endogenous rDNA fork blockage sites.
To this end, we deleted MPH1 in Nse5-Smc6 double degron cells. FACS profiles of both degron and degron mph1Δ showed identical first cell cycles (Fig 4A). Using the procedure described above, we found that mph1Δ increased levels of fully replicated Chr XII by about three-fold in degron cells (Fig 4B). This effect was similar to fob1Δ, although significant suppression was seen by 60’ for mph1Δ, but not fob1Δ. Furthermore, when examining the rDNA array within Chr XII by XhoI digestion, we found that its duplication in degron mph1Δ cells increased to levels similar to that of Chr XII (Fig 4C). We note an overall trend of weaker suppression of rDNA replication than Chr XII as a whole by mph1Δ, while fob1Δ had the opposite trend. This would be consistent with a role for Mph1, but not Fob1, at Chr XII loci outside rDNA.
After observing fob1 and mph1 suppression, we tested their genetic interactions. If Smc5/6 is required to limit Mph1 activity at Fob1-mediated fork pausing sites, we would expect combined fob1Δ mph1Δ to confer no additive effects on the rDNA replication phenotype of degron cells. PFGE and Southern blotting for the rDNA array showed that fob1Δ mph1Δ improved rDNA gel entry at 60, 120, and 180 min in Nse5-Smc6 double degron cells, with only a small proportion of rDNA signal remaining in the wells (Fig 5A and 5B). Quantification of several experiments shows that this improvement of rDNA replication by fob1Δ mph1Δ was not significantly greater than that shown by fob1Δ or mph1Δ single mutants (Fig 5C). In addition, the observed suppression reached a level similar to those of wild-type strains and double degron cells without IAA treatment (Fig 5C; S5 Fig). These data suggest that Mph1 and Fob1 function in the same pathways. We note that stronger effects for fob1Δ than mph1Δ at earlier time points may reflect additional roles played by Smc5/6 at rDNA beyond Mph1 regulation.
Our data so far support a premise that fork stalling by Fob1-RFBs in rDNA necessitates the presence of Smc5/6 to inhibit the pro-recombinogenic Mph1 activity; as such, removing Fob1 or Mph1 bypasses the need for Smc5/6. Such a potential role for Smc5/6 would mitigate recombination at RFBs and favor fork merging, an outcome less likely to cause rDNA repeat instability [29]. To test the above idea, we used 2D gel analyses to examine recombination structures formed at regions around RFB sites. An rDNA repeat fragment released by BglII restriction digest contains the RFB, rDNA replication origin (rARS), 5S rRNA gene, and part of the 35S rRNA gene (Fig 6A). As shown in previous studies, examining this fragment by 2D gel enables one to monitor levels of stalled forks at RFBs, regular replication forks (Y-shaped DNA), and recombination intermediates (X-shaped DNA or X-mols) [35] (Fig 6A).
Using the same experimental schemes as for the PFGE experiments (Fig 2A; S6A Fig), we first compared Nse5-Smc6 double degron cells and control cells with TIR1 alone. We examined samples from one S-phase time point (60 min) and one G2-phase time point (120 min) since double degron cells show reduced rDNA replication at both time points (Fig 2B–2E). Nse5-Smc6 double degron and control cells differed significantly in their X-mol or recombination intermediate levels at both time points (Fig 6B, arrows). Quantification of several experiments showed a ~1.5-fold increase at 60 min and 3-fold increase at 120 min for degron cells over controls (Fig 6C). A ~3-fold increase of RFB signals in degron cells at 120 min was also seen (Fig 6B; S6B Fig). These data suggest that the rDNA replication defect caused by Smc5/6 loss is associated with increased recombination and prolonged fork pausing.
We went on to ask whether increased recombination at RFB regions upon Smc5/6 loss is mediated by fork pausing and Mph1. 2D gel analyses found lower levels of recombination intermediates in both fob1Δ double degron and mph1Δ double degron strains relative to degron strains at 60 min and 120 min after G1 release; this reduction resulted in levels comparable to those of control cells (Fig 6B and 6C). As expected, fob1Δ also eliminated RFB-dependent fork pausing, while mph1Δ did not exert such an effect (Fig 6B). Moreover, the reduction in recombination intermediate levels for fob1Δ mph1Δ double degron cells was similar to that of fob1Δ or mph1Δ single mutants (Fig 6B and 6C), a finding consistent with our PFGE findings for rDNA array replication (Fig 5C). Together, our data suggest that both Mph1 activity and Fob1-mediated fork pausing contribute to recombination structure accumulation in the absence of Smc5/6.
The above data do not exclude additional mechanisms by which Smc5/6 could influence rDNA metabolism. It has been shown that the SUMO ligase activity of the Smc5/6 complex subunit Mms21 affects nucleolar function but is not essential for growth [2,36,37]. We thus tested whether Mms21 SUMO ligase function is directly linked to rDNA replication. We found that a SUMO ligase mutant of Mms21 did not affect rDNA replication at early time points after release from G1 (S7 Fig). This is different from our data regarding Smc5/6 loss, but consistent with previous findings that sumoylation is not required for Mph1 regulation [19,33]. At a later time point, the mms21 SUMO E3 mutant showed moderately reduced rDNA gel-entry, suggesting a late role for sumoylation. This observation corroborates a proposed role for the Mms21 SUMO ligase in dissolving recombination intermediates that block replication completion [38,39].
Despite being required for viability in multiple organisms, the role(s) played by Smc5/6 during mitotic growth remain poorly understood [1–3,40]. The varied phenotypes of chronic smc5/6 mutants have complicated the delineation of specific Smc5/6 functions. Acute Smc5/6 depletion offers a strategy for bypassing the obscuring secondary effects of chronic Smc5/6 loss. An inducible protein degradation system enabled us to investigate the effects of Smc5/6 loss on DNA replication and cell cycle progression within the first cell cycle after depletion (Fig 1C–1E). This system developed for analysis of the immediate effects of robust Smc5/6 loss in yeast can stimulate the development of similar approaches in other organisms.
Combining our acute Smc5/6 depletion system with chromosomal PFGE analyses, we show that Smc5/6 complex removal causes a striking Chr XII-specific replication defect (Fig 2B–2E). This is not associated with detectable changes in cell cycle progression or markers of DNA damage and checkpoint activation, suggesting that we have isolated a primary defect of Smc5/6 loss (Fig 1D and 1E). We further show that the observed replication defect is largely localized to the rDNA array on Chr XII (Fig 3D and 3E). Our findings are in line with previous reports for smc6-9, but not with the chromosome size-based model derived from studies of smc6-56 [5,6]. Considering that smc6-56 cells experience chronic genome stress based on their altered growth and dNTP levels, adaptive responses in these cells may contribute to their overall phenotype (S2A Fig). Alternatively smc6-56, but not smc6-9, may alter the Smc5/6 function in dealing with longer chromosomes. Our finding that the Smc5/6-dependent Chr XII replication defect begins in S phase (Fig 2E) suggests that Smc5/6 plays a role in this cell cycle phase. This conclusion corroborates chromatin immunoprecipitation data localizing Smc5/6 to replication forks [41,42]. It can also be reconciled with the ability of Smc5/6 expressed from a G2-cyclin module to sustain viability, since our data suggests that even low level expression (in S phase) may be functionally adequate (Fig 1A and 1B). Taking into consideration these findings, we conclude that Smc5/6 is required beginning in S phase, particularly for rDNA replication, through to post-replicative G2 events. Our phenotypic assessment of acute Smc5/6 loss clarifies the essential function of Smc5/6 in mitotic cells and enables more reliable interpretation of smc5/6 defects.
After redirecting focus towards an essential role of Smc5/6 in replicating at-risk rDNA loci, we provided genetic, PFGE, and 2D gel data to derive a mechanism by which Smc5/6 promotes rDNA replication. We show that Smc5/6 loss leads to increased levels of recombination intermediates at programmed rDNA fork pausing sites (Fig 6B and 6C). Importantly, removing the rDNA fork blocking protein Fob1 or the DNA helicase Mph1 in Smc5/6 degron cells reduces these intermediates and rDNA replication defects (Figs 3D, 3E, 4B, 4C, 6B and 6C). As fob1Δ and mph1Δ show similar and non-additive effects, the simplest interpretation is that upon fork stalling at RFBs in rDNA, Mph1-mediated recombination is a major contributor to defective rDNA replication when Smc5/6 is absent (Figs 5B, 5C, 6B and 6C). On the basis of these data, we propose a model in which Smc5/6 helps to manage replication forks paused at RFBs by inhibiting Mph1-mediated recombination (Fig 6D). As such, Smc5/6 is an important factor that influences the fates of stalled forks at this locus. When Smc5/6 is present, Fob1 at RFBs can prevent transcription-replication conflicts, and fork merging is favored. When Smc5/6 is absent, paused forks are vulnerable to recombination. The recombination intermediates thus generated are especially toxic, because Smc5/6 SUMOylation function is required for their dissolution [38,39]. This is consistent with our observation that mms21 SUMO ligase mutants affect rDNA replication at a later time point (S7 Fig). This model provides a straightforward explanation for fob1 and mph1 suppression, as the former reduces the number of forks that require Smc5/6 protection from Mph1 activity, while the latter reduces the potential for paused forks to undergo recombination. Both genetic changes decrease the Smc5/6 requirement at rDNA. As the fob1 and mph1 suppression patterns are not entirely identical, they must also play unique, yet-to-be determined roles in mediating Smc5/6 effects on rDNA replication.
Our data also reveal a previously unappreciated function of Mph1 at the rDNA locus. Although recombination at RFBs is toxic when Smc5/6 is lost, such repair could be useful for adjusting rDNA repeat numbers. Whether and how cells enable restricted use of Mph1 for this purpose will be interesting to investigate in the future. Mph1 and its homologs have been suggested to promote recombination at stalled replication forks via their ability to regress forks, which entails the annealing of two nascent strands accompanied by re-annealing of their template strands [32]. Replication fork regression can provide a mechanism for replication restart, but also generate DNA structures prone to cleavage or recombination [32]. As such, fork regression is negatively regulated by multiple mechanisms; Smc5/6 participates in one such mechanism by directly binding to and preventing Mph1 oligomerization at fork junctions [32, 33]. Though fork regression is mostly investigated in DNA damage conditions, our findings suggest that it can also occur in situations of endogenous fork pausing. In conjunction with previous findings of the Smc5/6-Mph1 relationship under DNA damage conditions, our data helps to establish the concept that a key Smc5/6 function is regulation of Mph1 in multiple conditions.
Overall, our results suggest that Smc5/6 plays a primary role in managing programmed fork pausing at rDNA by inhibiting pro-recombinogenic Mph1 activity. We find this role to be rDNA-specific, as other chromosomes harboring protein barriers [43,44] replicate proficiently upon Smc5/6 depletion and artificially introduced protein barriers on Chr III did not affect its replication in double degron cells [31] (S4 Fig). We suggest that multiple features unique to rDNA may help explain why Smc5/6 is particularly indispensable at this locus. First, the presence of many RFB sites in rDNA can generate much higher levels of replication fork pausing than any other locus. Second, some proteins that facilitate replication are known to be excluded from the nucleolus [45]. As such, Smc5/6, known to be enriched at rDNA [7], could be particularly important for managing stalled forks at this locus. When forks are paused outside rDNA, other pathways may compensate for the lack of Smc5/6 activity, so they do not manifest defects as robustly upon Smc5/6 loss. Third, rDNA has a unique chromatin environment and architecture connected to its activities in transcription, ribosome assembly, and mitotic exit. This high level of DNA, RNA, and protein transactions may generate a specific demand for Smc5/6 function. Finally, rDNA replication continues at the end of S and into G2 phase when other chromosomes have completed their replication [29]. This could further generate a demand for Smc5/6. In-depth examination of which and how these unique rDNA features influence its duplication and pose a critical requirement for Smc5/6 will provide additional insight into how this highly conserved repetitive sequence sustains stability across evolution.
In human cells, Smc5/6 also promotes replication and has been shown to localize to a subset of stalled replication forks [46,47]. Thus, our work also stimulates the examination of a similar role for human Smc5/6 in managing fork pausing in rDNA and other repetitive or highly organized regions where alternative mechanisms of fork management are ineffective. It is worth noting that disrupting rDNA replication has far-reaching consequences beyond this locus, through its strong influences on mitotic exit, nuclear organization, transcription, and translation [48,49]. Indeed, yeast Smc5/6 mutants show nucleolar fragmentation and stress [2,37]. As such, the importance of Smc5/6 function to rDNA replication may help explain the divergent phenotypes underlying its two human syndromes, which feature pleiotropic developmental defects across multiple lineages, including the musculoskeletal, endocrine and immune systems [4,50].
Yeast strains are derivatives of W1588-4C, a RAD5 variant of W303 (MATa ade2-1 can1-100 ura3-1 his3-11,15 leu2-3,112 trp1-1 rad5-535) [2]. At least two strains per genotype were examined in each experiment, and only strain is listed for each genotype in S1 Table. Standard methods were used for yeast strain construction. To tag subunits of the Smc5/6 complex, PCR products containing AID and FLAG tag sequences were generated with flanking homologous sequences to the insertion site. Standard PCR integration methods were used to generate fusion constructs, which were then fully sequenced to confirm the correct tagging.
Two main protocols were used for cell cycle studies, as shown in Figs 1D and 2A. In Fig 1D, asynchronous cultures were arrested in G1 phase by adding 5μg/ml α-factor (ThermoFisher) to media for 90 min. 1 mM IAA (Sigma) was added for 90 min to degrade Nse5-AID and Smc6-AID proteins. Subsequently, cells were released into fresh YPD containing 1 mM IAA to sustain Nse5 and Smc6 depletion in degron cells. Samples were collected at several time points after release as indicated. Note that the same procedure was applied to cells without degron alleles to ensure parallel treatment. A similar experimental procedure shown in Fig 2A has two alterations. One is BrdU (Sigma) addition after cells were released from G1 arrest to monitor new DNA synthesis. Another is nocodazole (US Biologicals) addition to cultures 45 min after G1 release to prevent first cell cycle exit. Standard FACS analyses were performed as described previously [51].
2D gel analyses were performed as previously described [52]. DNA was extracted and digested by BglII and separated by agarose gel electrophoresis in two dimensions. DNA was transferred onto Hybond-XL membranes (GE Healthcare) and analyzed by Southern blot using probes hybridizing specifically to rDNA. Primers used for probe amplification are available upon request. For quantification, the signals of 1N DNA were obtained from shorter exposures, while those of DNA intermediates came from longer exposures to ensure both types of signals fell within linear range of detection on the PhosphorImager.
PFGE was performed as previously described [53]. In brief, cells harvested from the indicated time points were embedded in agarose plugs, spheroplasted, and deproteinized. Plugs were loaded into 0.5X TBE gels and run on a Bio-Rad CHEF-DR III Pulsed Field Electrophoresis System for 12 hours to achieve chromosome separation. Gels were stained by ethidium bromide and Sytox (Molecular Probes) and then transferred onto Hybond-N+ membranes (GE Healthcare) using standard capillary transfer technique. Membranes were probed with anti-BrdU antibody (BD) and α-mouse secondary antibody (GE Healthcare). Membranes were scanned with Fujifilm LAS-3000 luminescent image analyzer, which has a linear dynamic range of 104 to achieve reliable quantification. The percentage of gel entry for each chromosome was calculated by dividing the chromosomal band signal by the sum of chromosomal band signal and well signal, after background subtraction. The positions of each chromosome were derived from [54]. Southern blotting of Chr XII, Chr III, and rDNA were performed using specific probes hybridizing to each region, and primers used for probe amplification are available upon request.
To detect protein levels, standard TCA protein extraction methods were used [55]. Protein samples were resolved on 3–8% or 4–20% gradient gels (Life Technologies and Bio-Rad) and transferred onto 0.2 um nitrocellulose membranes (G5678144, GE). Antibodies used were: α-Myc (9E10, Bio X Cell), α-HA (F-7, Santa Cruz), α-Flag M2 (Sigma), α-V5 (Life Technologies), α-Clb2 (y-180, Santa Cruz), α-Pds1 (gift of E. Schiebel), α-Rad53 (yC-19, Santa Cruz), α-H2A pS129 (Abcam), and PAP (P1291, Sigma). dNTP quantification was performed as previously described [56].
|
10.1371/journal.pntd.0002821 | Interruption of Infection Transmission in the Onchocerciasis Focus of Ecuador Leading to the Cessation of Ivermectin Distribution | Introduction: A clinically significant endemic focus of onchocerciasis existing in Esmeraldas Province, coastal Ecuador has been under an ivermectin mass drug administration program since 1991. The main transmitting vector in this area is the voracious blackfly, Simulium exiguum. This paper describes the assessments made that support the decision to cease mass treatment.
Methodology and Principle Findings: Thirty-five rounds of ivermectin treatment occurred between 1991–2009 with 29 of these carrying >85% coverage. Following the guidelines set by WHO for ceasing ivermectin distribution the impact on parasite transmission was measured in the two vector species by an O-150 PCR technique standard for assessing for the presence of Onchocerca volvulus. Up to seven collection sites in three major river systems were tested on four occasions between 1995 and 2008. The infectivity rates of 65.0 (CI 39–101) and 72.7 (CI 42–116) in 1995 dropped to zero at all seven collection sites by 2008. Assessment for the presence of antibodies against O. volvulus was made in 2001, 2006, 2007 and 2008 using standard ELISA assays for detecting anti-Ov16 antibodies. None of total of 1810 children aged 1–15 years (between 82 and 98% of children present in the surveyed villages) tested in the above years were found to be carrying antibodies to this antigen. These findings were the basis for the cessation of mass drug treatment with ivermectin in 2009.
Significance: This fulfillment of the criteria for cessation of mass distribution of ivermectin in the only known endemic zone of onchocerciasis in Ecuador moves the country into the surveillance phase of official verification for national elimination of transmission of infection. These findings indicate that ivermectin given twice a year with greater than 85% of the community can move a program to the final stages of verification of transmission interruption.
| Onchocerciasis has been known to be endemic in the northwestern coastal riverine jungle areas of the country since the early 1980's. A mass drug administration program with ivermectin was implemented in 1991, and in recent years has included consistent twice a year treatment. The impact of this program, and progress towards eliminating the transmission of Onchocerca volvulus from the endemic zone, was assessed by studying entomological parameters at sentinel sites in 1995, 2000, 2004 and 2008 using PCR detection of infective larvae in the vectors (Simulium exiguum and Simulium quadrivittatum); the survey of 2008 showed that all of the collection sites had reached a level consistent with the interruption of transmission. Serological assessment of children and adolescents in 2001–2008 also showed that transmission had been interrupted. These findings support the contention that the Ecuadorian Program has reached the post-treatment surveillance phase of the elimination program; mass drug administration of ivermectin was consequently stopped. The factors contributing to this successful achievement include ivermectin coverage of consistently around 85% or greater since 1998, careful consideration and control of possible expansions of the endemic area through migration, and the maintenance of a strong community health programme.
| Onchocerciasis was recognized in Ecuador some 30 years ago and extensive work during the 1980s identified the limits of the endemic focus in Esmeraldas Province in north west of coastal jungle area of the country [1], [2]. This was area of increasing population where residents were spreading along the various river systems, with the population in the endemic area by 2010 reaching around 26,000 people. The clinical disease resulting from infection with O. volvulus was described in the 1980's to be amongst the severest of all the American onchocerciasis foci with blinding disease and extensive onchodermatitis [3]–[5]. The vectors in the focus include S. exiguum and S. quadrivittatum [6]–[8], with S. exiguum being the most important, as it is a highly efficient vector for O. volvulus; S.exiguum has a vectorial competency comparable to forest cytotypes of S. damnosum sensu lato in terms of the percentage of flies developing infective stage larvae (L3s) and the numbers of L3 per fly [9]. The second vector species, S. quadrivitattum, is much less efficient due to the presence of a cibarial armature which damages microfilariae ingested during a blood meal.
In the 1980s the only approach to treatment and control of this infection was the removal of palpable nodules containing the adult worms; however this nodulectomy campaign did little to reduce the increasing prevalence of infection and did not reduce the increasing prevalence of clinical disease [10]–[12]. In 1991, as part of the global approach to control of onchocerciasis using the microfilaricidal drug ivermectin twice a year treatment was implemented in the endemic zone of Ecuador. This mass drug administration approach was aimed at eliminating the disease and its transmission, and the program and the monitoring activities have followed the international guidelines established by WHO [13].
The Ecuadorian program is based on a close relationship with the community and is supported by a strong health services approach, with constant communication with the field locations and with strong local support. Although the environment is a particularly difficult one in terms of access the national program has been able to maintain a constant link with the endemic population. This has been essential as concerns regarding migration and expansion of the focus have always been important questions. A number of the foci in Latin American onchocerciasis are found in isolated locations, and the Ecuadorian focus is also a relatively isolated jungle location, thus often making such drug interventions difficult to implement.
The assessment of success and progress towards control and elimination of transmission described here has been monitored through entomological and serological parameters [13]. Although clinical parameters are believed to be important indicators and provide useful programmatic information as to the progress towards control and elimination, entomology and serology have usually been used to define the actual transmission status of a program, and importantly to assist in any decision to cease mass drug administration. This present communication describes the entomological assessment of the Ecuadorian foci and the findings that led to the decision to cease drug treatment in this important Latin American focus of endemic onchocerciasis.
The Ecuadorian onchocerciasis control programme is carried out under the auspices and full approval of the National Ministry of Health and is thus fully supported by the Government of Ecuador. The distribution of drugs, the collection of skin and blood samples, as well as the collection of vectors, has been approved and monitored by the National Ministry of Health, and considered as ethical, safe, and to be of benefit to the local people. The study complied with current national and international regulations and standards for biomedical research in human subjects; informed consent was requested and obtained from the parents of the children for whom serological sampling was done.
Ecuador is one of the six countries endemic for onchocerciasis in Latin America. The infection is found in a single focus in the Esmeraldas Province of Northwestern Ecuador, where it is associated with several river basins (Figure 1). The locations where the disease and infection is most prevalent being the three major river systems, Rio Cayapas, Rio Santiago and the Rio Onzole. A total of approximately 25,900 people were at risk in the focus in 2008. Coverage figures were maintained by a rigorous accounting system that is based on an active regularly updated census of the villages. Assessments for clinical presentation, ocular and skin parasite loads were carried out on a number of occasions pre-ivermectin treatment and during the treatement period and will be, or have been, reported elsewhere [11], [12]. In the hyperendemic area the main monitoring site on the Rio Cayapa had community microfilarial levels of 81% in 1982 which had increased to over 90% by 1987. When assessments were carried out in 2004 the dermal load had dropped to 4% and by 2008 all dermal biopsies were free of microfilariae. The last time microfilariae were identified in the eyes of any of the residents was in 2000.
In 1987 Merck & Co. donated ivermectin for use in mass drug administration (MDA) programs for onchocerciasis [14]; Ecuador adopted this drug and began an MDA program in 1991. The Ecuadorian program has distributed drug once, or more usually twice, a year at 150 µg/kg to all eligible people (defined as those over 5 years of age, who are not pregnant or infirm); this is generally regarded to be approximately 80% of the total resident population. Drugs were distributed by local village health workers under the constant supervision of the Onchoerciasis National Program members from the central offices based in Quito and Borbon. The ivermectin distribution began in the main hyperendemic areas in 1991 with other communities added where necessary over the next two years.
Coverage rates were estimated from the detailed medical records kept on each resident in the treatment areas and expressed as % of the eligible individuals in the village. Throughout the period of mass drug administration (1991–2009) ivermectin was distributed in 35 rounds to eligible residents of the major endemic area (Figure 2) twice a year, except for 1995–7 when only one treatment was provided. In the majority of the 18 years during which treatment was given, at least one of the round of treatment exceeded this 85% level, and in most cases this level was achieved both times during the year. In the 12 year period, from 1998 to 2009, 24 bi-annual treatment rounds were administered, with a coverage of at least 85% of the eligible population in all but 2 of them (i.e. the second round in 2000–%, and in first round 2008–74%).
Black flies were collected at seven community sites in the Esmeraldas Province of Ecuador (Figure 1 & Table 1). Flies were collected using standard methods in the months of April-June, 2012; these months correspond to the peak of transmission. Parous females were not separated from nulliparous insects as these samples were used in a pool screening assay. Collections were carried out from 0800-1700 each day for 8 days per month. Flies were collected for 50 minutes each hour, allowing a 10-minute break to the collectors during each hour. The collected flies were sorted according to species (S. exiguum and S. quadrivitatum), divided into pools containing a maximum of 50 insects from a given species per pool, placed in isopropanol and stored at room temperature until analyzed by PCR.
For serology, samples were taken from children under the age of 17 from four sentinel communities (211 individuals in 2001; 519 in 2008) and eight extra-sentinel communities (318 individuals in 2006; 762 in 2007) for a total of 1941 individuals tested.
Heads and bodies from the collected flies were separated by freezing, agitation and separation through a 25 mesh sieve, as previously described [15]. DNA was prepared from the separated heads and bodies by proteinase K digestion, organic extraction and adsorption to a silica matrix, as previously described [16]. Pools were processed in groups consisting of 50 flies in each pool; one sham extraction served as an internal negative control. The resulting DNA preparations were used as templates in a PCR assay targeting an O. volvulus specific repeated sequence (O-150 PCR), as previously described [15], [16]. PCR products were detected by PCR-ELISA. Pools producing ELISA values which were equal to or greater than the mean plus three standard deviations of the values obtained from 10 negative control wells run on each plate were considered to be putatively positive for O. volvulus DNA. Putatively positive DNA samples were re-tested in an independent PCR procedure and samples that were positive in both assays classified as confirmed positives. Pools of bodies were initially screened, as bodies contain early stage larvae (microfilarial and L2 stages) and are the most sensitive indicator of parasite - vector contact. The prevalence of flies containing immature stages is 2 fold higher than the prevalence of flies containing infective stage larvae (L3) in S. exiguum and 20 fold higher in S. quadrivitatum [17]. DNA from head pools were screened if evidence for parasite-vector contact was observed in the screens of the body pools in order to obtain an estimate of the prevalence of flies containing L3. The upper bound of the 95% confidence interval for the prevalence of flies carrying O. volvulus parasites was calculated using the Bayesian algorithm of Poolscreen v 2.0. In undertaking these calculations, the mean number of L3s per infective fly was taken as 1, as reported to be the case in areas subject to effective control measures [16].
The Ov-16 ELISA assay uses a recombinant antigen of O. volvulus [18], [19], [20] to measure prevalence of IgG4 antibodies against the corresponding native antigen as a surrogate measure of exposure to the parasite. Blood spots for the Ov16 assay were collected in July 2007 and from December 2010 to January 2011. In both studies, sterile techniques were used to collect finger prick blood onto 5×5 cm area of filter paper (Whatman type 2). The saturated blood spots were dried, individually wrapped and transported at 4°C to the laboratory where they were stored at -20°C for further analysis. In the laboratory, sera were eluted from filter paper punches and used in standard ELISA assay as previously described [21]. A standard curve was used with each ELISA plate to identify positive samples and permit comparisons between plates and over time. The cut-off was chosen as 40 arbitrary units by identifying the value that optimized both sensitivity and specificity for Latin American and African positive and negative samples as previously reported [20], [21]. A sample is reported positive only after a second independent positive repeat test.
Poolscreen algorithm Versions 2.0 [17] was used to analyse the data and to calculate the infectivity rate (i.e. the prevalence of infection with L3), and associated 95% CI (Confidence Interval), in the vector populations. No statistics were required to compare serology results as all test were below the value set as positive [20], [21].
Samples taken in 1995 from two collection sites on the Rio Cayapas in the hyper-endemic zone showed infectivity rates (IR) of 72.7 and 65.0 (Table 1); these sites, and subsequently others, showed a reduction in IR for 2000 and 2004. In 2009 these and all tested sites were negative for O. volvulus DNA in the flies collected. The IR at each site declined in each additional year tested. In 2000 no evidence of the presence of infected flies was seen in any of the site tested in the Rio Santiago and testing was not continued on this river in subsequent years. Negative results were also achieved in the other two river systems in 2004 and 2008, although Rio Cayapas had two sites with low levels of IR in 2004, both of which became negative in 2008. All sites were negative by 2008.
Young residents under the age of 15 from 14 different communities in the major river system were tested during the period 2001–2008 on four different occasions; at no time was any individual found to be positive for anti-Ov-16 antibodies (Tables 2 and 3).
Nodulectomy, the approach originally used to control the increasingly prevalent clinical onchocerciasis seen in this endemic area before the introduction of ivermectin, was found to be singularly unsuccessful, with the prevalence of clinical disease, including severe eye problems, increasing significantly during the period from 1980 to 1989 [5], [12]. The introduction of the administration of ivermectin twice a year has had a remarkable effect on reducing all aspects of this disease complex and the findings described in this present paper indicate that the focus is moving successfully towards elimination of transmission of the infection.
The international community has developed a series of metrics, relying upon entomological and epidemiological indicators to determine when transmission had been suppressed [13]. Once transmission is determined to have been interrupted, mass drug distribution can cease. Three years after mass drug distribution ended, a post-treatment surveillance (PTS) should be undertaken, which uses entomological indicators to look for evidence of recrudescence of transmission [13]. If no evidence for ongoing transmission is uncovered during the PTS, it can be concluded that transmission had been eliminated. Thus these guidelines described the stages and testing required for an endemic area, be it a country or a site within a country, to be able to claim elimination. Essentially there are two major steps a program must make to achieve formal verification that transmission is ceased: first is the cessation of treatment based on evidence that vectors no longer carry the parasite, and the second is the maintenance of this interrupted state for three years (the post-treatment phase). The data presented in this paper fulfills the first phase of these requirements.
Despite the severity of the disease in the Ecuadorian onchocerciasis focus, and the voracity of the vector in this region, the strategy of high coverage and great emphasis on contact and communication with the affected communities has produced very satisfactory results. This success, achieved with arguably the most severe focus of onchocerciasis in Latin America, emphasizes the likelihood that the strategy of semiannual treatments will be effective in this region, and probably can also be successful in a number of foci in Africa and Yemen. Ecuador also faced a number of challenges that are to be expected with an isolated disease focus including the possibility of migration of infected people out of the focus and the attendant potential of establishment of new foci. This possibility was carefully monitored in the Ecuadorian programme and treatments carried out in any possible for areas outside the main focus, such as is shown in Figure 1, where three sites outside the main three river focus were monitored, and one tested in this present reported data. It is most important in establishing the elimination of onchocerciasis from the country as a whole to consider the question of dissemination of infection through population migration.
The first program to implement an ivermectin based elimination strategy for onchocerciasis was the Onchocerciasis Elimination Programme of the Americas (OEPA). OEPA's strategy was to provide semi-annual treatments with ivermectin at a coverage rate ≥85% of all eligible at risk individuals residing in the 13 foci of onchocerciasis in the six endemic countries in Latin America. By ensuring high coverage rates in the eligible population over a period of several years, it was believed that transmission of the parasite could be suppressed for a long enough period that the parasite population would eventually be pushed below the transmission breakpoint, and the parasite population would become locally extinct. The success of this approach is now being seen in most of the Latin American foci with Colombia being the first country to fulfill criteria for verification of elimination and others expected to follow in the next few years.
Arguably the most important lesson emerging from the success here is that high coverage with twice a year administration of ivermectin can successfully, albeit over a relatively long period of time, eliminate transmission of this infection. The fact that the main vector in this area, S. exiguum, is a very efficient insect in terms of promoting O. volvulus infection suggests that the success seen here in Ecuador may indeed be able to be translated from this relatively small focus to the more challenging locations in Africa.
The data presented here were in concordance with the guidelines established by the World Health Organization. In addition, clinical and pathological monitoring by the programme has supported the basic entomological and serological results. These findings led to the cessation of mass drug treatment with ivermectin in the Ecuadorian onchocerciasis focus, and the entry of the programme into post treatment surveillance phase. The maintenance of this status of interrupted transmission is needed for a further three years to complete successful elimination of this infection from the country [13].
|
10.1371/journal.pntd.0000940 | Ecological Modeling of Aedes aegypti (L.) Pupal Production in Rural Kamphaeng Phet, Thailand | Aedes aegypti (L.) is the primary vector of dengue, the most important arboviral infection globally. Until an effective vaccine is licensed and rigorously administered, Ae. aegypti control remains the principal tool in preventing and curtailing dengue transmission. Accurate predictions of vector populations are required to assess control methods and develop effective population reduction strategies. Ae. aegypti develops primarily in artificial water holding containers. Release recapture studies indicate that most adult Ae. aegypti do not disperse over long distances. We expect, therefore, that containers in an area of high development site density are more likely to be oviposition sites and to be more frequently used as oviposition sites than containers that are relatively isolated from other development sites. After accounting for individual container characteristics, containers more frequently used as oviposition sites are likely to produce adult mosquitoes consistently and at a higher rate. To this point, most studies of Ae. aegypti populations ignore the spatial density of larval development sites.
Pupal surveys were carried out from 2004 to 2007 in rural Kamphaeng Phet, Thailand. In total, 84,840 samples of water holding containers were used to estimate model parameters. Regression modeling was used to assess the effect of larval development site density, access to piped water, and seasonal variation on container productivity. A varying-coefficients model was employed to account for the large differences in productivity between container types. A two-part modeling structure, called a hurdle model, accounts for the large number of zeroes and overdispersion present in pupal population counts.
The number of suitable larval development sites and their density in the environment were the primary determinants of the distribution and abundance of Ae. aegypti pupae. The productivity of most container types increased significantly as habitat density increased. An ecological approach, accounting for development site density, is appropriate for predicting Ae. aegypti population levels and developing efficient vector control programs.
| Dengue infection is the leading cause of arbovirus illness worldwide with an estimated 2.5 billion people at risk. The primary dengue vector, Ae. aegypti, develops mainly in artificial containers in and around human dwellings. Often a small number of container types are responsible for a large proportion of adult mosquitoes in a region. To this point, most studies of Ae. aegypti production have failed to consider the spatial arrangement of development sites. For this investigation, mosquito populations and development sites were sampled in a spatially exhaustive manner over a four year time period in rural Kamphaeng Phet, Thailand. The data indicate that not only are some container types more productive than others, but that the local density of development sites has a large effect on the productivity of individual containers. Specifically, containers in areas of high development site density are more likely to be productive. The ecological setting and density of development sites should be incorporated in efforts to model and predict Ae. aegypti population levels. This understanding is vital in determining the feasibility of population control and the effort necessary to reduce vector populations below epidemic thresholds.
| The primary mosquito vector of dengue viruses (DENV), Aedes aegypti (L.), is well adapted to living with people and in much of the world is predominantly found among human settlements.[1], [2] Most dengue illness similarly occurs in urban and peri-urban environments, where humans are the only vertebrate host. Immature Ae. aegypti develop in artificial and natural water-holding containers located in and around human habitations. Reducing or eliminating larval habitat has been advocated as an important component of sustainable vector control programs.[3], [4], [5] Although none are currently commercially available, vaccines effective against all four DENV serotypes are reaching the final stages of development.[6], [7] In the near future, a combined dengue prevention strategy involving vaccine deployment and vector population reduction may significantly reduce the global burden of dengue illness.[8], [9]
A key to successful and sustainable vector control for dengue, whether it is done alone or with a vaccine, is a fundamental understanding of Ae. aegypti ecology that would allow predictions on their abundance through space and time.[8] A primary determinant of adult mosquito population density concerns the types and number of containers in a given environment. Adult production is unevenly distributed across potential larval development sites. In most cases, a few key types of containers are responsible for a large proportion of the pupal, and thus adult, production.[10], [11], [12] Protective measures such as lids, larvicide, removal of discarded and unused containers or biological agents have reduced adult vector population density.[2], [13] Container capacity, water temperature, source of water, and container location, all of which can vary seasonally,[14], [15], [16] have been cited as important ecological factors affecting production of adult Ae. aegypti.[12], [17] A number of studies have also found that Ae. aegypti abundance is not homogeneous among households, with disproportionate numbers of immature and adult mosquitoes clustered in key premises.[18], [19], [20] A study of Ae. aegypti production in Amercan Samoa found that containers were more productive on average in houses with a large number of containers.[21] To this point, the relationship between productivity and the spatial distribution of containers has not been rigorously examined. That is, how does the density of nearby development sites affect overall adult Ae. aegypti production?
After emergence, female mosquitoes mate and begin taking blood meals. The first gonotrophic cycle is completed several days later when eggs are oviposited into available containers. Egg development and oviposition continues, often with multiple blood meals per cycle.[22], [23] Mark-release-recapture studies have found that most adult Aedes aegypti do not disperse more than 200 m, and that many are captured within the house they were released or neighboring houses.[23], [24], [25], [26] It has also been observed that females are less likely to disperse from houses with a large number of available ovisposition sites.[27] Given that most Aedes aegypti do not disperse very far, we would expect that containers in close proximity to other productive containers are more likely to be oviposition sites and more likely to receive a large number of eggs. We hypothesize that, holding other attributes constant, containers in areas of dense larval habitat will have a greater probability of being productive and a greater abundance of pupae than areas where suitable, wet containers are rare and thus have a spatially dispersed distribution.
In this study we chose to focus on pupal production. In most environments, identification and enumeration of Ae. aegypti pupae is feasible and pupal counts can be correlated with adult Ae. aegypti density. Epidemiologically, pupae per person has been proposed as a measure of entomological risk for DENV transmission;[10] i.e., entomologic thresholds have been estimated for the minimum pupal density required to support epidemic DENV transmission.[28]
We used a hurdle regression model to test the hypothesis that density of local larval habitat is positively associated with production of Ae. aegypti pupae populations. We use presence and abundance of pupae in water holding containers as dependent variables representing pupal production. Our model was fit using 4 years of field data from rural Kamphaeng Phet, Thailand and it predicts pupal productivity at the individual container level. We used a hierarchical regression framework to account for variation in productivity among container types. This allowed us to determine differential effects of ecological factors across container types and different container densities. Many of the surveys we carried out were repeated in the same villages semiannually. By including varying intercepts by survey we could account for spatial and temporal dependence within surveys and understand the effects of repeated sampling on Ae. aegypti populations. Our overall aim was to better understand adult Ae. aegypti production at the scale of individual containers and to interpret our results in the context of targeted larval control; i.e., treatment, protection, and/or removal of the most productive containers.
Our data came from pupal surveys conducted in Kamphaeng Phet Province, Thailand (Figure 1). Sampled households were part of an epidemiologic study that included more than 8,000 households within five sub-districts in the vicinity of the provincial capital.[14] People living in the study area experience symptomatic DENV infections annually and all four serotypes have been recovered from the region.[29], [30], [31] The climate is tropical with marked rainfall seasonality (Figure 2A).[32] The area has an active vector control program, which includes larvicide application and insecticide fumigation. In some cases, larvicides are distributed as a preventative control measure. Focal fumigation and larvicide application are also initiated by local public health authorities upon identification of dengue cases in their catchment area.
Pupal surveys were carried out from 2004 to 2007 (Figure 2B).[14] Approximately 6,400 household surveys were made in 2,088 unique households. Written consent was obtained from an adult resident of each household before surveys were conducted. The study protocol and consent forms were approved by the AFRIMS Scientific Review Committee and the ethical review committees of the U.S. Army Surgeon General, Thai MoPH, and University of California, Davis. The location of each household was captured using a global positioning system (GPS) receiver. All water-holding containers in each of these household plots were examined. For each container the following attributes were recorded: container type, container dimensions, water depth, temephos, an organophosphate larvicide, status (with/without), cover status (with/without), filling method (rain/manual), location (indoor/outdoor with some shade/full sun), and fish status (with/without). All pupae were collected from each container and reared to adults in our field laboratory. The sex and species of each emerging adult was determined and totals were associated with each container. A previous study in this region indicated that fumigation can lead to significant, although short lived, reductions in Ae. aegypti adult populations.[33] Containers in households that have been fumigated within the 2 months prior to the sampling date were removed from our analysis. In total, 84,840 out of 98,862 container samples were retained for this analysis.
Two types of sampling methods were employed: cross-sectional and cluster sampling. Each of the sampling methods was spatially exhaustive, in that all occupied household plots in study area were included. Cross-sectional sampling was performed in two sub-districts: Kon Tee (16°22′ N, 99°38′ E) and Na Bo Kam (16°24′ N, 99°22′ E).[14] Within each sub-district one village with high housing density and one village with low housing density were selected for survey. Each of the four villages was sampled twice per year with first survey at the end of the dry season (late March to early May) and the second at the end of the wet season (September through early November). The number of participating households varied based on occupancy and participation, with an average of 543 houses surveyed in each of the eight samples. Households were also sampled as part of a cluster sampling methodology.[31] For these surveys, households within 100 m and including the home of a child with overt dengue illness or a non-dengue febrile illness were examined. Symptomatic febrile illness was detected using a school based surveillance program. Cluster studies took place between June and November of each year. Some households in clusters were also participants in our biannual entomological survey described above. In addition to the Kon Tee and Na Bo Kam sub-districts, cluster samples were carried out in Thep Na Korn (16°24′ N, 99°32′ E), Nakhon Chum (16°29′ N, 99°30′ E), and Nong Pling (16°32′ N, 99°30′ E) sub-districts (See Figure 1).
Our analyses focused on production of Ae. aegypti pupae. Our first dependent variable was whether or not there were any Ae. aegypti pupae in a container and is referred to as positive (POS). This variable is equal to one if at least one Ae. aegypti adult emerged from the pupae collected from a container and was otherwise equal to zero. Our second dependent variable was the total number of Ae. aegypti adults that emerged from pupae collected from a container. This variable is referred to as Ae. aegypti pupae (AEGPUP).
Containers were categorized to facilitate analysis. Container categories were based on five characteristics. The first concerned the type of container (jar, tank, tire, etc.). The remaining factors were temephos status, lid status, fill method, and location. Containers with fish were combined into one group. Classifications with fewer than 90 total containers were grouped into “Other.” A complete list of the 124 classifications used is provided in Table S1.
The PIPED variable indicates whether the container was located at a house with piped water or if water was drawn from a communal well. The samples were divided into three seasons that correspond approximately to the early season survey (EARLY), cluster surveys, and late wet season survey (LATE). The monsoon season cluster samples were our reference group. In some years cluster surveys continued into late fall. Those cluster surveys taking place in the 42nd week of the year or later are classified as LATE to preserve temporal consistency. Timing and coding of the samples is shown in Figure 2B.
Density of larval habitats in the vicinity of a container is included in the model as a function of the Euclidean distance to nearby containers and the average productivity of container types. This measure can be thought of as a spatially weighted count of the expected number of pupae positive development sites in the vicinity of an individual container. Containers within the same household were given a spatial weight of one. This weight declines exponentially until containers at households beyond 50 m have a spatial weight near zero. The spatial weighting function was chosen to reflect the relatively short dispersal distances of Ae. aegypti.[24] Maximum range was also limited by the size and shape of sampling regions. The density of nearby larval habitats (DENS) was computed with the equation:
In this equation, pj is the proportion of all containers of the same classification as container j that are Ae. aegypti pupae positive (see Table S1). The variable distij is the distance in meters between the household where container i was sampled and the household where container j was sampled. The minimum, maximum, mean, and standard deviation of the observed DENS values are 0.026, 15, 3.6, and 2 respectively.
Container classifications significantly more likely to contain Ae. aegypti pupae than the remaining containers were determined using Fisher's exact test.[34], [35] For each test, a 2×2 contingency table was examined where the rows represent whether or not a container is observed with pupae. The columns of the table are the containers of the classification under evaluation and all other containers. A one-tailed test was conducted for each container classification with experiment wise α = 0.05. A Bonferroni correction was used to account for multiple testing, which yields a per comparison α = 0.0004.[36] The R software system was used to compute the statistics.[37]
A varying coefficients negative binomial hurdle model was estimated to examine effects of habitat density and season on Ae. aegypti pupal production.[38] Units of analysis were samples of water holding containers. Hurdle models are a type of discrete mixture model used for count data with excess zeros. Excess zeros results when a dependent variable contains more zeros than would be expected for a Poisson or negative binomial distribution. If the excess zeros are ignored parameter estimates and standard errors may be biased.[39] In this implementation, a logistic regression model was used to predict the probability that an observation crosses the hurdle (has a non-zero count). The non-zero counts were modeled as a truncated negative binomial distribution, where values were restricted to be greater than or equal to 1.
Dependent variables in the model were POS and AEGPUP. Independent variables were DENS, PIPED, EARLY, and LATE. Separate intercepts and slopes for each container classification were estimated. In the case of the negative binomial model, a shape parameter was also estimated for each container classification. The shape parameter is inversely proportional to the overdispersion or extra-Poisson variation. Additionally, an intercept that varied by sample was included. This second varying intercept was included in the model to account for spatial and temporal dependence not explained by the independent variables. The logistic regression model is
In these equations, πi is the probability that pupae are observed in container i. a is an overall intercept parameter. aTj and aSk are group-level intercepts for container classifications and surveys respectively. The parameters bDj, bPj, bEj, and bLj are group-level slopes estimated for each container classification.
The truncated negative binomial portion of the hurdle model is given asµi is the expected value of the distribution and rj is the size parameter for each container classification. In this model, c, cTj, and cSk are the overall and group-level intercepts. The group-level slopes are dDj, dPj, dEj, and dLj. All four of the group-level intercept parameters were constrained to have an overall mean of zero; therefore, they represent deviations from the global intercept parameters. Non-informative prior probability distributions were assigned to each of the model parameters, indicating that no prior knowledge was incorporated into the parameter estimation.
The parameters were estimated using Bayesian Markov chain Monte Carlo (MCMC) methods available in the WINBUGS 1.4 software.[40] Convergence diagnostics and model evaluation were facilitated by the R2WinBugs package.[41] After a burn in period of 100,000 iterations, the posterior parameter distributions were sampled from the subsequent 20,000 iterations. Every 20th iteration was retained, and estimates of the median and 95% credible interval were based on 1,000 samples. Convergence was confirmed by running multiple chains and examining the potential scale reduction factor.[42] The fit of the model was examined through the use of summary measures and stochastic simulation. The concordance index, used to measure the fit of the logistic regression model, is an estimate of the probability that the predictions and outcomes are concordant, and is equivalent to the area under the receiver operating characteristic (ROC) curve.[34], [43] The stochastic simulation procedure uses the fitted model parameters to create realizations that were compared to the observed data, and is a fundamental methodology for checking model fit.[44] The distributions of pupal counts for simulated datasets were examined at the individual container level and for aggregate pupae counts by survey.
The 22 key container classifications that were significantly more likely to contain pupae than the remaining containers are summarized in Table 1. Together, these classifications make up 36% of all samples, 76% of the pupae positive samples, and 79% of the Ae. aegypti pupae collected. Container classes most likely to be positive were ant traps and jars that were manually-filled and unprotected by temephos or a lid (Table 1). Together, jars, tanks, and ant traps made up a large majority of the productive containers. Eighty percent of the sampled containers were manually-filled, and these accounted for approximately the same percentage of total pupae. Containers located in full sun were less likely to be positive than corresponding container classifications located indoors or outdoors with some shade.
The concordance index for the logistic portion of the hurdle model is 0.84, where perfect prediction of pupal presence would yield a value of 1.0 and random guessing or an intercept only model would yield a value of 0.5. One thousand realizations of pupal counts were generated using the estimated parameters of the entire hurdle model for each sampled container. Observed pupal counts were compared with the 2.5 and 97.5 percentiles of the realizations. Ninety-eight percent of the observed values were between the two percentile thresholds. All of the 76,137 observed zero counts were within the 95% simulation envelope, and 85% of the 8703 pupae positive containers were within the envelope. As pupal abundance increased container counts were less likely to fall within the envelope (Figure 3). More than 30 Ae. aegypti pupae were collected from 125 containers and all of those containers fell outside of the simulation envelope; nevertheless, most observed counts were not far above the simulation envelope. 52% of the observed counts above the envelope were within 3 pupae of the 97.5 percentile, and 99.8% of observed counts were less than the maximum simulated value (1000 realizations) for the container. At the container level, the overall model fit is adequate. The pupal counts for the super-producing containers, however, were underestimated.
Simulation values were aggregated by sample to assess model fit over time and in different villages. Aggregated simulation values are shown for cluster surveys with at least 20 containers in Figure 4. The observed number of pupae per cluster survey was above the 95% simulation envelope in 3 out of 91 included cluster surveys. In each of the repeated samples the observed sum of pupae fell within the simulation envelope (Figure 5). These simulations indicate that observed aggregate pupal counts per sample are reproduced by the estimated model.
Table 2 shows the population parameter estimates. The median of the 1000 retained samples is given to indicate the center of the posterior parameter distribution, and the precision of each estimate is indicated by the 2.5 and 97.5 percentiles. The intercept coefficients presented in Table 2 are the average intercepts for the entire population and group-level intercepts represent deviations from these values. Remaining coefficients in Table 2 are the population parameters or the weighted average of the individual container classification slope parameters. The average slope for the DENS variable is positive for both portions of the model. These values indicate that as container density increases, the probability that pupae are found in individual containers and the expected abundance of pupae counts are higher. For the logistic model, the median population parameter estimate for the DENS variable is 0.47. This indicates that holding all else equal, a change in DENS from 1.6 (one standard deviation below the mean) to 5.6 (one standard deviation above the mean) leads to a change in the probability that a container is pupae positive from 0.053 to 0.080; a 51% increase in probability of pupae presence. For some container classifications the effect is much larger, with a predicted 100% or greater increase in probability of being pupae positive with the same change in container density. For the negative binomial portion of the model, the median population parameter estimate is 0.11. This parameter value indicates that a change in the DENS variable from 1.6 to 5.6 leads to a change in the expected number of pupae in positive containers from 3.9 to 4.2 (8% increase).
Containers located in households with piped water had a lower probability of producing pupae, but there was not a significant effect on the pupae counts in positive containers. For the season variables EARLY and LATE, average effects are different for logistic and negative binomial models. In comparison to the cluster surveys during the monsoon season, containers sampled in the early part of the year were less likely to contain pupae, but the counts of positive containers are higher. The converse is true in the late season when the probability of observing pupae in containers was higher, but the pupae counts were lower on average.
The variability among group-level parameters is shown in Table 3. For both models container classification intercepts show the most variation. Distributions of group-level parameters for the logistic model are shown in Figure 6. The variations in container type intercepts represent the relative differences in probability that containers contain pupae. As expected, key containers (Table 1) have mostly large positive intercepts indicating that the probability of being positive is relatively high. The temephos-treated and lidded containers have mostly negative intercepts indicating relatively low probabilities of being positive. The group-level density parameters indicate the relative effect of local development site density on container productivity. In the logistic model, several of the largest DENS variable slopes were for temephos-treated containers. The PIPED slope parameters indicate the effect of piped water being available in a household on container productivity in the household. The overall effect of having piped water is for lower productivity, but several rain-filled container classifications had positive slopes, indicating that they were on average more productive in households with piped water. The largest positive PIPED slope parameter (0.467) was for the key classification, buckets that were unprotected, rain-filled, and outdoors. The lowest PIPED slope parameter (−0.772) was for bottles that were unprotected, manually-filled, and indoors, indicating that these were less productive in houses with piped water. Rain-filled container classes have mostly low values for EARLY season parameter, indicating that these containers are less productive during the dryer part of the year. Manually-filled key classifications make up most of the large EARLY parameters, indicating that these containers remain productive during the dry season. Many of the temephos treated classifications also have large EARLY slope parameters. There was less variation among the LATE slope parameters. This lack of variation indicates that there were no large changes in relative productivity among container types between the cluster season and the late season.
The distributions of group-level parameters for the negative binomial model are shown in Figure 7. The number of protected containers that were pupae positive and, therefore, included in the negative binomial model were relatively small. Given the small numbers of observations in each group, parameters tended to shrink towards the overall mean. Container type intercepts indicated differences in mean pupae counts among positive containers. The largest intercept (0.788) was for unprotected tanks that were rain-filled and outdoors. Positive containers with lids had relatively low pupae counts. The relationship between container group intercepts and the dispersion parameter is shown in Figure 8. Container types with higher average counts had larger variance. The data points in the upper right quadrant of the plot are different classifications of tanks, drums, and jars. These large size containers had higher average pupal counts and were more likely than other types to have unusually large pupal counts.
Variation among sample level intercepts indicates that there is spatial and temporal variation in container productivity that is not explained by the variables included in the model. The overall trend is for lower productivity over the study period (Figure 9). In most years, parameter values were higher in the early than late sample. Intercept values for the Kon Tee sub-district were generally higher than those for Na Bo Kam sub-district. There did not appear to be large and systematic differences between intercept estimates for villages with a high density of houses and villages with a low-density of houses.
Our results support the hypothesis that containers in close proximity to other larval habitats will have a greater probability of being productive and a greater abundance of pupae than areas where larval habitats are spatially dispersed. Figure 10 shows differences in the proportion of containers that were pupae positive as density increased across four container classifications. These classifications were chosen to represent very productive containers, containers that were protected, rain-filled containers, and infrequently productive containers. This density relationship is consistent and statistically significant across different container types, seasons, and locations within the study area. The use of a varying coefficients model allowed us to incorporate container characteristics, including shade, location, lids, and application of larvicide, that were collected for each container. These results do not rule out the possibility that an unobserved factor is correlated with container density and is responsible for part or all of the effect. The density of larval habitats should be incorporated into future studies of container breeding mosquitoes and the results presented here should be confirmed in other locations.
The findings also support the concept of targeted larval control. Isolation of potential oviposition sites reduced the likelihood that they would contain pupae, and reduced the average abundance of pupae found in containers. Incorporating this density effect is crucial to properly estimating the effort required to lower vector population levels below an epidemic transmission threshold. Density had a large effect on the productivity of the protected container types, indicating that effectiveness of control measures will diminish as larval habitat density increases. Further field studies and ecological studies from other regions of the world will be useful in explicitly describing the spatial extent and biological mechanisms underpinning the relationship between larval development site density and container productivity. For example, we expect that manually-filled containers used to store water for domestic use will be especially important. They sustain population levels during extended dry seasons when rain-filled containers are empty and egg survival is challenged. Careful and exhaustive control during dry periods has the potential to improve overall vector population reduction efforts during other times of the year.
Although our results are consistent with targeted larval control, the operational challenge is that the strategy must be applied thoroughly in time and space. If areas of dense larval habitat are missed or if there are breaks in intervention at critical times, prospects for a successful outcome will be reduced. However, if it can be applied properly, which is not a trivial operational feat, our analyses indicate that removal of key development sites will result in decreased productivity from treated/removed containers as well as other containers that remain nearby.
The consistently productive container types in the Kamphaeng Phet area were primarily domestic water storage containers, such as jars and tanks followed by ant traps, buckets, tires, and smaller domestic containers. This corroborates results from previous studies in Kamphaeng Phet and other regions in Thailand.[4], [11], [14] Hurdle regression model estimates indicated that container classification accounted for the largest proportion of the variation in individual container productivity. Ecological variables such as habitat density, water source, seasonal variations, and neighborhood conditions during surveys played a significant role in container productivity.
Overall productivity of containers declined during the course of the study. Repeated sampling is a likely explanation for the cross-sectional surveys, but a similar decline was observed in areas without repeated sampling. During 2004 and 2005 DENV-4 was the primary circulating virus and most of the illnesses were not severe.[31] In the following seasons (2006 and 2007), there was an increased incidence and severity of dengue illness. Use of vector control was incorporated into our data selection and modeling process, but unspecified factors such as behavior changes may have accounted for some of the spatiotemporal variation in productivity between surveys. A knowledge, attitude, and practice (KAP) survey carried out in our study area reported regional variation in the proportion of households with knowledge of dengue and application of vector control methods that could also partially explain the presence of positive containers.[45] Longitudinal, integrated studies of epidemiological, entomological, and behavioral factors would be necessary to examine changes in practice corresponding to dengue burden and the resulting effect on vector populations.
Our regression model provides reasonable estimates of container productivity at the individual container and sample level. The model did not, however, adequately predict the number of pupae emerging from a small number super-producing containers. The two-part hurdle model effectively accounted for the large number of zero pupal counts. We used a truncated negative binomial distribution to characterize the number of pupae per positive container with a separate dispersion parameter for each container classification. Larger containers had the highest means and dispersion parameters. The increased flexibility of this approach performs better than a single negative binomial model at characterizing the distribution of larval counts, but this rarely resulted in pupae estimates greater than 30 for an individual container. Dispersion parameter estimates were influenced by the large proportion of the pupae positive containers that produced a small number of pupae, and did not fully characterize the observed distribution of pupae per container. Underestimates for survey-wide pupae counts were also due to super-producing containers, but their influence and frequency declined over the course of the study period.
A closer examination of the containers mostly likely to produce large numbers of pupae illustrates the difficulties in effectively modeling the distribution of pupal counts in super producing containers. Eighty-one of the 125 container samples (65%) with more than 30 pupae were collected in untreated, un-lidded, and manually-filled jars and tanks. For these large containers, the distribution of positive containers and pupae counts are given in Table 4. Even among these large containers, only 3% of the positive containers had more than 30 pupae. The maximum number of emerging pupae collected from a single container was 298. The proportion of containers that are pupae positive, the proportion that contain more than 30 pupae and the maximum pupae counts increase with development site density. Even in the high density category, however, 88% of the positive containers have 10 or fewer pupae and 57% of the positive containers have 4 or fewer pupae. This large number of small pupal counts has a large weight in parameter estimation relative to the few very large pupae counts.
Super-producing containers tended to be relatively rare events, and may not be representative of the overall distribution of the standing crop of pupae. More importantly, the significance of super-producing containers in DENV transmission is not well documented. Further analysis is needed to examine the relationship between pupae production in general and the few very productive containers, especially with regard to the number of adults captured in nearby houses. Several field studies have implicated food limitation and larval competition as the primary regulating factors for Aedes aegypti larval development.[17], [46], [47] The frequency and pattern of water filling and removal may also play a role in the timing of pupal counts. These internal, ecological container characteristics were not captured in the current study, but should be incorporated along with local container density in ecological studies of Ae. aegypti production. Future modeling efforts should be designed to examine the likelihood of super-producing containers and better parameterize the distribution of pupal production in these habitats.
Piped water in a household had an overall negative effect on individual container productivity, but the classification level parameters varied considerably. The protective effect of piped water was strongest among manually-filled containers. A number of container classifications were more likely to produce pupae at households with piped water, primarily outdoor, rain-filled containers. Differences in water management related to household water source appear to have differential effects on container productivity. These results suggest that changes in water infrastructure may lead to changes in the relative productivity of different container types.
Population level parameters for the EARLY and LATE variables had opposite signs in the logistic and negative binomial models. The difference was driven by the relative productivity of manual and rain-filled containers. At the beginning of the wet season (EARLY), manually-filled containers were the primary producers. Most manually-filled containers were relatively large jars and tanks that when positive had higher pupae counts. During the wet season, rain-filled containers were more likely to be productive, but on average they produced fewer pupae. At the end of the wet season (LATE), almost all containers were again more likely to be productive, which further lowered overall productivity per container. For our logistic model, early season parameter estimates for larvicide treated containers (i.e., temephos) were on average high. This was likely due to the time elapsed between treatment during the previous transmission season and mosquito sampling almost one year later. Temephos packages were observed in the containers, but we suspect that their effectiveness had waned.
An improved understanding of the environmental determinants of container productivity is an important part of adaptive vector control efforts.[48] Studies are currently underway in tropical regions of the world to test the effect of container removal on Ae. aegypti populations.[49] Incorporating development site density into those efforts along with research on the ecology within larval development sites will lead to a more cohesive understanding of Ae. aegypti population dynamics and contribute to the design of increasingly effective vector interventions.
|
10.1371/journal.pgen.1000021 | The Mediator Subunit MDT-15 Confers Metabolic Adaptation to Ingested Material | In eukaryotes, RNA polymerase II (PolII) dependent gene expression requires accessory factors termed transcriptional coregulators. One coregulator that universally contributes to PolII-dependent transcription is the Mediator, a multisubunit complex that is targeted by many transcriptional regulatory factors. For example, the Caenorhabditis elegans Mediator subunit MDT-15 confers the regulatory actions of the sterol response element binding protein SBP-1 and the nuclear hormone receptor NHR-49 on fatty acid metabolism. Here, we demonstrate that MDT-15 displays a broader spectrum of activities, and that it integrates metabolic responses to materials ingested by C. elegans. Depletion of MDT-15 protein or mutation of the mdt-15 gene abrogated induction of specific detoxification genes in response to certain xenobiotics or heavy metals, rendering these animals hypersensitive to toxin exposure. Intriguingly, MDT-15 appeared to selectively affect stress responses related to ingestion, as MDT-15 functional defects did not abrogate other stress responses, e.g., thermotolerance. Together with our previous finding that MDT-15:NHR-49 regulatory complexes coordinate a sector of the fasting response, we propose a model whereby MDT-15 integrates several transcriptional regulatory pathways to monitor both the availability and quality of ingested materials, including nutrients and xenobiotic compounds.
| All organisms adapt their physiology to external input, such as altered food availability or toxic challenges. Many of these responses are driven by changes in gene transcription. In general, sequence specific DNA-binding regulatory factors are considered the specificity determinants of the transcriptional output. Here, we show that, in the roundworm Caenorhabditis elegans, one subunit of a >20 subunit, evolutionarily conserved, non-DNA binding co-factor termed Mediator, specifies a portion of the metabolic responses to a mixture of ingested material. This protein, MDT-15, is required for appropriate expression of genes that protect worms from the effects of toxic compounds and heavy metals. Our previous findings showed that the same protein also cooperates with other regulators to coordinate lipid metabolism. We suggest that MDT-15 may “route” transcriptional responses appropriate to the ingested material. This physiological scope appears broader and more sophisticated than that of any individual regulatory factor, thus coordinating systemic metabolic adaptation with ingestion. Given the evolutionary conservation of MDT-15 and the Mediator, a similar regulatory pathway may ensure health and longevity in mammals.
| Eukaryotic gene transcription requires the concerted interplay of many factors. DNA-binding factors nucleate specific regulatory complexes on individual genes, culminating in assembly of functional RNA polymerase II (PolII). These complexes also contain transcriptional cofactors that serve various functions, such as chromatin remodeling and chromatin modification. Within this complex machinery, the sequence specific regulatory factors are generally thought to be the primary determinants that specify transcriptional output in response to a certain signal.
The Mediator is a conserved multi-protein coregulatory complex that, at the minimum, serves a critical linking function between regulatory factors and the transcription initiation machinery [1]–[3]. Some Mediator subunits, such as yeast Med17 are required for essentially all PolII-driven transcription [4]. Similarly, in the nematode Caenorhabditis elegans the subunit MDT-14/RGR-1 is broadly required for early embryonic transcription [5]. However, in both organisms, some Mediator subunits are required only for expression of a restricted subset of all PolII-transcribed genes [4]–[7]. Indeed, many Mediator components influence specific physiological and/or developmental processes. For instance, mammalian MED1/TRAP220 is utilized by nuclear hormone receptors (NHRs) to implement programs such as adipogenesis (through peroxisome proliferators activated receptor γ (PPARγ) [8]) and systemic detoxification (through the pregnane-X-receptor (PXR) and the constitutive androstane receptor (CAR) [9],[10]). Likewise, the C. elegans Mediator subunits MDT-12/DPY-22, MDT-13/LET-19, and MDT-1.1/SOP-3 participate in vulva or male tail development [7], [11]–[13]. These studies raise the question of how individual components within the same regulatory complex can exert such separable effects. Also, it is unclear whether the specific functions of the coregulators are broader or narrower than those of the sequence specific regulatory factors with which they interact. Moreover, although certain Mediator subunits differentially affect related functions, the relationship between Mediator subunit utilization and a transcriptional mechanism or a physiological process is not known. Deciphering the mechanistic contributions of individual Mediator components to transcription is relevant in view of Mediator's conservation and its capacity to interact with numerous regulatory factors, thus influencing many biological processes [14].
In a previous study we found that the C. elegans Mediator subunit MDT-15 integrates expression of certain metabolic genes in NHR-49-dependent and -independent ways [15]. Others found that MDT-15 conveys regulation of fatty acid (FA) desaturases by the basic helix-loop-helix zipper protein SBP-1, the C. elegans ortholog of the mammalian sterol regulatory element binding proteins (SREBPs) that regulate FA and cholesterol metabolism [16]. Therefore, NHR-49 and SBP-1 appear to collaborate with MDT-15/MED15 to affect overlapping yet distinct sectors of metabolic genes. Hence, MDT-15 exhibits a broader spectrum of physiologic regulation than either individual regulatory factor, and could be viewed as an important node in a regulatory network that maintains metabolic homeostasis. Thus, analysis of individual Mediator components might reveal both upstream regulatory inputs and downstream regulatory mechanisms within a critical gene network.
To connect the function of MDT-15 in transcription to its precise physiologic role, we sought to more broadly define MDT-15's sphere of influence. As an initial step, we set out to globally discover new MDT-15-dependent genes in an unbiased fashion, and thus to identify previously unrecognized biological processes that lie downstream of MDT-15.
We previously showed that the C. elegans Mediator subunit MDT-15 impacts expression of select genes involved in fatty acid (FA) metabolism [15]. To identify MDT-15's physiological targets in an unbiased manner, we used expression microarrays to compare the transcriptional profiles in wild-type (N2) worms that were fed for 40 hr (i.e. allowed to synchronously develop to larval stage L4; see Materials and Methods) with bacteria harboring vectors for either control or mdt-15 RNA-interference (RNAi; reviewed in [17]). To identify statistically significant changes in gene expression following MDT-15 depletion we analyzed the raw data using “linear models for microarray data” (limma [18],[19]; see Materials and Methods). In total we found 187 genes that were downregulated, and 120 genes that were upregulated (P-value cutoff 0.05; Table S1). An alternative approach, “significance analysis of microarrays” [20], revealed similar numbers and similar sets of affected genes (data not shown). These numbers support the hypothesis that MDT-15 is essential for expression of only a subset of genes in C. elegans [15]. Given that MDT-15 and its mammalian orthologs have thus far been characterized as coactivators [15],[16],[21], we focused this study on genes with reduced mRNA levels in mdt-15(RNAi) animals.
The genes identified as downregulated in mdt-15(RNAi) animals included 12 previously identified MDT-15 target genes involved in lipid metabolism [15], and mdt-15 itself, thus validating our experimental approach (Figure S1, and Table S1). In order to verify previously unrecognized candidate MDT-15-targets, we performed quantitative real-time PCR (qPCR) analysis on total RNA from worms grown on either control RNAi or on mdt-15 RNAi bacteria (see Materials and Methods). In total we tested 85 of the 187 downregulated candidate MDT-15 targets by this method, primarily focusing on genes with orthologs in mammals, and genes predicted to participate in metabolism and detoxification (see below). We found that 63 of 85 genes (74%) were downregulated more than two-fold (Figure S1 and Table S2; note that, in this validation assay, the two-fold cut-off yields false negatives, as our limma analysis assessed significant change without a specific two-fold minimal change threshold). Thus, our microarray-based strategy was an effective tool to identify novel MDT-15 targets.
To corroborate the data obtained with RNAi depletion of endogenous MDT-15 we obtained a strain carrying a mutation in the mdt-15 gene, mdt-15(tm2182) (strain XA7702; see Materials and Methods and Figure S2A). This strain recapitulates the phenotypes evoked by mdt-15 RNAi, including shortened life span (Figure S2B), reduced fecundity (Figure S2C), clear appearance, and altered fat storage (determined by Nile Red staining [22]; data not shown). We determined by qPCR the relative mRNA levels of 97 MDT-15 targets (85 new candidate genes from the microarrays and the 12 previously known MDT-15 targets [15]) in wild-type N2 and mutant mdt-15(tm2182) worms harvested at the L4 stage. We found that 50 of 97 genes (52%) exhibited more than two-fold deregulation in mdt-15(tm2182) mutants (Figure S1 and Table S3); 47 of these 50 genes were also deregulated in mdt-15(RNAi) worms. In contrast, 28 genes deregulated more than two-fold in mdt-15(RNAi) worms were not affected to this extent in mdt-15(tm2182) mutants, although most genes were still downregulated to some extent (and only two were upregulated in mdt-15(tm2182) worms). Thus, mdt-15(tm2182) mutants recapitulate many but not all of the gene expression defects exhibited by mdt-15(RNAi) worms. Overall, the gene expression changes in mdt-15(tm2182) worms appear less severe than those presented by mdt-15(RNAi) worms, suggesting that mdt-15(tm2182) may represent a hypomorphic allele. Alternatively, as the mutation is predicted to produce a truncated MDT-15, and as mdt-15 mRNA levels appear normal in mdt-15(tm2182) worms (Table S3), it is possible that truncated MDT-15 dominantly interferes with Mediator complex function. In any case, the data support our results obtained by RNAi depletion of endogenous MDT-15 and also establish the mdt-15(tm2182) mutant as a valuable tool to study MDT-15 function.
To identify pathways and molecular functions common to the 187 genes observed by microarray analysis to be downregulated after MDT-15 depletion, we employed the DAVID gene ontology (GO) annotation tool [23]. As expected, we found a significant overrepresentation of GO-terms for functions related to lipid metabolism (Table 1). In addition, we used DAVID to query the same gene set for protein domains. We found that domains involved in lipid metabolism were overrepresented (e.g. FA-Δ9-desaturases and FA-oxidases). These results confirm the importance of MDT-15 in regulation of lipid metabolic genes.
Unexpectedly, we also discovered other enriched domains subject to MDT-15 dependence, specifically UDP-glucuronosyl/UDP-glucosyltransferases (UGTs; 15 genes of the 187 MDT-15 candidate MDT-15 dependent genes), glutathione S-transferases (GSTs; five genes), short-chain dehydrogenase/reductase SDRs (DHSs; five genes), and FAD-domains (two genes; see Table 2). Although proteins containing any of these domains may metabolize lipids, they are typically associated with systemic metabolism and clearance of endo- and xenobiotic compounds (see below). In addition to these predicted detoxification enzymes, the gene set exhibited overrepresentation of proteins containing the DUF227 domain (six genes). Drosophila melanogaster proteins harboring DUF227 domains are implicated in insecticide resistance [24], and some D. melanogaster DUF227 genes are induced by Phenobarbital [25]. In total, MDT-15 dependent genes are enriched for five protein families associated with detoxification.
In addition to the statistically overrepresented protein families, the MDT-15 dependent genes included three cytochrome P450s (CYP450s), three alcohol and aldehyde dehydrogenases, two ABC-transporters, and two acyl-CoA-synthetases (ACSs). These are all members of protein families associated with detoxification. Interestingly, the ACS proteins represent a class of lipid metabolizing enzymes, some members of which can metabolize xenobiotics [26]. Of note, we previously found that two other acs genes, acs-2 and acs-11, respond to short-term fasting in an MDT-15 dependent fashion [15],[27]. Finally, the list of MDT-15-dependent genes included four genes (a metallothionein, mtl-2; a cadmium-responsive gene, cdr-6; a predicted zinc-transporter, T18D3.3; and a predicted selenium-binding protein, Y37A1B.5) likely involved in heavy metal detoxification. In summary, 47 of 187 genes (25%) that our microarray analysis identified as significantly downregulated in mdt-15(RNAi) worms may be involved in metabolism or elimination of toxic substances (Table S1). This number is similar to the number of MDT-15 dependent genes related to energy metabolism (38 genes, i.e. 20%), suggesting that both processes compose important sectors within MDT-15's sphere of influence (Table S1).
DAVID analysis of the 120 genes upregulated in mdt-15(RNAi) worms revealed fewer enriched functions and domains (Table S4). Among these, upregulation of genes with functions/domains in energy metabolism and in protein folding/degradation may occur as compensation for (metabolic) challenges imposed by MDT-15 depletion. In addition, we noted the enrichment of CUB and saposin domains, both of which have been associated with pathogen defense [28],[29]; it may be interesting in future experiments to examine these genes in the context of MDT-15's biological role.
MDT-15 is highly expressed in the intestine [15],[30], which is an organ important for nutrient ingestion and digestion in C. elegans, and is also thought to be the principle organ for detoxification [31]. Thus, one might expect that MDT-15-targets preferentially exhibit intestinal expression. To address this notion we compared the set of genes identified by our microarrays as downregulated in mdt-15(RNAi) worms to published sets of genes exhibiting tissue specific expression in C. elegans [32],[33]. We found that many of the genes downregulated in mdt-15(RNAi) worms are intestine-selective (22%); in contrast, no MDT-15 target genes were enriched in muscle or germ line, and only 2% were pharynx-specific (Table S5). Comparison of our gene set to a C. elegans gene expression map integrating 553 individual microarray experiments [34] also revealed enrichment for intestinal genes, and depletion of germline, neuronal, and muscle specific genes (data not shown). Finally, the 120 genes induced in mdt-15(RNAi) worms also exhibited enrichment of intestine-selective genes (data not shown). This selective overrepresentation among MDT-15 targets of intestinally expressed genes fits well with the roles of MDT-15 in lipid metabolism [15],[16], and detoxification (see below).
MDT-15 is required for both basal and fasting-induced transcription of certain energy metabolism genes [15]. We reasoned that, because MDT-15 is essential for basal expression of several predicted detoxification genes (Table 3 and Tables S1,S2,S3), it might similarly be critical for toxin-activated gene expression. Consistent with the notion that C. elegans ugt, gst, dhs, and cyp genes (encoding CYP450s) are important for detoxification, several of these genes are transcriptionally induced when wild type worms are exposed to toxic chemicals (reviewed in [35]). Interestingly, this includes ugt-1 and cyp-35C1 [36] whose basal expression we found to be MDT-15 dependent (Table 3 and Tables S1,S2,S3).
To test whether MDT-15 is required to induce these detoxification genes, we fed N2 wild-type worms for 40 hr (i.e. L1–L4) with control or mdt-15 RNAi bacteria seeded on plates that contained toxic compounds (fluoranthene or β-naphthoflavone). We chose fluoranthene because it is a naturally occurring carcinogen, and β-naphthoflavone because it is a known inducer of numerous CYP450s [36],[37]. In these worms we then determined the relative mRNA levels of 80 MDT-15 targets by qPCR. We found that fluoranthene and/or β-naphthoflavone induced the mRNA levels of 21 MDT-15 target genes greater than two-fold in control(RNAi) worms (Table 3). Importantly, six of these genes exhibited reduced induction by toxins in mdt-15(RNAi) worms; for example, fluoranthene induced mRNA levels of gst-5 ten-fold in control(RNAi) worms, but only two-fold in mdt-15(RNAi) worms (Table 3). However, not all induced genes (e.g. ugt-25) exhibited MDT-15 dependent activation, suggesting that other regulatory pathways also contribute to gene induction by toxins. Moreover, when mdt-15(tm2182) mutants were exposed to fluoranthene for 40 hr (i.e. L1–L4), we observed defects in fluoranthene-induced mRNA accumulation for 13 of the 21 genes, corroborating the results obtained with RNAi (Table 4). The selectivity of MDT-15 action on these 21 detoxification genes is summarized in Figure S3, which demonstrates the weak correlation of toxin induction (particularly the response to fluoranthene) between control(RNAi) and mdt-15(RNAi) worms (panel A) and wild type and mdt-15(tm2182) worms (panel C). Taken together, these data demonstrate that MDT-15 is required not only for basal, but also for toxin-activated transcription of select detoxification genes.
Surprisingly, our microarray analysis revealed only three cyp genes to be MDT-15 dependent (of a total of 80 C. elegans cyp genes encoding CYP450s). This may be attributable, at least in part, to the low expression levels of many cyp genes. We therefore used qPCR to examine directly the expression of 43 C. elegans cyp genes, and to explore a potential role for MDT-15 in their regulation (Table S6). We found that 14 of these 43 genes required MDT-15 for basal expression in N2 L4 worms. Moreover, 15 of the 43 cyp genes were induced greater than two-fold by fluoranthene and/or β-naphthoflavone; this selectivity of cyp gene induction suggests that some cyp genes may respond to distinct toxins, or may function in roles other than detoxification. Importantly, the induction of nine of 15 toxin-responsive cyp genes was strongly reduced in mdt-15(RNAi) worms (Table S6, Figure S3B). These results show that MDT-15 is critical for both basal and activated transcription of detoxification CYP450s, thus strengthening the notion that it plays a vital role in the toxin response.
MDT-15 depletion reduces worm viability and causes larval arrest [15], raising the possibility that the reduced capability of mdt-15(RNAi) worms to express and appropriately induce detoxification genes is an indirect effect associated with arrested development and premature death. To address this issue, we determined relative mRNA levels of toxin-responsive genes in mdt-6(RNAi) worms. We focused on MDT-6 because, (i) like MDT-15, it is a Mediator subunit, (ii) it is required for stage- and gene-specific transcription in C. elegans, and (iii) mdt-6 knockdown evokes larval arrest and premature death, phenotypes reminiscent of but distinct from those of MDT-15 depletion [38],[39]. We found that, in N2 L4-stage worms, mdt-6 RNAi affected neither basal, nor fluoranthene- or β-naphthoflavone-induced expression of detoxification genes (but did reduce mRNA levels of fat-6 and fat-7, as described [15]), suggesting that MDT-6 is largely dispensable for toxin-induced gene transcription (Table S7, Figure S3A). We conclude that larval arrest and premature death alone are not sufficient to prevent the expression of detoxification genes in response to toxins.
To assess whether the defects in gene induction by toxins reflected abnormal growth of mdt-15(RNAi) worms, we employed a conditionally sterile strain, CF512 (fer-15(b26)II; fem-1(hc17)III) [40]. To assure that the fer-1 and fem-15 mutations do not compromise the detoxification per se, we first compared the toxin response in CF512 worms to that in N2 worms. We found that, at the L4 stage, CF512 worms exhibited a very similar response to fluoranthene and β-naphthoflavone as did N2 worms (Table S8). Next, we quantified the mRNA abundance of toxin-responsive MDT-15 targets in CF512 worms exposed to mdt-15 RNAi only after completion of embryonic and larval development (adult only RNAi). Under these conditions, 12 of the 20 tested genes exhibited MDT-15 dependence in the absence of fluoranthene. Furthermore, 13 genes were fluoranthene-responsive in these worms, and for four of these genes the induction was at least partially MDT-15 dependent, resembling our results obtained in N2 L4 worms (Table S9). Thus, we conclude that the defective detoxification gene expression in mdt-15(RNAi) and mdt-15(tm2182) worms reflects a direct action of MDT-15, rather than a secondary consequence of reduced viability or arrested development.
Given that MDT-15 binds to, and collaborates with NHR-49 and SBP-1 to express certain metabolic genes, it seemed conceivable that these two regulatory factors might also be responsible for expression of MDT-15 dependent detoxification genes. To test this hypothesis, we knocked down endogenous NHR-49 or SBP-1 in wild-type worms. As previously demonstrated, nhr-49(RNAi) and sbp-1(RNAi) worms exhibited reduced mRNA levels of fat-5, -6, and -7 genes (Table S10). In contrast, these worms were unaffected in their expression of the 20 tested detoxification genes, in unchallenged conditions as well as in the presence of fluoranthene or β-naphthoflavone (Table S10). Indeed, and unlike the pattern exhibited by mdt-15(RNAi) worms, the overall induction of detoxification genes by toxins strongly correlated with the induction in control(RNAi) worms (Figure S3A). This is noteworthy because sbp-1 RNAi evokes larval arrest and sterility, further strengthening the notion that developmental arrest per se does not impair detoxification. Moreover, worms carrying a mutation in the gene encoding a C. elegans NHR implicated in detoxification (nhr-8; strain AE501 [nhr-8(ok186)]; [41]) also failed to deregulate MDT-15 dependent detoxification genes in basal and xenobiotic-challenged conditions (Table S8, Figure S3C). We conclude that NHR-49, SBP-1, and NHR-8 are largely dispensable for MDT-15 dependent expression of detoxification genes, and that MDT-15 likely uses distinct regulatory factors to control expression of detoxification genes.
As MDT-15 is essential to activate certain detoxification genes, we hypothesized that mdt-15 depletion or mutation would render worms hypersensitive to toxins. To test this, we grew conditionally sterile CF512 worms on control and mdt-15 RNAi bacteria (RNAi from the L1 stage on) in the presence of various concentrations of toxins, and monitored their development and morphology until day four of adulthood. After this prolonged exposure to mdt-15 RNAi, worms were slightly thinner than control(RNAi) worms, consistent with previous results [15]. We found that fluoranthene synergized with mdt-15 RNAi, but not control RNAi, to evoke an arrest as small, scrawny adults (Figure 1A, B). In contrast, this was not the case for β-naphthoflavone (data not shown). Moreover, mdt-15(tm2182) mutants exhibited a similar adult arrest phenotype when grown on high concentrations of fluoranthene (Figure S2), whereas wild-type N2 worms showed only a mild developmental delay at the same concentrations of fluoranthene. We conclude that compromised detoxification capability as a result of reduced MDT-15 function causes growth defects in worms challenged with select toxic compounds.
Based on their molecular identity or previous studies, the MDT-15 targets cdr-6, mtl-2, T18D3.3, and Y37A1B.5 are predicted to contribute to heavy metal detoxification [42]–[45]. Interestingly, transcription of mtl-2 is induced by cadmium (Cd2+) and zinc (Zn2+), but not by copper (Cu2+) [44],[45]. Thus, we tested whether the previously uncharacterized genes Y37A1B.5 and T18D3.3 are metal-responsive also. We found that, whereas the Y37A1B.5 was not strongly induced by any of the tested metals, T18D3.3 was induced three- to four-fold by Cd2+and Zn2+, but not by Cu2+, thus establishing it as a novel metal-responsive gene (Figure 2A).
Given that two of the four metal detoxification genes are metal-responsive, MDT-15 might contribute both to basal and to metal-induced expression of these genes. This hypothesis is supported by the fact that the list of MDT-15-dependent genes (as determined by microarray analysis) overlaps in statistically significant manner with a list of Cd2+-responsive genes ([46]; Table S11). To further characterize MDT-15's possible role in metal detoxification, we fed wild-type N2 worms for 40 hr (i.e. L1–L4) with control or mdt-15 RNAi bacteria seeded on plates containing Cd2+, Zn2+ or Cu2+. We then probed for induction of the aforementioned genes, and other known heavy metal responsive genes (cdr-1, mtl-1), by qPCR. Consistent with published results, we observed mRNA accumulation of cdr-1, mtl-1, and mtl-2 in the presence of Cd2+, and of mtl-1 and cdr-1 in the presence of Zn2+ (Figure 2A). Importantly, the Cd2+- and Zn2+-dependent induction of mtl-1, mtl-2, cdr-1, and T18D3.3 was reduced in mdt-15(RNAi) worms (Figure 2A). Moreover, we found that mdt-15(tm2182) mutants were also severely defective for both basal and Cd2+-induced mtl-1, mtl-2, cdr-1, and T18D3.3 expression, and for basal Y37A1B.5 expression (Figure 2B). Finally, in conditionally sterile worms, adult-only mdt-15 RNAi resulted in reduced basal mtl-1, mtl-2, cdr-1, T18D3.3 and Y37A1B.5 mRNA levels and in reduced Cd2+-dependent accumulation of cdr-1 and T18D3.3 mRNA (Figure 2C). Together, these data demonstrate that MDT-15 plays a critical role in the transcriptional response to heavy metals in C. elegans.
In the above experiments, worms were exposed for ∼40 hr to heavy metals, corresponding to chronic conditions. In order to test an acute response to heavy metal, we grew adult worms for 48 hr on control and mdt-15 RNAi, and then challenged them with high concentrations of Cd2+ for five hr. We used a worm strain harboring a transcriptional Pmtl-2::GFP reporter, CL2122 [47]. This allowed us to rapidly assess mtl-2 promoter activity in vivo. Reminiscent of the results obtained with chronic metal exposure, Cd2+ caused induction of GFP-fluorescence. Importantly, the GFP signal was weaker in worms grown on mdt-15 RNAi, and Cd2+-dependent induction was not observed in these worms (Figure 2D). Thus, MDT-15 is apparently essential for both chronic and acute Cd2+ induced mtl-2 transcription. Moreover, MDT-15 was critical to induce mtl-2 promoter activity specifically in intestinal cells, further supporting the hypothesis that MDT-15 is particularly critical for intestinal gene expression.
We have previously reported that a specific nutritional condition, short-term fasting, relies on MDT-15 to induce select genes [15]. This raises the possibility that MDT-15 represents a general stress coregulator. However, our microarrays revealed that mdt-15(RNAi) worms upregulate known and putative heat-shock proteins (hsp-17 and hsp-16.2 are α-crystallins, T05E11.3 is related to HSP90). Thus, in contrast to the aforementioned ingestion-related stresses, MDT-15 might play a negative regulatory role in heat-induced transcription. To test this possibility we grew wild-type N2 worms for 40 hr (i.e. L1–L4) on control and mdt-15 RNAi bacteria at ambient temperature (20°C), subsequently applied heat-shock (i.e. five min, 15 min, and two hr at 35°C), and then quantified the mRNA levels of several hsp genes. At 20°C mdt-15(RNAi) worms exhibited increased mRNA levels of T05E11.3, hsp-16.2, and hsp-17 compared to control(RNAi) worms; we detected similar increases for the mRNAs of the HSPs F44E5.4 (HSP70), hsp-16.1, and hsp-16.49 (Figure 3). Heat-shock failed to affect hsp-4, hsp-17, and T05E11.3 mRNA levels, but induced F44E5.4, hsp-16.1, hsp16.2, and hsp-16.49 mRNA levels in control(RNAi) worms (Figure 3 and data not shown); in contrast, mRNA accumulation was not blocked in mdt-15(RNAi) worms. Instead, mdt-15(RNAi) worms induced these genes somewhat above the levels exhibited by control(RNAi) worms. Moreover, mdt-15(tm2182) mutants exhibited a qualitatively similar, yet milder gene expression defect (Figure S4). Thus, our data indicate that MDT-15 is dispensable for heat-shock gene activation, and that it instead suppresses expression of select hsp genes. We do not fully understand the upregulation of both basal and induced expression of these genes upon mdt-15 depletion or mutation, but it seems likely that this is an indirect effect, and that MDT-15 does not act equivalently in all stress responses. Indeed, the increased hsp expression may be an indirect consequence of the detrimental effects of mdt-15 RNAi (which evokes growth arrest and death) rather than a direct effect of deregulation of hsp gene modulators. Taken together, these results demonstrate that MDT-15 is not generally required for stress-activated gene expression, but rather may constitute an element in a regulatory system that specifically monitors availability and integrity of ingested material (see Discussion).
As mdt-15 depletion and mutation cause increased expression of heat-shock protein genes, we hypothesized that they might render animals resistant to high temperatures. To test this hypothesis we assessed the survival rate in control(RNAi) and mdt-15(RNAi) worm populations at 35°C. We found that both populations exhibited similar thermotolerance phenotypes. This was the case in N2 worms exposed to mdt-15 RNAi bacteria from L1–L4 stage (Figure 1C), and also in two-day old conditionally sterile worms exposed to mdt-15 RNAi from the L1 stage on (data not shown). Therefore, the altered heat-shock protein gene expression following mdt-15 depletion is not sufficient to increase resistance to elevated temperatures. This may be due to other gene expression defects in mdt-15(RNAi) worms. Alternatively, abnormally high heat-shock protein expression in mdt-15(RNAi) worms may not confer a marginal benefit over heat-shock protein in control(RNAi) animals.
A prevailing view concerning transcriptional regulation is that DNA-binding regulatory factors are the primary determinants of specificity and coordination within transcriptional regulatory networks. Our results support an additional mechanism in which a specific component of the Mediator complex serves such functions. Specifically, we demonstrate that the C. elegans Mediator subunit MDT-15 specifies a portion of the alternative metabolic responses to a complex mixture of ingested material.
Like most soil-dwelling animals C. elegans experiences a diversity of conditions, such as insufficient nutrient supply, potentially detrimental substances (endo- and xenobiotics), and harmful microorganisms. Toxic hydrophobic compounds and heavy metals represent a complex challenge as they can produce their adverse effects both acutely, and during chronic accumulation over prolonged exposure periods. Therefore, efficient detoxification mechanisms have evolved to ensure metabolism and elimination of harmful molecules. Here, we provide several lines of evidence that the conserved Mediator subunit MDT-15 is a key contributor towards regulation of systemic detoxification. First, even in the absence of toxic challenge, mdt-15 depletion or mutation results in downregulation of many detoxification-related genes. In addition, MDT-15 is required to induce select detoxification genes in response to hydrophobic toxins and heavy metals. Finally, depletion or mutation of MDT-15 renders worms hypersensitive to at least one toxic compound. Of note, MDT-15 affects expression of all three classes of detoxification genes, i.e. activation enzymes such as CYP450s (class I), transferase enzymes such as the UGTs and GSTs (class II), and transporters such as pgp-7 and pmp-5 (class III [48]). This highlights MDT-15's broad influence on detoxification and suggests that MDT-15 coordinates entire gene expression programs in response to individual toxic challenges. These data likely reflect a direct contribution by MDT-15, and not an indirect feature of sick/arrested mdt-15(RNAi) worms, because (i) larvally arrested mdt-6(RNAi) and sbp-1(RNAi) worms did not deregulate detoxification genes, (ii) the heat-shock response was unaffected, suggesting that stress-driven transcription can function, and (iii) numerous fluoranthene and cadmium-induced genes were also deregulated in worms exposed to mdt-15 RNAi only after the completion of larval development. Taken together, our data strongly suggest that MDT-15 critically contributes to at least two detoxification pathways in C. elegans.
MDT-15 integrates expression of metabolic genes by interacting with at least two metabolic regulatory factors, NHR-49 and SBP-1 [15],[16]. Our findings suggest that MDT-15 may similarly integrate the transcriptional response to environmental stressors such as xenobiotics and heavy metals by employing a different set of regulatory factors. Indeed, MDT-15, and not any single regulatory factor, may constitute the key determinant of the detoxification response in C. elegans. Analogous to metabolic regulation, MDT-15 may collaborate with NHR family regulatory factors to induce specific detoxification programs. Strikingly, the C. elegans genome is lavishly equipped with candidate NHRs [49]. Moreover, NHRs figure prominently in detoxification in metazoans [25], [41], [50]–[52]. In mammals, the NHRs PXR and CAR are particularly critical: by binding structurally diverse compounds [53], they induce expression of appropriate detoxification genes. Notably, both NHRs target the Mediator subunit MED1, which is necessary for induction of detoxification genes in response to drugs and/or toxins [9],[10]. Given that MDT-15 associates with NHRs [15], and modulates transcription in response to xenobiotic compounds (this study), hMED15, the human MDT-15 ortholog, may also contribute to systemic detoxification by CAR and PXR.
As the Mediator complex contributes to most PolII transcription, a simple view is that it might function through a single functional domain, or alternatively, that it may somehow operate with little specificity. Instead, the emerging view is that individual Mediator subunits appear to exhibit restricted specificity: For example, in D. melanogaster S2 cells, MED23 is specifically required for heat-induction of hsp26, whereas MED16 is critical to upregulate genes in response to LPS [6]. Similarly, serine 208 phosphorylation of yeast MED2 controls expression of select genes assuring growth under low-iron conditions [54]. More remarkably, our findings demonstrate that a single Mediator subunit appears to “route” transcription, selecting a physiological output appropriate for a given input, in this case energy metabolism in response to nutrient ingestion, or detoxification/transport in response to xenobiotic or heavy metal ingestion. To do so, MDT-15 likely interacts with a spectrum of sequence-specific regulators, as well as particular factors that directly or indirectly control polymerase activity. By incorporating this routing function, MDT-15 takes on a physiological scope that is broader and more sophisticated than any specific regulatory factor. Moreover, for an individual Mediator subunit to communicate with distinct regulatory factors and transcription components under different conditions (while being dispensable in still other circumstances), implies that it may itself receive signaling inputs that drive these distinct behaviors. Therefore, although the composition of Mediator may be similar in all cells and conditions, mere presence of a certain subunit within Mediator may not reflect its specific contributions to regulatory action. Indeed, it seems likely that many Mediator subunits may be subject to various post-translational modifications that could affect their functions. It would be interesting to assess systematically the relative mechanistic contributions of MDT-15 to Mediator's function under different physiologic conditions.
The apparent routing between detoxification and energy metabolism via MDT-15 is notable as others have found that contaminants such as cadmium affect the expression of energy metabolism genes [44],[46],[55]. Although deregulation of metabolic genes may result from perturbed signal transduction [56],[57], regulation of these two processes might conceivably have evolved in a coordinated manner. For example, responses to toxins may impose increased energy expenditure, as detoxification enzymes use ATP, NADH and other energy carriers as co-factors. Similarly, short-term fasting and long-term starvation could cause accumulation of harmful metabolic side-products that necessitate expression of detoxification genes. Accordingly, modulating the activity of a factor such as MDT-15 could ensure appropriate gene expression. Such functional regulation has been described for the mammalian PPARγ-coactivator 1 (PGC-1), which is upregulated by stimuli such as fasting, exercise, and cold exposure. Upon this induction, PGC-1 collaborates with several regulatory factors to confer a switch from reductive to oxidative metabolism [50],[58]. Similarly, detailed characterization of the network that defines MDT-15 action will provide new insights into metabolic homeostasis.
It is intriguing that many of the biological processes susceptible to MDT-15 regulation are linked to aging. We previously found that MDT-15 depletion dramatically shortens C. elegans life span, a phenotype partially attributable to reduced FA desaturation [15]. In our present study, we demonstrate a role for MDT-15 in regulation of detoxification, suggesting that compromised toxin elimination might also contribute to the short life span of these animals. Intriguingly, the expression of several genes that we identified as MDT-15 targets is affected by mutations that alter worm life span [59],[60]; moreover, stress tolerance is believed to contribute to life span extension in C. elegans [61],[62]. In summary, we speculate that MDT-15 is an integral part of a regulatory system that coordinates systemic adaptation to ingestion-related stress, thus ensuring maintenance of health and longevity.
MDT-15 affects expression of energy metabolism genes that respond to food supply [15],[27], as well as detoxification genes that respond to environmental contaminants (this study); in C. elegans, both activities are directly linked to eating. Although we cannot rule out that some toxins may be absorbed through the worm cuticle, this barrier is generally impervious [63]; moreover, metals are primarily taken up by feeding [64]. Thus, we suggest that MDT-15 is a component of an ingestion-related control system that monitors both the energy availability, as well as the integrity of the ingested material (Figure 4). Such a control system may be particularly beneficial for the soil-dwelling C. elegans, as its feeding during the growth periods may be accompanied by unavoidable co-ingestion of unfavorable substances. As worms may be unable to evade potential harm, or physically separate detrimental compounds, the need for an efficient defense system arises. We speculate that MDT-15 might be an active component of such a regulatory network. Accordingly, in fed worms MDT-15 may cooperate with SBP-1 to drive fat storage and adipogenesis [16], whereas in fasted worms MDT-15 would collaborate with NHR-49 to efficiently cope with short-term fasting [15]. In feeding worms, this same system might invoke “quality control”: hence, MDT-15 and yet unidentified regulatory factors would implement expression of proteins that appropriately metabolize and/or eliminate harmful xenobiotic substances. Moreover, MDT-15 targets such as lectins, GSTs, and lipid metabolism genes have been implicated in the response to pathogen exposure [28], suggesting that MDT-15 may also participate in host defense against microbial infection, another process resident in the worm gut [31]. In summary, we propose that MDT-15 is an essential component of a regulatory network that governs screening and routing of the ingested material to achieve its appropriate utilization, metabolism, and elimination (Figure 4).
C. elegans strains N2-Bristol (WT), CF512 [fer-15(b26)II; fem-1(hc17)III)] [40]), and CL2122 dvIs15 [mtl-2::GFP, pPD30.38 (unc-54 expression vector)] [47], and AE501 (nhr-8(ok186) [41]) were maintained as described [65]. The strain XA7702 was generated by out-crossing mdt-15(tm2182) worms (a gift from Dr. S. Mitani, Tokyo Women's Medical University) four times with wild-type N2 worms; after isolating homozygotes we verified the presence of the original genomic deletion by PCR with the following primers: 5′-aatgttgctgctcaacgtgc-3′ (forward primer), and 5′-cgatctcttccaattggtcc-3′ (reverse primer). For total mRNA isolation, worm embryos were allowed to hatch on unseeded nematode growth media (NGM)-lite plates overnight at 20°C, and then grown from L1–L4 stages (for 40 hr) on NGM-lite plates containing 25 µg/mL carbenicillin, 4 mM IPTG (Alexis 582-600), and 12.5 µg/mL tetracycline seeded with the appropriate RNAi bacteria. For validation of candidate MDT-15 targets, we grew worms as follows: control RNAi for 36 hr, control RNAi 24 hr followed by mdt-15 RNAi for 12 hr, control RNAi 12 hr followed by mdt-15 RNAi for 24 hr, or mdt-15 RNAi for 36 hr. Toxins were dissolved in DMSO and added at the following final concentrations (unless indicated otherwise): fluoranthene (Sigma F4418) at 5 µg/ml, and β-naphthoflavone (Sigma N3633) at 6 µg/ml. Heavy metals were dissolved in water and added at the following final concentrations: 3CdSO4*8H2O (Sigma C3266) at 25 µM, ZnCl2 at 100 µM, and CuSO4 at 10 µM. Life span and thermotolerance assays were performed as described [66],[67].
The mdt-15, mdt-6, nhr-49 and sbp-1 RNAi clones have been described [68],[69].
Worms were grown on NGM-lite RNAi plates seeded with E. coli strain HT115 carrying the appropriate RNAi vector. Worms were transferred onto 2% (w/v) agarose pads for microscopic examination. We captured images on a Retiga EXi Fast1394 CCD digital camera (QImaging, Burnaby, BC, Canada) attached to a Zeiss Axioplan 2 compound microscope (Zeiss Corporation, Jena, Germany), and used Openlab 4.0.2 software (Improvision, Coventry, U.K.) for image acquisition.
Starting with developmentally synchronized L1 stage larvae, N2 worms were grown on control or mdt-15 RNAi for 36 hr to yield synchronized L4 stage populations. Notably, this RNAi protocol did not result in visible death or sickness at the very stage worms were harvested for RNAi isolation. L4 larvae were harvested, washed 5 times in M9, and frozen in liquid nitrogen. RNA was isolated by the Trizol method [69], labeled with Cy3 or Cy5, and hybridized to the WUSTL C. elegans microarrays. We performed five independent biological repeats; one of the arrays was a dye swap. We extracted spot intensities with SpotReader (Niles Scientific), using the default parameters to flag bad spots. To normalize the data, we used RMA background correction and the “NormalizeWithinArrays” function of limma with the print tip loess correction [19]. For a gene to be considered in the final analysis, it had to be unflagged and have a background-subtracted intensity (mean of red and green channels) greater than 64 on at least three of the five arrays. These criteria are non-stringent, particularly the intensity criterion, and may have introduced some false positives, but enabled us to capture expression changes in genes with relatively low expression levels. We used empirical Bayes fitting in limma with multiple hypothesis correction method “BH”, and considered genes with a P-value<0.05 to be differentially expressed [18]. The microarray data have been deposited in the Gene Expression Omnibus database (http://www.ncbi.nlm.nih.gov/geo/) under the accession number GSE9720.
To determine overlaps between our set of MDT-15 dependent genes and the set of Cd2+-responsive genes [46], we proceeded as follows: As only 10,909 genes pass our stringent spot quality filter in the array analysis, we used this number to estimate the significance of the overlap. This produces a more conservative P-value than using the entire set of genes on the array. The expected fraction of genes that overlap is the fraction of MDT-15 regulated genes multiplied with the fraction of Cd2+-reponsive genes. We used the binomial distribution to calculate statistical significance [70].
Isolation, purification, and reverse transcription of C. elegans RNA have been described ([69] and http://www.ucsf.edu/krylab/Stefan-WormRNA_isolationqPCR.pdf). QPCR was performed in an ABI7300 PCR machine, and the data analyzed using the Ct method (Applied Biosystems Prism 7700 Users Bulletin No. 2 http://docs.appliedbiosystems.com/pebiodocs/04303859.pdf). Relative mRNA levels were normalized to act-1 mRNA levels. Primers for qPCR were designed using Primer3 [71]. Primers were tested on dilution series of cDNA, and analyzed for PCR efficiency [72]; primer sequences are available on request.
|
10.1371/journal.pgen.1008189 | Host-dependent nitrogen recycling as a mechanism of symbiont control in Aiptasia | The metabolic symbiosis with photosynthetic algae allows corals to thrive in the oligotrophic environments of tropical seas. Different aspects of this relationship have been investigated using the emerging model organism Aiptasia. However, many fundamental questions, such as the nature of the symbiotic relationship and the interactions of nutrients between the partners remain highly debated. Using a meta-analysis approach, we identified a core set of 731 high-confidence symbiosis-associated genes that revealed host-dependent recycling of waste ammonium and amino acid synthesis as central processes in this relationship. Subsequent validation via metabolomic analyses confirmed that symbiont-derived carbon enables host recycling of ammonium into nonessential amino acids. We propose that this provides a regulatory mechanism to control symbiont growth through a carbon-dependent negative feedback of nitrogen availability to the symbiont. The dependence of this mechanism on symbiont-derived carbon highlights the susceptibility of this symbiosis to changes in carbon translocation, as imposed by environmental stress.
| The symbiotic relationship with photosynthetic algae is key to the success of reef building corals in the nutrient poor environment of tropical waters. Extensive insight has been obtained from both physiological and “omics” level studies, yet, there are still gaps in our knowledge with respect to the metabolic interactions in this symbiotic relationship. In particular, the role of the host in nitrogen utilization and its potential link to symbiont population control still remains unclear. Using a meta-analysis approach on publicly available RNA-seq data and isotope-labeled metabolomics, we demonstrate the presence of a negative-feedback cycle in which the host uses symbiont-derived organic carbon to assimilate its own waste ammonium. This host-driven nitrogen recycling process might serve as a molecular mechanism to control symbiont densities in hospite. The dependence of this regulatory mechanism on organic carbon provided by the symbionts explains the sensitivity of this symbiotic relationship to environmental stress.
| The symbiotic relationship between photosynthetic dinoflagellates in the family Symbiodiniaceae [1] and corals is the foundation of the coral reef ecosystem. This metabolic symbiosis is thought to enable corals to thrive in the oligotrophic environment of tropical oceans by allowing efficient recycling of nitrogenous waste products in return for photosynthates from the symbionts [2]. Despite the general acceptance of this assumption, cumulative studies have raised discussions about the molecular mechanisms underlying host-symbiont metabolic interactions.
In particular, the role of nitrogen recycling from waste ammonium is still under debate. While it is generally assumed that ammonium assimilation is predominantly performed by the symbiont, some studies indicate that symbiont-growth is nitrogen limited in hospite [3–5], suggesting that the host might be able to control nitrogen availability. Moreover, it has been suggested that the host might be able to utilize organic carbon [5], in the form of glucose provided by the symbiont [6], to promote ammonium assimilation by itself, while suppressing ammonium production from deamination reactions [7]. Consequently, it has been proposed that recycling of ammonium waste by the host might serve as a mechanism to control symbiont densities [5]. Although this potential mechanism to control symbiont densities through nitrogen conservation by the host has been proposed for decades, it still remains highly contentious. Consequently, the coral research field still recognizes nitrogen recycling as a main function of the symbiont [8, 9].
To better understand these metabolomic interactions, the sea anemone Aiptasia (sensu Exaiptasia pallida) [10]—an anthozoan as corals—has emerged as a powerful model system because of the similar symbiotic relationship it forms with Symbiodiniaceae. Multiple symbiosis-centered transcriptomic studies have provided invaluable information on the interactions between Aiptasia and Symbiodiniaceae [11–13]. To generate a more concise set of high-confidence symbiosis-related genes, we adapted a meta-analysis approach, which is a statistical method developed from evidence-based medical research [14]. Because of its statistical power in integrating results from multiple sources, it has been recently applied to transcriptomic studies from both animals [15] and plants [16], and allows for the identification of high-confidence genes associated with certain biological processes.
By carrying out a meta-analysis on available symbiosis-centered RNA-seq datasets, we identified a core set of high-confidence genes and pathways involved in symbiosis establishment and maintenance. To further verify our conclusions made from expression changes of these core genes, we subsequently analyzed metabolomic profiles of symbiotic and non-symbiotic Aiptasia using 13C bicarbonate labeling. Through the integration of these two layers of omics-level information, we identified the pathways associated with nitrogen conservation in the host animal, and further revealed competition for nitrogen as a central mechanism in this relationship that is generally believed to be entirely mutualistic. Based on these findings, we propose a glucose-dependent nitrogen competition model that highlights the sensitivity of the symbiotic relationship to environmental stresses.
To carry out the meta-analysis, we collected symbiosis-centered transcriptomic data that was generated from the same clonal Aiptasia strain CC7. Since this was the strain used to sequence the genome, we expected to minimize background noise in our meta-analysis by mapping the reads from the different transcriptomic studies to the published reference gene models [12]. Based on these requirements, we identified 3 previous RNA-seq studies that met our criteria and provided 4 separate datasets, encompassing 17 biological replicates per symbiosis state (i.e., aposymbiotic and symbiotic) [11–13]. We named the four datasets after the initials of the respective first authors whose paper we obtained the data from—YL, SB, EML, and EML-36 (i.e. Yong Li [13], Sebastian Baumgarten [12], and Erik M. Lehnert [11], respectively). The meta-analysis was conducted on the expression levels of gene models that were quantified based on these data.
The initial PCA performed on samples from individual studies showed a clear separation of samples by symbiotic condition (S1 Fig). This implied that the symbiotic state was the main driver of expression changes in each of the individual studies. To further investigate the relationship between samples from different studies, we performed a principal component analysis (PCA) and a rank correlation analysis (RCA) on inter-sample normalized transcripts per million (TPM) values across all studies. Both the PCA (Fig 1A) and RCA (Fig 1B) showed clear grouping of samples by experiment rather than symbiotic state. This indicated that technical and/or experimental effects from each study exerted stronger effects on gene expression profiles than the actual symbiotic state of the animals.
Although the four datasets grouped distinctly in the PCA analysis (Fig 1), there was still a clear separation of symbiotic and aposymbiotic replicates within each of the datasets (S1 Fig). We hypothesized that this separation was due to the differential expression of core genes involved in symbiosis initiation and/or maintenance. To identify these genes, we performed four independent differential expression analyses using the exact same pipeline and parameters as described in Materials and Methods. These analyses identified between 2,398 to 11,959 differentially expressed genes (DEGs), corresponding to ~10–50% of all expressed genes in the respective studies (Table 1). Since the symbiotic state was supposed to be the main factor driving expression differences between the individuals in each study, we expected to find a great overlap between these lists of DEGs. However, the overlap was poor despite the large number of DEGs identified in the individual analyses: only 393 genes were consistently differentially expressed across all four studies. Out of these 393 genes, 166 were upregulated in symbiotic anemones in all comparisons, while 134 were found to be downregulated in symbiotic animals, relative to aposymbiotic controls (Table 1). Paradoxically, the remaining 93 genes (23.7% of all overlapped DEGs) were differentially expressed in all studies, but in different directions i.e. in some studies they were significantly upregulated while in others they were significantly downregulated.
To obtain a more robust set of core genes involved in symbiosis, we performed a meta-analysis with random effects across the four independent differential gene expression analyses (S1 Table). Using this approach, we identified 731 genes that exhibited a more consistent response to symbiosis.
To assess the robustness of these genes, we carried out a principal variance component analysis (PVCA). PVCA is an approach to partition the total variance present in the expression data by estimating the contribution of each experimental parameter (biological or technical) to the variance, and determine which of these sources are the most prominent [17]. By fitting the expression profiles and the different experimental parameters used in each study (such as feeding schedule, water source, temperature, etc., as shown in S2 Table) into the PVCA, we were able to detect correlations between expression changes and potential effect sources (Fig 2).
For the four individual studies, we found that the symbiotic state of the anemones accounted for a relatively small fraction (6.5% in raw data, 8.4% in normalized data) of the observed variance. Most of the variance was introduced by study-specific variables such as feeding frequency, days between feeding and sampling, water, light intensity, and temperature. We further noticed that a large proportion of the variance across these four datasets remained unaccountable, suggesting that technical variability, e.g. RNA extraction, library preparation and sequencing, also introduces substantial unwanted heterogeneity to gene expression profiles. When the PVCA was similarly applied to the 731 genes identified through our meta-analysis, we observed that these core genes had a significantly higher association with symbiosis. Symbiosis state accounted for 46.6% of the expression variance observed in these genes (Fig 2).
We noticed that smaller gene lists tended to have variances that were better explained by symbiosis state, exemplified by DEG_YL and DEG_EML-36 having better association with symbiosis than DEG_SB and DEG_EML (Fig 2). Thus, one could argue that the meta-analysis merely achieved better association with symbiosis as it had the fewest genes of interest. To assess this potentially confounding factor, we performed PVCAs on sets of 731 randomly picked genes from each of the DEG lists (DEG_YL, DEG_SB, DEG_EML, and DEG_EML-36). These were repeated 10,000 times (i.e., a Monte-Carlo approach). These simulations allowed us to estimate that the likelihood of our meta-analysis producing the observed 46.6% by random chance was p < 10−4 (0 of 40,000 trials had symbiosis state accounting for > 46.6% of the variance).
To assess the impact of the previously identified experiment-specific biases, we conducted Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses on the DEGs identified using the four independent differential gene expression analyses, respectively. Across the analyses of four independent experiments, 283–645 GO terms and 9–55 KEGG pathways were enriched. However, the functional overlap across all studies was poor: a large proportion of the putatively enriched terms were only identified in a single dataset (~75% in GO, and ~65% in KEGG) (S2 Fig). This finding reflected the previously observed poor overlap of differentially expressed genes across the studies and provided further evidence for the role of study-specific technical factors in driving gene expression profiles. Compared to the independent analyses, the GO and KEGG enrichment of the 731 symbiosis-associated core genes contained fewer significant GO terms (204) (S3 Table), but comparatively more significantly enriched KEGG reference pathways (31) (S4 Table).
Many of the enriched GO terms and KEGG reference pathways, as well as their associated genes, fit well with processes previously reported to be involved in symbiosis [11, 18], including symbiont recognition and the establishment of symbiosis, host tolerance of symbiont, and nutrient exchange between partners and host metabolism which are discussed separately (S1 Text). The enrichment of the KEGG nitrogen metabolism reference pathway (S3 Fig) concurs with previous studies that reported the symbiosis-induced upregulation of genes involved in glutamine synthetase / glutamine oxoglutarate aminotransferase (GS/GOGAT) cycle in Aiptasia [11, 18]. The GS/GOGAT cycle has been demonstrated to be the main pathway of ammonium assimilation in plants [19], bacteria [20] as well as in cnidarians [9, 11]. Moreover, we found that the GS/GOGAT cycle connects to several symbiosis-related processes that were previously overlooked; of these processes, pathways associated with amino acid metabolism exhibited some of the most extensive changes in response to symbiosis. These findings therefore suggested that amino acid biosynthesis pathways might play a previously undiscovered role in the maintenance of the symbiotic relationship.
Amino acid and protein metabolism represented a major symbiosis-related aspect in our meta-analysis. Nine of 31 enriched KEGG pathways (S4 Table) and 18 of 125 enriched biological process GO terms (S3 Table) were associated with amino acid and/or protein metabolism (Fig 3). A total of 97 DEGs were involved in these processes, of which 43 were upregulated in symbiotic animals. Interestingly, the DEGs involved in most of the enriched biological processes exhibited consistent expression changes (Fig 3A), i.e. the genes associated with the corresponding process were either exclusively upregulated or downregulated.
Further integration of these enriched biological processes and pathways revealed an amino acid metabolism hub in Aiptasia-Symbiodiniaceae symbiosis (Fig 4). We observed that genes catalyzing glycine/serine biosynthesis from food-derived choline were systematically downregulated in symbiotic anemones. In contrast, the genes involved in de novo serine biosynthesis from 3-phosphoglycerate, one of the glycolysis intermediates, and glutamine/glutamate metabolism were generally upregulated (Fig 4A). The resulting change in amino acid synthesis pathways suggested that symbiotic anemones utilize glucose and waste ammonium to synthesize serine and glycine, which are both main precursors for many other amino acids (S1 Text). Based on these findings, we hypothesized that the host might be using symbiont-derived glucose to assimilate waste ammonium into amino acids. To test this hypothesis, we further analyzed the metabolic profiles of anemones at different symbiotic states using 13C bicarbonate labeling, which can only be fixed by the symbiont through photosynthesis.
We first investigated metabolomes of symbiotic and aposymbiotic anemones using nuclear magnetic resonance (NMR) spectroscopy. Three metabolites in the de novo serine biosynthesis pathway were highly abundant in symbiotic Aiptasia (two of them significantly so, p < 0.05), while five out of the six intermediates in the alternative glycine/serine biosynthesis pathway using food-derived choline were significantly enriched in aposymbiotic anemones as predicted (Fig 4B, Fig 5A). However, as glucose produces multiple peaks in the 1H NMR spectrum, and most of these peaks overlap with many other potential metabolites in both symbiotic and aposymbiotic anemones, it was not possible to precisely determine glucose concentrations via NMR. Consequently, we performed 13C bicarbonate labeling experiments and compared metabolite profiles of symbiotic and aposymbiotic anemones using gas chromatography-mass spectrometry (GC-MS), in order to test if the glucose is indeed provided by the symbiont and if the downstream usage of symbiont derived organic carbon is in the host. Our experiments confirmed that symbionts provide large amounts of 13C-labeled glucose to the host (S4 Fig) and that the 13C-labeling was significantly enriched in many amino acids and their precursors in symbiotic anemones compared to aposymbiotic ones (S5 Table). Moreover, metabolite set enrichment analysis indicated that 13C was mainly enriched in several amino acid metabolism pathways (Fig 5B), which is consistent with the enrichment analysis of the 731 differentially expressed core-symbiosis genes. For the amino acids with good abundance in both symbiotic and aposymbiotic animals, we examined the proportion of 13C in each of them, respectively. Interestingly, we observed relatively stable increases (~1.5-fold) of 13C levels in symbiotic animals compared to aposymbiotic ones (Fig 5C). This constant increase suggests a single carbon source (photosynthesis-produced glucose) rather than multiple sources (glucose and symbiont-derived amino acids) involved in host amino acid biosynthesis. In the latter case, we would expect to identify more amino acid transporter genes being differentially expressed in response to symbiosis, which is not the case according to the meta-analysis. This provides further proof for symbiont derived glucose as the carbon source used for host amino acid synthesis.
Technical variation during experimentation may introduce strong bias in high throughput sequencing studies [21]. This is especially true in the study of symbiotic systems. Since such systems usually feature highly interdependent metabolic interactions, technical variations in culturing, sampling, and/or manipulation can be expected to introduce significant noise in the metabolic processes intertwined with real symbiosis-associated signals. However, this is often overlooked in transcriptomic studies, and especially so in non-model organisms. As we have noticed, the non-experimental parameters sometimes exerted stronger effects on the expression profiles than the symbiotic state in the Aiptasia transcriptomic studies. To reduce the high signal-to-noise ratio, we suggest two potential venues for differential expression studies. Firstly, future transcriptomic efforts should take extreme care to standardize all experimental conditions save for the one under study. For example, culture conditions should be identical across treatments except for the factor under study, treatments should further be performed on multiple independent batches, RNA extractions and library preparation should be carried out on all samples simultaneously. The prepared libraries should also be sequenced on the same run to further minimize technical variation. Secondly, one should not dogmatically adhere to the convention of using p = 0.05 as the cutoff for statistical significance. If a study considers one in every three genes as significantly differentially expressed, to a careful reader, the proclaimed significance of those genes is diminished. As the number of DEGs increase, the rate of type I errors would also increase, which makes the discovery of meaningful biological processes more difficult.
Consequently, meta-analyses have become a powerful approach for summarizing sequencing data from different trials in order to reduce the biases inherent to single experiments and to increase the statistical power for the identification of underlying common processes [15, 16]. By applying this approach to Aiptasia RNA-seq data, we were able to deal with the specific variances present in the individual studies, improve the precision in effect size estimation for each of the genes, and eventually identify a group of high-confidence symbiosis-associated candidates. As shown in our Monte Carlo simulations of PVCA, the genes we identified using meta-analysis exhibited significantly higher association with symbiotic state than any of the single experiment analyses. Moreover, the functional enrichment analyses of our core gene set presented more symbiosis-related GO terms and KEGG pathways, rather than the very broad terms identified from individual studies, being enriched. These terms were also enriched significantly with relatively smaller p values in meta-analysis-identified genes and assisted in the understanding of metabolic interactions between host and symbiont.
The metabolic interactions between host and symbionts have been of great interest in the study of the cnidarian-Symbiodiniaceae symbiosis [22, 23]. Among these interactions, the exchange of carbonic and/or nitrogenous compounds between the two partners is arguably a central process that has been the focus of many investigations [5, 7, 9, 24–26]. However, the connections between these two major compounds remains unclear and highly controversial. By combining a meta-analysis of transcriptomic data, metabolomics, and 13C profiling, we demonstrated a host-dependent negative feedback mechanism in the regulation of nitrogen availability to the symbionts, which is driven by symbiont-derived fixed carbon (Fig 6).
The systematic upregulation of genes involved in choline-betaine pathway highlights the heterotrophic state of aposymbiotic Aiptasia (Fig 6A). This also emphasizes the importance of regular feeding in the maintenance of aposymbiotic animals as previously stated [27]. The downregulation of choline transport in symbiotic Aiptasia indicates a decrease of the host’s demand on dietary choline (Fig 6B and 6C). Correspondingly, genes involved in the downstream conversion of choline to betaine and the production of glycine from betaine are also downregulated. The decrease of glycine caused by this downregulation is likely compensated by the metabolism of serine, which can be achieved by the observed upregulation of serine hydroxymethyltransferase (SHMT, AIPGENE4781), which catalyzes the interconversion of glycine and serine. Interestingly, our results suggest that serine is one of the key components in amino acid interconversion, as the genes involved in its de novo biosynthesis from 3-phosphoglycerate (one of the intermediates of glycolysis) were consistently upregulated. The conversion from glutamate to 2-oxoglutarate, catalyzed by the upregulated phosphoserine aminotransferase (PSAT, AIPGENE17104), may serve as the main reaction to provide amino groups for the biosynthesis of amino acids. Since 2-oxoglutarate is also one of the intermediates in the citrate acid cycle, an increase of glucose provided by the symbionts may also increase the overall activity of the cycle, hence raising the relative abundance of 2-oxoglutarate in symbiotic animals. High levels of 2-oxoglutarate have been reported to induce ammonium assimilation through glutamine synthetase / glutamate synthase cycle in bacteria [28]. Consistent with this finding, we observed upregulation of all the genes involved in this pathway for symbiotic anemones.
Metabolomic analyses of symbiotic and aposymbiotic anemones confirmed the predictions derived from our transcriptomic meta-analysis. Most of the metabolic intermediates in the de novo serine biosynthesis using symbiont-derived glucose were highly enriched in symbiotic anemones and showed increased 13C-labeling. Conversely, many of the metabolites from choline-betaine-glycine-serine conversion showed decreased abundance in symbiotic animals. Furthermore, we also identified many other amino acids with significantly increased abundance and 13C-labeling signals, suggesting that serine may serve as a metabolic intermediate for the production of other amino acids. Overall, these results highlight that symbiont-derived glucose fuels ammonium assimilation and amino acid production in the host and that serine biosynthesis acts as a main metabolic hub in symbiotic hosts (Fig 6B and 6C).
The strong shifts in host amino acid metabolic pathways induced by symbiont-provided glucose explains the interactions between nitrogen and carbon metabolism in the Aiptasia-Symbiodiniaceae symbiosis. The catabolism of glucose through pathways such as glycolysis, pentose phosphate pathway, and citric acid cycle, not only generates more energy (in forms of ATP, NADH, and NADPH), which is critical to ammonium assimilation, but also produces more intermediate metabolites that can serve as carbon backbones in many biosynthetic pathways such as amino acid synthesis. Our findings thus highlight nitrogen conservation, i.e. the host driven assimilation of waste ammonium using symbiont-derived carbon, as a central mechanism of the cnidarian-algal endosymbiosis [7]. This metabolic interaction might serve as a self-regulating mechanism for the host to control symbiont density through the regulation of nitrogen availability [5] in a carbon dependent manner. This allows for higher nitrogen availability in the early stages of infection (few symbionts translocating little carbon and requiring little nitrogen) and gradual reduction of nitrogen availability with increasing symbiont densities (many symbionts translocating more carbon and requiring more nitrogen). The strict dependence of this mechanism on symbiont-derived carbon highlights the sensitivity of this relationship to changes in carbon translocation from the symbiont to the host as imposed by environmental stresses (Fig 6D). Heat-challenged symbionts have been shown to retain significantly more carbon for their own proliferation using the increased nitrogen availability [29], while exhibiting a significant decline in light utilization efficiency [30]. This indicates that the balance of the negative-feedback system is tipped by climate change-induced heat stress, because such stress disrupts carbon translocation from the symbionts to the host while increasing the amount of nitrogen available to the symbionts. Overall, this sensitive metabolic equilibrium presents a potential molecular mechanism underlying the establishment, maintenance, and breakdown of symbiotic relationships between cnidarian hosts and Symbiodiniaceae.
To collect data for a meta-analysis, we screened for transcriptomic study that focused on cnidarian-Symbiodiniaceae symbiosis using the clonal Aiptasia strain CC7. We obtained 3 previous RNA-seq studies that met our criteria and provided 4 separate datasets [11–13]. All the datasets were generated on the same platform (Illumina HiSeq 2000). Three of the datasets contained 101 bp paired-end reads, while the last one contained 36 bp single-end reads. Samples were labeled based on the initials of the first author of published papers: YL, SB, EML, EML-36.
As all raw data from Lehnert et al. [11] was provided as a monolithic FASTQ file, a custom Python script was written to split the reads into its constituent replicates, as inferred from the FASTQ annotation lines.
To avoid biases stemming from the use of disparate bioinformatics tools in calling DEGs, data from the four datasets were processed with identical analytical pipelines.
Gene expressions were quantified (in TPM, transcripts per million) based on the published Aiptasia gene models [12] using kallisto v0.42.4 [31]. DEGs were independently identified in the four datasets using sleuth v0.28.0 [32]. Genes with corrected p values < 0.05 were considered differentially expressed.
To enable direct comparisons of gene expression values between datasets, another normalization with sleuth was carried out on all samples (n = 17 aposymbiotic and n = 17 symbiotic). Principal component analysis (PCA) and ranked correlation analysis (RCA) were carried out on these normalized expression values to assess the relationship between samples and reproducibility of these studies.
Principal variance components analysis (PVCA), a technique that was developed to estimate the extent of batch effects in microarray experiments [17], was used several times in our study. A PVCA was carried out on raw data to estimate the batch effects in the combined dataset and their possible source in the original experimental designs. Consistently, the normalized data was also assessed for the reduction of batch effects post-normalization. We also performed PVCA on normalized expression values of the DEGs identified in each independent analysis or the final meta-analysis to detect the robustness of DEG calling.
For every gene with at least two studies with significant differential expression values, a meta-analysis was performed to determine the overall effect size and associated standard error. Effect sizes from each study i (represented as wi) were calculated as the natural logarithm of its expression ratio (ln Ri), i.e. geometric means of all expression values in the aposymbiotic state divided by the geometric means of all expression values in the symbiotic state. Conveniently, this value is approximately equal to the βi value provided by sleuth. As sleuth also calculates the standard error of βi, the variance of ln Ri was not calculated via the typical approximation—instead, the variance vi was directly calculated as
vi=SEβi2∙ni
where ni represents the number of replicates in study i.
To combine the studies, a random-effects model was used. While the use of this model is somewhat discouraged for meta-analyses with few studies as it is prone to produce type I errors [33], we still opted for its use over the fixed-effects model due to the substantial inter-study variation evident in the PCAs performed previously. Also, the type I error rate could be controlled by setting a more conservative p threshold, if required.
The DerSimonian and Laird [14] method was implemented as described below. Studies with individual effect sizes mi were weighted (w*) by a combination of the between-study variation (τ2) and within-study variation (vi), according to the formula
wi*=1vi+τ2
The between-study variation (τ2) across all k studies was calculated as
τ2=max{Q−dfC,0}
where
Q=∑wi(Ti−T¯)2
C=∑wi−∑wi2∑wi
The weighted mean (m*) was calculated as
m*=∑wi*Ti∑wi*
while the standard error of the combined effect was
SE(m*)=1∑wi*
The two-tailed p-value was calculated using
p=2[1−Φ(|m*SE(m*)|)]
and then subsequently corrected for multiple hypothesis testing with the Benjamini-Hochberg-Yekutieli procedure [34, 35] using a Python script. Genes with corrected p < 0.05 were considered differentially expressed. For transparency, calculations for all equations were implemented manually in Microsoft Excel (S1 Table) following established guidelines [36].
Gene ontology (GO) and KEGG pathway enrichment analyses were both conducted on five DEG lists: one each from the four independent datasets, and one from the results of the meta-analysis.
Identification of enriched GO terms were conducted using topGO [37] by a self-developed R script (https://github.com/lyijin/topGO_pipeline). A GO term was considered enriched only when its p value was less than 0.05.
KEGG pathway enrichment analyses were performed using Fisher’s exact and subsequent multiple testing correction via false discovery rate (FDR) estimation. A KEGG pathway was deemed enriched (or depleted) only when its FDR was less than 0.05. The results of enrichment analyses were visualized using GOplot [38].
Aposymbiotic Aiptasia strain CC7 and the same strain in symbiosis with Breviolum minutum strain SSB01 (formerly Symbiodinium minutum SSB01) [1] were used for metabolic profiling. All the symbiotic and aposymbiotic anemones were maintained in the laboratory in autoclaved seawater (ASW) at 25°C in 12-hour light/12-hour dark cycle with light intensity of ~30 μmol photons m-2s-1 for over three years. Anemones were fed three times a week with freshly hatched Artemia nauplii, and water change was done on the day after feeding.
Anemones were rinsed extensively to remove any external contaminations, and starved for two days in ASW and transferred into ASW with 10 mM 13C-labelled sodium bicarbonate (Sigma-Aldrich, USA) for another two days before sampling. The four-day starvation period ensured all Artemia had been digested and consumed, hence there was no contamination from Artemia in the samples for NMR and GC-MS. The samples were drained completely on clean tissues to remove any water on surface, then snap frozen in liquid nitrogen to avoid any further metabolite changes in downstream processing.
To compare metabolomic profiles of anemones at different symbiotic states, four replicates of each state (n = 30 individuals per replicate), were processed for metabolite extraction using a previously reported methanol/chloroform method [39]. The free amino acid-containing methanol phase was dried using CentriVap Complete Vacuum Concentrators (Labconco, USA).
For NMR metabolite profiling, samples were dissolved in 600 μl of deuterated water (D2O), and vortexed vigorously for at least 30 seconds. Subsequently, 550 μL of the solution was transferred to 5 mm NMR tubes. NMR spectrum was recorded using 700 MHz AVANCE III NMR spectrometer equipped with Bruker CP TCI multinuclear CryoProbe (BrukerBioSpin, Germany). To suppress any residual HDO peak, the 1H NMR spectrum were recorded using excitation sculpting pulse sequence (zgesgp) program from Bruker pulse library. To achieve a good signal-to-noise ratio, each spectrum was recorded by collecting 512 scans with a recycle delay time of 5 seconds digitized into 64 K complex data points over a spectral width of 16 ppm. Chemical shifts were adjusted using 3-trimethylsilylpropane-1-sulfonic acid as internal chemical shift reference. Before Fourier transformations, the FID values were multiplied by an exponential function equivalent to a 0.3 Hz line broadening factor. The data was collected and quantified using Bruker Topspin 3.0 software (Bruker BioSpin, Germany), with metabolite-peak assignment using Chenomx NMR Suite v8.3, with an up-to-date reference library (Chenomx Inc., Canada).
For 13C-labelling investigation using GC-MS, dried samples were re-dissolved in 50 μl of Methoxamine (MOX) reagent (Pierce, USA) at room temperature and derivatized at 60°C for one hour. 100 μl of N,O-bis-(trimethylsilyl) trifluoroacetamide (BSTFA) was added and incubated at 60°C for further 30 min. 2 μl of the internal standard solution of fatty acid methyl ester (FAME) were then spiked in each sample and centrifuged for 5 min at 10,000 rpm. 1 μl of the derivatized solution was analyzed using single quadrupole GC-MS system (Agilent 7890 GC/5975C MSD) equipped with EI source at ionization energy of 70 eV. The temperature of the ion source and mass analyzer was set to 230°C and 150°C, respectively, and a solvent delay of 9.0 min. The mass analyzer was automatically tuned according to manufacturer’s instructions, and the scan was set from 35 to 700 with scan speed 2 scans/s. Chromatography separation was performed using DB-5MS fused silica capillary column (30m x 0.25 mm I.D., 0.25 μm film thickness; Agilent J&W Scientific, USA), chemically bonded with 5% phenyl 95% methylpolysiloxane cross-linked stationary phase. Helium was used as the carrier gas with constant flow rate of 1.0 ml min-1. The initial oven temperature was held at 8°C for 4 min, then ramped to 300°C at a rate of 6.0°C min-1, and held at 300°C for 10 min. The temperature of the GC inlet port and the transfer line to the MS source was kept at 200°C and 320°C, respectively. 1 μl of the derivatized solution of the sample was injected into split/splitless inlet using an auto sampler equipped with 10 μl syringe. The GC inlet was operated under splitless mode. Metabolites in all samples were identified using Automated Mass Spectral Deconvolution and Identification System software (AMDIS) with the NIST special database 14 (National Institute of Standards and Technology, USA). The mass isotopomer distributions (MIDs) of all compounds were detected and their 13C-labelling enrichment in symbiotic Aiptasia were investigated using MIA [40]. Pathways associated with these 13C-enriched metabolites were explored using MetaboAnalyst v3.0 [41].
|
10.1371/journal.pbio.2001663 | Late Maastrichtian pterosaurs from North Africa and mass extinction of Pterosauria at the Cretaceous-Paleogene boundary | Pterosaurs were the first vertebrates to evolve powered flight and the largest animals to ever take wing. The pterosaurs persisted for over 150 million years before disappearing at the end of the Cretaceous, but the patterns of and processes driving their extinction remain unclear. Only a single family, Azhdarchidae, is definitively known from the late Maastrichtian, suggesting a gradual decline in diversity in the Late Cretaceous, with the Cretaceous–Paleogene (K-Pg) extinction eliminating a few late-surviving species. However, this apparent pattern may simply reflect poor sampling of fossils. Here, we describe a diverse pterosaur assemblage from the late Maastrichtian of Morocco that includes not only Azhdarchidae but the youngest known Pteranodontidae and Nyctosauridae. With 3 families and at least 7 species present, the assemblage represents the most diverse known Late Cretaceous pterosaur assemblage and dramatically increases the diversity of Maastrichtian pterosaurs. At least 3 families—Pteranodontidae, Nyctosauridae, and Azhdarchidae—persisted into the late Maastrichtian. Late Maastrichtian pterosaurs show increased niche occupation relative to earlier, Santonian-Campanian faunas and successfully outcompeted birds at large sizes. These patterns suggest an abrupt mass extinction of pterosaurs at the K-Pg boundary.
| Pterosaurs were winged cousins of the dinosaurs and lived from around 200 million years ago to 66 million years ago, when the last pterosaurs disappeared during the Cretaceous-Paleogene extinction that wiped out the dinosaurs. The pterosaurs are thought to have declined in diversity before their final extinction, suggesting that gradual processes played a major role in their demise. However, pterosaur fossils are very rare, and thus, it is unclear whether pterosaurs were really low in diversity at this time or whether these patterns merely result from a paucity of fossils. We describe new pterosaur fossils from the end of the Cretaceous in Morocco, including as many as 7 species. They represent 3 different families and show a large range of variation in size and skeletal proportions, suggesting that they occupied a wide range of ecological niches.
| Pterosaurs first appear in the fossil record in the Late Triassic [1–3], tens of millions of years before birds took wing [4]. Like birds, pterosaurs were archosaurs capable of powered flight; unlike birds, they flew on membraneous wings, supported by an elongate fourth digit, and walked or climbed on all fours [2,5,6]. After appearing in the Triassic, pterosaurs radiated in the Jurassic [2,7–9], followed by a second radiation of advanced, short-tailed pterodactyloid pterosaurs in the Early Cretaceous [2,7–12]. By the mid-Cretaceous, pterosaurs had evolved aerial insectivores, carnivores, piscivores, durophages, and filter feeders [2,5,6] and exploited habitats from forests [13], lakes [12], coastal plains [14], and deserts [15,16] to shallow seas [2] and the open ocean [17]. The smallest pterosaurs had a wingspan of 50 cm or less [13,18]; the largest had wingspans of 10–11 m and weighed 200–250 kg [19], making them the largest flying animals ever to evolve.
How and why this long-lived, diverse clade became extinct remains unclear. Pterosaur diversity declined in the mid-Cretaceous [5,7,8], but at least 4 clades—Azhdarchidae [2], Nyctosauridae [2], Pteranodontidae [2], and Tapejaridae [16]—and perhaps a fifth lineage, represented by the enigmatic Piksi barbarulna [20], persist into the final 25 million years of the Cretaceous, before seeming to gradually disappear towards the end of the Cretaceous. Only a single family, Azhdarchidae, is definitively known from the Maastrichtian [2,5,21]. The youngest pteranodontids are early Campanian in age [2,18]. Nyctosaurids persisted until the Campanian at least, but the youngest nyctosaurid, “Nyctosaurus” lamegoi, lacks formation-level provenance data [22] and may be Campanian or Maastrichtian [22], making the timing of extinction uncertain [22]. When Tapejaridae became extinct is also unclear. The Santonian Bakonydraco galaczi [23] has been interpreted as a tapejarid [24], extending tapejarids into the middle Late Cretaceous [23]; the tapejarid Caiuajara dobruskii [16] could be as young as Campanian or as old as Turonian [16]. The enigmatic Piksi is late Campanian [25] in age.
Along with a decline in number of families towards the Cretaceous–Paleogene (K-Pg) boundary, pterosaurs’ species richness [8] and morphological disparity [9] are thought to decrease prior to their ultimate extinction. These patterns have been interpreted as showing a gradual decline in pterosaur diversity in the late Cretaceous [26]. If so, the K-Pg extinction may have been the final blow to a group whose extinction had long been underway and was perhaps inevitable [6].
However, the pterosaur record is highly incomplete, raising the possibility that sampling artifacts drive these patterns. Sampling effects can cause abrupt extinctions to appear gradual [27], an artifact known as the Signor-Lipps Effect: the last fossil of a lineage appears some point before its extinction. When this artifact affects many species at once, it can cause catastrophic extinctions to appear drawn out [27]. The Signor-Lipps Effect should be strongest for groups with a highly incomplete record. Pterosaurs represent an extreme case, because their thin-walled, hollow bones have low preservation potential [2]. The gradual disappearance of pterosaur families, therefore, could be a sampling artifact. Similarly, observed declines in diversity [8] and disparity [9] could be driven by changes in the quality of the fossil record [8,9], given that the number of formations preserving pterosaurs declines from the Campanian to the Maastrichtian [8]. The completeness of pterosaur fossils also decreases [28] such that the available fossils may provide less information on species richness and disparity. Furthermore, the pterosaur record is dominated by Lagerstätten [8,11,28,29]—localities with exceptional preservation. Pterosaur diversity is concentrated in these Lagerstätten, notably the Solnhofen [2], Yixian, Jiufotang [11], Romualdo [2,30,31], Crato [10], Cambridge Greensand [32,33], and Niobrara [34,35] formations [2,5,10], such that a dozen such formations account for around half of known diversity [28]. However, no Lagerstätten are known from the final 15 million years of the Cretaceous. Finally, end Cretaceous pterosaurs are primarily known from terrestrial horizons, with few occurrences in marine settings [22,36,37], which may provide an incomplete record of marine lineages.
These processes—the Signor-Lipps Effect and changes in the quality of the fossil record—may drive the apparent decline in pterosaurs. If so, improved sampling should reveal additional diversity and disparity in the latest Cretaceous. To test this hypothesis, we studied a remarkable new collection of pterosaurs from the late Maastrichtian [38,39] phosphates of the Khouribga Plateau in Morocco (Fig 1A), North Africa [40]. Here we provide a preliminary description of this fauna and explore its implications for pterosaur extinction.
The fossils described here come from the upper Maastrichtian phosphates of the Ouled Abdoun Basin, in northern Morocco. Commercial exploitation of the phosphates has uncovered large numbers of marine vertebrates [41] from the Maastrichtian and early Paleogene [41]. The Cretaceous fauna includes an extraordinary diversity of marine reptiles, including mosasaurs, plesiosaurs, and turtles [41,42], abundant and diverse bony fish [42], sharks [38], and pterosaurs [43], as well as rare dinosaurs [44,45]. Preliminary studies indicate that the fauna is the most diverse and abundant known Maastrichtian marine vertebrate assemblage.
These beds have not been formally assigned to a formation; instead, a series of beds or “Couches” are informally designated for the purposes of the mining industry (Fig 1C). Couche III is Late Cretaceous in age, and Couche I and Couche II are early Paleogene. Vertebrate biostratigraphy places Couche III in the upper Maastrichtian [38], and carbon and oxygen isotope chemostratigraphy constrain Couche III to the latest Maastrichtian, within approximately 1 Ma of the K-Pg boundary [39]. The fauna therefore provides a picture of a marine ecosystem just before the K-Pg extinction.
Until now, the pterosaur record from the assemblage comprised a single specimen, the holotype of the azhdarchid Phosphatodraco mauritanicus [43]. Over the past 3 years, we have worked with the local fossil industry to assemble a collection of pterosaurs that includes over 200 specimens, ranging from isolated bones to partial skeletons. This collection is currently the largest and most diverse collection of Maastrichtian pterosaurs in the world. Fossils primarily occur as disarticulated bones, but associated bones and, rarely, partial skeletons have also been recovered. Most come from dense, laterally extensive bonebeds in the middle of Couche III at Sidi Daoui, and a minority come from Couche III at Sidi Chennane (Fig 1B). Several specimens originate in a lower layer, about 2 m below Couche III, which is characterized by a fine, pale grey matrix and white bone. The age of these fossils is unknown, and thus, they are not described here.
The fauna comprises a minimum of 7 species, including 1 species of Pteranodontidae, 3 species of Nyctosauridae, and 3 species of Azhdarchidae.
Archosauria Cope, 1869
Pterosauria Kaup, 1834
Pterodactyloidea Plieninger, 1901
Ornithocheiroidea sensu Kellner, 2003 [46]
Pteranodontoidea sensu Kellner, 2003 [46]
Pteranodontia sensu Unwin, 2003 [47]
Pteranodontidae Marsh, 1876
urn:lsid:zoobank.org:act:E5036E72-9C55-4BD8-9315-552F572781F
Etymology. The genus derives from Tethys, in reference to the Tethys Sea, and the Latin draco, “dragon.” The species name is derived from the Latin regalis, “royal.”
Holotype. FSAC-OB 1, left humerus (Figs 2 and 3).
Horizon and locality. Middle Couche III; Sidi Daoui, Khouribga Province, Morocco.
Referred material. FSAC-OB 199 ulna (Fig 4A), FSAC-OB 200 ulna (Fig 4B); FSAC-OB 201, femur (Fig 5A), and FSAC-OB 202 femur and tibia (Fig 5B).
Diagnosis. Pteranodontid with a deltopectoral crest that is small and proximally placed, terminating just past the end of the ulnar crest; a very broad, triangular distal end of the humerus; an ectepicondyle with a prominent dorsal projection; and an entepicondyle that is enlarged and proximally extended. The ulna is proportionately short and broad, with a massively expanded proximal end.
Description. The humerus measures 231 mm long but is missing the head. The length of the humerus implies a wingspan of approximately 5 m, making Tethydraco comparable to other pteranodontids in size.
The shaft is broad proximally, narrows past the deltopectoral crest, and distally becomes a broadly expanded, triangular structure. As in Pteranodon [34] and ornithocheirids [30], the deltopectoral crest is trapezoidal in shape; the base is broad, but it tapers toward the apex. However, the deltopectoral crest is smaller and more proximally placed than in Pteranodon (S1 Fig), with the base of the crest extending slightly past the ulnar crest and the apex terminating above the ulnar crest. This feature is an autapomorphy that distinguishes Tethydraco from other Pteranodontidae.
The deltopectoral crest is massively constructed. The proximal and distal margins of the crest are thin, but then the deltopectoral crest becomes thicker towards the middle, with a broad, massive pillar of bone extending along the ventral surface of the crest from its base to its apex. The apex of the deltopectoral crest is also broadly expanded. This thickened crest is absent in Azhdarchidae or Nyctosauridae but is seen in Pteranodon [34] and Ornithocheirae [30]. The deltopectoral crest is strongly curled such that its tip actually hooks backward. This feature is not seen in nyctosaurids or azhdarchids. This feature may be present in Pteranodon but is difficult to assess given the crushing of the material, but a strong posterior curling of the deltopectoral crest is seen in ornithocheirids [30], if not to the degree seen in Tethydraco.
The deltopectoral crest differs from that of Azhdarchidae and Nyctosauridae but resembles that of other Pteranodontidae [34] and ornithocheirids [30] in being warped. The entire deltopectoral crest curls posteriorly, but this curvature is more strongly developed distally than proximally, such that the deltopectoral crest is twisted about its long axis: at its tip, the dorsal surface of the crest is angled to face distally, and the apex of the deltopectoral crest is rotated so that its long axis lies at an angle of 45° to the humerus, rather than lying parallel to the long axis as in other Late Cretaceous families such as Nyctosauridae and Azhdarchidae.
The distal half of the humeral shaft is expanded and triangular and closely resembles Pteranodon (S2 Fig). This derived feature of pteranodontids [48] is shared with nyctosaurids (see below) but is absent in azhdarchoids, in which the middle of the shaft is relatively cylindrical, and the distal one-third of the humeral shaft is strongly expanded [49]. The distal expansion is extreme, with the distal width being more than twice the width of the shaft at its narrowest point and perhaps as much as one-third the total humerus length, giving the end of the humerus a paddle-like shape. Neither Pteranodon nor any other pterosaur has a similarly extreme shape, and it appears to be an autapomorphy of Tethydraco.
The lateral condyle resembles that of Pteranodon [48]; the medial condyle is damaged. A large oval pneumatic foramen appears to lie between the distal condyles as in other pterosaurs, including Pteranodon [48].
The ectepicondyle is prominent and projects strongly laterally. This strong lateral projection is a derived feature shared with Pteranodon [48], but it is better developed in Tethydraco, an autapomorphy that helps to diagnose Tethydraco.
The supracondylar process is proximally positioned, lying about one-quarter of the way from the end of the humerus. The crest is long and narrow but relatively low, unlike the prominent flange seen in nyctosaurids.
The entepicondyle projects medially and distally. The strong distal projection of the entepicondyle is a derived feature shared with both Pteranodon and the Nyctosauridae, although it is better developed in nyctosaurids than pteranodontids. The broad medial projection of the entepicondyle is shared with Pteranodontidae to the exclusion of other pterosaurs and represents a pteranodontid feature.
The distal end of the humerus is strongly compressed. Although this condition may be exaggerated by crushing, the rest of the humerus is relatively 3-dimensional, suggesting that the end of the humerus was dorsoventrally flattened in life.
The humerus is subtriangular in distal view, with a relatively straight dorsal margin, an ornithocheiroid feature, rather than a D-shaped one as would be expected for an azhdarchid, in which the dorsal margin is strongly convex. In distal view, there is a deep olecranon fossa with a prominent ulnar tubercle. There is a pneumatic foramen below the medial condyle. On the posterior surface of the shaft is a shallow trough for the passage of the triceps brachii.
Two pteranodontid ulnae are known from the phosphates. FSAC-OB 200 (Fig 4A) is smaller but better preserved than FSAC-OB 200 (Fig 4B); it compares well with Pteranodon in shape (S3 Fig). The proximal end of the ulna in FSAC-OB 199 is extremely broad, about twice the minimum diameter of the shaft. This feature may be autapomorphic and presumably corresponds with the strong distal expansion of the humerus. The shaft is relatively short and robust as in other Pteranodontidae [34,48,50] and gradually expands in diameter distally. Distal expansion of the shaft is seen in other pteranodontids [34,48,50] but is better developed in Tethydraco, such that there is no clear demarcation between the shaft and the distal end of the ulna as seen in other taxa.
The referred femora, FSAC-OB 201 and 202, resemble those of Pteranodon (S4 Fig) [34] and are short and robust compared to azhdarchoids. The femoral head is proximally directed as in other pterosaurs and is well developed and ball-like, in contrast to nyctosaurids, in which the head is reduced [35]. A narrow neck connects the femoral head to the shaft; by contrast, the neck is distally expanded and massive in nyctosaurids [35]. The greater trochanter is large and developed as a prominent proximal prong set off from the femoral head by a deep notch, as in Pteranodon. The greater trochanter is reduced in nyctosaurids [35]. The shaft is distinctly bowed medially as in other Pteranodontidae and Nyctosauridae. The shaft is relatively constant in diameter along its length, in contrast to Nyctosauridae, in which the distal half of the shaft is expanded [35].
The tibia, FSAC-OB 202, is proportionately short and robust relative to that of Pteranodon.
Comments. Tethydraco represents both the first report of a pteranodontid from the Maastrichtian and the first known from Africa. Numerous features allow referral to Pteranodontidae. The humerus is distinguished from nyctosaurids by the deltopectoral crest, which lacks a hatchet-shaped distal expansion, and from both nyctosaurids and azhdarchids by the massive construction, curling, and warping of the deltopectoral crest. The humeral shaft differs from Azhdarchidae in that the distal half is strongly expanded. The humerus differs from both nyctosaurids and azhdarchids in having a prominent, laterally projecting ectepicondyle and a large, triangular, medially projecting entepicondyle; in these features, it resembles Pteranodon. The ulna resembles pteranodontids in being robust, with a distally expanded shaft. The femur is characteristic of pteranodontids in having a combination of a bowed shaft, a derived character of the Pteranodontidae and Nyctosauridae, with a well-developed femoral head, a prominent greater trochanter, and an unexpanded distal femoral shaft, plesiomorphies absent in nyctosaurids.
Despite the similarities, Tethydraco differs from Pteranodon in having a smaller, proximally positioned deltopectoral crest and a more prominent ectepicondyle. The shape is different as well, with the shaft being more elongate and more strongly expanded distally. The ulna has a strong expansion of the proximal end and distal expansion of the shaft. These features distinguish Tethydraco from Pteranodon. The femur is indistinguishable from Pteranodon.
Referral of the ulna to Tethydraco is made on the basis of size and the strong transverse expansion of the elbow joint, which corresponds to the expanded end of the humerus seen in the holotype. Referral of the femur is more tentative, being made on the basis of size and provenance. It is entirely possible that more than one pteranodontid existed in the assemblage but in the absence of any evidence for more than one species, it is provisionally referred to Tethydraco. Associated material is required to test this referral.
Nyctosauridae Nicholson and Lyddekker 1889
urn:lsid:zoobank.org:act:928D37AF-8C3E-4168-BF88-6531B3EC520B
Etymology. The genus name is from Alcyone of Greek mythology, who was turned into a seabird, and the species name is from the Greek elaino, “to stray or wander.”
Diagnosis. Small nyctosaur. The scapula and coracoid are subequal in length. The humerus is short and robust, with a strongly expanded proximal end; proximal pneumatic fossa and foramen are absent. The deltopectoral crest is positioned proximally, close to the head of humerus, with strong constriction at midlength producing an exaggerated hatchet shape and an acutely pointed distal prong. It has a very large, proximally positioned supracondylar process. The entepicondyle is hypertrophied and distally projecting. The antebrachium and metacarpal IV are short and robust. The femur is short and robust.
Holotype. FSAC-OB 2 (Fig 6), partial skeleton including humerus, sternum, scapulocoracoid, and femur.
Type locality. Middle Couche III; Sidi Daoui, Khouribga Province, Morocco.
Referred material. FSAC-OB 217, metacarpal IV (Fig 6D); FSAC-OB 156 mandible (Fig 7); FSAC-OB 4, partial wing including humerus, radius, ulna, parts of metacarpal IV and phalanx IV-1 (Fig 8), and additional postcrania including humeri, ulnae, radii, scapulocoracoids, and synsacra (S1 Table).
Description. The referred mandible (Fig 7) resembles other nyctosaurids, with the fused dentaries forming a long, slender, Y-shaped element in dorsal view. Teeth are absent; the occlusal margins of the beak have sharp edges for a rhamphotheca. In the anterior third of the jaw, the occlusal edges are nearly parallel, giving the beak a needle-like shape. Near the middle of the jaw, the occlusal edges diverge, and a broad, shallow trough, triangular in shape, is developed between them.
The scapulocoracoid in the holotype is fused (Fig 6E), suggesting somatic maturity [51]. The bone is boomerang shaped as in other nyctosaurids [35] and, to a lesser degree, in Pteranodon [34]. The scapula is straight with a robust acromion process. The coracoid is gently curved and bears a triangular flange below the glenoid.
The sternum (Fig 6C) bears a prominent, triangular cristospina, as in other nyctosaurids [35] and pteranodontids [34]. An oval sternocoracoid articulation is preserved and appears to have had a constriction posterior to it.
The humerus (Figs 6A, 8 and 9) has a straight shaft with a sharp inflection proximal to the deltopectoral crest, deflecting the humeral head dorsally. The head’s posteroventral face bears a deep fossa just proximal to the deltopectoral crest but lacks a pneumatic foramen. The ulnar crest is a rounded rectangular flange projecting posteriorly.
The deltopectoral crest is oriented obliquely with respect to the humeral shaft but lacks the curling and twist seen in pteranodontids [34] and Ornithocheirae [30] or the thickening of the deltopectoral crest and its apex seen in those taxa. Instead, it is a thin, flat plate that projects ventrally from the humeral shaft. Its lateral surface is gently concave, and its medial surface (Fig 9A) bears a ridge running from base to apex, the ventral pillar. Running from the ventral pillar to the anterodistal edge of the deltopectoral crest is an oblique ridge, probably a muscle scar.
The deltopectoral crest has the hatchet shape characterizing nyctosaurids [35,52–55], narrowing at midlength and expanding again distally. The deltopectoral crest is strongly constricted in Alcione, with its width at the narrowest point being about 65% of its distal width. A similar constriction is seen in Nyctosaurus nanus YPM 1182 [53] and Nyctosaurus FMNH P 25026 [35] but is more weakly developed in Nyctosaurus “bonneri” SM 11311 [52], Muzquizopteryx coahuilensis [54,55], and “N.” lamegoi [22]. The distal prong of the apex is prominent and acutely pointed, and the anterior margin of the deltopectoral crest is strongly concave, as in N. nanus [53] and FMNH P 25026 [35]. The proximal prong is a prominent, bluntly rounded process.
The supracondylar process of the ectepicondyle is shifted well proximal to the distal condyles of the humerus and is developed as a large flange, a derived feature of nyctosaurids. The entepicondyle is developed as a prominent posteroventrally projecting flange, as in pteranodontids, but differs in projecting distally well past the condyles, an autapomorphy of Alcione. The humerus’s distal end bears a small groove between the condyles that terminates in a pneumatic foramen. The distal condyles are well ossified in most specimens, implying that they are near or at maturity [51].
The antebrachium is short and robust (Figs 6B and 8), being 128% the length of the humerus in the type; the antebrachium of Nyctosaurus is 170%–184% the length of the humerus.
The femur (Fig 6F) resembles other nyctosaurids [35,55] and, to a lesser degree, pteranodontids. The femoral head is highly reduced, but the neck is expanded distally where it meets the shaft. The femoral shaft is strongly bowed medially. A small pneumatic foramen is present between the femoral neck and a distinct greater trochanter, with a larger foramen positioned between the distal condyles on the posterior surface. The femoral shaft is distally expanded and lacks distinct epicondyles, as in Nyctosaurus [35,55].
Comments. Alcione is the first nyctosaurid assignable to the late Maastrichtian and the first nyctosaurid from Africa. It is referred to Nyctosauridae based on the hatchet-shaped deltopectoral crest with a distinct ventral ridge and the lack of a warped deltopectoral crest. It differs from other nyctosaurids in numerous characters. The proximal end of the humerus is broader, the distal prong of the deltopectoral crest is acutely pointed, the antebrachium and wing metacarpal are much shorter, and the femur is unusually short and robust. All of these features show that it is a new taxon.
Within Alcione, there is a high degree of variation, especially in the shape of the deltopectoral crest (S5 Fig). The deltopectoral crest ranges from narrow, with a straight apex (Fig 7A), to broader, with a more convex apex (Fig 8). The significance of this variation is unclear. Most individuals are similar in size and, based on bone texture, appear to be mature; this suggests that this shape variation is not ontogenetic. Intraspecific variation is another possibility. A third possibility is that the current sample includes 2 or more species. This might not be surprising given that in many extant marine bird faunas, a given genus may include 3 or more co-occurring species; examples include the albatross Phoebastria, the frigatebird Fregata, the booby Sula, the gull Larus, and the tern Sterna [56], in which multiple species have overlapping ranges. Additional, associated remains are needed to test these hypotheses.
The abbreviated distal wing elements in Alcione indicate a specialized flight style. The short, robust proportions suggest reduced wingspan and increased wing loading, implying distinct flight mechanics and an ecological shift. Short wings would increase lift-induced drag at low speeds, but reduced wing areas would decrease parasite drag at high speeds [57], suggesting that Alcione may have been adapted for relatively fast flapping flight compared to other nyctosaurids. Alternatively, reductions in wingspan might represent an adaptation to underwater feeding, i.e., plunge diving of the sort practiced by gannets, tropicbirds, and kingfishers, where smaller wings would reduce drag underwater.
urn:lsid:zoobank.org:act:CBA04F2E-D7BA-47BC-A76B-1096B1FE4354
Etymology. The genus name refers to the Simurgh, a flying beast from Persian mythology. The species name is from the Latin robusta, “robust.”
Holotype. FSAC-OB 7 (Fig 10).
Locality and horizon. Middle Couche III, Sidi Daoui, Khouribga Province, Morocco.
Diagnosis. Large nyctosaurid, with a humerus that is approximately 165 mm long. The deltopectoral crest is proportionately short and broad, but the apex is strongly expanded with a strongly convex apical margin giving it a fan shape. The ventral pillar is shifted to the proximal margin of the deltopectoral crest. The humeral shaft is proportionately robust and distally expanded. The supracondylar process is hypertrophied and triangular.
Description. Simurghia is a large nyctosaurid most closely resembling “N.” lamegoi in terms of size and morphology. The humeral head and ulnar crest are broken away. The humeral shaft is robust, with its distal half expanded and subtriangular as in other Pteranodontia. The deltopectoral crest resembles “N.” lamegoi in being proportionately short and broad, but with a prominent terminal expansion and a strong proximal prong of the apex, such that the proximal margin of the deltopectoral crest is strongly convex. As in other nyctosaurids, there is a strong ridge on the ventral surface of the deltopectoral crest, the ventral pillar, terminating in a tubercle at the apex. However, the ventral pillar lies along the proximal edge of the deltopectoral crest, as in “N.” lamegoi [2,22], whereas the pillar is more distally located in Nyctosaurus, Alcione, and Barbaridactylus. A muscle scar extends proximodistally across the neck of the deltopectoral crest, again as in Alcione.
The distal condyles are not preserved, but there is an unusually large, subtriangular supracondylar process, an autapomorphy of Simurghia.
Discussion. Simurghia is referred to Nyctosauridae on the basis of the deltopectoral crest, which is hatchet shaped, with a ventral pillar, and weakly curved but not warped. It differs from other nyctosaurids in the broad fan-shaped deltopectoral crest (S6 Fig), the position of the ventral pillar along the proximal margin of the crest, and the hypertrophied supracondylar process.
Although Simurghia resembles Alcione, it is unlikely to represent an adult Alcione. Specimens referred to Alcione are all subadults or mature: bones have the dense, avascular surface texture that characterizes adult pterosaurs [51], the condyles are well ossified, the holotype scapulocoracoid is fused, and the synsacrum is fused in a referred specimen. Alcione humeri average 93 mm (n = 12) long and reach a maximum of 102 mm, versus an estimated 165 mm for Simurghia. Assuming isometric scaling, Simurghia would weigh 560% the mass of the average Alcione. Such an extreme size discrepancy exceeds what is expected for intraspecific variation or sexual dimorphism. Furthermore, there are no humeri that are intermediate in length, implying that the sample comes from 2 distinct populations of adults.
Finally, Simurghia exhibits features—large size, the very broad, fan-like deltopectoral crest, and the position of the ventral pillar along the medial edge of the deltopectoral crest—that suggest affinities with “N.” lamegoi, not Alcione.
urn:lsid:zoobank.org:act:626E3A52-74AB-401B-BCD9-8849FD4D43F0
Etymology. The genus name refers to North Africa’s Barbary Coast region and the Greek dactylo, “finger.” The species name is from the Latin grandis, “great.”
Locality and horizon. Middle Couche III, Sidi Daoui, Khouribga Province, Morocco.
Holotype. FSAC-OB 232 (Fig 11), associated skeleton including left humerus, radius and ulna, right femur, left scapulocoracoid, partial right mandible.
Referred specimens. FSAC-OB 8, 9, 10 (Fig 12), and 11, humeri.
Diagnosis. Large nyctosaurid, with a humerus that is up to 225 mm long. The humerus is slender, with the deltopectoral crest well distal to the humeral head. The deltopectoral crest is short, broad, and subrectangular, with weak constriction; warping of the deltopectoral crest is weakly developed. The humeral head has large ventral pneumatic fossa and foramen/foramina. Small pneumatic foramina are proximal to the lateral condyle. The bones of the antebrachium are slender—130% of the humerus’s length. The femur is 85% of the humerus’s length, with a slender shaft and a moderately developed greater trochanter.
Description. The type (Fig 11) preserves parts of the left and right mandible. There is a small cotyle posteriorly; ahead of this, the jaw becomes deeper and plate-like, with a thick ventral margin and a sharp occlusal margin. The dorsal margin is gently concave, indicating that the upper and lower jaws were upcurved as in other nyctosaurids.
A single cervical is preserved. It is proportionately short and broad, as in other nyctosaurids [35] and pteranodontids.
The scapulocoracoid is preserved in medial view. It resembles other nyctosaurids, being a boomerang-shaped element with the robust scapula and coracoid meeting at an angle of 60°. The 2 elements are fused, suggesting skeletal maturity [51].
The humeral head has a semicircular dorsal margin and a concave anterior ventral. The humeral shaft (Figs 11A and 12) is long and slender and sigmoidal in anterior view. The deltopectoral crest is distally placed relative to the humeral head, unlike Alcione but as in Nyctosaurus [35]. The deltopectoral crest is constricted at midlength and distally expanded to give it the characteristic hatchet shape. However, the deltopectoral crest is unusually short and broad. The distally expanded tip, which gives nyctosaurids the distinctive hatchet-shaped crest, is weakly developed, a primitive feature. In anterior view, the crest is slightly warped. This feature is very weakly developed compared to pteranodontoids such as Pteranodon but better developed than in other nyctosaurids. Ventrally, there is a prominent ventral pillar running from the apex down the shaft. Anterior to this is a muscle scar, running obliquely to the distal prong of the deltopectoral crest.
The ventral surface of the humeral heads bears a large pneumatic fossa, with either a foramen or several small foramina. The presence, size, and position of this foramen are unique to Barbaridactylus among nyctosaurids. The ulnar crest is well developed and subtriangular in shape; it projects ventrally.
Distally, there is an enlarged supracondylar process as in other nyctosaurids. There is a depression between the medial and lateral condyles, with a pneumatic foramen inside the depression, beneath the lateral condyle. A pair of small, elongate pneumatic foramina are present proximal to the lateral condyle, which appear to be unique to Barbaridactylus among nyctosaurs.
The antebrachium resembles that of Nyctosaurus [35]. The ulna is relatively slender, in contrast to the robust ulna of Alcione, and weakly expanded at either end. The radius is about two-thirds of the diameter of the ulna.
The femur resembles other nyctosaurids and pteranodontids [35] in having a sigmoidal shaft with a strong dorsal projection of the humeral head and a weakly developed greater trochanter. In contrast to Nyctosaurus [35] and Alcione, the femur lacks the strong distal expansion of the shaft. The end of the shaft is more gently expanded, as in Pteranodon [35].
Comments. Barbaridactylus is referred to Nyctosauridae on the basis of the deltopectoral crest, which is hatchet-shaped, with a ventral pillar, and weakly curved rather than warped in end view. It is distinguished from other nyctosaurids by its large size, long and slender humerus, quadrangular deltopectoral crest, and foramen/foramina on the anterior surface of the humerus below the deltopectoral crest. The pneumatic foramen is variable in Barbaridactylus, and it differs in size and morphology in all the individuals studied; in some individuals, it is developed as a single foramen, and in others, it is developed as a pair of foramina. Nevertheless, this feature is seen, where exposed, in all specimens referred to Barbaridactylus and is absent in other nyctosaur material.
Barbaridactylus resembles “N.” lamegoi from the Campanian-Maastrichtian of Brazil [22] in terms of size and the broad deltopectoral crest and the proximally shifted ventral pillar. These affinities are supported by phylogenetic analysis (see below). However, it lacks the strongly pointed proximal prong of the deltopectoral crest seen in “N.” lamegoi, indicating that the 2 are distinct.
Azhdarchoidea Kellner 2003 [46]
Neoazhdarchia Kellner 2003 [46]
Azhdarchidae Nessov 1984
Referred material. FSAC-OB 12 (Fig 13), cervical vertebra C5, and FSAC-OB 13, cervical vertebra.
Description. The centrum of the referred cervical vertebra (Fig 13) measures 190 mm, and the maximum length of the vertebra is 210 mm. These measurements closely compare with cervical 6 of P. mauritanicus, which measures 196 mm and 225 mm, respectively [43].
The cervical is long and slender, as typical of Azhdarchidae [21]. The length of the vertebra is approximately 400% the width across the prezygapophyses, matching the proportions of cervical 6 of the holotype of Phosphatodraco, OCP DEK/GE 111 [43]. The centrum lacks lateral pneumatic foramina, as in other azhdarchids. Prominent postexapophyses project ventrolaterally.
As in other Azhdarchidae, a hypapophysis projects anteriorly beneath the cotyle [21]. The anterior cotyle is flanked by grooves beneath the prezygapophyses, implying that the cervical ribs had not fused to the centrum to form transverse foramina. The animal may have been near maturity but was not skeletally mature.
The neural arch is reduced and confluent with the body of the centrum. The prezygapophyses project anterodorsally, while the postzygapophyses project almost laterally. The neural spine is reduced as in other azhdarchids [21]; it forms small anterior and posterior blades, but between the 2 blades, the neural spine is not developed.
Comments. The azhdharchid P. mauritanicus has previously been described from the phosphates [43]. The vertebra described here is consistent with Phosphatodraco in size and proportions, supporting referral to that taxon.
Material. FSAC-OB 14, cervical vertebra C5 (Fig 14).
Description. The cervical vertebra comes from a small azhdarchid; the centrum measures 153 mm. However, the surface bone lacks vascularization, and the cervical ribs are fused to the centrum. These features indicate that despite the animal’s small size, it was somatically mature [51].
The centrum is typical of Azhdarchidae in being elongate and lacking pneumatic foramina piercing the lateral surfaces. The elongate proportions of the vertebra indicate that it comes from the middle of the neck. Based on comparisons with Quetzalcoatlus [58], the vertebra is probably C4 or C5, most likely C5 given the similarities in shape. Centrum length is 440% of the width across the prezygapophyses, more elongate than the C5 of Phosphatodraco [43], where length is 356% of the prezygapophyseal width.
The centrum differs from Phosphatodraco in being broad anteriorly but very narrow posteriorly; however, this feature is typical of Quetzalcoatlus [59]. Posteriorly, the centrum bears a deep ventral depression; a similar depression is present but much more weakly developed in Phosphatodraco. Lateral to the depression, a pair of prominent, subtriangular postexapophyses project laterally. Those of Phosphatodraco are more weakly developed.
As in other Azhdarchidae, the neural arch is confluent with the centrum, with no clear separation between them. Anteriorly, the prezygapophyses are long and narrow, while those of Phosphatodraco are shorter and more ovoid. Two laminae extend from the prezygapophyses to the neural spine, forming a “V.” A similar feature is seen in Phosphatodraco, but here the ‘V’ extends further back so that the dorsal surface of the centrum is broadly exposed, which does not occur in Phosphatodraco. A pair of faint ridges extend back from the prezygapophyses onto the dorsal surface of the vertebra. Postzygapophyses are laterally projecting and shifted anteriorly.
The neural spine is highly reduced as in other azhdarchids. Anteriorly, it forms a low ridge, which then becomes shallower posteriorly until near the middle of the centrum it forms a faint line running along the dorsal surface of the centrum.
Comments. The new azhdarchid is smaller than Phosphatodraco but exhibits fusion between the cervical ribs and centrum and an avascular surface bone texture. This suggests that it is mature [51] and not a juvenile Phosphatodraco. The new azhdarchid also differs from P. mauritanicus in the strong tapering of the centrum posteriorly, the deep ventral fossa on the posterior end of the centrum, and the larger postexapophyses. These are derived features that are shared with Quetzalcoatlus [59] to the exclusion of Phosphatodraco and suggest affinities with that genus. Phylogenetic analysis places this azhdarchid as sister to a clade including Zhejiangopterus linhaiensis, Arambourgiania philadelphiae, Hatzegopteryx thambena, and Quetzalcoatlus spp.
Material. FSAC-OB 203, left ulna missing proximal end (Fig 15).
Description. The ulna comes from a very large pterosaur. The preserved part of the bone, comprising the distal end and the middle of the shaft, measures 362 mm in length and may have been 600–700 mm long when complete. The shaft is 40 mm in diameter at its narrowest and 65 mm at its distal end. These proportions suggest a wingspan approaching 9 m.
Overall, the shaft is proportionately long and slender. The shaft is broad proximally, narrows distally, and then gradually expands towards its distal end. The distal end bears a broad tubercle, as in Azhdarcho [21]. The ventral margin bears a long, low flange, again as in Azhdarcho. There is an articular surface dorsal to the tubercle, separated by a notch.
Comments. The broad tubercle and low, proximodistally elongate ventral crest support azhdarchid affinities; the tubercle of Pteranodontidae is smaller, the ventral crest is much more pronounced, and the ulna is proportionately shorter and broader.
Affinities with either Phosphatodraco or aff. Quetzalcoatlus appear unlikely given that the bone texture of both indicates that they are somatically mature. Affinities with or referral to the giant azhdarchid A. philadelphiae seem more likely given that both are known from late Maastrichtian deposits of the Tethys sea [37], but more complete material is needed to test this assignment. The animal approached Quetzalcoatlus in size, but it was much more lightly built and presumably weighed much less. These proportions presumably indicate a distinct flight mode and ecological niche, suggesting that giant pterosaurs occupied a range of niches.
We undertook a morphological phylogenetic analysis of the new Moroccan species and all diagnostic Late Cretaceous species based on a previous character-taxon matrix [60] (Fig 16). Curves of taxic diversity (raw species counts over time) and phylogenetic diversity (species plus ghost lineages) were generated from the results, with minimum divergence times used (Fig 17).
Tethydraco is recovered as a pteranodontid while Alcione, Simurghia, and Barbaridactylus are recovered as nyctosaurids. Phosphatodraco and the new small azhdarchid are recovered as different lineages within Azhdarchidae.
In the current analysis, 3 or perhaps 4 other lineages extend into the final 25 million years of the Cretaceous. These include 2 lineages of Tapejaromorpha, 1 represented by Caiuajara and 1 represented by Bakonydraco, and a third lineage giving rise to Piksi. The Cenomanian Alanqa saharica and the Campanian-Maastrichtian Aerotitan sudamerica are recovered as thalassodromids, which would imply survival of thalassodromids into the latest Cretaceous. Both are known from very fragmentary material, and thus, these results are not well supported and will require testing by the recovery of more complete remains.
To test the hypothesis of a Late Cretaceous decline in pterosaur diversity, we assessed pterosaur functional diversity in the final 20 Ma of the Cretaceous. In contrast to disparity, which measures the range of morphologies regardless of function, functional diversity uses functional characters or correlates to quantify the diversity of function [61,62], e.g., inferred diet, locomotor style, or habitat. Functional correlates used here include morphological features such as size, jaw shape, or limb proportions that are likely to correlate with or influence function. Based on analogy with living birds, jaw shape likely reflects diet, wing proportions will reflect locomotion, and body size is likely to be associated with a wide range of variables including diet, locomotion, physiology, and life history. Given the incompleteness of the fossils and the difficulty of inferring function, functional diversity cannot capture the full range of ecological niches but can act as a proxy for niche occupation.
Functional diversity has advantages over taxon counting. First, it should provide a more accurate picture of the range of ecological niches and ecosystem structure than simply counting species [63], because species richness need not correlate with niche occupation. Second, because functional diversity measures dissimilarity among fossils, rather than using the number of taxonomic divisions as a proxy for diversity, it is less affected by taxonomic lumping or splitting. Furthermore, because many species can occupy the same niche, one need not find all or even most taxa to provide a relatively complete picture of ecosystem structure; one need only find representatives of each niche. It follows that functional diversity should be robust to sampling versus taxon counting, which is particularly important with poorly studied groups such as pterosaurs.
To quantify functional diversity, we created a matrix combining discrete and continuous characters to capture ecological and functionally significant morphological variation. Characters include habitat (continental, brackish, or marine), size (wingspan), and morphological characters including jaw curvature, cervical elongation, hindlimb elongation, ulna:humerus ratio, and metacarpal IV:humerus ratio. Wingspan and limb ratios were treated as continuous characters, whereas others were treated as discrete. Functional morphospaces can effectively characterize functional diversity with as few as 4 characters [61]; thus, the use of relatively few characters is appropriate here. Data were taken from fossils described in this study, previous studies of pterosaur size [18], and the literature (S2 Data). Some characters are conserved within clades, e.g., azhdarchids all have elongate cervicals, and thus, based on phylogenetic inference, it is possible to code them when the character is not preserved.
Functional diversity was quantified using principal coordinates analysis (PCoA). PCoA is used instead of principal component analysis (PCA) because unlike PCA, it can use both discrete and continuous characters, and it can accommodate missing data. PCoA rotates and replots data points along a series of axes that summarize the dissimilarity of the dataset. Gower’s distance was used to calculate the distance between points. Calculations were done in PAST [64].
Pterosaurs were assigned to 2 time bins: Santonian-Campanian and Maastrichtian. This was done because the Campanian has relatively few marine pterosaurs, which might artificially depress diversity. The 2 bins span unequal intervals (6.1 Ma versus 14.2 Ma) [65]. Because the Santonian-Campanian interval encompasses over twice the amount of time as the Maastrichtian, time averaging across this interval should artificially increase diversity and should therefore bias the result in favor of higher pre-Maastrichtian diversity.
Our results suggest that despite this, Maastrichtian niche occupation is comparable to or higher than that of the Campanian-Santonian interval (Fig 18). This is true whether functional diversity is quantified in terms of the product of ranges (Maastrichtian = 0.01545, Campanian = 0.01141), proportional to the volume of functional space occupied, or sum of ranges (Maastrichtian = 4.0078; Campanian = 3.7993), proportional to the total spread along the various functional axes. Most of the Santonian-Campanian functional space is also occupied by the Maastrichtian taxa, but new functional space is occupied in the Maastrichtian by giant Azhdarchidae such as Quetzalcoatlus and Hatzegopteryx in continental ecosystems and Arambourgiania and the Sidi Chennane giant in marine environments, driving an increase in functional occupation.
To test whether the higher diversity of the Maastrichtian is driven by sampling, we rarefied the data, randomly resampling without replacement from the set of occurrences 5,000 times at varying levels of sampling using a custom script in R and then calculating the average functional diversity and 95% confidence intervals at various levels of sampling (Fig 19). Both Maastrichtian and Santonian-Campanian functional diversity increase rapidly with sampling and fail to asymptote. This, perhaps unsurprisingly, implies that Late Cretaceous pterosaur diversity remains undersampled and that further sampling is likely to reveal that pterosaurs occupied a wider range of niches than currently known. Sampling effects are extreme in both intervals given the high diversity of pterosaurs and poor sampling (Fig 19B); the 95% confidence intervals estimated strongly overlap. Our analyses therefore show that there is no support for a decline in pterosaur diversity from the Santonian-Campanian into the Maastrichtian. However, given the strong sampling effects seen, the increased diversity of the Maastrichtian could represent a sampling effect. Further data are needed to test the hypothesis of a Maastrichtian diversity increase. However, the Maastrichtian appears to be more poorly sampled than the Santonian-Campanian as shown by the steeper slope of the sampling curve towards the right end of the graph. If so, increased sampling may increase rather than decrease the difference between the 2 intervals.
The new pterosaurs from Morocco, with previously described pterosaurs, indicate that diversity was high and niche occupation may have increased in the late Maastrichtian. Pterosaurs were a diverse and important part of Cretaceous ecosystems up to the K-Pg boundary, consistent with a catastrophic extinction driven by the Chicxulub impact. Three families, Pteranodontidae, Nyctosauridae, and Azhdarchidae, are now known from the late Maastrichtian (Fig 16). Nyctosauridae not only survived into the late Maastrichtian but did so at high diversity. Other evidence for Maastrichtian nyctosaur diversity comes from an isolated femur from the early-mid(?) Maastrichtian Peedee Formation in Maryland [66]. It is referred to Nyctosauridae on the basis of its small size, massive femoral neck, and distally expanded femoral shaft. The Peedee nyctosaurid differs from Nyctosaurus in having a robust shaft, but this feature is shared with Alcione (Fig 7), suggesting affinities with that genus.
Azhdarchidae are the most diverse group, perhaps because they occur in terrestrial and marine strata, allowing them to occupy more niches and increasing preservation potential. As many as 10 late Maastrichtian species are known. These include Quetzalcoatlus northropi [59] Q. sp. [59] and an unnamed species [2] from the Javelina Formation of Texas, a slender-necked azhdarchid from the Hell Creek Formation of Montana [14], a small azhdarchid from the Lance Formation of Wyoming [67], A. philadelphiae from Jordan [37], H. thambema [68,69] from Romania, a giant species from Mérignon, France [70], and P. mauritanicus and aff. Quetzalcoatlus from the marine phosphates of Morocco [43]. The giant Sidi Chennane azhdarchid could represent Arambourgiania or another species.
Maastrichtian pterosaurs both occupied a range of habitats and show a range of morphologies, suggesting diverse ecologies (Figs 18 and 20). Among pteranodontidians, a diversity of wing morphologies implies a wide range of flight styles. Pteranodontids and nyctosaurids were medium-to-giant marine forms with high-aspect ratio wings and low wing loading [71]. Nyctosaurids resemble frigatebirds, marine thermal soarers, in morphology and flight performance, and probably exploited flap gliding and thermal soaring [71]. However, the short-winged Alcione may have been better adapted for flapping flight, whereas the large, slender-winged Barbaridactylus was probably specialized for gliding. Pteranodon was most likely a glider specialized to exploit marine thermals [71,72], and Tethydraco probably had a similar ecology. Like modern seabirds, including frigatebirds, gulls, petrels, and tropicbirds, pteranodontids and nyctosaurids would have flown long distances in search of food, feeding on the wing [72] and perhaps while floating on the water [73].
Azhdarchids had short wings but long arms and legs and likely foraged on the ground or wading in shallow water, as suggested by trackways [58,78]. Their ecology is controversial, and azhdarchids have variously been interpreted as scavengers, piscivores, probers, skimmers, or terrestrial predators [58]. Repeated occurrences of azhdarchids in brackish water and marine environments [36,37,43] suggests at least some species exploited marine resources. The short wings of azhdarchids [58] would have been inefficient for aerial foraging; instead, the long, pointed jaws [59] and long necks of azhdarchids, especially Quetzalcoatlus and kin, are consistent with hunting while wading in shallow water [79,]. Azdharchids might have fished by grasping prey with the beak, as in storks, or using heron-like spear fishing [73]. Probe traces associated with azhdarchid tracks [80] suggest a stork-like foraging strategy for at least some species. Azhdarchidae are also well represented in continental deposits, however, [81], and some species had short necks [69,78] and robust beaks [2], suggesting a distinct feeding strategy. Similarly, azhdarchid skeletons imply variation in their flight. The sheer size of the largest azhdarchids implies a flight style distinct from that of the smaller species, but variation is seen even within giant azhdarchids, from slender wing bones in the Sidi Chennane giant, to light but robust bones in Quetzalcoatlus [49], to the massively built Hatzegopteryx [78], showing that diverse locomotor strategies existed even within the giants. Given variation in beak shape [78], neck proportions [78], wing proportions, and body size, as well as their occurrence in a range of environments, azhdarchids probably exploited a variety of niches; they are best seen not as a specialist lineage but as a Late Cretaceous adaptive radiation.
The closest analogues to azhdarchids among modern organisms are long-necked, long-beaked birds such as cranes, herons, and bustards. Rather than being highly specialized for a particular niche, these birds are opportunistic feeders and occupy a wide range of niches [56,82–84]. By analogy, it is unlikely that azhdarchids specialized on a single niche. Instead, some may have foraged for vertebrates and invertebrates in marginal marine habitats such as bays, lagoons, mudflats and estuaries, like many herons and storks [83]. Others may have been terrestrial predators [69,78], similar to the cattle egret, white stork [83], and ground hornbill [84]; scavengers [69] analogous to the Marabou stork; or plant-dominated omnivores similar to bustards and cranes [82].
These qualitative assessments are borne out by quantitative analyses of pterosaur functional diversity (Figs 18 and 19). PCoA of functional characters suggests that Maastrichtian pterosaurs occupied a range of ecologies, and resampling suggests that functional diversity is undersampled (Fig 19). Just as sampling the Moroccan phosphates has revealed previously unknown diversity, future finds in other localities may reveal previously unrecognized taxic and ecological diversity.
Our data, along with previous discoveries, show that pterosaur diversity did not decline in the latest Cretaceous and may have been increasing prior to the K-Pg extinction (Figs 16–19). Taxic diversity peaks in the Early Cretaceous with a pterodactyloid radiation, followed by a decline in the middle Cretaceous. Subsequently, however, pterosaurs radiated up to the K-Pg boundary; Azhdarchidae diversified in terrestrial and nearshore marine environments, while Pteranodontia radiated in nearshore and offshore marine habitats. Similarly, pterosaur functional diversity appears to actually increase in the latest Cretaceous. This increase is driven primarily by the appearance of giant azhdarchids such as Quetzalcoatlus and Arambourgiania, but the appearance of the short-winged Alcione and the large Barbaridactylus shows nyctosaurs radiated as well.
The apparent latest Cretaceous decline of diversity [8] and disparity [9] seen in previous studies results from the Signor-Lipps Effect and variation in the completeness of the fossil record. Similarly, while phylogenetic diversity estimates show a modest decline (Fig 17), this is a predictable result of the Signor-Lipps Effect: the absence of post-Cretaceous pterosaurs means that ghost lineages cannot be inferred in the late Maastrichtian. Again, the patterns recovered here support a catastrophic extinction at the K-Pg boundary.
Given the rarity of pterosaur fossils, it is possible that additional lineages survived up to the K-Pg boundary. Several candidates exist. One such clade is Tapejaridae. Previously restricted to the Early Cretaceous, C. dobruskii [16] from the Turonian-Campanian of Brazil shows that tapejarids survived into the Late Cretaceous. Our analysis also follows previous analyses [24,85] in recovering B. galaczi, previously referred to Azhdarchidae [23], as a Late Cretaceous tapejarid.
Another candidate lineage is represented by P. barbarulna from the Campanian of Montana [25]. Piksi is controversial; first described as a bird, it has been reinterpreted as a small pterosaur [20] and then as a theropod [86]. Piksi shows no characters of birds or theropods that are not also seen in pterosaurs [20] but does show pterosaur synapomorphies and autapomorphies (see S1 Text). Piksi may represent a distinct lineage of small pterosaur that extends into the late Cretaceous. If so, it is unclear whether it belongs to a known family or a previously unknown lineage.
A third clade that may extend into the latest Cretaceous is Thalassodromidae. Our analysis recovers the Cenomanian A. saharica [87] and the Campanian/Maastrichtian A. sudamericanus [88] as thalassodromids. This result is poorly supported, and given the lack of other latest Cretaceous thalassodromiids, these fossils could represent azhdarchids.
Extending these clades to the K-Pg boundary would involve range extensions of 10–30 Ma. However, as shown by the discovery of late Maastrichtian Pteranodontidae, the Signor-Lipps Effect exerts a powerful influence on taxa with a poor fossil record. Long range extensions are possible with poor sampling, and the terrestrial record is particularly incomplete. Here most pterosaurs are known from rare, dissociated remains [14,67,69,88,89], and few associated specimens are known [25,59,90]. Given this, we hypothesize that pterosaur diversity remains undersampled and predict that further sampling will reveal additional lineages in the late Maastrichtian.
Birds did not drive pterosaurs extinct directly. Rather than competing, pterosaurs and birds appear to engage in size-based niche partitioning, avoiding competition. No known Late Cretaceous birds exceeded 2 m in wingspan or a few kg in mass [18,91]. Meanwhile, Late Cretaceous pterosaurs were mostly large bodied, ranging from 2 to over 10 m in wingspan [18], with the possible exception of Piksi. This pattern holds in marine ecosystems, where small Ichthyornithes coexisted with large pteranodontids and nyctosaurids, and terrestrial habitats, where small birds [18,91] lived alongside large and giant azhdarchids (Fig 20). Birds apparently outcompeted pterosaurs at small sizes, but the absence of large (>5 kg) birds suggests that the birds could not compete with pterosaurs at large size; here, pterosaurs dominated. This is not to say that no large birds or small pterosaurs existed, but they must have been rare to escape discovery.
A similar pattern is seen with nonavian dinosaurs and mammals. Dinosaurs occupied large-bodied niches as predators and herbivores, and mammals diversified at small body sizes [92]. The fate of the pterosaurs also mirrors the fate of the dinosaurs. Birds and mammals, which were most diverse at small sizes, survived the K-Pg extinctions; pterosaurs and nonavian dinosaurs, which were most diverse at large sizes, did not. While avian competition may not have directly driven pterosaurs extinct, the absence of small pterosaurs resulting from avian competition may have left pterosaurs vulnerable to an extinction event that selected against large size [18].
Finally, pterosaur extinction may have contributed to avian radiation. The extinction of pterosaurs and archaic birds, including enantiornithes and stem ornithurines [74,93], left the surviving birds with few competitors. This created an opportunity for the emergence of a diverse fauna of birds in the early Paleogene [94–96]. Strikingly, within 10 million years of the extinction of the pterosaurs, marine birds diversified. Tropicbirds [97] and the first large marine soaring birds, the Pelagornithidae, appeared in marine ecosystems [98,99], large soaring pelecaniforms appeared in freshwater habitats [100], and large lithornithid palaeognaths appeared in terrestrial habitats [101]. These patterns suggest that the extinction of pterosaurs in these environments allowed birds to evolve large size.
The high diversity seen in the late Maastrichtian of Morocco suggests that pterosaur diversity remained stable or increased prior to the end-Cretaceous mass extinction. These patterns are consistent with a catastrophic extinction of pterosaurs at the K-Pg boundary, driven by the Chicxulub impact. Pterosaurs and birds engaged in size-based niche partitioning, and pterosaur extinction provided a competitive release that helped drive avian radiation in the Early Cenozoic.
Phylogenetic analysis used the character-taxon matrix from Andres et al. [60], updated with new characters and all diagnostic taxa from the Late Cretaceous, for a total of 134 taxa and 271 characters. Ordered and unordered characters were used and equally weighted. Continuous characters were rescaled to unity using the “nstates stand” command. Inapplicable features were reductively coded [102], with multistate coding used to denote variation within species or instances in which all but a couple of the possible states could be excluded. Analysis was conducted with TNT v.1.5 [103] using discrete and continuous character partitions. Ambiguous branch support was not used, zero-length branches were automatically collapsed, and the resultant trees were filtered for best score. Basic tree searches of 2,000 random addition sequence replicates were conducted with and without the parsimony ratchet. For time calibration, the timescale is from Gradstein et al. [104], and ages for species were taken from the literature.
Diversity curves were created from the time-calibrated phylogeny. The timescale was converted into 1 Ma bins and divided into stages and early, middle, and late substages. A taxic diversity curve of species counts and a phylogenetic diversity curve including species and ghost lineage counts were generated. A species was counted as present for the entire length of its possible occurrence, with the exception of Domeykodactylus ceciliae [105], which is only constrained to the Early Cretaceous and was given the earliest occurrence of its relative Noripterus parvus [106]. Minimum divergence dates of 0 million years were used for the ghost taxon and lineage extensions.
|
10.1371/journal.pntd.0007649 | A high-throughput and multiplex microsphere immunoassay based on non-structural protein 1 can discriminate three flavivirus infections | The explosive spread of Zika virus (ZIKV) and associated complications in flavivirus-endemic regions underscore the need for sensitive and specific serodiagnostic tests to distinguish ZIKV, dengue virus (DENV) and other flavivirus infections. Compared with traditional envelope protein-based assays, several nonstructural protein 1 (NS1)-based assays showed improved specificity, however, none can detect and discriminate three flaviviruses in a single assay. Moreover, secondary DENV infection and ZIKV infection with previous DENV infection, both common in endemic regions, cannot be discriminated. In this study, we developed a high-throughput and multiplex IgG microsphere immunoassay (MIA) using the NS1 proteins of DENV1-DENV4, ZIKV and West Nile virus (WNV) to test samples from reverse-transcription-polymerase-chain reaction-confirmed cases, including primary DENV1, DENV2, DENV3, WNV and ZIKV infections, secondary DENV infection, and ZIKV infection with previous DENV infection. Combination of four DENV NS1 IgG MIAs revealed a sensitivity of 94.3% and specificity of 97.2% to detect DENV infection. The ZIKV and WNV NS1 IgG MIAs had a sensitivity/specificity of 100%/87.9% and 86.1%/78.4%, respectively. A positive correlation was found between the readouts of enzyme-linked immunosorbent assay and MIA for different NS1 tested. Based on the ratio of relative median fluorescence intensity of ZIKV NS1 to DENV1 NS1, the IgG MIA can distinguish ZIKV infection with previous DENV infection and secondary DENV infection with a sensitivity of 88.9–90.0% and specificity of 91.7–100.0%. The multiplex and high-throughput assay could be applied to serodiagnosis and serosurveillance of DENV, ZIKV and WNV infections in endemic regions.
| Although there was a decrease of Zika virus (ZIKV) infection since late 2017, the specter of congenital Zika syndrome and its re-emergence in flavivirus-endemic regions emphasize the need for sensitive and specific serological tests to distinguish ZIKV, dengue virus (DENV) and other flaviviruses. Compared with traditional tests based on envelope protein, several nonstructural protein 1 (NS1)-based assays had improved specificity, however, none can discriminate three flaviviruses in a single assay. Moreover, secondary DENV infection and ZIKV infection with previous DENV infection, both common in endemic regions, cannot be distinguished. Herein we developed a high-throughput and multiplex IgG microsphere immunoassay using the NS1 proteins of four DENV serotypes, ZIKV and West Nile virus to test samples from laboratory-confirmed cases with different primary and secondary flavivirus infections. Combination of four DENV NS1 assays revealed a sensitivity of 94.3% and specificity of 97.2%. The ZIKV and WNV NS1 assays had a sensitivity/specificity of 100%/87.9% and 86.1%/78.4%, respectively. Based on the signal ratio of ZIKV NS1 to DENV1 NS1, the assay can distinguish ZIKV infection with previous DENV infection and secondary DENV infection with a sensitivity of 88.9–90.0% and specificity of 91.7–100.0%. This has applications to serodiagnosis and serosurveillance in endemic regions.
| Despite a marked decrease of Zika virus (ZIKV) infection since late 2017, the specter of congenital Zika syndrome (CZS) and its re-emergence in flavivirus-endemic regions highlight the need for sensitive and specific diagnostic tests [1–4]. Similar to the laboratory diagnosis for other flaviviruses, detection of nucleic acid as soon as possible post-symptom onset (PSO) is considered as the gold standard to confirm ZIKV infection, [5,6]. Since many individuals test for ZIKV infection beyond the period when RNA is detectable and most (~80%) of ZIKV infections are asymptomatic, serological tests remain as a key component of ZIKV confirmation [5,6]. Furthermore, ZIKV can be transmitted sexually or following asymptomatic infection [7–9].
ZIKV is a member of the genus Flavivirus of the family Flaviviridae, which includes several pathogenic mosquito-borne viruses in different serocomplexes. The four serotypes of dengue virus (DENV) belong to the DENV serocomplex; West Nile virus (WNV) and Japanese encephalitis virus (JEV) to the JEV serocomplex; yellow fever virus (YFV) as a single member; and ZIKV10. Given that the envelope (E) protein is the major target of antibody response after flavivirus infection, different E antigens such as recombinant E protein, inactivated virions or virus-like particles have been developed for serological tests [10–13]. Due to the presence of several highly conserved residues of flavivirus E proteins, anti-E antibodies in serum are commonly cross-reactive to different flaviviruses [13–17]. The guidelines of Centers for Disease Control and Prevention (CDC) recommend that positive or equivocal results of E protein-based IgM tests require further testing with time-consuming plaque reduction neutralization tests (PRNT) [5,6]. However, PRNT can confirm ZIKV-infected individuals who acquire ZIKV as the first flavivirus infection, known as primary ZIKV (pZIKV) infection, but often can only be interpreted as unspecified flavivirus infections for those who have experienced previous DENV or other flavivirus infections, limiting its application for ZIKV serodiagnosis in flavivirus-endemic regions.
When 795 sera that were IgM positive for ZIKV antigen by ELISA were tested for flavivirus neutralizing antibodies by PRNT, 45% were positive for ZIKV and at least one other flavivirus [18]. This non-specificity may be an inherent property of the early post-infection response to ZIKV or reflect prior flavivirus experience. A large number of Americans (7 million) have experienced a WNV infection since 1999 [19] and ~8 million traveled to yellow fever endemic countries in 2015 [20,21]. Thus, a sensitive, specific and multiplex serological test that can distinguish ZIKV and other flavivirus infections is needed in both U.S. and flavivirus-endemic countries [18]. Moreover, several studies have shown that anti-DENV or WNV antibodies can enhance ZIKV infection in vitro [22–26] and in small animals, in which administration of DENV-immune plasma resulted in increased viremia and mortality in stat2 knock out mice [27]. This is known as antibody-dependent enhancement, in which antibody at suboptimal concentration for neutralization can enhance DENV, ZIKV or other flavivirus entry and replication in Fcγ receptor-bearing cells such as monocytes and is believed to contribute to disease pathogenesis [28]. Despite ADE of ZIKV by previous DENV immunity was not supported by two studies in non-human primates [29,30], more in-depth studies of DENV immunity on ZIKV disease outcome and complication in humans are warranted [31–33]. Thus, serological tests that can distinguish pZIKV infection (p = primary) from ZIKV infection with previous DENV (ZIKVwprDENV, wpr = with previous) infection are crucial to understand the pathogenesis of ZIKV and CZS in regions where ZIKV and DENV co-circulate.
Compared with traditional E protein-based assays, several enzyme-linked immunosorbent assays (ELISAs) based on ZIKV nonstructural protein 1 (NS1), including a recently reported blockade of binding ELISA, have shown improved specificity [34–39]. However, secondary DENV (sDENV) and ZIKVwprDENV infections, of which both were common in endemic regions, cannot be discriminated [34–39]. Moreover, none can detect and distinguish ZIKV, DENV and other flavivirus in a single assay.
With its high-throughput and multiplex (up to 100-plex) capacity, microsphere immunoassay (MIA) has been employed in the detection of cytokines, transplantation and transfusion antigens, and various bacterial and viral pathogens [40–43]. Previously, we reported that a combination of ELISAs based on the NS1 proteins of DENV and ZIKV can distinguish various DENV and ZIKV infections [44,45]. In this study, we developed a high-throughput and multiplex IgG MIA using NS1 proteins of DENV1 to DENV4, ZIKV and WNV, and showed that the NS1 IgG MIA can detect and distinguish not only primary DENV, ZIKV and WNV infections but also sDENV and ZIKVwprDENV infections.
The Institutional Review Boards (IRB) of the University of Hawaii approved this study (CHS #17568, CHS#23786). S1 Table summarizes the numbers, serotypes, sampling time and sources of different panels of serum or plasma samples, including those from primary DENV1 (pDENV1), primary DENV2 (pDENV2), primary DENV3 (pDENV3), primary WNV (pWNV), pZIKV, sDENV and ZIKVwprDENV infections as well as flavivirus-naïve individuals. Samples collected <3 months or ≥3 months PSO were designated as convalescent- or post-convalescent-phase samples, respectively. Samples from reverse transcription-PCR (RT-PCR) confirmed Zika cases were from the Pediatric Dengue Cohort Study (PDCS) and the Pediatric Dengue Hospital-based Study in Managua, Nicaragua between July 2016 and March 2017 [46,47]. The Zika cases that were DENV-naïve or previously DENV-exposed were defined as pZIKV (p = primary) or ZIKVwprDENV (wpr = with previous) panels, respectively. The DENV-immune status was based on anti-DENV antibody testing by an inhibition ELISA at entry and annually of the PDCS [44–47]. Parents or legal guardians of all participants provided written informed consents, and participants ≥6-year old provided assents. These studies were approved by the IRBs of the University of California, Berkeley, and Nicaraguan Ministry of Health. Thirty-six plasma samples from blood donors, who were tested WNV-positive by the transcription-mediated amplification (a sensitive nucleic acid detection method used in blood bank), IgM and IgG antibodies between 2006 and 2015, designated as pWNV infection, were provided by the American Red Cross at Gaithersburg, Maryland [48]. Pre-2015-16 ZIKV epidemic convalescent- and post-convalescent-phase samples from RT-PCR confirmed cases with different primary DENV infections (pDENV1, pDENV2, and pDENV3) or sDENV infection were from Taiwan, Hawaii and Nicaragua; 53 flavivirus-naïve samples from a seroprevalence study in Taiwan were included as control in this study [44,45,49–52]. Samples from cases with primary DENV4 infection were not available. Primary DENV or sDENV infection was determined by IgM/IgG ratio or focus-reduction neutralization tests as described previously [49–51].
The NS1 gene (corresponding to amino acid residues 1–352) of ZIKV (HPF2013 strain) with a His-tag at the C-terminus was codon-optimized (Integrated DNA Technologies, Skokie, IL) and cloned into pMT-Bip vector to establish a Drosophila S2-cell stable clone [44]. ZIKV-NS1 protein from supernatants of the stable clone was purified by fast purification chromatography system (AKTA Pure, GE Health Care Bio-Science, Pittsburg, PA) [44]. Purified DENV1-4 and WNV NS1 proteins were purchased from The Native Antigen (Oxford, UK).
Ten μg each of the 6 purified NS1 proteins, bovine serum albumin (BSA) and PBS (as negative antigen control) were coupled individually onto 8 types of magnetic carboxylated miscrosphere beads (1.25 X 106 each) containing different fluorophores (MagPlexTM-C) (Luminex, TX, Austin) using two-step carbodiimide process at room temperature [53,54]. The antigen-conjugated microspheres were stored in 250 uL PBN buffer (PBS with 1% BSA and 0.05% sodium azide, Sigma Aldrich) at 4°C until use.
Eight types of microsphere beads coupled with different NS1 proteins, BSA or PBS were combined and diluted in PBS-1% BSA. Fifty μL of the mixture (containing ~1250 beads of each type) were added to each well of a flat-bottom 96-well plate, and incubated with 50 μL diluted serum or plasma (1:100 dilution in PBS-1% BSA) at 37°C for 30 min in the dark, followed by wash with 200 μL of PBS-1% BSA twice, incubation with 50 μL of red phycoerythrin-conjugated anti-human or anti-mouse IgG (Jackson Immune Research Laboratory, West Grove, PA) at 37°C for 45 min in the dark, and wash with 200 μl of PBS-1% BSA twice [54]. Microspheres were then resuspended in 100 μl of PBS-1% BSA, incubated for 5 min and read by Luminex 200 machine (Austin, TX). All incubations were performed on a plate shaker at 700 rpm and all wash steps used a 96-well magnetic plate separator (Millipore Corp., Billerica, MA) [54]. Each plate includes two positive controls (confirmed-ZIKV or DENV infection), four negative controls (flavivirus-naïve samples), samples, and mouse anti-His mAb (all in duplicates). The median fluorescence intensity (MFI) was determined for 100 microspheres for each well. The MFI values for each antigen were divided by the mean MFI value of one positive control (MFI~104) and multiplied by 104 to calculate to rMFI for comparison between plates (S1 Fig). The cutoff rMFI for each antigen was defined by the mean rMFI value of 19 flavivirus-naïve samples plus 5 standard deviations, which gave a confidence level higher than 99.9% from 4 negatives [55]. Each MIA was performed twice (each in duplicate). New batch of conjugated antigens was tested with flavivirus-naïve panel to determine the cutoff rMFI.
DENV1-, DENV2-, DENV3-, and ZIKV-NS1 IgG ELISAs have been described previously [44,45]. Briefly, purified NS1 proteins (16 ng for individual NS1 protein per well) were coated on 96-well plates at 4°C overnight, followed by blocking (StartingBlock blocking buffer, Thermo Scientific, Waltham, MA), incubation with primary antibody (serum or plasma at 1:400 dilution) and secondary antibody (anti-human IgG conjugated with horseradish peroxidase, Jackson Immune Research Laboratory, West Grove, PA), and wash [44,45]. After adding tetramethylbenzidine substrate (Thermo Scientific, Waltham, MA) followed by stop solution, the optical density (OD) at 450 nm was read with a reference wavelength of 630 nm. Each ELISA plate included two positive controls (confirmed-ZIKV or DENV infection), four negative controls (flavivirus-naïve sample), and samples (all in duplicate). The OD values were divided by the mean OD value of one positive control (OD close to 1) in the same plate to calculate the relative OD (rOD) values for comparison between plates [44,45]. The cutoff rOD was defined by the mean rOD value of negatives plus 12 standard deviations, which gave a confidence level of 99.9% from 4 negatives [55]. Each ELISA was performed twice (each in duplicate).
Two-tailed Mann-Whitney test was used to determine the P values between two groups, the two-tailed Spearman correlation test the relationship between the rOD and rMFI values, and the receiver-operating characteristics (ROC) analysis the cutoffs of the rMFI and rOD ratios (GraphPad Prism 6). The 95% confidence interval (CI) was calculated by Excel.
We first employed the multiplex NS1 IgG MIA to test samples from primary DENV (pDENV1, pDENV2 and pDENV3), pZIKV and pWNV infection panels. Compared with flavivirus-naïve panel, the pDENV1 panel recognized the NS1 proteins of DENV1 (100%) and other DENV serotypes (33.3 to 61.9%), but not those of different serocomplexes (ZIKV and WNV NS1 proteins) (Fig 1A and 1B). Similarly, the pDENV2 and pDENV3 panels recognized the NS1 protein of the homologous serotype (DENV2, DENV3) better than those of other serotypes (Fig 1C and 1D), but did not recognize ZIKV or WNV NS1 protein except two samples (recognizing WNV, 2/13). The pZIKV panel recognized ZIKV NS1 protein but not those of WNV and DENV except two sample recognizing DENV2 (2/38), whereas the pWNV panel recognized WNV proteins rather than those of ZIKV and DENV except one sample (recognizing DENV4, 1/36) (Fig 1E and 1F). Taken together, these findings suggested that primary infection panels recognized the homologous (infecting serotype) NS1 protein better than other NS proteins within the same serocomplex, and in general did not recognize an NS protein of different serocomplexes (Fig 1G).
We next tested samples from sDENV and ZIKVwprDENV panels. For convalescent-phase samples, sDENV panel not only recognized NS1 proteins of DENV1-4 (66.7 to 100%) but also those of ZIKV and WNV (45.8 to 54.2%) (Fig 2A). The ZIKVwprDENV panel recognized ZIKV NS1 protein (100%) as well as DENV1-4 and WNV NS1 proteins (60.0 to 90.0%) (Fig 2B). A similar trend was observed for post-convalescent-phase samples (Fig 2C and 2D). These findings were in agreement with our previous reports based on NS1 IgG ELISAs [44,45], and suggested that after repeated flavivirus infections, such as sDENV and ZIKVwprDENV infections, anti-NS1 antibodies cross-reacted to multiple NS1 proteins, including those from prior exposure or sometimes those with no prior exposure.
Previously we reported that sDENV panel not only recognized DENV1 NS1 protein but also ZIKV NS1 protein in IgG ELISA (95.8 and 66.7%, respectively); similarly the ZIKVwprDENV panel recognized both ZIKV and DENV1 NS1 proteins (95.0 and 85.0%, respectively) [44]. Using the rOD ratio of ZIKV NS1 to DENV1 NS1 with a cutoff at 0.24, we can distinguish ZIKVwprDENV and sDENV panels. Since the same sDENV and ZIKVwprDENV panels recognized both DENV1 and ZIKV NS1 proteins in IgG MIA (Fig 2A and 2B), we calculated the ratio of relative median fluorescence intensity (rMFI) of ZIKV NS1 to that of DENV1 NS1 and found that a cutoff of the rMFI ratio at 0.62, as determined by ROC analysis, can distinguish these two panels with a sensitivity of 88.9% and specificity of 91.7% (Fig 2E). Since both panels also recognized DENV2 NS1 protein, we further calculated the ratio of rMFI of ZIKV NS1 to DENV2 NS1; interestingly a cutoff of the rMFI ratio at 0.62 was able to distinguish these two panels with a sensitivity of 94.4% and specificity of 90.9% (Fig 2F). Similar observations were found for post-convalescent-phase sDENV and ZIKVwprDENV panels; these two panels can be distinguished by a cutoff (0.62) of the rMFI ratio for ZIKV NS1 to DENV1 NS1 or DENV2 NS1 with a sensitivity/specificity of 90.0/100% or 83.3/100%, respectively (Fig 2G and 2H).
Since these panels have been tested with individual DENV1 to DENV4 and ZIKV NS1 IgG ELISAs previously [45], we compared the detection rates for each NS1 protein between ELISA and MIA. For the pZIKV panel, ZIKV NS1 ELISA had a detection rate of 100%, comparable to that of MIA, for the post-convalescent-phase samples, but only 5% for the convalescent-phase samples, which was much lower than that of MIA (100%) (Fig 3A and 3B). Although 19 convalescent-phase pZIKV samples were tested negative by ZIKV NS1 IgG ELISA, the relative optical density (rOD) values were positively correlated with the rMFI values (correlation coefficient r = 7464, P = 0.0002) (Fig 3C), suggesting that ZIKV NS1 MIA was more sensitive than ELISA. A positive correlation was also found between rOD and rMFI values for the post-convalescent-phase samples (r = 8922, P<0.0001) (Fig 3D). For pDENV1 panel, DENV1 NS1 ELISA and MIA had comparable detection rates (100%) for both convalescent and post-convalescent-phase samples (Fig 3E and 3F). Similarly, a positive correlation was found between rOD and rMFI values (Fig 3G and 3H).
For ZIKVwprDENV panels, ZIKV NS1 IgG ELISA and MIA had comparable detection rates for both convalescent and post-convalescent-phase samples (Fig 4A and 4B). A positive correlation was found between rOD and rMFI values for ZIKV NS1 as well as DENV1, DENV2, DENV3 and DENV4 NS1 tested (Fig 4C–4E). Similar observations were found for sDENV panels (S2 Fig).
Table 1 summarizes the results of all samples tested with different NS1 proteins (DENV1, DENV2, DENV3, DENV4, DENV1, 2, 3 or 4, ZIKV and WNV) in the IgG MIA. For statistical analysis comparing different panels, one sample from each participant was included (S2 Table). The overall sensitivity of each DENV (DENV1, DENV2, DENV3) NS1 IgG MIA to detect different DENV infections ranged from 73.6 to 90.1% and specificity from 98.1 to 100% (Table 2). Interestingly, combination of four DENV NS1 IgG MIA increased the sensitivity to 94.5%, while maintaining the specificity of 97.2%, suggesting that this multiplex assay can be applied to detect DENV infections rather than distinguish different DENV serotypes. For the ZIKV NS1 IgG MIA, the overall sensitivity was 100% and specificity 87.9%. For the WNV NS1 IgG MIA, the overall sensitivity was 86.1% and specificity 78.4% (Table 2).
In this study, we developed a high-throughput and multiplex IgG MIA using NS1 proteins of DENV1 to DENV4, ZIKV and WNV to detect and distinguish various DENV, ZIKV and WNV infections. Based on the results, we propose an algorithm to discriminate primary DENV, pZIKV and pWNV infections, sDENV infection and ZIKVwprDENV infection (Fig 5). Previous studies of flavivirus serodiagnosis mainly focused on two flaviviruses. Compared with a recent study of IgG MIA containing ZIKV and DENV antigens, our multiplex IgG MIA consists of 6 antigens (DENV1 to DENV4, WNV and ZIKV NS1 proteins) plus two controls (BSA and PBS) [56]. To our knowledge, this is the first report of a single serological test to detect three flavivirus infections. Our findings that combination of DENV1 to DENV4 NS1 IgG MIA increased the sensitivity to 94.3% while maintaining a specificity of 97.2% and that the rMFI ratio of ZIKV NS1 to DENV1 or DENV2 NS1 can distinguish ZIKVwprDENV and sDENV infections with a sensitivity of 83.3–94.4% and specificity of 90.9–100.0% have important applications to serodiagnosis and serosurveillance of DENV and ZIKV infections in regions where both viruses co-circulate.
Generally in agreement with our recent study of individual DENV NS1 ELISAs [45], we found that DENV1 and DENV3 NS1 IgG MIAs can detect primary DENV infection of the homologous serotype with a sensitivity (100%) higher than that for heterologous serotypes (25.0 to 100%) (Table 2). DENV1, DENV2 and DENV3 NS1 IgG MIAs can detect secondary DENV infection with a sensitivity of 95.5 to 100%. This was also consistent with our previous study using Western blot analysis, in which anti-NS1 antibodies recognized NS1 protein predominantly of the infecting serotype after primary DENV infection and multiple NS1 proteins after secondary infection [13]. Taken together, due to the variable and extensive cross-reactivity of anti-NS1 antibodies after primary and secondary DENV infections, respectively, it is difficult to use a single NS1 IgG MIA or ELISA to identify the infecting DENV serotype. Notably, the combination of four DENV NS1 IgG MIA can detect different primary and secondary DENV infections with a sensitivity of 94.3% and specificity of 97.2% (Table 2), suggesting the feasibility and application of this multiplex NS1 IgG MIA to detect DENV infection rather than distinguish DENV serotypes.
The overall sensitivity of the ZIKV NS1 IgG MIA was 100% and the specificity was 87.9%, primarily due to the cross-reactivity of the sDENV panel (Table 2). The sensitivity (100%) was higher than or comparable with those previously reported (79 to 100%) using the Euroimmun ZIKV NS1 IgG ELISA kit [34–37]. The ZIKV NS1 blockade of binding ELISA had an overall specificity of 91.4–92.6%, which reduced to 77.6–90.5% when comparing with sDENV panel [38,39]. A recently reported ZIKV NS1 IgG3 ELISA had a sensitivity of 97% based on samples from Salvador, but it reduced to 83% when comparing with samples outside of Salvador [32]. A previous study of multiplex IgG MIA including ZIKV NS1 reported a sensitivity of 100% and specificity of 78% for pZIKV panel based on PRNT results, however, the sDENV and ZIKVwprDENV panels were not distinguished [56]. For the WNV NS1 IgG MIA, the overall sensitivity was 86.1% probably due to sampling during the early convalescent-phase for this pWNV panel (S1 Table), and the specificity was 78.4%, mainly due to the cross-reactivity from the sDENV and ZIKVwprDENV panels (Table 2). Using the rMFI ratio of ZIKV NS1 to DENV1 or DENV2 NS1, we can distinguish ZIKVwprDENV and sDENV panels with a sensitivity of 83.3–94.4% and specificity of 90.9–100.0%. This was consistent with our previous reports of IgG ELISAs using the rOD ratio of ZIKV NS1 to DENV1 NS1 or mixed DENV1-4 NS1 to distinguish these two panels with a sensitivity of 91.7–94.1% and specificity of 87.0–95.0% [44,45]. It is worth noting since DENV3 and DNV4 NS1 proteins were not recognized by several samples from the sDENV and ZIKVwprDENV panels (Fig 2A and 2D), they were not included in the analysis of the rMFI ratio.
Comparing the results of individual NS1 IgG MIA in this study and those of NS1 IgG ELISA reported previously [45], we found comparable detection rates between MIA and ELISA, and positive correlations between the rMFI and rOD values for both convalescent-phase and post-convalescent-phase samples of most panels tested including pDENV1, sDENV and ZIKVwprDENV panels except pZIKV panel (Figs 3 and 4 and S2 Fig). Of note, the IgG MIA detection rates for DENV1-4 for the post-convalescent-phase ZIKVwprDENV panel were much lower than those for the sDENV panel (Fig 4E and S2E Fig), suggesting that prior DENV exposure of the ZIKVwprDENV panel may have been only to a single DENV serotype. For the convalescent-phase pZIKV panel, the higher detection rate of ZIKV NS1 IgG MIA (100%) than that of ELISA (5%) and the positive correlation between rOD and rMFI values suggest that MIA was more sensitive than ELISA (Fig 3A–3C). Thus, we did not observe a trend of increased detection rates of NS1 IgG MIA from convalescent to post-convalescent phases for primary infection panels (pZIKV, pDENV1) (Fig 3B and 3F) as previously reported for NS1 IgG ELISA and blockade of binding of NS1 ELISA [38,45]. Notably we incubated 16 ng antigen coated on each well with 50 μL of serum (1:400) in ELISA, whereas we incubated ~10 ng antigen (in 1250 beads) with serum (final dilution 1:200) per well in MIA. The higher concentration of serum and more surface area of antigen coupled on beads may account for the higher sensitivity of the IgG MIA compared with IgG ELISA for the pZIKV convalescent-phase panel.
Although neutralization tests are still considered a confirmatory assay, they are time-consuming and can be performed only in reference laboratories. Compared with PRNT and ELISA, the multiplex MIA requires less time (2.5 h vs. 7 h for ELISA and 5–6 days for PRNT) and less sample volume (1 μL vs. 8 μL for ELISA and 144 μL for PRNT for 8 antigens or viruses). The newly developed multiplex NS1 IgG MIA could have wide-ranging applications, such as serodiagnosis, blood screening, serosurveillance of ZIKV, DENV and WNV infections, and retrospective study of ZIKV infection among pregnant women with CZS [57,58]. The current octaplex (6 NS1 antigens plus PBS and BSA controls) IgG MIA serves as a “proof-of-concept” assay to demonstrate that NS1-based MIA can distinguish three flavivirus infections; incorporation of other antigens would increase the detection capacity for different clinical settings and studies. These together would further our understanding of the epidemiology, pathogenesis and complications of ZIKV in regions where multiple flaviviruses co-circulate [1–4].
There are several limitations of this study. First, due to limited samples of < 3 months PSO from patients with primary DENV infection (S1 Table), the study focused on NS1 IgG MIA. Future studies on NS1-based IgM MIA are warranted. Second, despite the availability of two-time point samples for the pZIKV and ZIKVwprDENV panels, future studies involving more sequential samples are needed to validate these observations. Additionally, the sample size in each panel with well-documented infection is small. Third, although this multiplex assay can distinguish various panels of samples with three flavivirus infections, future tests that can distinguish other pathogenic flaviviruses such as JEV, YFV and tick-borne encephalitis virus (TBEV) remain to be exploited [59,60]. Moreover, samples with well-documented repeated flavivirus infections such as DENV with previous ZIKV infection and sequential DENV and WNV infections are lacking and remain to be investigated in the future. In light of the successful implementation of several flavivirus vaccines and vaccine trials in flavivirus-endemic regions, serological tests that can distinguish ZIKV infection from vaccinations with DENV, JEV, YFV and TBEV vaccines are warranted [59,60].
|
10.1371/journal.pcbi.1003216 | Characterizing Changes in the Rate of Protein-Protein Dissociation upon Interface Mutation Using Hotspot Energy and Organization | Predicting the effects of mutations on the kinetic rate constants of protein-protein interactions is central to both the modeling of complex diseases and the design of effective peptide drug inhibitors. However, while most studies have concentrated on the determination of association rate constants, dissociation rates have received less attention. In this work we take a novel approach by relating the changes in dissociation rates upon mutation to the energetics and architecture of hotspots and hotregions, by performing alanine scans pre- and post-mutation. From these scans, we design a set of descriptors that capture the change in hotspot energy and distribution. The method is benchmarked on 713 kinetically characterized mutations from the SKEMPI database. Our investigations show that, with the use of hotspot descriptors, energies from single-point alanine mutations may be used for the estimation of off-rate mutations to any residue type and also multi-point mutations. A number of machine learning models are built from a combination of molecular and hotspot descriptors, with the best models achieving a Pearson's Correlation Coefficient of 0.79 with experimental off-rates and a Matthew's Correlation Coefficient of 0.6 in the detection of rare stabilizing mutations. Using specialized feature selection models we identify descriptors that are highly specific and, conversely, broadly important to predicting the effects of different classes of mutations, interface regions and complexes. Our results also indicate that the distribution of the critical stability regions across protein-protein interfaces is a function of complex size more strongly than interface area. In addition, mutations at the rim are critical for the stability of small complexes, but consistently harder to characterize. The relationship between hotregion size and the dissociation rate is also investigated and, using hotspot descriptors which model cooperative effects within hotregions, we show how the contribution of hotregions of different sizes, changes under different cooperative effects.
| Within a cell, protein-protein interactions vary considerably in their degree of stickiness. Mutations at protein interfaces can alter the interaction between protein pairs, causing them to dissociate faster or slower. This may lead to an alteration in the dynamics of the cellular networks in which these proteins are involved. Therefore, the calculation and interpretation of mutants, which affect the rate of dissociation, is critical to our understanding of complex networks and disease. A key characteristic of protein–protein interfaces is that a subset of residues are responsible for most of the binding energy, such residues are called hotspots and effectively represent the sticky points of the interaction. In this work, we exploit both hotspot energies and organization and use them for the calculation of off-rate changes upon mutations. The insights gained provide us with a clearer understanding of the critical regions of stability and how they change for complexes of different sizes. Moreover, we provide a comprehensive map of the key determinants responsible for the accurate characterization of different classes of mutations, complexes and interface regions. This paves the way for more intelligent computational-interface-design algorithms and provides new insight into the interpretation of destabilizing mutations involved in complex diseases.
| Protein-Protein interactions are at the core of all biological systems and the rates at which biomolecules associate and disassociate are the major driving forces behind the complex time-dependent signaling observed in many biological processes. Ordinary Differential Equations (ODEs) are generally used to model these processes [1]–[3]; however, ODEs are bottlenecked by the limited availability of the relevant experimental rate constants [4]. Therefore, the accurate calculation of the kinetic rate constants holds significant application in our understanding of complex networks involved in diseases such as cancer [5]–[7]. Kinetic rate constant prediction is also central to effective drug design [8]–[10]; in vivo scenarios, where the concentration of a drug-like ligand exposed to its target receptor is not constant, as usually it is in vitro, the drug efficacy is no longer well described by the in vitro measured dissociation constant, but rather depends on the association (kon) and dissociation (koff) rate constants [8]. Whereas the enhancement of the on-rate is limited by the diffusion rate and several pharmacological factors, off-rate optimization is independent of such factors and entirely dependent on the short-range interactions between the bound monomers in question [8]. Hence the calculation and minimization of dissociation rate constants becomes a critical objective in drug design optimization [11]. At the other end of the spectrum, most disease causing mutations which are not in the protein core, occur at the interface regions and result in complex destabilization [12] and a number of studies have shown that changes in the binding free energy are largely the result of changes in the off-rate as opposed to minimal changes in the on-rate [13], [14]. While several aspects of biomolecular association have been investigated [10], [15]–[17], work on dissociation rate is still very limited [18]. Moreover, up to the analysis reported in this work, which attempts to calculate off-rate variations upon mutations in a high throughput context, calculation of dissociation rate constants has been limited to wild-type complex studies [19], [20].
The koff of a complex may be estimated using Molecular Dynamic (MD) simulations starting from the bound structure and allowing for dissociation to occur [21]. MD simulations typically allow simulation times of ns to μs, which are below the time-scales necessary for natural dissociation. Although steered molecular dynamics (SMD) simulations provide an alternative means to estimate the dissociation of protein complexes [21]–[23], such methods bias the dissociation process through a forced pathway in the direction of the applied force, and computational complexity limits their applicability. In our recent work, the wild-type kinetic rate constants of a number of complexes were predicted, using empirical scoring functions, with a number of molecular descriptors, describing various aspects of protein-protein interaction [19]. Whereas many descriptors showed high correlations with the association rate, particularly those calculated using the unbound structures, significant correlations for the dissociation rate could not be found.
Given the limited predictive ability of the current molecular features for koff [19], instead of trying to characterize off-rate mutations directly using such molecular features only, a different approach is taken here, one which exploits the synergistic and distributional information available in hotspot residues. Hotspots refer to a subset of residues at the interface which are able to significantly destabilize the binding free energy by more than 2 kcal/mol when mutated to alanine [24]. So far hotspot research has mainly focused on their identification [25]–[35], residue-level properties [24], [36] and distributional properties [37]–[39]. However, work on their practical application is still very limited [40]. Here, the relationship between hotspot energetics and the dissociation rate constant is investigated. We put to test the notion of whether the ΔΔGs of single-point mutations to alanine, as traditionally trained upon and predicted by hotspot prediction algorithms, can be used to quantify changes in Δkoff. The key point of interest here is that mutations, such as those we would like to quantify the changes in off-rate for, are not limited to single-point alanine mutations, as are in hotspot prediction algorithms. Therefore, a direct estimation of Δkoff using ΔΔG will not suffice. To address this, an unconventional approach is taken and computational alanine scans of the interface pre- and post-mutation are performed using hotspot predictor algorithms. Using the ΔΔGs of the single-point alanine mutations generated from these scans, a set of 16 hotspot descriptors are designed and calculated. The hotspot descriptors are then used as features to quantify off-rate changes of single-point, and more importantly multi-point, mutations to alanine and also non-alanine. A key advantage of using such hotspot descriptors, is not only the fact that non-alanine and multi-point mutations can now be characterized using single-point alanine mutations, but higher-order or rather, global effects of a given mutation can now be addressed. These include changes in the size and distribution of hotregions (clusters of hotspots), cooperative effects within hotregions and changes in localized stability regions such as the core, rim and support regions. All of which, as shown in this work, play varyingly important roles in the determination of the off-rate of a given mutation.
Our results confirm that indeed, using hotspot descriptors, the energies of single-point mutations to alanine can be used to describe effects of mutations other than alanine and also multi-point mutations. Machine learning models using such hotspot descriptors show consistently higher predictive abilities in the fine-grained and coarse-grained prediction of off-rate changes upon mutation, than models without hotspot descriptors. We find that hotspot descriptors tend to be broadly predictive for different classes of mutations, whereas molecular descriptors can be highly specific to small subsets of mutations. Our investigations also highlight differences in the distribution of stable regions at the interface for complexes off different sizes and interface areas and show the effects of cooperativity, on the stability provided by hotregions of various sizes.
In the first part of this work, sets of hotspot descriptors are generated, where each set represents hotspot descriptors generated from a particular hotspot predictor. The hotspot predictors tested include; two hotspot prediction servers (KFC2 [30] and Hotpoint [28]) and also two hotspot predictors developed in this work (RFSpot and RFSpot_KFC2). The hotspot descriptors' ability to characterize changes in off-rate due to mutations is assessed on a set of 713 experimental off-rates taken from wild-type and mutated proteins in the SKEMPI database [41]. Experimental off-rates in the dataset cover a range of Δlog10(koff) of −8.5 to 6.8, with koff units of s−1, and represent a diverse set of interactions as listed in the Supplementary Information (Dataset S1). As a relative performance measure, a benchmark set of 110 molecular descriptors (Text S1) is also included in the analysis and compared to the performance of the hotspot descriptors. The molecular descriptor set consists of a complex and comprehensive set of structure related descriptors characterizing various aspects of protein-protein interactions and their energetics; a subset of which have already proven to be successful in our previous work on predicting wild-type protein-protein binding free energies and kinetic rate constants [19], [42] and therefore serves as a thorough benchmark comparison. All descriptor analysis in the initial section is independent of any machine learning models trained on off-rate data. Rather, the aim here is to uncover the individual predictive power of each descriptor in estimating off-rate mutations. The Pearson's Correlation Coefficient (PCC) is used to evaluate fine-grain predictive ability, i.e. the ability to make numerical predictions. On the other hand, the Mann Whitney U-Test and several classification measures described in Materials and Methods are used to evaluate the coarse-grain ability to detect stabilizing mutations from neutral and destabilizing mutations.
In the second part, the prediction of off-rates using machine learning models is investigated. Here, several models using both hotspot and molecular descriptors are built, and their prediction patterns and anomalies highlighted. In order to uncover similarities in their predictions, the 713 off-rate dataset is categorized into what we term as data regions. Such data regions represent mutations that have a common physical property, or come from a similar type of complex or region on the interface. Mutations within a data region in turn might hold different predictive difficulty than mutations in another. This data region analysis enables us to identify current strengths in the prediction of off-rates and conversely, mutations which are consistently harder to characterize.
In the third part of this work, the use of specialized models specific to different data regions is investigated. By doing so we are able to identify descriptors of which their predictive value is highly specific to subsets of mutations, regions on the interface, or types of complexes. The specialized models are generated using a Genetic Algorithm running Feature Selection (GA-FS) with either linear (Linear Regression, LR) or non-linear (Support Vector Machines, SVM) learning models.
In the latter sections, the effects of complex size and interface area on the distribution of stability regions at the interface are investigated. Issues related to cooperativity and conformational changes, in the determination of off-rates, are also highlighted.
One of the main motivations behind this work is to explore the use of currently available descriptors (physics-based and knowledge-based potentials) and design a new class of descriptors (hotspot descriptors) for describing changes in off-rates. On the design of a new class of descriptors, our proposition is that interface hotspots can be seen as the anchor points responsible for the stable longevity of a complex. Namely, changes in the number of hotspots, hotspot energies and their distribution across the interface brought upon by structural mutations directly relates to changes in off-rate. Our approach of using hotspot predictions and subsequently hotspot descriptors for characterizing off-rates is depicted in Figure 1. First a pre-mutation alanine scan is performed; essentially this translates to using a hotspot predictor of choice on each residue at the interface. This generates a collection of single-point alanine ΔΔGs that are then employed differently depending on the hotspot descriptor in question (See Table 1). For example if we are using Int_HS_Energy, then this hotspot descriptor will sum all the energies of only the hotspot residues. After all the hotspot descriptors for the wild-type complex are calculated, the mutation in question is applied using FoldX [43], such as the Arg to Leu mutation in Figure 1. Then, using a hotspot predictor as in the wild-type scan, another computational alanine scan is performed on the mutated interface. Again, all single-point alanine ΔΔGs are then fed into the hotspot descriptors. Continuing with the example of Int_HS_Energy as a hotspot descriptor, here the ΔΔGs of only the hotspot residues on the mutated interface are summed, and the final descriptor value will be the change in the sum of the single-point ΔΔGs to alanine of all hotspot residues pre- and post-mutation. This value is then correlated to ΔkoffLeu→Arg.
The motivations and calculation for each of the 16 hotspot descriptors is detailed in Materials and Methods. In summary (See Table 1); Int_HS_Energy, is the difference in the sum all the energies of hotspot residues pre- and post-mutation. HSEner_PosCoop and HSEner_NegCoop are identical to Int_HS_Energy except that, in order to account for positive and negative cooperativity effects between hotspots within a hotregion, the hotspot energies are down-weighted and up-weighted accordingly to the size of hotregion they are in. CoreHSEnergy, RimHSEnergy and SuppHSEnergy, are similar to Int_HS_Energy, except that changes in hotspot energies are limited to the given region on the interface. Each of the 6 descriptors also has its coarse-grain counterpart (No_HS, HS_PosCoop, HS_NegCoop, CoreHS, RimHS and SuppHS), where only hotspot counts instead of energies are used in the calculations. Other hotspot descriptors include the change in the size of the largest hotregion (MaxClusterSize), the number of hotregions (No_Clusters), the spread of the hotspots at the interface (AVG_HS_PathLength) and Int_Energy_1 that characterizes changes in all single-point alanine mutations at the interface.
A number of hotspot predictors are investigated for the generation of hotspot descriptors, and in total 6 sets of hotspot descriptors are generated (See Table 2). These include hotspot descriptors generated from available hotspot prediction servers, KFC2a, KFC2b [30], RFHotpoint1 and RFHotpoint2 [28], along with the hotspot descriptors generated from hotspot prediction algorithms developed in this work (RFSpot, RFSpot_KFC2). Explanation of each hotspot prediction algorithm, its features, and performance comparisons can be found in Materials and Methods. In summary, KFC2a and KFC2b are SVM hotspot prediction models developed in [30] and use a combination of solvent accessibility and plasticity features. RFHotpoint1 and RFHotpoint2 are random forest models using the features from the original Hotpoint [28] hotspot predictor, but re-trained on a larger dataset from SKEMPI (Table S16 in Text S4). RFSpot is a random forest model that employs a large set of molecular descriptors and RFSpot_KFC2 adds to this feature set, features from the original KFC2a and KFC2b models. The use of multiple hotspot predictors enables us to probe consistencies and anomalies in the predictive abilities of the hotspot descriptors.
Confirming that energy estimates of single point-alanine mutations can be used to describe the effects of off-rate changes of single- and multi-point mutations not limited to alanine, we assess whether the whole set of 16 hotspot descriptors from each hotspot prediction algorithm can be combined synergistically in a model for off-rate prediction to achieve even higher correlations. A separate Random Forest (RF) regression model is trained on the 713 off-rate mutant dataset using the descriptors generated by each hotspot predictor (RFSpotOff-Rate, RFSpot_KFC2Off-Rate, RFHotpoint1Off-Rate, RFHotpoint2Off-Rate, KFC2aOff-Rate and KFC2bOff-Rate). In addition, models that add the set of 110 molecular descriptors to the hotspot descriptors (RFSpot+MolOff-Rate, RFSpot_KFC2+MolOff-Rate, RFHotpoint1+MolOff-Rate, RFHotpoint2+MolOff-Rate, KFC2a+MolOff-Rate and KFC2b+MolOff-Rate) are also built for comparison. Note that the ‘Off-Rate’ subscript is used to distinguish the off-rate predictor trained on hotspots, from the actual hotspot predictor generating the hotspot descriptors in question. The 20-Fold Cross-Validation (20-Fold CV) results are concatenated to form of a set of 713 test predictions and their PCC with Δlog10(koff) are shown in Figure 5A (See Table S1 for list of predictions for each model). The best performing off-rate predictor (RFSpot_KFC2Off-Rate, R = 0.79, see Figure 6A) combines the hotspot descriptors generated from RFSpot_KFC2 hotspot predictor and the molecular feature set. In general, the models which combine both hotspot and molecular descriptors achieve higher correlations to the hotspot descriptor models, though which on their own, the latter still achieve correlations of R>0.7 using only 16 hotspot descriptors. Off-rate models using hotspot descriptors (Figure 6A and B), have more stabilizing mutations in the lower left quadrant, and hence such mutations tend to be less underestimated, than a model using molecular descriptors (Figure 6C).
Previous analysis has been performed using models trained on all the 713 off-rate mutations in the dataset, of which the predictions were then subdivided into data regions for separate analysis. Here, descriptors, which are specific to the prediction of mutations within each data region, are investigated. To do so, separate models are built for the different data regions of the dataset using a Genetic Algorithm for Feature Selection (GA-FS) as described in Materials and Methods. All 110 molecular descriptors and 16 hotspot descriptors generated from the RFSpot_KFC2 hotspot predictor are available for feature selection. The feature set size is set to 5 features to avoid over-fitting and both non-linear (using Support Vector Machines, SVM) and linear (using Linear Regression, LR) models are investigated. For every data region, 50 separate GA-FS runs are performed; an inner-cross validation loop is used for FS (And SVM parameter optimization), whereas an outer-cross validation loop is used for testing the final model, of which the results are summarized in Figure 8E (blue and red). The GA-FS models built on rim and support region mutations achieve markedly lower correlations than core region models, though a non-linear model increases the accuracy of the latter two models. There are no notable differences in the ability to model LIA and SIA complexes; however, multi-point mutations are markedly better predicted than single-point mutations. Polar and charged mutations show good correlation which decreases when considering hydrophobic residues.
One advantage of using hotspot descriptors to estimate off-rates is the ability to localize interface regions of high stability and assess how mutations affect the distribution of stabilities, within these regions. The importance of the core interface region is implicated largely due to the tendency of hotspots to preferentially occur in this region [24]. On the other hand rim residues seem to play a more secondary role of solvent shielders by providing an ideal dielectric constant for better interactions at the core [24]. In this section we analyze hotspot energies at specific regions of the interface, namely the core, rim and support regions and evaluate whether complex stability can be effectively disrupted homogenously across the interface or preferentially in a particular region. More specifically the role of rim residues is re-investigated in the light of off-rate changes upon mutations on complexes of various sizes and interface-areas.
CoreHSEnergy, RimHSEnergy and SuppHSEnergy represent the change in total hotspot energies limited to each region upon mutation. Effectively, the PCC of these descriptors with the off-rate expresses how well changes in the given region show themselves as changes in log10(koff) - irrespective of changes in hotspot energies in any other region. Therefore, by assessing the relative PCCs of the three regions we can gauge whether a given region acts independently and dominates in its contribution to complex stability compared to other regions. Given that we have 6 instances of each hotspot descriptor, as generated per each hotspot predictor, the correlations for each descriptor shown are the mean of each descriptor's correlation under the 6 hotspot predictors. Hence results can be considered to be independent of the hotspot predictor generating the hotspot descriptors. From the PCCs of the three hotspot region specific descriptors (CoreHSEnergy |R| = 0.48, RimHSEnergy |R| = 0.20 and SuppHSEnergy |R| = 0.38), it is observed that changes in the hotspot energies at the core affect the off-rate more significantly than the rim (p<<0.01) and support region (p<0.01). Given that 355 mutations affect hotspot energies in the core region compared to 148 and 182 for rim and support regions respectively, results may however be biased. For example, if fewer events are observed at the rim region, there is less chance of the rim region playing a significant role in off-rate changes, when looking at it globally over a population of complexes as is done presently. To remove this potential bias, the subset of mutations, which affect all three regions simultaneously, is extracted and the PCC recalculated. The PCCs still suggest dominance from the core region (|R| = 0.53), more significantly than the rim region (|R| = 0.22 p<<0.01).
In this work we have shown that indeed changes in the energies of hotspots upon mutations have a direct relationship with the off-rate. More so, changes at certain regions of the interface such as the rim may affect the off-rate differently depending on its size, whereas the core is a critical stability region for complexes of a wide range of size and interface areas. Hotspots tend to cluster into tightly packed regions and the conservation of this type of organization suggests that they are important for protein-protein association [37]. The aforementioned analysis however is not performed in relation to binding free energies or off-rates for protein-protein interactions. Therefore, it is still not clear to which extent, the presence, number and size of hotregions is advantageous to complex stability. Using the hotspot descriptors and the experimental off-rates, some insights into this can be gained.
Predictions of off-rate models are analyzed separately for mutations on complexes which undergo significant backbone conformational changes. The subset of complexes for which the unbound crystal structures of the wild-type complex are available, were singled out and their I_RMSD values for backbone conformational rearrangements were extracted from [66].This subset of complexes for which unbound crystal structures are available, amounts to 17 complexes and 332 mutations. 67 mutations on 4 complexes show significant conformational changes with (I_RMSD >1.5 Å) as defined in [66], and if the threshold is lowered to (I_RMSD >1 Å), this results in 119 mutations on 6 complexes. The PCCs for the off-rate model predictions with Δlog10(koff) are shown under three conformational change categories (Figure 11). The PCC, for complexes which show little to no conformational change (I_RMSD <1.5 Å), averaged over all prediction models, shows a correlation of R = 0.86, which decreases to R = 0.58 at (I_RMSD >1 Å) and R = 0.28 at (I_RMSD >1.5 Å). Though for the latter category, RFSpotOff-Rate achieves a correlation of R = 0.43. Changes in the different models are more apparent at complexes with higher conformational changes, most notably is the discrepancy in PCC between Molecular and RFSpotOff-Rate off-rate prediction models. This discrepancy is minimal at complexes with little conformational changes, ΔR = 0.01I_RMSD <1.5 Å and increases to ΔRI_RMSD >1 Å = 0.11 and ΔRI_RMSD >1.5 Å = 0.24 for complexes with significant conformational changes. Reduction in the prediction accuracies for wild-type binding free energy prediction for complexes which undergo conformational change have also been noted [42], constituting an important challenge. Several factors may contribute to this, for example, complexes that are natively unstructured/disordered in local regions, may still remain disordered even in the bound state [67], [68]. Binding site variability has also been observed in certain complexes where the variability is not explained by experimental or procedural inaccuracies [69] and the off-rate may also be affected directly by the unbinding mechanism [70]. In all these examples, having a single snap-shot i.e. one conformational state for the complex we wish to calculate off-rate changes for, may not provide a picture comprehensive enough to predict off-rates. Methods for modeling conformational changes which in turn can be used to generate relevant snap-shots, are still one of the main limitations in current docking algorithms [71]. The generation of relevant snap-shots might also possibly involve the characterization of encounter complexes and their stability, where both the computational generation and experimental measurement of such states is still major challenge [50].
In this work we take a comprehensive look at the determinants of complex dissociation in relation to interface hotspot energies and organization. Though the ΔΔG of a mutation may manifest itself as change in the off-rate as well as the on-rate [72], several lines of evidence suggest a dominant contribution from the off-rate [44]–[46]. Using experimental values on 713 mutations, in this work we also find evidence for a stronger relationship of ΔΔG with Δlog10(koff). More importantly, our investigations show that the change in the off-rate of a protein-protein interaction can be sufficiently explained by the re-distribution of hotspot energies caused by that mutation. Hence, the ΔΔG of single-point alanine mutations, and readily available hotspot predictors, can indeed be used as a starting point for the estimation of off-rate mutations to any residue type and also multi-point mutations. Given this, the novelty in our approach is in the way we quantify the effects of a mutation on the dissociation rate of a protein-protein interaction. Namely, instead of directly calculating a number of features pre- and post-mutation, a complete computational alanine scan is performed at the interface pre- and post-mutation. Using the single-point alanine energies from the scans we generate a set of hotspot descriptors which describe both local and global changes caused by the mutation in question. These include changes in the size and distribution of hotregions, cooperative effects within hotregions and changes in localized stability regions such as the core, rim and support regions. Using these sets of hotspot descriptors and a number of computational experiments, we are able to gain new insights into the determinants of protein-protein dissociation.
The predictive ability of the hotspot descriptors, in estimating Δlog10(koff), is first assessed independent of a learning model. Emphasis is given, both to numerical estimation and detection of stabilizing mutations (Δlog10(koff)<−1). As a benchmark comparison, the performance of the hotspot descriptors is compared to a diverse set of molecular descriptors, varying from physics-based energy terms to coarse-grain and atom-based statistical potentials. Here we find consistently higher predictive abilities for the hotspot descriptors, in estimating Δlog10(koff). The results suggest that both the synergistic and distributional information within hotspot energies may be exploited to uncover the more causative changes in complex stability. More importantly, it proposes an alternative way of modeling single-point and multi-point mutations to any residue type, which is that of mapping them to functions using only alanine ΔΔG energies.
To assess the predictive abilities of hotspot descriptors when combined in learning models, several machine learning models trained on Δlog10(koff) are also investigated. The best regression model, which combines both molecular and hotspot descriptors, RFSpot_KFC2Off-Rate+Mol, achieves a PCC of 0.79 with experimental off-rates. Model predictions are also assessed on different subsets of mutations defined as data regions. The data regions enable us to identify, classes of mutations which are consistently harder to characterize, data set biases and prediction patterns. We find that core and multi-point mutations are the most accurately predicted; however, mutations at rim regions are consistently harder to characterize. In terms of the prediction of stabilizing mutations, a pattern emerges where mutants to alanine which stabilize the complex are harder to detect. To uncover relationships between different subsets of off-rate mutations and descriptors, we develop linear and non-linear feature-selection models trained on data-regions. Descriptor-data region networks generated from these models, enable us to identify descriptors highly specific to certain classes of mutations and those which are broadly important to a number of different regions simultaneously.
The results gained in this work are particularly useful from a computational design perspective. Off-rate classification models for stabilizing mutation prediction (Δlog10(koff)<−1), achieve a MCC of 0.59, which increases to 0.82 when neutral mutations are excluded. We find that hotspot descriptors which are able to capture the intricacies of off-rate changes related to the re-distribution of hotspot energies and positive cooperative effects play a key role in detecting such mutations. Secondly, we underline the importance of performing a computational alanine scan, if possible, before optimizing an interface. This presents a distributional context that one may exploit and apply mutations accordingly, and thus adopt a biomimetic design strategy mirroring that taken by evolution. For example, our results indicate that the distribution of the critical stability regions across protein-protein interfaces is a function of complex size. Though large-size complexes investigated here show more robustness to mutations than small-size complexes, here we show the insensitivity to mutations is not shared equally across all parts of the interface, as changes in the core can still significantly affect complex unbinding for large complexes. Conversely for small complexes, the increase in insensitivity to mutations is distributed homogenously across the interface, with hotspots in the rim region becoming jointly critical for complex longevity. This suggests that the accurate characterization of rim hotspots is important in the design of small complex interfaces. Further advances in characterization of off-rate mutations are likely to be achieved upon improved modeling of cooperative effects within hotregions and that of conformational changes.
Six hotspot prediction algorithms (RFSpot, RFSpot_KFC2, RFHotpoint1, RFHotpoint2, KFC2a and KFC2b) are used for the generation of hotspot descriptors, which are subsequently used for the prediction of off-rates. The method is explained in Figure 1 and requires that the hotspot predictor in question generates a prediction for each residue at the interface, both pre- and post-mutation, akin to an alanine-scan. The energies from single-point alanine mutations of the pre- and post-mutation scans are then used to calculate a set of 16 hotspot descriptors. For each hotspot predictors, its own set of hotspot descriptors is generated. The hotspot descriptors enable us to use the energies from single-point alanine mutations, particularly those which are hotspots, in order to describe the effects of off-rate mutations to non-alanine mutations and also multi-point mutations.
The prediction of hotspots is an active area of research and several hotspot prediction algorithms have been developed [25]–[35]. One short-coming of these algorithms is that they have been trained and tested on very limited alanine scanning databases, namely ASEdb [73] and BID [74]. The shortcoming of these datasets as benchmarks has been highlighted in [25], [41]. To address these limitations we recently assembled the largest database of mutations to date, with 3047 experimentally determined structures and binding kinetics, including free energy changes, dissociation/association rates and enthalpies/entropies where available [41]. All single-point alanine mutations, limited to the complex interfaces, were extracted from the SKEMPI database. This totals to a set of 635 non-redundant mutations with experimental ΔΔG in 59 different complexes and 154 hotspot residues with ΔΔG > = 2 kcal/mol (Table S16 in Text S4). All hotspots represent the positive training examples and anything, which is not a hotspot (ΔΔG <2 kcal/mol) as negative training examples.
As depicted in Figure 1, for any given complex, a computational alanine scanning is first performed on the wild-type interface using a hotspot prediction algorithm. This enables calculation of the set of hotspot descriptors described in Table 1. The respective single-point or multi-point mutation is then applied using FoldX [43], and another computational alanine scan is performed on the mutated interface, again using the same hotspot prediction algorithm invoked for the wild-type scan, from which a new set of hotspot descriptors are calculated. The energetic value contributed by each hotspot descriptor is then the difference in its energetic value pre- and post-mutations:(5)
The hotspot descriptors are calculated for a set of 713 mutations from the SKEMPI database [41]. Therefore, in total, for each hotspot prediction algorithm, we make 50 wild-type and 713 mutant computational alanine scans. To ensure that off-rate predictions are not made via hotspots models trained on the same examples, all 713 computational alanine-scans made by RFSpot, RFspot_KFC2, RFHotspoint1 and RFHotspoint2 are strictly 20-Fold-test predictions for mutations common between the off-rate and hotspot datasets, and test predictions for the rest. Therefore, all hotspot predictions on which the hotspot descriptors are calculated are unbiased and not susceptible to over-fitting. Each mutation in the 713 off-rate mutant dataset has available the experimental wild-type and mutant off-rates and the respective PDB structure. This off-rate dataset is the largest assembled to date and experimental off-rates within this set range cover a range of Δlog10(koff) of −8.5 to 6.8, with koff units of s−1 and represent a diverse set of interactions as listed in (Dataset S1).
Hotspots provide a very rich source of information, which can be exploited on many levels. Firstly, the occurrence of a hotspot is not limited to any particular physical phenomena. Instead hotspots are the result of the synergistic effect of different phenomena together. These may include evolutionary pressures, along with physicochemical and structural properties [76]. Thus, mapping all the critical points for each to an interface produces a complex distribution. However, the description of an interface though hotspots is conceptually much simpler. From a computational stand-point, the advantage is that one is able to represent an interface with a much smaller set of features without compromising accuracy, as the effects of several phenomena is still encompassed within the hotspots themselves. This reduction in feature set size is also particularly attractive in the context of learning algorithms. A second attractive attribute of hotspots is their distributional properties. Hotspots tend to cluster into hotregions, within which, hotspots are suggested to be energetically cooperative [37], [65]. It has also been shown that hotspots tend to occur more at the core regions as opposed to the rims; however, low solvent accessibility is not a sufficient property for a residue to be a hotspot [24]. Understanding how these two aspects of hotspot structure and organization, relate to the off-rate of a complex, is critical for an accurate characterization of changes in the off-rate caused by mutations. The aim of the hotspot descriptors designed here (Table 1) is therefore to present hotspots in different positional contexts, which may affect complex destabilization to differing degrees. The relevance of each descriptor to off-rate variation is then assessed with different feature importance measures and the key determinants of the dissociation process reported.
The importance of the descriptors used in this work, in relation to the dissociation rates, is assessed using three methods. The first method is the global correlation of a given descriptor with the target variable, which in this case is the experimental off-rate Δlog10(koff). To calculate this, the Pearson's Correlation Coefficient (PCC) is used. A second method is the Mann-Whitney U-test, which checks whether a set of two independent observations have smaller or larger values than the other. The test is used to assess the coarse-grain predictive power of our descriptors in discriminating between stabilizing mutants from neutral to destabilizing mutants. Several other classification related measures are used for this same purpose also, namely:
True-Positive-Rate (TPR)/Recall:
False-Positive-Rate (FPR):
Specificity:
Precision:
Accuracy:
Matthew's Correlation Coefficient (MCC):
F1-Score:
where TP = True-Positive, FP = False-Positive, TN = True-Negative, FN = False-Negative. A third method used is an assessment of descriptor importance in the context of a learning model where several of the descriptors are combined together to make a prediction. For this the built-in Random Forest Feature Importance measure (RFFI) is used [75]. Note that unlike the PCC, Mann-Whitney U-test and the above mentioned classification measures, the RFFI calculates feature importance as a function of other features in the model.
The 713 off-rate mutations from SKEMPI are also subdivided into the following data regions for analysis: Single-Point (SP) alanine mutations, 361; SP non-alanine mutations, 155; SP mutations, 516; Multi-Point (MP) mutations, 197; SP mutations to polar (Q, N, H, S, T, Y, C, M, W) residues, 39; SP mutations to hydrophobic (A, I, L, F, V, P, G) residues, 309; SP mutations to charged (R, K, D, E) residues, 68; mutations exclusively on core regions, 272; rim regions, 79; support regions, 114; mutations on complexes of Large-Interface-Area (>1600 Å2) , 355 and Small-Interface-Area (<1600 Å2), 358.
The GA-FS algorithm runs feature selection on subsets of the off-rate mutation dataset defined as data regions. Two separate GA-FS runs are performed, one for Linear Regression models and another for Support Vector Machine (RBF) Regression Models (using the LIBSVM package). Two separate 10-fold cross-validation loops are used. One to assess prediction accuracy on the off-rate mutations for the given data region and the second to derive the optimal feature set. A 10-fold inner-cross validation loop is used within the GA-FS fitness function to drive the feature selection process with reference to Pearson's Correlation Coefficients. After the GA has converged, the LR/SVM model is tested for its accuracy on the outer-loop fold. This process is repeated 10 times such that all 10 outer loop folds are used as a test set validation for the final model. Therefore, the accuracy of the final model is tested on data which is not used to derive the feature set. As an initial feature set available for selection, 110 molecular descriptors and 16 hotspot descriptors from the best performing off-rate prediction model RFSpot_KFC2 are available. A fixed feature set size of 5 is chosen so as to avoid overfitting on smaller sized data regions. Therefore, the genome size for the GS-FS (LR) is 5 whereas that for GA-FS (SVM) is 7 to also optimize the cost and gamma parameters of the RBF. Available cost parameters values are quantized into 111 bins ranging from 2−5 to 26. Gamma parameter values are quantized into 1300 bins ranging from 2−8 to 25. The GA's initial population size was set at 1000 individuals, and generated such that the initial population included at least one instance of each of the 126 features. Tournament selection was employed with a size of 8 individuals. Uniform random crossover was used with a crossover fraction set to 50% and a mutation rate that exponentially decreased as the number of generations applied increased. Note that for each data region 50 separate GA-FS runs were performed.
To assess the discriminatory power of the hotspot and molecular descriptors, the 713 off-rate mutations are partitioned into (Δlog10(koff)<−1), representing the stabilizing portion of the dataset, and (Δlog10(koff)>0), representing the neutral to destabilizing portion of the dataset (referred to as CDS1 –Classification Dataset 1). The motivations behind the thresholds of CDS1 are two-fold. Firstly, previous error estimates show that experimental noise in the data can be as high as 2kcal/mol [41], [42]. Experimental noise causes miscategorization errors when converting Δlog10(koff) from continuous values to categorical bins, and therefore, the exclusion of data-points within [−1, 0] should reduce sufficiently the number of miscategorization errors between stabilizing and neutral/de-stabilizing mutations. Secondly, being able to detect stabilizing mutations from neutral ones is an important aspect of interface design. As shown in Figure S1, the range within [0, 1] contains 43% of the data. Therefore, the removal of Δlog10(koff) within the range [−1,0] still allows a sufficient amount of neutral mutations. This data subset, results in a dataset of 501 neutral to destabilizing mutations (referred to as non-stabilizing mutations) and 31 stabilizing mutations (See Dataset S2). To further investigate the discrimination ability of the descriptors, an additional threshold satisfying |Δlog10(koff)| >1 is also investigated (Dataset S3). This dataset which removes most of the neutrals is referred to CDS2.
|
10.1371/journal.pcbi.1000993 | Spike-Timing-Based Computation in Sound Localization | Spike timing is precise in the auditory system and it has been argued that it conveys information about auditory stimuli, in particular about the location of a sound source. However, beyond simple time differences, the way in which neurons might extract this information is unclear and the potential computational advantages are unknown. The computational difficulty of this task for an animal is to locate the source of an unexpected sound from two monaural signals that are highly dependent on the unknown source signal. In neuron models consisting of spectro-temporal filtering and spiking nonlinearity, we found that the binaural structure induced by spatialized sounds is mapped to synchrony patterns that depend on source location rather than on source signal. Location-specific synchrony patterns would then result in the activation of location-specific assemblies of postsynaptic neurons. We designed a spiking neuron model which exploited this principle to locate a variety of sound sources in a virtual acoustic environment using measured human head-related transfer functions. The model was able to accurately estimate the location of previously unknown sounds in both azimuth and elevation (including front/back discrimination) in a known acoustic environment. We found that multiple representations of different acoustic environments could coexist as sets of overlapping neural assemblies which could be associated with spatial locations by Hebbian learning. The model demonstrates the computational relevance of relative spike timing to extract spatial information about sources independently of the source signal.
| There is growing evidence that the temporal coordination of spikes is important for neural computation, especially in auditory perception. Yet it is unclear what computational advantage it might provide, if any. We investigated this issue in the context of a difficult auditory task which must be performed quickly by an animal to escape a predator: locating the source of a sound independently of the source signal. Using models, we found that when neurons encode auditory stimuli in spike trains, the location-specific structure of binaural signals is transformed into location-specific synchrony patterns. These patterns are then mapped to the activation of specific neural assemblies. We designed a simple neural network model based on this principle which was able to estimate both the azimuth and elevation of unknown sounds in a realistic virtual acoustic environment. The relationship between binaural cues and source location could be learned through a supervised Hebbian procedure. The model demonstrates the computational relevance of relative spike timing in a difficult task where spatial information must be extracted independent of other dimensions of the stimuli.
| Animals must be able to rapidly estimate the location of the source of an unexpected sound, for example to escape a predator. This is a challenging task because the acoustic signals at the two ears vary with both the source signal and the acoustic environment, and information about source location must be extracted independently of other causes of variability. Psychophysical studies have shown that source localization relies on a variety of acoustic cues such as interaural time and level differences (ITDs and ILDs) and spectral cues [1]. At a neuronal level, spike timing has been shown to convey information about auditory stimuli [2], [3], and in particular about source location [4], [5]. Although it is well accepted that ITDs can be extracted from phase-locked responses, it is unknown how information beyond this could be extracted from the spike timing of neurons. In addition, the potential computational advantages of a spike timing code in this task are unclear.
The sound S produced by a source propagates to the ears and is transformed by the presence of the head, body and pinnae, and possibly other aspects of the acoustic environment (such as reflections). It results in two linearly filtered signals FL*S and FR*S (linear convolution) at the two ears, where the filtering depends on the relative position of the head and source. Because the two signals are obtained from the same source signal, the binaural stimulus has a particular structure, which is indicative of source location. When these signals are transformed into spike trains, we expect that this structure is transformed into synchrony patterns. Therefore, we examined the synchrony patterns induced by spatialized sounds in neuron models consisting of spectro-temporal filtering and a spiking nonlinearity, where binaural signals were obtained using a variety of sound sources filtered through measured human head-related transfer functions (HRTFs). We then complemented the model with postsynaptic neurons responding to both sides, so that synchrony patterns induced by binaural structure resulted in the activation of location-specific assemblies of neurons. The model was able to precisely encode the source location in the activation of a neural assembly, in a way that was independent of the source signal.
Several influential models have addressed the mechanisms of sound localization at an abstract level [6]–[9]. Considerable progress has also been realized in understanding the physiological mechanisms of cue extraction, in particular neural mechanisms underlying ITD sensitivity [10]–[14]. These studies mostly used simplified binaural stimuli such as tones or noise bursts with artificially induced ITDs. Several purely computational models [15]–[17] address the full problem of sound localization in a virtual acoustic environment with realistic sounds, although these do not suggest how neurons might perform this task. Here we propose a binaural neural model that performs the full localization task in a more realistic situation, based on the idea that synchrony reflects structural properties of stimuli, which in this setting are indicative of source location.
Consider a sound source located at azimuth θ and elevation φ. The signal S(t) arrives at the two ears as two linearly filtered signals SL = HRTFL(θ,φ)*S and SR = HRTFR(θ,φ)*S (Figure 1, A). Other aspects of the acoustic environment such as reflections and distance would also impact the binaural signal, but their effect can always be expressed with linear filters. In general, the signals at the two ears are filtered versions of the source signal, where the filters are determined by the relative positions of the head and source in the acoustic environment. What are the correlates of this filtering at neural level? Let us consider a neuron A which responds to sounds from the left ear. A simplified way to model its response is to consider that the sound is transformed into spike trains after filtering through the neuron's spectro-temporal receptive field NA (Figure 1, B–C), that is, the filtered signal NA*SL = NA* HRTFL(θ,φ)*S is followed by a spiking nonlinearity (Figure 1, D–E). Since we are interested in precise spike timing, we consider that the spiking nonlinearity is represented by a neuron model, e.g. integrate-and-fire or more complex models, rather than by a Poisson process. The response of a neuron B to sounds from the right ear would similarly be modeled as spike trains produced from the signal NB*SR = NB* HRTFR(θ,φ)*S.
None of these individual neurons expresses spatial tuning, but we now consider the pair of neurons A and B and ask ourselves when these two neurons fire in synchrony. More precisely, we define the synchrony receptive field of this neuron pair as the set of stimuli that induce synchronous firing in these neurons. Synchrony occurs when the input signals to the two neurons match, that is, when the following identity is met: NA* HRTFL(θ,φ) = NB* HRTFR(θ,φ) (Figure 1, left column). In the left column of Figure 1, the synchrony receptive fields of pairs of monaural neurons contain the presented location, while in the right column they do not. This identity expresses the condition that the combinations of acoustical and neural filtering match on both sides. Thus, the synchrony field is defined independently of the source signal S, as long as the signal contains energy in the neurons' receptive fields. It is a set of pairs of acoustical filters (HRTFL, HRTFR), which defines a spatial field: synchrony between the two neurons signals spatial information independently of the source signal. For example, if neuron A has receptive field NA = HRTFR(θ*,φ*) and neuron B has receptive field NB = HRTFL(θ*,φ*), then the synchrony field of the pair contains the location (θ*,φ*). This example corresponds to a recent signal processing method designed by MacDonald [16], which was found to be very accurate in estimating the azimuth of a sound source. The same holds true if the receptive fields are band-pass filtered in the same frequency band, i.e., NA = K*HRTFR(θ*,φ*) and NB = K*HRTFL(θ*,φ*) (Figure 1, D–E). Thus, any given location will elicit a specific pattern of synchrony in a way that is independent of the signal S. Now consider a postsynaptic neuron that receives inputs from these two neurons A and B: if it is sensitive to the relative timing of its inputs then it will fire preferentially when the two inputs are synchronous, that is when the stimulus is in the synchrony receptive field of its inputs (Figure 1, F–G).
To better understand how the neural pattern of synchrony encodes source location, consider that each signal or filter in the processing chain is described by a set of columns of various heights, in the same way as a graphic equaliser (Figure 2): each column represents the level or phase of an individual frequency component. In Figure 2, signals are in pink, acoustical filters (HRTFs) in green and neural filters (receptive fields) in blue. Combining two filters (or filtering a signal) corresponds to adding each column of the first filter on top on the corresponding column of the second filter. The first two columns illustrate the case when the source location X is not in the synchrony receptive field of the neuron pair (A, B): when the signal is combined with the left HRTF and with the receptive field of neuron A, it does not match the combined signal on the other side (combination of signal, right HRTF and receptive field of neuron B). Thus neurons A and B do not fire in synchrony. On the other hand (next two columns), when location Y is presented, the two signals match and the neurons fire in synchrony, which can make a postsynaptic binaural neuron fire. From this illustration, it clearly appears that the two signals would also match if the signal S (pink cubes) were different, and that the synchrony receptive field contains more than a single pair of HRTFs – therefore higher spatial selectivity requires several different neuron pairs. The next two columns illustrate the synchrony receptive field of two other neurons C and D, which contains location X but not Y. Thus location X induces synchrony between A and B, while location Y induces synchrony between C and D. Therefore the pattern of synchrony indicates the location of the sound source, independently of the source signal S. This idea of binaural matching was recently implemented in a signal processing method to estimate the azimuth of a sound source [16], with excellent performance. In that algorithm, the set of “neural filters” corresponds to all possible HRTFs (no band-pass filtering) and the “synchrony pattern” is reduced to a single neuron pair, where synchrony and coincidence detection are replaced by maximum correlation. On the other hand, if neural filters consist only of gains and delays, then the binaural matching would correspond to the Equalisation-Cancellation model [18]–[20], where synchrony and coincidence detection are replaced by cancellation.
A given source location, then, will activate a specific assembly of postsynaptic neurons – all those neurons for which the synchrony field of their inputs contains that location – so that source locations are mapped to the activation of (possibly overlapping) neural assemblies (Figure 3A). To test this principle, we simulated a virtual acoustic environment using measured HRTFs and implemented a spiking network model which responded to the binaural signals as described above (Figure 3). A variety of short sounds (noise bursts, musical instruments, voices, tones; Figure 4) was filtered through pairs of human HRTFs to reproduce the natural acoustical filtering of sounds due to the presence of the head, body and pinnae. Neural filtering was modeled as band-pass filtering followed by some additional linear transformations (Figure 3B). These were either all the transformations that we may possibly need to represent all locations, i.e., the complete set of HRTFs (the ideal model), or only delays and gains (the approximate model). The ideal model should not be taken to imply that the auditory system actually implements HRTF filtering, which would be physiologically unreasonable, but that the neural receptive fields may correspond to band-pass filtered HRTFs. Figure 5 shows 12 examples of such neural filters, which look very similar to gammatone filters, except for slight changes in their envelopes. Nevertheless, these filters may be too diverse to be represented in the auditory system, which motivated the approximate model: in narrow frequency bands, filters can be well approximated by a more restricted set including a range of gains and delays. The resulting signals were then transformed into spike trains with noisy spiking models (mostly integrate-and-fire neurons, with more complex models in one case). Finally, postsynaptic neurons received inputs from two monaural neurons, and were modeled in the same way (Figure 3B). For each source location, we assigned a neural assembly by selecting all the neurons for which the synchrony field of the inputs contained that location, making one neuron per frequency channel (Figure 3C,D). The output of the model was the assigned location of the maximally activated assembly (Figure 3E), and the goal was to predict the actual location of the source. The ideal model is conceptually close to a recent signal processing method designed by MacDonald [16], while the approximate model resembles the Equalisation-Cancellation model [18]–[20], although these two techniques are not neuron models (and the former was only applied on the broadband signal rather than in multiple frequency bands). We describe their relationship with our model in more detail in the Discussion.
Figure 6B and C show the activation (total spike count) of all location-specific neural assemblies in the ideal model for two particular sound presentations, at locations indicated by black crosses. In both examples, the assigned location of the maximally activated assembly (indicated by white crosses) is indeed the actual source location. Although in these figures we represented model outputs on a map, this topographical representation is not present in the model itself. Figure 6D–F shows the activation of three particular neural assemblies as a function of source location. It appears that these assemblies are spatially tuned in that they fire more when the source is at their assigned location. The spatial receptive fields of these assemblies also reveal ring-like structures: these correspond to the cones of confusion [1], where the distances to the two ears are constant, so that interaural cues are very similar (the rings correspond to circles around the approximate symmetry axis that goes through the two ears). If the head were perfectly spherical, there would be no way to distinguish between these locations, but the model does so thanks to the irregular shape of the head and to the presence of the body (see also Figure 7).
Quantitatively, for the ideal model with 80 frequency channels, the average estimation error for white noise was almost zero degrees for azimuth (Figure 6G) and 1.5 degrees for elevation (Figure 6H). Other types of sounds did not have as much power in all frequency bands (Figure 4), and consequently the estimation error was larger for these sounds, but remained very small for speech (1 degree azimuth and 3 degrees elevation) and musical instruments (2 degrees azimuth and 6 degrees elevation), even though they were not used to build the model. This performance for previously unknown sounds is the key property we were expecting to see in the model, as the synchrony patterns are location-specific and independent of the source signal. The error was substantially higher for pure tones, particularly for elevation (6 degrees azimuth and 28 degrees elevation, which is close to the chance level of 36 degrees given the distribution of source locations). This is not surprising since for high frequencies the ITD cues are ambiguous due to periodicity, and for low frequencies ILD cues are very weak, giving only one dimension in the binaural cues.
Not surprisingly, the model could correctly categorize the sound as coming from the left or right (100% success rate for all types of sounds), but it also performed well in more difficult categorization tasks: discriminating between front and back (70% for pure tones and 90–98% for other sounds), and between up and down (64% for pure tones and 93–98% for other sounds), as shown in Figure 6I. This indicates that different neural assemblies were activated for front and back locations and that the model was thus able to exploit small specific interaural differences, as we show below.
It might be surprising that the model can estimate elevation and even discriminate between front and back while it only uses binaural cues. Figure 7 explains how across-frequency integration (i.e., looking at the spatial tuning of neural assemblies rather than that of individual neurons) allows the model to estimate both azimuth and elevation. The mechanism is illustrated with ITD cues but the same holds for ILD cues. When the ITD is estimated in a fine frequency band, it varies with frequency for the same source location, because of sound diffraction by the head [21]. Specifically, it is larger at lower frequencies. Because the head is not spherical, this frequency-dependence of ITDs is location-specific. When the ITD is observed is a single frequency band, it is consistent with many possible locations (Figure 7A, solid curve), because two dimensions (azimuth, elevation) are mapped to a single one (ITD). In a spherical head model, these possible locations form the cone of confusion [1]. When the ITD is observed in another frequency band, it is also consistent with a whole set of locations (Figure 7B, dashed curve), but this set is frequency-dependent. Therefore if the sound contains the two frequency components, the intersection of the two sets of possible locations is a single point corresponding to the true location of the source (Figure 7C). In other words, if source location is two-dimensional, it can be estimated using two independent observations. Pure tones cannot benefit from this disambiguation, consistently with the poor performance seen in Figure 6G–H (magenta bars). There is experimental evidence that humans can indeed use binaural cues to estimate elevation [22], which we comment on in the Discussion.
It might be unrealistic to assume that the receptive fields of auditory neurons have so much diversity as to include all possible HRTFs. Besides, it is not straightforward to see how the auditory system could learn these acoustical filters. To address this issue, we tested an approximate model in which neural filtering consisted only of band-pass filtering with various gains and delays, which could be produced by many mechanisms: axonal or dendritic propagation, inhibition, voltage-gated conductances, etc. Within fine frequency bands, such simple transformations can approximate but not completely match HRTF filtering (Figure 8A) and therefore performance might be expected to drop. However, we found that the location-dependent responses of neural assemblies were very similar to those seen in the ideal model, but slightly less specific (Figure 8B–D). Quantitatively, estimation errors were still small, although not as much as in the ideal model (2 to 7 degrees azimuth and 7 to 20 degrees elevation, excluding pure tones; Figure 8 G–I and Supplementary Figure S1). In this approximate model, it might be thought that binaural neurons perform a cross-correlation of delayed monaural inputs, as in the classical Jeffress model [6]. However, because the inputs to these neurons are precise spike trains rather than Poisson processes, the operation that they perform is more accurately described as a similarity operation (firing when the inputs are similar) than as a cross-correlation. In particular, this operation includes level differences as well as timing differences. We comment on this issue in the Discussion.
In the approximate model, neurons in a location-specific assembly can be fully described by their preferred frequency, interaural delay and gain difference (in log scale). Figure 8E shows the preferred delay of neurons as a function of their characteristic frequency for two neural assemblies tuned to the same location but with front and back reversed, and Figure 8F shows the preferred interaural level difference as a function of frequency. It appears that the preferred delay is approximately constant at high frequencies but irregular at low frequencies (<1 kHz). By construction, the preferred delay corresponds to the ITD in the neuron's frequency channel at these locations, as measured as the peak of crosscorrelation of the band-passed filtered HRTFs. These frequency-dependent patterns are consistent with previous measurements in HRTFs [21], [23] and with theoretical predictions: for high frequencies, acoustical waves behave as light rays and the ITD is determined by the difference in the shortest paths from the source to the ears, but for low frequencies (with wavelength larger than the size of the head) sound propagation is governed by diffraction, which predicts larger and frequency-dependent ITDs [21]. The interesting consequence for sound localization is that the frequency-dependent pattern of ITDs is location-specific, and is therefore a cue for both azimuth and elevation, and can also be used to discriminate between front and back, as is illustrated in Figure 7. Psychophysical studies showed that, indeed, head diffraction and torso reflections provide elevation cues even when pinnae cues are absent [22].
The estimation error decreased as the number of frequency channels in the model was increased (Figure 8, J–L). For example, for white noise, the estimation error in azimuth was halved using 240 channels instead of 80 (Figure 6A). Except for pure tones, the performance did not seem to have converged to an asymptotic value, so we expect the error to be even smaller with more channels. A human cochlea has 3,000 inner hair cells, of which around 1800 have characteristic frequencies between 150 Hz and 5 kHz. For pure tones, the performance did appear to be approaching an asymptotic value, which is not surprising as there are limitations in the available acoustical cues.
In many previous studies, possible source locations were constrained in the horizontal plane (for example [16]). For comparison, we show in Supplementary Figure S2 the estimation error of the model in this case (with 80 frequency channels). The performance was significantly better, especially for the ideal model, which made zero errors except for pure tones (which provide ambiguous information at high frequencies).
The model relies on selective synchronization and sensitivity to synchrony, which might require specific neural properties, such as low intrinsic noise and short membrane time constant. Figure 9A shows how the estimation error depends on the level of intrinsic neuronal noise in the model. It appears that the performance in azimuth estimation is very robust to noise, and that elevation estimates are reasonably accurate with a noise level up to about 2 mV (standard deviation of the membrane potential). Since the model must be able to resolve submillisecond differences in spike timing, we expected that the membrane time constant of neurons should be small. The results we previously showed were obtained with a membrane time constant of 1 ms, and Figure 9B shows that the model gave reasonable estimates up to about 4 ms, which is not unreasonably short for auditory neurons [24]–[26]. Since the models did not explicitly include voltage-gated conductances [27] and coordinated inhibition [28], which both shorten the integration time constant, this value should be understood as the effective time constant that accounts for these effects. When the time constant was larger than 10 ms, performance was close to chance level, which indeed confirms that the model relied on precise spike timing.
In all previous figures, the performance of the model was only tested in quiet, although physiological noise was included in the neural models. In Figure 9C, we added uncorrelated white noise to the left and right signals and we tested how accurately the approximate model could localise white noise (500 ms). It should be stressed that source locations were not constrained to the horizontal plane, but that the model was nonetheless robust to moderate levels of distracting noise.
All previous results were obtained with simple integrate-and-fire models. However the model is based on the principle that neurons synchronize when their inputs are similar, which should be mostly independent of the specific way in which these inputs are transformed into spike trains. We checked this idea by replacing the neuron models by adaptive exponential integrate-and-fire models tuned to regular spiking cortical cells [29]. This model can predict the spike trains of cortical neurons in response to somatic time-varying current injection [30] and includes two major features of cortical pyramidal cells: spike-frequency adaptation and realistic spike initiation [31]. As we expected, the model was still able to accurately estimate source location, with quantitatively similar results (Figure 9D–F).
In the model, source location is indicated by the activation of a specific neural assembly. Thus, estimating the source location requires that each physical location has been assigned to a neural assembly. We suggest that this assignment could be obtained through Hebbian learning, for example by associating neural activation with visual cues. In Figure 10, we show the response of a population of postsynaptic neurons with various preferred frequencies, interaural delays and gains to a long broadband sound (20 s) played at a particular location. Picking the maximally active neuron in each frequency channel (white crosses) defines a neural assembly that is indeed very close to the choice we previously made from the knowledge of HRTFs (black crosses). We estimated the performance of the model when the location-specific assemblies were learned in this way from 7 seconds of white noise played at each location (Figure 10, C–E). The error in azimuth, elevation and categorization was only slightly worse (compare with Figure 8, G–I). Note that the training data consisted of white noise while the test data included different types of sounds (speech and musical instruments).
Since source location is encoded in the identity of (possibly overlapping) neural assemblies, learning consists in assigning a correct label to an assembly rather than in tuning parameters. Therefore, the same network can encode several acoustic environments: it is only the mapping from assembly activation to physical location that is environment-specific. Humans can learn this mapping when their acoustical cues change, for example when molds are inserted into their ears [32]–[34] (although these experiments mainly modify spectral cues rather than binaural cues). Interestingly, although learning a new mapping can take a long time (several weeks in the first study), the previous mapping is instantly recovered when the ear molds are removed, meaning that the representations of the two acoustical environments do not interfere, consistently with our model. We tested this idea by using two different sets of HRTFs (corresponding to two different human subjects) with the same network model. Figure 11A–D shows that the same source location activates two different neural assemblies depending on the HRTF set. We defined two mappings from neural activation to physical location, one for each HRTF set, by associating a neural assembly with each location, as previously. When a sound was presented through a particular HRTF set, it maximally activated a neural assembly assigned to the correct HRTF set at the correct location (Figure 11E), so that the model was still able to accurately estimate the source location (Figure 11F), as well as to identify the acoustical environment.
The acoustical transformation of a sound between source and ear is a linear filter that depends on their relative positions. The binaural stimulus resulting from the two ears receiving different filtered signals of the same source has a structure that is indicative of source location. We looked for correlates of this binaural structure in neuron models described by spectro-temporal filtering and spiking nonlinearity. We found that the synchrony receptive field of a pair of monaural neurons, defined as the set of stimuli that induce synchronous spiking in these neurons, defined a set of source locations (more precisely, filter pairs) independently of source signal. This is a very interesting property for source localization because an animal must be able to estimate the location of an unexpected sound. These location-specific synchrony patterns are then mapped to the activation of location-specific assemblies of postsynaptic neurons. The spiking model we implemented was indeed able to accurately locate a variety of sound sources in a virtual acoustic environment, and was robust to significant changes in model properties, including changes in the neuron models themselves, and to both physiological and acoustical noise. It demonstrates the computational relevance of relative spike timing to extract spatial information about sources independently of the source signal. In the model, source location is encoded into overlapping neural assemblies, and the mapping from neural activation to source location could be learned by Hebbian mechanisms (for example by association with visual cues). Because only this latter mapping depends on the acoustical environment, the model could simultaneously store several representations of different environments and both accurately estimate source location and identify acoustical environment.
Our model only uses binaural cues, while it is known that the dominant cue for elevation is monaural spectral information in high frequencies [1], [35]. It may be surprising that the model can estimate elevation with only binaural cues, but psychophysical results show that human subjects can extract information about elevation from signals that are low-pass filtered below 3 kHz, where monaural spectral information is minimal [22], except in the median plane. Indeed, for any elevation in the median plane, the left and right signals are close to identical, therefore binaural cues do not provide information about elevation. Interestingly, we find the same pattern in our model (Figure 12A): performance in estimating elevation is close to chance level in the median plane. The azimuth error was also larger away from the median plane (Figure 12B), which is consistent with psychophysical experiments (see Figure 4A in [36]). Since the model does not use high-frequency monaural spectral information, it might be more relevant to compare psychophysical results with our model performance when both are constrained to the low-frequency range. Supplementary Figure S3 shows the performance of the approximate model when frequency ranges from 150 Hz to 3 kHz (as in reference [22]), which is close to the results shown earlier (Figure 8G–I). The same estimation error patterns were also seen in this case. It could be argued that the model indirectly uses monaural cues in the form of elevation-dependent spectral notches, but it only does so by comparing the two binaural signals rather than extracting spectral information from each monaural signal, which explains why it cannot estimate elevation in the median plane.
Previous neural models focused on the mechanisms of cue extraction such as ITD sensitivity, using delayed tones or noise bursts (as in e.g. [13]), whereas we addressed the problem of estimating the location of arbitrary unknown sounds in a realistic acoustical environment. Since our model relies on coincidence detection between monaural inputs, it could be compared to the Jeffress model [6], where a neuron is maximally activated when acoustical and axonal delays match - although the Jeffress model is restricted to azimuth estimation. However, because in our model the inputs to these binaural neurons are precise spike trains rather than Poisson processes, the operation that they perform is more accurately described as a similarity operation (firing when the inputs are similar) than as a cross-correlation. In particular, this similarity operation includes level differences as well as timing differences. Indeed, there is considerable difficulty in implementing the Jeffress model with neuron models when realistic acoustical cues are considered, because ILDs always co-occur with ITDs and disturb spike timing. Figure 13A–C shows that model performance indeed drops when neural filtering is restricted to band-pass filtering and delays (no gains). In our model, the sensitivity of binaural neurons to ILDs comes from the fact that monaural neurons fire earlier as sound level increases, which is unavoidable if spikes are triggered when a threshold is reached. This “time-intensity trading” has been observed in the auditory nerve [37]: the effect is small but within the sensitivity of the binaural system. It has not yet been measured in bushy cells, which would correspond to the monaural neurons in our model (see below).
Several computational algorithms address the full problem of sound localization in a virtual acoustic environment with realistic sounds, but without a neural implementation. In our model, two monaural neurons fire in synchrony when the combinations of acoustical and neural filtering match on both sides. This is conceptually similar to the Equalisation-Cancellation (EC) model [18]–[20], in which compensating interaural gain and delay are chosen so that the two signals maximally match (i.e., by minimizing the difference). In the original EC model, this was done for the broadband signals but later versions of the model used multiple frequency bands [15]. Besides the fact that our model has a straightforward interpretation in terms of neural responses, we highlight two conceptual differences. Firstly, the signal transformations are not restricted in principle to delays and gains (as in our approximate model), but could include any sort of filtering (as in the ideal model). Although the performance of the approximate model was good, the ideal model was more accurate, in fact almost perfect when locations where restricted to the horizontal plane. When considering more complex environments (with reflections), the difference could be even more important. Secondly, our model provides an online, instantaneous estimation of source location, which could potentially change if the source moves. Additionally, the fact that the spike trains of the location-specific assemblies are locked to the signal may be useful to bind the spatial information with other types of information, for example to listen to a source at a particular location, or to integrate multimodal information (e.g. binaural information with visual motion). Recently, MacDonald [16] proposed a signal processing method (with no neural implementation) to localize sound sources in the horizontal plane that is conceptually very similar to our ideal model, where coincidence detection is replaced by Pearson correlation between the two transformed monaural signals. Interestingly, the estimation was found to be quite robust to background noise. Our model provides a neural implementation and works in multiple frequency channels rather than on the entire signal. It also provides signal-locked estimations. Perhaps more importantly, our approximate model provides a framework for learning to localize without explicit knowledge of HRTFs.
The model we implemented includes the following components: acoustical environment, auditory periphery, and central neurons.
First, we modeled acoustical stimuli by using a variety of recorded sounds (noise, speech, musical instruments), and the acoustical environment was reproduced using human HRTFs, measured in anechoic conditions. This includes the diffraction of sounds by the head, pinnae and torso, which is much more realistic than using fixed ITDs and ILDs: even without considering the high frequency spectral notches introduced by the pinnae, ITDs are frequency-dependent for a given source location [21]. We also checked that model performance was robust to additional acoustical noise (Figure 9C). Thus, it can be considered as a reasonably realistic reproduction of the acoustical environment of humans in anechoic conditions. A more realistic model would include reflections, at least from the ground. If the delay between the direct sound and the reflection is large, then psychophysical studies suggest that the reflected sound is suppressed by the auditory system [38], but it would presumably require specific mechanisms, which we did not address. If the delay is short, it could change the binaural cues but because the physical laws of sound propagation are linear, their effect could still be modeled as location-dependent linear filtering, simply with different HRTFs than in the anechoic case. Therefore it should not impair the performance of the model, as long as the acoustical environment is known.
Our model of the auditory periphery is rather simple, compared to recent detailed models of auditory nerve fibers [39]. The main reason is practical. In the model we simulated up to 240 channels, with various sounds, each one played at all measured locations (azimuth and elevation, almost 200 locations), totaling up to 90,000 filters and 106 neurons. Simulations took several days even after being accelerated using graphics processing units (GPUs). Using more realistic models is possible in principle, but would require faster hardware or substantially improved numerical techniques. The second reason is to keep the model simple enough to clearly demonstrate the underlying principle. Nonetheless, it includes the following ingredients: outer ear filtering (implicitly included in the HRTFs), band-pass filtering, compression, half-wave rectification, physiological and acoustical noise and spiking. Middle ear filtering was not included, but since it affects both ears equally (and therefore does not affect interaural differences), it should not have any impact on the performance of our model, which we checked with white noise stimuli (not shown). To generate spikes from the filtered signals, we used noisy neuron models (integrate-and-fire or more complex models in Figure 9D–F) rather than Poisson processes. These models implicitly include low pass filtering of the input signal (via the leak current). In Figure 13, we checked that the model also worked if sounds only contained frequencies below 3 kHz, where the temporal fine structure of sounds is still represented in the firing of auditory nerve fibers (we also found similar results for sounds below 1.5 kHz). To decompose sounds into frequency bands, we used gammatone filters. Other filters could be used, such as gammachirps, but the principle of the model does not rely on these details. Finally, the model included strong nonlinearities (half-wave rectification and compression) but they did not seem to affect the performance of the model.
There were two types of neurons in our model: monaural neurons and binaural neurons. In our description of the synchrony receptive field, we considered that the responses of monaural neurons consist of linear filtering followed by spiking nonlinearity. While this is clearly an approximation, it seems reasonable for the earliest neural structures in the auditory system. In the specific model we implemented, the strong nonlinearities in signal filtering (half-wave rectification and compression) did not seem to affect the principle of the model. We found that model estimations were still accurate when postsynaptic potentials were as long as about 4 ms, which is consistent with electrophysiological measurements of neurons in many structures in the auditory system [24]–[26]. The model was also robust to rather large levels of intrinsic noise (Figure 9A). However two assumptions restrict the set of candidate neural structures where these neurons could reside: neurons should be mainly monaural and their firing should be precisely time-locked to the stimulus. Most likely candidates are neurons in the cochlear nucleus, such as bushy cells. These cells are indeed essentially monaural and their spikes are precisely time-locked to sound stimuli [40], [41]. In the approximate model, we assumed that the receptive field of these neurons can be modeled as a band-pass filter (gammatone) with various gains and delays. Differences in input gains could simply arise from differences in membrane resistance, or in the number and strength of the synapses made by auditory nerve fibers. Delays could arise from many causes: axonal delays (either presynaptic or postsynaptic), cochlear delays [42], inhibitory delays [43]. In the ideal model, we assumed a larger diversity of receptive fields, in fact we assumed that all combinations of HRTFs and band-pass (gammatone) filters were represented, which might seem unrealistic. However, it does not mean that HRTFs themselves are represented in the auditory system. As is seen in Figure 5, these combined filters look very much like gammatone filters, but with variable envelopes. The variability of receptive fields of bushy cells could perhaps be characterized using reverse correlation techniques.
Spherical bushy cells project to binaural neurons in the medial superior olivary nucleus (MSO) [44]. Thus it seems natural to identify the binaural neurons in our model with these cells (the inferior colliculus (IC) and the dorsal nucleus of the lateral lemniscus (DNLL) also contain neurons with similar properties). In small mammals (guinea pigs, gerbils), it has been shown that the best phases of binaural neurons in the MSO and IC are scattered around ±π/4, in constrast with birds (e.g. barn owl) where the best phases are continuously distributed [45], [46]. In larger mammals such as cats, best IPDs in the MSO are more continuously distributed [47], with a larger proportion close to 0 (Figure 18 in [47]). It has not been measured in humans, but the same optimal coding theory that predicts the discrete distribution of phases in small mammals predicts that best delays should be continuously distributed above 400 Hz (80% of the frequency channels in our model). Figure 14 shows the distribution of best phases of binaural neurons in the approximate model as a function of preferred frequency (nearly identical results were obtained for the ideal model). It appears that the distribution is consistent with these predictions in humans (Figure 3 in [12]). However, one fact that is in contradiction with our model is that the best delays in both birds and mammals (including humans, based on fMRI studies [48]) are almost always smaller than half the characteristic period, i.e., they are within the π-limit. To check whether this was a critical element of our model, we estimated the performance of the approximate model when all best delays were constrained to the π-limit, and we found that it was essentially unchanged (Supplementary Figure S4). This is not very surprising since best delays above the π-limit are mostly redundant.
The HRTFs used in our virtual acoustic environment were recorded at a constant distance, so that we could only test the model performance in estimating the azimuth and elevation of a sound source. However, in principle, it should also be able to estimate the distance when the source is close (when the source is far, binaural cues are not informative of distance because the sound wave becomes a plane wave). A more difficult problem is that of reflections. In principle, our framework applies equally well to any acoustical environment, whether anechoic or not, but the mapping between neural assemblies and physical location must be known, meaning that the acoustical environment must be familiar. To estimate the location of a source in an unknown environment, one possibility would be to isolate the direct sound from the reflections, but this requires additional mechanisms, which probably underlie the precedence effect [38]. Finally, a challenging question is how the auditory system might perform this task in the presence of multiple sources or ambient noise [49], or use localization cues to listen to a particular source [50], [51]. In our model, binaural neurons respond only when their inputs receive consistent signals, so that spectro-temporal regions where noise dominates the signal should be ignored. Thus, we suggest that our model could address this more challenging task by “listening in the dips” of the noise to extract reliable information, in the same way as humans are thought to understand speech in noisy environments [52].
All programming was done in the Python programming language, using the “Brian” spiking neural network simulator package [53]. Simulations were performed on Intel i7 Core processors with dual NVIDIA GTX295 graphics processing units (GPUs). Linear filtering was carried out in parallel on the GPUs with a custom algorithm designed for large filterbanks (around 30,000 filters in our simulations, or 90,000 in the simulations for Figure 8, J–L), reducing computation times for each sound from hours to minutes. The largest model (Figure 10) involved approximately one million simulated neurons. The overall structure and architecture of the model is illustrated in Figure 3.
Sound sources used were: broadband white noise; recordings of instruments and voices from the RWC Music Database (http://staff.aist.go.jp/m.goto/RWC-MDB/); recordings of vowel-consonant-vowel sounds [54]; and pure tones between 150 Hz and 5 kHz, uniformly distributed in ERB scale (Figure 4). All sounds were of 500 ms duration, cut to this length in the case of the VCVs (from around 600ms), and repeated twice in the case of the instruments (of length 250ms), and were presented at 80 dB SPL. Sounds were filtered by head-related impulse responses (HRIRs) from the IRCAM LISTEN HRTF Database (http://recherche.ircam.fr/equipes/salles/listen/index.html). HRIRs from this and other databases do not provide sufficiently accurate timing information at frequencies below around 150Hz, and so subsequent cochlear filtering was restricted to frequencies above this point.
Head-filtered sounds were passed through a bank of fourth-order gammatone filters with center frequencies distributed on the ERB scale, modeling cochlear filtering [55], [56]. Additional linear filters were then applied: either the entire set of head filters (ideal model) or only gains and delays (approximate model). 80 channels were used in all models except for Figure 8, J–L, in which 240 channels were used. Center frequencies were chosen from 150Hz to 5kHz in all models except for Supplementary Figure S3, in which an upper limit of 3kHz was used.
For Figure 8C–F, an AdEx neuron was used. We used the equations and parameters for a regular spiking AdEx neuron from [29], with an additional white noise current (). The leak conductance was adjusted so that the membrane time constant was 1 ms, as previously (giving gL = 281 nS). The acoustical input to the encoder neurons was also scaled to provide the same input level as before, relative to threshold. The standard deviation of the noise and the synaptic weights were doubled so as to represent the same proportion of the distance between threshold and rest as in the previous model (σ = 2 mV and W = 10 mV). Finally, the model had the same refractory properties as in the previous model.
For a given location and frequency channel with corresponding HRIRs L and R (after gammatone filtering), the gains (gL, gR) and delays (dL, dR) of the two presynaptic monaural neurons were chosen to minimize the RMS differencesubject to the conditions max(gL, gR) = 1, dL≥0 and dR≥0. The RMS difference is minimized when the delays correspond to the maximum of the cross-correlation between L and R, , so that C(dR−dL) is the maximum, and .
Each location is assigned an assembly of coincidence detector neurons, one in each frequency channel. Each of these neurons has a pair of presynaptic neurons for which the synchrony field contains the given location (see Results). In the approximate model, this is obtained by selecting appropriate gains and delays for the presynaptic neurons as explained above; in the ideal model, the filters of the presynaptic neurons are the gammatone-filtered HRTFs for the given location. When a sound is presented to the model, the total firing rate of all neurons in each assembly is computed. The estimated location is the one assigned to the maximally activated assembly.
In the model of learning shown in Figure 10, location-specific assemblies are learned by presenting unknown sounds at different locations to the model, where there is one coincidence detector neuron for each choice of frequency, relative delay and relative gain. Relative delays were uniformly chosen between −0.8ms and 0.8ms, and relative gains between −8 dB and 8 dB uniformly on a dB scale. In total 69 relative delays were chosen and 61 relative gains. With the 80 frequency channels, this gives a total of roughly 106 neurons in the model. When a sound is presented at a given location, we define the assembly for this location by picking the maximally activated neuron in each frequency channel, as would be expected from a Hebbian learning process.
|
10.1371/journal.pcbi.1004241 | Synaptic Homeostasis and Restructuring across the Sleep-Wake Cycle | Sleep is critical for hippocampus-dependent memory consolidation. However, the underlying mechanisms of synaptic plasticity are poorly understood. The central controversy is on whether long-term potentiation (LTP) takes a role during sleep and which would be its specific effect on memory. To address this question, we used immunohistochemistry to measure phosphorylation of Ca2+/calmodulin-dependent protein kinase II (pCaMKIIα) in the rat hippocampus immediately after specific sleep-wake states were interrupted. Control animals not exposed to novel objects during waking (WK) showed stable pCaMKIIα levels across the sleep-wake cycle, but animals exposed to novel objects showed a decrease during subsequent slow-wave sleep (SWS) followed by a rebound during rapid-eye-movement sleep (REM). The levels of pCaMKIIα during REM were proportional to cortical spindles near SWS/REM transitions. Based on these results, we modeled sleep-dependent LTP on a network of fully connected excitatory neurons fed with spikes recorded from the rat hippocampus across WK, SWS and REM. Sleep without LTP orderly rescaled synaptic weights to a narrow range of intermediate values. In contrast, LTP triggered near the SWS/REM transition led to marked swaps in synaptic weight ranking. To better understand the interaction between rescaling and restructuring during sleep, we implemented synaptic homeostasis and embossing in a detailed hippocampal-cortical model with both excitatory and inhibitory neurons. Synaptic homeostasis was implemented by weakening potentiation and strengthening depression, while synaptic embossing was simulated by evoking LTP on selected synapses. We observed that synaptic homeostasis facilitates controlled synaptic restructuring. The results imply a mechanism for a cognitive synergy between SWS and REM, and suggest that LTP at the SWS/REM transition critically influences the effect of sleep: Its lack determines synaptic homeostasis, its presence causes synaptic restructuring.
| Sleep is important for long lasting memories. There exists, however, a controversy regarding the mechanisms by which sleep modifies synapses to consolidate enduring memories. One theory posits that sleep weakens synapses, leading to the forgetting of all but the strongest memories. The alternative theory proposes that sleep promotes both weakening and strengthening of different connections, the latter through a process known as long-term potentiation (LTP). Here we measured the levels of a phosphorylated protein related to LTP during the sleep cycle of rats and used the data to build models of sleep-dependent synaptic plasticity. By feeding one model with spikes recorded from the rat hippocampus, we observed that LTP during sleep not merely strengthens certain connections, but actually reorganizes how these connections are ranked in strength, leading to substantial changes of the overall pattern. A more detailed model of hippocampus and cortex showed that the interaction of the mechanisms predicted by the competing theories promotes a more efficient control of which memories are stored. Our results provide a step forward in the understanding of the cognitive role of sleep by indicating that the current competing theories are not mutually exclusive. Instead, each constitutes an important stage of memory consolidation.
| In the hippocampus, slow-wave sleep (SWS) is characterized by large amplitude, low-frequency oscillations of the local field potential (LFP), concomitant with a phasic regime of neuronal firing, with relatively low mean firing rates and intermittent synchronization [1–4]. In contrast, rapid-eye-movement sleep (REM) displays small amplitude, high-frequency oscillations that underlie a tonic firing regime, with relatively high mean firing rates and decreased synchrony [1–4]. Both sleep states play a role in the consolidation of hippocampus-dependent memories [5, 6], but the mechanisms remain poorly understood.
Two theories are in dispute. The synaptic homeostasis hypothesis (SHY) proposes that SWS causes generalized synaptic weakening [7–9], leading to the down-selection of weak synapses [10]. The notion that synaptic depression is determinant for off-line memory processing departs from the conventional Hebbian learning rule, by which connections among simultaneously firing neurons are reinforced [11]. On the other hand, the synaptic embossing hypothesis postulates the combination of non-Hebbian rescaling and Hebbian potentiation of synaptic weights in complementary circuits during REM [6, 12, 13]. The core of the dispute is the controversy on whether long-term potentiation (LTP) occurs during sleep. Empirical studies diverge considerably, with molecular, electrophysiological and morphological evidence for [14–22] and against [4, 7, 23–25] it.
The theories also differ on the roles of the different sleep states in memory consolidation. SHY only considers SWS and does not propose any role for REM [4, 26], while the embossing theory encompasses both states [6, 12, 13]. The substantial differences in the firing and correlation regimes of SWS and REM suggest that the two states should be separately and sequentially modeled [6, 12, 27]. One study has suggested that SWS leads to general memory reinforcement while REM leads to forgetting of all but the strongest memories traces [28]. LTP during SWS has been proposed to amplify the synaptic changes acquired during WK [29], with further processing of the resulting synaptic weights during REM [30, 31]. In fact, plasticity factors such as protein kinases and transcription factors encoded by immediate-early genes are up-regulated during REM [14–16, 20, 32, 33]. Therefore, it is possible that a complete sleep cycle traversing SWS and REM leads to important perturbations in the pattern of synaptic weights [6, 12, 27], rather than to the simple weight convergence observed during SWS alone [4].
To address this debate, we first assessed phosphorylated Ca2+/calmodulin-dependent protein kinase II (pCaMKIIα) in the hippocampus of rats exposed (or not) to novel objects, and killed immediately after subsequent WK, SWS or REM (S1 Fig). CaMKIIα phosphorylation is one of the earliest mechanisms with a critical role in LTP, memory and learning [34, 35]. CaMKIIα undergoes conformational changes towards the active form within seconds after the beginning of synaptic stimulation [36], triggering later events that include the up-regulation of immediate-early genes required for long-term synaptic remodeling, such as Zif-268 [37]. Given the very fast kinetics of CaMKIIα phosphorylation [36] in comparison with Zif-268 [38], pCaMKIIα levels were hypothesized to show experience-dependent changes immediately after SWS and/or REM, while Zif-268 protein levels were expected to be invariant immediately after any given state. The protein levels of total CaMKIIα and Actin were also assessed as negative controls expected to show invariant levels across groups, given their much slower transcriptional and translational regulation. To gain insight in the state dependency of pCaMKIIα regulation, we also investigated the relationship between pCaMKIIα levels and electrophysiological markers of SWS (delta oscillations), REM (theta oscillations) or the SWS/REM transition (neocortical spindles) [39].
Since both SHY and the synaptic embossing theory have empirical support, computational work may be particularly insightful. SHY has been modeled at the single neuron and network levels, but without real neurophysiological inputs [4, 10]. A recent SHY model found deleterious effects for memory when instantaneous potentiation was switched on during sleep [10], but the synaptic embossing hypothesis is yet to be simulated with realistic LTP onset and dynamics. To that end, we computationally investigated the network consequences of LTP triggered during sleep, by feeding simulated neurons with action potentials recorded from the rat hippocampus across the sleep-wake cycle. LTP was applied at REM onset as constrained by the empirical results. Synchronization-based LTP was calculated from coincident spiking during SWS only or SWS+REM, while an alternative model captured the notion that short-term changes in pCaMKIIα levels determine long-term increases in synaptic weights. Finally, we investigated the interaction of synaptic homeostasis and embossing mechanisms by simulating the dynamics of memory formation during a sleep cycle in a canonical hippocampo-cortical model.
For quantification of pCaMKIIα and Zif-268 levels, adult male Wistar rats (n = 27, 300–350 g) were housed, surgically implanted and recorded according to National Institutes of Health (NIH) guidelines and Edmond and Lily Safra International Institute of Neuroscience of Natal (ELS-IINN) Committee for Ethics in Animal Experimentation (permit # 05/2007). Implanted animals were housed in individual home cages with food and water ad libitum, and were kept on a 12h light/dark schedule with lights on at 06:00. At the end of the experiment, rats were anesthetized with isoflurane 5%, and decapitated.
The network was implemented as a modified Boltzmann machine [44]. The total synaptic current for each neuron i is defined as Ii(t)=wieei(t)+1N−1∑j=1Nwijvj(t) where N is the number of neurons, ei(t) is an external input and vj(t) is the state of pre-synaptic neuron j; wie and wij are the corresponding synaptic weights. The neuron state is binary (0 or 1) and stochastically updated by the adjusted sigmoid function P(vi(t)=1)=11+e(Kt−KsIi), where Kt = 6 and Ks = 11 are constants; with these values the mean firing rate of the spontaneous network activity (when wie = 0) is around 0.5 and 1Hz (S9 Fig). During all simulations, we used wie = 0.5 and pre-synaptic weights wij were randomly initiated following a uniform distribution between 0 and 1, except when i = j, in which case the synaptic weight is set to 0 throughout the simulation.
The network was exposed to 2 types external inputs (ei(t)): Real spike data and Poisson spike trains with same mean firing rates as those observed during WK, SWS and REM (S4A Fig). When Poisson spike trains were used as inputs, the network was composed of 150 neurons. When real spike data were used, the network was composed by the same number of neurons as recorded in each animal (Rat1 = 45, Rat2 = 39, Rat3 = 39, Rat4 = 34, Rat5 = 22 and Rat6 = 13). Despite the lack of inhibitory neurons, this simplified network replicates the synaptic rescaling dynamics (see below) of a more complex network with both excitatory and inhibitory processing units [4].
We built a network of 45 leaky integrate-and-fire neurons with network feedback inhibition [50–52] (S6 Text). Each neuron received synaptic connections from 200 input units and was not directly connected to other neuron. Half of the input units were assigned to memory A (MA = [1 … 100]) and the other half to memory B (MB = [101 … 200]). Input unit i emitted actions potentials (S[t] = 1 if there is a spike at time t, S[t] = 0, otherwise) following a Poisson distribution with mean frequency factive = 20 Hz when in an active state and finactive = 10 Hz when in an inactive state. Input pattern was determined in 125 ms windows by randomly selecting one memory to be in an active state. Each cell gets selective to a specific memory through the synaptic weight dynamics driven by the plasticity mechanisms described below.
Six groups were defined based on prior experience (with or without exploration of novel objects), and on the state immediately before killing (WK, SWS or REM) (S1 Fig). In unexposed control groups, no significant differences in hippocampal pCaMKIIα levels were detected across states (Fig 1A; Kruskal-Wallis p = 0.2958, Dunn’s test for consecutive states, adjusted p values: WK vs. SWS p>0.9999; SWS vs. REM p = 0.5615). However, in animals previously exposed to novel objects, hippocampal CaMKIIα phosphorylation significantly increased from SWS to REM (Fig 1A; Kruskal-Wallis p = 0.0396, Dunn’s test, adjusted p values: WK vs. SWS p = 0.0954; SWS vs. REM p = 0.0473). The number of cells with pCaMKIIα somatic labeling was counted to determine whether the densitometric changes observed could be attributed to a sheer increase on the total number of cells engaged in the pCaMKIIα response during REM. No significant differences were detected when labeled cells were counted. This indicates that the changes observed from SWS to REM in exposed animals did not reflect increased numbers of responsive neurons, but rather increased pCaMKIIα levels in the neuropil and soma.
To control for potential a priori inter-group differences in baseline CaMKIIα phosphorylation levels, we assessed the protein levels of total CaMKIIα, Actin and Zif-268 in adjacent brain sections. The rationale for this comparison was the fact that the regulation of these proteins is downstream of CaMKIIα phosphorylation, with much slower kinetics [48, 49, 53]. Since killing occurred immediately after specific sleep-wake states, we did not expect the levels of total CaMKIIα, Zif-268 or Actin to be differentially modulated across groups, unless there were spurious inter-group differences unrelated to states. We found no significant differences across groups for these proteins, irrespective of previous exposure to novel objects (Fig 1A; Kruskal-Wallis: Zif-268 Exposed p = 0.5387, Zif-268 Control p = 0.8775; total CaMKIIα Exposed p = 0.8270, total CaMKIIα Control p = 0.6505; Actin Exposed p = 0.6907; Actin Control p = 0.8781). The even distribution of the levels of total CaMKIIα, Zif-268 or Actin across groups indicated that the increase in CaMKIIα phosphorylation from SWS to REM was not an artifact of group sorting.
It is clear that the significant increase of CaMKIIα phosphorylation from SWS to REM could only occur because of a reduction from WK to SWS, which nevertheless only reached a near-significant trend (p = 0.0954). To grasp this effect, we calculated quadratic fits of the data across wake-sleep states (Fig 1A, red curves, curvatures indicated by numbers in the bottom). The pCaMKIIα data show a “U” effect in the exposed group but not in the control group, while no such effect was observed for total CaMKIIα, Actin or Zif-268. Schematic pCaMKIIα profiles predicted by the synaptic homeostatic hypothesis (monotonic decrease), the synaptic embossing hypothesis (“U”shape) and the null hypothesis (stable levels) are shown in Fig 1B. The data fit the picture of homeostatic rescaling during SWS (decrease in pCaMKIIα levels) followed by experience-dependent LTP during REM (increase in pCaMKII levels).
To investigate the electrophysiological correlates of CaMKIIα phosphorylation during REM, we assessed the correlation of pCaMKIIα levels with LFP power in the delta (0.5 to 4.5 Hz) and theta (4.5 to 12 Hz) bands during the last 15 min before killing. Neither frequency band showed significant correlations with pCaMKIIα levels (for theta, SWS group: R2 = 0.05322 and p = 0.5504, REM group: R2 = 0.02431 and p = 0.6888; for delta, SWS group: R2 = 0.002432 and p = 0.8997, REM group: R2 = 0.1463 and p = 0.3097).
Next we calculated the correlation of CaMKIIα phosphorilation with cortical spindle counts during the last 15 min before killing (Fig 2A and 2B). Spindle occurrence in the SWS and REM groups was assessed on spectral ratio maps (Fig 2C and 2D). While animals in the SWS group did not show any significant correlation (Fig 2E left panel), animals allowed to enter REM showed a positive correlation between spindle counts and pCaMKIIα levels (Fig 2E right panel). Interestingly, while pCaMKIIα levels were correlated to the sum of all spindles that occurred in the transition from SWS to REM (Fig 2E right panel, black dots and regression, R2 = 0.484, p = 0.0375), no correlations were observed for spindles occurring exclusively within SWS (Fig 2E right panels, blue dots and regression, R2 = 0.116, p = 0.369), nor for spindles sampled only from the IS state immediately before REM (Fig 2E right panels, magenta dots and regression, R2 = 0.007, p = 0.825).
S3A Fig shows that REM animals spent significantly more time in IS than SWS animals. REM animals did not have significantly more spindles than SWS animals (S3B Fig), but spindles lasted longer in IS than in SWS (S3C Fig). Altogether, more IS time in REM animals and longer spindles in IS explain why SWS animals displayed significantly less time with spindle occurrence (S3D Fig).
The results above support the notion that LTP is triggered at the SWS/REM transition. To model the consequences of this phenomenon, we first investigated how state-dependent variations in firing regimes affect the synaptic weights of a fully-connected network comprising an excitatory population of stochastic binary units (see Materials and Methods). The synaptic weights were initialized with a random uniform distribution, and therefore with maximum entropy. A stable Hebbian learning rule based on pairwise synchrony was used to update synaptic weights over time [45]. Synchrony was evaluated in 4-ms bins, well within the interval of maximum STDP [47], and lack of synchrony led to a fixed amount of synaptic weight weakening. The simulations were generated by feeding the network with 2 kinds of external inputs (S1 Table; representative example): Poisson spike trains at various rates; and actual spike trains recorded from the rat hippocampal field CA1 during WK, SWS and REM [20].
The data statistics conformed to the expected state-dependent changes across the sleep-wake cycle [1–3]: SWS featured low mean firing rates (Fig 3A; representative example) and decreased firing synchronization (Fig 3B; representative example) compared with WK. REM was characterized by mean firing rates in between those of WK and SWS (Fig 3A), with very low firing synchronization (Fig 3B). For data from 5 other rats see S4 Fig. Fig 3C depicts the durations of intervals separating consecutive SWS/REM transitions (left panel; n = 6), and a cumulative plot showing that 91,6% of these intervals are shorter than 30 min (right panel).
Two major scenarios were implemented, with and without LTP during sleep. For statistical robustness, all simulations were independently repeated 25 times. The dynamics of the synaptic weight patterns were quantified using 2 metrics: the Similarity Index measured the sum of absolute differences between a given synaptic weight pattern and a reference pattern, while the Spearman´s correlation quantified ranking changes among synaptic weights compared to the reference pattern. Altogether, these metrics allowed us to estimate the rescaling and restructuring of the synaptic weight patterns over time. Rescaling was indicated by a reduction in the Similarity Index without major changes in Spearman´s correlation. Restructuring was indicated by a reduction in the Similarity Index accompanied by a reduction of Spearman´s correlation.
We began by characterizing the rate-dependency of synaptic weight dynamics during a regime of non-correlated external inputs. The network was fed Poisson surrogated spike trains with mean rates between 5 and 10 Hz, which represent the real data range (Figs 3A and S4A). The distribution of synaptic weights was rescaled over time to a narrow range of values (S5A and S6A Figs), exactly as observed previously for a single cell model [45], as well as a network model with both excitatory and inhibitory units [4]. The mean synaptic weight at the convergence time point (S5A Fig, dashed black lines) depended on the mean firing rate of the inputs (S6B Fig, blue curve). By the same token, the time required for the synaptic pattern to converge was inversely proportional to the input rate (S6B Fig, red curve).
For simulations with low rate inputs typical of SWS (Figs 3 and S4A), mean synaptic weights at the time of convergence were smaller than 0.5, resulting in net weakening (Mw<0, see Material and Methods) of the synaptic weights (S5A and S6A Figs, distributions for 3, 5 and 7Hz). However, synaptic weights that were initially below the mean at the time of convergence became strengthened (S5A Fig, bottom panels). Therefore, net weakening of synaptic weights does not imply that all synaptic weights decay over time, since weak synaptic connections are potentiated. Conversely, for simulations in which inputs had higher rates typical of REM or WK (>7Hz), synaptic weights at the time of convergence were larger than 0.5, resulting in net potentiation of synaptic weights (S5A Fig, for 10Hz and its corresponding gray distribution in S5C and S6A Figs, distributions for 12, 20 and 40Hz). Yet, synaptic weight values initially above the convergence range were reduced over time (S5A Fig, bottom panels). In summary, when the network was fed with non-correlated Poisson inputs, the wide range of synaptic weights used to initialize the network converged to a narrow and stable distribution, producing net weakening or net potentiation of the synapses for low and high firing rates, respectively.
To further characterize the state-dependency of synaptic weights, we fed the network with spike data from concatenated WK, SWS or REM episodes (S5B Fig). The simulations confirmed that the mean firing rates of the external inputs determine the mean value reached over time by the distribution of synaptic weights P(w), as shown for Poisson data in previous work [45] and in the preceding section. Periods of increased spike rate and synchronization, such as WK, resulted in a smaller standard deviation of the synaptic weights when the network reached steady state, in comparison with periods of reduced spike rate and correlation, such as SWS or REM (S1 Table and S5B and S5C Figs green distributions). Note that the standard deviations at steady state were even smaller for non-correlated Poisson data of identical mean rates, and also obeyed the relationship WK<REM<SWS (S1 Table and S5C Fig, gray distributions).
Next we investigated how external inputs with the real dynamics of state alternation affected the distribution of synaptic weights. Fig 4 shows results when the network inputs were real spike data recorded over 5 hours from one rat (Rat1) cycling freely through the sleep-wake cycle, i.e. containing the natural alternation of WK, SWS and REM (Fig 4A, hypnogram). As expected, population firing rates were markedly state-dependent throughout the recording (Fig 4A, Pop. rate). The model displayed net potentiation during WK and net weakening during sleep (Fig 4B). We also observed that most synaptic connections did not reach extreme values (close to 0 or to 1) but converged to the middle range of possible values (Fig 4B).
Two alternative LTP models triggered near the SWS/REM transition were investigated. Since LTP is tightly associated with firing synchrony [11, 47, 54], one model attempted to capture the notion that SWS causes LTP through firing synchronization [55–57] and enhanced calcium influx [19, 30, 58], leading to calcium accumulation during SWS that would then lead to increased pCaMKIIα levels at REM onset. To simulate this scenario, LTP1 full SWS model applied a long-term bonus to each synapse starting at the SWS/REM transition, but according to the amount of pairwise synchrony observed throughout an entire SWS episode (87.06 ± 47.47, mean ±SEM in seconds, n = 6 rats).
The second model (LTP2) simulated short-term changes in pCaMKIIα levels as short-term increases in synaptic weights at the SWS/REM transition, and then used these short-term increases to determine long-term increases in synaptic weights. This model is compatible with the empirical data (Figs 1 and 2), and with the evidence of REM-dependent upregulation of plasticity factors [14–16, 20]. For each synaptic weight, LTP2 model coupled short-term changes at the SWS/REM transition to a gradual long-term bonus. The rationale for this coupling was the fact that the balance between high and low calcium influx in the millisecond scale is reflected in the balance between kinase and phosphatase activation in the seconds range, in particular CamKIIα, and determines the subsequent activation for over 30 min of Rho GTPases such as cdc42, and consequently to changes in gene regulation and protein synthesis in the hours scale [59–63].
To simulate this scenario, short-term changes in synaptic weights assessed for 60s at the transition from SWS (30s) to REM (30s) determined long-term bonuses applied for 30 min to the synaptic weights (see Material and Methods). Specifically, the angle formed by the synaptic weight trajectory at the transition from SWS to REM determined a long-term bonus. To comply with the notion that LTP requires positive calcium transients, LTP was applied exclusively to synaptic weights whose trajectory showed a positive slope during REM (LTP2 permissive 60s SWS/REM).
The long-term bonus applied in both models consisted of a Gaussian curve with onset at a reference SWS/REM boundary and peak value at 30 min, to match the duration of the spine-specific signaling cascade Ca2+–CaMKIIα–Cdc42 [64].
The evolution of synaptic weight patterns varied according to the LTP model employed. When LTP1 full SWS model was fed real data as inputs (Fig 4C), about 85% of the synapses underwent potentiation, in comparison with 21% in the No-LTP control. This resulted in a marked modulation of synaptic weight values (S7D Fig, 1st and 2nd panels), with a substantial net increase of the mean weight (S7A Fig, top panel, red curve) and increased spreading towards high synaptic weight values (S7C Fig, red with black edge distribution). Only 8% of the synapses underwent weakening, in comparison with 64% in the No-LTP control (S6A Fig, 1st and 2nd panels). The selection obeyed a uniform distribution across the synaptic weight range, so that even weak synaptic connections had a 16% chance of being potentiated. Overall, LTP model 1 led to a net potentiation of synaptic weights across their entire range (S7D Fig, middle panel), greatly exceeding that observed without LTP (S7B Fig, top panel, red curve). The distributions of synaptic weight changes (ΔWij) showed a marked upward shift, with a concavity change on the quadratic fit of LTP1 full SWS model, in comparison with the fit for the No-LTP model (compare red and black curves in S7E Fig).
When LTP2 permissive model was fed real data (Fig 4D), about 48% of the synapses were recruited to undergo LTP (S7D Fig, right panel, red bar), substantially less than in the case of LTP1 full SWS model. LTP2 permissive model also led to net synaptic potentiation across the entire range of weights, but many more connections were de-potentiated in comparison with LTP1 full SWS model (compare with middle panel). The distributions of differences between final and initial synaptic weight values (S7E Fig, blue) show an upward shift without concavity change in which most changes affect weak connections, i.e. the shift was largest for the lowest initial synaptic weights, and decreased gradually for larger initial weights.
The different consequences of the LTP models (LTP1 full SWS and LTP2 permissive) are shown in Fig 5, which depicts the temporal evolution of a fully connected network of 16 neurons under Poisson (Fig 5A) or real data regimes (Fig 5B). The initial synaptic weight pattern (Fig 5A, 1st column) was quickly erased when Poisson-distributed data were used as inputs (Fig 5A, 2nd and 3rd columns). LTP1 full SWS model uniformly enhanced all synaptic weights, while LTP2 permissive model embossed a new pattern (Fig 5A, 4th and 5th columns).
Real data allowed for a much longer persistence of the initial pattern, which evolved in distinct ways for LTP1 full SWS and LTP2 permissive. While the pattern changed monotonically over time in LTP1 full SWS model (Fig 5B, 2nd row), LTP2 restrictive model caused a major revamp of synaptic weights (Fig 5B, 3rd row), so that the pattern that emerged at the end of the simulation was very different from the initial pattern.
Were the differences between LTP1 and LTP2 models related to the different durations of the inputs, to the role assigned to the SWS/REM transition, or to intrinsic mechanistic differences between the models? To address these questions, we futher simulated LTP1 using not an entire SWS episode, but either a limited 30s SWS period immediately before REM (LTP1 - 30s SWS end), or that 30s SWS period plus the following 30s of REM (LTP1 - 60s SWS/REM). We found that short sleep periods near the SWS/REM transition produced less net synaptic potentiation for LTP1 than full SWS episodes (Fig 6A and 6B). They also produced more pattern restructuring than full SWS episodes (i.e. decrease in Spearman´s correlations, Fig 6C), partially replicating the results of LTP2 permissive model. Next we compared LTP2 permissive model with an alternative in which LTP2 was applied only when the REM slope was positive and larger than the SWS slope (LTP2 restrictive) (see Material and Methods). This more restrictive version of LTP2 led to less potentiation (Fig 6A, last two panels; Fig 6B, right panel), and less pattern restructuring (Fig 6C), than the original, more permissive LTP2.
Of note, the case presented so far is representative of 5 other animals, despite the high variability of the spike datasets used as inputs, which resulted from traversing quite different real sleep-wake cycles (S8 Fig). Spearman´s correlations and mean synaptic weights (Fig 7), used respectively to characterize restructuring and rescaling, confirm across animals that LTP during sleep leads to pattern restructuring, especially when assessed at the SWS/REM transition.
Since both SHY and the synaptic embossing hypothesis have empirical support, it is possible that the use of both mechanisms is synergistic. To clarify this point, we used a network model based on a canonical hippocampal-cortical circuit capable of developing specificity in response to concurrent inputs [50–52] (Fig 8). The model was adapted by implementing plasticity mechanisms analogous to SHY and the synaptic embossing hypothesis (see Material and Methods). The effect of sleep on memory was studied by comparing the patterns of synaptic weights post-sleep with pre-sleep (to assess the level of memory restructuring) and with the pattern reinforced by sleep-dependent LTP (to assess the influence of the SWS/REM transition pattern over the established memory).
Synaptic homeostasis was implemented by modulating the STDP rule [47, 54] during sleep through weakening of potentiation and strengthening of depression [4]. As observed in our empirical data, the rate of the input units was reduced during sleep. The effect of sleep on synaptic restructuring was measured by assessing the number of neurons whose pattern selectivity was either stable or switched. We observed that an increase in STDP modulation led to a number of stable memories, as expected by a random permutation of the pattern selectivity (Fig 9A). Although this experiment demonstrates that synaptic homeostasis shuffles synaptic weights, there is no mechanism to determine the output synaptic structure, i.e., SHY allows an acquired memory to be erased, but it cannot promote a specific pattern into the synaptic weights.
Embossing was implemented by evoking LTP on selected synapses during sleep. The LTP pattern was set proportional to the synaptic weights of another neuron selected randomly at a time previous to sleep onset. LTP control over the synaptic organization could be measured by comparing the memory of the neuron after sleep with the memory of the neuron selected as reference for LTP just before sleep onset. We observed that an increase of the intensity of LTP modulation led to a full control of the memory by the evoked LTP (Fig 9B).
We then implemented both hypotheses and observed the development of synaptic reorganization in the model. We simulated sleep cycles for a range of values of STDP modulation and LTP intensity and observed the number of memory hits between the pattern following the sleep cycle and the reference pattern with evoked LTP (Fig 9C). With an increasing level of STDP modulation, a lower LTP intensity was required to enhance LTP control, although some level of LTP was still needed to ensure LTP control over the synaptic organization. This result indicates that synaptic homeostasis facilitates controlled synaptic restructuring during synaptic embossing.
The experience-dependent increase of pCaMKIIα levels in hippocampal cell layers from SWS to REM indicates that the latter triggers synaptic potentiation within circuits selected by waking experience. Our data also show that CaMKIIα phosphorylation during sleep was proportional not to slow-wave or theta oscillations, respectively markers of SWS or REM [39], but rather to the amount of cortical spindles that occur at the SWS/REM transition. Spindles are 7–14 Hz oscillations that occur during sleep [39, 42], are abundant at the transition from SWS to REM [43], gate hippocampo-cortical communication [65], and correlate strongly with sleep-dependent learning [66–68]. Cortical spindles have been proposed to play an important role in calcium-dependent plasticity [30], perhaps by coupling the changes in memory traces produced by SWS to REM-triggered gene expression required for long-term storage [20].
The tight relationship between spindle counts and hippocampal pCaMKIIα levels likely reflects the key role played by spindles in firing synchronization across the hippocampo-cortical axis [65, 69]. Neither SWS or IS alone showed correlation between spindle count and pCaMKIIα levels, which suggests that cortical spindles spanning SWS and IS are critical for the coupling of the neuronal changes produced during SWS to hippocampal CaMKIIα phosphorylation at REM onset. A similar relationship between LFP power in the spindle range and Zif-268 mRNA expression 30 min after REM [20] supports a causal chain linking spindles, CaMKIIα phosphorylation and Zif-268 induction. Non-REM sleep has recently been implicated with learning-related formation of dendritic spines [21], but REM deprivation in this study was only partial (about 80% of the total), and spared SWS/REM transitions. Determination of the precise role of these transitions for sleep-dependent plasticity requires further investigation. Overall, our results are compatible with the notion that sleep spindles “open the molecular gates to plasticity” [30].
Based on the empirical data, we simulated a homeostatic excitatory network fed with spike data from the hippocampus of behaving rats to compare alternative theories regarding plasticity during sleep. In the 1st scenario, which did not include sleep-dependent LTP and was compatible with SHY, synaptic weights inexorably converged to the center of the distribution. In the 2nd scenario, which included sleep-dependent LTP, synaptic weights showed marked changes compatible with either orderly synaptic up-scaling (for synchronization-based LTP computed over entire SWS episodes) or with disorderly synaptic embossing (for LTP triggered at the SWS/REM transition).
Rescaling occurred in the absence of LTP during sleep, and was most pronounced when non-correlated Poisson inputs were used (Figs 4A and S4A). Synaptic weights quickly converged to a narrow range of values as previously described for different networks that reach an equilibrium state through a homeostatic process [7], including a network with inhibitory inputs [4]. Net weakening or potentiation depended on how low or high was the external activity. Non-correlated Poisson activity with high rate caused a faster decrease in the diversity of synaptic weight values than correlated spiking at lower rates (S4A Fig). Similarly, REM-only inputs were more effective in rescaling than SWS-only inputs (S4B and S4C Fig), which contradicts SHY [7] but agrees with the empirical findings of decreased firing rates after REM [25].
When the network received correlated real data inputs, the resulting synaptic weight distributions were much wider but still converged to the center (S4B Fig), with a tight relationship between input rates and the mean synaptic weight at the end of the simulation. These results conform to the notion that a network that goes through sleep without LTP remains stable, avoiding extremely weak or strong synaptic weights [45]. Synaptic weight convergence was orderly, preserving synaptic weight ranks as indicated by the relatively stable values of Spearman´s correlations over time (S6A Fig, bottom panel, black curve, 10,000s to 14,500s interval). While synaptic weight ranks were preserved in the absence of LTP, mean synaptic weights progressively decreased (S6A Fig, top panel, black curve), reaching the lowest value during sleep. The same occurred for the Similarity Index (S6A Fig, middle panel, black curve). Without LTP during sleep, net synaptic weights were downscaled, and the initial pattern was rescaled.
When LTP was modulated by the spike synchrony exhibited during an entire SWS episode, synapses were uniformly selected for potentiation (LTP1 full SWS model). When synapses were potentiated based on the angle formed by the synaptic weight trajectories at the SWS/REM transition (LTP2 model), synaptic recruitment was also quite uniform but the magnitude of potentiation was stronger for weights that were initially low (S6E Fig, blue curve). In both models, synaptic weight values were scattered away from the convergence range (Figs 4C, 4D and S6C middle, right panel, distribution with black edges).
Real neuronal activity imposes network relations that do not exist for Poisson inputs, including inhomogeneous firing rate variability and synchronous firing among neurons. These conditions determine that specific connections undergo markedly different dynamics under LTP1 and LTP2 models. The most interesting cases are those in which synaptic weights undergo opposite changes. About 85% of the connections undergo potentiation under LTP1 full SWS model but no modulation or weakening under LTP2 permissive model (S6E Fig, triangles). In contrast, only 8% of the connections undergo no modulation or weakening under LTP1 model, but show potentiation under LTP2 model (S6E Fig, squares). Interestingly, the results produced by LTP1 and LTP2 models converged when the former was evaluated around the SWS/REM transition; or when the latter was more restrictive (Fig 6).
The results support the notion that LTP during sleep allows weaker synaptic connections to also play a role in mnemonic processing [70]. Weight-dependent plasticity of hippocampal neurons, favoring weak over strong synapses, has been shown in vitro [45, 47]. Here, the combination of real hippocampal inputs and LTP during sleep triggered long-term changes in the synaptic weights that were specific of the particular LTP mechanism simulated. The consequences of LTP1 model during sleep were overall synaptic potentiation, increased similarity with the initial pattern and preservation of synaptic weight ranks. For LTP1 and LTP2 models evaluated at the SWS/REM transition, sleep gave rise to new synaptic weight patterns, as indicated the low Spearman´s correlation values (Fig 5B, last row, from 10,000s to 14,000s). Given that the typical period of the sleep cycle in rats is below 2 min (Fig 3C, left panel), with 91,6% of the episodes under the 30 min LTP Gaussian peak (Fig 3C, right panel), the sleep-dependent synaptic weight changes introduced by LTP models are prone to stagger from cycle to cycle, leading to progressively different patterns over time.
SHY does not seem to account for all the cognitive effects of sleep, which not only protects memories passively from retroactive interference, but can actively enhance specific memories in detriment of others, and lead to novel insights [71–79]. A simplified integrate-and-fire SHY model exclusively composed of excitatory neurons was recently used to simulate gist extraction and integration of new with old memories without the need of LTP during sleep [10,80]. These properties derive directly from the down-selection of weak synapses and protection of strong synapses, which end up eliminating weak memories (supposedly “spurious” information) and therefore increasing the signal-to-noise ratio of memory retrieval. However, by the same token, such mechanisms are not bound to explain the additional information content that arises from memory restructuring during sleep [71–79]. For instance, a recent study of perceptual learning in humans found that REM rescues memories from interference, preferentially strengthening weak memories [81].
In contrast, the results presented here for Model 1 show that LTP calculated over short periods near the SWS/REM transition lead to pattern restructuring compatible with the synaptic embossing theory. Furthermore, the more realistic hippocampal-cortical architecture with both excitatory and inhibitory synapses (Model 2) showed that the restructuring of synaptic patterns due to LTP was enhanced by homeostasis, which could explain the sleep-dependent enhancement of specific memories. These results highlight the importance of the SWS/REM coupling, providing strong support for the sequential hypothesis of mnemonic processing during sleep [6, 12, 27]. The joint occurrence of Hebbian and non-Hebbian plasticity during sleep, which leads to new synaptic patterns embossed over a background of homeostatically rescaled synaptic weights, may ultimately explain the positive role of sleep in the cognition.
|
10.1371/journal.ppat.1000020 | The Effect of Immune Selection on the Structure of the Meningococcal Opa Protein Repertoire | The opa genes of the Gram negative bacterium Neisseria meningitidis encode Opacity-associated outer membrane proteins whose role is to promote adhesion to the human host tissue during colonisation and invasion. Each meningococcus contains 3–4 opa loci, each of which may be occupied by one of a large number of alleles. We analysed the Opa repertoire structure in a large, well-characterised collection of asymptomatically carried meningococci. Our data show an association between Opa repertoire and meningococcal lineages similar to that observed previously for meningococci isolated from cases of invasive disease. Furthermore, these Opa repertoires exhibit discrete, non-overlapping structure at a population level, and yet low within-repertoire diversity. These data are consistent with the predictions of a mathematical model of strong immune selection upon a system where identical alleles may occupy different loci.
| Neisseria meningitidis is a globally important pathogen that causes 2,000–3,000 cases of invasive meningococcal disease annually in the United Kingdom. The meningococcal Opa proteins are important in mediating adhesion to and invasion of human tissues, and are important for evasion of the host immune response. They are encoded by a repertoire of 3–4 genomic loci in each meningococcus and exhibit high levels of sequence diversity. Here we analyzed the Opa repertoires of a large, well-characterised, asymptomatically carried meningococcal isolate collection. We found that the Opa repertoires were specific to individual meningococcal genotypes, similar to that observed in isolates from cases of invasive disease. These repertoires exhibited discrete, non-overlapping structure at a population level, and yet low within-repertoire diversity. These data were consistent with the predictions of a mathematical model of strong immune selection, suggesting that the collective immune response of the host population shapes the antigenic diversity of the meningococcal Opa repertoire. This study provides new insights into Opa-mediated meningococcal pathogenesis and the effect of host population immunity on the biodiversity and population structure of bacterial pathogens. These data may also have implications for the design of new meningococcal vaccines based on surface proteins.
| The Opacity (Opa) proteins of the bacterial pathogen Neisseria meningitidis mediate adhesion to and invasion of the human nasopharyngeal epithelium [1] via interaction with cell surface saccharides [2] and members of the carcinoembryonic antigen cell adhesion molecule (CEACAM) family of proteins [3],[4]. The opa gene repertoire comprises 3–4 loci per meningococcus (opaA, opaB, opaD and opaJ) [5]–[8]. These are constitutively transcribed and their expression is controlled by stochastic changes in a phase variable, pentameric repeat tract within the reading frame of the genes [9]. Varying numbers of opa loci may be expressed at different times and in different combinations, providing both functional flexibility and a possible mechanism for immune evasion.
Opa proteins are highly diverse [8],[10] with the majority of sequence changes localised in three regions which correspond to surface exposed loops in the proposed protein structure. It is thought that different sequences in the semivariable (SV) and two immunodominant hypervariable (HV) regions [10],[11] confer different receptor specificities to the protein [12],[13]. Diversity is generated by gene conversion, mosaicism and also modular exchange of variable regions, with the consequence that different opa loci in the same meningococcus may encode identical, similar or diverse HV regions [14],[15].
It has been shown that the Opa repertoire is highly structured among the hyperinvasive lineages of meningococci that are responsible for the majority of global disease [16]. Isolates from the same hyperinvasive clonal complexes (as defined by MLST) have been shown to possess similar and often identical Opa repertoires, despite being sampled from disparate geographical locations and temporal periods [8]. Little information is available, however, on the extent of the diversity of the Opa repertoire in carried populations of meningococci which contain the majority of meningococcal biodiversity. In this investigation, we analysed the Opa repertoires of a geographically and temporally related collection of asymptomatically carried meningococci to determine whether the association between clonal complexes and particular combinations of these adhesins, as observed in hyperinvasive lineages, was present in non-disease causing meningococci.
We analysed the data using a theoretical model of immune selection which incorporated the particular features of this antigenic system including its phase variable nature and the modular exchange of variable regions within genotypes. We found the patterns of diversity evident at both the population level and within individual repertoires to be indicative of strong immunological selection acting in addition to the forces of functional adaptation in influencing the structure of the Opa repertoire.
The four known opa loci were analysed in the 216 meningococcal isolates from a carried population sample from the Czech Republic: a total of 864 loci. In 784 loci (90.74%) an intact opa sequence was detected; these contained a total of 222 alleles (nucleotide p distance: 13.59%). These encoded 76 HV1 variants (amino acid p distance: 47.8%) which fell into 19 families and 93 HV2 variants (amino acid p distance: 37.6%) which fell into 21 families. A total of 212 opa loci were also analysed in a contemporaneous collection of 53 isolates from invasive disease. In 185 loci (87.26%) an opa sequence was detected, these contained a total of 75 alleles (nucleotide p distance: 14.26%). These encoded 41 HV1 variants (amino acid p distance: 48.4%) which fell into 15 families and 44 HV2 variants (amino acid p distance: 40.1%) which fell into 17 families.
In both the carriage and disease collections, we found that genetically related isolates, whether belonging to hyperinvasive clonal complexes or not, often had identical Opa repertoires (see Text S1 and Tables S1 and S2). For example, for the ST-11 complex, the opa gene alleleic repertoire opaA 83, opaB 11, opaD 132 and an insertionally inactivated opaJ locus was present in 27 of 32 carried isolates and 16 of 20 disease isolates. The remaining isolates from this complex in each collection had highly related repertoires, differing at only one or two loci. These repertoires were highly similar to those of isolates belonging to the same clonal complexes observed in a global collection of hyperinvasive meningococci (see Text S1 and Tables S1 and S2)[8].
The modular exchange of the immunodominant HV regions among different opa loci [14],[15] makes the system unusual in the context of immune selection, since the same hypervariable region variants may be present at multiple opa loci within the same isolate, as well as in different isolates. We extended a multi-locus mathematical model (see Materials and Methods for details) developed by Gupta et al.[17] to incorporate this feature by allowing the two HV regions (HV1 and HV2) at each locus to contain two possible amino acid sequence epitopes (‘a’ and ‘b’ for HV1 and ‘x’ and ‘y’ for HV2) as shown in Figure 1. Thus, possible combinations of HV regions in Opa proteins expressed by different meningococci could be: ‘ax’, ‘ay’, ‘bx’, ‘by’, ‘axy’, ‘bxy’, ‘abx’, ‘aby’, and ‘axby’. The behaviour of this model under different levels of immune selection is shown in Figure 2. These simulations indicate that the system shows a tendency to self organise at a population level into discrete antigenic types as the strength of immune selection increases, as previously observed [17] for multi-locus systems without modular exchange of variable regions. When immunological selection (as measured by cross-protection between pathogen types sharing variable regions) is weak, all antigenic types coexist at the similar abundances as shown in Figure 2a. By contrast, when immunological selection is high, a subset of two strains expressing two non-overlapping HV1/HV2 region combinations (for example, ‘ax’ and ‘by’) dominates, excluding all other strains, as exemplified by Figure 2c. Between these two extremes, we observe cyclical dynamics with strains expressing subsets of non-overlapping HV variants successively dominating the population (Figure 2b).
Figure 3 shows the combinations of HV1 and HV2 present at all loci for all isolates. Each opa locus is treated independently, so each isolate can contribute more than one combination. Variants for which only a single isolate was found were excluded from this analysis (see Table S3 for full details). To determine which of these population structures best described these data, a simple metric (f*) was developed to assess the extent of overlap between two epitopes among different isolates (see Materials and Methods for the derivation, and Text S1 for model validation): f* scores close to 1 indicate a highly non-overlapping structure, expected when cross-immunity is high, whereas scores close to 0 occur when strains have completely overlapping antigenic repertoires. Scores obtained from the opa loci in the dataset were compared to scores from housekeeping genes belonging to the same isolates. The f* metric for the data shown in Figure 3 is 0.9737, whereas pairwise comparisons of the housekeeping gene loci yielded a mean f* score of 0.3453 and a maximum of 0.4578. These scores indicate the non-overlapping nature of the Opa HV1/HV2 combinations as compared to the housekeeping loci, and reflect the diagonal pattern observed in the figure.
A total of 124 HV1/HV2 combinations were observed out of a possible total of 7068 (76 HV1 variants multiplied by 93 HV2 variants). Discrete, non-overlapping combinations of HV1 and HV2 are clearly dominant, despite the presence of rare combinations generated by frequent recombinational exchange. These observations support the model structure described above in which strong immune selection is responsible for structuring Opa repertoires.
Another important feature of the simulations presented in Figure 2 which is unique to a system with modular exchange between loci is that immune selection paradoxically leads to a reduction in diversity within individual opa repertoires. In other words, if more than one opa locus is expressed in vivo, selection will favour those strains expressing multiple loci encoding the same combination of HV regions, rather than different, more diverse variants. It is evident from Figure 2a that even under low levels of immune selection, the prevalence of strains expressing three or four antigenic determinants is suppressed. The magnitude of suppression increases with the degree of cross-protection (represented in the model by the parameter γ), such that these more diverse types are entirely absent in Figure 2c. This suppression occurs because strains expressing more than two HV variants are less likely to encounter hosts who have not previously been exposed to one or more of their epitopes, and are therefore at a disadvantage within the population. Thus, at very high levels of immune selection, we observe only meningococci that expressed a single opa locus, or multiple loci encoding the same combination of HV regions (i.e. ‘ax’ at locus 1 and ‘ax’ at locus 2). This is because the pathogen population, and therefore the background of host immunity, is dominated by two non-overlapping strains, say ‘ax’ and ‘by’, so that more diverse strains (such as ‘axy’) are more likely to be recognized by hosts who have encountered either one of the dominant strains.
To investigate the effect of host immunity on the structuring of the HV region repertoire diversity in individual isolates, we analysed the HV combinations of different Opa proteins within the same isolate for those that had full opa sequences at more than one locus (not including those that had frame-shift mutations or insertional inactivations). Figure 4 shows the proportion of isolates with identical HV1/HV2 combinations at different opa loci within the same isolate, compared to a hypothetical pathogen population in which the same combinations found in the data were distributed randomly within and among isolates. Only unique Opa repertoires were included in the analysis, to control for bias due to particularly prevalent sequence types and clonal complexes.
Our results showed that significantly more isolates contained two or more of the same HV1/HV2 combination than would be expected by chance given the same overall prevalence of variants (p<0.0001). Furthermore, they were not always the same HV1/HV2 combinations that were identical, with 28 different combinations occurring more than once within isolates. Finally, the probability that two or more were identical increased with the number of opa loci at which a full length opa allele was detected for each isolate (see Figure 4).
The Opa repertoire structure observed in the carried meningococci from the Czech Republic, and its relationship to clonal complex, was similar to that previously described in an isolate collection representing the diversity of meningococci causing disease globally in the latter half of the 20th century [8]. This does not imply that the Opa repertoire has no role in meningococcal pathogenesis since other factors, such as differences in expression patterns among meningococci and host susceptibility, are likely to influence the outcome of infection.
Despite evidence for extensive recombination of opa loci among meningococci, only a fraction of all possible combinations of HV1 and HV2 were observed. . These combinations exhibited a non-random and non-overlapping structure, which was consistent with a model of immunological selection in which competition between pathogen types leads to a pathogen population dominated by non-overlapping combinations of antigenic variants [17]–[19]. The low frequency off-diagonal elements shown in Figure 3 may be attributed either to the point prevalent nature of the data set (ie. that these combination are shortlived) or reflect the fact that certain variants possess immunological similarities, and are therefore equally likely to occur in combination with certain others. It is also possible that there are functional constraints in operation here since particular HV1/HV2 combinations influence receptor tropism and potentially also avidity [12],[13],[20]. It has been suggested that expression of CEACAM on host cell surfaces may allow evasion of antibody responses by Opa-mediated entry into epithelial cells [4] and modulation of the host immune responses by interaction with CD4+ T cells [21]. The specificity of these interactions is likely to constrain allowable HV1/HV2 combinations and may explain why particular combinations are entirely absent in our data.
Non-overlapping patterns of epitope combinations have also been observed among meningococcal PorA variable regions [17],[19]. Unlike PorA however, the Opa repertoire is a four-locus system [5],[8], and has been suggested to play a role in immune evasion [22],[23]. For the Opa proteins, individual repertoires exhibited more identical HV variants than would be expected under the assumption that antigenic diversity of a surface component prolongs infection (Figure 4). This result is, however, consistent with the predictions of a mathematical model of strong immune selection upon a system where identical alleles may occupy different loci. Within this framework, isolates expressing diverse repertoires are at a disadvantage because they are more likely to encounter hosts with previous exposure to one or more of their epitopes. This results in selection for identical variants at multiple loci: in other words, a reduction in the diversity of the Opa repertoires within individual meningococci. The prevalence of identical Opa variants within repertoires implies that multiple opa loci are expressed in vivo; if expression were restricted to a single opa locus, there would be no selective disadvantage of carrying a diverse repertoire. An alternative explanation for the low within-repertoire diversity is that identical HV combinations reflect genetic duplication events that are followed by specialisation of duplicates for new functions [24],[25].
An exception to the pattern of diversity within the Opa repertoires in most clonal complexes was that of the ST-11 complex which did not have any identical HV variants among its loci. This may be due to the recent entry and rapid spread of this clonal complex into the population of the Czech Republic, where it was responsible for a rapid increase in the incidence, mortality and morbidity of invasive meningococcal disease in 1993 [26]. Retrospective monitoring of isolates since 1970 suggested that this strain was not present in the country before 1993 and consequently the Czech population may have been immunologically naïve, allowing these meningococci to spread through the population. Thus, high Opa repertoire diversity may be selectively advantageous for the invasion of new communities of hosts with variable immunological backgrounds. During prolonged carriage in the same host population however, increased diversity may become costly as the proportion of immunologically naïve hosts decreases. This would inevitably cause a reduction in the range of receptor tropism, but this would be offset by the gain in probability of transmission. To date, the majority of Opa proteins tested bind at least CEACAM1 [27], suggesting that the repertoire retains binding of this major receptor.
Intriguingly, the number of opa loci differ among the Neisseria species, with 3–4 in Neisseria meningitidis [6],[7], 11–12 loci in Neisseria gonorrhoeae [28] and two in Neisseria lactamica [29]. The reasons for these differences are unclear, but our analyses in this study suggest a theory based on population prevalence and immunological cross-protection. For example, whereas N. meningitidis is transmitted by aerosol inhalation, N. gonorrhoeae is transmitted sexually and consequently has a much lower population prevalence. The likelihood of N. gonorrhoeae encountering an immunologically naïve host may be much higher, therefore, and the diversity-reducing effect from the host population's immunological responses less pronounced than for the meningococcus. A more diverse Opa repertoire with more loci may be more advantageous in these circumstances. Further information on the antigenic diversity of the gonococcal Opa repertoire and immunological responses against both pathogenic Neisseria species would be required to test this hypothesis.
In conclusion, this analysis demonstrates that particular Opa repertoires are associated with meningococcal clonal complexes irrespective of geographic or temporal sampling, whether isolated from asymptomatic carriers or invasive disease cases. The repertoires exhibit discrete, non-overlapping structure on a population level and low within-repertoire diversity, indicating that immune selection plays an important role in shaping Opa repertoires.
A total of 216 meningococcal isolates were obtained from an asymptomatically carried population of meningococci collected in the Czech Republic between March and June of 1993 [30]. A full description of these meningococci, including year and location of isolation, MLST and antigen gene sequencing data appears online at http://pubmlst.org. Genomic DNA was prepared by culturing isolates as previously described [30] before extracting with a DNA mini kit (Qiagen, Crawley, UK) according to the manufacturer's instructions. The opa loci were isolated in separate, locus-specific PCR amplifications, their nucleotide sequences determined at least once on each strand and their variable regions identified as previously described [8]. Nucleotide and amino acid sequence data are available in an online database located at http://neisseria.org/nm/typing/opa/. For analyses of diversity, by uncorrected nucleotide or amino acid percentage (p) distance, sequences were aligned and diversity calculated using the program DAMBE: Data Analysis and Molecular Biology and Evolution [31].
A non-overlapping strain structure results in a matrix of allelic associations between two antigenic loci in which each allele at locus 1 should be predominantly associated with only one allele at locus 2, and vice versa. This means that the most prevalent strains should dominate both the ‘row’ and ‘column’ of their allelic association matrix. The level of overlap within such a matrix can therefore be measured by assessing the dominance of the most prevalent allele combinations;The dominance of the most prevalent allele combination in each column (fa) is calculated, where locus 1 expresses allele i, and locus 2 expresses allele j. fi is the frequency of the most prevalent strain expressing allele i at locus 1, with respect to all strains expressing that allele (ie. the ‘column’ dominance), fj is the frequency of that strain with respect to all strains expressing allele j at locus 2 (ie. the ‘row’ dominance), and fij is the frequency of allele combination ij overall. These are calculated as follows:The sum over all fa gives the overall overlap between two loci:such that f* varies between 0 and 1. For a completely non-overlapping matrix, with no combinations found except for those that do not overlap, f* will be exactly one. As this structuring breaks down, the f* score will decrease rapidly.
Three differential equations, based on a model by Gupta et al. [17],[18], describe the system:The model states that once infected with a particular strain, the host gains partial immunity to any other strains expressing shared antigenic determinants (the subset of strains i' above) as specified by the parameter γ. For each strain i, the host population consists of three overlapping compartments; the proportion infectious to other hosts, xi; the proportion exposed (and therefore immune) to strain i, zi, and the proportion exposed to any strain sharing antigenic determinants with i, wi. It was assumed that the duration of infectiousness (1/σ) was short compared to the average host life-span (1/μ), and that immunity was life-long. All strains were assumed to have the same transmission coefficient, β. The effect of recombination was not explicitly included in the model, however all possible strains were present from the start in order to investigate the competitive interactions between them. Note that in this model there was no dose-dependence; two loci expressing Opa proteins with identical HV regions was taken as being the same as if only one locus expressed the protein. |
10.1371/journal.pntd.0003566 | Comparative Analysis of Field-Isolate and Monkey-Adapted Plasmodium vivax Genomes | Significant insights into the biology of Plasmodium vivax have been gained from the ability to successfully adapt human infections to non-human primates. P. vivax strains grown in monkeys serve as a renewable source of parasites for in vitro and ex vivo experimental studies and functional assays, or for studying in vivo the relapse characteristics, mosquito species compatibilities, drug susceptibility profiles or immune responses towards potential vaccine candidates. Despite the importance of these studies, little is known as to how adaptation to a different host species may influence the genome of P. vivax. In addition, it is unclear whether these monkey-adapted strains consist of a single clonal population of parasites or if they retain the multiclonal complexity commonly observed in field isolates. Here we compare the genome sequences of seven P. vivax strains adapted to New World monkeys with those of six human clinical isolates collected directly in the field. We show that the adaptation of P. vivax parasites to monkey hosts, and their subsequent propagation, did not result in significant modifications of their genome sequence and that these monkey-adapted strains recapitulate the genomic diversity of field isolates. Our analyses also reveal that these strains are not always genetically homogeneous and should be analyzed cautiously. Overall, our study provides a framework to better leverage this important research material and fully utilize this resource for improving our understanding of P. vivax biology.
| In this study we compare the genome sequences of Plasmodium vivax collected directly from patients with those of parasites propagated in laboratory monkeys. We show that the adaptation and continuous propagation of Plasmodium vivax in monkeys does not induce systematic changes in the genome and, therefore, that these parasites constitute an unbiased resource for studying this important pathogen. Our analyses also reveal that some monkey-adapted Plasmodium vivax strains are not genetically homogenous and retain multiple genetically different parasites present in the original patient infection. Overall, our study confirms the utility of monkey-adapted Plasmodium vivax strains for malaria research but also shows that this resource should be analyzed cautiously as different samples of the same strain might provide different biological material.
| Today approximately 2.5 billion people are at risk of Plasmodium vivax malaria [1]. While transmission of P. falciparum is slowly decreasing in many countries committed to malaria elimination, vivax malaria displays surprising resilience in a majority of these countries [2]. This difference, likely resulting from the important biological differences between the two parasite species (e.g., the existence of a dormant stage in P. vivax), calls for specific elimination strategies targeting P. vivax more efficiently. However, our understanding of P. vivax biology remains limited by the difficulties of culturing P. vivax in vitro. The lack of an in vitro culture system notably hampers investigations of parasite cell and developmental biology, biochemistry, and the physiology of host cell and parasite interactions by decreasing the availability of the parasite to most laboratories. Rapidly advancing genomics technologies have led to a growing number of P. vivax whole genome sequences [3–6]. In-depth characterization of multi-gene families [3], identification of single nucleotide polymorphisms [3,5], gene rearrangements [7] and previously uncharacterized genes [8] have for example, provided the molecular foundations to prompt new hypotheses and studies on this important parasite. However, testing these hypotheses in vivo remains difficult and, currently, our best opportunity to investigate P. vivax biology may be through P. vivax parasites that have been adapted for propagation in New World monkeys [9].
Monkey-adapted P. vivax strains are typically generated by direct injection of parasitized erythrocytes from patients or, after passage through mosquitoes, by the injection of sporozoites dissected from infected mosquito salivary glands into Saimiri or Aotus monkeys [10]. Once infections are stably established by serial passage, these strains can be continuously propagated in monkeys by initiating further infections using sporozoites or infected erythrocytes, which can be cryopreserved for later use. These parasites are extremely useful to obtain large amount of proteins or nucleic acids from a single strain and can be shared among researchers to investigate various aspects of the parasite biology. However, important questions regarding their biological relevance and homogeneity remain unanswered.
It notably remains unclear whether the host switch, from humans to New World monkeys, induces or requires specific genomic changes. While P. vivax-like parasites have been identified in great apes, to date, genomic studies have indicated that these parasites belong to a clearly distinct sister clade, basal to the human P. vivax [11] and suggest that P. vivax are specific to humans. In addition, many attempts to adapt P. vivax to New World monkeys fail to result in detectable levels of the parasite [12], alter the parasite life cycle [13] or are only successful in a specific monkey species or subspecies [13,14]. (Note that once a strain has been successfully adapted, it can typically be more easily propagated in subsequent monkeys.) These observations suggest that the molecular mechanisms used by P. vivax to invade and survive the metabolic environment of red blood cells (RBCs) and evade the host innate and adaptive immune responses have been tuned to humans by thousands of years of evolution and might be maladapted to New World monkey physiology and RBCs. Successful adaptation to the new environment of New World monkey RBCs could therefore require subtle changes throughout the genome. Interestingly, P. vivax does not seem to be able to infect Old World monkeys, although these primates are more closely related to apes than the New World monkeys. On the other hand, there are clear indications that Old World primate malaria parasites can infect humans [15–18] despite consequent differences in genome sequences [19,20]. Note however that these infections are not usually as robust as in the natural hosts and that these parasites’ genomes have not been examined after passage in humans. Independently of the host switch, the continuous serial blood stage propagation of adapted parasites in New World monkey may also induce genomic changes as some genes become dispensable in this setting. For example, the Vietnam IV Palo Alto strain is not able to infect mosquitoes [21] suggesting that some genes underlying infectivity to mosquitoes might have been altered during propagation in monkeys. In this regard, it is important to note that many genomic rearrangements have been documented during the propagation of P. falciparum in in vitro cultures [22,23]. Finally, once an isolate becomes a monkey-adapted strain it is often unknown whether it consists of a single homogeneous clonal parasite population (i.e., a single “genotype”, later referred to as a clone) or a complex infection as observed in genomic analysis of field isolates [5] and numerous field studies (see e.g., [24]).
In this study, we compare the genomes of seven monkey-adapted strains with the genomes of six field isolates to characterize genomic changes that potentially occur during adaptation to New World monkeys and continuous propagation. We also analyze six different samples collected during the generation of the Mauritania-I and Mauritania-II strains. These analyses provide additional insights regarding the homogeneity of monkey-adapted strains and the changes that occur during the establishment and propagation of these strains.
For our analyses, we used genome sequence data previously generated from seven monkey-adapted strains: the Salvador-I [25], Belem [5], Chesson [8], Brazil-I [3], India-VII [3], Mauritania-I [3], and North Korean [3] strains. We compared these sequences with data from six previously sequenced field isolates from Cambodia and Madagascar (M08, M15, M19, C08, C15, and C127) [5,7]. For some of the analyses, we focused on four of these field isolates (M08, M15, C08, and C127) that carry one single highly dominant clone and therefore allow inference of the entire haploid genome sequence (see supplemental information in [5] for details). Several sequencing runs were independently produced for the samples sequenced at the Broad Institute and we used, for most of our analyses, those generated using 101 bp paired-end reads (as these are most similar to the data we generated). The remaining libraries were only used to assess sequencing error hotspots and unannotated paralogous sequences (see below). Detailed information on the samples and sequencing libraries used is provided in S1 Table.
In addition, we analyzed sequences from DNA extracted from additional blood samples collected during the generation of the Mauritania-I and Mauritania-II P. vivax strains [26]. Three blood samples (AI-3221, AO-521 and WR-1714) were collected from Aotus nancymaae monkey infections derived directly from the original patient infection in February 1995. The infection in WR-1714 was initiated by sporozoites collected from mosquitoes fed on blood from the patient’s initial infection. DNA of the stabilate of the Mauritania-I strain sequenced by the Broad Institute [3] came from infections of two Saimiri boliviensis boliviensis monkeys, SI-3095 and SI-3097. We also analyzed the blood sample from the patient when a relapse occurred in October 1995 and blood from an Aotus nancymaae monkey (AI-3218) infection derived from this relapse after five direct serial passages in monkeys. The AO-521, WR1714, AI-3218 and patient samples were collected in 1995 and cryopreserved at the Division of Parasitic Diseases of the Centers for Disease Control and Prevention (Atlanta, GA). The AI-3321 specimen was collected in 2006 from a monkey infected by parasitized erythrocytes from AI-653 (that had been cryopreserved since 1995). For all samples, we extracted DNA from 200 μl of cryopreserved blood using the Qiagen DNeasy Blood and Tissue kit according to the manufacturer’s instructions.
We mapped sequencing reads from all samples to the P. vivax Salvador-I [25] reference genome using bowtie2 [27]. We mapped each end of all read pairs independently and considered as correctly mapped only reads best mapped to a single genomic location. Only read pairs for which both ends fulfilled this criterion were included for further analyses. We also identified read pairs that mapped to the exact same positions and randomly discarded all but one pair to eliminate reads representing DNA molecules amplified during the library preparation. In total we examined 13 strains; seven monkey adapted isolates and six human field isolates.
We screened for single nucleotide variants (SNVs) at all nucleotide positions covered by at least 20 reads with a base quality score greater than 30 in all analyzed samples. Regions of high DNA sequence similarity were excluded from our analysis as previously described [5]. Overall, 19.7 Mb or 87% of the Salvador-I reference genome sequence were analyzed. Mismatches (i.e., SNVs) between reads generated from a given sample and the reference genome sequence were determined using samtools mpileup [28] and the extended base alignment quality computation. Positions were considered variable only if at least 10% of the reads from a given sample supported an allele different from the reference nucleotide.
We screened each genome for DNA sequence rearrangements as described in [7]. Briefly, we analyzed all read pairs that did not map in the expected configuration (i.e., head-to-head within 1 kb from each other) and might be indicative of deletions (reads mapping head-to-head but distant by more than 1 kb), inversions (reads mapping in a head-to-tail configuration) and tandem duplications (tail-to-tail) (see S1 Fig. of [7] for details). We then identified regions of the genome with more read pairs in unusual configurations than we would expect by chance (as modeled by a Poisson distribution). To avoid including artifacts occurring during library preparation, we focused on rearrangements greater than 1 kb but smaller than 100 kb.
We also searched for large deletions by scanning for chromosomal regions greater than 100 kb where the sequence coverage was less than 50% of the genome average coverage of the sample. To avoid including regions where reads systematically mapped poorly (due to high DNA sequence divergence or high repeat content), we restricted our analyses to loci that displayed low sequence coverage in some, but not all, of the samples.
To identify positions in the P. vivax genome prone to systematic sequencing errors, we analyzed sequence data generated by the Broad Institute: for some of the monkey-adapted strains, the sequence data came from several independent sequencing reactions generated from the same library (S1 Table). We examined the reference allele frequency (RAF) for nucleotide positions sequenced by more than 50 reads in each of three independent sequencing runs of the Brazil-I, North Korean, and Mauritania-I strains. We then catalogued genome positions that displayed a RAF between 1–10% and 90–99% in all three runs of these three samples. We focused for this analysis on positions where less than 10% of the reads differed from the main allele (i.e., from all other reads) as this corresponds to the peak of RAF observed in Mauritania-I (see Results). These “consistently variable” positions may represent sequencing error hotspots or unannotated paralogous sequences and were filtered out. We considered that remaining nucleotide positions sequenced at high coverage (>150 total reads) and with a RAF between 1–10% and 90–99% in Mauritania-I represented positions where a previously unreported minor clone differed from the major clone. We then reconstructed the haploid genome sequence of this second (minor) clone using the minor allele at these positions (i.e., we assumed that only one minor strain was present in this sample).
In addition to the Mauritania-I strain data generated by the Broad Institute [3], we analyzed blood samples from four additional monkey blood samples collected during the generation of the Mauritania-I strain (derived from the initial patient infection) and Mauritania-II strain (derived from a relapse of the same patient) [26]. We also analyzed blood directly collected from the patient during the relapse. The quality of the DNAs (frozen since 1995 for most samples) and the lack of leukocyte depletion before freezing prevented whole genome sequencing. We therefore designed primers to amplify 38 SNVs distributed across the P. vivax genome and for which we observed two alleles in the Mauritania-I Broad Institute sequence data (S2 Table). Each primer was designed to include a 5’ oligonucleotide tail for barcoding and high-throughput sequencing (see below). We amplified each locus with the following conditions: 94°C for 3 min; 40 cycles of 94°C for 45 sec, 56°C for 45 sec and 72°C for 45 sec; and final extension at 72°C for 3 min. We pooled the 38 amplification products obtained from each blood sample together, purified the DNA pools with Qiagen QIAquick columns and labeled them with an individual oligonucleotide barcode (i.e., one barcode per blood sample) by a second amplification (with the same conditions as previously but with only 10 cycles) using primers targeting the 5’ oligonucleotide tail and containing the Illumina adapter sequence and the unique barcode sequence. The barcoded samples were then pooled together at equal DNA concentrations and sequenced simultaneously on an Illumina MiSeq to generate 32,283,840 paired-end reads of 150 bp (4.4–10.8 million pairs per sample). We mapped the reads on the Salvador I reference genome sequence using bowtie2 and analyzed allelic variations at the SNVs targeted. We discarded from our analysis 11 out of the 38 targeted SNVs due to allelic dropout or insufficient read coverage (<100 X).
We analyzed 19.7 million nucleotide positions (87% of the Salvador I reference genome sequence) that have been previously sequenced at more than 20 X in seven monkey-adapted strains and six field isolates and identified 140,949 variable positions. We refer to these variable nucleotide positions as single nucleotide variants (SNVs) as they may include variants that occurred during the adaptation and propagation of the strains in New World monkeys as well as single nucleotide polymorphisms (SNPs). In all monkey-adapted strains and four of the field isolates, one clone of P. vivax accounted for >80% of all P. vivax sequences enabling reconstruction of the entire haploid genome sequence for this clone. To assess whether adaptation to a new host induced systematic genomic changes, we first performed a principal component analysis of all dominant clones using all SNVs identified. Interestingly, P. vivax parasites clustered according to their geographic origin and not to the host species from which the sample was obtained (Fig. 1). Further analyses of the first ten principal components (accounting for 94% of the variance) did not reveal any clustering of samples according to their host. This observation indicated that, at the genome level, the host switch was not a major determinant of the genetic diversity.
Even if host switch did not alter the genetic diversity of P. vivax at the genome-scale, it is possible that a few critical protein coding genes or regulatory elements were systematically modified during the parasite passage from human to monkey hosts. We therefore examined every SNV throughout the genome and tested whether its alleles segregated according to the host. Throughout 19.7 Mb covered by more than 20 high quality reads in all samples, we did not find a single variant (out of 140,949 SNVs) for which one allele was fixed in all monkey-adapted strains and the other allele was fixed in all human field isolates (e.g., a position where all monkey-adapted strains would carry an A and all human isolates a T). Analyses of DNA sequence insertions, deletions or inversions [7][8] also failed to reveal any DNA sequence rearrangement systematically present in all samples from one group and absent from all samples from the other. Overall, our analyses suggested that adaptation to a New World monkey host did not induce systematic genomic changes nor did it leave any consistent signature in the P. vivax genome among the strains evaluated here.
Once adapted to a different host, P. vivax strains can be propagated for years through successive infections of New World monkeys. We therefore wanted to determine whether this propagation could lead to genetic changes. If monkey-adapted strains accumulate mutations during propagation, we would expect that they differed more from each other or from a set reference than field isolates. In contrary, our results showed that there were, on average, 36,297 nucleotide differences (16,683–47,597) between a given monkey-adapted strain and the Salvador-I reference genome sequence and 40,730 differences (38,520–45,306) between human isolates and the reference (p = 0.4).
Another way to test whether long-term propagation in New World monkeys result in the accumulation of mutations is to compare genome sequences from the same strain generated from DNA isolated years apart. We have independently [5] produced sequencing data from the Salvador-I strain used for generating the reference genome sequence [25]. Out of the ~12.2 million bases covered by 20 reads or more in our data and after filtering out ~360 kb of repetitive or potentially paralogous regions (see [5] for details), we observed 3,116 possible SNVs (i.e., positions where >10% of the reads differed from the reference allele) between the genomes of this same strain collected at two time points. However, there were only 8 positions where >90% of the reads generated differed from the reference Salvador I sequence (note that these figures are slightly different from those presented in [5] as we used here a better read mapping algorithm). It is important to note that these differences represented a combination of sequencing errors and possible genuine differences. Overall, these observations suggested that propagation in New World monkeys was unlikely to lead to accumulation of many mutations in the P. vivax genome.
In three of the seven monkey-adapted strains (Belem, Brazil-I and North Korean), we noticed that very few (if any) reads mapped to a 130 kb region at the subtelomeric end of chromosome 7 (Fig. 2). While telomeric and subtelomeric regions are enriched in repeated sequences and therefore difficult to assemble, resequence and analyze, this particular deletion extended far beyond the typical repeat- and AT-rich region and was successfully sequenced in other P. vivax strains. In addition, the GC content along this subtelomeric region gradually decreases with the most abrupt change (from ~40% to ~28% GC) occurring around position 1,411,000, roughly 35 kb downstream of the deletion boundary (Fig. 2). The deleted region contains 22 annotated protein coding genes including a cytoadherence linked asexual protein (CLAG, PVX_086930), an early transcribed membrane protein (ETRAMP, PVX_086915), a Phist protein (PVX_086910), ten hypothetical proteins and nine vir genes. In the Belem and Brazil-I strains, no sequence reads could be aligned to this region suggesting that the entire end of the chromosome had been deleted. The exact demarcation of the deletion did not appear to be identical between these samples, with the deletion starting at base ~1,367,000 in the Belem strain and 6 kb later, at base ~1,373,000, in the Brazil-I and North Korean strains (Fig. 2). This could indicate independent deletion events or continuous trimming of the telomere. Evidence of this subtelomeric deletion in the North Korean strain was supported by a significant, but not complete, reduction in coverage (~75% less reads), suggesting that, within the North Korean strain, some parasites carried the deletion while some had the entire subtelomeric sequence. Interestingly, the reference allele frequency (RAF) profile for the North Korean strain (Fig. 3, light blue) suggested that the two clones in this sample (with or without the deletion) were otherwise genetically identical. This observation suggested that the subtelomeric deletion occurred recently in a clonal population of parasites and that the North Korean strain of P. vivax is not genetically homogeneous anymore. It is important to note that this telomere shortening was not exclusive to monkey-adapted strains but was also observed in one of the minor clones of a field isolate from Cambodia (C15, Fig. 2).
In a previous study we showed that the Salvador-I and Belem displayed reference allele frequency (RAF) distributions consistent with the presence of a single clone [5]. For these samples, all reads covering a given genome position either carried a nucleotide identical to the reference allele or all carried a same but different nucleotide (the alternative allele), with minor alleles represented by less than 5% of the reads likely representing sequencing errors. This pattern was also observed in three out of four monkey-adapted strains sequenced by the Broad Institute [3] (Fig. 3).
In contrast, in the Chesson sample, we detected the presence of a second clone that accounted for approximately 10% of all reads (Fig. 3). At all positions that harbored two alleles for this sample, the minor allele was always identical to the Salvador-I reference allele. In addition, we did not observe a single position with a RAF of 0% (which occurs when both clones are identical and differ from the reference genome) suggesting that, throughout the entire genome, the minor clone sequence never differed from the Salvador-I reference genome sequence. These observations suggested that the Chesson sample we sequenced had likely been contaminated by Salvador-I DNA.
The RAF spectrum of the Mauritania-I strain (Fig. 3) also clearly indicated the presence of a minor clone accounting for ~5% of the P. vivax sequences. Overall, we identified 2,255 nucleotide positions where the two clones present in the Mauritania-I sample differed. This number of differences was much lower than we would expect for two unrelated clones (typically around 30,000 nucleotide differences) and suggested that these clones were likely related (Fig. 4). Analysis of the spatial distribution of these genetic differences revealed that the SNVs differentiating the two clones of Mauritania-I were not randomly spread throughout the genome (as would be expected from a unrelated clone) but instead appeared to be clustered in distinct “blocks” (Fig. 5): 1,969 out of 2,255 differences (87%) were located in 153 regions ranging from 5 kb to 165 kb (and accounting for 3.78 Mb or 20% of the genome sequence). One possible explanation for this block pattern is that several P. vivax clones were present in the original patient infection and that they recombined during the passage through Anopheles mosquitoes in the laboratory (Fig. 6) and that the Mauritania-I sample originally sequenced by the Broad Institute is a mixture of a parental and a recombinant clone.
To confirm the presence of multiple clones in the Mauritania-I sample, we analyzed blood samples collected at different time points along the generation of the Mauritania-I strain (see Fig. 6 and Methods for details). We selected 38 SNVs differentiating the major and minor clones present in the Mauritania-I genome sequence and located in the recombinant blocks. We genotyped these SNVs (see Methods) in three monkey P. vivax infection samples derived from the initial malarial episode (AI-3221, AO-521 and WR-1714, which was infected by sporozoites from mosquitoes fed on the patient’s blood) and two samples from a subsequent relapse of the same patient (blood from the relapsing patient and, after serial passage in New World monkeys, from a later monkey-adapted stabilate, AI-3218). These samples are collectively referred to as Mauritania-I (for the samples derived from the initial infection) and Mauritania-II (for the samples derived from the relapse).
At each of the 27 position successfully genotyped, the monkey samples AI-3221 and AO-521 showed 100% of the reads carrying the same allele indicating that these samples were infected by a single clone (referred to as P1). The sample from the patient relapse and the monkey sample derived from this relapse (AI-3218) also showed genotypes consistent with infection by parasites with the same single genotype as one another. However, this genotype was different from the P1 genotype noted in AI-3221 and AO-521 at 19 of the 27 successfully genotyped SNVs (Fig. 6) indicating that the patient’s relapse parasites and the parasites passaged through AI-3218 were a distinct clone (referred to as P2). Finally, the genotypes generated from the sample WR-1714 showed two alleles at 24 out of 27 positions indicating the presence of multiple P. vivax clones in this sample. This observation confirmed that the two clones detected in the Mauritania-I genome sequence data were genuine (and not the result of a laboratory contamination) since the passage lineage of the sample sequenced by the Broad Institute derives from WR-1714 (Fig. 6). WR-1714 displayed genotypes consistent with the presence of both P1 and P2 clones as well as a third clone (P3) at a much lower frequency (<5%). Overall, our analyses are consistent with the presence of at least three clones (P1, P2, and P3) in the original infection, a single clone (P2) in the patient’s relapse blood specimen (and the subsequent infected monkeys) and the presence of two clones (a predominant P1 clone and a minor recombinant clone of P1 and P3) in the sample sequenced by the Broad Institute (Fig. 6). Note that, since we selected SNPs differentiating the recombinant clone from the major clone (P1) from the Mauritania-I genome data, the recombinant genotype is identical to its parental genotype (P3) at these markers.
The main purpose of this study was to determine whether the adaptation of the human malaria parasites P. vivax to New World monkey hosts resulted in systematic genetic or genomic changes. Overall, our analyses suggested that monkey-adapted strain genomes were not significantly altered and remained representative of the original P. vivax parasite genomes circulating in the blood of the infected patient. In particular, we did not detect any fixed nucleotide differences between field isolates and monkey-adapted strains suggesting that the host switch did not lead to systematic genetic changes. Our analyses relied on the comparison of existing monkey-adapted P. vivax genomes to those of field isolates. A more elegant and straight-forward approach would be to directly compare the genomes of the same P. vivax strain generated from DNA isolated from the original patient and from an infected New World monkey after adaptation. Unfortunately, few laboratories are able to perform such host switch and they do so irregularly, and no matched DNA pairs from previous adaptations were available for genome sequencing.
We have also tested whether monkey-adapted strains accumulate mutations during continuous propagation in monkeys. The mutation rate during asexual reproduction of P. vivax remains unknown and long-term culture studies similar to those performed in P. falciparum [29] are not necessarily comparable to in vivo propagation. However, analysis of the genome of the monkey-adapted Salvador-I strain sequenced from two New World monkeys separated by at least five consecutive passages revealed a small number of putative genetic changes suggesting a low asexual mutation rate (note that these differences could also originate from sequencing errors). Importantly, most of these nucleotide differences between the Salvador-I reference genome and our later sequence were only supported by a small proportion of the reads and only 8 nucleotide differences were supported by 90% or more of the reads (out of 12 Mb sequenced at more than 20 X in Salvador I). This observation suggested that, despite likely population bottlenecks occurring during the propagation of the Salvador-I strain in different monkeys, few novel mutations (if any) have drifted to fixation and that most of the possible differences observed are only present in a subset of the otherwise clonal parasite population. Studies including multiple passages will be required to confirm these findings and provide a rigorous estimate of the mutation rate during asexual reproduction.
One limitation of our analyses is that we excluded regions of the P. vivax genome where high DNA sequence homology or unannotated paralogous sequences greatly complicates unambiguous read mapping and SNP calling. While we analyzed here 87% of the P. vivax reference genome, it is possible that unidentified mutations occurred, during adaptation and propagation of these strains in monkeys, in the remaining non-unique regions of the P. vivax genome. Similarly, we did not consider short indels for technical reasons and these might represent another source of possible genetic differences unaccounted for in our study.
During our analyses, we observed a large deletion at the subtelomeric end of chromosome 7 in three out of the seven monkey strains, as well as in one Cambodian field isolate. While telomeres are typically difficult to sequence and assemble (and are partially missing in the Salvador-I reference genome sequence), this deletion mostly included unique DNA sequences and contained little repeated sequences. Similar subtelomeric deletions have been reported in P. falciparum, both in field isolates and in vitro cultures (e.g., [22,23]). Interestingly, the chromosome 7 subtelomeric deletion displayed different boundaries in different samples suggesting that i) it resulted from independent events that occurred in the P. vivax population prior to adaptation to New World monkeys or ii) that the telomere was slowly being eroded. In addition, in the North Korean strain we observed genetic heterogeneity for this rearrangement suggesting that a proportion of the parasites in the sequenced sample carried the deletion while the rest of them had the full-length chromosomal sequence. This observation suggested that the subtelomeric loss was recent in this strain (i.e., post adaptation to monkeys) and that it remained polymorphic in this otherwise clonal parasite population. This finding also raised questions regarding the presumed genetic homogeneity of monkey-adapted strains.
One technical factor may artificially influence the heterogeneity of the strains: DNA samples collected from multiple individual monkeys infected with the same strain are often pooled together to obtain enough genetic material for genome sequencing. This procedure may result in laboratory contamination with another strain, especially since these strains are not differentiable without the use of genetic markers. For example, we detected a contamination of the Chesson sample by the Salvador-I strain. Such cross-contamination could have important consequences: sequencing a particular gene may, for example, reveal two different DNA sequences and suggests that there are multiple copies of that gene in this strain.
Finally, we observed in the Mauritania-I sample sequenced by the Broad Institute [3] evidence of genetic heterogeneity, with the presence of at least two genetically distinct clones. Analysis of additional Mauritania-I samples confirmed that multiple clones were present in the original patient infection and revealed that different clones became isolated (or dominant) in different monkeys during the propagation. This observation raises important concerns on the use of monkey-adapted P. vivax strains as different aliquots of the same monkey-adapted strain might actually contain genetically different parasites and therefore might respond differently in in vitro or in vivo assays (e.g., of drug resistance, infection efficiency or virulence). On the other hand, our study illustrates the potential advantages of applying genomic tools to studies of monkey-adapted strains. Identification of multiple clones in a sample is traditionally conducted by genotyping a small number of microsatellites (typically between 5 and 10), which does not have the sensitivity necessary to differentiate closely related clones or identify clones making up less than 10% of the parasites [30]. The resources provided by genomic data now enable genotyping of several dozen of highly informative SNPs and might help in solving phenotypic discrepancies among samples from the same monkey-adapted strain. In addition, the observation of a recombinant clone in the Mauritania-I sample sequenced by the Broad Institute illustrates how application of genomic tools could guide the generation of P. vivax genetic crosses which could lead to major advances in gene mapping in P. vivax (but see also [31]).
The development and maintenance of monkey-adapted P. vivax strains has and will continue to be an essential tool for the study of this important malaria parasite. While we have highlighted some of the hidden problems of monkey-adapted strains, our study also provides great prospects for studying this important resource. The extensive information generated by genome sequencing provides numerous genetic markers that can easily be genotyped in a given sample to monitor the identity, complexity and purity of a given strain and improve studies of monkey-adapted strains.
|
10.1371/journal.pntd.0001795 | Combining Hydrology and Mosquito Population Models to Identify the Drivers of Rift Valley Fever Emergence in Semi-Arid Regions of West Africa | Rift Valley fever (RVF) is a vector-borne viral zoonosis of increasing global importance. RVF virus (RVFV) is transmitted either through exposure to infected animals or through bites from different species of infected mosquitoes, mainly of Aedes and Culex genera. These mosquitoes are very sensitive to environmental conditions, which may determine their presence, biology, and abundance. In East Africa, RVF outbreaks are known to be closely associated with heavy rainfall events, unlike in the semi-arid regions of West Africa where the drivers of RVF emergence remain poorly understood. The assumed importance of temporary ponds and rainfall temporal distribution therefore needs to be investigated.
A hydrological model is combined with a mosquito population model to predict the abundance of the two main mosquito species (Aedes vexans and Culex poicilipes) involved in RVFV transmission in Senegal. The study area is an agropastoral zone located in the Ferlo Valley, characterized by a dense network of temporary water ponds which constitute mosquito breeding sites.
The hydrological model uses daily rainfall as input to simulate variations of pond surface areas. The mosquito population model is mechanistic, considers both aquatic and adult stages and is driven by pond dynamics. Once validated using hydrological and entomological field data, the model was used to simulate the abundance dynamics of the two mosquito species over a 43-year period (1961–2003). We analysed the predicted dynamics of mosquito populations with regards to the years of main outbreaks. The results showed that the main RVF outbreaks occurred during years with simultaneous high abundances of both species.
Our study provides for the first time a mechanistic insight on RVFV transmission in West Africa. It highlights the complementary roles of Aedes vexans and Culex poicilipes mosquitoes in virus transmission, and recommends the identification of rainfall patterns favourable for RVFV amplification.
| Rift Valley fever (RVF) is a zoonotic disease that affects domestic livestock and humans. During inter-epizootic periods, the main infection mechanism is suspected to be through bites by infected mosquitoes, mainly of Aedes and Culex genera. In East Africa, RVF outbreaks are known to be closely associated with heavy rainfall events, unlike in the semi-arid regions of West Africa where the drivers of RVF emergence remain poorly understood. This study brings mechanistic insight to explain why reported RVF outbreaks in Northern Senegal cannot be correlated directly to rainfall. This is done through the use of a rainfall-driven model of RVF vector populations that combines a hydrological model to simulate daily water variations of mosquito breeding sites, with mosquito population models capable of reproducing the major trends in population dynamics of the two main vectors of RVF virus in Senegal, Ae. vexans and Cx. poicilipes. Results show that RVF occurs during years when both species are present simultaneously in high densities. Simulations of inter-annual variations in mosquito populations successfully explained the dates of RVF outbreaks observed between 1961 and 2003.
| Rift Valley fever (RVF) is a vector-borne disease caused by a virus (RVFV) belonging to the Bunyaviridae family, genus Phlebovirus, that affects domestic livestock (e.g., sheep, cattle, camels, and goats) and humans. In humans, RVF can take different forms [1]. Most human cases are characterized by a ‘dengue-like’ illness with moderate fever, joint pain, and headache. In its most severe form, the illness can progress to hemorrhagic fever, encephalitis, or ocular disease with significant death rate. In livestock, it causes abortion and high mortality of newborns and thus induces important direct and indirect economic impacts.
Since the first isolation of RVFV in Kenya in 1930 [2], major RVF outbreaks have been reported in Egypt in 1977–1978 [3] and 1993 [4], in the Senegal River Valley in 1987 [5], [6], in Madagascar in 1990 [7] and 1992 [8], and in northern Kenya and Somalia in 1997, 1998 and 2007 [9]. In 2000, RVF cases were reported for the first time outside the African continent, in Saudi Arabia and Yemen [10]. Recently, a new wave of RVF epidemics occurred in 2006 and 2007 in East Africa (Kenya, Somalia and Tanzania) [11], [12], in Sudan in 2007 [13], in Madagascar in 2008 [14], and in Southern Africa in 2010 [15].
Two main modes of transmission of RVFV are suspected: i) a direct transmission from infected ruminants to healthy ruminants or humans, (ii) an indirect transmission through the bites of infected mosquito vectors [16]. The respective contribution of these different transmission routes remain unevaluated [17]. However, it is assumed that the transmission by the bite of infected mosquitoes is the main infection mechanism during inter-epizootic periods [18].
The number of mosquito species potentially involved in RVFV transmission is very large (more than 30 species), with the main vectors belonging to the Aedes and Culex genera [19]. Because mosquitoes are highly dependent on environmental conditions, the distribution in space and time of RVF is also related to climatic and landscape features. Until now, the ecological areas associated with RVFV transmission were either irrigated or flooded areas located in bushed or wooded savannas of semi-arid areas [20], although a recent study on RVF outbreaks in Madagascar showed possible transmissions in a temperate and mountainous region [17]. In semi arid areas, natural water bodies which are generally full during the rainy season allow the development of Aedes and Culex species [20], [21]. Based on this, climate based models have been developed to predict RVF outbreaks in Eastern Africa [22], [23], and a strong correlation was found between extreme rainfall events and RVF outbreak occurrences in the Horn of Africa [24].
In West Africa, there is strong evidence that the disease is endemic [18]: different RVF outbreaks were reported in ruminants since the severe outbreak in the Senegal River basin in 1987 [25], [26], [27], [28], and RVFV was isolated from mosquitoes [21], [29] (Figure 1a). However, using a statistical approach, the correlation found in East Africa is not valid in the semi-arid regions of West Africa [30], [31] where the drivers of RVFV transmission dynamics remain poorly understood. There, temporary water bodies (ponds) constitute the main oviposition sites of different mosquito species [32], [33] and mosquito population dynamics are assumed to mainly depend on water availability and on pond dynamics, themselves driven by rainfall [34].
In this study, we use a mechanistic modelling approach to better understand the dynamics of RVF transmission in Northern Senegal, in relation to the population dynamics of its two main mosquito vectors in Senegal, Aedes (Aedimorphus) vexans arabiensis [21], [33] and Culex poicilipes [29]. These two species are considered as the main RVF vector in the area because i) they were proven experimentally to be competent for RVF virus transmission [35], [36], [37]; ii) they were frequently found infected in nature and are the most abundant species in our field site [21], [38]; iii) their interaction with the RVF vertebrate hosts (sheep, goats, and cattle) is very important [39]. The dynamics of the two vector species is modelled by combining a hydrological model of the dynamics of the water bodies, with mosquito population models describing different stages of the mosquito life cycle. Once calibrated and validated on recent rainfall, pond water levels, and entomological data, the combined model can be used to simulate the evolution of the two species' populations during the period 1961–2003, using only rainfall data as input. The comparison of model simulations with recorded prevalence rates and RVF outbreaks in the region is then analyzed and discussed.
The study area is an agropastoral zone of northern Senegal (Figure 1b). It is representative of the Ferlo region and is characterized by a complex and dense network of ponds that are filled during the rainy season (from July to mid-October). These water bodies are focal points where humans and livestock have access to water during the rainy season and are also the main breeding sites for Aedes vexans arabiensis and Culex poicilipes mosquitoes.
We used a hydrologic pond model that simulates daily spatial and temporal variations (surface, volume, and height) of temporary ponds in arid areas [40]. The model consists in a daily water balance model taking into account the contribution from direct rainfall, the runoff volumes of inflows and the water loss through evaporation and infiltration. The relation between water volume, surface and height of a given pond depends on the 3D shape of that pond and is modelled by two volume-depth and area-depth empirical equations. Parameters of the model were estimated using detailed bathymetry of representative ponds of the study area and remotely sensed data such as a Digital Elevation Model (DEM) and a very high spatial resolution Quickbird image.
The model was calibrated and validated with field data (water height data and shape profile) collected during the rainy season 2001 and 2002 in the Barkedji area. The application of the model to the ponds (98) of the study area gave fair results both for water height and water area predictions. The comparison of simulated and observed water areas show significant correlations with a coefficient of determination (r2) of 0.89. More details of the hydrologic model are given in [40].
In this study, two sets of rainfall data were used as model input: i) daily rainfall data recorded during the rainy seasons (July–December) 2002 and 2003 with an automatic meteorological collector located in Barkedji village (Figure 1b); and ii) daily rainfall data recorded from January 1961 to December 2001 by the Linguère meteorological station located 30 km from Barkedji (Figure 1a). The output of interest of the hydrologic model for modelling mosquito population dynamics is , the water surface of any pond P at time t.
The mosquito life cycle involves aquatic (egg, larva, and pupa) and aerial (adult) stages. It begins with an egg, which hatches as a larva. Depending on the species and environmental conditions, hatching may occur immediately or may be delayed. The larvae then mature through four stages before entering pupation. After pupation, the mosquito emerges as an adult (imago) at the surface of water. Adults rapidly mate after emergence and females then seek a blood meal necessary for developing their eggs. Following egg development of about three days, females lay eggs on specific humid surfaces (oviposition sites), proceed to a new blood meal, and perform a new gonotrophic cycle, which corresponds to the period between 2 successive egg layings.
The bioecology of Ae. vexans and Cx. poicilipes differs. Cx. poicilipes eggs are deposited directly on water surfaces and immediately proceed through development into larvae; they do not survive dessication. In contrast, Ae. vexans females lay their eggs on the soil just above the current water level [33]. To hatch, the eggs must first dry out for a minimum number of days before being submerged in water. Moreover, in dry Sahelian regions, Cx. poicilipes populations may survive unfavourable conditions of the dry period as adults in dormancy (diapause) whereas Ae. vexans survive as eggs in desiccated mud, that will hatch during the next rainy season [33].
In the context of data scarce regions, we developed a simple model that captured the main features of Ae. vexans and Cx. poicilipes dynamics at the scale of a pond. The sole dynamic input was the water surface area of pond P at a daily time step t, written as . Only female mosquitoes are modelled and the two mosquito populations of each pond are assumed independent. We followed the theoretical framework proposed by Porphyre et al. [41] for Cx. poicilipes populations, and we extended this model to better take into account specificities of the bioecology of Ae. vexans.
The dynamics of the number of adult female mosquitoes of pond P, time step t, , is described by:(1)where is the daily mortality rate, T the developmental period, i.e. the elapsed time during which a newly hatching egg undergoes its development until the emergence of an adult, the number of hatching eggs in the pond P, time step t, and Tdiapause the date when mosquitoes enter into diapause. The production rate of new adults from a pool of hatching eggs is expressed as the product of the mosquito production capacity of the breeding site, , and of the availability function of the pond P, .
The hydrologic model and both Cx. poicilipes and Ae. vexans models were run for two ponds in the study area, Niaka and Furdu (Figure 1b). The two ponds were considered representative of the water bodies in the area, Niaka (363 525 m2) being a large pond located in the main stream of the Ferlo Valley, and Furdu (9 603 m2) being a smaller pond located outside the main stream [40].
The initial Cx. poicilipes adult population was defined proportionally to the pond perimeter covered by vegetation, with an initial density of adults of 1 adult.m−1. The initial number of Ae. vexans eggs was defined proportionally to the pond surface, with an initial density of 1000 eggs.m−2. Simulations started June 1st, at the beginning of the rainy season. The date of diapause was October 1st, according to [46].
A sensitivity analysis was carried out to assess the robustness of the mosquito population model. We used the OAT (one-factor-at-a-time) Morris's method [47], as revised by Campolongo (1999), allowing the estimation of the two-factor interaction [48], [49]. The input parameters and their ranges based on the literature data were used in the analysis. When information was unavailable, the parameters space variation was defined using nominal values ±10% and a uniform distribution. Three outputs have been tested for each species: (1) the cumulated annual abundance, (2) the maximum abundance, and (3) the date of the peak of abundance.
We used field mosquito collection data during two periods, 1991–1996 and 2002–2003 [21], [33], in an area surrounding Barkedji village to 1) calibrate and 2) assess the goodness of fit of the population dynamics models using the coefficient of determination to measure how well the predicted Ae. vexans and Cx. poicilipes abundance values fit with a set of observed mosquito data. The latter were collected at Furdu and Niaka ponds near Barkedji village, every 20 days during the 2002 and 2003 rainy seasons (Figure 1b, Table 2) [34]. The mean number of Culex and Aedes collected per trap over the consecutive nights of a trapping session (between 5 and 9 days) was calculated. The mosquito population model was calibrated for the two species using 2002–2003 Furdu entomological data collection. The parameters identified as most sensitive by the sensitivity analysis were calibrated. The calibration was then performed with a systematic exploration of the input parameters space (Table 3). Other parameter values were determined based on literature data and expert knowledge (Table 1). To validate the model, we then compared observed and simulated relative abundances of Ae. vexans and Cx. poicilipes mosquito populations for the Niaka pond, 2002–2003 period. The degree of association between the temporal series was assessed by the calculation of the cross-correlation coefficient. This statistical index allows to test whether two temporal series are correlated. It returns values ranging from −1 (negative correlation) to 1 (positive correlation).
Between 1991 and 1996, mosquitoes were collected each year monthly between July and November in the Barkedji area with different kinds of traps at different locations [21] (Table 2). We computed the mean number of Cx. poicilipes and Ae. vexans collected per CO2 light trap and per night over the different locations. We used only one type of trap to avoid any trap related bias in the measure of mosquito abundance. CO2 light traps collections were used because those traps were used evenly each year. The degree of association between observed and simulated abundances for each mosquito species was assessed by calculating the cross-correlation coefficient.
Once validated, the models were run over a 63-year period, from 1961 to 2003, using rainfall historical records provided by the meteorological station of Linguère. As output, we considered the dynamics of each mosquito species expressed in relative values, as well as the product of the two temporal series. The latter index expresses the synchronicity of the Ae. vexans and Cx. poicilipes populations and higher values are obtained when the two mosquito populations are both abundant at the same time. It is subsequently referred as the Index of Simultaneous Abundance (ISA).
Finally, we compared and discussed the outputs of the model with the occurrence dates of RVF outbreaks or seroconversion rates reported in Northern Senegal and Southern Mauritania between 1987 and 2003 (Figure 1a) and with the annual prevalence rates recorded between 1989 and 2003 by the FAO sentinel herd system [50].
The sensitivity analysis (SA) allows identifying the key parameters of the population dynamics models for Ae. vexans and Cx. poicilipes species (Figure 2). Overall, the SA showed that the development period T and daily larval survival rate γ, which are both linked to the larval stage, are the parameters with the most effects on model outputs for the two species. Other parameters identified as influential for Cx. poicilipes were Emax and λ, two parameters concerning the oviposition, whereas the other key parameters for Ae. vexans, φ and Td, were related to the desiccation phase. These eight parameters were thus more accurately estimated through the calibration process.
The T, γ, Emax, λ, φ and Td parameter values were estimated from model calibration for Cx. poicilipes and Ae. vexans species on the Furdu pond (Table 2). The comparison of Cx. poicilipes and Ae. vexans observed abundances in 2002–2003 with outputs of the model showed that the model, driven only by rainfall data, reproduces well the major trends in the intra- and inter-annual population fluctuations (Figure 3). With cross-correlation values of 0.78 for Culex, to 0.52 for Aedes, the results of the simulations regarding the dates of the peaks and the proportion of abundance are consistent with entomological field data. When considering Ae. vexans populations, for both years the model reproduces well the first abundance peak of catches occurring at the beginning of the rainy season (July), generally after the first effective rainfall [33]. Moreover, the model simulates well the dates of maximum abundance at the end of the rainy season for Cx. poicilipes in 2002 and 2003. Finally, the model is able to correctly simulate the relative levels of abundance between the two years for the two species (higher Cx. poicilipes and Ae. vexans densities in 2003 than in 2002) (Figure 3).
The comparison of observed and simulated mosquito abundances from 1991 to 1996 confirmed the capacity of the model to assess the inter-annual variability of Cx. poicilipes populations (Figure 4). For instance, the year of highest abundance of Cx. poicilipes observed during this six years period (1993) was clearly identified by the model. However, it failed to simulate the high abundances of Ae. vexans populations observed in 1991 and 1996 (Figure 4), suggesting that the model would only detect very high inter-annual variations in Ae. vexans abundances, like between the years 2002 and 2003. The cross-correlation coefficient values were fair (cor = 0.43 for both species). Finally, considering both population dynamics, the model reflects well the temporal interval between Ae. vexans and Cx. poicilipes dynamics, the former appearing at the very first rain, while the latter is stronger at the end of the rainy season, taking over from the declining Ae. vexans population.
The modelled dynamics of Ae. vexans and Cx. poicilipes populations depict a high inter-annual variability over the studied period (Figure 5). Simulations put into evidence that the abundance of both species vary greatly between years. Moreover, the model shows that the peak of abundance of Ae. vexans populations generally occurs before the peak of Cx. poicilipes populations, depicting Aedes-before-Culex population cycles. Variations of the ISA reveal the variations in the temporal lag between Ae. vexans and Cx poicilipes populations.
The two major RVFV circulation events in northern Senegal and southern Mauritania were recorded in 1987 [25] and 2003 [28]. For these two years the model predicted high ISA values of Ae. vexans and Cx. poicilipes populations. According to this index, 1989 and 1993 also appear as years of simultaneous abundant mosquito populations (Figure 5). This is in agreement with the results of several sero-surveys conducted in the area. Serosurveys in small ruminants performed after 1988 showed an active transmission of RVFV till 1989 [26]. In October 1993, active RVFV transmission was detected in several locations of southern Mauritania, in association with an increase of abortions in small ruminant populations [26] (Figure 1a). That same year, RVFV was isolated from Ae. vexans and Ae. ochraceus species, and from one sheep in Barkedji village [27]. Between 1993 and 2003, no epizootic event was observed but virus circulation was detected in 1998 from Cx. poicilipes populations [29].
The results of our modelling approach are consistent with those of previous studies [21], [29], [34], [51], which argue that the two vector species Ae. vexans and Cx. poicilipes play a major synergistic role in RVFV transmission in Senegal, and that the years of high virus circulation levels coincide with years of high abundances of both mosquito species. In Figure 5 it can be seen that since 1961, years of RVF outbreaks do not coincide with years of highest total rainfall. Previous studies have shown that in West Africa, Ae. vexans and Cx. poicilipes abundance and total rainfall were not correlated [30], [31]. Rainfall variability was suggested to be more important than total rainfall for explaining mosquito populations, as the amount of Ae. vexans and Cx. poicilipes generation depends on the alternation of rainy and dry periods [33]. Our results come in support of these findings and suggestions, by providing evidence that present knowledge on the hydrology of temporary ponds and on mosquito population dynamics, as formalised in a model, is able to explain a large part of the observed mosquito abundance temporal variability. According to the yearly simulations, exceptionally high Aedes population densities were present in 1987 and 2003 (Figure 5). This result strengthens the hypotheses that RVFV may either be introduced by transhumant herds at the beginning of the rainy season or transmitted vertically in Aedes populations (which would explain the maintenance of the virus during inter-epizootic periods [21], [27], [28]), and would be amplified by Aedes populations, relayed by the Cx. poicilipes species [33], when both species are present abundantly at the same time. To a lesser extent, the same pattern can be observed in 1993 (Figure 5).
Due to the limited number of animals monitored, the RVF surveillance system showed limited capacities to correctly detect RVFV circulation and may have failed to detect animal cases [18], [28]. In 1993, RVF outbreaks were reported in Mauritania [26], whereas according to the surveillance system based on sentinel herds, only one sheep specimen was found infected in Barkedji in Senegal [27]. As confirmed by observation data [21], the small simulated Ae. vexans population may explain why no clinical cases were reported that year in Barkedji, suggesting again that the Ae. vexans population does play a major role in the amplification of the virus.
In 1987, the modelled mosquito abundances were the highest for the 1961–2003 period. In 1989, the Ae. vexans and Cx. poicilipes ISA was also very high, although no outbreak was detected. This can be explained by the probably high immunity rate of the ruminant populations following the 1987 outbreak, when animals may have been infected but remained asymptomatic cases. Moreover, since 1987 no other epizootic event led to an epidemic. Thus, although the simulated inter-annual variations in mosquito populations may explain the dates of RVF outbreaks observed between 1961 and 2003, others factors may drive the transition from an epizootic to an epidemic event. One strong possibility is the date of the Eid al-Kabir celebration, which favour very high ruminant concentrations [52], [53] and numerous contacts between humans and potentially viremic animals. Moreover, the co-occurrence in time of the Ae. vexans populations and the arrival of transhumant herds in the study area at the beginning of the rainy season may be crucial for the amplification of RVFV: if there are only few domestic ruminants available at the emergence of Ae. vexans populations, the virus will not spread.
Given the huge and dramatic socio-economic impacts of RVF, as well as its increasing global importance, there is an urgent need to develop appropriate mathematical tools for disease forecasting [18]. Our modelling approach which integrates presently available knowledge on RVF vector biology, is a first step towards the development of a climate-based early-warning system in Senegal which could allow prediction of at-risk periods for RVF, but certainly not the epidemic extent which is driven by human factors [54], [55].
Our results highlight that rainfall, as main driver of the hydrologic dynamics of the main breeding sites of RVF vectors, is a predictive factor of RVF in the studied area. In this respect, RVF in East and West Africa present very similar transmission processes, with water availability driving mosquito populations of the Aedes and Culex genera which have almost the same breeding sites and trophic behaviour [21].
More improvement on the model itself can be sought, as different simplifications have been made to develop a simple and robust model in a context of data poor areas. Improvements of the hydrological model have been discussed in [40]. To model the mosquito population dynamics, we considered water availability as the main constraint driving the population dynamics. Nevertheless, other variables, such as temperature, humidity, and vegetation cover, could be taken into account in the mosquito population model. These variables might impact the survival rates of mosquitoes in aquatic and aerial stages, as well as the RVFV development. Moreover, values of the different parameters, such as the date of diapause, could be better estimated from entomological data relative to Ae. vexans and Cx. poicilipes in Senegal.
For the first time, mechanistic insight is provided in this study to explain why reported RVF outbreaks in Northern Senegal cannot be correlated directly to rainfall, as it is the case in East Africa. This is done through the use of a rainfall-driven model of RVF vector populations that combines a hydrological model to simulate daily water variations of mosquito breeding sites, with mosquito population models capable of reproducing the major trends of population dynamics of the two main vectors of RVFV in Senegal, Ae. vexans and Cx. poicilipes. Results show that RVF occurs during years when both species are present simultaneously in high densities. These occur when the rainfall temporal patterns result in water variations in the pond that are favourable for the reproduction of both mosquito species, i.e., abundant rains occurring at regular intervals throughout the rainy season. The combined model can now be used in simulation studies for identifying which rainfall patterns would result in the simultaneous abundance of both species (high ISA), so that operational real-time rainfall-based monitoring systems can be developed.
|
10.1371/journal.pbio.1001735 | Heat Shock Transcription Factor σ32 Co-opts the Signal Recognition Particle to Regulate Protein Homeostasis in E. coli | All cells must adapt to rapidly changing conditions. The heat shock response (HSR) is an intracellular signaling pathway that maintains proteostasis (protein folding homeostasis), a process critical for survival in all organisms exposed to heat stress or other conditions that alter the folding of the proteome. Yet despite decades of study, the circuitry described for responding to altered protein status in the best-studied bacterium, E. coli, does not faithfully recapitulate the range of cellular responses in response to this stress. Here, we report the discovery of the missing link. Surprisingly, we found that σ32, the central transcription factor driving the HSR, must be localized to the membrane rather than dispersed in the cytoplasm as previously assumed. Genetic analyses indicate that σ32 localization results from a protein targeting reaction facilitated by the signal recognition particle (SRP) and its receptor (SR), which together comprise a conserved protein targeting machine and mediate the cotranslational targeting of inner membrane proteins to the membrane. SRP interacts with σ32 directly and transports it to the inner membrane. Our results show that σ32 must be membrane-associated to be properly regulated in response to the protein folding status in the cell, explaining how the HSR integrates information from both the cytoplasm and bacterial cell membrane.
| All cells have to adjust to frequent changes in their environmental conditions. The heat shock response is a signaling pathway critical for survival of all organisms exposed to elevated temperatures. Under such conditions, the heat shock response maintains enzymes and other proteins in a properly folded state. The mechanisms for sensing temperature and the subsequent induction of the appropriate transcriptional response have been extensively studied. Prior to this work, however, the circuitry described in the best studied bacterium E. coli could not fully explain the range of cellular responses that are observed following heat shock. We report the discovery of this missing link. Surprisingly, we find that σ32, a transcription factor that induces gene expression during heat shock, needs to be localized to the membrane, rather than being active as a soluble cytoplasmic protein as previously thought. We show that, equally surprisingly, σ32 is targeted to the membrane by the signal recognition particle (SRP) and its receptor (SR). SRP and SR constitute a conserved protein targeting machine that normally only operates on membrane and periplasmic proteins that contain identifiable signal sequences. Intriguingly, σ32 does not have any canonical signal sequence for export or membrane-integration. Our results indicate that membrane-associated σ32, not soluble cytoplasmic σ32, is the preferred target of regulatory control in response to heat shock. Our new model thus explains how protein folding status from both the cytoplasm and bacterial cell membrane can be integrated to control the heat shock response.
| The heat shock response (HSR) maintains protein homeostasis (proteostasis) in all organisms. The HSR responds to protein unfolding, aggregation, and damage by the rapid and transient production of heat shock proteins (HSPs) and by triggering other cellular protective pathways that help mitigate the stress. Although the specific HSR is tailored to each organism, chaperones that mediate protein folding and proteases that degrade misfolded proteins are almost always included in the core repertoire of induced protein and are among the most conserved proteins in the cell. These HSPs maintain optimal states of protein folding and turnover during normal growth, while decreasing cellular damage from stress-induced protein misfolding and aggregation. Malfunction of the HSR pathway reduces lifespan and is implicated in the onset of neurodegenerative diseases in higher organisms [1]–[3].
In E. coli and other proteobacteria, σ32 mediates the HSR by directing RNA polymerase to promoters of HSR target genes [4]–[9]. Given the importance of this response and the necessity for a rapid but transient increase in expression of HSPs, it is not surprising that regulation of the HSR across organisms is complex. σ32 is positively regulated by a feed-forward mechanism in which exposure to heat melts an inhibitory mRNA structure enabling high translation of σ32 mRNA [10],[11] and is negatively regulated by two feedback loops [12] mediated through members of the σ32 regulon (Figure 1A). σ32 activity is coupled to the cellular protein folding state via a negative feedback loop executed by the two major chaperone systems, DnaK/J/GrpE and GroEL/S. There is extensive support for the model that free chaperones directly inactivate σ32 and that these chaperones are titrated by unfolded proteins that accumulate and bind chaperones during a HSR. Depletion of either chaperone system or overexpression of chaperone substrates leads to an increase in σ32 activity, and conversely, overexpression of either chaperone system decreases σ32 activity [13],[14]. Inhibition is likely direct, as DnaK/J and GroEL/S bind σ32 in vitro and inhibit its activity in a purified in vitro transcription system [13],[15]–[17]. σ32 stability is controlled by the inner membrane (IM) protease FtsH: deletion of the protease stabilizes σ32 [18]–[20], and FtsH degrades σ32 in vitro, albeit slowly [18],[20]. DnaK/J and GroEL/S also regulate stability, as their depletion leads to σ32 stabilization in vivo [13],[14],[21], although this finding has not yet been recapitulated in vitro [22].
Despite the regulatory complexity of the current model, it inadequately addresses two issues that are central to our understanding of the circuitry controlling the HSR, motivating us to search for additional players in the response: (1) Exhaustive genetic screens for mutations in σ32 that result in misregulation have identified a small cluster of four closely spaced amino acid residues (Leu47, Ala50, Lys51, and Ile54), of which three are surface exposed, as well as a somewhat distant fifth residue that abuts this patch in the folded σ32 structure. When these residues are mutated, cells have both increased level and activity of σ32, indicating that this region is involved in a central process required for operation of the negative feedback loops that control both the activity and degradation of σ32 (Figure 1A) [23]–[25]. However, the phenotypes of these mutants are not recapitulated in vitro, where both FtsH degradation and chaperone-mediated inactivation of mutant and WT σ32 are experimentally indistinguishable [25],[26]. Thus, we do not understand how this “homeostatic control region” of σ32 functions. (2) σ32 is thought to monitor the folding status of IM proteins as well as cytoplasmic proteins, but the mechanism for this additional surveillance is unknown. Their close connection is indicated because (1) the IM protease, FtsH, not only degrades σ32, but also maintains quality control in the IM by degrading unassembled IM proteins; (2) induction of the HSR is a very early response to perturbations in the co-translational membrane-trafficking system that brings ribosomes translating IM proteins to the membrane [27]–[29]; and (3) IM proteins are significantly overrepresented both in the σ32 regulon [30] and in an unbiased overexpression screen for HSR inducers [30].
In this report, we identify the co-translational protein targeting machinery, comprised of the Signal Recognition Particle (SRP; Ffh protein in complex with 4.5S RNA; Figure 2A) and the SRP Receptor (SR; FtsY), as a regulator of σ32. We show that SRP preferentially binds to WTσ32 compared to a mutant σ32 with a defective homeostatic control region. We further show that a fraction of σ32 is associated with the cell membrane and that both the SRP-dependent machinery and the homeostatic control region of σ32 are important for this localization. Lastly, the regulatory defects in HSR circuitry caused by mutation of either the σ32 homeostatic control region or the co-translational targeting machinery are circumvented by artificially tethering σ32 to the IM. We propose that SRP-dependent membrane localization is a critical step in the control circuitry that governs the activity and stability of σ32. Membrane localization is widely used to control σ factors, but this is the first case where the IM-localized state is used for dynamic regulation rather than as a repository for an inactive protein.
To identify additional players involved in activity control of σ32, we carried out a genetic screen for transposon mutants with increased σ32 activity under conditions that inactivate σ32 in wild-type cells (see Methods). To impose a condition that mimics the negative feedback control of σ32, the DnaK/J chaperones were overexpressed from an inducible promoter at their chromosomal locus. Under these conditions, a σ32-regulated lacZ chromosomal reporter (PhtpG-lacZ) is expressed so poorly that cells do not make sufficient β-galactosidase to turn colonies blue on X-gal indicator plates. We screened for blue colonies, indicative of a defect in σ32 inactivation. A conceptually similar screen previously identified mutations in the DnaK/J chaperones—key negative regulators of the σ32 response [31]. In addition to re-identifying these components, we found an insertion in the promoter region of ftsY (pftsY::Tn5), located 39 bp upstream of the ftsY open reading frame. The pftsY::Tn5 strain had a 3- to 4-fold reduction in the level of FtsY, the SR, and a ∼7-fold increase in the activity and amount of σ32 relative to WT (Table 1). Defects were complemented by a plasmid carrying ftsY. Unlike WT, in the pftsY::Tn5 strain σ32 activity did not respond to increased chaperone expression. Upon chaperone overexpression in WT cells, the specific activity (S.A.) of σ32 fell to 0.3, relative to that in cells growing without chaperone overexpression. In contrast, upon chaperone overexpression in pftsY::Tn5 cells, the S.A. of σ32 did not change, suggesting a defect in chaperone-mediated activity control in that strain (Table 1). This finding raised the possibility that the high activity of σ32 in pftsY::Tn5 resulted from disruption of activity control of σ32, rather than reflecting a cellular response to accumulation of unassembled membrane proteins.
We tested whether σ32 binds to either FtsY (SR) or to Ffh, the protein component of SRP. Ffh is a two-domain protein, comprised of an M-domain that binds the signal sequence and 4.5S RNA, and an NG-domain that binds to SR, the ribosome, and GTP (Figure 2A). We first used co-immunoprecipitation analysis. Interacting proteins were immunoprecipitated with antibodies against either FtsY or Ffh and, following resolution on SDS-PAGE, antibodies against σ32 or σ70 were used to probe for the presence of these proteins. σ32 was detected in the immunoprecipitations (Figure 2B, lanes 7 and 8), and this signal was dependent on the presence of σ32 in the strain (Figure 2B, lanes 1–4). By contrast, σ70, although much more abundant than σ32 in the cell, did not interact with either SRP or SR (Figure 2B, Lanes 3,4 and 7,8), indicating that interaction with SRP is not a general property of σs. It was not surprising that σ32 was co-immunoprecipitated with both SRP and SR, as the latter two components interact in vivo. To determine the direct binding partner of σ32, purified Ffh and FtsY were resolved on SDS-PAGE, transferred to nitrocellulose, and incubated with purified σ32. Antibodies against σ32 detected σ32 present at the molecular weight corresponding to Ffh but not SR (Figure 2C). In a reciprocal experiment, purified σ32 was resolved on SDS-PAGE, transferred to nitrocellulose, and incubated with purified Ffh or SR. Ffh, but not SR, bound σ32 (unpublished data). Similar studies did not reveal an interaction between σ70 and either Ffh or SR (unpublished data). We determined which Ffh domain binds σ32 by partially-proteolyzing Ffh to produce an 18 kDa M-domain and a 38 kDa NG-domain, resolving the mixture by SDS-PAGE, transferring to nitrocellulose, and probing with σ32. σ32 was detected at the position of full-length Ffh and the M-domain, but not at the position of the NG-domain (Figure 2D), indicating that the M-domain contains the determinants mediating the σ32-interaction.
We used in vivo crosslinking to validate the direct interaction of SRP (Ffh+4.5S RNA) and σ32. We created a σ32 derivative with an N-terminal 6×HIS-tag and a photoreactive amino acid analog (pBPA) at amino acid position 52 (6×HIS-σ32T52pBPA; see Methods), which is active as WTσ32 in expression of the σ32 reporter PhtpG-lacZ (activity is 150% that of WT; within the range of the variability of the assay; unpublished data). Following UV irradiation of whole cells, anti-Ffh immunoblotting of the whole cell lysate detected one predominant crosslinked product, which was dependent on UV-irradiation (Figure 3A, lanes 1 and 2) and pBPA at position 52 (Figure 3A, lanes 2 and 4). This UV- and pBPA-dependent product was also detected with anti-σ32 immunoblotting (Figure 3A, lane 6). To determine whether the crosslinked product represented 6×HIS-σ32T52pBPA-Ffh, we determined whether this product was identified both by co-immunoprecipitation with anti-Ffh antisera (Figure 3B) and by affinity purification of 6×HIS-σ32T52pBPA on a TALON resin (Figure 3C). Upon immunoprecipitation with anti-Ffh antisera, we detected a single higher molecular mass band, which reacted with both anti-Ffh (Figure 3B, lane 2) and -σ32 (Figure 3B, lane 6). Upon affinity purification on a TALON resin, anti-Ffh identified the same predominant UV- and pBPA-dependent Ffh-containing crosslinked product (compare Figure 3B and 3C, lane 2). Importantly, no free Ffh was recovered following TALON purification, indicating that the recovery of the Ffh conjugate was mediated by the covalently linked 6×HIS-σ32, rather than interaction with either the TALON resin or another protein. These results strongly suggest that σ32 directly interacts with Ffh in vivo. Although only a faint band was seen at the same position using anti-σ32 immunoblotting, this was likely a result of high background in this area of the gel, possibly because of extensive interaction between chaperones and σ32 (Figure 3C, lanes 5–8).
The function of the homeostatic control region of σ32 is not known [25]. I54Nσ32 is a mutation located in this region is severely compromised in both activity and degradation control, but the mechanism responsible for this phenotype had not yet been determined [25]. We therefore compared the binding of WTσ32 and I54Nσ32 to SRP using gel filtration. We incubated WTσ32 or I54Nσ32 either alone or in combination with SRP and subjected the mixture to gel filtration. Analysis of the elution profiles demonstrated that most WTσ32 was shifted towards the higher molecular weight region in the presence of SRP, and additionally, a fraction of σ32 eluted at a higher molecular weight than that of SRP alone, indicative of an SRP–σ32 complex [compare A280 profiles of σ32, SRP, and SRP-σ32 (Figure 4A) with immunoblotting for σ32 (Figure 4B; rows 1,2)]. σ32 present at a molecular weight between σ32 and SRP likely represents transient forms of the σ32–SRP complex. In sharp contrast, an interaction between I54Nσ32 with SRP was almost undetectable [compare A280 profiles of I54Nσ32 and SRP (Figure 4A) with immunoblotting for I54Nσ32 (Figure 4B; rows 3,4)], indicating that I54Nσ32 bound more weakly to SRP than WTσ32. Neither WTσ32 nor I54Nσ32 interacted detectably with Ffh, indicating that differential binding is dependent on the formation of SRP (Ffh+4.5S RNA), the biologically relevant cellular species of Ffh.
The biological function of SRP is co-translational protein targeting, leading us to test whether σ32 may be targeted to the IM through an SRP-dependent mechanism. Rapid degradation by FtsH normally keeps σ32 levels very close to the detection limit (∼20–50 molecules/cell; [8]), making reproducible detection following fractionation very difficult. Therefore, we performed fractionation experiments (Figure 5), either in cells expressing an enzymatically inactive mutant of the FtsH protease (FtsH E415A) or in cells lacking FtsH altogether (ΔftsH). Approximately 44% of σ32 fractionated to the membrane in a ΔftsH strain, and this fraction was increased to ∼58% in the FtsH E415A strain, raising the possibility that FtsH itself may participate in retention of σ32 at the IM. As the β′ subunit of RNA polymerase, a known interaction partner of σ32, also fractionated with the membrane, we next tested whether σ32 association with the IM was dependent on its association with RNA polymerase. To this end, we used σ32Δ21aa, which is defective in interacting with RNA polymerase [32]. We confirmed that σ32Δ21aa did not detectably interact with RNA polymerase (Figure S1A,B). Yet endogenous WTσ32 and ectopically expressed σ32Δ21aa fractionated equivalently to the IM both in ΔftsH cells (∼39%) and in FtsH E415A cells (∼58%) (Figure S2), indicating that σ32 transited to the membrane independent of RNA polymerase.
We next tested whether the pftsY::Tn5 mutation or the homeostatic control region mutation of σ32 disrupted membrane partitioning of σ32. Both WTσ32 and ectopically expressed σ32Δ21aa were defective in partitioning to the IM in pftsY::Tn5 cells (Figure 5). To look at the effect of disrupting the homeostatic control region on membrane fractionation, we expressed I54Nσ32 as a σ32Δ21aa variant (I54Nσ32Δ21aa). The size difference allowed us to compare I54Nσ32Δ21aa and WTσ32 in the same cells (Figure S2). Whereas WTσ32 exhibited normal fractionation, I54Nσ32Δ21aa showed a severe localization defect, comparable to that of pftsY::Tn5 cells (Figure 5). We conclude that σ32 targeting to the IM is dependent on both SRP/SR and the σ32 homeostatic control region.
SecA is an ATP-fueled motor protein that recognizes signal peptides, drives the translocation of secreted proteins through the Sec translocon [33]–[37], and collaborates with the SRP/SR for integration of a subset of IM proteins into the membrane [33],[38]. We previously found that σ32 activity is increased in a SecA(ts) strain [39]. This observation motivated us to explore the relationship of SecA to IM trafficking of σ32. Indeed, using a SecA(ts) mutant with general defects in protein export (SecAL43P) [40],[41], we observed that cells displayed a significant defect in membrane localization of σ32 (Figure 5), as well as increased σ32 activity ([39] and unpublished data). In addition, purified SecA, resolved on SDS-PAGE and transferred to nitrocellulose, showed binding affinity for σ32, suggesting that these two proteins interact (Figure S3). We conclude that SecA participates in trafficking of σ32 to the IM.
SecY forms the core of the SecYEG IM translocon. This multidomain protein has a large cytoplasmic domain (C5) that functionally interacts with SR [42], SecA, and the ribosome [43]–[50] (Figure 6A). We tested whether 10 previously described secY mutations located in various domains of SecY (Figure 6A) [51] perturb chaperone-mediated control of σ32 activity and trafficking of σ32 to the IM (Figure 6B). All mutants had enhanced σ32 activity. This result was not surprising as secY mutants are expected to accumulate secretory protein precursors that titrate chaperones [52]. Importantly, four mutants (secY124, secY351, secY40, secY129) were also defective in chaperone-mediated control of σ32 activity (Figure 6B), as indicated by a lack of down-regulation of σ32 activity in response to overexpression of one or both of the chaperone systems. We examined the secY351 mutant, which had both high σ32 activity and a significant defect in chaperone-mediated inactivation, and found it to be defective in IM trafficking of σ32 (Figure 5). secY40 and secY351 affect domain C5 (Figure 6A), implicated in the interaction of SecY with SR, raising the possibility that this interaction is important for both homeostatic control and IM targeting of σ32.
Alkaline phosphatase is active only in the periplasm, where it forms the disulfide bonds necessary for its activity. Therefore, translational fusions to alkaline phosphatase (PhoA) lacking its own export signal are commonly used as an indicator of membrane targeting by the appended N-terminal sequence [53]. If the appended N-terminal sequence has either an export or insertion sequence, the fusion protein will exhibit alkaline phosphatase activity in vivo because it is partly transported to the periplasmic side of the membrane through the SecYEG translocon. Although σ32 has neither a membrane insertion nor an export sequence, it may contain a sequence that targets it to the cytosolic face of the IM. There is some evidence that the secretory apparatus can recognize the mature domains of exported proteins at low efficiency [54]. If so, proximity of PhoA to the translocon resulting from the IM targeting signal might enable transit of some fraction of PhoA to localize to the periplasmic side of the membrane, where it is active. By random insertion of the transposon probe TnphoA into rpoH, encoding σ32 (see Materials and Methods), we found that a phoA fusion to the first 52 amino acids of σ32 (N52-σ32-PhoA) showed ∼10-fold greater PhoA activity than signal-less PhoA itself, indicating that the N-terminus of σ32 facilitates PhoA export (Table 2). Moreover, PhoA activity enhancement is dependent both on the SRP/SR-dependent trafficking system and on SecY, as both pftsY::Tn5 and secY351 decreased the PhoA activity ∼2-fold, whereas leaderless PhoA exhibited little response to these perturbations (Table 2). Thus, this assay is consistent with the idea that the N-terminus of σ32 carries an IM-trafficking sequence and that the targeting process is dependent on SRP and SecY.
The I54Nσ32 mutant and mutants in the IM-targeting machinery (pftsY::Tn5, secA(ts), secY351) were both defective in proper regulation of σ32 and in σ32 association with the IM. This convergence motivated us to test whether artificially tethering σ32 to the IM could restore homeostatic control. To this end, we exploited the bacteriophage Pf3 coat protein. With the addition of three leucine residues in its membrane-spanning region, 3L-Pf3 translocates spontaneously in an orientation-specific manner to the IM, where it inserts in an N-out/C-in orientation [55]. We modified rpoH (encoding σ32) at its chromosomal locus to encode a σ32 variant with the 3L-Pf3 membrane-insertion signal attached to its N-terminus (schematized in Figure S4A). Strains carrying 3L-Pf3-σ32 (IM-WTσ32) or 3L-Pf3-I54Nσ32 (IM-I54Nσ32) as their sole source of σ32 were viable, even though 99% of IM-WTσ32 was inserted in the membrane as judged by fractionation studies (Figure S4B). Thus, σ32 functions when it is tethered to the IM.
We determined whether IM-WTσ32 was subject to homeostatic control circuitry exhibited by WTσ32. σ32 is maintained at a low level by FtsH degradation, and its activity is decreased by chaperone-mediated inactivation. Both phenotypes are evident by comparing the amount and activity of σ32 in a WT versus a ΔftsH strain. In a ΔftsH strain, the level of WTσ32 increases ∼30-fold because the major protease degrading σ32 is removed (Table 3; Figure S5 [compare lanes 1 and 3]; and [25]). However, the activity of σ32 increases only 3-fold as a consequence of chaperone-mediated activity control, leading to a 10-fold reduction in the S.A. of σ32 in ΔftsH cells relative to that in WT cells (Table 3 and [56]). Both the level and S.A. of WTσ32 and IM-WTσ32 were closely similar in a ΔftsH strain, indicating that the chaperone-mediated activity control circuit is active in IM-WTσ32 (Table 3 and Figure S5 [compare lanes 3 and 4]). Additionally, the level of IM-WTσ32 was significantly lower in ftsH+ than in a ΔftsH strain, indicating that IM-WTσ32 was efficiently degraded by FtsH (Table 3 and Figure S5 [compare lanes 2 and 4]). The presence of a contaminating band prevented absolute quantification of IM-WTσ32 levels via Western blot analysis (Figure S5). However, if the relative S.A. of IM-WTσ32 and WTσ32 are equivalent in the ftsH+ strain as we found in the ΔftsH strain, then the 2-fold decrease in activity of IM-WTσ32 relative to WTσ32 implies a slight increase in the rate of degradation of IM-WTσ32 relative to WTσ32. Note that the 3L-Pf3 membrane-insertion tag itself is not a signal for FtsH degradation, as the stability of the FliA σ factor, which is closely related to σ32, was unchanged when expressed as 3L-Pf3-FliA (Figure S6). In summary, both the chaperone-mediated activity control circuit and the FtsH-mediated degradation control circuit are active on IM-tethered σ32.
Next, we asked whether the forced and stable tethering of σ32 to the IM bypassed the regulatory defects of I54Nσ32 and the reduced-level SR mutant pftsY:::Tn5. I54Nσ32 is degraded poorly by FtsH as its level was 11-fold higher than that of WTσ32 (Table 3; Figure S5 [compare lanes 1 and 6] and [25]). I54Nσ32 also had compromised chaperone-mediated activity control as the high chaperone levels in this strain did not reduce the S.A. of I54Nσ32 (Table 3; and [25]). In stark contrast, both degradation and activity control were restored when I54Nσ32 was converted to IM-I54Nσ32. FtsH efficiently degraded the membrane-tethered variant: IM-I54Nσ32 was undetectable in ftsH+ cells but present at a high level in ΔftsH cells (Table 3 and Figure S5 [compare lanes 5 and 7]). Additionally, IM-I54Nσ32 and IM-WTσ32 exhibited comparable reductions in relative S.A. of σ32 in ΔftsH cells (Table 3). Stable tethering of σ32 to the IM also bypassed the regulatory defects of pftsY::Tn5 as IM-WTσ32 in the reduced-level SR background was degraded and subject to chaperone-mediated activity control. Indeed, IM-WTσ32 behaved identically in WT and pftsY::Tn5 strains, exhibiting comparable σ32 activity at a protein level below detection (Table 3 and Figure S5 [compare lanes 8 and 9]). Finally, IM-tethering relieved the growth defects of both I54Nσ32 (Figure S7A and C) and of pftsY::Tn5 (Figure S7B, C, and D). In summary, stable tethering of σ32 to the IM restored both homeostatic control and normal growth to cells with a defective σ32 homeostatic control region and to cells with a compromised SRP/SR co-translational targeting apparatus.
Our work has led to a revised model of the HSR circuitry (Figure 1B). σ32 first transits to the IM via an SRP/SR-dependent process and is then subjected to the chaperone-mediated activity control and FtsH-mediated degradation control that have been previously described. This revised model enables the homeostatic control circuit to integrate information on both cytosolic and IM status. Importantly, the efficiency of co-translational protein targeting depends on the cumulative effect of multiple SRP checkpoints including differences in cargo binding affinities, kinetics of SRP-SR complex assembly, and GTP hydrolysis [57]. Multiple checkpoints and the fact that SRP is sub-stoichiometric relative to translating ribosomes (∼1∶100; SRP molecules to translating ribosomes [58]) may allow SRP to modulate the extent of IM-localization of σ32 during times of stress and/or increased protein flux. Thus, σ32 down-regulation through its localization to the membrane could be alleviated when the IM is disturbed or SRP is overloaded in assisting membrane protein biogenesis. This feed-forward mechanism allows the σ32 homeostatic control to sense the state of cytosolic and IM proteostasis before unfolded proteins accumulate to a significant extent. Interestingly, ffh (encoding the protein subunit of the SRP) is a σ32 regulon member as its expression increases at least 3-fold following induction of σ32 either by heat shock or by deletion of dnaK/J ([30] and unpublished data). This could provide an additional connection between σ32 and protein flux to the IM. Finally, and more speculatively, given the demonstrated involvement of SecA in IM targeting of σ32 and its direct interaction with σ32, the σ32 homeostatic control circuit may also monitor protein flux through SecA to the periplasm and outer membrane.
The idea that the high activity of σ32 in the I54Nσ32 homeostatic control mutant and in SRP/SR mutants (eg. pftsY::Tn5) results from σ32 mislocalization to the cytosol and consequent homeostatic dysregulation, rather than from chaperone titration by a buildup of unfolded proteins, is supported by our data. First, forced IM-tethering overcomes the inviability of the I54Nσ32 mutation in the ΔftsH strain background (Table 3), as well as the growth defects of I54Nσ32 and pftsY::Tn5 (Figure S7), suggesting that high expression of σ32 is aberrant and deleterious to cells, rather than required to remodel misfolded proteins. This is reminiscent of previous findings that reduced-function σ32 mutants suppress physiological defects of a ΔdnaK strain [59] and that overexpression of HSPs was deleterious to growth [13],[60]. Second, secY mutants dysregulated in chaperone-mediated activity control were not distinguished by their extent of σ32 induction. This is contrary to the prediction of the chaperone titration model, which posits that secY mutants with the highest σ32 induction would have the highest level of unfolded proteins. These mutants would then be refractory to activity control because the additional chaperones resulting from chaperone overexpression would actually be needed to remodel the misfolded protein burden. We conclude that homeostatic dysregulation of σ32 results from σ32 mislocalization, rather than from the buildup of unfolded proteins.
The molecular nature of IM-localized σ32 remains unclear. Prediction programs [61],[62] do not detect either a signal peptide-like or transmembrane sequence in σ32. We favor the idea that following transit to the IM, σ32 is maintained at the membrane via interactions with other proteins and/or lipid head groups during its short half-life in the cell (30–60″). Indeed, we have already demonstrated interactions between σ32 and several membrane-associated or IM proteins, including SRP, SecA, and FtsH itself. Moreover, the chaperone systems regulating σ32 (DnaK/J/GrpE and GroEL/S) show partial distribution to the membrane [63]–[68], whereas other potential membrane-associated protein partners have not yet been tested for σ32 interaction (e.g., SecY and additional members of the Sec machinery). Each of these proteins could result in partial membrane localization of σ32, as was shown for FtsH where deletion of the protein decreased localization relative to cells with the protease-dead mutation FtsH E415A. Importantly, if σ32 is membrane associated via transient protein–protein and/or protein–lipid interactions, some σ32 may dissociate from the membrane during cell lysis, as was demonstrated for FtsY, another peripheral membrane protein [69],[70]. Therefore, although we report that ∼50% of σ32 is membrane-associated, the fraction of σ32 that is actually IM-localized may be significantly higher.
IM-associated σ32 may provide regulatory flexibility not possible for IM-tethered σ32. For example, during times of high stress, σ32 may be able to dissociate from the membrane to escape homeostatic control. These excursions could be transient if SRP were able to transport σ32 posttranslationally, a possibility suggested by the fact that full-length, fully folded σ32 binds to SRP (Figures 2 and 3 and Figure S1). Additionally, IM-tethered σ32 is more rapidly degraded than IM-associated σ32, suggesting that tethering makes σ32 a better FtsH substrate. This could diminish the ability of the cell to regulate the rate at which FtsH degrades σ32, which is of physiological significance during temperature upshift [8]. The transient reduction in σ32 degradation following increased temperature contributes significantly to the rapid build-up of σ32 during heat shock [8].
Membrane localization is widely used to control σ factors [71],[72]. The inactive B. subtilis SigK pro-protein is membrane inserted; cleavage of its N-terminal pro-sequence releases SigK [73],[74]. Cleavage is coordinated with passage of a checkpoint in spore development to provide just-in-time SigK activity [75]. Additionally, many σ factors are held in an inactive state at the membrane by cognate membrane-spanning anti-σ factors and released as transcriptionally active proteins when stress signals lead to degradation of their anti-σ [71],[76]. IM-localization of σ32 serves a conceptually distinct role as σ32 is equally active in the cytoplasm or at the IM. Instead, the localization process itself is the key regulatory step in two ways: localization is both regulated by protein folding status and is prerequisite for proper function of the homeostatic control circuit.
The SRP-SR co-translational targeting system has an important role in maintaining proteostasis. SRP-SR minimizes aggregation and misfolding of the approximately 20%–30% of proteins destined for the IM, by making their translation coincident with membrane insertion. Our finding, that SRP/SR-mediated transit of σ32 to the IM is also critical for proper control of the HSR, points to a significant new regulatory role for the co-translational targeting apparatus in protein-folding homeostasis. This finding also raises important mechanistic questions. Our in vitro interaction results suggest a direct, but weak, interaction between full-length σ32 and the M-domain of SRP. The prevailing paradigm suggests that the M-domain interacts only with nascent polypeptides with particularly hydrophobic signal sequences. It is possible that σ32 is detected co-translationally, as the Region 2.1 N-terminal α-helical structure, which resembles a hydrophobic signal sequence, may be recognized by the SRP. Alternatively, we note that the SRP chloroplast homolog (cpSRP54) has a dedicated posttranslational targeting mechanism for several fully translated membrane proteins [77], and E. coli SRP, alone or in combination with additional accessory factors (e.g., other σ32 interactors, such as chaperones or SecA), may target mature σ32 to the membrane in vivo. It remains to be determined whether an interaction between full-length σ32 and SRP, or a novel co-translational targeting interaction by the SRP-SR system, mediates transit of σ32 to the membrane.
All strains used were derivatives of the E. coli K-12 strain MG1655, CAG48238 [25],[39]. For chaperone overexpression experiments, mutations were transduced with phage P1 into strains carrying chromosomal Para-groEL/S [78] or PA1/lacO-1-dnaK/J-lacIq [14]. Mutant alleles in secY [51] and secA [39] were transferred to various strain backgrounds through P1 transduction. The SecAL43P mutant used here is a SecA(ts) allele, with general defects in protein export [40],[41]. For propagation and transfer of the R6K pir plasmid, pKNG101, strains DH5σ λpir and SM10 λpir were used, respectively. Plasmids pET21a and pTrc99A were used as expression plasmids. For construction of pRM5 (6×HIS-rpoH), the rpoH gene was PCR-amplified from the chromosomal DNA of W3110 and cloned into the EcoRI-SalI sites of pTTQ18 [79]. Then, the T52amber mutation was introduced into pRM5 by site-directed mutagenesis, yielding pRM17 (6×HIS-σ32T52amber). pEVOL-pBpF (Addgene) carried evolved Methanocaldococcus jannaschii aminoacyl-tRNA synthetase/suppressor tRNA for incorporation of a photoreactive amino acid analog, p-benzoylphenylalanine (pBPA), into the amber codon site. All strains were grown in LB medium. When required, antibiotics were added to the medium as follows: 100 µg/ml ampicillin, 30 µg/ml kanamycin, 20 µg/ml chloramphenicol, and 25 µg/mL streptomycin.
Strain CAG48275 [25], which is ΔlacX74, contains the prophage JW2 (PhtpG-lacZ), and a chromosomal dnaK/J locus driven from PA1/lacO-1 under control of lacIq [14] was grown in LB, induced with 1 mM IPTG to overexpress DnaK/J chaperones, treated with Tn5, and plated at 30°C on X-gal indicator plates containing kanamycin to select for strains containing Tn5. Blue colonies were picked and tested for higher σ32 activity and for feedback resistance to excess DnaK/J [25]. Tn5 insertion sites were determined by DNA sequencing.
Overnight cultures (LB medium) were diluted 250-fold and grown to exponential phase (OD600 = 0.05–0.5). Samples were taken at intervals starting at OD600 = 0.05, and σ32 activity was monitored by measuring β-galactosidase activity expressed from the σ32-dependent htpG promoter, as done previously [25].
The following proteins were purified essentially as described: 6×H-tagged, Strep-6×H-tagged, and untagged WTσ32 or I54Nσ32 [80], FtsY, Ffh, 4.5S RNA [81], and SecA [82]. Chaperones were removed from σ32 with an additional wash containing 10 mM ATP, 10 mM MgCl2, and 25 uM of both peptides, CALLLSAARR and MQERITLKDYAM, synthesized by Elim Biopharmaceuticals, Inc (Hayward, CA).
Cells were grown to OD600∼0.35 in LB medium at 30°C, harvested, washed two times with 1× PBS, resuspended in Lysis Buffer (20 mM Hepes-KOH, 150 mM NaCl, 10 mM EDTA, 10% glycerol, pH 7.5), and lysed by passing 4× through an Avestin EmulsiFlex-C5 cell homogenizer at 15,000 psi. Cellular debris was spun out and the supernatants were incubated with anti-Ffh or anti-FtsY antibodies at 4°C for 14 h by rotation. TrueBlot anti-Rabbit Ig IP Beads (eBioscience) were added and the supernatants rotated for an additional 2 h at 4°C. Immunocomplexes were isolated by centrifugation and washed 5× in Lysis Buffer without EDTA, and eluted in TCA Resuspension Buffer (100 mM Tris (pH 11.0), 3% SDS) containing LDS Sample Buffer (Invitrogen). Proteins were separated by 10% SDS-PAGE, analyzed by immunoblotting using anti-σ70 and anti-σ32 antibodies, and imaged using fluorescent secondary antibodies (as described below).
Detection of a direct protein–protein/domain interaction was carried out exactly as previously described [83]. Proteins were separated on 10% SDS-PAGE. Partially proteolyzed Ffh was obtained by incubating 400 µg of purified Ffh with 4 µg of Glu-C endopeptidase (New England Biolabs) at 25°C in 10 mM Na-HEPES (pH 7.5), 150 mM NaCl, 1 mM DTT, 10 mM MgCl2, and 10% glycerol. An aliquot of the reaction was taken out at various times (0, 5, 10, 15, 30, 45, 60, 120, 180, and 330 min) and stopped by addition of 5× volume of 5× SDS-sample loading buffer. The samples were then analyzed by blot overlay with σ32 as the probe.
In vivo crosslinking experiments were carried out essentially as described previously [84]. Strains of CAG48238 carrying pEVOL-pBpF were further transformed with pRM5 or pRM17. Cells were grown at 30°C in L medium containing 0.02% arabinose and 1 mM pBPA, induced with 1 mM IPTG for 1 h, and UV-irradiated for 0 or 10 min at 4°C. For analysis of whole cell samples, total cellular proteins were precipitated with 5% trichloroacetic acid, solublized in SDS sample buffer, and analyzed by 7.5% SDS-PAGE and immunoblotting.
Co-immunoprecipitations were carried out as follows: UV-irradiated cells were suspended in 10 mM Tris-HCl (pH 8.1) and disrupted by sonication at 0°C. After removal of total membranes by ultracentrifugation, proteins were precipitated with 5% trichloroacetic acid, washed with acetone, and solubilized in buffer containing 50 mM TrisHCl (pH 8.1), 1% SDS, 1 mM EDTA. The samples were then diluted 33-fold with NP40 buffer (50 mM TrisHCl (pH 8.1), 150 mM NaCl, 1% NP40). After clarification, supernatants were incubated with anti-Ffh antibodies and TrueBlot anti-Rabbit Ig IP Beads (eBioscience) at 4°C for 13 h with rotation. Immunocomplexes were isolated by centrifugation, washed 2 times with NP40 buffer and then once with 10 mM TrisHCl (pH 8.1), and dissolved in SDS sample buffer. Proteins were separated by 7.5% SDS-PAGE and analyzed by immunoblotting using anti-Ffh and anti-σ32 antibodies, TrueBlot anti-Rabbit IgG (eBioscience), and Can Get Signal immunoreaction enhancer solution (TOYOBO Life Science, Japan).
For 6×HIS-tag affinity isolation, UV-irradiated cells were suspended in 10 mM Tris-HCl (pH 8.1) containing 6 M urea and disrupted by sonication at 0°C. After clarification by ultracentrifugation, the soluble fraction was loaded onto the TALON resin (TAKARA BIO, Inc., Japan). After washing the resin with wash buffer (50 mM TrisHCl (pH 7.0), 300 mM KCl, 6 M urea, 20 mM imidazole), bound proteins were eluted with wash buffer containing 300 mM imidazole. Proteins were precipitated with 5% trichloroacetic acid, solublized in SDS sample buffer, and analyzed by 7.5% SDS-PAGE and immunoblotting.
Purified proteins were run on a Superdex 200 PC 3.2/30 column, pre-equilibrated with Buffer A (20 mM Tris-HCl pH 8.0, 150 mM NaCl, 10 mM MgCl2, 2 mM DTT). Purified proteins or protein complexes were run with Buffer A at a flow rate of 40 µL/min, and collected fractions were analyzed by SDS-PAGE and immunoblotting for σ32. SRP was formed by incubating purified Ffh with 1.5× molar excess of purified 4.5S RNA on ice for 10 min. To form SRP-σ32 complexes, 3 µM of purified WTσ32 or I54Nσ32 was mixed with 10× molar excess of SRP; proteins were incubated on ice for 30 min before analysis by gel filtration.
A 52-σ32-Tn5PhoA fusion was initially isolated by random screening for PhoA+ clones on PhoA indicator plates—using a strain carrying a TnphoA transposon probe [85] on a low-copy plasmid and Plac-rpoH (encoding σ32) on a multicopy plasmid. The fusion used in this article (N52-σ32-PhoA lacking the transposon but containing the first 52 amino acids of WTσ32) was subsequently constructed by standard recombinant DNA techniques. Direct construction of fusions past amino acid 52 of σ32 was very unstable, precluding their analysis.
Cells were grown to OD600 = 0.3–0.4, harvested, and resuspended in ice-cold Buffer B (10 mM Tris-Acetate (pH 7.4), 10 mM Mg(OAc)2, 60 mM NH4Cl, 1 mM EDTA, supplemented with 1 mM PMSF) to an OD600 of 15. Cells were immediately lysed by passaging the extracts through an Avestin EmulsiFlex-C5 cell homogenizer at 15,000 psi, and subjected to low-speed centrifugation to remove cell debris and un-lysed cells. Membranes were collected by ultracentrifugation in an Optima benchtop centrifuge (Beckman–Spinco) with a TLA 100.3 rotor (60 min; 52,000 rpm; 4°C). The supernatant was saved as the soluble fraction, while the pellet was washed 3× with Buffer B and then resuspended in Buffer C (50 mM HEPES-KOH pH 7.6, 50 mM KCl, 1 mM EDTA, 1 mM EGTA, 0.5% n-Dodecyl β-D-maltoside, and 5% glycerol). Both the soluble and membrane fractions were precipitated in trichloroacetic acid (13% vol/vol), incubated on ice for 30 min, and then overnight at 4°C. Precipitated proteins were then washed with ice-cold acetone and analyzed by SDS-PAGE and immunoblotted for σ32 (Neoclone), β′ (Neoclone), σ70 (Neoclone), RseA [86], and RuvB (Abcam) with fluorescent secondary antibodies (LI-COR Biosciences) used for detection. The percentage of σ32 in each fraction was determined by direct scanning and analyzing bands with ImageJ software (National Institutes of Health).
Cells were grown to OD600 = 0.35–0.45, harvested, and resuspended in ice-cold Buffer D (50 Tris-HCl, pH 8.0, 0.1 mM EDTA, 150 mM NaCl, and 5% glycerol) to an OD600 of 20. Lysozyme was added to 0.75 mg/mL and cells were incubated on ice for 30 min, followed by sonication, then subjected to low-speed centrifugation to remove cell debris and unlysed cells. Lysates were then incubated with pre-equilibrated, pre-blocked (Buffer D containing 5% Bovine Serum Albumin, 0.1 mg/mL dextran) Softag 4 Resin (Neoclone) overnight at 4°C. Bound proteins were washed 3× with Buffer D and eluted with 4× LDS NuPAGE Buffer (Life Technologies). To collect lysates and eluted proteins, 0.05 µM of Strep-6×H-tagged σ32 was added as a loading and blotting control during analysis by SDS-PAGE and Western blotting against σ32.
The 3L-Pf3 genetic sequence was created by carrying out standard polymerase chain reaction using the following overlapping oligos: 5′-atgcaatccgtgattactgatgtgacaggccaactgacagcggtgcaagc-3′, 5′-taccattggtggtgctattcttctcctgattgttctggccgctgttgtgctggg-3′, 5′-aaagaattgcgctttgatccagcgaatacccagcacaacagcggccagaa-3′, and 5′-aagaatagcaccaccaatggtagtgatatcagcttgcaccgctgtcagtt-3′. The stitched oligos were then cloned using TOPO TA cloning (Invitrogen) and sequenced. To construct chromosomal 3L-Pf3-σ32, PCR was carried out to stitch the 3L-Pf3 gene sequence flanked by the first 500 base pairs of the σ32 open reading frame and 500 base pairs upstream of the start codon, and subsequently cloned into the pKNG101 suicide vector. The 3L-Pf3 sequence was then integrated 5′ and in-frame with the chromosomal rpoH gene by double homologous recombination. Counterselection of sacB on pKNG101 was carried out on 10% sucrose media (5 g/L Yeast Extract, 10 g/L Tryptone, 15 g/L Bacto Agar, 10% sucrose) [25],[87]. Clones were sequenced to verify chromosomal integration of the 3L-Pf3 sequence in the correct reading frame.
To construct pTrc99A expressing 3L-Pf3-FliA, flgM and fliA (in that order) were cloned as an operon, with the sequence 5′-ccgtctagaattaaagAGGAGaaaggtacc-3′ added between the two genes in the vector; the Shine-Dalgarno site is designated in uppercase. Two plasmids were created—one with just flgM and fliA, unmodified, and one where the 3L-Pf3 sequence was cloned 5′ to and in-frame with fliA. Clones were sequenced to verify correct sequences and proper reading frame. Expression was from the leaky pTrc promoter, and experiments were only carried out after fresh transformation into the parental CAG48238 strain. Levels of FliA were analyzed by SDS-PAGE and immunoblotting with antibodies against FliA (Abcam).
Cells were re-suspended in equal volumes of Buffer C, with the addition of trichloroacetic acid (final 13% vol/vol), kept on ice overnight, and the precipitate collected by centrifugation. Pellets were washed with acetone and resuspended in 1× LDS NuPAGE Buffer (Life Technologies). Serial dilutions of WT and mutant samples were loaded onto a polyacrylamide gel, and proteins transferred to nitrocellulose membranes. The blots were first probed with primary antibodies and then with anti-primary fluorescence-conjugated secondary antibody (Licor). Immunoblots were scanned at the appropriate wavelengths for detection. Fold increase (protein level experiments) was estimated by comparison with a dilution series of samples from the WT strain. Fold decrease after addition of chloramphenicol (protein stability experiments) was determined by direct scanning and analyzing bands with ImageJ software (National Institutes of Health).
|
10.1371/journal.pgen.1007602 | Homozygous loss-of-function mutations in MNS1 cause laterality defects and likely male infertility | The clinical spectrum of ciliopathies affecting motile cilia spans impaired mucociliary clearance in the respiratory system, laterality defects including heart malformations, infertility and hydrocephalus. Using linkage analysis and whole exome sequencing, we identified two recessive loss-of-function MNS1 mutations in five individuals from four consanguineous families: 1) a homozygous nonsense mutation p.Arg242* in four males with laterality defects and infertility and 2) a homozygous nonsense mutation p.Gln203* in one female with laterality defects and recurrent respiratory infections additionally carrying homozygous mutations in DNAH5. Consistent with the laterality defects observed in these individuals, we found Mns1 to be expressed in mouse embryonic ventral node. Immunofluorescence analysis further revealed that MNS1 localizes to the axonemes of respiratory cilia as well as sperm flagella in human. In-depth ultrastructural analyses confirmed a subtle outer dynein arm (ODA) defect in the axonemes of respiratory epithelial cells resembling findings reported in Mns1-deficient mice. Ultrastructural analyses in the female carrying combined mutations in MNS1 and DNAH5 indicated a role for MNS1 in the process of ODA docking (ODA-DC) in the distal respiratory axonemes. Furthermore, co-immunoprecipitation and yeast two hybrid analyses demonstrated that MNS1 dimerizes and interacts with the ODA docking complex component CCDC114. Overall, we demonstrate that MNS1 deficiency in humans causes laterality defects (situs inversus) and likely male infertility and that MNS1 plays a role in the ODA-DC assembly.
| Although no clear explanation is yet provided for the correct establishment of the left-right body asymmetry in human, animal studies have clearly shown that tiny-hair-like organelles in the ventral node of the embryo—called motile nodal monocilia—beat regularly and play an important role in this process. To date, different genetic variants were identified to cause laterality defects but the list is still incomplete. Here, we describe two loss-of-function variants in MNS1 in individuals suffering from laterality defects and infertility. This finding is particularly important because laterality defects can also be associated with increased risk of congenital heart disease. By examining a specific case with loss-of-function variants in MNS1 and DNAH5 (the latter encodes for DNAH5, a main component of the motor protein complexes providing the ciliary beating), we show that MNS1 plays a role in stabilizing structures in the motile cilium that helps the attachment of motor proteins–such as DNAH5- to the microtubules to ensure a correct beating of the cilia. Further studies of the function of MNS1 may implicate new biological pathways affecting susceptibility to laterality defects and infertility.
| Cilia assemble on most cell types of the human body to perform diverse biological roles [1]. Non-motile primary cilia function in mechano- and chemosensation as well as in photoreception and olfaction, in addition to an essential role in several signal transduction pathways (noncanonical Wnt and Hedgehog pathways) [2]. Motile cilia and flagella exhibit several tissue and cell-type specific functions.
Establishment of the left-right body axis in vertebrates is an evolutionarily conserved process for which in mammals the embryonic node plays an essential role. A rotational leftward flow established by rotational movement of motile cilia is thought to result in transport of signaling molecules to the correct side of the embryo where this signal defines the establishment of the left-right body axes [3]. Analyses of mouse mutants affected by impaired left-right body axes’ development have proven that dysmotility or abnormal differentiation of nodal monocilia correlate directly with laterality defects [4]. Further, laterality defects can also be observed as a consequence of defective signal reception at the embryonic node and/or defects within the left-right patterning signaling pathways itself such as NODAL, Bmp or FGF signaling [3]. Situs abnormalities include situs inversus totalis (mirror-image reversal), left/right isomerism or situs ambiguous also known as heterotaxy. If heterotaxy occurs, congenital heart disease (CHD) can be frequently observed [5–6].
The underlying genetic defects associated with randomization of body laterality in human are therefore heterogeneous, including motile- as well as non-motile ciliary causes and non-ciliary causes. To date, human mutations have been identified in a number of genes including LEFTY1 (MIM 603037), LEFTY2 (MIM 601877) [7], ACVR2B (MIM 602730) [8], CFC1 (MIM 605194) [9], NEK8 (MIM 609779) [10], INV (MIM 243305) [11] and CFAP53 (CCDC11 (MIM 614759) [12–13].
However, laterality defects often occur (>50%) in individuals harboring mutations in genes which are necessary for the proper structure and function of the motile cilia—resulting in a condition clinically known as Primary Ciliary Dyskinesia (PCD; MIM# 244400) due to defective mucociliary clearance in the respiratory system [14]. PCD associated with chronic sinusitis, bronchiectasis and situs inversus is also described as Kartagener’s syndrome (KS) [15]. Male infertility can also be associated with PCD resulting from dysmotile/immotile sperm flagella. Mutations in the majority of the genes cause a combined phenotype comprising randomization of left/right body asymmetry, male infertility and defective ciliary clearance of the airways.
MNS1, a meiosis specific nuclear structural 1 protein, has been identified in the proteomes of human bronchial epithelium [16]. More recently, Mns1-/- mice have been reported that exhibit: i) randomization of left/right asymmetry and laterality defects, ii) male infertility with sperm immotility and sperm flagellar defects, and iii) partial defects of outer dynein arms (ODAs: motor proteins providing the mechanical force for ciliary movement) in tracheal cilia [17].
In this study, through the collaboration of three centers in Germany, Israel and The Netherlands, we analysed 85 individuals with laterality defects but not classical PCD and identified by whole exome sequencing identical loss-of-function (LOF) mutations in MNS1 (NM_018365) in four individuals from three unrelated families. Additionally, we identified a second MNS1 homozygous mutation in a KS affected individual through linkage analysis and homozygosity mapping. Sanger sequencing of 134 PCD-affected individuals with or without laterality defects did not reveal any additional mutations in MNS1.
Given the genetic heterogeneity of laterality disorders and known genes only explaining a subgroup of cases, we performed independent whole exome sequencing (WES) in 85 affected individuals diagnosed with situs abnormalities with or without heart defects but lacking clinical criteria for PCD, under the hypothesis of a recessively inherited and rare causal allele (Table 1). 8 unrelated families were recruited and evaluated in Israel. In 5 families, more than one member was affected. We performed WES in 8 affected singletons, one from each family. Another 36 families were recruited in Muenster (Germany), and DNA from 37 affected individuals including 2 brothers was analyzed by WES. The remaining 40 Turkish individuals are descendants from 40 families recruited in Turkey and WES was performed and analyzed in the Netherlands. The analysis revealed the same homozygous stop mutation rs185005213, c.724T>C, p.Arg242* in MNS1 in three affected individuals: 2 from Israel (AL-IV-III, BG-II-1) and 1 from Turkey (MS-II-1) (Fig 1A–1D, Table 1). Lists of variants left after filtering are provided in S1–S4 Tables. Genotyping of all available family members by Sanger sequencing showed segregation of the allele with the disease phenotype in all families, revealing an additional homozygote for this mutant allele: affected individual AL-III-9.
rs185005213 is carried in the heterozygous state by 67 of the ~140,000 healthy individuals whose exome and genome analyses were deposited at gnomAD website (no homozygous LOF individuals were present in this cohort [18]) and has a frequency below 0.1 percent in dbSNP, ExAc, 1000 Genomes Project and the National Heart, Lung and Blood Institute–Exome Sequencing Project.
AL-IV-3 had dextrocardia associated with congenitally corrected transposition of the great arteries, mitral atresia and pulmonic atresia (Table 2). AL-III-9, BG-II-1 and MS-II-1 (Fig 1E) had situs inversus totalis without a heart defect and no history of respiratory symptoms (Table 2). Additionally, AL-III-9 and BG-II-1 suffered from infertility. In AL-III-9, the sperm count was reported at 22 X 106 per ml (normal > 15 X 106 per ml). However, under light microscopic examination, less than 1% of sperm cells were found to have normal morphology with the majority having abnormally short tails and some with abnormal head morphology. Furthermore, only 7% of spermatozoa had progressive motility (normal >32%), while 88% did not present any flagellar motility (S5 Table). In BG-II-1, the sperm count was reported at 80 X 106 per ml; only 10% of sperm cells were found to have normal morphology by light microscopy while only 20% of spermatozoa had progressive motility (S5 Table). AL-IV-3 and MS-II-1 are children and precluded sperm analysis.
Thus, homozygous MNS1 mutations in humans result in laterality defects and likely cause infertility, consistent with the phenotype of Mns1-/- mouse [17].
In humans and mice, MNS1/Mns1 mutations are associated with laterality defects; therefore, we hypothesized that MNS1 may be essential for the embryonic nodal function as left/right organizer during early embryonic development, either playing a role for nodal monocilia function or for cilia independent processes in the embryo. To verify nodal expression, we performed in situ hybridization analyses of 8,25 dpc (days post coitum) mouse embryos (Fig 1F). At this stage, the left/right organizer or ventral node is present at the posterior end of the midline and the left/right body axis develops. By utilizing a probe complementary to Mns1 mRNA, we detected expression of Mns1 within the gastrulating mouse embryo. Compared to other embryonic tissues at this developmental stage, Mns1 expression is strongly enriched in the ventral node (Fig 1F).
Because we detected Mns1 expression in the ventral node of mouse embryos, indicating a potential role of MNS1 for node monocilia function, we went further to check if MNS1 is expressed in human ciliary tissues as the axonemes of respiratory epithelial cells and sperm flagella. Using a rabbit polyclonal antibody specific to MNS1, we found by Western blot analysis that MNS1 is expressed in human nasal respiratory epithelial cells and in whole sperm lysates (Fig 2A and 2C). This antibody recognized a single protein with the expected size (60 kDa) in both lysates, while one additional band of approximately 45 kDa was detected in sperm lysates (Fig 2C), likely indicating an isoform of MNS1. This isoform likely corresponds to the sequence listed in AceView, cloned from testis with 363 amino acids and with a molecular weight of 45 kDa (AK057542).
Moreover, we analyzed the localization of MNS1 in human motile respiratory cilia by immunofluorescence microscopy (IF) and determined that in contrast to the control (Fig 2B), MNS1 was undetectable in the respiratory cilia of individual AL-III-9, thus confirming the LOF mutations in MNS1 and supporting antibody specificity. We subsequently determined by IF that MNS1 is detectable in spermatozoa flagella from human control sperm, prominently localizing to the midpiece and the principal piece (Fig 2D).
Moreover, we analyzed whole genome linkage and haplotype analysis data in our PCD cohort recruited in Muenster (Germany) on the basis of respiratory symptoms, laterality defects and/or infertility. Linkage analysis was performed in 161 different consanguineous families. In only 4 families, homozygosity maps revealed homozygous regions encompassing MNS1. After Sanger sequencing of all exons in 4 affected individuals (one from each family), we identified in individual OI-11 II6 a homozygous nonsense mutation (c.607C>T) in MNS1 predicting a premature termination of translation (p.Gln203*). OI-11 II6 is descendant of first-degree consanguineous parents from the multiplex Israeli family (Fig 3A and 3B, S1 Fig). Both unaffected parents were heterozygous for this mutation (Fig 3B). The detected mutation in MNS1 is not listed in gnomAD, absent in 180 controls of European ancestry and not found in the dbSNP, ExAc, 1000 Genomes Project or the National Heart, Lung and Blood Institute–Exome Sequencing Project human polymorphism databases.
X-ray analysis showed that OI-11II6 has situs inversus totalis, consistent with findings reported in Mns1-/- mice, in which the genetic defect causes randomization of left/right body asymmetry [17]. This individual is a 15 year old female and could not be assessed for fertility. OI-11 II1, the other affected sibling, was heterozygous for this mutation.
Whereas the affected individuals AL-IV-3, AL-III-9, BG-II-1 and MS-II-1 (described above) did not present a respiratory phenotype, OI-11 II6 exhibited classical PCD symptoms (Table 2), indicating that she might be carrying additional mutations in a PCD-related gene.
In PCD, the most frequent abnormalities affect ODA composition and function. Mutations in the axonemal heavy chain dynein gene DNAH5 (MIM 603335) cause PCD with ODA defects and account for more than 50% of the total ODA deficiency cases [19–20]. In effect, by screening the five hot-spot exons of DNAH5 [21], we also identified a homozygous deletion of four nucleotides in DNAH5 (c.13432_13435delCACT) leading to a frameshift and premature termination of translation (p.His4478Alafs3*) in OI-11 II6 (Fig 3A and 3B). Both parents were heterozygous carriers for this DNAH5 mutation (Fig 3B). Thus, OI-11 II6 carries bi-allelic homozygous LOF mutations in both MNS1 and DNAH5. Moreover, we identified the identical homozygous frameshift DNAH5 mutations in OI-11 II1 (sister of OI-11 II6), OI-14 II1, OI-24 II1 and OI-24 II2 (PCD-affected individuals from two consanguineous families closely related to family OI-11, Fig 3A).
We then analyzed MNS1 localization in respiratory cells of OI-11 II6 (mutations in DNAH5 and MNS1) and OI-24 II1, which carries the same DNAH5 mutations as OI-11 II6 but is wildtype for MNS1 (Fig 3C). As expected, MNS1 was absent from OI-11 II6 axonemes. Interestingly, MNS1 localization in DNAH5 mutant cells (OI-24 II1) was normal, indicating that the ODA protein DNAH5 is not essential for axonemal MNS1 localization.
Because ODA defects were reported in Mns1-deficient mice, we subsequently sequenced all exons of MNS1 in 134 PCD-affected individuals from different origins (60 from Germany, 45 from Israel, 24 from Denmark, 2 from Hungary, 2 from Greece and 1 from Turkey). These individuals were characterized to have an ODA defect by TEM and/or by IF and with no mutations in reported PCD-related genes. Interestingly, no MNS1 mutations could be found in any of these individuals.
Furthermore, we analyzed respiratory cilia of individual AL-III-9 (MNS1 mutations) and OI-11 II6 (combined MNS1 and DNAH5 mutations) by TEM. The ciliary ultrastructure of individual AL-III-9 displayed a slight reduction of ODAs attached to outer doublets (up to 3–4 ODAs missing out of 9), indicating a partial defect of ODA assembly (control mean ODA: 8.7; affected individual mean ODA: 6). As expected, TEM of individuals OI-24 II1 (DNAH5 mutations) and OI-11 II6 (combined MNS1 and DNAH5 mutations) showed absent ODAs (Fig 4A), a typical finding caused by DNAH5 mutations [19].
Further IF analyses with antibodies targeting DNAH5 [19] and the axonemal ODA intermediate chains DNAI1 [21] and DNAI2 [22] confirmed normal axonemal localization of the three ODA proteins in respiratory cilia of individual AL-III-9 (MNS1 mutations), whereas in respiratory cells of individual OI-11 II6 (combined MNS1 and DNAH5 mutations), DNAH5 as well as DNAI1 and DNAI2 were completely absent from ciliary axonemes (S2 and S3 Figs), as expected due to the DNAH5 mutations and consistent with the ODA defects as demonstrated by TEM. Normal axonemal localization of the inner dynein arm (IDA) light chain DNALI1 and the nexin dynein regulatory complex (N-DRC) component GAS8 (S3 Fig) confirmed absence of other ciliary ultrastructure defects due to MNS1 mutations. Accordingly, PCD-affected individual OI-24 II1 (biallelic DNAH5 mutations without MNS1 impairment) showed the same pattern of staining as OI-11 II6 (biallelic DNAH5 and MNS1 LOF mutations) (S2 and S3 Figs). These findings are consistent with reported PCD-affected individuals with DNAH5 mutations [19–20].
Interestingly, by carefully analyzing ultrastructural cross-sections of OI-24 II1 (biallelic DNAH5 mutations without MNS1 impairment) by TEM, we noticed residual projections on the doublet microtubules (Fig 4A, white arrows), known as the projections of the ODA docking complex (ODA-DC) system, which facilitate the attachment of ODAs onto microtubules [23] but are not affected by DNAH5 mutations. The same residual ODA-DC projections were present in DNAH5-mutant cilia from other PCD individuals with bi-allelic DNAH5 mutations [24]. In contrast, TEM analysis of OI-11 II6 (DNAH5 and MNS1 mutations) showed these projections were absent, demonstrating that a combined DNAH5 and MNS1 deficiency likely disrupts correct ODA-DC assembly.
To address whether MNS1 supports ODA-DC function, we examined the localization of the ODA-DC component CCDC114 [25–26] in respiratory cilia of individuals AL-III-9 (MNS1 mutations), OI-24 II1 (DNAH5 mutations) and OI-11 II6 (combined MNS1 and DNAH5 mutations). In control as well as in MNS1-deficient cilia of individual AL-III-9 and DNAH5-deficient cilia of OI-24 II1, CCDC114 normally localizes along the entire axonemes (Fig 4B). In contrast, CCDC114 staining is reduced in combined MNS1- and DNAH5-mutant cilia, where it localized only to the proximal region of axonemes (Fig 4B). In addition, we examined the localization of the ODA-DC associated protein ARMC4 [27] in respiratory cilia of all three individuals mentioned above. ARMC4 localized normally in MNS1-deficient cilia and in DNAH5-deficient cilia but like CCDC114, it was absent in the distal axonemes of cilia with both MNS1 and DNAH5 mutations (S4 Fig). Thus, the combined deficiency of MNS1 and DNAH5 results in abnormal ODA-DC assembly (proximal CCDC114 and ARMC4), consistent with the absence of ODA-DC projections in individual OI-11 II6 detected by TEM.
Considering that the ODA-DC protein CCDC114 assembly was absent from the distal ciliary axonemes in respiratory cells with a combined MNS1 and DNAH5 deficiency, we tested for possible interactions between MNS1 and ODA and ODA-DC proteins including CCDC114. Using myc- and FLAG-tagged proteins that were co-expressed in HEK293 cells, we found by co-immunoprecipitation that MNS1 interacted with CCDC114 (Fig 5A–5D) but not with ODA proteins DNAI1, DNAI2, and DNAL1, that when mutated also result in ODA defects. We confirmed the reciprocal interaction between CCDC114 and MNS1 by yeast two-hybrid analysis (Y2H) (Fig 5E). We could also show by Y2H that human MNS1 self-dimerized, in agreement with the report of mouse MNS1 dimerization [17] (Fig 5E). Overall, these data indicate that MNS1 plays also a functional role in ODA docking.
Consanguineous families are a resource for the identification of disease-associated genes. Within the framework of our cardio-genetic project, we have identified several CHD genes [28–29] and reached a diagnosis in ~70% of the study families. Here we report the identification of LOF mutations in MNS1 in individuals with laterality defects ranging from situs inversus totalis to dextrocardia with congenitally corrected transposition of the great arteries and mitral atresia (see Table 2). Zhou et al. reported laterality defects in Mns1-deficient mice: accordingly, of 36 Mns1-/- mice tested for laterality defects, 8 (22%) presented with situs inversus totalis, 6 (17%) exhibited left isomerism and the remaining had situs solitus [17]. Because approximately half of the mutant mouse offspring (39%) exhibited laterality defects at birth, we assume that the genetic defect causes randomization of left/right body asymmetry, as observed in other mouse models with abnormal nodal ciliary motility function [27;30]. Our observation of Mns1 expression at the mouse left/right organizer and that mutations in MNS1/Mns1 in human and mouse are associated with laterality defects indicates that MNS1 might be a functional component of motile node monocilia relevant for laterality development.
MNS1 was reported by western blot to be highly expressed in mouse testis, consistent with the encoded protein being a component of the sperm flagella [31]. Using WB and IF analyses, we detected two isoforms of MNS1 and determined the localization of MNS1 throughout the human sperm flagellum, predominantly in the mid and principal piece where the accessory structures are located. Male Mns1-deficient mice were sterile, and their epididymis sperm count was reduced to only 8% of the wild type level [17]. Most of the sperm cells had very short tails with no motility while the head morphology appeared normal [17]. Two affected males in this study have infertility problems with severely reduced flagellar motility and abnormal morphology of the sperm tails. The third and fourth males are still too young for analysis. Interestingly, five males related to the reported individuals with MNS1 mutations (AL-III-9 and BG-II-1) suffered also from infertility, but these individuals were not included in our genetic analysis.
Respiratory cilia in humans with MNS1 mutations displayed only partial ODA defects, similar to previously reported findings in Mns1-deficient mice [17]. Most ODAs were still present in patient respiratory cilia, indicating that their ciliary motility was not significantly altered, consistent with only mild or no respiratory complaints reported in MNS1-deficient individuals.
We have previously shown that while ODA-DC deficiency affects the attachment of ODAs to the axonemes, deficiency of ODAs does not affect ODA-DC assembly [24; 27; 30]. Here we observed for the first time that ODA (DNAH5) deficiency causes a defect in ODA-DC assembly if accompanied with MNS1 deficiency. Analyses in a PCD-affected individual with combined MNS1 and DNAH5 deficiency showed that the ODA-DC protein CCDC114 and the ODA-DC associated protein ARMC4 were absent from the distal ciliary axonemes, suggesting that MNS1 functions in ODA docking. We have previously reported at least two discernible ODA types in ciliary axonemes of human respiratory cilia -type 1 proximal compartment that contain DNAH5 and DNAH11, and type 2 distal compartments that contain DNAH5 and DNAH9 [32–33]. Our observation that a combined MNS1 and DNAH5 deficiency affects CCDC114 and ARMC4 localization only in the distal axonemes raises the possibility that distinct docking complexes exist for the two types of ODAs in humans.
It was recently shown that the ODA-DC is more likely to be a flexible stabilizer of ODAs where it strengthens the electrostatic interactions between ODAs and microtubules, rather than a molecular ruler on which ODAs attach [34]. In contrast to the prior assumption that ODA-DC is an intermediate between ODAs and microtubules, it seems possible that ODAs at this point also interact directly with microtubules. Thus, ODAs, ODA-DC and microtubules interact altogether to stabilize both the attachment of ODAs and the attachment of ODA-DC to the microtubules. Based on the data presented in this paper, we can hypothesize that in the distal axonemes, a complex involving MNS1, ODA-DC and microtubules might be also playing a role in stabilizing the attachment of ODA-DC to the microtubules. When ODAs (DNAH5) are absent, MNS1 could still maintain the attachment of ODA-DC to the microtubules and when MNS1 is absent, ODAs (DNAH5) could also maintain the attachment of ODA-DC to the microtubules at the distal part of the cilia. However, when both ODAs (DNAH5) and MNS1 are absent, ODA-DC fails to assemble in the distal axonemes, indicating that DNAH5 and MNS1 compensate for each other in the stabilization of ODA-DCs to the microtubules at the distal part of the cilia.
The physical interaction between MNS1 and CCDC114 demonstrated here by IP and Y2H supports this hypothesis. Additional intermediate partners interacting with MNS1, CCDC114 and DNAH5 that contribute to correct axonemal localization of ODAs and the regulation of ODA-docking in mammals remain to be determined.
In summary, we identify MNS1 mutations as the cause of laterality defects as well as reduced male fertility and identify an axonemal and flagellar localization of MNS1 in human respiratory cilia as well as sperm cells. We confirm partial outer dynein arm (ODA) abnormalities in respiratory cilia from an MNS1-deficient individual, as observed in Mns1-/- mice. Analyses indicate that MNS1 likely functions also in the process of outer dynein arm docking (ODA-DC). Further functional analyses regarding MNS1 function will help to better understand the molecular mechanisms involved in laterality defects as well as male infertility.
All research complies with the ethical standards in Israel. The ethical committees of Hadassah Medical Center and the Israeli Ministry of Health approved this study; approval number is 0306-10-HMO. All research complies with the ethical standards in Germany and with the rules as set by the European Union; the local Ethics committee Ethikkommision der Ärztekammer Westfalen-Lippe und der Medizinischen Fakultät der Westfälischen Wilhelms-Universität approved this study, approval number is 2010-298-b-S. WES in Nijmegen was perfomred diagnostically in the Human genetics department under the Innovative Diagnostics programme.
Prior to participation, written informed consent was obtained from all patients and family members.
All experiments utilizing animals were performed under the approval of local government authorities (Landesamt für Natur, Umwelt und Verbraucherschutz Nordrhein-Westfalen, Germany (Az.: 84–02.05.20.12.164, and Az.: 84–02.05. 5.15. 012).
This study involved 85 individuals with laterality defects without evidence of PCD. Nine affected individuals -whose clinical data are presented in Table 2- originating from three unrelated consanguineous families, are presented here (Fig 1A–1C):
On the basis of the recurrence of similar clinical features, and reportedly healthy parents, we hypothesized an autosomal-recessive mode of inheritance in the families mentioned. To investigate the molecular underpinnings of disease, we collected DNA samples after obtaining informed patient or parental consent. The study was performed with the approval of the ethical committees of Hadassah Medical Center and the Israeli Ministry of Health as well Turkish ethics committee and Institutional Ethics Review Board at the University of Muenster.
In parallel, we searched our large PCD cohort in Muenster (Germany) for consanguineous families with linkage analysis showing big homozygous regions on chromosome 15 encompassing MNS1. We found 4 consanguineous families (3 from Israel and 1 from Germany) showing linkage to MNS1 and performed Sanger sequencing in all affected individuals.
Among these families, family OI-11 is a part of a large Israeli consanguineous family with 5 affected individuals (Fig 4 and Table 2):
Individual OI-11 II6 is a 15 year old girl exhibiting situs inversus totalis and classical PCD symptoms including bronchiectasis, recurrent pneumonia, recurrent sinusitis, recurrent otitis media and chronic cough. Nasal NO levels were very low (6 nl/min), consistent with PCD [35]. Individual OI-11 II1 shared the same symptoms with her sister OI-11 II6.
OI-14 II1, OI-24 II1 and OI-24 II2 from two Israeli consanguineous families related to family OI-11, had the same respiratory symptoms as OI-11 II6.
Independently, based on phenotypic data including laterality defects, infertility and ODA defects reported in Mns1-deficient mice, we considered MNS1 a reasonable candidate for PCD in humans. Subsequently, we screened 134 PCD-affected individuals characterized to have an ODA defect documented by transmission electron microscopy (TEM) and/or by immunofluorescence analysis (IF) for mutations in MNS1.
Exomic sequences from DNA samples of AL-IV-3 and BG-II-1 were enriched with the SureSelect Human All Exon 50 Mb V.5 Kit (Agilent Technologies, Santa Clara, California, USA) and SureSelect Human All Exon V.6 Kit was used for sample MS-II-1. AL-IV-3 and BG-II-1 sequences (100-bp paired-end) were generated on a HiSeq2500 and MS-II-1 sequences were generated on a Hiseq PE150 (Illumina, San Diego, California, USA). Read alignment and variant calling were performed with DNAnexus (Palo Alto, California, USA) using default parameters with the human genome assembly hg19 (GRCh37) as reference.
The analysis in affected individuals AL-IV-3 and BG-II-1 from consanguineous families A (Palestine) and B (Jordan) yielded 43.0 and 47.3 million mapped reads with a mean coverage of 84x and 82x, respectively. Following alignment and variant calling, we performed a series of filtering steps. In AL-IV-3 and BG-II-1, we excluded variants which were called less than 8x, off-target, heterozygous, synonymous, and had MAF>0.5% at ExAC (Exome Aggregation Consortium, Cambridge, MA) or MAF>2% at the Hadassah in-house database (~1000 ethnic matched exome analyses). In MS-II-1, the following filters are used: MAF 0.5% in ExAc, 1000 genome project and esp6500 databases, coding variant or within 20 bp of exon-intron boundaries, genes carrying bi-allelic variants with prioritization of stop homozygous variants.
We performed total-genome scans by using single-nucleotide polymorphism (SNP) arrays (Affymetrix GeneChip Human Mapping 10K Array v.2.0 [Affymetrix, Santa Clara, CA, USA]) in 161 consanguineous families. Relationship errors were evaluated with the help of the program Graphical Relationship Representation. The program PedCheck was applied for detecting Mendelian errors. Non-Mendelian errors were identified with the program MERLIN. Linkage analysis was performed under the assumption of autosomal-recessive inheritance, full penetrance, and a disease gene frequency of 0.0001. Multipoint LOD scores were calculated with the program ALLEGRO29 and presented graphically with Homozygosity-Mapper (http://www.homozygositymapper.org/).
Genomic DNA was isolated by standard methods directly from blood samples or from lymphocyte cultures after Epstein-Barr virus transformation. Amplification of genomic DNA was performed in a volume of 50 μl containing 30 ng DNA, 50 pmol of each primer, 2 mM dNTPs, and 1.0 U GoTaq DNA polymerase (Promega Corporation, #M3001) or 1.0 U MolTaq polymerase (Molzym Corporation, #P-016). PCR amplifications were carried out by an initial denaturation step at 94°C for 3 min, and 33 cycles as follows: 94°C for 30 sec, 58–60°C for 30 sec, and 72°C for 70 sec, with a final extension at 72°C for 10 min. PCR products were verified by agarose gel electrophoresis, purified and sequenced bi-directionally. Sequence data were evaluated using the CodonCode software.
Mutations in OI families were identified by Sanger sequencing, and sequencing primers for all exons analyzed are available on request.
An ejaculate from individual AL-III-9 was analyzed by light microscopy:
The spermiogram is summarized in S5 Table.
Immunofluorescence analysis was performed as described [26]. Polyclonal rabbit anti-DNALI1, anti-DNAH5, anti-DNAI1 and monoclonal anti-DNAI2 were reported previously [35] as well as monoclonal anti-GAS8 [36]. Monoclonal mouse anti acetylated-α-tubulin (T7451) was obtained from Sigma (Germany). Polyclonal rabbit anti-MNS1 (HPA039975) and anti-CCDC114 (HPA042524) were obtained from Atlas Antibodies (Sweden). Anti-mouse Alexa Fluor 488 and anti-rabbit Alexa Fluor 546 were used as secondary antibodies (Molecular Probes, Invitrogen). DNA was stained with Hoechst 33342 (Sigma). Images were taken with a Zeiss Apotome Axiovert 200 and processed with AxioVision 4.8 and Adobe Creative Suite 4 software.
TEM was performed as described [37].
A Gateway NTAP-MNS1 construct used for tandem affinity purification was described [38]. MNS1 was subcloned into Gateway destination vectors for co-immunoprecipitation and yeast two hybrid analysis via LR Clonase reaction as described [39]. All MNS1 constructs were confirmed by sequencing and matched gene accession number: NM_018365.2. CCDC114 Gateway constructs used for interaction studies with MNS1 were described [29].
Co-IP and WB were performed as previously described [39]. Briefly, HEK293 cells were transfected with plasmids encoding myc- and FLAG-tagged cDNA constructs using Gene Juice (Novagen) at approximately 0.1 μg DNA per ml of media. Within 24 hrs, cells were collected in 1× PBS and lysed in 1 ml of the following buffer: 50 mM Tris-Cl, pH 8.0, 150 mM NaCl, 1% IGEPAL, 0.5 mM EDTA, and 10% glycerol supplemented with protease (Roche Complete) and phosphatase inhibitors (Cocktails 2 and 3, Sigma Aldrich). Lysates were centrifuged at 16,000 × g for 30 min. at 4°C. Approximately 2 mg of each lysate was precleared with 4 μg of rabbit control IgG antibody for 2 hrs. at 4°C, and then incubated with MagSi/protein A beads (MagnaMedics, Germany) for 1 hr. Lysates were then incubated with 4 μg of rabbit anti-FLAG or anti-myc antibody overnight at 4°C, and then incubated with MagSi/protein A beads for 1 hr. to capture immunoprecipitates. Bead complexes were washed four times in lysis buffer and then resuspended in 1× LDS buffer supplemented with DTT (1/8 lysis volume) and heated for 10 min. at 90°C. Lysates were electrophoresed in NuPAGE 4–12% Bis-Tris gels, transferred to PVDF filters, and subsequently immunoblotted with either anti-myc (A7) or anti-FLAG (M2) mouse monoclonal antibodies. PVDF filters were washed three times in TBS-T (10 minutes each) before blocking in 5% BSA for 2 hours at room temperature. Filters were then washed three times (10 minutes each) before incubation with primary antibody (diluted in TBS-T) overnight at 4°C. Filters were washed three times (10 minutes each) and then incubated with secondary antibody for 1 hour at room temperature. Filters were then washed four times and developed by ECL using Prime Western Blotting Detection Reagent (Amersham). Images were digitally acquired using a FUSION-SL Advance Imager (PeqLab) and modified for contrast using Adobe Photoshop v. CS4. All wash and incubation steps were performed with gentle shaking.
Western Blots in respiratory and sperm lysates were performed similarly but PVDF filters were blocked overnight with 5% milk at 4°C and incubated with primary antibody (diluted in 5% milk) for 3 hours.
The following antibodies were used: Rabbit polyclonal anti-MNS1 (HPA039975), rabbit polyclonal anti-CCDC114 (HPA042524) and rabbit polyclonal anti-myc (1:25; clone A-14, Santa Cruz).
Binary interaction between CCDC114 and MNS1 was tested as described [39].
Using primers Mns1-F: 5`-aagaagcgtgaggagatgga-3´ and Mns1-R: 5`cggccttgctatgaagactc-3´, a 728 bp fragment of Mns1 (NM_008613.3) was amplified from mouse testis cDNA. The fragment was ligated into the pCRII-TOPO vector by TOPO cloning reaction (Invitrogen; Thermo Fisher Scientific) and the resulting plasmid subsequently transformed into chemically competent E.coli cells. Positive clones were selected for plasmid preparation (plasmid midi kit; Qiagen) and the isolated plasmid was checked for the correct insert by Sanger sequencing.
NotI or KpnI (Fermentas; Thermo Fisher Scientific) digested plasmids served as templates for in vitro transcription of sense and antisense probes, respectively by T7 or SP6 polymerase incorporating Digoxigenin (DIG) -11-UTPs into the newly synthesized probes.
All experiments utilizing animals were performed under the approval by local government authorities (Landesamt für Natur, Umwelt und Verbraucherschutz Nordrhein-Westfalen, Germany (AZ 84–02.05.20.12.164, and AZ 84–02.05. 5.15. 012).
Wildtype CD-1 mice were mated in the evening and the females checked for vaginal plugs in the morning of the following day. Pregnant mice were sacrificed at day 7 after plug detection, corresponding to embryonic day 8.25 dpc (days post coitum) and embryos were dissected in ice cold 1xPBS. After fixation in 4%PFA in 1x PBS overnight, embryos were transferred to methanol and stored in methanol at -80°C until use.
Mouse embryos were transferred to PBT (1xPBS + 0,1% Tween) and bleached in H2O2 in PBT for 1 hour. Digestion with proteinase K (10μg/ml) was performed and stopped after 8 minutes by the use of glycine (2mg/ml). After fixation with fixing solution (4%PFA and 0,2% Glutaraldehyde in PBT), embryos were rinsed and washed with PBT and preincubated with hybridization solution (50% formamide, 0,5% CHAPS, 1,3x SSC, 5 mM EDTA, 50μg/ml Yeast t-RNA; 700U/ml Heparin, 0,2%Tween) at 65°C for few hours. Hybridization with sense or antisense probes was performed in hybridization solution at 65°C overnight. The following day, after washing twice with hybridization solution and twice with MABT (Maleic acid (500mM), NaCl (750 mM), NaOH (1M), 0,2%Tween) embryos were blocked in blocking solution (2% Boehringer blocking reagent (Roche; Merck) in MABT) for 1 hour, and blocking solution with 20% lamp serum for additional 2–3 hours. Anti-DIG antibody was diluted 1:2000 in blocking solution with 20% lamp serum. Binding of the antibody was performed in blocking solution with 20% lamp serum at 4°C overnight. The following day, embryos were washed three times with MABT and twice with NTMT (NaCl (100mM), Tris (100mM), MgCl2 (50mM), Tween (0,2%)). Color development was performed utilizing NBT/ BCIP (Roche; Merck) in NTMT. After fixation in fixing solution and transfer to 80% glycerol in PBT, embryos were imaged utilizing a Nikon Digital Sight DS-L3 camera mounted on a Nikon SMZ1000 stereomicroscope and images were processed using creative suite (Adobe).
Aceview, https://www.ncbi.nlm.nih.gov/IEB/Research/Acembly/
Homozygosity-Mapper, http://www.homozygositymapper.org/
Exome Aggregation Consortium (ExAC), http://exac.broadinstitute.org/
Genome Aggregation Database (gnomAD), http://gnomad.broadinstitute.org/
Database of Genomic Variants, http://projects.tcag.ca/variation/
dbSNP, http://www.ncbi.nlm.nih.gov/SNP/
1000 Genomes Project human polymorphism database, http://www.1000genomes.org/
National Heart, Lung and Blood Institute–Exome Sequencing Project, http://evs.gs.washington.edu/EVS/
Online Mendelian Inheritance in Man, http://www.omim.org/
|
10.1371/journal.pgen.1001166 | Evidence for a Xer/dif System for Chromosome Resolution in Archaea | Homologous recombination events between circular chromosomes, occurring during or after replication, can generate dimers that need to be converted to monomers prior to their segregation at cell division. In Escherichia coli, chromosome dimers are converted to monomers by two paralogous site-specific tyrosine recombinases of the Xer family (XerC/D). The Xer recombinases act at a specific dif site located in the replication termination region, assisted by the cell division protein FtsK. This chromosome resolution system has been predicted in most Bacteria and further characterized for some species. Archaea have circular chromosomes and an active homologous recombination system and should therefore resolve chromosome dimers. Most archaea harbour a single homologue of bacterial XerC/D proteins (XerA), but not of FtsK. Therefore, the role of XerA in chromosome resolution was unclear. Here, we have identified dif-like sites in archaeal genomes by using a combination of modeling and comparative genomics approaches. These sites are systematically located in replication termination regions. We validated our in silico prediction by showing that the XerA protein of Pyrococcus abyssi specifically recombines plasmids containing the predicted dif site in vitro. In contrast to the bacterial system, XerA can recombine dif sites in the absence of protein partners. Whereas Archaea and Bacteria use a completely different set of proteins for chromosome replication, our data strongly suggest that XerA is most likely used for chromosome resolution in Archaea.
| Bacteria with circular chromosome and active homologous recombination systems have to resolve chromosomal dimers before segregation at cell division. In Escherichia coli, the Xer site-specific recombination system, composed of two recombinases and a specific chromosomal site (dif), is involved in the correct inheritance of the chromosome. The recombination event is tightly regulated by the chromosome translocase FtsK. This chromosome resolution system has been predicted in most bacteria and further characterized for some species. Intriguingly, most archaea possess a gene coding for a recombinase homologous to bacterial Xers, but none have homologues of the bacterial FtsK. We identified the specific target sites for archaeal Xer. This site, present in one copy per chromosome, is located in the replication termination region and shows sequence similarities with bacterial dif sites. In vitro, the archaeal Xer recombines this site in the absence of protein partner. It has been shown that DNA–related proteins from Archaea and Eukarya share a common origin, whereas their analogues in Bacteria have evolved independently. In this context, Eukarya and Archaea would represent sister groups. Therefore, the presence of a shared Xer-dif system between Bacteria and Archaea illustrates the complex origin of modern DNA genomes.
| In Bacteria, homologous recombination is essential during DNA replication to resume stalled replication forks and to repair DNA double and single strand breaks. Odd numbers of homologous recombination events between circular chromosomes generate dimers, which need to be resolved to ensure proper segregation in daughter cells. In Escherichia coli two paralogous site-specific tyrosine recombinases XerC and XerD were shown to convert chromosome dimers to monomers [1] by acting at a specific DNA recombination site, dif, located close to the replication termination region [2]–[4]. Homologues of XerCD are widespread in the bacterial domain, and dif sites have been characterized in several Proteobacteria and Firmicutes [5]–[10]. dif sites are semi-conservative inverted repeats formed by two arms (Xer protein binding sites) of 11 base pairs [11], separated by a spacer of 6 bp and are fairly conserved among Bacteria [5]. In Proteobacteria the XerCD activity is tightly regulated by the cell division protein FtsK, a DNA translocase anchored at the division septum [12]–[15]. In E. coli, 8 bp G-rich polar sequence elements (KOPS) direct FtsK translocation on DNA [16]–[19]. KOPS are oriented from the origin of replication towards dif where their polarity is precisely inverted. FtsK DNA translocation is therefore always oriented towards the dif site and dif sites carried on a chromosome dimer are brought together at midcell. FtsK further controls chromosome dimer resolution by activating XerD activity through protein-protein interactions [20], [21]. In several Lactococcus and Streptococcus strains, the canonical bacterial XerCD-dif system has been replaced by a single tyrosine recombinase, XerS (distantly related to XerCD) whose gene is located next to its specific dif-like site and localized at the terminus of replication [22]. Strikingly, this XerS-dif-like system still depends on the KOPS-oriented FtsK activity to form the synaptic complex for recombination [22].
Archaea harbour circular chromosomes and have an active homologous recombination system [23]. Therefore, they are expected to resolve chromosomal dimers to ensure proper chromosome segregation. It was previously reported that most archaeal genomes encode a single protein homologous to bacterial XerCD [24]; however, none encode a FtsK homologue. It is thus unclear whether archaeal Xer-like proteins (hereafter called XerA) are involved in chromosome resolution in Archaea, as in Bacteria.
In order to determine whether XerA is involved in chromosome resolution, we performed an in silico search for XerA specific recombination dif-like sites in four closely related archaeal genomes from Thermococcales. We identified a highly conserved 28 bp sequence that shares 14 out of 28 bases with characterized bacterial dif sites. The predicted dif sites are systematically located in the replication termination regions of the four genomes. The same analysis performed on three Sulfolobales genomes revealed that a similar site is also present in this crenarchaeotal species. We further identified short polarized sequences that point towards the predicted dif sites in Thermococcales genomes. We validated the in silico predictions by showing that a purified recombinant XerA protein from Pyrococcus abyssi specifically recombines plasmids carrying the predicted dif site of this archaeon. The recombination activity did not require the presence of any protein partner, in contrast to bacterial Xer-mediated recombination. Our data strongly suggest that XerA is most likely used for chromosome resolution in Archaea.
The majority (88%) of archaeal genomes sequenced so far (KEGG database [25]) harbour single orthologues of the bacterial XerCD recombinases. Alignments of several bacterial XerCD proteins with XerA proteins from different Archaea revealed that they share a conserved C-terminal domain where the catalytic site (Figure S1) is located. The six catalytic residues (R-K-H-R-[H/W]-Y) characteristic of tyrosine recombinases [26] are perfectly conserved in archaeal XerA proteins [24].
The more variable outer sequences of the bacterial dif sites are the place of specific amino-acids/bases contacts that drive protein-DNA interaction specificity. Several amino acids residues involved in these contacts were identified, which led to the definition of a dif-binding region within the C-terminal domain of the Xer recombinases [27]–[29]. The dif binding motif is also conserved in XerA proteins. Notably, the key residues that define binding specificity for the XerC or XerD binding sites are distinct from XerC or XerD in XerA (Figure 1A and Figure S1).
To search for conserved putative archaeal dif sites, we selected as first candidates four closely related genomes of Thermococcales since their XerA proteins [Pyrococcus abyssi (Pab0255), Pyrococcus horikoshii (PH1826), Pyrococcus furiosus (PF1868) and Thermococcus kodakaraensis (TK0777)] are the most similar to bacterial XerCD (35%–39% identity; Figure S2) among archaeal XerA. The putative dif-binding motif of these XerA proteins shares numerous conserved positions with both XerC (10 out of 19 positions) and XerD (11 out of 19 positions) proteins (Figure 1A). These remarkable sequence similarities suggest that one may expect to identify conserved dif-like sequences in these four closely related archaeal genomes based on known properties of bacterial dif sites. Finally, XerA proteins are well conserved between these four species (above 85% similarity, Figure S2). Thermococcales XerA proteins are thus expected to recognize similar dif sites.
We built an algorithm to search for any potential tyrosine recombinase-binding site. We searched for imperfect inverted repeats of 11 to 15 bp separated by spacers ranging from 4 to 10 bp. A total of 481,319 sequences were recovered after this analysis. In order to reduce the sequences to one single most likely dif candidate, we selected only sequences that were conserved above 80% similarity in the four genomes. We found six sequences fulfilling this criterion: two were shawn to be spacer sequences of CRISPRs [30], [31], three were imperfect inverted repeats of 11 bp separated by long spacers (one of 8 bp and two of 10 bp), and only one sequence in each genome was composed of 11 bp imperfect inverted repeats separated by a 6 bp spacer (Figure 1B). Strikingly, these sequences are 100% conserved between P. horikoshii and P. furiosus, have three mismatches with that of Pyrococcus abyssi and seven with that of T. kodakaraensis. Moreover, these four predicted sites share many positions with the bacterial dif consensus sites (Figure 1B). The three Pyrococcus sites show 14 out of 28 conserved positions of the bacterial dif-consensus, and the T. kodakaraensis site shows 18 out of 28 (Figure 1B). The same site search was performed on the T. gammatolerans, T. onnurineus and T. sibiricus genomes and led to the identification of unique sites showing the same level of conservation with the bacterial dif-consensus (Figure S3). Further analysis of the dif sites environment in Thermococcales genomes revealed that all dif sites are surrounded by conserved flanking sequences (Figure S3).
As in the canonical bacterial model, Thermococcales dif sites predicted by our analysis were not located next to the xerA genes (Figure 2 and Figure S4). The position of the xerA genes relative to the replication origins (oriC) was highly variable (Table S1), whereas the predicted dif sites were located within the second quarter of the genome for P. horikoshi, P. furiosus and T. kodakaraensis (135°, 122° and 130° from oriC, respectively) and in the third quarter (−142° from oriC) for P. abyssi (Figure 2). In the latter case, the difference in position could be a consequence of the large fragment inversion containing oriC that recently occurred in this species [32]. The conservation of dif site positions relative to oriC in the four Thermococcales (between 122° and 142°) is especially striking since these genomes have been extensively rearranged by chromosome recombination, as indicated by the patterns obtained from whole genome alignments (Figure S5).
The dif-like sites identified in our analyses do not localize precisely at 180° from oriC. However, the predicted dif site of P. abyssi is located into a late replicating fragment of the genome [33]. dif sites positions are therefore compatible with a localization in the terminus region of chromosome replication. It is not known if the two replication forks always meet at the same point in Thermococcales. A precise site for the terminus of DNA replication in Thermococcales genomes cannot be predicted by using GC skew analysis because, in contrast to the sharp peak observed at oriC, the potential termination region appears as a broad distribution [33], [34]. The terminus region appears to be especially prone to chromosomal rearrangement in Thermococcales [32], possibly explaining this lack of resolution.
We next extended our analysis to archaeal genomes outside of the Thermococcales group. Using the dif sites and flanking sequences found in Thermococcales, we constructed a Hidden Markov Model with HMMER2 [35] and searched other archaeal genomes for potential dif sequences. As an example, a single statistical-significant sequence matching the Thermococcales dif sites was found in the Methanosphaera stadtmanae genome. This site is located at about 180°C from oriC (Figure S6), and is surrounded by imperfect inverted repeats. We then selected three Sulfolobales genomes to search for dif sites in Crenarchaeota. Sulfolobus species possess multiple replication origins [36], [37] raising the possibility that an alternative to the canonical Xer recombination system may occur in these organisms. We used the same initial methodology that was applied to Thermococcales genomes. Unique dif sites were found for S. acidocaldarius and S. tokodaï, whereas two copies of this site were found at the same chromosomal location in the S. solfataricus genome (Figure S7). As opposed to Euryarchaeota, the predicted dif sites localized close to xerA genes, and were flanked on only one side by a short conserved sequence of 13 bp.
In Bacteria, the directionality of FtsK-mediated DNA translocation is determined by octamers that are polarized on each arm of the chromosome with their orientation switching at the dif site [16]–[18]. Although Archaea lack an FtsK homologue, we searched for the most frequent and skewed sequences in Thermococcales genomes by using the R'MES program ([38], see Materials and Methods). We used the E. coli genome to validate this methodology and, as expected, we found that KOPS are the most over-represented skewed 8nt-long sequences. The corresponding diagram shows a sharp peak corresponding to the position of the E. coli dif site (Figure 2). We then analyzed all possible sequences of 4 to 8 nucleotide-long in the four Thermococcales genomes. We identified GTTG as the most over-represented and skewed sequence in the P. abyssi, P. horikoshi and T. kodakaraensis genomes, and GTTC in the P. furiosus genome. For Thermococcales genomes, the cumulative frequency of these 4 nucleotide sequences (Archaea Short Polarized Sequences, ASPS) does not give a perfect triangle-shaped diagram as observed for Proteobacteria (Figure 2) or Firmicutes where dif sites locate exactly at the KOPS skew inversion. Nevertheless, in the case of the three Pyrococcus genomes, the cumulative ASPS skews diagrams displayed a sharp optima precisely located next to the dif-like sites identified in silico (Figure 2). The optimum was located very close to the dif sites in the cases of P. horikoshi and P. furiosus, whereas it was located more to the left in the case of P. abyssi (around 160° instead of 142°). This shift could be due to a recent transposition of two chromosomal segments that occurred in the terminus region of this species [32]. In the case of T. kodakaraensis, we only obtained the sharp minimum corresponding to the replication origin, whereas the opposite region appeared as a broad peak containing the predicted dif site. The same analysis was extended to other archaeal genomes where dif sites were predicted, and revealed that both euryarchaeal and crenarchaeal genomes harbour ASPS (Figure 2; Figures S4, S6, S7). Remarkably, the M. stadtmanae ASPS skew displays the triangle-shaped diagram observed in Bacteria, with the predicted dif site precisely located at the skew inversion (Figure S6).
To test our in silico predictions, we purified to homogeneity the P. abyssi XerA protein (Pab0255) as a recombinant protein. Recognition of the predicted dif site by the P. abyssi XerA protein was first evaluated by Electrophoretic Mobility Shift Assay (EMSA). Double stranded oligonucleotides corresponding to the P. abyssi dif site were incubated with increasing amounts of P. abyssi XerA protein at 20° and 65°C (Figure 3). As a control, P. abyssi XerA protein was also incubated in presence of a non-specific DNA site corresponding to the minimal recombination site (attP) of another archaeal tyrosine recombinase, the SSV1 integrase [39]. XerC and XerD from E. coli have only been shawn to bind to an oligonucleotide containing the E. coli dif site [40]. In contrast, P. abyssi XerA was able to bind to both substrates, with two protein-DNA complexes detected in each case (Figure 3). However, complex migration was different between the two DNA substrates, with P. abyssi XerA/dif complexes showing a higher mobility than P. abyssi XerA/attP complexes. Furthermore, the P. abyssi XerA protein presented a preference for the P. abyssi dif site as compared to the non-specific substrate, with a 4 fold increase in complex formation at 20°C and an 8 fold increase at 65°C (Figure 3). At the P. abyssi optimal growth temperature (90°C), XerA binding to the dif site should therefore be highly specific. Competition experiments further confirmed that XerA has a much higher affinity for its dif site than for a heterologous tyrosine recombinase binding site (Figure S8).
We next searched for full or partial site-specific recombination activity. In the case of XerCD, in vitro recombination at dif sites requires the C-terminal domain of FtsK [15]. However, dif-dependent DNA relaxation has been observed for XerC and XerD [41]. In order to test for such activity, we cloned an oligonucleotide containing the predicted dif site into a plasmid vector. After incubation of XerA with this substrate, reactions products were analyzed by agarose gel electrophoresis. Whereas incubation of the control plasmid (without dif site) with P. abyssi XerA protein did not reveal any reaction product, addition of the protein to the dif-containing plasmid led to the appearance of several new bands of lower mobility than the open circular (Moc) form of the substrate (Figure 4A). The migration of the major product suggested that it may correspond to the supercoiled form of a dimeric plasmid (Dsc), while the other products may correspond to increasing multimers of the dif-containing plasmid. The reaction was strongly dependent on temperature (Figure 4A), as expected for a reaction catalyzed by a protein from a hyperthermophilic organism. A single product, migrating slightly above the open circular substrate form appeared when the incubation was performed at 20° or 35°C. The amount of product strongly increased when the incubation was performed at 50° and 65°C, reaching an amount roughly equivalent to that of the remaining monomeric supercoiled (Msc) substrate. Several new products of low mobility were detected when the reaction was performed above 50°C, and their relative amounts increased from 50° to 65°C. To determine whether the reaction products generated by the P. abyssi XerA were indeed multimers of the dif-containing plasmid, we took advantage of a unique HindIII restriction site present on the plasmid. We assumed that partial HindIII digestion of the reaction products would produce linear multimeric plasmids that could be identified by their size. The products of a P. abyssi XerA catalyzed reaction performed for 20 minutes at 65°C were incubated with one unit of HindIII for one hour, either at 37°C for full digestion or at 20°C for partial digestion (Figure 4B). At 37°C, digestion of reaction products produced only linear DNA of the monomeric size (LM, 2.6 kb), indicating that all reaction products were multimers of dif-containing plasmids which were cleaved at all available HindIII sites (Figure 4B). HindIII digestion at 20°C produced two additional bands of linear DNA with the expected molecular weight for linear dimers (LD, 5.2 kb) or linear trimers (LT, 7.8 kb) of the dif-containing plasmid. This result indicates that P. abyssi XerA can recombine dif-containing plasmids in the absence of protein partners, producing multimeric forms of the initial substrate. The specificity of the reaction was further controlled by using as substrate a plasmid containing the attP site (Figure S9). No recombination activity could be detected on this substrate, further indicating that binding of XerA to the attP site is non specific.
We followed the time course formation of multimeric plasmids by P. abyssi XerA on the dif-containing plasmid at 65°C. Plasmid dimers were obtained after five minutes incubation and plasmid multimers were detected after 10 minutes (Figure S9). After 30 minutes reaction time, the relative intensity of all bands remained fairly constant suggesting either that the recombination activity had reached enzymatic equilibrium, or that the protein was rapidly inactivated upon incubation at 65°C. However, pre-incubating the protein alone for up to 40 min at 65°C did not reduce the extent of recombination (Figure S9) thus ruling out protein denaturation during the time course assay. This suggests that at the reaction equilibrium, production of multimers from monomers is equivalent to multimer resolution into monomers.
To further validate the resolution activity of the P. abyssi XerA protein, we constructed a substrate with two dif sites in direct repeat (Figure 4C). The reaction products were analyzed by PCR as both integration and resolution events can occur on this substrate. Resolution events reduced the distance between the two primers from 1058 bp to 885 bp (Figure 4C). Even though integration events were still favoured, as attested by the appearance of plasmid multimers (not shown), a PCR product with a size around 900 bp was detected (Figure 4C). This product indicates that XerA was able to assemble a recombination proficient synaptic complex and that although at low level resolution events also occurred. The P. abyssi XerA protein is therefore able to catalyse both resolution and integration depending on the substrate provided in the reaction.
We have shown that the P. abyssi XerA protein (homologous to the bacterial XerCD recombinases) specifically recombines a plasmid containing a predicted dif sequence present in the P. abyssi genome. This dif sequence was identified in silico, taking into account known features of tyrosine recombinase recombination sites and searching for sequences present in the genomes of four closely related Thermococcales. Importantly, whereas the location of xerA genes relative to the replication origin (oriC) varies from one genome to the other, the positions of the dif sites with respect to oriC is relatively conserved in all four genomes. These observations strongly suggest that archaeal XerA proteins may be involved in the resolution of chromosome dimers at the terminus of replication.
Interestingly, although Archaea lack a FtsK homologue, we could identify polarized sequences of four nucleotides (ASPS) that define the replication termination region and point towards the predicted dif sites in Euryarchaeota. The ASPS are shorter than the KOPS used by FtsK in Bacteria. Strikingly, in most genomes the predicted dif site localized at the summit of curves obtained by ASPS cumulative skew analyses. Archaea may therefore use a functional analogue of the bacterial FtsK-KOPS mechanism to produce a dif-synaptic complex in vivo. It was suggested that the archaeal bipolar DNA helicase HerA, which is probably involved in the processing of double-strand breaks for homologous recombination [42]–[44] may also be a functional analogue of FtsK in Archaea on the basis of their common ATPase domain used for DNA translocation [45]. Even though our results show that P. abyssi XerA does not require an accessory protein to catalyse recombination in vitro, as opposed to bacterial XerCD which only recombine dif sites in the presence of FtsK [12], [15], they do not exclude that a protein partner could either regulate synaptic complex assembly or XerA activity in vivo. Indeed, in Thermococcales genomes, dif sites are flanked by inverted repeats (Figure S4) that may be binding sites for such partner.
Using different in silico methodologies, we were able to predict dif sequences in other euryarchaeal genomes and crenarchaeal genomes that harbour xerA genes. Archaea lacking a xerA gene, such as Thaumarchaea and Pyrobaculum species, may have recruited another tyrosine recombinase, as happened in some bacterial groups [5], [22].
Although the archaeal and bacterial Xer/dif systems use similar dif sites, they differ in terms of biochemical properties and reaction mechanisms. Whereas E. coli and B. subtilis XerCD do not bind a DNA fragment without dif site [8], the P. abyssi XerA protein can bind with a lower affinity to a heterologous tyrosine recombinase site. However this site is not recognized as a recombination substrate. More significantly, the P. abyssi XerA protein can recombine plasmids carrying the dif sequence in the absence of protein partner. In contrast, the XerCD activity depends in vitro and in vivo on FtsK to recombine the chromosomal dif site or on other partners (PepA and ArcA or ArgR) to recombine the plasmidic recombination sites psi on pSC101 or cer on ColE1 [46]–[49]. The ability of P. abyssi XerA to recombine in vitro a dif-containing plasmid without any accessory protein suggests that XerA may also work alone in vivo. However we cannot rule out that a protein partner may bind the conserved dif-flanking sequences. Such a partner may either control the directionality of the reaction towards resolution or coordinate the recombination activity with the progression of the cell cycle. Alternatively, XerA activity may be limited to replication termination and chromosome segregation by regulating XerA expression at the transcriptional level. In agreement with this view, the xerA gene of the crenarchaeon Sulfolobus acidocaldarius (Saci 1490) is induced during the G1/S phase and reaches its maximal expression level in the G2 phase of the cell cycle [50].
Our results show that Archaea possess a Xer/dif system similar to its bacterial counterpart that is likely involved in chromosome dimer resolution. However, the archaeal Xer system displays differences, such as the involvement of a unique protein and the ability to perform site-specific recombination in vitro in the absence of accessory proteins. To get a better view of the evolution of the Xer system, we performed a phylogenetic analysis including bacterial and archaeal Xer proteins and a subset of bacteriophage encoded tyrosine recombinases (Figure 5). Unfortunately, the resulting tree is not resolved at most basal nodes, preventing a clear view of the evolutionary relationships of these proteins. However, it shows that bacterial and archaeal homologues are not intermixed, indicating that no recent horizontal gene transfer has occurred between domains. Our data thus suggest that a Xer/dif system was present in the common ancestor of Archaea and Bacteria, suggesting that this ancestor had a circular double-stranded DNA genome. However, this raises further issues such as why the replication machinery of Archaea and Bacteria are now composed of non homologous proteins [51]. Alternatively, homologous viral tyrosine recombinases may have been recruited independently in Archaea and Bacteria to be used as Xer/dif systems after transition from RNA to DNA genomes [52]. In any case, the presence of a shared Xer-dif system in Bacteria and Archaea illustrates the complex origin of modern DNA genomes. Further studies of Xer-dif systems in different archaeal and bacterial groups will be now necessary to test alternative scenarios for the origin and evolution of Xer proteins.
A model searched for all inverted repeats (11 to 15bp) separated by a spacer (4 to 10bp) present in non-coding genomic regions. A consensus sequence was deduced from the alignment of predicted dif sites and represented as a sequence logo [53].
Statistical significant skewed words in the four Thermococcales genomes were determined using the R'MES program (http://mig.jouy.inra.fr/logiciels/rmes/) [38]. Since R'MES reads only one single strand of DNA at time, we artificially defined several ends of replication, starting at 120°, and then moving by 5° steps up to 200° from oriC. From these artificially-defined ends of replication to oriC we reverse complemented the genomic sequence. The most skewed word for each analyzed genome was selected by comparing the R'MES results. Cumulative skews were calculated using the formula:
Genomes of the four Thermococcales where aligned by dot-plot analyses based on BlastP searches (e-value of 1×10−10).
Xer homologues were searched by BLASTP at the NCBI (http://www.ncbi.nlm.nih.gov/) against complete sequenced archaeal genomes using E. coli XerC and XerD proteins as query (threshold of 1×e−10). The retrieved homologues were aligned using Muscle and a specific HMM profile was calculated for exhaustive detection of homologues. Using the HMMER program (http://hmmer.janelia.org/) we performed iterative searches until no new homologues were detected in the archaeal genomes. From the resulting dataset, a preliminary phylogenetic analysis allowed to select 62 representative XerC and XerD sequences from several bacterial species and several tyrosine recombinases from plasmids/viruses or from mobile elements integrated into cellular genomes. The final alignment was trimmed to remove ambiguously aligned positions, leading to 222 conserved residues for phylogenetic analysis. A maximum likelihood tree was obtained by using PHYML [54], with the WAG evolutionary model including correction for heterogeneity of evolutionary rates (4 categories+invariant) and statistical support at nodes was calculated by non parametric bootstrap on 100 resampling of the original dataset by PHYML.
The Pab0255 gene was amplified by PCR using P. abyssi genomic DNA as a template, Phusion high-fidelity DNA polymerase (Finnzyme) and the following primers:
5′-GGGAACATATGCACCATCACCATCACCATGAGGAGAGGGAGGAGAGAGTGAGGGATGATACAATTG-3′
5′-TTTTTGCGGCCGCTTAGGAACCCCCGATG-3′
The forward primer allows the addition of six histidine codons in frame with the ATG start codon (underlined). The PCR product was digested by NdeI and NotI and cloned into a derivative of a pET vector (Novagen). The resulting recombinant plasmid was sequenced prior to being transformed into the E. coli expression strain Rosetta(DE3)pLys (Novagen).
Cells were grown in 2xYT medium (BIO101Inc.) at 37°C to A600nm = 1 and expression of Pab0255 was induced by the addition of 0.5 mM IPTG (final concentration). Four hours after induction, cells were harvested by centrifugation, and the pellets resuspended in 40 ml of 50 mM Tris-HCl, pH 8.0, 1 M NaCl and 5 mM β-mercaptoethanol and stored at −20°C. Cells were lysed by sonication and centrifuged at 13, 000× g for 30 min at 25°C. The supernatant was collected and heated for 15 minutes at 70°C. After centrifugation at 13,000× g, the supernatant was loaded onto a Ni2+ affinity column (Ni-NTA agarose, Qiagen) pre-equilibrated in the same buffer. The His-tagged Pab0255 was eluted at 20 mM imidazole, and loaded onto a 2 ml HiTrap Heparin (Amersham Biosciences) column pre-equilibrated in a buffer containing 50 mM Tris-HCl pH 7.0, 200 mM NaCl, 1 mM DTT. A NaCl linear gradient (200 mM to 2 M) was developed, and the protein eluted at about 800 mM NaCl. Finally, the protein was loaded onto a cation-exchange SP Sepharose column (Amersham Biosciences) pre-equilibrated in the same buffer, and eluted by a NaCl linear gradient. The purified protein was dialysed against 50 mM Tris pH7.0, 300 mM NaCl, 50% glycerol prior to being stored at −20°C.
The following 43 nt long oligonucleotides containing the top and bottom strands of the predicted dif site from P. abyssi (Pab) or minimal attP site from SSV1 [39] were purchased from Eurogentec:
Pab-Dif-Top 5′-gttaactatATTGGATATAATCGGCCTTATATCTAAAgtgttg-3′
Pab-Dif-Bottom 5′-caacacTTTAGATATAAGGCCGATTATATCCAATatagttaac-3′
attP-Top 5′-gttaactaTCTTTTCCGCCTCCGGAGCCGGAGGTCCCgtgttg-3′
attP-Bottom 5′-caacacGGGACCTCCGGCTCCGGAGGCGGAAAAGAtagttaac-3′
For binding assays, top strand oligonucleotides were 5′-end labeled by using [g-32P]ATP and T4 polynucleotide kinase. Unincorporated nucleotides were removed by spin dialysis, and the labeled oligonucleotide was then hybridized with a 2-fold excess of unlabeled complementary strand in TE buffer (10 mM Tris, pH 8.0, 1 mM EDTA).
The DNA binding reactions were carried out in 20 µl of a mixture composed of 0.5 µM 5′-end labeled dif substrate or attP substrate, increasing amounts of Pab0255 in a binding buffer composed of 50 mM Tris pH 7.5, 30 mM NaCl and 0.5 µg poly(dIdC).poly(dIdC). Incubation was performed for 30 min at either 20°C or 65°C, and then 5 µl of 5× loading buffer (10 mM Tris pH 7.5, 1 mM EDTA, 20% glycerol, 0.1 mg/ml BSA, 0.1% xylene cyanol) was added to the binding reactions. The samples were loaded onto 8% polyacrylamide gels (30∶0.5 acrylamide∶bisacrylamide), and electrophoresis performed in 1× TGE buffer (50 mM Tris, 8 mM Glycine, 0.1 mM EDTA) at 4°C for 4 h at 7 V/cm. The DNA-protein complexes were visualized by autoradiography and phosphorimaging.
The 43 bp double stranded oligonucleotide containing the predicted P. abyssi dif and the 43 bp double stranded oligonucleotide containing the attP site were cloned into the pBend2 (2.6 Kbp) vector [55]. pBend2, pBend2-dif and pBend2-attP plasmids were purified on CsCl gradients. Recombination reactions were performed in 20 µl of reaction mixture consisting of 30 mM Tris pH 7.5, 50 µg/ml bovine serum albumin, 50 mM NaCl, 500 ng of plasmid and 40 pmol of XerA protein. The reaction mixture was incubated at 65°C (unless otherwise stated) and at the times indicated, quenched with SDS (0.5% final) and 10× loading buffer (100 mM EDTA, 5% SDS, 40% glycerol, 0.35% Bromphenol blue) was added. Reaction mixes were loaded on a 1.2% agarose gel, and electrophoresis performed in 1× TAE buffer at room temperature for 3 hr at 4 V/cm with buffer circulation. DNA was visualized by staining with ethidium bromide.
|
10.1371/journal.pcbi.0030147 | Distributed Representations Accelerate Evolution of Adaptive Behaviours | Animals with rudimentary innate abilities require substantial learning to transform those abilities into useful skills, where a skill can be considered as a set of sensory–motor associations. Using linear neural network models, it is proved that if skills are stored as distributed representations, then within-lifetime learning of part of a skill can induce automatic learning of the remaining parts of that skill. More importantly, it is shown that this “free-lunch” learning (FLL) is responsible for accelerated evolution of skills, when compared with networks which either 1) cannot benefit from FLL or 2) cannot learn. Specifically, it is shown that FLL accelerates the appearance of adaptive behaviour, both in its innate form and as FLL-induced behaviour, and that FLL can accelerate the rate at which learned behaviours become innate.
| Some behaviours are purely innate (e.g., blinking), whereas other, “apparently innate,” behaviours require a degree of learning to refine them into a useful skill (e.g., nest building). In terms of biological fitness, it matters how quickly such learning occurs, because time spent learning is time spent not eating, or time spent being eaten, both of which reduce fitness. Using artificial neural networks as model organisms, it is proven that it is possible for an organism to be born with a set of “primed” connections which guarantee that learning part of a skill induces automatic learning of other skill components, an effect known as free-lunch learning (FLL). Critically, this effect depends on the assumption that associations are stored as distributed representations. Using a genetic algorithm, it is shown that primed organisms can evolve within 30 generations. This has three important consequences. First, primed organisms learn quickly, which increases their fitness. Second, the presence of FLL effectively accelerates the rate of evolution, for both learned and innate skill components. Third, FLL can accelerate the rate at which learned behaviours become innate. These findings suggest that species may depend on the presence of distributed representations to ensure rapid evolution of adaptive behaviours.
| Both evolution and learning may be considered as different types of adaptation. Learning occurs within a lifetime, whereas genetic change occurs across lifetimes [1]. Whereas genetic change ensures that a task can be executed innately, learning permits even the most rudimentary innate ability to be honed into a useful skill.
In an environment that fluctuates from generation to generation, learning permits an innate ability to be adapted to the particular physical environment into which each generation is born. If the environment ceases to fluctuate, then genetic assimilation [2] can transform a rudimentary innate ability, which requires much learning, into an innate skill, which requires minimal learning. This transformation is more likely to occur if the cost of learning is high [3,4], and, in this case, computer simulations suggest that learning can accelerate the rate of genetic assimilation [5] via the Baldwin effect [6]. However, if learning is sufficiently inexpensive, then genetic change may not occur at all [7,8]. Overall, there appears to be a tradeoff between learning and genetic assimilation, such that learning can subsidize genetic change, especially if learning is inexpensive.
All but the most primitive organisms learn in order to survive, and organisms which learn quickly are at a selective advantage relative to those that learn slowly. Therefore, a mechanism which reduces the time required to learn a given behaviour confers a selective advantage. One candidate for such a mechanism is FLL [9,10].
As explained below, FLL ensures that in the process of learning one set of associations or behaviours another set of associations is usually learned. These associations could comprise either perceptual skills (such as face recognition, predator recognition [11], or prey recognition), or motor skills (such as catching prey, flying, seed pecking, or nest building).
Before considering how FLL can accelerate evolution of certain types of behaviours, FLL will be described in its original context of spontaneous recovery of memory in humans [9] and in neural network models [10]. Note that FLL is not unique to a specific class of network architectures, although it does assume that associations are learned using a form of supervised learning.
In humans, FLL has been demonstrated using a task in which participants learned the positions of letters on a nonstandard computer keyboard [9]. After a period of forgetting, participants relearned a proportion of these letter positions. Crucially, it was found that this relearning induced recovery of the non-relearned letter positions.
More recently, a set of theorems provided a formal characterization of FLL in linear neural network models [10]. In essence, FLL occurs in neural network models because each association is distributed amongst all connection weights (synapses) between units (model neurons). After partial forgetting, relearning some of the associations forces all of the weights closer to pre-forgetting values, resulting in improved performance even on non-relearned associations; a general proof is provided in [10]. A geometric demonstration of FLL for a network with two connection weights is given in Figure 1. Networks with multiple input and output units can be considered without loss of generality [10].
The protocol used to examine FLL in neural networks is as follows (see Figure 2). A network with n input units and one output unit has n connection weights. This network learns a set A of m ≤ n associations, where A = A1 ∪ A2 comprises two subsets A1 and A2 of n1 and n2 associations, respectively (note that m = n1 + n2). After the associations A have been learned and then partially forgotten, performance error on subset A1 is measured (forgetting is induced by adding isotropic noise to weights). Finally, only A2 is relearned, and then performance error on A1 is remeasured. FLL occurs if relearning A2 improves performance on A1. It has been proven that the probability of FLL approaches unity as the number of weights increases [10]. For the sake of brevity, this is reflected in phrases such as “learning A2 usually improves performance on A2” in this paper.
Now consider an organism b2 which is born with a genetically specified set of neuronal connections [12]. These connections are organised such that, if b2 learns one subset A2 of associations then another subset A1 is usually learned. In other words, the organism b2 happens to be born with neuronal connections similar to the connections of an organism b1 which had once learned and then forgotten subsets A1 and A2 (e.g., isotropically distributed around w0 in Figure 1). Just as FLL ensures that if organism b1 relearns A2 then subset A1 is usually relearned (see Figure 2), so if b2 learns A2 then A1 is usually learned. In both cases, FLL ensures that learning one subset of associations induces learning of the other subset. Critically, whereas the FLL exhibited by organism b1 depends on previous learning and forgetting, FLL in organism b2 depends on being born in a state such that the first time A2 is learned, the associations A1 are also usually acquired. Such a network can be evolved using a genetic algorithm, as shown below.
The use of two distinct subsets in this paper is clearly unrealistic when considered in the context of skill learning. However, the use of two subsets lies at one extreme along a continuum of tasks. At one extreme, associations are learned one by one in a strict order, and at the other extreme, all associations are learned simultaneously. In a biological context, the components of a skill which are learned first act as “scaffold” for others, and this effectively imposes a temporal order to the acquisition of different skill components. This is the type of scenario assumed for the simulations reported in this paper. Essentially, learning A2 is assumed to consist of a subset of skill components which provide a scaffold for the skill components in A1.
This section is a brief account of the basic geometry underlying FLL, in the absence of its interactions with evolution. For the present, and without loss of generality (see [10]), we assume that the network has one output unit and two input units, which implies n = 2 connection weights, and that A1 and A2 each consist of n1 = n2 = 1 association, as in Figure 1. Input units are connected to the output unit via weights wa and wb, which define a weight vector w = (wa,wb). Associations A1 and A2 consist of different mappings from the input vectors x1 = (x11,x12) and x2 = (x21,x22) to desired output values d1 and d2, respectively. If a network is presented with input vectors x1 and x2, then its output values are y1 = w · x1 = wax11 + wbx12 and y2 = w · x2 = wax21 + wbx22, respectively. More generally, network performance error for k associations is defined as
The weight vector w defines a point in the (wa,wb) plane. For an input vector x1, there are many different combinations of weight values wa and wb that give the desired output d1. These combinations lie on a straight line L1, because the network output is a linear weighted sum of input values. A corresponding constraint line L2 exists for A2. The intersection of L1 and L2 therefore defines the only point w0 that satisfies both constraints, so that zero error on A1 and A2 is obtained if and only if w = w0. Without loss of generality, we define the origin w0 to be the intersection of L1 and L2. A general prerequisite for FLL is that L1 is not orthogonal to L2.
We now consider the geometric effect of partial forgetting of both associations, followed by relearning A2. This geometric account applies to a network with two weights (see Figure 1), and depends on the following observation: if the length of the input vector x1 is unity, then the performance error E(w,A1) = (d1 − y1)2 of a network with weight vector w when tested on association A1 is equal to the squared distance between w and the constraint line L1 [10]. For example, if w is in L1, then E(w,A1) = 0, but as the distance between w and L1 increases so E(w,A1) must increase. For the purposes of this geometric account, we assume the length of the input vectors is unity.
If partial forgetting is induced by adding isotropic noise g to the weight vector w = w0, then this effectively moves w to a randomly chosen point w1 = w0 + g on the circle C of radius r, where r is the length of g, and represents the amount of forgetting. For a network with w = w1, learning A2 moves w to the nearest point w2 on L1 [10], so that w2 is the orthogonal projection of w on L2. Before relearning A2, the performance error E(w,A1) on A1 is the squared distance p2 between w1 and its orthogonal projection on L1. After relearning A2, the performance error Epost is the squared distance q2 between w2 and its orthogonal projection on L1. The amount of FLL is δ = E(w1, A1) − E(w2, A1), and (for a network with two weights) is also given by Q = p2 − q2. The probability P(δ > 0) of FLL given L1 and L2 is equal to the proportion of points on C for which δ > 0 (or equivalently, for which Q > 0). For example, it can be shown that the mean value of this proportion is P(δ > 0) for a two-weight network like the one shown in Figure 1A. Given the particular configuration shown in Figure 1A, the critical point wcrit is defined such that the performance error before and after learning is the same (i.e., δ = 0).
Given a network with n weights w and two subsets A1 and A2 of n1 and n2 associations, respectively, it is shown that weights w* exist such that learning associations A2 is guaranteed to yield zero performance error on A1, provided n ≥ n1 + n2.
Consider a network with n = 2 weights and subsets A1 and A2, each of which comprises a single association. Each association defines a constraint line L1 and L2, respectively (see Figure 1). If the weight vector w is in L1, then performance error on A1 is zero, and if w is in L2, then performance error on A2 is zero. Clearly, if and only if w is at the intersection w0 of L1 and L2, then performance error on both A1 and A2 is zero. If w is not in L2, then learning A2 moves w from its current position to its orthogonal projection onto L2 [10]. Crucially, if w = w* in Figure 1, then learning A2 moves w to the optimal weight vector w0. In this case, learning A2 reduces performance error to zero on both A1 and A2, and therefore learning A2 implies perfect performance on A1.
This line of reasoning generalises to networks with more than two weights, as follows. If a network has more than two input units, then subsets A1 and A2 can have n1 > 1 and n2 > 1 associations. If n ≥ n1 + n2, then A1 and A2 define an (n − n1)-dimensional subspace L1 and an (n − n2)-dimensional subspace L2, respectively. The intersection L12 of L1 and L2 corresponds to weight vectors which generate zero error on A = A1 ∪ A2. In this case, the circle in Figure 1 corresponds to an n-dimensional hypersphere, with its centre w0 in L12. Given that learning A2 provides an orthogonal projection of w onto L2, and that there exists a w = w* such that its orthogonal projection onto L2 is w0, it follows that learning A2 in a network with w = w* yields zero performance error on both A2 and A1.
Given that the weight vector w is genetically specified with finite precision, a network is necessarily born with its weight vector w = w1 at a non-zero distance r from the optimal weight vector w0. This finite precision defines a hypersphere C around w0, and the location of w1 on C determines the amount of FLL. If a network is born with w1 = w*, then learning A2 induces perfect performance on A1. If fitness depends on performance on both A1 and A2 after learning only A2, then there is selective pressure for networks to be born with weight vectors close to w*, given a specific degree of genetic precision r. More generally, there is pressure for networks to be born with weight vectors close to the subspace with contains w0 and w*.
As this paper deals with subtle combinations of evolution and learning, involving two distinct subsets of associations (A1 and A2), it is important to be clear about terminology. Specifically, we need to be careful about which subset is being referred to, and whether we are referring to innate performance or not. Accordingly, performance error on A1 before learning A2 is called just that, “innate performance error on A1.” Behaviours that are induced by learning A2 are called FLL-induced behaviours, because they are not innate, nor are they learned, so that performance error on A1 after learning A2 is called “FLL-induced performance error on A1.” If learning A2 does not affect performance on A1 (as in condition NoFLL, below), then this is referred to as “post-learning performance.” More generally, performance will be quoted with a specified context (e.g., performance on A1 after learning A2).
The effect of FLL on evolution was tested by measuring performance on A1 after learning A2 across generations. To eliminate the possibility that the observed results are artefacts, the effects of FLL were compared with two control conditions (described below).
Each generation consisted of 1,000 neural networks, each of which consisted of 20 input units and one output unit. The genome of each network was defined by a one-to-one mapping of the n = 20 weight values in the network to a single string of n genes, where the value of each gene was set to the value of a corresponding network weight. The number of offspring generated by each network was proportional to its fitness, which depended only on its ability to provide the correct desired output value for each of 20, n-element input vectors. The mapping from each input vector to its output value defines one association (see Figure 1).
A network's output yi is a weighted sum of input values
, where xij is the jth value of the ith input vector xi, and each weight wi is one input–output connection.
The fitness of each network was assessed with respect to its performance error on a single common set A = A1 ∪ A2 of m = 20 associations, where A1 and A2 are two disjoint subsets of n1 = 10 and n2 = 10 associations, respectively. The m associations in A were allocated randomly to the two subsets, A1 and A2. The subsets A1 and A2 were intended to represent different components of a task, and were therefore the same for all networks, and across all generations. In the first generation, each network's weight values were chosen from a Gaussian distribution (see below for details).
The desired output value di for each input vector xi was drawn from a Gaussian distribution with variance 1/n. An analytic method was used to solve for the optimal weight vector w0 which maps inputs to outputs: di = w0 · xi (see below). Given the variances of the inputs and outputs, the expected length of w0 is unity.
Each new generation was formed from 1,000 matings between 1,000 pairs of networks. The Kth network was chosen for mating according to its fitness F(K) with probability p(K). Networks were chosen with replacement to ensure that the number of offspring from a given network was proportional to p(K), which is defined as
where the denominator ensures ∑kp(k) = 1. Half the weights of each offspring were copied from (randomly chosen) corresponding weight locations in one parent network, and half from the other parent. Aside from mutations, weight values inherited by an offspring were the same as those inherited by its parents (i.e., inheritance was Darwinian, not Lamarckian).
Mutation was applied to each weight with a probability of 0.05, using a uniform probability density function. Then Gaussian noise with a standard deviation of 0.05 was added to the value of those weights that had been chosen for mutation.
There were three conditions: FLL, NoFLL, and NoLearn, with corresponding fitness functions FFLL, FNoFLL, and FNoLearn. The initial randomly chosen weight values (see Network Learning Algorithm) of the population of networks were the same in all conditions. Networks were selected for mating according to their performance on the combined set A = A1 ∪ A2, according to Equation 2 for all fitness functions, as described next.
Networks that exhibited high levels of FLL were preferentially selected for mating. Only associations A2 were learned, but the fitness of each network depended on its performance on both the learned associations A2 and on the unlearned associations A1. The fitness FFLL(K) of the Kth network is defined in terms of its innate performance error Epre on A = A1 ∪ A2, and on its performance error Epost on A after learning A2:
where Epre and Epost are:
where
and
are the network's output errors in response to the ith input vector before and after learning A2 (respectively). The parameter c = 0.05 defines the balance between performance error on innate versus post-learning (e.g., FLL-induced) behaviours, and is interpreted as a cost-of-learning parameter (see below). The network's fitness error Di is a function of the difference ei = yi − di between the network's response yi to the ith input vector and the desired output value di:
This ensures that output errors above Dthresh have a disproportionately large and detrimental effect on fitness, as shown in Figure 3. This, in turn, ensures that only those networks with “good” performance are likely to be selected for reproduction. The value of Dthresh was set to 0.01.
For later use, we also define the fitness errors Epre = Epre(A1) + Epre(A2), and Epost = Epost (A1) + Epost(A2) ≈ Epost(A1). The approximation here and in Equation 5 emphasises the fact that the total fitness error is attributable almost exclusively to A1 after learning A2 (because error on A2 is then almost zero).
The inclusion of innate performance error Epre in FFLL ensures that the cost of learning is taken into account when assessing fitness. If c is small, then Epre tends to be large, so that much learning is required to increase fitness. Conversely, if c is large, then Epre tends to be small, so that little learning is required to increase fitness. Thus, Epre is an implicit measure of the cost of learning (where c multiplies Epre), and ensures that networks which require minimal learning have high innate fitness (although the small value of c used in Equation 3 defines a relatively low cost of learning).
This was identical to condition FLL, except that the effects of FLL were precluded by making all input vectors in A mutually orthogonal, whilst retaining the length of each vector as in condition FLL, using Gram-Schmidt orthogonalisation. This makes the input vectors in A1 orthogonal to those in A2, which ensures L1 and L2 are orthogonal. This, in turn, ensures that learning A2 cannot affect performance on A1. The fitness function FNoFLL was the same as in condition FLL (i.e., FNoFLL = FFLL).
No learning occurred, so that improvement in performance over successive generations was due only to selection of innate performance on A1 and A2. Fitness was defined as FNoLearn = 1/ Epre, where Epre is defined in Equation 4. This is equivalent to setting c = 1 in Equation 3.
The network learning algorithm used here involves a type of supervised learning. Note that Equation 1 defines the network error used for learning, whereas Equation 3 defines the fitness of a network.
Each network was initialised with n weight values drawn randomly from a Gaussian distribution with unit variance. This was then divided by n1/2, which ensures that the expected length of weight vectors in the population is unity.
Given a network with n input units and one output unit, the set A of m, n-element input vectors xi: i = 1…m and m desired scalar output target values di were chosen randomly from a Gaussian distribution with unit variance. Each input vector was then divided by n1/2 so that the expected length of input vectors was unity (i.e., the variance of input values was 1/n).
In conditions FLL and NoFLL, each network learned n2 associations. Rather than using the iterative weight update normally associated with the delta rule, an analytic solution was obtained. Learning n2 associations consists of finding the orthogonal projection operator which projects the initial weight vector w1 to its nearest point in the subspace (e.g., L2) defined by the n2 input vectors being learned. The end result w2 is the same as that obtained using the standard delta rule for infinitesimal learning rates [10]. As with the standard delta rule, this yielded a value of approximately zero for post-learning performance error on the learned associations A1. This type of learning is most plausibly associated with motor learning in the cerebellum and basal ganglia [10].
The results are based on ten computer simulation runs for each of the three conditions, FLL, NoFLL, and NoLearn, described above, and graphs show the mean of these ten runs. Each run involved a different fixed set A of 20 associations. As a reminder, the two free parameters are: 1) the cost-of-learning parameter, which was set to c = 0.05, and 2) the threshold of the fitness error function, which was set to Dthresh = 0.01.
The main results are shown in Figure 4, which is a summary of more detailed results in Figure 5. Condition FLL yields a FLL-induced error (i.e., error on A1 after learning A2) of approximately zero after 30 generations, whereas condition NoFLL requires about 60 generations to achieve an error of less than unity. Condition NoLearn (dotted curve) yields the slowest innate learning, and is included for comparison.
The proportion of networks that exhibit FLL over generations is shown in Figure 6A, and the amount of FLL is shown in Figure 6B. The proportion and amount of FLL increases in condition FLL, as indicated by the solid line in each figure. The zero prevalence of FLL in condition NoFLL (dashed line) is associated with zero FLL as indicated in Figure 6B. More detailed results are shown in Figure 5A–5C.
Performance on A1 (solid line in Figure 5A) after learning A2 is better in condition FLL than in condition NoFLL (solid line, 5B). Innate performance on A1 is also better in condition FLL (dashed line, Figure 5A) than in conditions NoFLL (dashed line, Figure 5B) and NoLearn (dashed line, Figure 5C). Together, these results suggest that FLL accelerates both the rate at which FLL-induced behaviours (A1) appear, as well as the rate at which FLL-induced behaviours (A1) become genetically assimilated.
Innate performance on subset A2 is better in condition FLL (dotted curve, Figure 5A) than in conditions NoFLL (dotted curve, Figure 5B) and NoLearn (dotted curve, Figure 5C).
Learning A2 reduces error on subset A1 even in the first generation in conditions FLL (Figure 5A). Additionally, the proportion of networks showing FLL is greater than 0.5 (Figure 6A), and the amount of FLL is greater than zero (Figure 6B) in the first generation. These effects are not due to any special properties of the networks nor of the associations. Indeed they are entirely expected, and are consistent with the theoretical analysis in [10]. In essence, FLL is observed in the first generation because it is very unlikely that the mainfolds L1 and L2 defined by A1 and A2 (respectively) are orthogonal, so that learning A2 usually reduces error on A1 (albeit by a small amount in the first generation).
Before discussing results in detail, it is important to clarify precisely what is being claimed here. The main claim is that, given a population of organisms which can learn, the presence of FLL accelerates the rate at which a given set of advantageous behaviours evolves relative to populations which 1) can learn but which do not have FLL (e.g., condition NoFLL) and 2) cannot learn (e.g., condition NoLearn). Specifically, it is claimed that FLL accelerates the appearance of adaptive behaviour, both in its innate form and as FLL-induced behaviour, and that FLL can accelerate the rate at which learned behaviours become innate.
It is also claimed that FLL increases the rate at which a set of behaviours (e.g., A) is acquired within a lifetime. Clearly, if learning one subset (e.g., A2) induces learning of another subset (e.g., A1), then the amount of learning required to learn both subsets (e.g., A) is reduced.
It is worth noting that FLL is not related to generalisation (see [10]), which cannot therefore be responsible for the effects reported here.
The task was purposely made difficult, such that network outputs which were not close to desired target values were assigned an error value of unity. This heavily penalises networks that do not generate near-correct responses. This type of task may emulate tasks for which being “almost correct” provides no fitness benefit. Such tasks are exemplified by a predator which almost catches prey (e.g., a kingfisher almost catching a fish, or where each failed attempt yields a large fitness cost), or where learning is incremental and stepwise (e.g., learning to catch progressively larger prey). Such tasks give rise to “needle-in-a-haystack” search spaces [5], which have rugged or uncorrelated landscapes [13].
A cogent critique of research by Nolfi et al. [14] argues that accelerated evolution (specifically, assimilation) is a generic consequence of learning per se [15]. In results not shown here, replacing A2 with a new, randomly chosen subset every generation in condition FLL yields a more gradual evolution of FLL-induced and innate behaviours than is obtained in any of the conditions used here. This effectively excludes the possibility that the accelerated evolution reported here is due to learning per se.
In terms of evolutionary theory, FLL-induced behaviours can be considered as the establishment of a new reaction norm. The specific “environment” that induces the reaction norm is learning a particular subset of behaviours (A2), and the phenotypic reaction to this environment is another subset of behaviours (A1).
FLL does not necessarily force FLL-induced behaviours to become genetically assimilated. In fact, there is a tradeoff between the amount of acceleration induced by FLL and the extent to which behaviours become innate. If the cost-of-learning parameter is set to c = 1, then there is no incentive for FLL to increase over generations. In contrast, if c ≈ 0 (as in the simulations reported here), then the rapid evolution of FLL-induced behaviour shown in Figure 5A (solid line) is obtained, alongside the slower evolution of innate behaviour (dashed line in Figure 5A). Thus, even the small value of c (0.05) used here puts pressure on learned behaviours to become innate, as indicated by the decreasing innate performance errors on A1 (dashed line) and A2 (dotted line) in Figure 5A.
In practice, learning always has a non-zero fitness cost, if only in terms of the time required for that learning to occur. This is because time spent learning is time spent not eating, or time spent being eaten, both of which reduce fitness. Thus, the small value of c used here represents one value along the spectrum of learning costs. It therefore seems likely that even the simplest learned behaviours have a tendency to become innate, and that this tendency increases with the cost of learning. For example, in results not shown here, increasing the cost-of-learning parameter c decreases the rate at which FLL-induced performance on A1 improves, and increases the rate at which performance on A1 and A2 becomes innate (innate performance with c = 1 is effectively obtained in condition NoLearn (see Figure 5C)). It is therefore not easy to classify the effect reported here as a clear-cut example of the Baldwin effect [6], although these effects are almost certainly related.
The basic geometry which underpins FLL within a lifetime (as in [10]) and across lifetimes (as here) can also be applied in two other contexts: 1) evolution of innate behaviours without learning, and 2) evolution of general phenotypic traits. These two cases are considered in the next two paragraphs. Both of these effects require the presence of environmental conditions that fluctuate over successive generations (e.g., fluctuations in temperature induced by ice ages, salinity, prey numbers, or predation pressure).
1) Accelerated evolution of innate behaviours without learning can be understood by considering an organism that has no learning ability, and which relies on genetic specification of its neuronal connections [12]. Natural selection ensures that its neuronal connections at birth yield innate behaviour matched to its environment. If the environment changes, then natural selection will induce a corresponding shift to a new set of innate connections. If the environment then shifts back to its original state, then organisms' connections will tend to revert to their original values. Let us assume that some connections revert faster than others over successive generations. For example, some connections may be specified by genes linked to other innate behaviours, and this genetic linkage would tend to reduce the rate of genetic change. In fact, for simplicity, assume that half of the connections revert quickly and half revert slowly. If the required behaviours are encoded as distributed representations, then this connection reversion will induce a FLL-type effect, such that all associations benefit from the reversion of a proportion (half here) of connection values.
2) Accelerated evolution of general phenotypic traits can be understood if we assume an extreme form of pleiotropy: that each of a given set of genes affects every phenotypic trait. This is equivalent to assuming that the genome is a distributed representation of the phenotype. Consider a population in which the fittest organism has a genome w0 which is perfectly adapted to its environment e0. If the environment changes to e1, then the fittest organism's genome will eventually evolve to a new state w1 that is suited to e1 (this is analogous to forgetting in FLL). Now, consider what happens if the environment changes back to e0. The fittest organism's genome will be forced back toward w0, but inevitably some genes will revert faster than others. For the sake of argument, assume that a subset G2 of genes revert to their original values, while others G1 remain as they were in w1 (this is analogous to relearning only A2). Because each gene in G2 contributes to every phenotypic trait, the reversion of genes in G2 to their original values will push the entire phenotype back toward its state in the original environment e0. Thus, the reappearance of an entire set of phenotypic traits (e.g., changes in size) can occur more quickly if those traits are encoded within a set of pleiotropic genes than if each trait is represented by a non-pleiotropic gene, and suggests a form of free-lunch evolution.
It has been demonstrated that FLL accelerates the evolution of behaviours in neural network models. Given that FLL appears to be a fundamental property of distributed representations, and given the reliance of neuronal systems on distributed representations, FLL-induced behaviours may constitute a significant component of apparently innate behaviours (e.g., nest-building). Results presented here suggest that any organism that did not take advantage of such a fundamental and ubiquitous effect would be at a selective disadvantage. Finally, if FLL accelerates evolution in the natural world, then it may have been involved in the Cambrian explosion, an explosion that began when brains (and therefore learning) first appeared. |
10.1371/journal.pgen.1003707 | Hard Selective Sweep and Ectopic Gene Conversion in a Gene Cluster Affording Environmental Adaptation | Among the rare colonizers of heavy-metal rich toxic soils, Arabidopsis halleri is a compelling model extremophile, physiologically distinct from its sister species A. lyrata, and A. thaliana. Naturally selected metal hypertolerance and extraordinarily high leaf metal accumulation in A. halleri both require Heavy Metal ATPase4 (HMA4) encoding a PIB-type ATPase that pumps Zn2+ and Cd2+ out of specific cell types. Strongly enhanced HMA4 expression results from a combination of gene copy number expansion and cis-regulatory modifications, when compared to A. thaliana. These findings were based on a single accession of A. halleri. Few studies have addressed nucleotide sequence polymorphism at loci known to govern adaptations. We thus sequenced 13 DNA segments across the HMA4 genomic region of multiple A. halleri individuals from diverse habitats. Compared to control loci flanking the three tandem HMA4 gene copies, a gradual depletion of nucleotide sequence diversity and an excess of low-frequency polymorphisms are hallmarks of positive selection in HMA4 promoter regions, culminating at HMA4-3. The accompanying hard selective sweep is segmentally eclipsed as a consequence of recurrent ectopic gene conversion among HMA4 protein-coding sequences, resulting in their concerted evolution. Thus, HMA4 coding sequences exhibit a network-like genealogy and locally enhanced nucleotide sequence diversity within each copy, accompanied by lowered sequence divergence between paralogs in any given individual. Quantitative PCR corroborated that, across A. halleri, three genomic HMA4 copies generate overall 20- to 130-fold higher transcript levels than in A. thaliana. Together, our observations constitute an unexpectedly complex profile of polymorphism resulting from natural selection for increased gene product dosage. We propose that these findings are paradigmatic of a category of multi-copy genes from a broad range of organisms. Our results emphasize that enhanced gene product dosage, in addition to neo- and sub-functionalization, can account for the genomic maintenance of gene duplicates underlying environmental adaptation.
| Existing genetic diversity reflects evolutionary history, but it has rarely been possible to probe for footprints of selection at loci known to functionally govern adaptive traits. Both naturally selected metal hypertolerance and extraordinary leaf metal accumulation of the extremophile Arabidopsis halleri require strongly enhanced transcript levels of Heavy Metal ATPase4 (HMA4) encoding a PIB-type ATPase that pumps Zn2+ and Cd2+ out of specific cells. By comparison to the metal-sensitive A. thaliana, highly elevated HMA4 expression results from a combination of gene copy number expansion and cis-regulatory modifications. But how do these findings, which were based on a single accession, relate to species-wide HMA4 sequence diversity in A. halleri? Addressing this question, we detect positive selection in the promoter regions of three tandem A. halleri HMA4 paralogs, which are uniformly cis-activated. The accompanying hard selective sweep, however, is segmentally eclipsed as a consequence of recurrent ectopic gene conversion among HMA4 protein-coding sequences, which undergo concerted evolution. Together, this constitutes an unexpectedly complex profile of polymorphism as a result of natural selection. Our observations can serve as a blueprint for future analyses of duplicated genes that have undergone selection for more of the same gene product.
| Analyses of nucleotide sequence variation bear great promise for advancing our understanding of evolutionary processes. However, such analyses have so far rarely targeted loci of experimentally established roles in naturally selected adaptive traits, and, instead, have mostly been conducted on candidate loci or even anonymous sequences [1]–[3]. Among the highest selection pressures known in ecology are those encountered by plants on metalliferous soils, which contain high, toxic levels of heavy metals from geological anomalies or anthropogenic contamination [4]. Examples of metalliferous soils are the widespread ultramafic (serpentine) soils rich in Ni, Co and Cr, and calamine soils containing high levels of Zn, Cd, and Pb. The extremophile species Arabidopsis halleri is one of the few plant taxa capable of colonizing calamine metalliferous soils [5]. In addition to its hypertolerance to Zn, Cd and likely Pb, A. halleri groups among approximately 500 known taxa of so-called hyperaccumulators of metals such as Ni, Co, Zn or Cd [6], [7]. Hyperaccumulators are characterized by leaf metal concentrations exceeding those of ordinary non-accumulator plants by more than two orders of magnitude. Metal hyperaccumulation contributes to metal hypertolerance and has been proposed to act as an elemental defense against biotic stress [8], [9].
A. halleri is closely related to Arabidopsis lyrata and to the genetic model plant Arabidopsis thaliana, both of which are non-hyperaccumulators and exhibit only basal metal tolerance common to all vascular plants [10]. Different from A. thaliana, A. halleri is an outcrossing, stoloniferous perennial, with a nuclear genome of 2 n = 16 chromosomes [6]. In an attempt to address the molecular basis of Zn and Cd hyperaccumulation and associated hypertolerance in A. halleri, cross-species transcriptomics approaches employing the accession Langelsheim (Germany) established dozens of candidate genes with potential functions in metal homeostasis, of which transcript levels were elevated in A. halleri when compared to A. thaliana [11]–[13]. Functional characterization through various molecular approaches supported a role for several of these genes including HEAVY METAL ATPASE4 (HMA4) [8], [12], HMA3 [11], METAL TRANSPORT PROTEIN1 (MTP1) [11], [14], NICOTIANAMINE SYNTHASE2 (NAS2) [13], [15], and IRON-REGULATED TRANSPORTER3 (IRT3) [6], [7], [16]. Transcript abundance of HMA4 was highest of all identified candidate genes, with more than 100-fold higher transcript levels in both roots and shoots of A. halleri than in A. thaliana or A. lyrata [12], [17]. The HMA4 protein is a plasma membrane transport protein acting in ATP-driven cellular export-mediated detoxification of Zn2+ and Cd2+, as well as root-to-shoot translocation of both metals [8], [18]. The strongly enhanced HMA4 transcript levels present in A. halleri were shown to be necessary not only for metal hypertolerance but also for metal hyperaccumulation, by employing RNA interference-mediated silencing in the A. halleri accession Langelsheim. The introduction into A. thaliana of an AhHMA4 promoter fused to an AhHMA4 cDNA suggested that AhHMA4 alone, however, is not sufficient to generate either metal hypertolerance or hyperaccumulation [8]. In agreement with these findings, genetic studies identified HMA4 and MTP1 to be located within rather large QTL regions for metal hypertolerance in a segregating back-cross 1 population of an inter-specific hybrid cross between A. halleri (accession Auby, France) and A. lyrata [17], [19]. Moreover, HMA4 co-localized with one out of several major QTL for leaf Zn and Cd hyperaccumulation, respectively, in a segregating F2 population [20], [21]. Among the candidate genes of A. halleri characterized in detail to date, HMA4 thus makes the largest contribution to both metal hyperaccumulation and metal hypertolerance.
High HMA4 transcript levels were shown to be attributable to a combination of tandem gene triplication and cis-activation in the Langelsheim accession of A. halleri [8]. Promoter-reporter fusions suggested approximately equivalent quantities and localizations of promoter activity for all three A. halleri HMA4 gene copies, in agreement with copy-specific transcript quantification through quantitative real-time RT-PCR [8]. Because of almost identical protein-coding sequences, the functions of the three HMA4 protein isoforms of A. halleri have not been individually characterized. All these findings supported a critical role of enhanced HMA4 gene product dosage in naturally selected metal hyperaccumulation and hypertolerance of A. halleri [8]. Interestingly, high HMA4 transcript levels, copy number expansion and cis-activation were also reported in Noccaea caerulescens [22], [23], another Zn/Cd hyperaccumulator in the Brassicaceae family, in which metal hyperaccumulation and associated hypertolerance must have evolved independently. Moreover, copy number expansion appears to be common among additional highly expressed metal hyperaccumulation/hypertolerance candidate genes of A. halleri, for example the ZINC-REGULATED TRANSPORTER, IRON-REGULATED TRANSPORTER-RELATED PROTEIN (ZIP) genes ZIP3, ZIP6 and ZIP9 [12], MTP1 [14], [24] and PLANT DEFENSIN (PDF) genes [25].
Gene duplication is known as a major driver of genome evolution over long timescales [26]. In eukaryotic genomes, gene duplications occur spontaneously at rates that are between 100 and 10,000 times higher per locus than those of base substitutions per site [27], [28], thus explaining the presence of substantial gene copy number variation polymorphism in genomes. For example, per haploid genome and generation, S. cerevisiae was estimated to spontaneously acquire about 0.002 non-synonymous base substitutions within coding regions and 0.02 gene duplications [28]. A number of genetic diseases of humans are caused by gene duplication events [29], [30]. Current theory predicts the rapid loss of recent duplicates unless they undergo neo- or sub-functionalization, with few exceptions [26], [31], [32]. However, the factual contribution of gene duplication to evolutionary adaptation as an outcome of natural selection remains poorly understood. Functional diversity and evolutionary dynamics of multigene families are of particular importance in plant and animal immunity, as exemplified by plant Resistance (R) and human Major Histocompatibility Complex (MHC) genes [33]. Natural selection for increased gene product dosage was implied to account for copy number expansion of the BOT1 boron tolerance locus of barley [34], the MATE1 aluminum tolerance locus of maize [35] and the human salivary amylase gene (AMY1) [36]. However, these reports were based merely on functional data encompassing genotype-phenotype relationships, without evidence for selection from an analysis of sequence polymorphism.
Here, we address two gaps in present knowledge, namely whether a signature of selection can indeed be identified at a locus known to functionally govern an adaptive trait and, more specifically, whether positive selection for increased gene product dosage can result in the fixation of gene duplications [37]. We detect positive selection at the copy-number expanded HMA4 metal hypertolerance locus of Arabidopsis halleri. Moreover, we show that the profile of polymorphism is unexpectedly complex as a result of ectopic gene conversion. This work can act as a guide for related studies on other duplicated genes, and warrants caution in targeted analyses as well as genome-wide scans of polymorphism when dealing with presently or historically copy-number expanded loci.
For an analysis of intra-specific nucleotide sequence diversity across the triple HMA4 genes of A. halleri, we sequenced from multiple individuals (Table 1) series of 13 genomic DNA segments positioned consecutively along the 150-kb HMA4 region and in flanking regions (Figure 1, Figure S1A and Table S1). In more detail, amplicons of between 492 and 2245 bp in length (see Table S1) were designed based on published sequence data, and sequenced from between 15 and 20 individuals (see Table S2; http://www.ebi.ac.uk, accession nos. HE995813 to HE996227). The number of alleles observed per genotype never exceeded expectations of a maximum of two for any of the amplicons (see Table S2, lower section; see Materials and Methods section ‘Sequencing, Sequence Assembly and Assignment of Consensi’). We confirmed leaf metal accumulation in these same individuals by Inductively-Coupled Atomic Emission Spectrometry analysis of field-collected leaves. Maximal concentrations exceeded 10,000 µg Zn g−1 leaf dry biomass and 100 µg Cd g−1 leaf dry biomass in individuals from both non-metalliferous and metalliferous sites that are characterized by toxic levels of metals in the soil and a specialist vegetation (Table 1). For comparison, we also obtained nucleotide sequence data from single individuals of the Zn/Cd-hypertolerant and -hyperaccumulating subspecies A. halleri ssp. gemmifera [38] from East Asia and the closely related Zn/Cd-sensitive, non-hyperaccumulating Arabidopsis lyrata. The genome of A. lyrata contains a single functional HMA4 gene (Figure S1B) in a region that is overall syntenic to A. halleri (Figure 1 and S1A) and A. thaliana (Figure S1C) [39]. In addition, the A. lyrata genome uniquely contains a second, 5′-truncated HMA4-like pseudogene in a non-syntenic position.
If a novel mutation confers a strong selective advantage, the corresponding haplotype is likely to sweep through a population. This reduces or even eliminates pre-existing nucleotide sequence diversity at the affected locus and – proportionately to the extent of genetic linkage – at flanking loci through genetic hitchhiking [40]. In order to test for evidence of a selective sweep in the HMA4 genomic region of A. halleri, we calculated statistics of genetic diversity. At distant control loci (S1 and S13) and loci flanking the ∼150-kb HMA4 genomic region (S2 and S12), average pairwise nucleotide sequence diversity (π) was between 4.9 and 9.1‰ (Figure 1A and Table S2), and thus within the published range for random neutral loci in A. halleri [41]–[43]. Comparable studies on A. halleri ssp. halleri have reported a median π of 3.9‰ (between 0.3 and 37.7‰) for 24 randomly chosen loci [42] and a median π of 4.3‰ (between 1.8 and 32.7‰) for a total of 8 loci [41] (Figure S2). Indeed, this was in sharp contrast with much lower values for π of between 0.1 and 1.8‰ for segments comprising sequences in the promoter regions of the three paralogous HMA4 gene copies (S4, S6, S9; Figure 1A, Figure S1D and Table S2). Compared to the distant and flanking control loci, π decreased gradually towards and within the HMA4 region and reached a minimum of 0.1‰ at the HMA4-3 promoter (S9), yielding a profile as expected upon a hard selective sweep. This characteristic profile of nucleotide sequence diversity was found to be interrupted, however, by elevated π values of between 3.2 and 5.2‰ for segments positioned within the coding sequences of the three HMA4 gene copies (S5, S7, S10) and also for the additional segment S8, all comprising sequences that are present in two or more, almost identical copies in the HMA4 genomic region (Figure 1A, Tables S2 to S4). The overall profile of nucleotide sequence diversity across the HMA4 region was robust against error in sequence assignment to S5, S7 and S10 (see Materials and Methods, Table S4A), as well as towards a regionally separate analysis of individuals from the Harz Mountains and the Thuringian Forest (Table S4B).
To further substantiate the evidence for positive selection in the genomic HMA4 region of A. halleri, we conducted statistical tests of molecular population genetics by calculating Tajima's D, Fu and Li's D*, and Fu and Li's F* [44], [45]. For segments in the promoter regions of HMA4-1 (S4), HMA4-2 (S6) and HMA4-3 (S9), these three tests unanimously indicated an excess of rare polymorphisms resulting from a depletion of higher-frequency, ancestral polymorphisms. A statistically significant deviation from expectations under neutral evolution was detected at the promoter of HMA4-3 (S9; Figure 1B and Table S2), diagnostic of positive selection. Indeed, diversity statistics indicated a unique combination of a very low value for π with a highly negative Tajima's D for S9 (see Figure S2). In agreement with these results, there were fewer long and intermediate-length branches in the topologies of maximum likelihood phylogenetic trees for HMA4 (S4, S6, S9) than observed for control loci on either side of the HMA4 region (S1, S2, S12, S13; Figure 2, Figure S3) [44]. In the region of extremely low sequence diversity in the promoter region of HMA4-3 (S9) of A. halleri ssp. halleri (see Figure 1A), for example, all polymorphisms were unique to single observations (e.g., Figure 2B, 1.1-2, 1.3-2, 5.1-1, 7.2-1; see also Figure 1B). Taken together, statistical tests of sequence diversity, molecular population genetics and sequence phylogenies concordantly support a hard selective sweep centered on the promoter of HMA4-3, with genetic hitchhiking [40] covering a total of 250 kb. This is comparable to previously reported selective sweeps, which affect chromosomal regions of between 60 and 600 kb in length linked to domestication loci of crop plants [46].
As demonstrated in a single individual of A. halleri [8], the combination of gene copy number expansion and cis-regulatory divergence results in strongly enhanced steady-state HMA4 transcript levels that are necessary for metal hyperaccumulation and hypertolerance. If this was selected for in the entire species A. halleri, as indicated by the diversity statistics (see Figure 1), then we would expect high HMA4 transcript levels in all A. halleri individuals. Indeed, we observed between 20- and 130-fold higher HMA4 transcript levels across individuals of A. halleri from different collection sites, when compared to A. thaliana (Figure 3). This result supports a substantial increase in HMA4 gene product dosage in all A. halleri ssp. halleri and ssp. gemmifera individuals analyzed here, by comparison to A. lyrata and A. thaliana.
For segments located within coding sequences of the three HMA4 gene copies (S5, S7, S10), relationships among haplotypes differed from those for segments located in HMA4 promoters (S4, S6, S9). Phylogenetic reconstructions of S5, S7 and S10 did not recover three distinct groups of haplotypes as expected for three independently evolving paralogs (Figure S4). Instead, the genealogy resembled a network-like structure, with complex relationships between HMA4 haplotypes at different loci (Figure 4A). For example, out of a total of 25 haplotypes, three were found at two or more of the paralogous HMA4 genes (h13, h20, h25; Figure 4B). These results demonstrate a recurrent transfer of genetic information between the coding sequences of different HMA4 gene copies of A. halleri.
Segmental transfer of genetic information between paralogous sequences can arise in somatic cells during homologous recombination-based repair of double-strand breaks, addressed here as ectopic gene conversion (EGC, also termed interlocus or non-allelic gene conversion), or alternatively result from unequal crossing-over events during meiosis [29], [47], [48]. Quantitative PCR analysis of genomic DNA of A. halleri individuals from different collection sites was consistent with the species-wide presence of three HMA4 gene copies per haploid genome (Figure 5). Average gene copy number was estimated at 3.2±0.2 for A. halleri, compared to 1.8±0.2 and 1.0±0.1 for A. lyrata and A. thaliana, respectively (arithmetic means ± SD), whereby one of the two gene copies detected in A. lyrata is a truncated pseudogene in a non-syntenic position (see Figure S1B). A total of three HMA4 gene copies is in agreement with our observations of a maximum of six alleles observed per individual upon joint PCR amplification of all 3′ HMA4 coding sequences (S5/S7/S10; Table S2), and a maximum of two alleles observed in the promoter region of each HMA4 gene copy (S4, S6, S9; Table S2). The lack of evidence for HMA4 copy number variation among A. halleri individuals suggests that recurrent EGC events account for the segmental transfer of genetic information between paralogous HMA4 coding regions. EGC is known to be common among some genes, for example rRNA genes [29], [47], [49]–[51]. Paralogous genes of eukaryotes have been reported to exchange sequence information at per-locus frequencies even higher than those of spontaneous gene duplications [52], [53], thus contributing significantly to human disease [54]. The contribution of EGC to adaptation, however, is poorly understood.
EGC is predicted to transfer a newly arisen mutation from the site of its origin in one HMA4 paralog to the corresponding sites in the other two paralogs, thus cumulatively enriching species-wide sequence diversity in each individual HMA4 gene copy [55]. This explains the higher levels of nucleotide sequence diversity detected at S5, S7, S8 and S10, when compared to S4, S6 and S9 (see Figure 1) [29], [32]. Simultaneously, EGC suppresses between-copy sequence divergence and thus results in the concerted evolution of the affected loci [29]. Our findings imply that EGC accounts for the high extent of 99 to 99.3% inter-copy sequence identity among A. halleri HMA4-1 to -3 coding sequences (Table S3) [8], consistent with the prevalence of EGC among duplicates of >95% sequence identity known in other organisms [29], [47].
Hallmarks of EGC were also detected in the multi-copy portion of segment S8 outside the HMA4 coding sequence (Figure S5A, Table S3), again with a network-like genealogy (Figure S5B–D) and a comparably high π of 9.5‰ (as opposed to π of 1.7‰ for the single-copy 3′-portion of S8). As in A. halleri ssp. halleri, EGC was also evident among the coding sequences of HMA4 gene copies of A. halleri ssp. gemmifera (Figure S5D and S6), with an apparent additional EGC event between the promoters of HMA4-2 and -3 that was uniquely observed in this individual (S6, S9; see Figure S1D and compare Figure 2B and Figure S3C and S3D).
Population genetics theory and simulations have been developed for small multigene families undergoing concerted evolution [56]–[59]. Nucleotide substitution rates were predicted to be strongly enhanced with increasing gene copy number for beneficial mutations, whereas gene copy number had no effect on substitution rates for selectively neutral mutations [57]. Indeed, the AhHMA4 protein-coding sequences represented in S5/S7/S10, which correspond to the cytoplasmic C-terminal regulatory domain of the HMA4 protein [60], show an over-proportionately high nucleotide sequence divergence of 22% from A. thaliana [8]. By comparison, within coding regions in general, average divergence of both A. halleri and A. lyrata from A. thaliana is around 6%. In the corresponding region of HMA4, A. lyrata is 9% divergent from A. thaliana and 22% divergent from A. halleri. This suggests an enhanced rate of fixed sequence alterations in 3′ AhHMA4 coding sequence of S5/S7/S10, which – according to theoretical considerations – is likely to constitute evidence for positive selection [57]. Different from predictions, however, there is no prevalence of non-synonymous over synonymous nucleotide substitutions in this region, but a prevalence of indel polymorphisms instead. Nonetheless, these considerations suggest that HMA4 gene copy number expansion is not only a result of selection for enhanced gene product dosage, but – in combination with EGC – accommodates an enhanced evolutionary rate of HMA4 under positive directional selection.
Regions of the human genome hosting multigene families that undergo segmental exchange of sequence information have been addressed as hypermutable [8], [30]. Similarly, sequence exchange was proposed to contribute to the unusually high levels of sequence diversity among plant disease Resistance (R) genes, which typically belong to multigene families and are often present in the genome as tandem arrays of multiple paralogous genes [61]. Alongside unequal crossing over and illegitimate inter-allelic recombination, EGC was implicated in the generation of novel pathogen recognition specificities [33], [61]–[64]. The pervasiveness of sequence differences between paralogous R genes, despite sequence exchange, was attributed to small exchanged tracts of sequence of mostly <100 bp among multiple paralogs [61], [63], [64], to the suppression of unequal crossing over within R gene clusters of homozygotes [61], to the occurrence of inter-allelic rather than inter-locus gene conversion [33], or to the past discontinuation of sequence exchange [65]. By comparison to the high sequence diversification among paralogous R genes, the concerted evolution of A. halleri HMA4 paralogs is in stark contrast. This could be interpreted to indicate a prominent role for selection in determining the outcome of inter-locus sequence exchange, a process that appears to be common at least in some classes of multigene families [61], [62], [64].
The evolutionary events reflected in the profile of nucleotide sequence diversity across the HMA4 region of A. halleri occurred concurrent with or after the divergence from the A. lyrata lineage. Whereas nucleotide sequence diversity within A. halleri ssp. halleri was not positively correlated with the genetic divergence from A. lyrata across the HMA4 region (Figure S7A), we detected shared ancestry of nucleotide sequence diversity profiles in the two subspecies of A. halleri, ssp. halleri and ssp. gemmifera. This is supported by a positive correlation between inter-subspecies sequence divergence and sequence diversity π within ssp. halleri (Figure S7B), by the grouping of ssp. gemmifera alleles among ssp. halleri alleles in genealogies (Figures 2, S1D, S3, S4), and by shared polymorphisms among the two A. halleri subspecies in the coding sequences of HMA4 genes (Figure S6) as well as downstream of HMA4-2 (S8) (Figure S5D). These findings also indicate that our sampling captured a large proportion of sequence diversity within A. halleri, which was further confirmed by a larger genetic diversity of A. halleri within collection sites than between collection sites or between regional subgroups of collection sites according to analyses of molecular variance (AMOVA) (Table S5).
Our results support two consecutive duplications of HMA4 with or after the split of the lyrata and halleri lineages, which was estimated at between 2 mio. years ago according to sequence divergence [66] and around 0.34 mio. years ago according to approximate Bayesian computation [43]. Previous estimates of the timing of HMA4 duplication events 0.36 and 0.25 mio. years ago, respectively, are likely to require downward adjustment as they were based on single A. halleri sequences for each of the three gene copies and did not take into account EGC [43].
Enhanced HMA4 gene product dosage is known to functionally underlie the environmental adaptations of heavy metal hyperaccumulation and hypertolerance in the wild plant A. halleri [8]. Here, we detect positive selection in HMA4 promoter regions of A. halleri, incurred by either activating cis-regulatory mutations or gene copy number expansion of HMA4, and likely by both. Furthermore, we identify ectopic gene conversion to effect the concerted evolution of paralogous HMA4 coding sequences, a finding that adds unexpected complexity to the profile of sequence polymorphism. We expect that, together, our results coin a class of multi-copy genes associated not only with instances of environmental adaptation in plants [6], [51], [67], but also more generally with eukaryotic adaptation [29], [32], [36], [37], [68]. Thus, this work will stimulate the development of crop breeding strategies based on gene copy number variation [34], [69]. In the future, complex profiles of nucleotide sequence polymorphism, as exemplified by the HMA4 region of A. halleri, will deserve designated attention in advanced targeted studies as well as in large-scale genome scanning approaches [2], [3], [70]. Subsequent to gene duplication events [27], alongside neo- and sub-functionalization, selection for more of the same gene product is of higher evolutionary relevance than previously appreciated [26], [31], [32], [71].
Leaf tissues and soil samples were collected in the field from 18 randomly selected A. halleri ssp. halleri individuals at 7 European sites (Table 1). A minimum distance of 2 m was kept between sampled individuals to avoid sampling clones because A. halleri is stoloniferous. From a subset of collected genotypes, clones were propagated vegetatively and maintained in a greenhouse. For element analysis by Inductively-Coupled Plasma Atomic Emission Spectrometry (ICP-AES), leaf material was washed with ultrapure water and dried at room temperature (RT) for >1 week, followed by processing of samples and measurements as described [11], [12]. For the determination of extractable soil metal concentrations, soil cores were taken down to 0.05 m depth within 0.1 m distance from each individual. Three g of air-dried, sieved soil (2 mm particle size) were extracted in 25 ml of 0.1 M HCl with rotary shaking at 150 rpm at RT for 0.5 h.
For DNA extraction, leaf tissues were frozen in liquid nitrogen immediately after harvest, kept on dry ice for up to 20 h, and stored at −80°C until further processing. Additionally, previously characterized greenhouse-cultivated, clonally propagated genotypes were included in some experiments: the BC1 parent individual from Auby (individual 8.1) [17], [72], individuals 1.1/Lan 3.1 [8], [12] and 1.4/W504 [13] from Langelsheim, and an individual (9.1) of A. halleri ssp. gemmifera [38] (Table 1).
Genomic DNA was extracted using the DNeasy Plant Mini Kit (Qiagen, Venlo, The Netherlands) from 100 mg of frozen leaf material of each genotype. The thirteen amplicons designed to analyze sequence diversity (S1 to S13) comprised either non-coding (i.e., promoter, UTR and intron) or both non-coding and coding sequences, and were positioned within all three of the HMA4 gene copies and at loci of increasing distances upstream and downstream of HMA4 (Figure 1, Figure S1A, Tables S1 and S2). No additional amplicons could be designed in the repeat- and transposon-rich genomic regions between HMA4 genes [8]. Primer sequences for amplicons S2 to S12 were designed based on available A. halleri BAC sequences (Genbank accession numbers EU382073.1 and EU382072.1) (Table S1) [8]. Primer design for S1 and S13 was based on the Arabidopsis thaliana and Arabidopsis lyrata ssp. lyrata genome sequences [39], [73]. In A. thaliana and A. lyrata, S1 is located 116 and 198 kb upstream of S2, and S13 is located 113 kb and 2.47 Mbp downstream of S12, respectively. Amplicons comprising the 3′-portions of AhHMA4-1 (S5), AhHMA4-2 (S7) and AhHMA4-3 (S10) were simultaneously amplified in each of three independent PCRs using primer pairs that were not copy-specific (Table S1). In contrast, primers for S8 amplified only the 3′-end of AhHMA4-2 and additional downstream intergenic sequence, taking advantage of copy-specific sequence polymorphisms in the design of the reverse primer (see Figure S5A).
For PCR amplification, 2 µl of genomic DNA were used with GoTaq DNA polymerase (S1, S2, S11 and S13, Promega, Leiden, The Netherlands), Bio-X-Act Long DNA polymerase (S4, S6, Bioline/Gentaur, Brussels, Belgium) or a mix of both enzymes (S3, S9, S12 and S5/S7/S10), the respective primer pairs (0.5 µM each) (Table S1) and dNTPs (200 µM each) (Fermentas, St. Leon-Rot, Germany) in a final volume of 25 µl, the latter enzyme allowing more efficient amplification. PCR reactions were carried out as follows: 3 min at 95°C, followed by 30–32 cycles of 30 s at 95°C, 30 s at 58°C, 1 min per kb at 70–72°C, and a final extension step of 7 min at 70–72°C. PCR products were gel-purified and cloned into the pGEM-T easy vector (Promega, Leiden, The Netherlands) before transformation of E. coli DH5α.
In order to ensure with high probability that both alleles were sampled in heterozygous individuals through DNA sequencing, plasmid DNA was isolated from overnight cultures of at least eight independent bacterial colonies per amplicon and genotype, 20 clones for S6 and a total of 56 clones for S5/S7/S10, respectively, before sequencing of inserts by the Sanger method on an ABI 3730xl automated sequencer (Applied Biosystems, Darmstadt, Germany) using vector-specific and additional locus-specific primers when required (Table S1). For two individuals, 48 additional clones from two further independent PCRs were sequenced for S5/S7/S10 to resolve remaining sequence ambiguities.
For the S6 amplicon (corresponding to the promoter region of HMA4-2), a set of substantially divergent sequences was initially obtained, and, including these, a total of more than the expected maximum of two types of S6 sequences, corresponding to two alleles expected per individual at this single locus, were found in several A. halleri individuals. Using a combination of PCR, BAC end sequencing and DNA gel blot analyses of previously isolated A. halleri BACs harboring HMA4 and related sequences [8], the divergent set of sequences was unequivocally attributed to the promoter of AhHMA2, which was found to occur in tandem with AhHMA3 on a BAC clone, but this BAC did not contain any AhHMA4 coding sequence. AhHMA2 and AhHMA3 are orthologs of AtHMA2 and AtHMA3 that are located in tandem on chromosome 4 of A. thaliana whereas AtHMA4 is on chromosome 2. HMA4, HMA2 and HMA3 genes all encode divalent transition metal cation-transporting P1B-type ATPases [18], [74].
Sequence assembly was conducted with DNASTAR (DNASTAR Inc., Madison, USA). First, a consensus sequence was generated for each clone. Then, each consensus was compared to all other consensi from the same amplicon in a given individual and to all consensi of the same amplicon from all other individuals to i) correct Taq polymerase errors, ii) identify recombinant chimeras that resulted from template switches during PCR amplification [75] and iii) distinguish heterozygous from homozygous loci.
For the 3′-regions of the three HMA4 gene copies (S5, S7, S10, S8) more than 800 sequences were obtained in total. Among these, sequences were considered to be authentic when the same sequence was observed at least three times from one PCR reaction or in at least two independent PCRs of the same genotype. After removal of chimeras (which accounted for ca. 5% of the sequences), a total of 25 consensus sequences were retained for the 3′-regions of HMA4 gene copies. These consensi were assigned to the three HMA4 loci taking advantage of i) the copy-specific sequence information for AhHMA4-2 via the overlap between S7 and S8 for each individual (see Figure S5A), ii) position information available from two completely sequenced BACs [8], and iii) step-wise inference using a strictly parsimonious approach, similar to the strategy used to solve a SUDOKU in two times three double-blind independent replicates to ensure reproducibility.
After sequence assembly and alignment, DnaSP v5 [76] was used to calculate sequence diversity (π), Tajima's D [45], Fu and Li's D* and F* [44], and to conduct other statistical tests of molecular population genetics. MEGA v5 was used for phylogenetic analyses [77]. The ML trees shown throughout were constructed using a general time-reversible model. Rates among sites were assumed to be gamma-distributed with invariant sites, and 5 discrete categories of gamma were used. All sites were used. To estimate bootstrap support for the nodes, 1000 replicates were calculated. Neighbor joining methods yielded essentially the same results for tree branching orders. Genome sequence information from A. lyrata ssp. lyrata was used as a reference [39]. Network analyses for HMA4 genes and for S8 were conducted with TCS v1.21 using a connection limit of 95% [78]. Alignment gaps were re-coded with nucleotides to reflect the exact number of mutational steps between sequences in the respective sequence portion. AMOVA (Analysis of Molecular Variance) was carried out with Arlequin 3.5 [79] to compare the contribution of three hierarchical levels to genetic variance: among the geographic regions of the Thuringian Forest (A. halleri ssp. halleri), the Harz Mountains (A. halleri ssp. halleri), and Japan (A. halleri ssp. gemmifera), among geographic collection sites in each of these three regions, and within single geographic collection sites. A total of 1000 permutations were carried out for each locus, with equal weights of 1 for transitions and transversions, and a deletion weight of 0.
Quantitative PCR reactions were performed on 5 ng of genomic DNA in 384-well plates with an ABI Prism 7900HT system (Applied Biosystems, Brussels, Belgium) using MESA GREEN qPCR MasterMix (Eurogentec, Liège, Belgium). Mean reaction efficiencies were determined from all reactions for each amplicon (>270 reactions, Table S6) [80] and used to calculate relative gene copy number by normalization with the qBase software [81] using (i) multiple single-copy reference amplicons and (ii) A. thaliana genomic DNA (Col-0) as a calibrator [82]. Three single-copy reference amplicons were selected and designed at the 5′- and 3′-ends of the AhFRD3 gene [12] and in the S13 amplicon (this study), respectively. Their adequacy to normalize gene copy number in our experimental conditions was validated using the geNorm module in qBase (gene stability measure M = 0.309, pairwise variation CV = 0.121) [83].
Fresh cuttings of greenhouse-grown A. halleri and A. lyrata genotypes were cultivated hydroponically in 0.1× Hoagland solution for about 2 weeks [13]. After rooting, plants were transferred to pots with soil and further grown in a greenhouse with temperature settings of 22°C (day)/20°C (night) and a photoperiod of 16 h light and 8 h dark. Leaf material was harvested twice independently from the same individuals at an interval of eight weeks, immediately frozen in liquid nitrogen and stored at −80°C. A. thaliana and A. lyrata plants were grown from seeds as described, with harvest of leaves from 6-week-old plants, alongside harvest of A. halleri tissues [12]. Total RNA was extracted with TRIzol Reagent (Invitrogen, Karlsruhe, Germany), cDNA was synthesized from 1 µg of DNaseI-treated (Invitrogen) total RNA using oligo-dT and the SuperScript First-Strand Synthesis System (Invitrogen). Quantitative PCR was conducted in 96-well plates with a MyiQ Single Color Real-Time PCR Detection System (Bio-Rad, Munich, Germany) using SYBR Green qPCR Master Mix (Eurogentec, Cologne, Germany). A total of three technical repeats were run per cDNA and primer pair combination. Data were analyzed using iQ5 Optical System Software version 2.0 (Bio-Rad). Relative transcript levels of HMA4 were calculated by normalization to EF1α as a constitutively expressed reference gene [12]. Primers were as follows: AhHMA4 primers (5′- GCTGCAGCGATGAAAAACAAAC-3′ and 5′-TCCATACAACATCCCGAGGAAC-3′; amplification efficiency: 1.88); AlHMA4 primers (5′- TGAAGGTGGTGGTGATTGCA-3′ and 5′-CTCTCCACATTGACCAACTTTG-3′; amplification efficiency: 1.90). AtHMA4 and EF1α primers were described earlier [12].
Sequence data are available through EBI (http://www.ebi.ac.uk), accession nos. HE995813 to HE996227.
|
10.1371/journal.pcbi.1003228 | Sensitive Detection of Viral Transcripts in Human Tumor Transcriptomes | In excess of % of human cancer incidents have a viral cofactor. Epidemiological studies of idiopathic human cancers indicate that additional tumor viruses remain to be discovered. Recent advances in sequencing technology have enabled systematic screenings of human tumor transcriptomes for viral transcripts. However, technical problems such as low abundances of viral transcripts in large volumes of sequencing data, viral sequence divergence, and homology between viral and human factors significantly confound identification of tumor viruses. We have developed a novel computational approach for detecting viral transcripts in human cancers that takes the aforementioned confounding factors into account and is applicable to a wide variety of viruses and tumors. We apply the approach to conducting the first systematic search for viruses in neuroblastoma, the most common cancer in infancy. The diverse clinical progression of this disease as well as related epidemiological and virological findings are highly suggestive of a pathogenic cofactor. However, a viral etiology of neuroblastoma is currently contested. We mapped transcriptomes of neuroblastoma as well as positive and negative controls to the human and all known viral genomes in order to detect both known and unknown viruses. Analysis of controls, comparisons with related methods, and statistical estimates demonstrate the high sensitivity of our approach. Detailed investigation of putative viral transcripts within neuroblastoma samples did not provide evidence for the existence of any known human viruses. Likewise, de-novo assembly and analysis of chimeric transcripts did not result in expression signatures associated with novel human pathogens. While confounding factors such as sample dilution or viral clearance in progressed tumors may mask viral cofactors in the data, in principle, this is rendered less likely by the high sensitivity of our approach and the number of biological replicates analyzed. Therefore, our results suggest that frequent viral cofactors of metastatic neuroblastoma are unlikely.
| Many human cancers are caused by infections with tumor viruses and identification of these pathogens is considered a critical contribution to cancer prevention. Deep sequencing enables us to systematically investigate viral nucleotide signatures in order to either verify or exclude the existence of viruses in idiopathic human cancers. We have developed Virana, a novel computational approach for identifying tumor viruses in human cancers that is applicable to a wide variety of tumors and viruses. Virana firstly addresses several important biological confounding factors that may hinder successful detection of these pathogens. We applied our approach in the first systematic search for cancer-causing viruses in metastatic neuroblastoma, the most common form of cancer in infancy. Although the heterogeneous clinical progression of this disease as well as epidemiological and virological findings are suggestive of a pathogenic cofactor, the viral etiology of neuroblastoma is currently contested. We conducted an analysis of experimental controls, comparisons with related approaches, as well as statistical analyses in order to validate our method. In spite of the high sensitivity of our approach, analyses of neuroblastoma transcriptomes did not provide evidence for the existence of any known or unknown human viruses. Our results therefore suggest that frequent viral cofactors of metastatic neuroblastoma are unlikely.
| To date, pathogenic agents are known to be causally related to 20% of human cancer cases [1] and significantly affect the global health burden of this disease [2]. The majority of these agents comprise oncogenic viruses such as human papilloma virus (HPV), Epstein-Barr virus (EBV), hepatitis B virus (HBV), and hepatitis C virus (HCV) [3]. Characterizing the oncogenic potential of viral pathogens has important consequences for prevention, diagnosis, and treatment of malignant neoplasms [4], [5]. Tumor viruses in particular have received renewed attention in the context of recent global efforts to characterize the etiology of cancer [6], [7]. Consequently, viral cofactors for several idiopathic cancers are currently investigated [8] and epidemiological indicators suggest that additional human tumor viruses remain to be discovered [9].
Neuroblastoma is a heterogeneous embryonal tumor [10], [11] that is accountable for 15% of deaths caused by malignant conditions in children [12]. The disease is associated with an exceptionally low median age of presentation of months [13] and is often diagnosed in utero. Metastatic neuroblastoma has two biologically divergent subtypes. Stage 4S is characterized by an age of presentation between in utero and months, metastases confined to liver, skin, lymph nodes and bone marrow, and its ability to regress spontaneously [14], [15]. In contrast, stage 4 tumors are presented at any age, demonstrate high infiltration rates in bone marrow and bone, and are most often progressive [10], [16]. While genes related to neuronal differentiation have been described to be upregulated in stage 4S in comparison to stage 4 neuroblastoma, thereby indicating distinct levels of neuronal differentiation [17], little is currently known about the differences between molecular etiologies of stage 4 and stage 4S neuroblastoma.
The variation of clinical outcomes between neuroblastoma subtypes indicates distinct genetic and environmental factors affecting the development of this malignancy. Interestingly, the early onset of the disease overlaps with periods of high susceptibility to viral infections and is reminiscent of acute lymphoblastic leukemia – another pediatric tumor with uncertain etiology for which an infective cofactor has long been suspected [18]. Furthermore, epidemiological studies have associated reduced neuroblastoma risk with immunologic indicators such as previous childhood infections, day care attendance, and breast feeding [19], [20] that are suggestive of an infective cofactor [21]. While transforming polyomaviruses such as JCV and BKV were previously identified within neuroblastoma samples and other pediatric embryonal tumors [22]–[24], newer studies seem to render these associations inconclusive [25]. Therefore, the role of pathogenic cofactors of neuroblastoma oncogenesis remains unresolved.
In general, the search for suspected viral cofactors of idiopathic diseases requires systematic screening of human tissues for viral biomarkers such as virus-derived nucleotide sequences. Unfortunately, viruses are of polyphyletic origin and thus lack common universal marker genes as they are frequently exploited in metagenomics studies targeting cellular microorganisms. Consequently, it is not currently possible to specifically PCR-amplify viral nucleotide sequences within a given tissue without prior information about the infective agent being sought [26]. As a result, several systematic assays for pathogen detection have been developed that do not rely on targeted PCR-amplification of viral factors [27] and were employed to identify Kaposi's sarcoma-associated herpes virus (KSHV) as a human tumor virus [28]. These systematic approaches were recently supplemented by sensitive deep sequencing technologies [27]. These technologies were recently applied to exclude several cancer-virus associations based on negative evidence [29], [30] and aided in the identification of MCPyV, a human polyomavirus, as a cofactor of Merkel cell carcinoma [31].
Deep sequencing technologies have enabled detection of both known and novel viruses with unprecedented sensitivity [32]. However, the large numbers of sequence fragments (“reads”) generated by these methods necessitate data reduction approaches for filtering and condensing the list of putative viral transcripts. Two such approaches are currently represented in the literature: digital transcript subtraction that discards human sequence homologs from the sequence data and considers the remaining transcripts as potential viral signatures [30], [31], [33]–[39], and de-novo sequence assembly that aims to reconstruct whole viral genomes from overlapping reads [40]–[43]. Recently, variants of these of two approaches have been implemented in several computational pipelines such as PathSeq [44], RINS [45], and CaPSID [46].
Identification of tumor viruses in particular poses several important challenges to existing computational pipelines. Confounding factors such as loss of viral genetic material from progressed tumors as well as limited replication competence or latent replication strategies often result in low or selective transcription of tumor viruses [5]. In addition, viral oncogenes homologous to human factors and chimeric transcripts originating from proviral insertion sites may share significant sequence similarity with human transcripts [47], thus making unequivocal identification of viral factors difficult. Last, high rates of viral sequence divergence from (dsDNA viruses) up to (ssRNA viruses) substitutions per site and year [48], [49] hinder recognition of known viruses based on known reference sequences.
We have developed Virana, a novel computational approach specifically tailored to detecting low-abundance transcripts that diverge from known viral reference sequences or share significant sequence homology with human factors. In particular, our method maps sequence reads to a combined reference database comprising the human genome and all known viral reference sequences. The approach is configured to allow for high mismatch rates and mappings to multiple reference sequences (‘multimaps’). By using this combined and sensitive mapping strategy, our approach is especially well suited for detecting human-viral chimeric transcripts and viruses diverging from known references. In contrast to existing subtractive approaches for viral transcript discovery, our method abstains from discarding reads homologous to the human genome from further analysis. Instead, Virana exploits multimaps to assign sequence reads to a homologous context comprising human reference transcripts and viral reference genomes. These homologous regions retain the full, unfiltered information contained in the raw sequence data while also being amenable to further analyses by multiple sequence alignments, human-viral phylogenies, and orthogonal taxonomic annotations, thus greatly aiding in the interpretation of the results.
We applied our novel approach on an overall number of deep sequencing transcriptomes of stage 4 and stage 4S metastatic neuroblastoma in order to identify putative viral cofactors associated with this idiopathic disease.
Primary neuroblastoma samples from stage 4 (progressive) patients () and stage 4S (regressive) patients ( were obtained prior to treatment from the central neuroblastoma tumor bank at the University Hospital of Cologne, Germany. None of the tumors harbored amplification of the MYCN proto-oncogene as determined by two independent laboratories for each case by fluorescence in situ hybridisation (FISH) and Southern blot [50]. Only neuroblastoma samples with a tumor cell content of above % as assessed by a pathologist were selected for deep sequencing. Integrity of RNA was evaluated using the Bioanalyzer (Agilent Technologies) and only samples with an RNA integrity number of at least were considered for further processing. Quality of all neuroblastoma samples and related deep sequencing data was additionally confirmed by an orthogonal computational analysis focusing on human gene expression in the context of differential splicing [51].
All patients were enrolled in the German Neuroblastoma trials with informed consent. In order to validate our approach we additionally employed a positive control panel consisting of tumors with known viral cofactors. An EBV-positive B-cell-lymphoma (BCL) was received from the Pediatric Oncology and Hematology Department of the Hannover Medical School. Deep-sequencing reads obtained from full transcriptome libraries of two HPV18-positive HeLa samples (HeLa) and a HPV16-positive primary cervical squamous cell carcinoma (ceSCC) were downloaded from the Short Read Archive (SRA) and preprocessed as specified in the original publication [30]. Transcriptome data of a HBV-positive hepatocellular carcinoma (HCC) HKCI-5 cell line with confirmed HBV integration events was downloaded from the SRA based on information in the original publication [52]. A negative control panel consisting of a normal brain transcriptome generated as part of the Illumina BodyMap 2.0 project was obtained from the SRA at run accession number ERR030882.
mRNA libraries of the EBV-positive B-cell lymphoma and neuroblastomas were prepared following the Illumina RNA Sample Preparation Kit and Guide (Part # Rev. A). For each sample, g high-quality total RNA was processed for mRNA purification, chemical fragmentation, first strand synthesis, second strand synthesis, end repair, ′-end adenylation, adapter ligation, and PCR amplification. Validated libraries underwent gel size selection and final paired-end sequencing with an effective read length of bp on the Illumina Genome Analyzer IIx following Illumina standard protocols. Additionally, libraries for two of the neuroblastoma samples were generated using the same protocols and sequenced with an effective paired-end read length of bp on a Illumina HiSeq . All libraries had insert size distributions approximating bp, bp as later confirmed by read mapping. The data were filtered according to signal purity by the Illumina Realtime Analysis (RTA) software.
In this study we employ simulated sequencing data from three viral genomes that are homologous to human factors. Reads originating from the ABL1-homologue of the Abelson murine leukemia virus (A-MuLV, GI:9626953, positions ), from the the gag region of HERVK22I (obtained from Repbase [53], positions ), and from Bo17, a GCNT3-homolog of the bovine herpesvirus 4 (BoHV-4, GI:13095578, positions ) were generated in silico by dwgsim, a read simulator based on wgsim [54]. In addition, we produced simulated chimeric transcripts by fusing each of the aforementioned sequence regions to the human TP53 gene, a known proto-oncogene (UCSC build hg19, GRCh37, chr17, positions 7572926–7579569). These artificial fusion transcripts were generated using Fusim [55] based on TP53 exon models obtained from the UCSC refGene database [56]. Fusion transcripts were then used as templates for generating simulated data sets with dwgsim. In all cases, dwgsim was applied using the default empirical error model. Paired-end read lengths and insert size distributions were chosen according to the neuroblastoma sequencing data (see above). Additional simulated sequencing data generated by a related publication were analyzed as described in Section “Estimation of read mapping sensitivity”.
Sample panels containing neuroblastoma transcriptomes sequenced at bp and bp effective read lengths are denoted as NB1 and NB2, respectively. While the NB1 panel contains seven transcriptomes of neuroblastoma stages 4 and 4S each, the NB2 panel contains one sample of stages 4 and 4S each (see Table 1). Positive control panels of human cancer transcriptomes with known viral cofactors (BCL, HeLa, ceSCC, and HCC) are denoted as POS. The negative control panel consisting of a normal human brain transcriptome is denoted as NEG.
The current assembly of the human reference genome (UCSC build hg19, GRCh37) as well as corresponding refGene splice-site annotations were obtained from UCSC. Splice variant annotations and cDNA sequences for the human genome were downloaded from Ensembl [57]. A set of all available complete viral reference genomes and their taxonomic lineages were obtained from NCBI via the E-utilities web service [58] and the database query: “Viruses[Organism] AND srcdb_refseq[PROP] NOT cellular organisms [ORGN]”. In addition, we obtained consensus reference sequences for all human endogenous retroviruses (HERV-K/HML-2) represented in Repbase (Primate HERV, HERVK11DI, HERVK11I, HERVK13I, HERVK22I, HERVK3I, HERVK9I, HERVKC4)) [53]. All reference genomes were combined into a single human-viral reference database for Virana. Since RINS and CaPSID cannot use such a combined database, human and viral reference sequences were collected within two separate databases for these approaches.
Paired-end reads from the neuroblastoma panels and positive control panels were quality-controlled with an in-house sequence analysis framework in order to identify sample contamination, adapter contamination, and batch effects. After quality control, the sequence data consisted of Gbp (NB1), Gbp (NB2), Gbp (POS), and Gbp (NEG) of sequence reads, respectively (see Table 1).
All data were mapped against a combined human-viral reference database with the splicing-aware and gapped read mapper STAR [59] in paired-end mode. While Virana considers the read mapper to be a replaceable component, in principle, we decided to employ STAR due to its mapping speed, high sensitivity settings, and its consideration of putative chimeric transcripts. We configured the mapper for high sensitivity by following recommendations of the author of STAR (personal communication). In particular, we set the rate of acceptable mismatches to times the length of each read and the seedSearchStartLmax and winAnchorMultimapNmax parameters to and , respectively. The minimum length of chimeric segments (chimSegmentMin) was reduced to in order to detect fusion transcripts at short read lengths. Known splice sites from splice annotations of the human reference genome as well as canonical splice sites were considered in the mapping. For each read, multiple mapping locations with alignment score distances of up to ranks relative to the best score were permitted (‘multimaps’). Read alignments were stored in standardized BAM files. STAR supports detection of chimeric transcripts by reporting discordant read pairs whose ends map to different chromosomes. These discordant read pairs were employed in further analyses as detailed in the next section.
In order to identify putative new viral transcripts, read pairs with at least one unmapped read end were extracted from BAM files by the Samtools suite [54] and assembled into longer contigs by the de-novo transcriptome assemblers Trinity [60] and Oases [61] using default parameters. Oases was configured for using different k-mer values in order to facilitate reconstruction of low-abundance viral transcripts. Contigs of length less than bp were considered to be spurious assemblies and excluded from further processing.
Virana supports detection of human-viral chimeric transcripts in two different manners. First, the read mapper employed in our study is able to partially align reads that contain a human-viral chimeric breakpoint to multiple reference sequences. Consequently, these partially aligned reads can be detected by Virana within the generic analysis of homologous regions (see below). The second, more sensitive approach to detecting chimeric transcripts is based on paired-end read information. Since the STAR mapper assigns reads to a combined reference database comprising both human and viral reference sequences, ends of paired-end reads whose inserts span the breakpoint of a chimeric transcript will be aligned to different reference sequences. These discordant read pairs are reported by STAR during read mapping (see above) and can further be filtered by mismatch score or sequence complexity in order to yield a high-confidence list of chimeric transcripts.
A distinguishing feature of Virana is its ability to automatically reconstruct the homologous context of reads that map to both viral and human reference sequences. This homologous context is constructed in four steps:
Consensus sequences can be further processed by phylogenetic analyses. For generating phylogenies, Virana employs the software PhyML [64] following the maximum likelihood approach and using default parameters recommended by the HIV sequence database (http://hiv.lanl.gov, GTR model of nucleotide substitution, transition/transversion ratio: 4, gamma shape parameter: 1, number of substation rate categories: 4, approximate Likelihood Ratio Test (aLRT) using SH-like supports where applicable). We note that the topology of the phylogenetic trees constructed in this manner is stable with regard to the model choice; while more complex model parameters may yield better likelihoods in some instances, these differences do not influence interpretation of our results.
In this study, we additionally compare consensus sequences of aligned HOGs as well as de-novo assembled sequence contigs to nucleotide (NCBI NT) and protein (NCBI NR) reference archives in order to assign transcripts to a taxonomic origin. To this end, we employ several BLAST [58] search strategies (BLASTN, BLASTX, and TBLASTX) with sensitive word sizes (, , and , respectively). TBLASTX bypasses synonymous mutations during similarity search and is particularly suited for detecting functionally conserved homologs. This approach is therefore recommended for discovering remote similarities [65] and is widely used in environmental metagenomics [66]. A permissive E-value threshold of is used for all comparisons in order to reduce the possibility of missing true viral hits. For each query transcript and search strategy, the three highest-scoring reference sequences are extracted from the BLAST results. Subsequently, descriptions, taxonomic information, and available gene annotations for high-scoring reference hits are pooled and query transcripts are assigned a putative viral, human, or ambiguous origin based on the pooled information. In order to limit the search space of the computationally intensive TBLASTX procedure, we constrain the allowed taxonomic origin of reference sequences to only viral (NCBI taxon ID ) or human (NCBI taxon ID ) hits while excluding artificial sequences (NCBI taxon ID ) using the NCBI database query “(((txid10239 [ORGN]) OR (txid9606 [ORGN]) OR (human [ORGN])) NOT (txid81077 [ORGN]))”.
We quantify the ability of our novel method Virana and the related methods RINS [45] and CaPSID [46] at detecting diverged viral transcripts among human sequence data by employing a recently published validation data set [46]. This data set consists of a negative control background set of reads simulated from the human reference genome that is spiked with four sets of reads simulated from viral reference genomes. Nucleotide positions within reads of each of the four viral spike-in data sets are mutated randomly independently and uniformly with a set-specific probability before being merged with the background data set. The set of viral reference sequences represents different viral families that infect plants (Cherry green ring mottle virus, Cestrum yellow leaf curling virus, Elm mottle virus, East African cassava mosaic virus), birds (Gallid herpesvirus 1), insects (Cotesia congregata bracovirus), bacteria (Guinea pig Chlamydia phage), amphibians (Frog adenovirus 1), and mammals (Rat coronavirus Parker, Banna virus).
All five data sets (non-spiked human negative control and four human-viral spike-in sets) are analyzed by Virana, RINS, and CaPSID using identical reference sequences as described in Section “Reference genomes”. Sensitivity (fraction of correctly identified viral reads among all viral reads) and specificity ( fraction of falsely identified human reads among all human reads) of viral read detection are determined for each method and data set. Analyses are performed with either default parameters (Virana), parameters published in the original validation data set (CaPSID), or settings adapted by us in order to maximize sensitivity (RINS: minimal contig length decreased to , read lengths and insert size distributions according to input data).
Since all methods map to the same complete viral reference set, reads from a particular viral genome of the validation data set may be distributed across several closely related reference genomes, all of which may be considered valid mappings. For this reason, we added post-processing steps to CaPSID and RINS and performed this validation on the level of viral taxonomic families rather than on the level of single viral species. We note, however, that results of all tested methods including Virana retain information on single viral species throughout the analysis. In particular, sensitivity and specificity of the methods change only minimally if data is analyzed on the single species level.
Analysis of the human-viral homologous regions and chimeric transcripts based on simulated read data (see Section “Simulated sequencing data”) was conducted by configuring CaPSID, RINS, and Virana analogous to the previous section. For the validation of fusion transcript detection, the number of true positives is set to the number of all reads originating from the human-viral fusion transcript. Since all detection methods in this validation are configured to only report reads mapping to the viral part of the fusion transcript, sensitivity estimates are scaled down equally for all methods in this particular validation. Analysis of discordant read ends in order to detect the origins of chimeric transcripts was performed as described before (see Section “Detection of chimeric transcripts”).
Expanding on related work [34], [35], we quantify the theoretical sensitivity of Virana by estimating the number of viral transcripts per cell that are required for achieving a certain minimal sequencing coverage at a probability of at least 95%. Based on human genome annotations obtained from UCSC, we determined an average length of human coding sequences (CDS) of bp. By conservatively assuming that an idealized cell contains mRNAs [34] of average length fragmented at bp as a result of library preparation, an expected number of cDNA fragments are generated per cell. For a given viral transcript of length and a viral transcript abundance per cell, we expect a number of viral transcript fragments. Assuming a theoretical, unbiased sequencing process, the probability of sequencing a viral transcript fragment among the overall transcript fragments is . Given a single-end read length of , a number reads are required to achieve a sequence coverage of that viral transcript. The probability of observing at least reads during sequencing with a sequencing depth is specified by the cumulative binomial distribution function with parameters , and . Due to numerical instabilities of computing the cumulative binomial distribution for large values , we exploit the Central Limit Theorem and estimate by the Camp-Paulson normal approximation to the binomial distribution. This approach has a negligible approximation error of , where [67]. Our approach further depends on successfully reconstructed homologous regions, each requiring an empirically determined minimum number of transcripts separated by no more than base pairs.
Although the probability of a homologous region being successfully constructed from viral transcripts at a given sequence coverage can be derived analytically for a special case [68], this solution neither considers edge effects occurring for small transcripts nor takes into account the distribution of insert sizes of paired-end reads. We therefore approach the problem empirically by in silico simulation of paired-end reads that are assigned randomly independently and uniformly to transcripts of different lengths and at varying coverages. This simulation process addresses the aforementioned confounding factors by considering transcript boundaries and sampling insert sizes from a normal distribution parametrized according to neuroblastoma sequence data employed in this study (see Section “Library preparation and sequencing”). An mean estimator for and its standard error were derived by averaging the success rates of homologous region constructions across simulations for each transcript length, read length, region linkage, and read coverage.
All sequence data generated in this study are publicly available in the European Nucleotide Archive (ENA) at study accession number PRJEB4441. Software implementations of our method and all validation procedures are available at http://mpi-inf.mpg.de/∼sven/virana.
This study presents a novel approach to detecting viral transcripts in human tumor transcriptomes. In contrast to related approaches such as RINS and CaPSID that rely on subtracting reads homologous to human transcripts from the analysis, our novel method Virana assigns sequence reads to a combined human-viral reference database without discarding homology information (see Figure 1). By employing a particularly fast and sensitive read mapper, Virana gains sensitivity at discovering highly divergent and chimeric viral transcripts. In addition, this configuration allows for exploitation of multimaps (e.g., sequence reads mapping to several reference genomes with varying mismatch rates) to discover the homologous context of sequence reads with regard to viral and human reference sequences. Last, Virana employs chimeric alignments as well as de-novo assembly of unmapped sequence reads followed by taxonomic annotation in order to discover proviral integration events and novel viruses, respectively.
In order to compare Virana and the two subtractive approaches CaPSID and RINS in a controlled environment we rely on a previously published simulated data set consisting of a negative control data set free of viral reads, here denoted as background set. The background set is used to construct four additional validation data sets spiked with viral reads at increasing rates of sequence divergence (0%, 5%, 10%, 25%, see Materials and Methods). Performance is quantified in terms of sensitivity and specificity (see Materials and Methods). Applying all three viral detection methods on the validation data sets reveals comparatively high rates of correctly detected viral reads for CaPSID and RINS at low sequence divergences between 0% and 5%. Specifically, the two subtractive methods achieve fold higher sensitivities compared to Virana (sensitivities of versus for subtractive approaches and Virana, respectively, see Figure 2). In contrast, Virana substantially surpasses subtractive approaches at higher rates of viral sequence divergence (10–25%), offering comparatively stable sensitivities between -fold and -fold higher than Capsid and RINS, respectively (sensitivities of versus for subtractive approaches and Virana, respectively, see Figure 2, left panel). Notably, while subtractive approaches fail to identify 20–90% of viruses in settings of high sequence divergence, Virana is the only approach able to reliably detect the full set of viruses in all validation scenarios (see Figure 2, right panel). As a result of Virana's ability to detect human-viral transcript homologs, reads originating from several human endogenous retroviruses (HERVs) that are part of the human reference genome but technically also belong to the viral family Retroviridae are detected in validation data at all levels of sequence divergence. Since the detected HERV reads originate from the human rather than from the viral part of the validation data, these reads classified as false positive (FP) hits for the purpose of this validation. As a result of this artifact, Virana exhibits a slightly lowered specificity compared to subtractive approaches (0.99985 versus 1.0 for Virana and CaPSID/RINS, respectively). However, we note that HERV reads are correctly classified by Virana during homologous region construction and by optional BLAST-based taxonomic annotation. These reads can therefore be safely and automatically ignored in subsequent analyses if HERV expression is of no interest to the researcher.
In spite of the involved construction process of homologous regions, Virana is fastest among the three viral detection approaches, requiring only about half an hour per sample analyzed. In contrast, RINS and CaPSID require two to times longer per sample, respectively (see Figure 3). Interestingly, the majority of time spend by CaPSID is lost on subtraction, indicating that this step is a limiting factor of subtractive approaches. We note than reported times are based on analyses using a single compute core. Since all evaluated methods benefit from multithreading, dedicating additional compute cores to the analysis allows for further reduction in processing time.
Having established Virana's ability to detect reads sampled at comparatively high coverage from viral genomes with low or no human-viral sequence similarity, we next test the sensitivity of the viral detection methods in a more challenging scenario involving gene regions of animal viruses that have close human homologs and are sampled at low sequencing coverages. Three such human-viral homologs are used in the analysis: V-ABL of the acutely transforming retrovirus A-MuLV, Bo17 of herpesvirus BoHV-4 (a model virus for oncogenic gammaherpesviruses such as EBV and KSHV and implied in several animal cancers [69]) and gag of HERV-K(HML2)22I, a class of human endogenous retroviruses associated with some forms of breast cancer [70]). Validation is based on simulated sequencing data and split into two scenarios (see Materials and Methods for details). Within the first scenario, simulated sequencing reads are sampled directly from human-viral homologs while in the second scenario reads are generated from artificial fusion transcripts that each involve one of the three homologs fused to the human TP53 proto-oncogene. The resulting human-viral fusion transcripts mimic transcriptional signals indicating retroviral integration or homologous recombination of viral DNA next to a human gene which may result in activation of the latter by insertional mutagenesis.
We apply the viral detection methods Virana, CaPSID, and RINS on these two validation data sets in order to evaluate sensitivity at detecting viral genes that are similar to human factors either due to natural sequence homology or due to gene fusions. Performance is quantified by detection sensitivity, specificity, as well as by the absolute number of reads correctly detected. While all methods performed at a perfect specificity of , only Virana detects viral transcripts at all coverages and with two to three-fold higher sensitivities compared to competing methods (Figure 4). In particular, sequence reads originating from endogenous retroviruses were almost always subtracted from the analysis by RINS and CaPSID. In addition, RINS seemed to be confounded by low sequencing coverage, a fact most probably resulting from its heavy reliance on de-novo transcript assembly. Subsequent analysis of discordantly mapped read pairs by Virana (see Materials and Methods) correctly identified the TP53 gene as fusion partner of both V-ABL and Bo17, indicating that detection of human-viral chimeras is reliable even at low twofold coverage. Due to the repeat nature of the HERV-K sequence in the human genome and the resulting re-occurrence of HERV-K homologs at multiple loci in the human reference it was not possible to unambiguously identify the fusion partner of the HERV-K gag gene.
Due to a variety of factors (see Discussion) human tumor viruses often replicate at very low levels within the infected cell. Determining the required sequencing depth for detecting viral transcripts present at specific cellular abundances is therefore crucial for planning transcriptome experiments designed to identify tumor viruses. Based on statistical arguments and average mRNA sizes (see Materials and Methods), we inferred the minimal abundances of viral transcripts required in an average cell required for detection depending (1) on the length of the transcript being sought and (2) on the sequencing depth employed in the experiment. Here we report results for an average viral cDNA-transcript ( bp), an average viral transcript region analyzed in the validation of human-viral homologs (Bo17 and vABL, bp, see previous section), an average length human CDS ( bp), and the genome size of a small tumor virus (A-MuLV, bp). Based on these estimates and given an average sequencing depth as employed in the NB1 analysis panel, Virana requires a minimum twofold sequence coverage of an average viral cDNA transcript in order to detect the transcript within a homologous region with % probability (Figure 5, upper left quadrant, dashed blue vertical line). This sequence coverage is produced with % probability if at least one viral transcript is present per cell, on average (Figure 6, upper left quadrant, dashed blue vertical line). The number of viral transcripts per cell required for detection is inversely related to transcript length and sequencing depth, in principle: at a transcript length corresponding to a small viral genome ( bp) and a per-sample sequencing depth of % of the sequencing depth generated in the NB1 panel, a transcript coverage of and at least viral transcripts per cell are required for reliable detection (Figure 6, upper right panel, dotted black vertical line).
In order to evaluate Virana on experimental data we conducted an analysis of several positive and negative control samples with a cumulative size of Gbp. The negative control sequencing data originates from a normal brain transcriptome that is suitable as a control for neuroblastoma data. Positive controls span a range of cancer transcriptomes that are associated with several viral cofactors such as a hepatocellular carcinoma (HCC) cell line with proviral integration of Hepatitis B virus, a cervical squamous cell carcinoma (ceSCC) and two HeLa cell line samples with associated human papillomavirus (HPV), and a Ebstein-Barr virus (EBV) positive B-cell lymphoma (BCL).
As displayed in Figure 7 (upper part), analysis of the brain negative control sample demonstrates that viral transcription is ubiquitous even in normal (non-cancerous) samples. Specifically, several bacteriophages of the taxonomic families Microvirodae, Myoviridae, Podoviridae, and Siphovoridae indicate sample contamination with bacteria as well as technical spike-ins (http://res.illumina.com/documents/products/technotes/technote_phixcontrolv3.pdf). Remarkably, the Coliphage phi-X174 genome of the family Microvirodae could be fully assembled by Virana's homologous region construction, yielding a single fragment of 99% sequence identity and 100% coverage compared to the phi-x174 reference genome. In addition, several retroviral and flaviviral hits at low abundances of reads per million reads mapped (RPMM) highlight human factors such as HERV-Ks (endogenous retroviruses) as well as human proto-oncogenes SRC/ABL and DNAJC14/RP11 that have close homologs in the viral families Retroviridae and Flaviviridae, respectively. The taxonomic ambiguity of these regions is automatically identified during Virana's homologous region construction and confirmed by optional BLAST-based annotation compared to NCBI nt and nr databases (as indicated by thinner bars in Figure 7).
Analysis of positive control samples resulted in homologous regions (HORs)spanning five viral families (see Figure 7, lower part). Viral cofactors associated with each of the cancer samples are correctly recovered at a high dynamic range of read abundances between RPMM (HCC with integrated HBV provirus) and RPMM (HeLa cell line associated with HPV18). In addition, several viral fragments were successfully reconstructed within HORs of the positive control samples, such as a bp long EBV segment containing latency-associated factors EBNA 3b, 3c, and 4a (80% sequence identity with the wild type genome) as well as a bp long HBV fragment containing the oncogenic HBV-X gene (98% sequence identity compared with Hepatitis B virus isolate HK1476). Similar to results on the negative control brain sample, several HORs with lower abundances assigned to the taxonomic families Retroviridae and Flaviviridae represent human-viral sequence homologies that are automatically flagged to be of ambiguous taxonomic status by Virana.
Interestingly, the HCC sample was also investigated in recent work focusing on detecting viral integration events [52]. In this recent study, the authors confirmed one integration event by Sanger sequencing while alluding to two additional events still awaiting experimental validation. By analyzing discordantly mapped read ends, Virana could correctly identify all three HBV fusion events involving human genes TRRAP (11 read pairs), ZNF48 (11 read pairs), and PLB1 (6 read pairs) as part of the primary mapping procedure.
Deep-sequencing of neuroblastoma samples on two sequencing platforms yielded Gbp (NB1) and Gbp (NB2) of mapped read pairs (including multimaps), respectively (see Table 2). While samples were sequenced independently and marked with unique identifiers to allow for sample tracking at each step of the analysis, reads from each sample panel and each tumor stage (4 or 4S) were pooled for analysis. Processing the pooled sample panels with Virana resulted in homologous regions representing four viral families (see Figure 8). All HORs were associated with low relative read abundances of RPMM compared to confirmed viral signatures of experimental positive controls ( RPMM, see Figure 7). Several homologous regions assigned to bacteriophage viral families Baculoviridae and Myoviridae are attributable to sample contamination.
Reads assigned to viral families Retroviridae and Flaviviridae were determined to originate from either endogenous elements (HERVs) or from human proto-oncogenes that have close homologs in pestiviruses and acutely transforming retroviruses. HORs associated with these viral families were automatically assigned human or ambiguous taxonomic origin by Virana, as indicated by narrower bars in Figure 8. We undertook manual investigation of homologous relationships within each ambiguous HOR by analyzing multiple sequence alignments and phylogenetic trees of the respective regions. These analyses revealed unambiguous clusterings of neuroblastoma sequence reads near human or endogenous factors in all cases (see Figure 9 for an example phylogeny).
No significant differences in viral expression signatures between neuroblastoma 4 and 4S stages could be detected except for HERV-K endogenous retroviruses which display higher abundances in stage 4S (NB1: 56 RPMM, NB2: 28 RPMM) than in stage 4 (NB1: 41 RPMM, NB2: 15 RPMM) neuroblastomas. All reads assigned to homologous regions were further analyzed for evidence of chimeric transcription (see Materials and Methods). While several read pairs with putative chimeric mappings could be identified, all viral chimeric read ends were clustered within low-complexity regions of the viral genomes. Analyses revealed that these putative chimeric mappings represent sequencing errors and low-complexity templates that non-specifically attracted reads of similarly low sequence complexity. No cluster of chimeric reads located at a specifically viral genome location and representing a human-viral breakpoint could be identified.
In order to identify transcripts of novel viruses that do not map to known references, we generated de-novo transcriptome assemblies of all unmapped reads. We applied the two de Bruijn graph based assembly methods Oases[61] and Trinity[60] that demonstrated best-in-class performance in recent evaluations [71] on sequencing data of the NB2 panel. This sequencing data is especially amenable to assembly due to its long read length (see Table 1). Assembly resulted in and reconstructed neuroblastoma 4S contigs for Oases, and Trinity, respectively (see Figure 10). Assembly of the neuroblastoma 4 sample yielded and contigs from the same methods. Results of Oases and Trinity assemblies are comparable in terms of contig length. All contigs were subjected to taxonomic annotation using high-sensitivity TBLASTX annotation based on human and viral content of the NCBI nt and nr databases (see Materials and Methods). Overall, contigs ( of contigs of any specific assembly) were identified to be of putative viral origin. contigs were assigned to bacteriophage references and excluded from further analysis. Based on searches against the full NCBI nr and nt databases followed by manual inspection, all remaining contigs were determined to display higher similarities to bacterial or human sequences than to any viral reference.
Neuroblastoma is a pediatric tumor of the sympathetic nervous system that represents the most common form of cancer in infancy. It is characterized by a striking diversity in biology and clinical behaviour of its subtypes. This heterogeneity as well as supporting epidemiological findings are highly suggestive of infectious cofactors involved in genesis and maintenance of the disease [19], [20]. While several studies utilizing technologies with lower sensitivity compared to our approach have identified human polyomaviruses in neuroblastoma and pediatric embryonal tumors [22]–[24], newer investigations seem to render these associations inconclusive [25]. However, viral commensals of the families polyomaviridae and adenoviridae are indeed suspected to acquire rare transforming properties as a consequence of viral latency or defective replication [72] and to encode oncogenes [73], [74] whose carcinogenic potential in human is currently investigated [8], [75]. We undertook the first systematic search for known and unknown viruses in transcriptomes of metastatic neuroblastoma by analyzing deep sequencing RNA-Seq data of metastatic neuroblastomas from two tumor stages as well as positive and negative experimental controls.
Several high-throughput methods for detecting viral sequence reads among human RNA-Seq data have been developed. Among these methods, PathSeq, CaPSID and RINS are most prominent due to their design as reusable computational pipelines. In this study we selected CaPSID and RINS due to their high performance and public availability and compared their detection performance with that of our novel method Virana. Both CaPSID and RINS follow a subtractive approach, e.g. they separately map input data to viral and human reference sequences and subtract viral read mappings that are similar to the human genome from the analysis. While CaPSID is conceptualised as a generalised framework that supports the subtraction process by means of a database and a web server, RINS features an integrated pipeline that splits input reads into shorter fragments in order to increase mapping sensitivity, followed by transcriptome assembly of putative viral reads into full length transcripts.
Both RNA and DNA viruses may share considerable sequence homology to human factors due to reasons such as lateral gene transfer, oncogene capture, ancestral endogenization, or insertional mutagenesis leading to chimeric transcripts [47]. Such homologous transcripts may display human-viral sequence similarities of 86% (Bovine Herpes virus) and up to 92% (acutely transforming retroviruses). Subtractive approaches silently discard these transcript from the analysis due to their similarity to the human reference genome. In contrast, our novel method Virana follows a radically different approach. Instead of separate mapping to viral and human reference database followed by digital subtraction, Virana undertakes a particularly sensitive read mapping to a combined set of human and viral references. By allowing for multimaps, this mapping strategy facilitates discovery of viral transcripts regardless of their similarity to human factors. Apart from being conceptually simpler by relying on only one mapping step and discarding the subtraction procedure that is both possibly erroneous and computationally costly, this approach empowers the mapper to make informed decisions about relative alignment quality by weighing different human and viral reference positions against each other. As a direct consequence of this increased mapping quality, paired-end reads can be mapped across human and viral references, allowing for detection of human-viral chimeric transcription and proviral integration events.
We quantitatively validated Virana's approach both in settings involving simulated reads as well as in real-world scenarios involving experimental positive and negative controls. In these validations, Virana displays significantly higher detection sensitivities than competing approaches especially at high rates of viral sequence divergence exceeding % that are common for tumor viruses [76]–[78]. As a consequence, Virana was the only method able to detect all viral families independent of sequence divergence in the validation data set. In spite of the additional processing undertaken by our method, Virana features between and two and three times faster execution speeds compared to related methods.
Interestingly, viral reads analyzed in the sequence divergence validation originate from a broad array of viral species, only two of which infect mammalian hosts and none of which display significant human-viral sequence homology. As a consequence, this validation favors subtractive approaches by reducing the danger of erroneous subtraction of viral reads that are similar to the human genome. In addition, the sequence divergence validation contained reads sampled at high coverage. However, transcripts of tumor viruses are often expressed at only low cellular abundances and are thus expected to have low sequence coverage. We therefore next validated the ability of viral detection approaches to detect viral transcripts homologous to human factors at varying levels of sequence coverage. Virana, by virtue of not relying on digital subtraction, demonstrated superior sensitivity at this validation both in settings of natural sequence homology as well as in cases of human-viral chimeric transcription. Specifically, Virana was the only method able to detect evidence for all viruses even at low twofold coverages. We observed that both RINS and CaPSID discarded a substantial amount of human-viral homologous transcripts due to their high similarity to the human reference genome, a fact that explains the lower performance of these methods in this validation scenario.
Analysis of positive and negative experimental controls further reveals that Virana is able to detect viral transcripts associated with four types of cancer at a high dynamic range of relative abundances. While Virana displays a slightly reduced specificity in simulated and experimental evaluations, these false positive hits are limited to only two viral families (Flaviviridae and Retroviridae) that display high sequence similarity to human factors. These hits are additionally annotated with an ambiguous taxonomic origin by Virana. In addition, Virana provides extensive support for investigating such ambiguous viral hits by analyzing the homologous context of putative viral reads in a context of multiple sequence alignments and phylogenies.
In principle, several biological confounding factors may hinder detection of viral transcripts by any sequence-based method. Low concentration and extratumoral location of viral producer cells [8] or selection of growth-autonomous cells in progressed tumors [79] can significantly dilute the number of viral transcripts in a sample. Additionally, known tumor viruses such as high-risk HPV strains, EBV, and MCPyV selectively transcribe their genome during viral latency (HPV: E6/7 [80], [81], EBV: EBNA1/2 [82]–[84], MCPyV: large T antigen [31], [85]), thus generating only low abundances of tens (MCPyV [31]) to hundreds (KSHV [86], EBV [87]) of transcripts per cell. Last, transcription of human oncogenic factors modulated by viral [88] or endogenous [89], [90] retroviral promoters as well as ‘hit-and-run’ mechanisms of viral oncogenesis that imply loss of viral material [91], [92] may predispose cells to transformation without requiring maintenance of viral transcripts.
Our approach aims to counteract these confounding factors by two strategies: first by sequencing neuroblastoma transcriptomes at comparatively high depth in order to detect rare transcripts and second by using several biological replicates at different tumor stages, thus reducing the probability of total loss of viral material from all analyzed samples. Based on statistical estimations concerning Virana's homologous region construction process and the sequencing depth of our experimental data, we can conclude that our approach requires minimal abundances of only two average-length viral transcripts per cell even under adverse conditions such as high viral divergence or extensive human-viral sequence homology. While representing a theoretical sensitivity that may be altered by sequencing biases [93], these copy numbers compare very favorably with related estimates reporting minimal abundances of one to several complete viral genomes per cell [27], .
After applying Virana to several positive control panels of human cancers with known viral cofactors and accurately reconstructing large fragments of viruses that are causally related to the respective tumors, we analyzed neuroblastoma transcriptomes at high sequencing depth and using two different sequencing platforms. Analyses of neuroblastoma transcriptomes resulted in the detection of putative viral transcripts with high local sequence similarity to several viral families. However, automatic taxonomic annotation as well as detailed manual inspection of homologous regions pertaining to these families revealed the human or bacteriophage origin of all transcripts. While we could find differences in the abundance of HERV-K transcripts between neuroblastoma stages 4 and 4S, the causative role of HERV transcription with regard to oncogenesis is currently unclear [94] and, as to our knowledge, only tentative associations with specific cancers have been made as to date [70]. Apart from these tentative differences in HERV-K abundances, no quantitative difference between neuroblastoma stages 4 and 4S could be identified with regard to viral transcription.
In conclusion, our observations provide negative evidence regarding the contested question of putative viral cofactors of metastatic neuroblastoma by suggesting that viruses are unlikely to be frequent cofactors in the maintenance of metastatic neuroblastoma.
|
10.1371/journal.pgen.0040002 | Dissecting the Genetic Components of Adaptation of Escherichia coli to the Mouse Gut | While pleiotropic adaptive mutations are thought to be central for evolution, little is known on the downstream molecular effects allowing adaptation to complex ecologically relevant environments. Here we show that Escherichia coli MG1655 adapts rapidly to the intestine of germ-free mice by single point mutations in EnvZ/OmpR two-component signal transduction system, which controls more than 100 genes. The selective advantage conferred by the mutations that modulate EnvZ/OmpR activities was the result of their independent and additive effects on flagellin expression and permeability. These results obtained in vivo thus suggest that global regulators may have evolved to coordinate activities that need to be fine-tuned simultaneously during adaptation to complex environments and that mutations in such regulators permit adjustment of the boundaries of physiological adaptation when switching between two very distinct environments.
| The mammalian intestine is a privileged physiological site to study how coevolution between hosts and the trillions of bacteria present in the microbiota has shaped the genome of each partner and promoted the development of mutualistic interactions. Herein we have used germ-free mice, a simplified albeit ecologically relevant system, to analyse intestinal adaptation of a model bacterial strain, Escherichia coli MG1655. Our results show that single point mutations in the ompB master regulator confer a striking selective adaptive advantage. OmpB comprises EnvZ, a transmembrane sensor with a dual kinase/phosphatase activity, and OmpR, a transcription factor controlling more than 100 target genes. In response to environmental changes, EnvZ modulates the phosphorylation and thereby the transcriptional activity of OmpR. We further show that the selective advantage conferred by OmpB mutations is related to their additive and independent effects on genes regulating permeability and flagellin expression, two major set of genes controlled by OmpR. These results suggest that global regulators may have evolved to coordinate physiological activities necessary for adaptation to complex environments and that mutations offer a complementary genetic mechanism to adjust the scale of the physiological regulation controlled by these regulators in distinct environments.
| Bacterial populations are powerful model to explore the mechanisms of evolution. Several in vivo experiments have pointed to the possible important role of pleiotropic adaptive mutations, but their molecular basis remain in most of cases largely elusive [1–3]. Here we have used gnotobiotic mice that offer a simplified and controlled albeit ecologically relevant experimental environment model to analyse the adaptation of E. coli MG1655 to the gut, as E. coli is usually the first colonizer of the mammalian newborn germ-free intestine [4,5]. Taking advantage that this laboratory strain is entirely sequenced and easily accessible to genetic manipulations, we could design a study that allowed deciphering the beneficial effects of pleiotropic mutations during intestinal colonisation.
The mammalian intestine is a privileged physiological site to study how coevolution between hosts and the trillions of bacteria present in the microbiota has shaped the genome of each partner and promoted the development of mutualistic interactions. Genetic adaptation to the host over the millions years of coevolution has translated into physiological regulatory pathways that are rapidly mobilized in response to intestinal colonization [6–9]. In the microbiota, the contrast between the considerable number of species, more than a thousand, and the small number of bacterial divisions [10], indicates that coevolution has selected bacterial genera possessing the genetic gear to adapt to the host environment, a notion supported by recent evidence that gut habitats in different host species dictate distinctive structures of intestinal bacterial communities [11]. Yet, the intestine is a complex and highly dynamic ecosystem composed of a large diversity of niches that vary in space and time, where bacteria face a permanent adaptive challenge. Furthermore, intestinal bacteria must be able to hurdle between their hosts across the exterior environment and for certain such as E. coli to switch between two entirely distinct natural environments. Gnotobiotic animals that offer a simplified albeit relevant model to study reciprocal mechanisms of adaptation between bacteria and their hosts, within a few days, the host can only adapt via physiological changes, whereas bacteria can adapt both by gene regulations and adaptive mutations. Indeed, we have previously demonstrated that adaptive mutations are central for efficient intestinal colonization by E. coli MG1655 [12]. Here we show that adaptation of this strain of E. coli during intestinal colonization entails rapid and parallel evolution in the EnvZ/OmpR two-components transduction system [13]. The gain of fitness provided by the diverse mutations selected in this global regulator during in vivo colonization results mainly from two distinct and measurable effects on motility and permeability that are both reduced in the mutant strains selected in the gut environment. These findings suggest that evolutionary pressures can put a diverse set of physiological functions facilitating adaptation under the control of one global regulator, and that mutations permit to adjust the scale of the physiological regulation controlled by this regulator in a given environment.
We have shown that adaptive mutations play a critical role in the success of the E. coli MG1655 strain in colonizing of the mouse gut [12]. A possible clue to the nature of the mutation(s) selected during colonization ensued from our subsequent observation of bacteria with a reduced motility phenotype in the feces of all gnotobiotic mice colonized with the wild type MG1655 strain (WT) (Figure 1A). The colonies displayed a new small and granular morphotype (SG) distinct from the large and smooth morphotype (LS) of the WT inoculated strain (Figure S1). SG colonies forming bacteria, undetected in the initial inocula, appeared in the feces within two days, and reached a prevalence of 90% within seven days (Figure 1B). Their phenotype remained stable when grown in vitro over many generations, indicating that it was heritable and may result from the rapid in vivo selection of mutation(s).
In order to identify the potential mutations responsible for the SG morphotype, a clone forming SG colonies (SG1) isolated from mouse feces two days post-colonization, was transformed with a genomic DNA plasmid library generated from the parental WT strain. All plasmids that restored the ancestral WT LS morphotype carried the ompB locus, coding for the membrane sensor EnvZ and the transcriptional regulator OmpR of a two-component signal transduction system central to the osmolarity-dependent regulation of genetic expression [13]. A chloramphenicol resistance gene (cat), inserted downstream the ompB locus in the chromosome of the WT ancestral and the SG1 strains, co-transduced with a 95% frequency with the morphotype (LS or SG), indicating that in the SG1 strain, the DNA region surrounding cat was responsible for the SG morphotype. This region was sequenced for one SG and one LS clone harvested from the feces of each of the 8 independent mice inoculated with either MG1655 or an MG1655 E. coli strain carrying a yellow fluorescent protein (YFP) as reporter of fliC expression (MG1655pfliC-YFP) (see below). While no mutation was detected in LS clones, all SG clones displayed a different missense point mutation, seven located in envZ, and one in ompR (Table1). The independent systematic and rapid selection of mutations in the same genes under identical experimental conditions is evidence for a strong selective advantage of the mutants during gut colonization [1].
To confirm and estimate the relative fitness of the SG1 mutant versus the ancestral strain in the mouse gut, we performed in vivo competition experiments between strains isogenic except for the point mutation present in the envZ gene of the SG1 strain (SG1 mutation) and the inducible fluorescent marker (RFP vs. GFP). Prior experiments have indicated that these inducible markers do not induce any selection bias [14]. The ratio of mutant (GFP) to WT (RFP) colonies was defined after culture of the feces and ex vivo induction of the fluorescent marker. Competition experiments using initial ratios of mutant to WT strain of 1:1, 1:100 and 1:1,000 indicated that the SG1 mutation confers a considerable fitness gain (Figure 1C). With the assumption that the mean generation time for E. coli in the gut is 60 minutes [15], the selective advantage of the SG1 mutation was estimated to be 24% when the mutant to WT strain ratio remained under 1:10 (Table S1). These data explained how adaptive mutations in envZ, that are likely to happen at a frequency below 10−7, can be very rapidly selected upon colonisation with the WT strain. The selective advantage of the SG1 mutation decreased to approximately 10% when the ratio of mutant to WT strain increased over 1:10, indicating that the selective advantage conferred by the mutation is frequence-dependent, consistent with the observation that the WT strain is not entirely displaced in the mono-colonization experiment (Figure 1B).
Importantly, the selected mutants did not exhibit the same motility phenotype as null mutations, since strains deleted for envZ, ompR or both kept the wild type LS morphotype (Figure S1). The membrane receptor kinase-phosphatase EnvZ forms a two-component pair with its cognate response regulator, OmpR, that enable cells to sense external changes of osmolarity [13]. The native receptor exists in two active but opposed signalling states, the OmpR kinase-dominant state and the OmpR-P phosphatase-dominant state. The balance between the two states determines the level of intracellular OmpR-P, which in turn determines the level of transcription of the many target genes [13].
One important bacterial function controlled by OmpR is motility, as OmpR regulates transcription of the flhDC operon, the master regulator of flagellar biosynthesis [16]. Several mutations identical to those selected in vivo during colonization were previously shown, by in vitro mutational analysis of EnvZ activities, to switch on the EnvZ kinase-dominant state [17,18] (Figure 2), resulting in increased levels of phospho-OmpR and repression of the flhDC operon [16].
Consistent with repression of flagellin expression in all SG mutants, no flagellin could be detected in cell lysates or supernatants obtained from stationary phase cultures, while the ancestral WT strain and the LS colonies (that kept the wild-type motility phenotype after mouse colonization) synthesised large amounts of flagellin in the same in vitro conditions (Figure 3B). We have previously shown that the WT ancestral E. coli strain induces a potent NF-κB-dependent inflammatory response in intestinal epithelial cells that hinges on the interaction of flagellin with Toll receptor 5 [19]. Consistent with impaired flagellin expression, culture supernatants of SG strains in stationary conditions, failed to induce any inflammatory signal in monolayers of epithelial cells (Figures 3A and S2).
These in vitro observations showing repression of flagellin synthesis in SG mutants were thus compatible with the observed defective motility morphotype. This morphotype was however clearly distinct from the pin point morphotype of the ΔfliC strain lacking the gene encoding flagellin, the primary flagellar subunit (Figure S1). In order to confirm that flagellin was downregulated by SG mutants in vivo in the intestine, germ-free mice were inoculated with an MG1655 E. coli strain carrying a yellow fluorescent protein (YFP) as reporter of fliC expression. The bacterial fluorescence in the feces was monitored in the feces by flow cytometry. Fluorescence decreased rapidly in mice inoculated with the WT strain, demonstrating in vivo down modulation of flagellin (Figures 4 and S3). Fluorescence monitoring after plating confirmed this result. Thus, in mice inoculated with the WT strain, the fraction of fluorescent colonies decreased to an average of 10% within 8 days, consistent with the selection of SG mutants described above (Figure 4B). Furthermore, all bacteria forming non-fluorescent colonies tested on motility plate exhibited an SG morphotype, while those forming fluorescent colonies retained the LS morphotype (Figure 4B).
As OmpR/EnvZ controls many activities, we looked for other effects of the selected mutants. The characteristic motility phenotype of the SG selected mutants could be a result of an enhanced aggregation of bacteria to each other via the production of curli fibres encoded by the csgBA operon whose expression is regulated by the OmpR regulated csgD gene [20]. However, in contrast to the previously described ompR mutant of E. coli K12 that promotes biofilm formation via the derepression of the csgA gene [21], none of the SG mutants exhibited changes in csgA gene expression and their biofilm formation was reduced compared to the WT strain (data not shown).
Another essential function of the two-component system envZ/ompR is to modulate membrane transport and permeability in response to medium osmolarity [22]. In particular, OmpR affects the reciprocal transcription of the small pore OmpC and large pore OmpF porins [23], the two E. coli porins that are thought to play a central role in the adaptation of E.coli to the hyperosmotic conditions of the intestine [24]. Consistent with mutations that switch on the OmpR kinase-dominant state of EnvZ, selected SG mutants had decreased ompF and increased ompC mRNA and membrane protein levels compared to the WT ancestral strain (Table 1, Figure 3C), i.e. a reduced permeability phenotype [23]. Membrane permeability is central for both stress protection and nutritional competence [25]. It has been postulated that reduced permeability would be favourable in the environmental conditions of the gut, consisting of high osmolarity, low oxygen pressure and the presence of bile salts [24]. Indeed, all SG mutants grew much better than the ancestor in medium containing bile salts, the ancestor being entirely displaced within 7 hours of growth (data not shown).
Transcriptome analysis has pointed to the potential role of the two-component EnvZ/OmpR system in the regulation of multiple genes, including genes involved in transport across membranes and cell metabolism [22], which may perhaps promote intestinal adaptation of E. coli. We therefore assessed the importance of flagellin repression and/or porins regulation on the parallel selection of envZ-ompR mutations.
To analyse the role of flagellin in the selection of SG mutants, germ-free mice were inoculated with either the WT or the ΔfliC strain carrying a fluorescent protein (YFP) as reporter of fliC expression. Flow cytometry analysis of the feces showed that in situ fluorescence decreased faster and more extensively in mice inoculated with the WT than with the ΔfliC strain (Figures 4A and S3), a result confirmed by fluorescence monitoring after plating (Figure 4B). Thus, in mice inoculated with the ΔfliC strain, the fraction of fluorescent colonies had decreased to only 50% on day 8 as compared to 10% in mice inoculated with the WT strain and the kinetics of selection was slower (Figure 4B). Altogether, these results point to a strong impact of flagellin on the selection of EnvZ mutations. However, mutations downregulating fliC expression could still be selected despite the absence of flagellin, presumably because of the pleiotropic effect of these mutations. Sequencing the ompB locus in non-fluorescent clones harvested from 4 mice inoculated with the ΔfliC strain revealed missense point mutations (Table 1). Three were located in envZ, including one identical to a mutation found in a clone isolated from a mouse inoculated with the WT strain. The fourth one was located in the same codon of ompR as the mutation identified in a clone derived from the WT strain (Table 1). These results show that the adaptive advantage conveyed by selected mutations is only partially flagellin-dependent, suggesting that selected mutations provide further advantage resulting from the modulation of other genes controlled by OmpR.
One likely candidate was the large porin encoding gene ompF. Indeed we have observed that this gene expression is downmodulated by the selected envZ-ompR mutations, resulting in a reduced permeability phenotype known to be associated with increased resistance to bile salts [26], as observed for SG mutants. To assess the role of OmpF in the selection of EnvZ mutations, mice were inoculated with a ΔompF mutant that expresses OmpC but no OmpF protein (Figure 3C) and carries the YFP reporter of fliC expression. Although the impact of OmpF deletion alone was not as strong as the one of flagellin, selection of non fluorescent mutants studied in the feces after plating was significantly less efficient than in mice colonized with the WT E. coli strain (Figure 5). In one out of five studied mice, all non-fluorescent mutants exhibited an SG phenotype in soft agar plates. In two other mice, the non-fluorescent colonies had a totally nonmotile (NM) pinpoint phenotype comparable to the ΔfliC-engineered strain (Figure S1). In the last two mice, both SG and NM morphotypes were observed. Sequencing the ompB locus revealed a missense mutation in envZ in all SG clones tested (Table 1). In contrast, NM clones forming pin-point colonies had a normal envZ sequence but contain large deletions from 1.5 to 12 kb between the otsA and cheB loci, encompassing the flhDC operon and thereby precluding any expression of the whole flagellum operons (Figure 6). Interestingly, all deletions had occurred immediately upstream of an Insertion Sequence (IS1) located just upstream the flhDC operon, and probably reflecting an imprecise excision of the IS [27]. The deleted genes, that all belong to the chemotaxis/motility pathway, failed to be amplified by PCR (data not shown), showing that they were indeed lost rather than inserted ectopically. These results show that in mutants with reduced permeability, the major fitness gain results from repression of gene(s) controlled by FlhDC, probably flagellar genes and in particular the fliC gene encoding flagellin.
To confirm this hypothesis, mice were inoculated with double ΔfliC ΔompF mutants carrying the YFP reporter of fliC expression and expressing the fluorescent CFP protein under the control of a constitutive promoter. Strikingly, combining deletions of porins and flagellin had additive effects and almost entirely abolished the in vivo selection of EnvZ/OmpR mutants (Figure 5). At day 11 post-inoculum, only 9% of clones were YFP-negative. None had mutation in the ompB locus, a deletion in the flhDC region or a mutation in the pfliC-YFP construct. These YFP-negative clones were all CFP positive and remained CFP positive during the 100 days of observation. Since YFP and CFP were expressed at the same level in the inoculated strain, the hypothesis that YFP expression was costly for the bacteria and eliminated by mutations is unlikely.
Therefore, our results show that the selective pleiotropic advantage conferred EnvZ/OmpR mutation predominantly results from a combined effect of modulation of fliC expression and membrane permeability, but does not exclude minor additional effect(s) of (an)other as yet uncharacterized gene(s) under the control of EnvZ/OmpR.
Due to their high growth rate and large population size, microbes have a remarkable capacity to evolve and diversify by generation and spread of mutations that improve their fitness in a given environment [1]. We have previously observed that within a few days a mutant strain with a high mutation rate increased in frequency to the expense of the parental commensal E. coli MG1655 strain during gut colonization. In contrast, the mutator strain lost the competition against a clone collected from the feces of mice colonized for 40 days with the parental commensal E. coli MG1655 [12]. These results suggested that adaptive mutations enable bacteria to rapidly and efficiently cope with the drastic environmental changes encountered during gut colonization. Our novel results identify the central role of the EnvZ/OmpR regulon in the physiological adaptation of E. coli MG1655 to the gut environment, and show that adaptive mutations in this two-component system provide an additional gear to adjust precisely the scale of the physiological regulation controlled by this regulator to the gut environment. Furthermore our results provide the molecular basis of the beneficial effects of the pleiotropic mutations in EnvZ/OmpR in adaptation of E. coli MG1655 to the mouse gut.
Mutations in the envZ/ompR locus were systematically detected in 90% of bacteria harvested from independent mice feces within a week of colonization with WT E. coli MG1655. Except for one mutation in its cognate transcription factor OmpR, all mutations were found in the membrane sensor EnvZ. The major fitness gain conferred by these mutations was confirmed by in vivo competitions between the ancestor WT strain and an isogenic mutant strain harboring the prototype SG1 envZ mutation. The emergence of distinct point mutations at the same two-component locus in bacterial populations evolving in different colonized mice suggested a comparable impact on the physiological effects mediating the fitness gain due to these mutations. Indeed all mutations resulted in profound repression of flagellin expression and modulation of OmpF versus OmpC porin expression yielding a reduced permeability phenotype. This phenotype is typical of mutations that switch the phosphatase/kinase membrane sensor EnvZ toward a OmpR kinase-dominant state. Indeed several of the missense mutations selected during in vivo colonisation were previously identified by in vitro mutational analysis as turning on this functional state [17,18]. Mutations selected during colonization were not restricted to the catalytic domains of EnvZ, but were also found in the periplasmic sensor and cytoplasmic linker domains, highlighting the participation of all of the protein's domains in the control of gene regulation (Figure 2). Interestingly in mice colonized with the ΔfliC mutant, where adaptive mutants were mainly selected on their reduced permeability phenotype, mutations were still exclusively found in the EnvZ/OmpR system, a result that underscores the prominent role of the EnvZ/ompR system in the regulation of membrane permeability of E. coli MG1655 during intestinal colonization.
Notably, colonization with the WT E. coli did not select for mutations inactivating genes specifically controlling motility or permeability. Yet, selection of mutants with deletion of flhD/C operon was observed during colonization by the ΔompF strain, a result reminiscent of observations in streptomycin-treated mice [28,29]. Clonal interference [30] thus likely prevents the selection of mutations affecting only one function, presumably associated with smaller selective value than the pleiotropic mutations in envZ/ompR modulating simultaneously functions as different as permeability and motility. Indeed, using reporter mutant bacteria carrying a fluorescent protein under the control of the fliC promoter, we could clearly demonstrate that the selective advantage conveyed by mutations in envZ/ompR resulted from their pleiotropic and additive effects on the repression of flagellin production and OmpF porin expression. The almost complete abolition of adaptive selection of envZ/ompR mutations in mice colonized with a double mutant E. coli strain that lacks both fliC and ompF, underscores the major contribution of the pathways controlled by envZ and ompR in the intestinal adaptation of E. coli. The precise elucidation of the selective forces is beyond the scope of this study, but likely scenarios are briefly discussed below. Flagellin downregulation could be selected for via its pro-inflammatory role [19,31–35], via its direct energetic cost [28,36], or via still non-identified mechanisms. The fitness gain conveyed by reduced permeability was suggested by in vitro analysis indicating that, similar to ΔompF mutants, all ompR/envZ mutants grew much better in medium containing high concentrations of bile salts, a major stress factor for bacteria in the intestinal lumen. Interestingly, it has been reported that the concentration of biliary salts in the intestinal lumen decreases upon colonization [37,38]. A lower concentration of biliary salts in mice treated by streptomycin which empties the enterobacteriae niche but does not deplete completely the intestinal flora, might explain the predominant selection of mutants in the flhD/C operon in this mouse model [28,29].
In E. coli, stress protection comes at the cost of nutritional competence through the regulation of membrane permeability [39]. In the gut rich environment, bacterial nutrient intake is likely sufficient even if permeability is restrained, so that the growth rate is not significantly affected [25]. Yet, the extent of physiological regulation allowed by wild type EnvZ/OmpR might not be optimal to respond to our experimental mice gut conditions. Thanks to adaptive mutations in EnvZ/OmpR, the trade-off between self-preservation and nutritional competence (SPANC balance) might easily be switched to either better resistance or faster growth [25]. To mutate may thus represent a complementary genetic gear to adjust precisely the scale of physiological regulation controlled by a global regulator when switching between complex environments.
Notably, the selective advantage conferred by the envZ mutations was frequency dependent, consistent with the observation that in mice colonized with the WT strain, the mutation invades rapidly and massively the population, but does not go to fixation, as a minor part of the population kept the original colony morphotype (and genotype for envZ-ompR). These results suggest a mechanism causing the coexistence of ancestral and evolved form, perhaps because the ancestral phenotype confers some advantage to colonize a specific niche. Work is in progress to address this issue.
Experiments with microbial populations have been largely used to gain insight into the mechanics of evolution and have pointed to the possible important role of pleiotropic adaptive mutations [1]. Thus, finding mutations in regulatory genes is a recurrent observation both in natural populations and during in vitro experimental evolution, that led to postulate that mutations affecting regulators are more likely to promote adaptation and evolution than those improving a single enzymatic step [1,25,40]. Our results obtained in an in vivo model of bacterial evolution supports this hypothesis. As mutations in global regulators affect the regulation of many genes, they must be pleiotropic and are thus expected to result in the expression not only of beneficial but also of detrimental traits. The molecular mechanisms responsible for the selection of such pleiotropic mutations have therefore remained largely elusive in most systems. A recent study in a simple ecological in vitro model [41], has shown that adaptive mutations allow P. fluorescens to occupy a novel ecological niche at the air-liquid interface [42]. All selected strains had pleiotropic loss-of-functions mutations in one gene encoding a putative methyl-esterase in the wsp operon [2,3]. Drawing analogy with the che operon of E. coli that encodes proteins homologous to the wsp operon, the authors suggested that this protein acts in concert with a putative methyl-transferase to adjust the activity of a kinase. The mutations may thus destroy the capacity of the pathway to fluctuate between activity states, producing instead a steady state output allowing niche specialization. Our results, combined with previous biochemical works [17], provide direct evidence that a distinct scenario promotes the in vivo adaptation of an E. coli MG1655 to the gut of germ-free mice. In the case of EnvZ/ompR, the two opposed enzymatic activities are exerted by the cytoplasmic domain of EnvZ and are modulated in response to signals sensed by the external domain of the protein. Mutations in EnvZ, that directly affect the balance between two activities, are selected because of their independent and additive effects on genes controlling flagellin expression and membrane permeability. Dissecting the fitness gain due to these independent pathways allowed us to demonstrate that the EnvZ/OmpR global regulator orchestrates the physiological adaptation of E. coli MG1655 to the gut environment. More generally, the observation that the EnvZ/OmpR system gathers under its control genes central to promote intestinal colonization leads us to suggest that global regulators may have arisen during evolution to optimize the coordination of genes that collaborate to adapt to a given niche. Mutations in such global regulators may provide a complementary genetic tool that allows bacteria to extend the scale of the physiological regulation and promotes their rapid adaptation when confronted to very specific environments.
All strains were derived from the commensal flagellated E. coli K12 MG1655 sequenced strain [43]. The MG1655 ΔfliC E. coli isogenic mutant has been described [19]. To construct the reporter WT pfliC-YFP strain used to monitor in vivo activity of fliC promoter, sequence encoding the fluorescent protein YFP++ [44] was cloned downstream the upstream region of the fliC gene (pfliC: from nucleotides −230 to +5 relative to the translation start). The fragment (pfliC YFP, T1T2 and cat) was flanked by 40 nucleotides sequences homologous respectively to the 5′ and 3′ of the IS2 and IS30 insertion sequences interrupting the ybdA E. coli gene and by KpnI and SphI restriction sites and cloned in p5Y, a pUC-18-derived plasmid. After plasmid amplification, the fragment was inserted into the ybdA gene of the MG1655 E. coli chromosome replacing the IS sequences following method already described [45]. MG1655 ΔfliC pfliC-YFP was constructed by P1 phage co-transduction of the pfliC-YFP-v+ and the cat alleles from MG1655 pfliC-YFP into MG1655 ΔfliC strain.
The MG1655 ompB-cat and SG1 ompB-cat E. coli strains (used to assess the link between the ompB locus and the motility phenotype) were constructed by inserting the FRT flanked cat gene of the pKD3 plasmid [45] between the envZ and pck genes as described [45], using PCR primers that contained a 40 bases-5′ end extension centered on the translation stops of the envZ or pck gene. Insertion of the PCR product was monitored using primers respectively identical or complementary to the nucleotides 1562 to 1582 of pck and 1238 to 1258 of the EnvZ gene. These strains kept the motility phenotype of the MG1655 and SG1 strains respectively.
The MG1655 ptet-GFP ompBSG1-cat and MG1655 ptet-RFP ompB-cat E. coli strains (used to measure the relative fitness of the SG1 strain in vivo) were constructed by introducing by P1 phage co-transduction of the ompB region from the SG1 ompB-cat strain and the cat allele into the MG1655 ptet-GFP and the MG1655 ptet-RFP strains respectively (described in [14]). The MG1655 ptet-GFP ompBSG1-cat strain was selected among granulous transductants (SG morphotype) in motility agar whereas the MG1655 ptet-RFP ompB-cat was selected among transductants that kept the WT motility phenotype (LS morphotype).
The ΔompF, ΔompC, ΔompR, ΔenvZ, and ΔompB strains were constructed by replacing the ompF, ompC, ompR, envZ and envZ and ompR open reading frame respectively from start to stop codon by the FRT flanked cat gene of the pKD3 in the E. coli MG1655 strain following method already described [45]. The MG1655 ΔompF pfliC-YFP strain was constructed by P1 phage co-transduction of the ΔompF and the cat alleles from MG1655 ΔompF into MG1655 pfliC-YFP p2rrnB-CFP strain. The MG1655 ΔfliC ΔompF pfliC-YFP strain was constructed by P1 phage co-transduction of the ΔompF and the cat alleles from MG1655 ΔompF into MG1655 ΔfliC pfliC-YFP p2rrnB-CFP strain.
To construct the reporter p2rrnB-CFP, sequence encoding the fluorescent protein CFP++ was cloned upstream of the promoter p2 of the rrnB operon (p2rrnB: from nucleotides 152 to 94 relative to the translation start of the rrsB gene). The fragment (prrnB-cfp, T1T2 and cat) was flanked by 40 nucleotides sequences homologous respectively to the 5′ and 3′ of the IntC (IntS) E. coli gene and by KpnI and PacI restriction sites and cloned in a pUC-18-derived plasmid. After plasmid amplification, the fragment was inserted into the IntC gene of the MG1655 E. coli chromosome following method already described [45].
Genomic DNA from E. coli MG1655 strain was prepared with the Wizard Genomic DNA Preparation kit (Promega, Charbonnières, France) and partially digested with the Sau3AI restriction enzyme. Fragments ranging from 2 to 6 kb were eluted from agarose gel (Gel extraction kit, Promega), and cloned into BamHI-digested and dephosphorylated pACYC184 plasmid. The purified ligation reaction was used to electro-transform DH5-α E. coli. Transformants were selected on LB plates containing chloramphenicol. Ligation efficiency was 95% and average size of genomic inserts 3 Kb. Plasmids were extracted from about 1.5 × 104 pooled colonies (Miniprep kit, Promega). The SG1 clone was transformed with the genomic library and transformants were selected on motility plates supplemented with chloramphenicol. The clones with a wild type motility phenotype (LS) were isolated and the E.coli MG1655-derived locus carried by the transforming plasmids was determined by sequencing with primers flanking the cloning site.
Sequencing of the ompB locus (from the greB translation stop to the pck translation stop) and pfliC-YFP construction of the MG1655 strain and of the clones isolated from mouse feces was carried out on purified PCR amplification products using standard procedures in the Institut Cochin sequence facilities.
The following primers were used to define the size of the deletions in MG1655 ΔompF pfliC-YFP non motile (NM) mutants: otsAup (5′-GTGCAACTCAGGCATCATGG-3′) either in association with CheBdwn (5′-CGTATGGTGGAAAAGTCATCC-3′) for clones NM2 and NM4, with CheAdwn (5′-cgctgaagccaaaagttcctgc-3′) for the clone NM1 and with ArgSdwn (5′-CTAACGGCATGATGGGAGTTG-3′) for clone NM3.
Bacterial motility was monitored in soft agar plates (4.5 g/L agar in Luria broth medium (LB)) at 30 °C for 24 h. Enumeration of fluorescent bacteria was made on solid LB agar plates (15 g/L) after a 48-h incubation at 37 °C using a lighting system (LT-9500–220 Illumatool, Lightools Research). YFP fluorescence was detected in colonies using 470-nm excitation wavelengths and 530-nm reading filters. Fluorescence detection in feces was performed on dilutions of freshly passed feces using a BD-LSR flow cytometer (Becton Dickinson). Data were analyzed with Cell Quest software (Becton Dickinson).
Strains were grown in LB for 16 h, and population sizes were determined by plating appropriate dilutions of the culture on LB plates. 50 μL of a 1 × 104 fold dilution of the pre-culture of the mutant and of the reference parental strain were inoculated in 5 mL of LB and LB supplemented with bile salts (Bile salts N°3, Difco) at 5% (M/W) and incubated at 37 °C under agitation. Mutant and parental population sizes were determined after 7h30 of culture by counting SG and LS populations on motility plates (for SG1 and SG2 against MG1655 competitions), or fluorescent and non-fluorescent populations as described above.
Conventional and germ-free C3H/HeN mice were bred at the INRA facilities. Germ-free and gnotobiotic mice were reared in isolators (Ingenia) in individual cages and fed ad libitum on a commercial diet sterilized by gamma irradiation (40 kGy) and supplied with autoclaved (20 min, 120 °C) tap water. For colonization experiments, 8–12-week-old germ-free mice were inoculated per os with 104 bacteria from the chosen strain in 0.5 mL 10−2 M MgSO4. Colonization was monitored by bacterial counts in individual freshly harvested fecal samples as described [12].
For in vivo bacterial competition, MG1655 ptet-GFP ompBSG1-cat (containing the envZ SG1 I281S mutant allele) and MG1655 ptet-RFP ompB-cat E. coli (containing the envZ wild type allele) were grown in LB for 16 h and mixed at the 1:1, 1:100, 1:1,000 SG to WT ratios. 0.5 mL of a 1 × 104 fold dilution in 10−2M MgSO4 of these mix were used for mouse colonization. Mutant and WT population sizes were determined every 12 h by counting Red and Green fluorescent CFU on plates containing anhydrotetracycline (50 μM Acros) during 5 days following colonization. The maximal relative fitness was estimated by fitting an exponential curve to the evolution of the SG/WT ratio between 12 and 36 hours following colonization with the initial ratios 1:100 and 1:1,000.
All procedures were carried out in accordance with the European guidelines for the care and use of laboratory animals.
Stimulation of monolayers of the human IEC line HT29-19A with live bacteria, preparation of epithelial cell nuclear extracts, electrophoretic mobility shift assay (EMSA), determination of CCL-20 mRNA level by real-time quantitative PCR after a 6-h stimulation and determination of IL-8 concentrations in epithelial cell supernatants by enzyme-linked immunosorbent assay (ELISA) (Duoset kits, R&D Systems) were all performed as previously described [19].
Total RNA was extracted from 5 ml of stationnary phase culture (at 37 °C, with agitation) using the RNeasy kit (Qiagen), according to the manufacturer's instructions. RNA was treated with four units of the Turbo DNA-free (Ambion) for 1h at 37 °C. RNA was quantified by measuring the optical density at 260 nm and checked for degradation by an agarose gel electrophoresis. The cDNA synthesis was performed using 2 μg RNA with random hexamers (12.5 ng/ml) and the Superscript II RNAse H− kit 5 (invitrogen) according to the manufacturer's instructions.
The real-time PCR experiments were performed using the SYBRgreen PCR Master Mix (Applied Biosystems) to quantify the expression level of the ompC and ompF genes. The rpoD gene was chosen as a reference gene for data normalization. The primers RpoD1RT (5′-GTAGTCGGTGTTCATATCGA-3′), RpoD1FT (5′-CGTCTGATCATGAAGCTCT-3′), OmpC2RT (5′-GTCAGTGTTACGGTAGGT-3′), OmpC2FT (5′-CGACTACGGTCGTAACTA-3′), OmpF2RT (5′-CCTGTATGCAGTATCACCA-3′) and OmpF2FT (5′-CCAGGGTAACAACTCTGAA-3′) were designed by the Primer Express software (Applied Biosystems). Amplification and detection of the specific products were carried out with the 7300 Real Time PCR System (Applied Biosystems). Data analysis was performed with the 7300 System Software. For each target gene, the average Ct value was calculated from triplicate reactions for RNA samples. The difference between Ct of the target gene and Ct of the endogenous reference gene (rpoD) was defined as the ΔCt. The ΔΔCt value described the difference between the ΔCt of the wild type strain and the mutant strain. The difference in expression was calculated as 2ΔΔCt, and a twofold difference was considered as significant.
Bacterial proteins were obtained from culture supernatants precipitated by 10% trichloroacetic acid and from bacterial pellets sonicated in PBS containing an anti-protease cocktail (Roche Diagnostics) and 1% Triton X100 (Sigma). Twenty μL of 25-fold concentrated bacterial supernatants or 20 μg of total proteins from bacterial lysates were electrophoresed on 10% SDS-PAGE gels and transferred onto PVDF membranes (Amersham Biosciences, Saclay, France). Membranes blocked with 5% nonfat dry milk in 20 mM Tris pH 7.5, 150 mM NaCl, and 0.05% Tween-20, were incubated overnight with a 1:2,000 dilution of monoclonal antibody 15D8 against E. coli flagellin (Bioveris Europe), and then for 1 h with a 1:8,000 dilution of HRP-conjugated goat anti-mouse immunoglobulins (Amersham Biosciences). HRP was revealed with ECL-Plus light (Amersham Biosciences) using a luminescent image analyzer LAS-1,000plus (Fujifilm).
Cultures (20 ml) grown at 37 °C with agitation were harverested and washed in 20 mM sodium phosphate buffer, pH 7.4. The pellet were suspended and sonicated in 10 mM Hepes buffer, pH 7.4 (Vibra-cell, Bioblock Scientific). Sarkosyl was added to a final concentration of 0.5% and the detergent extraction was carried out at room temperature for 1 hour. The unbroken cells were removed by centrifugation at 3,000 rpm for 10 min, and the outer membrane fraction was obtained by an ultracentrifugation at 40,000 rpm for 1 hour. The outer membrane proteins were suspended in 10 mM Hepes and quantified by Bradford Method (Biorad). Samples were analyzed by SDS-polyacrylamide gel electrophoresis containing 8M urea as described previously [46]. Gels were transferred onto PVDF membranes at 200 mAmp for 50 minutes. Membranes blocked with 5% nonfat dry milk in 20 mM Tris pH 7.5, 150 mM NaCl, and 0.05% Tween-20, were incubated overnight with a 1/1,000 dilution of an rabit anti-OmpC/F (gift from Roland Lloubès, CNRS UPR 9027, Institut de Biologie Structurale et Microbiologie, Marseille), and then for 1 h with horseradish peroxidase conjugated anti-rabbit immunoglobulins (Cell Signaling Technology). |
10.1371/journal.pbio.2004986 | Canonical PRC2 function is essential for mammary gland development and affects chromatin compaction in mammary organoids | Distinct transcriptional states are maintained through organization of chromatin, resulting from the sum of numerous repressive and active histone modifications, into tightly packaged heterochromatin versus more accessible euchromatin. Polycomb repressive complex 2 (PRC2) is the main mammalian complex responsible for histone 3 lysine 27 trimethylation (H3K27me3) and is integral to chromatin organization. Using in vitro and in vivo studies, we show that deletion of Suz12, a core component of all PRC2 complexes, results in loss of H3K27me3 and H3K27 dimethylation (H3K27me2), completely blocks normal mammary gland development, and profoundly curtails progenitor activity in 3D organoid cultures. Through the application of mammary organoids to bypass the severe phenotype associated with Suz12 loss in vivo, we have explored gene expression and chromatin structure in wild-type and Suz12-deleted basal-derived organoids. Analysis of organoids led to the identification of lineage-specific changes in gene expression and chromatin structure, inferring cell type–specific PRC2-mediated gene silencing of the chromatin state. These expression changes were accompanied by cell cycle arrest but not lineage infidelity. Together, these data indicate that canonical PRC2 function is essential for development of the mammary gland through the repression of alternate transcription programs and maintenance of chromatin states.
| The formation of mammary glands requires the tight regulation of many genes that govern cell fate decisions in the cells that form them. However, most of these genes remain undefined. The Polycomb repressive complex 2 (PRC2) has a role in gene silencing, and it is comprised of several subunits, which include either Enhancer of Zeste homolog 2 (Ezh2) or Ezh1 in combination with Suppressor of Zeste 12 protein homolog (Suz12) and embryonic ectoderm development (EED). Dysregulation of these subunits can lead to breast cancer. Although previous studies have analyzed the contribution of these complexes in mammary epithelium, failure to inactivate all canonical PRC2 complexes has made this task difficult. Deletion of Suz12 resulted in nonfunctional PRC2 and led to growth defects in mammary epithelial cells in vivo and in vitro. Here, we have used a 3D mammary organoid system to circumvent the lethality associated with Suz12 loss and studied chromatin dynamics and gene expression of PRC2 complexes. Our results suggest that loss of all canonical PRC2 complexes results in failure to repress transcriptional programs associated with early commitment and differentiation of mammary epithelial cells.
| A central question in biology is how different cell types maintain distinct cell fates despite containing the same genetic material. Organization of DNA into open or closed chromatin states by posttranslational modifications (PTMs) of histones has emerged as a critical mechanism underpinning cell diversity and reflecting lineage-specific gene expression, developmental programs, or disease processes [1]. The highly conserved Polycomb repressive complex 2 (PRC2), which catalyzes trimethylation of histone 3 on lysine 27 (H3K27me3), is associated with global gene repression and suppression of alternative differentiation programs. The canonical PRC2 complex is composed of the intimately associated core proteins histone methyltransferase Enhancer of Zeste homolog 2 (Ezh2) or an alternative related subunit Ezh1 [2], embryonic ectoderm development (Eed), Suppressor of Zeste 12 protein homolog (Suz12), and histone-binding protein accessory proteins [3]. Upon recruitment of PRC2 to chromatin, Ezh2/Ezh1 deposits the H3K27me3 mark associated with chromatin compaction [3]. PRC2 is required for deposition of H3K27me3 and for maintenance of this PTM upon cell division [4]. Additionally, the requirement of PRC2 activity for H3K27 mono- and dimethylation (H3K27me1 and H3K27me2) remains unclear [5]. Early studies showed that both Suz12 and Eed are nonredundant and essential for a functional PRC2 complex [6]. Deletion of Suz12 or Eed resulted in elevated expression of Hox genes in Drosophila and mammalian cells and marked increases in gene networks that control developmental lineages [7,8] because of loss of PRC2 integrity and H3K27me3-mediated repression [9]. In contrast to Eed and Suz12, Ezh2 function can be compensated, partially [10] or completely [11], by Ezh1. Nevertheless, like mice lacking Suz12 or Eed, Ezh2-deficient mice are not viable and die during early implantation stages [6].
Members of PRC2—in particular, Ezh2—are often found dysregulated in human cancers. In breast cancer, Ezh2 overexpression is associated with aggressive breast cancers and poor prognosis and inversely correlated with H3K27me3 expression [12]. It remains unclear whether Ezh2 overexpression is a consequence or cause of breast oncogenesis [13]. High levels of Ezh2 may not be sufficient to induce mammary tumors in mice [14], suggesting additional driver mutations are required. To further understand the basis of this dysregulation in cancer, it is imperative to determine the normal functions of PRC2 and Ezh2 in maintaining gene expression programs in the mammary gland.
The mammary gland in both humans and mice is a bilayered structure composed of two cellular lineages: an inner luminal layer and an outer myoepithelial layer that contacts the basement membrane [15]. There is increasing evidence for a differentiation hierarchy composed of stem cells, committed progenitors, and mature epithelial cells [15]. In the steady state, mouse mammary stem cell (MaSC)/basal, luminal progenitor, and mature luminal cell subsets display distinct patterns of H3K27me3 [16]. The MaSC/basal subset demonstrates the lowest levels of H3K27me3. Higher levels of H3K27me3 correlate with reduced gene expression and increase upon cell specialization [16]. These data support a model whereby mammary epithelial cell (MEC) differentiation requires narrowing of transcriptional programs and suppression of alternate cell fates. Accordingly, genetic ablation or knock-down of Ezh2 in the mammary gland resulted in a developmental delay [16,17] but did not entirely prevent mammary gland development. This is likely due to residual H3K27 methylation, presumably established through compensatory methyltransferase activity of Ezh1. It is therefore probable that studies to date have underestimated the importance of PRC2 in directing mammary cell fate.
To reexamine the contribution of the canonical PRC2 complex to mammary gland development, we deleted Suz12 in vivo and in mammary organoids [18]. Here, we show a nonredundant function for Suz12 in mammary progenitor cells due to loss of PRC2 function. Similar findings were made upon deletion of Eed. Through the application of assay for transposase-accessible chromatin using sequencing (ATAC-seq) to probe chromatin accessibility in Suz12-deficient organoid cultures, we have identified regions of PRC2-dependent chromatin compaction and consequent changes in gene expression. Interestingly, the chromatin state in basal-derived organoids was reliant on PRC2 function, and Suz12 deletion led to gene de-repression, thus highlighting a crucial role for PRC2 in the maintenance of chromatin states within the mammary epithelial hierarchy.
Mammary gland development proceeds through distinct phases that include puberty and cycles of pregnancy, lactation, and involution. Suz12 expression, like Ezh2 [16,17], was detected at all stages of mammary gland development (S1A Fig) but was particularly high during puberty (4–6 week old mice). To examine the role of Suz12 in the mammary gland, we crossed mice bearing floxed Suz12 alleles with Cre transgenic mice that express Cre under control of the mouse mammary tumor virus promoter (MMTVcre). Genotyping of offspring revealed that two-thirds of MMTVcreT/+Suz12fl/f mice did not survive to weaning (S1 Data), and examination of newborn pups revealed abnormal lung development (S1B Fig), likely due to activity of MMTVcre in this tissue [19], and consistent with the reported role for PRC2/Ezh2 in lung [20].
MMTVcreT/+Suz12f/f mice that survived beyond birth appeared normal and did not differ from wild-type (Wt) mice with respect to bodyweight (S1C Fig). Examination of whole mounts and histological sections of mammary glands from MMTVcreT/+Suz12f/f mice during puberty revealed a heterogeneous phenotype. Some glands were indistinguishable from Wt or heterozygous littermates, while others comprised a small ductal tree, characteristic of the rudimentary ductal tree found in newborn mice (n = 4) (Fig 1A and S1D Fig). These small but otherwise normal mammary glands indicate that gene deletion leads to severe impairment of ductal growth during puberty [21,22]. Notably, Suz12 mRNA expression in MMTVcreT/+Suz12f/f and MMTVcreT/+Suz12f/+ glands was found to be comparable (S1E Fig), demonstrating that these ductal structures were derived from epithelial cells in which Suz12 had not been deleted. Moreover, protein expression of Ezh2, Suz12, and H3K27me3 was retained in sections from 5–6 week old MMTVcreT/+Suz12f/f mammary glands, indicative of strong selection against Suz12-deleted cells and retention of a functional PRC2 complex in the mammary epithelium (Fig 1B). Cell fate was also retained as assessed by immunostaining for estrogen receptor (ER), progesterone receptor (PR), and forkhead box A1 (Foxa1) (S1F Fig).
We next determined whether deletion of the second core subunit of canonical PRC2 member Eed produced a similar phenotype to deletion of Suz12. As seen with Suz12 deletion, conditional deletion of Eed with MMTVcre resulted in perinatal lethality of mice (S1 Data), and surviving mice appeared normal, except for a pronounced absence or delay in growth of the ductal tree (S2A Fig). Similar to Suz12 deletion, Eed mRNA expression was found to be indistinguishable between mammary epithelium from MMTVcreT/+Eedf/f and Wt mice (S2B Fig). In addition, Eed, Ezh2, and H3K27me3 proteins were detected by immunofluorescence (IF) and immunohistochemistry (IHC) at comparable levels to those in Wt mammary glands (S2C and S2D Fig). Taken together, these results suggest that canonical PRC2 complexes are critical for the proliferation and/or survival of MECs, and cells deleted for PRC2 function cannot contribute to the developing mammary gland.
To investigate which MEC subsets were affected by Suz12 loss, we acutely deleted Suz12 using the inducible and conditional Rosa26-creERT2 (R26creERT2) mouse model. R26creERT2KI/+Suz12f/f mice injected with tamoxifen die within 2 weeks because of hematopoietic failure (personal communication, S. Lee to E. Michalak), thus precluding the use of this model for in vivo studies. Basal/MaSC-enriched (basal, CD29hiCD24+), committed luminal progenitor (CD29loCD24+CD14+), and mature luminal (CD29loCD24+CD14−) cell subsets were sorted from R26creERT2KI/+Suz12f/f and control mice and subjected to in vitro colony-forming assays to determine the effect of Suz12 deletion on the activity of mammary progenitor cells following induction by 4-hydroxytamoxifen (4OHT). Consistent with expression of Suz12 in all mammary epithelial subsets (Fig 2A), we observed fewer colonies in 2D colony forming assays on irradiated feeder cells (Fig 2B) upon induction of Suz12 deletion. In contrast to the hematopoietic system, in which loss of one allele of Suz12 [14] or Eed [23] enhances the activity of stem cells, deletion of one Suz12 allele did not affect the clonogenic activity of progenitor cells (Fig 2B). Unsorted MECs efficiently deleted Suz12 and yielded sufficient protein for western blot analysis (Fig 2C). As expected, Ezh2 protein was completely lost upon Suz12 deletion, while Ezh2 mRNA levels were not reduced (S3A Fig), suggesting rapid Ezh2 protein degradation leads to the absence of functional PRC2 upon Suz12 deletion. This is supported by the loss of H3K27me2 and H3K27me3 that accompanies Suz12 deletion (Fig 2C). Conversely, expression of the active H3K27me1 mark [4] did not change. No change in cleaved poly ADP ribose polymerase (PARP) was detected, suggesting that Suz12-deleted MECs do not undergo appreciable apoptosis. However, we observed increased expression of cyclin-dependent kinase p16 and p19 alternate reading frame (p19Arf), consistent with a known role for Ezh2/PRC2 in regulating cdkn2a gene expression (S3B Fig). These data suggest that Suz12 exerts an essential function in the mammary gland through maintenance of functional tri- as well as dimethylation of target loci.
To develop a system more amenable to molecular studies, we employed a 3D mammary organoid system, in which organoids grown from single cells in defined medium recapitulate many features of mammary tissue architecture and function [18]. Single basal or luminal progenitor cells sorted from R26creERT2KI/+Suz12f/f mice were plated, and deletion of Suz12 was induced by addition of 4OHT on day 1. Suz12 loss resulted in diminished numbers and smaller basal- and luminal progenitor–derived organoids over a 2 week culture period. Notably, PCR analysis of the resulting organoids revealed that selection against cre-mediated excision of both floxed alleles of Suz12 had occurred (S3C Fig). To circumvent this issue, small organoids were allowed to form before induction of deletion. Indeed, addition of 4OHT on day 4 resulted in smaller cystic-like organoids (Fig 3A and 3B) composed of Suz12-deleted cells (Fig 3C and S3D and S3E Fig) that had markedly reduced levels of H3K27me3 (S3E Fig). There were no notable differences in cleaved caspase 3 (CC3) in 2 week old Suz12-deleted organoids, but rather, a striking decrease in proliferative (Ki67+) cells was evident (Fig 3D). Consistent with the idea that cells lacking Suz12 are nonproliferative, repassaging resulted in the expansion of rare cells that had escaped Suz12 deletion (S3F and S3G Fig). Together, these data suggest that loss of Suz12 reduces the fitness and proliferation of mammary organoids.
To determine potential target genes that are dependent on Suz12 and therefore canonical PRC2 function, we performed RNA sequencing (RNA-seq) analysis of Suz12 Wt or Suz12-deleted basal- and luminal-derived mammary organoids grown from R26creERT2KI/+Suz12f/f mice. Suz12 was significantly down-regulated in both populations (Fig 4A, S4A Fig and S2 Data), and deletion of exon 5 was confirmed at the targeted locus (S4B Fig). In both basal- and luminal-derived organoids, the majority (>86%) of differentially expressed (DE) genes were up-regulated rather than down-regulated upon Suz12 deletion, consistent with the repressive role of PRC2 (Fig 4A, S4A and S4C Fig). A large overlap was apparent between up- and down-regulated DE genes (S4C Fig), as well as a strong correlation with gene expression changes (S4D Fig) in Suz12-deleted basal- versus luminal-derived organoids. This is consistent with a global effect of Suz12 deletion on PRC2 stability in MECs. Gene ontology (GO) enrichment analysis found that common DE down-regulated genes were significantly enriched for metabolic processes and H3K27 methylation, while common DE up-regulated genes were enriched for development and morphogenesis GO terms (S4C Fig).
To explore whether the changes in gene expression accompanying loss of Suz12 reflect alterations in chromatin compaction associated with loss of repressive H3K27me2/me3 marks, we performed global mapping of chromatin accessibility using ATAC-seq. Differential accessibility analysis identified 2,767 windows, of which 2,415 (87.3%) were more accessible upon Suz12 deletion (Fig 4B and S3 Data). Further, 1,377 of these windows overlapped with the transcription start site or gene body of 766 independent genes in Suz12-deleted basal organoids (referred to as differentially accessible [DA] genes from hereon). Nevertheless, the frequency of reads mapping to nucleosome-free regions versus mononucleosomes was not significantly altered. We observed only minor changes to insert length, suggesting no gross changes in nucleosome occupancy (S5A and S5B Fig), and a comparison of GC enrichment indicated a minor increased frequency of highly GC-rich regions, usually found at regulatory elements within promoters (Fig 4C). Moreover, model-based analysis of ChIP-seq (MACS) peaks analysis showed a small number of unique peaks upon Suz12 deletion (Fig 4D). Next, we compared the DA windows identified by ATAC-seq with gene expression from their associated gene to determine if changes in chromatin compaction elicited by Suz12 deletion were sufficient to induce changes in gene transcription. A strong correlation was evident between open chromatin and transcriptional up-regulation in Suz12-deleted basal-derived organoids (Fig 4E). However, analysis of Wt organoids did not reveal a correlation between genes assigned from ATAC-seq analysis and changes in their expression. Almost one-third (28%) of genes that became more accessible upon Suz12 deletion were significantly up-regulated by RNA-seq (Fig 4F). These 218 common genes included Cdkn2a (S6A Fig and S4 Data), a known target of Ezh2-mediated repression that encodes p16 and p19Arf, and transcription factors (TFs) involved in diverse lineage commitment, including Foxd1 and homeobox TFs (S6B and S6C Fig), and accordingly, represented sequence-specific DNA-binding GO terms (S5 Data). Notably, confocal imaging of organoids confirmed the persistence of basal and luminal cells in Suz12-deleted organoids (S6D Fig), suggesting lineage fidelity was maintained.
A similar ATAC-seq analysis was applied to Suz12-deleted luminal-derived organoids (S7A and S7B Fig), which revealed an increase in reads associated with nucleosomes (S5A and S5B Fig). These longer fragments were enriched in promoter-flanking and transcribed regions [24] consistent with increased expression of DE genes. Despite detecting many unique DA peaks (S7C Fig), particularly associated with promoters and intragenic/gene bodies (S7D Fig), the corresponding windows were less strongly associated with DE genes found by RNA-seq (S7E and S7F Fig).
Comparison of the Suz12-deleted gene expression signatures derived for both basal and luminal organoids showed that they were most similar to claudin-low and normal-like breast tumors (S7G Fig), similar to that observed for the basal/MaSC cell signature [25]. Likewise, the gene expression profile of Ezh2-deleted MECs was most concordant with claudin-low breast tumors [16]. The striking similarity between gene expression signatures of organoids lacking key PRC2 complex genes and claudin-low breast tumors—which exhibit a metaplastic, clinically aggressive phenotype in patients—suggests that PRC2 may play a role in tumor phenotype and behavior.
Using in vivo and in vitro studies, we show that deletion of a core component (either Suz12 or Eed) of the canonical PRC2 complexes leads to a complete block in mammary gland development in vivo and markedly curtails progenitor cell activity in vitro. Strikingly, Suz12- or Eed-deleted cells were strongly selected against and could not contribute to the developing mammary gland. This resulted in very small glands consisting of Wt cells. These findings are consistent with previous reports for nonredundant genes that are required for stem cell function [26]. Assays using primary MECs and organoids suggest that this is due in part to de-repression of cdkn2a, resulting in increased p16 and p19 expression and reduced proliferation owing to cell cycle arrest. Thus, previous studies [16,17] have underestimated the importance of PRC2 in governing mammary progenitor activity and differentiation, owing to some functional redundancy between Ezh1 and Ezh2. Moreover, these results are consistent with PRC2 promoting mammary epithelial expansion, rather than inhibiting it [27]. By exploiting mammary organoids to bypass the severe developmental phenotype associated with loss of Suz12, we examined changes in chromatin structure in the presence and absence of PRC2 and correlated these with gene expression changes.
PRC2 complexes are reported to colocalize with H3K27me3 on the promoters of around 10% of all genes [5]. Upon Suz12 deletion, we observed comparable changes in gene expression (7%–11%) in both basal- and luminal-derived organoids. There was significant overlap between the up-regulated genes in basal- and luminal-derived organoids, and the magnitude of change was highly correlated. Assessment of chromatin accessibility with ATAC-seq indicated that the chromatin of basal-derived organoids is more open upon Suz12 loss. Recent findings in human breast epithelial subsets [28] predict that bivalent promoters and primed enhancers would be the most affected by loss of H3K27me3 in Suz12-deleted basal-derived organoids. Since we did not observe a significant increase in GC content of DA reads identified by ATAC-seq, the mechanism is most likely to involve enhancers. Coupling RNA-seq with ATAC-seq analysis in basal-derived organoids revealed that deletion of Suz12 led to more than one quarter of genes becoming significantly transcriptionally active coincident with more accessible chromatin. We have previously shown that the MaSC/basal subset demonstrates the lowest levels of H3K27me3 across transcriptional start sites (TSSs) [16] while also producing slightly less RNA overall [29]. This observation supports the notion that this gene subset marked by H3K27me3 is transcriptionally regulated through H3K27me3-mediated repression accompanied by chromatin compaction and is dependent on canonical PRC2 function. Our results indicate that open chromatin is a good predictor of gene activation in basal-derived organoids. In the luminal compartment, there is an increased level of H3K27me3 [16], while these cells are more transcriptionally active overall [29]. In luminal-derived organoids, loss of PRC2 resulted in increased chromatin accessibility and an increase in reads mapping to regions with high GC content. In contrast to basal-derived organoids, the low correlation with gene expression in luminal-derived organoids suggests that additional mechanisms beyond chromatin accessibility serve to ensure proper gene repression. This may be due to DNA methylation or the combination of additional activating and repressive histone marks that maintain gene repression in this more committed cell type, as observed for human luminal cells [28].
Our studies indicate that PRC2 is responsible for both H3K27me3 and H3K27me2 marks in MECs, as evidenced by their loss in Suz12-deleted cells. Notably, we show here that the absence of these repressive marks through deletion of a single protein results in large changes in gene expression, consistent with the notion that chromatin states are the sum of a limited repertoire of PTM combinations on histone tails. While these data do not exclude noncanonical functions in the presence of a functional PRC2 complex [30], they suggest that Suz12 is a limiting factor for expression of Ezh2 protein and thus a functional PRC2 complex in the mammary epithelium.
In the context of breast cancer, our data indicate that Suz12-deleted organoids, like Ezh2-null basal cells [16], displayed similarities with signatures of claudin-low breast cancers, which are thought to arise from MaSCs. While Ezh2 up-regulation was initially associated with aggressive breast cancers, several studies now indicate that Ezh2 overexpression may be a consequence rather than a cause of breast cancer [13]. Indeed, deletion of Ezh2 accelerated tumors in a mouse model of Brca1-deleted breast cancer [31] and in a breast cancer model of Notch activation [13]. Additionally, the status of the cdkn2a locus of an individual tumor might predict its response to Ezh2 loss. The observation that Suz12 deficiency in normal mammary cells leads to de-repression of cdkn2a, paralleling that seen with loss of Ezh2 [32], would predict that loss of PRC2 function in breast cancer would result in decreased tumor proliferation via up-regulation of p16Ink4a and p14Arf expression. However, if this locus is perturbed, as is seen in many breast cancers [33], then loss of Ezh2 might have quite different consequences. Further work will be required to determine if loss rather than overexpression of PRC2 contributes to tumor progression in other cellular contexts.
In summary, our findings establish an essential role for PRC2 in the maintenance of mammary progenitor function and point to a critical function for PRC2 in maintaining chromatin states to ensure appropriate gene expression in MECs.
All animal experiments were conducted using mice bred at and maintained in our animal facility according to the Walter and Eliza Hall Institute of Medical Research Animal Ethics Committee guidelines, approval number 2017.002.
Suz12f/f [34], Eedf/f [35], Ezh2f/f [36], MMTV-cre (line A) [37], and R26creERT2 [38] gene-targeted mice have been described previously. CD4cre-Eedf/f T lymphocytes [39] were a gift from R. Allan. MMTV-cre mice were maintained as a pure strain on an FVB/N background, R26creERT2 mice were on a C57BL/6 background, and Ezh2, Eed, and Suz12 conditional knockout mice were on either a C57BL/6 or mixed FVB/N and C57BL/6 background. For analysis of postnatal day 1 lungs, adult female mice were subjected to timed pregnancies and injected with 0.2 mg progesterone on days 17 and 18 of pregnancy to delay parturition. At 19.5 days of pregnancy, pups were recovered by caesarian section and monitored for breathing for 1 hour with constant stimulation before collection of lungs into Bouin’s solution.
Primary mouse MEC cultures were isolated from both inguinal and/or thoracic mammary glands and prepared as described [17]. MEC suspensions and flow cytometry were performed as previously described [40]. Antibodies against mouse antigens were purchased from Biolegend and included CD24-PB (#101820), CD31 (#102410), CD45 (#103112), and Ter119 (#116212) conjugated to APC, CD29–FITC (#102206), and CD14-PE (#123309). Cells were sorted on a FACSAria or FACSDiva (BD PharMingen) and manually counted prior to plating on irradiated fibroblast feeders as described [40]. After 7 days, colonies were fixed and stained with Giemsa and manually counted. To induce R26creERT2-mediated deletion, 0.1 μM 4OHT was added to culture medium for 20 hours on day 1 of culture.
Organoids were cultured as described previously [18] in 8 μl BME (Cultrex) drops (70 basal and 60 luminal progenitor cells per drop) on nontreated 24 well plates and covered by Advanced DMEM/F12 supplemented with growth factors excluding Wtn3a or FGF2. Rock inhibitor (Y-27632) was added to culture medium for the first 4 days, and the medium was refreshed every 2–3 days. To induce R26creERT2-mediated deletion, 0.1 μM 4OHT was added to culture medium for 16–20 hours on day 1 or day 4 of culture. Organoids were photographed using best focus projection images on the Nikon TiE System software after 12–14 days in culture. Prior to proliferation measurements, RNA-seq, or ATAC-seq analysis, 12–16 day old organoids were dissociated into single cells using TrypLE express (Thermo Fisher Scientific). Cell suspensions were mixed 1:1 with CellTiter-Glo Luminescent Cell Viability Assay (Promega) prior to detection of luminescence. Genomic DNA was extracted from organoids using DNeasy Blood and Tissue kit (Qiagen) and PCR used to distinguish the Wt, floxed, and recombined (deleted, referred to as “del”) Suz12 alleles as described [34].
Mice were injected with BrdU Cell Labeling Reagent (0.5 mg/10 g body weight, Amersham Biosciences) 1.5 hours prior to collection. For histology, tissues were fixed in 4% paraformaldehyde overnight and embedded in paraffin. Sections (5 μm) were stained with hematoxylin–eosin (HE). For whole-mount analysis, mammary glands were harvested and fixed in Carnoy’s solution (6:3:1 of 100% ethanol, chloroform, and glacial acetic acid) and stained with Carmine alum. The extent of ductal outgrowth was measured on inguinal whole mounts as the distance from the center of the lymph node to the leading edge of the ductal mass.
IHC and IF were performed as described [16]. Paraffin-embedded sections (5 μm) were dewaxed in xylene and rehydrated through an alcohol series, blocked with 3% hydrogen peroxide, and subjected to antigen retrieval by boiling in 10 mM citrate buffer pH 6.0 for 30 seconds using a DAKO pressure cooker. Immunostaining was performed using the streptavidin-biotin peroxidase detection system as per the manufacturer’s instructions (ABC reagent, Vector Laboratories) and 3,3-diaminobenzidine was used as substrate (DAKO). In all cases, an isotype-matched control IgG was used as a negative control. The following antibodies were used: anti-BrdU (Bio Rad OBT0030), anti-Ezh2 (BD Biosciences #612667), anti-ERα (Santa Cruz sc-543), anti-PR (Santa Cruz sc-538), anti-Foxa1 (Abcam Ab23738), anti-H3K27me3 (Millipore #07–449), anti-Suz12 (Diagenode pAB-029-050), anti-Eed (R&D Systems AF5827-SP), anti-CC3 (Cell Signaling #9664L), and anti-Ki67 (Cell Signaling #12202S). Secondary antibodies were biotin-conjugated anti-rabbit IgG, anti-rat IgG and anti-mouse IgG (Vector Laboratories), anti-mouse alexa-488 (Invitrogen), and anti-rabbit alexa-647 (Invitrogen), and DAPI was used to detect nuclei (Thermo-Fischer Scientific). As described in [18], 3D imaging of organoids was performed on a SP8 Confocal microscope after staining with anti-cytokeratin 14 (Thermo-Fisher Scientific) and anti-cytokeratin 8/18 (TROMA-I, DSHB) antibodies.
Lysates were prepared in RIPA buffer [16], and western blotting was performed as described [17]. The following antibodies were used for western blot analysis: anti-Ezh2 (BD Biosciences #612666), anti-Suz12 (Cell Signaling 3737S), anti-Eed (Millipore 05–1320), anti-H3K27me1 (Millipore 07–448), anti-H3K27me2 (Millipore 07–452), anti-H3K27me3 (Millipore 07–449), anti-Histone 3 (Millipore 07–690), anti-β-actin (Sigma A5441), anti-Gapdh (Sigma G8795), anti-ERα (Millipore 07–690), anti-Ezh1 (Millipore 07–690), anti-PARP (Cell Signaling #9542), anti-p16 (Santa Cruz M-156), and anti-p19 (Rockland Immunochemicals #200-501-891). Secondary antibodies included HRP-conjugated anti-rabbit and anti-mouse (Southern Biotech), both 1:10,000.
ATAC-seq was performed as described [24] with the following adaptations. Organoid cultures were dissociated, and 50,000 single cells were lysed and nuclei collected at 500 g for 10 minutes in lysis buffer containing 0.1% NP-40. Pelleted nuclei were tagmented with Nextera Tn5 Transposase (TDE1, Illumina FC-121-1030) for 20 minutes at 22 °C and 30 minutes at 37 °C. Transposed DNA was purified using Qiagen MinElute kit (28204) and fragments PCR amplified as described [24]. ATAC libraries were sequenced on a NextSeq using a 150H kit with 75 bp paired-end reads, and total reads were collated from 2 runs to obtain approximately 60 × 106 reads per sample. Sequencing runs were pooled for each replicate and sample, and adapters were trimmed with Trim Galore! (http://www.bioinformatics.babraham.ac.uk/projects/trim_galore/) and mapped to the mm10 mouse genome using bowtie2 [41], allowing for fragments <2,000 bases in length. Mitochondrial and duplicate reads were removed using Picard-Tools (http://broadinstitute.github.io/picard/), and bam files were run through macs2 (doi: 10.1186/gb-2008-9-9-r137) for peak calling using these parameters: “—nomodel—shift—75—extsize 150—qvalue 0.05.” For DA analysis, bam files were loaded in SeqMonk (v1.37.1, https://www.bioinformatics.babraham.ac.uk/projects/seqmonk/), and probes were created using a tiling approach with 150 bp windows end to end across the genome. Raw counts were then processed through the differential gene expression pipeline within edgeR [42], and DA regions were called with an exact test and FDR < 0.05. Reads were quantified by reads per million and log2 transformed for visualizing graphically. For each gene, the 5 Kb region upstream of intron 1 was used in combination with the gene body to define the TSSs plus gene body. GC enrichment was quantified on mapped bam files using DeepTools [43] computeGCBias with an effective genome size of 2,150,570,000. These data have been deposited in the Gene Expression Omnibus (accession code GSE116431).
Total RNA was extracted from basal- or luminal-derived organoids grown from single sorted luminal or basal populations from the mammary glands of R26creERT2/Suz12f/f female mice using the RNeasy Mini kit (Qiagen). Two biological replicates were prepared of the basal-derived organoids and 3 biological replicates of luminal-derived. RNA-seq was carried out on an Illumina Nextseq 500 to produce 20–65 million 80 bp reads per sample. Read pairs were mapped to the mouse genome (mm10) using the subread aligner [44] implemented in the Rsubread software package. Read counts for Entrez Genes were obtained using featureCounts [45] and its inbuilt mm10 annotation. Gene information was downloaded from the NCBI on 1 February 2017. Statistical analysis used the limma [46] and edgeR [42] software packages. Genes with at least 0.5 read counts per million (cpm) in at least 2 samples were considered to be expressed and were kept in the analysis. Immunoglobulin receptor segments, ribosomal genes, predicted and pseudogenes, and obsolete Entrez IDs were filtered out. Trimmed mean of M-values (TMM) scale normalization [47] was applied, and read counts were transformed to log2-cpm with a prior count of 3. Linear models were used to test for expression differences between 4OHT treated versus untreated samples from luminal cells and from basal cells. Each organoid sample was treated as a random block, allowing for correlation between repeats [48]. Differential expression was assessed using the Treat method [49], computing empirical Bayes moderated t statistics relative to a fold change threshold of 1.5, and allowing for an abundance trend in the standard errors and for robust estimation of the Bayesian hyperparameters [50]. The Benjamini and Hochberg method was used to control the FDR. These data have been deposited in the Gene Expression Omnibus (accession code GSE116431).
Microarray expression profiles of breast tumors were downloaded from Gene Expression Omnibus series GSE18229 [51]. Probe intensities were normexp background corrected with offset 50 [52] and loess normalized using the limma package. Mouse Entrez Gene IDs were mapped to human using HUGO Gene Nomenclature Committee orthology predictions downloaded November 2016. Suz12-deficient expression signatures were computed for each tumor using a previously described method [25]. Briefly, the sum of products of RNA-seq log2-fold-changes with microarray log2-normalized intensities was computed for all genes DE in the Suz12 deficient basal- or luminal-derived organoids.
For qRT-PCR analysis, total RNA from MECs was reverse-transcribed using Superscript III (Invitrogen), and cDNA were analyzed on a LightCycler 480 (Roche). Input cDNA concentrations were normalized to GAPDH. Product accumulation was evaluated using the comparative Ct method (2−ΔΔCT). Primer sequences were: Ezh2 For: 5′-ACATCCCTTCCATGCAACACC-3′; Ezh2 Rev: 5′-TCCCTCCAGATGCTGGTAACA-3′; Eed For: 5′-GTGAACAGCCTCAAGGAAGAT-3′; Eed Rev: 5′- ATAAGGTTACTCTGTGCTTC-3′; Gapdh For: 5′-TGACATCAAGAAGGTGGTGAAG; Gapdh Rev: AAGGTGGAAGAGTGGGAGTTGC-3′.
|
10.1371/journal.pgen.1006027 | Chloroplast RNA-Binding Protein RBD1 Promotes Chilling Tolerance through 23S rRNA Processing in Arabidopsis | Plants have varying abilities to tolerate chilling (low but not freezing temperatures), and it is largely unknown how plants such as Arabidopsis thaliana achieve chilling tolerance. Here, we describe a genome-wide screen for genes important for chilling tolerance by their putative knockout mutants in Arabidopsis thaliana. Out of 11,000 T-DNA insertion mutant lines representing half of the genome, 54 lines associated with disruption of 49 genes had a drastic chilling sensitive phenotype. Sixteen of these genes encode proteins with chloroplast localization, suggesting a critical role of chloroplast function in chilling tolerance. Study of one of these proteins RBD1 with an RNA binding domain further reveals the importance of chloroplast translation in chilling tolerance. RBD1 is expressed in the green tissues and is localized in the chloroplast nucleoid. It binds directly to 23S rRNA and the binding is stronger under chilling than at normal growth temperatures. The rbd1 mutants are defective in generating mature 23S rRNAs and deficient in chloroplast protein synthesis especially under chilling conditions. Together, our study identifies RBD1 as a regulator of 23S rRNA processing and reveals the importance of chloroplast function especially protein translation in chilling tolerance.
| Compared to cold acclimation (enhancement of freezing tolerance by a prior exposure to low non-freezing temperature), the tolerance mechanism to non-freezing chilling temperatures is not well understood. Here, we performed a genome-wide mutant screen for chilling sensitive phenotype and identified 49 candidate genes important for chilling tolerance in Arabidopsis. Among the proteins encoded by these 49 genes, 16 are annotated as having chloroplast localization, suggesting a critical role of chloroplast function in chilling tolerance. We further studied RBD1, one of the four RNA-binding proteins localized to chloroplast. RBD1 is only expressed in the green photosynthetic tissues and is localized to nucleoid of chloroplasts. Furthermore, RBD1 is found to be a regulator of 23S rRNA processing likely through direct binding to the precursor of 23S rRNA in a temperature dependent manner. Our study thus reveals the importance of chloroplast function especially protein translation in chilling tolerance at genome-wide scale and suggests an adaptive mechanism involving low temperature enhanced activities from proteins such as RBD1 in chilling tolerance.
| Low temperature inhibits plant growth in general and limits the geographical distribution of plants. Earlier studies have identified numerous physiological and cellular changes associated with chilling (more than 0°C) or freezing (less than 0°C) conditions, such as alterations in membrane composition, calcium signals, metabolite composition, photosynthesis, and protective molecules [1,2]. Most of these changes are thought to help plants to cope with low temperature stresses. Plants differ in their abilities to tolerate chilling stresses. Low temperature often inhibits photosynthesis and reduces carbon uptake and allocation to developing sink tissues [3,4]. Many tropical and subtropical plants including maize, rice, and tomato do not survive at 4°C because they cannot undergo photosynthesis and carbon metabolism under this condition. Arabidopsis, as well as some overwinter cereals, can grow at the low temperature due to their biochemical and physiological adaptations which may include acclimation of photosynthetic metabolism [5,6].
Translation in chloroplast appears to be especially sensitive to chilling stresses. Chilling slows down protein biosynthesis in plastids by eliciting frequent ribosome pausing in tomato [7]. The otherwise chilling tolerant Arabidopsis plants become chilling sensitive when they are defective in chloroplast ribosomal biogenesis and RNA processing [8–12]. For instances, loss of the translation elongation factor SVR3, the rRNA maturation factor NUS1, and chloroplast RNA binding proteins CP29A and CP31A, all lead to increased chilling sensitivity through affecting chloroplast biogenesis [9,10,13]. In addition, the loss of chloroplast ribosomal subunits reduces the ability of plants to recover from prolonged chilling periods [11,12].
Chloroplast function is carried out by genes coded mostly by the chloroplast genome [14]. Transcription of such genes relies on two plastid RNA polymerases: nucleus-encoded RNA polymerase (NEP) and plastid-encoded RNA polymerase (PEP) [15,16]. Chloroplast RNAs need to be processed to become functional rRNAs and mRNAs. Many of the processing factors for RNA cleavage, splicing, editing or stability are RNA-binding proteins [13,17–19]. They are all coded by the nuclear genome. One family has pentatricopeptide repeats (PPR) and it usually carries out specific RNA processing in chloroplasts [19]. Another family contains RNA recognition motif/RNA-binding domain/ribonucleoprotein (RRM/RBD/RNP) domain and these proteins, referred to as RNPs, are suspected to regulate larger sets of RNAs [20]. Among the chloroplast RNPs, CP31A and CP29A are associated with a large pool of chloroplast transcripts and influence their stability, processing, and splicing [13].
While chilling tolerance mechanism is not well understood, cold acclimation, an enhancement of freezing tolerance by a prior exposure to low non-freezing temperature, has been intensively studied in Arabidopsis thaliana [21–25]. Cold acclimation is mediated in part by C-repeat binding factors (CBFs) which regulate the expression of a large number of Cold Responsive (COR) genes, some of which are thought to confer freezing tolerance. The upregulation of the CBF genes by low temperature is critical for this acclimation and is mainly modulated by ICE1 (Inducer of CBF expression 1), ICE2 and the three closely related CAMTAs (Calmodulin binding Transcription Activators) [26–29]. Although genes regulated by CBF proteins play vital roles in freezing tolerance, they only represent a small percentage of the COR genes. CBF independent regulations of cold acclimation or freezing tolerance have been found in Arabidopsis [30,31]. It is yet to be determined whether or not chilling tolerance and cold acclimation have shared mechanisms.
To have a better understanding of chilling tolerance in Arabidopsis, we carried out a chilling sensitive mutant screen on all available T-DNA insertion mutants that represent half of the total Arabidopsis genes. Interestingly, mutants defective in chloroplast localized proteins are overrepresented, indicating the importance of chloroplast function in chilling tolerance. Detailed characterization of one such mutant of a RNA binding protein supports a critical role of chloroplast translation in chilling tolerance.
In order to identify genes important for chilling tolerance in Arabidopsis at the genome scale, we analyzed 11,000 T-DNA insertion mutants covering half of the total Arabidopsis genes for chilling sensitive phenotypes. Each of them is putative homozygous T-DNA knockout or knockdown mutants in indexed genes generated by the SALK Institute [32] and available from the Arabidopsis Biological Resource Center (http://abrc.osu.edu/). The use of these indexed knockout or knockdown lines over chemical mutagenized population enabled the assessment and verification of phenotypes under different growth conditions and allowed uncovering of conditional lethal mutants and avoiding temperature-sensitive alleles of essential genes. In this screen (S1 Fig), four mutant seeds of each line were germinated and grown side by side with the wild-type Col-0 on vertical plates under a 16 hours (h) light/day photoperiod first at 22°C for 8 days and then transferred to 4°C for two months. Lines showing abnormal growth phenotypes compared to the wild type only at 4°C but not 22°C were selected as chilling sensitive mutants. These phenotypes include albino, yellow or purple leaves, abnormal leaf shapes, smaller leaves, and shorter roots. Because not all of these T-DNA lines were homozygous for the T-DNA insertion mutations, we further analyzed those lines where not all four seedlings exhibited mutant phenotypes to exclude those whose mutant phenotypes are not correlated with the T-DNA insertion. From this screen, a total of 54 lines showed growth defects at 4°C but not 22°C and we defined these mutants as chilling sensitive (Fig 1A).
We then examined the genes indexed to be disrupted by T-DNA insertions in these lines. Those with insertion in the promoter region but not exon, intron, or UTRs were removed as they might not affect the function of the indexed gene and the phenotype is unlikely due to a defect in the gene. This leaves us with 49 chilling sensitive mutant lines where the function or expression of the indexed genes are disrupted. None of the 49 genes are in the characterized CBF cold acclimation pathway, and loss of function mutants of CBF1, CBF2, CBF3, ICE1, HOS1 and SIZ1 were not in the collection. Although causal genes for chilling sensitivity may not be the indexed gene in a small proportion of the lines, we decide to analyze functional categories of these candidate genes as a group to reveal potential important chilling tolerance mechanisms. Among proteins encoded by the 49 candidate genes, 16 were annotated as localized to the chloroplast (Table 1), 10 to the nucleus, 6 to the cytosol, 4 to the mitochondria, and the rest either to other locations or without localization information. There is thus a high representation of chloroplast related genes in these mutants, suggesting a critical role of chloroplast function in chilling tolerance.
Among the 16 proteins with chloroplast localization, 4 were chloroplast RNP proteins and three were previously characterized: ORRM1 (AT3G20930), CP29A (AT3G53460) and CP31A (AT4G24770). Mutants of CP29A and CP31A were chilling sensitive and defective in processing various RNAs in chloroplast [13]. ORMM1 is required for plastid RNA editing in Arabidopsis and maize [33], but its role in chilling tolerance has not been analyzed. The other RRM/RBD/RNP coding gene, AT1G70200, was not characterized before, and we named it RBD1. Mutants of all of these 4 genes showed bleaching in newly emerging leaves at 4°C. A mutant of another RNP coding gene AML1 (AT5G61960) also had a bleaching phenotype. AML1 is predicted to localize in the nucleus but it may also have a chloroplast localization signal from the e-FP data. It plays a role in meiosis as well as in vegetative growth in Arabidopsis thaliana [34], but its role in chilling tolerance was not analyzed. We therefore further analyzed chilling sensitive phenotypes of mutants of these 5 RNP coding genes. When grown on soil at the normal temperature 22°C, the rbd1, cp29a and cp31a mutants exhibited a wild-type phenotype, while orrm1 and aml1 had a smaller size than the wild type Col-0 (Fig 1B). The aml1 had yellow cotyledons at germination but recovered in a few days, and it often had white strips on the leaves. Consistent with visual phenotypes, the rbd1, cp29a and cp31a mutants has a fresh weight similar to that of the wild type at 22°C, but orrm1 and aml1 mutants displayed a 35% and 58% reduction compared to the wild type (Fig 1C). When shifted to 4°C after grown at 22°C for three weeks, new leaves emerged from all of these mutants were yellow (Fig 1B). When these mutants were grown constantly at 4°C in soil from germination, all their leaves were yellow and the plants were much smaller than the wild type with only 10% to 35% of the wild-type fresh weight (Fig 1C).
We chose RBD1 for further analysis because there were no prior reports on this gene. The RBD1 gene encodes a protein of 538 amino acids (aa) containing one RNA recognition motif/ RNA-binding domain (Fig 2A). We confirmed that the chilling sensitive phenotype observed in the mutant line SALK_041100 is due to the loss of the RBD1 function through analysis of additional RBD1 mutants and complementation test. The SALK_041100 line (named rbd-1) has the T-DNA inserted in the third exon of the RBD1 gene, while the T-DNA insertion line SALK_012657 (named rbd1-2) has an insertion in the 5’-UTR of RBD1 (Fig 2B). Both mutants were loss-of-function mutants of RBD1 as no full length transcripts could be amplified by Reverse Transcription (RT)-PCR (Fig 2C). The rbd1-2 mutant, like rbd1-1, produced yellow emerging leaves at 4°C (Fig 2D). In addition, RNA interference (RNAi) was used to reduce the expression of RBD1, and 13 of 15 RNAi transgenic lines produced yellow leaves under chilling conditions (Fig 2D). Furthermore, when the full-length RBD1 cDNA driven by the cauliflower mosaic virus (CaMV) 35S promoter was transformed into the rbd1 mutants, the chilling sensitive phenotype was suppressed in 30 of 32 the transgenic lines (Fig 2D). Therefore, the loss of RBD1 function does confer chilling sensitivity.
We characterized the growth phenotypes of the rbd1-1 mutants in more detail. When plants were shifted from three weeks’ growth at 22°C to 4°C, the yellowing or bleaching became visible only after 2 weeks of chilling treatment and only happened in leaves emerged at 4°C. There is a gradient of yellowing decreasing from the youngest leaf to the oldest leaves (Fig 2D). This yellowing phenotype is reversible, when plants were moved from 4°C back to 22°C, the pale green phenotype disappeared in 2 days (Fig 2E). The rbd1-2 mutant also had a lighter green appearance compared to the wild type when grown constantly at 22°C. Using spectrophotometer, we found that the two rbd1 mutants had a 10–20% reduction of chlorophyll a, chlorophyll b and carotenoids compared to the wild-type Col-0. This phenotype was not observed in the complemented lines (Fig 2F). Therefore, the rbd1 mutants are compromised in chloroplast function especially under chilling conditions.
Because the rbd1 mutants displayed chilling sensitivity, we analyzed whether chilling induction of the CBF and COR genes was altered in the rbd1 mutants. CBF1, CBF2, and CBF3 are induced by 4°C treatment at 6 hours to the same extent in Col-0 and rbd1-1 (Fig 3A). Similarly, these genes fall to the basal level after 3 weeks of 4°C treatment in both the wild type and the mutant (Fig 3A). COR15A, COR47, and KIN1 genes are regulated by the CBF genes, and they were induced by 4°C treatment at 6 hours and stayed elevated during the subsequent chilling treatment for four weeks (Fig 3B). The rbd1-1 mutant had the same induction kinetics and amplitudes of these three COR genes by 4°C treatment (Fig 3B). Therefore, the rbd1-1 mutant is not impaired in its capacity to induce the CBF or COR expression by 4°C and the chilling sensitivity of the mutants is likely independent of the CBF pathway.
The RBD1 gene is expressed in green tissues. Using RT-PCR analysis, RBD1 mRNA was detectable in leaves, stems, flowers and siliques but not in roots (Fig 4A). The expression pattern of RBD1 was further investigated in transgenic lines containing a ß-glucuronidase (GUS) reporter gene fused to the RBD1 promoter. GUS expression was observed in all green tissues throughout development (Fig 4B). In 3-day-old seedlings, GUS staining was primarily detected in emerging cotyledons (Fig 4B). In 14-day-old seedlings, it was detected in most of the plant but not in roots, and was particularly strong in cotyledons (Fig 4B). GUS staining was also observed in the stems and siliques, but not in the matured seeds (Fig 4B). Within the flower, it was strong in the sepals and carpels, but not in the stamens and petals (Fig 4B). In addition, older green tissues usually had higher expression levels than younger green tissues, suggesting that RBD1 might contribute to the function of mature green tissues. Altogether, these expression patterns show that RBD1 is highly expressed in the photosynthetic tissues, which was in agreement with its putative roles in chloroplast function.
RBD1 is not induced by light according to the expression data from Arabidopsis eFP Browser (http://bbc.botany.utoronto.ca). It is not induced by chilling treatment either. RT-PCR revealed that RBD1 is expressed at the same level during the chilling treatment from day 0 to day 21 (Fig 3C). The RBD1 gene therefore might be constitutively expressed at the transcript level.
RBD1 is annotated as a chloroplast targeted protein, which could explain its expression in green tissues. To verify the subcellular localization of the RBD1 protein, we expressed GFP (Green fluorescent Protein) fusion of RBD1 under the 35S promoter (RBD1:GFP) in protoplasts isolated from wild-type Arabidopsis leaves and monitored fluorescence by confocal microscopy. While the GFP protein alone was present in the cytoplasm and nucleus, the RBD1:GFP fusion protein was found exclusively in chloroplasts (Fig 4C). In addition, the RBD1:GFP fusion protein was dispersed as small fluorescent particles within chloroplast (Fig 4C), which is reminiscent of nucleoid localization [35]. This localization pattern does not appear to be temperature-dependent. In both 22°C and 4°C, RBD1:GFP is only present in the nucleoid structure in chloroplast (S2 Fig).
The nucleoid localization pattern and the yellowish leaf mutant phenotype promoted us to investigate the role of RBD1 gene in chloroplast RNA processing because nucleoid is the site of rRNA processing and ribosome assembly [36]. We first analyzed total RNAs separated on formaldehyde denaturing agarose gel stained by ethidium bromide. Plastid rRNAs (23S, 16S, 5S, and 4.5S) and cytosolic rRNAs (25S, 18S, 5.8S, and 5S) can be easily observed on such gels because they are very abundant [37]. When the same amount of total RNAs was loaded on gel, rbd1-1 and rbd1-2 mutants showed significant reduced accumulation of the 1.1 and 1.3-kb species of the 23S rRNA compared to the wild type under chilling stress conditions, but exhibited almost the same amount of rRNAs as wild type at 22°C (Fig 5A).
RNA gel blot hybridization was used to further analyze the defects in 23S rRNA processing. The 23S rRNA with a full length of 2.9-kb is cleaved internally at two sites, yielding fragments of 0.5, 1.1, and 1.3-kb in wild type [38]. With a probe detecting the 3’-end fragments, we found that 4°C treatment did not affect the processing of 23S rRNA transcripts in the wild type (Fig 5B). In contrast, the chilling treated rbd1 mutants accumulated the partially processed 2.9-kb and the 2.4-kb of 23S rRNA at a much higher level than the wild type, while the fully processed product of 1.1-kb rRNA was greatly reduced (Fig 5B).
Consistent with a defect in processing of 23S rRNA which is a component of the chloroplast ribosome, the rbd1 mutants had a reduction in chloroplast translated protein. When total leaf proteins were analyzed on a SDS polyacrylamide gel, the chloroplast Rubisco Large Subunit (RbcL) protein, which is synthesized in chloroplast, was found to have a drastic reduction in rbd1-1 and rbd1-2 mutants compared to the wild type by 4°C treatment but not at 22°C (Fig 5C). We further analyzed the chloroplast translation efficiency in the rbd1 mutants. Spectinomycin is a chemical inhibitor of chloroplast translation as it prevents translocation of the peptidyl-tRNA from the A site to the P site on the 30S subunit of 70S ribosomes [39]. When seeds were sowed and grown on 1/2 MS supplied with spectinomycin at a concentration of 3mg/L at 22°C, both the rbd1-1 and rbd1-2 mutants exhibited total yellowing while the wild type and the complemented rbd1 mutants maintained some green tissues (Fig 5D). This suggests that translation in chloroplast is compromised in the rbd1 mutants even under non-chilling condition.
We further analyzed the transcripts of 9 chloroplast RNAs in the rbd1 mutants under normal and chilling conditions to assess how broad a role the RBD1 gene might play in chloroplast RNA regulation. They are 16S rRNA, PEP transcribed ndhF, psaA, rbcL, psbB, psbF and petB, NEP transcribed ycf3, and PEP and NEP transcribed rps4. Among these, 16S rRNA, ndhF, psbB, petB, ycf3 and rps4 need to be processed to become mature transcripts.
In contrast to that of 23S rRNA, the processing of 16S rRNA was only slightly affected in the rbd1 mutants compare to Col-0 at 4°C but not 22°C (S3B Fig). The transcripts for ndhF, psaA and rbcL showed decreased levels under chilling stress compared to the wild-type plant, but displayed similar levels to wild type at 22°C (S3A Fig). For psbB, psbF and petB, low temperature down regulated their transcripts levels, but the transcripts were at similar levels in the rbd1 mutants and wild type at 22°C and 4°C (S3B Fig). The transcripts of ycf3-ex2 and rps4 showed an over-accumulation under chilling condition in the rbd1 mutants compared to the wild type, but exhibited the same levels at 22°C (S3C Fig). In all, the loss of RBD1 mainly alters processing of 23S rRNAs but not other chloroplast RNAs we analyzed. It also reduces the expression level of some PEP transcribed genes, but increases the expression of some NEP transcribed genes at chilling temperature.
The leaf yellowing phenotype in the rbd1 mutants likely results from an accumulative defect over time because it only became visible two weeks after plants were shifted from 22°C to 4°C. To identify earlier defects in the mutants, we monitored the molecular events after the plants were shifted from 22°C to 4°C at 0 hour (h), 6 h, 1 day (d), 3 d, 7 d, 14d, 21 d and 28 d. Among the 6 genes that showed altered processing and expression in the mutants, 23S rRNA was the first to show defects. Its processed 1.1-kb transcript was significantly lower than the wild type after 7 days of 4°C treatment (S4A Fig). The transcript of psaA showed defects in the rbd1 mutants after 2 weeks of cold treatment (S4B Fig). The transcripts of the rest ndhF, rbcL, ycf3-ex2 and rps4 genes began to show significant change after 3 or 4 weeks of chilling treatment (S4 Fig). Therefore, the rbd1 mutants exhibit a defect in 23S rRNA processing very early on (S4B, S4C and S4D Fig), which would lead to reduced chloroplast translation and bleaching phenotype under chilling conditions.
To determine whether or not RBD1 protein is associated with the 23S rRNA and the association is temperature regulated, we performed RNA co-immunoprecipitation (IP) assay at both 22°C and 4°C. RBD1 was fused to the GFP (RBD1:GFP) and expressed under the strong 35S promoter in Arabidopsis protoplasts. As a control, the signal peptide of ORRM1 that targets the protein to chloroplast was fused with GFP (sORRM1:GFP) for protoplast expression as well. After transformation, the protoplasts were incubated at 22°C for 8 h before being split for further incubation at 22°C and 4°C respectively for 12 h to assess temperature dependency of binding. Total proteins from protoplasts were then IPed by the anti-GFP antibodies, and the co-IPed RNAs (after DNase I treatment) were reverse transcribed and detected by quantitative Real Time-PCR. We analyzed four transcripts: 23S rRNA, 16S rRNA, psbF, and rbcL and found binding only to 23S rRNA (Fig 6). Three regions of the 23S rRNA precursor were analyzed: 5’ end, middle part, and 3’ end, residing in three different processed RNAs. There was an 8, 11 and 6 fold increase of PCR products for the three regions of 23S rRNA precursor in the RBD1:GFP sample compared to the sORRM:GFP sample at 22°C (Fig 6A). Furthermore, the fold increase for the three regions were 23, 39 and 19 fold in the RBD1:GFP sample compared to the sORRM:GFP at 4°C (Fig 6B). Therefore, RBD1 binds to three regions of 23S rRNA precursor, and the binding is at least two times more at 4°C than at 22°C. In contrast, no difference was detected for the 16S rRNA, psbF and rbcL between RBD1:GFP sand sORRM:GFP at either 22°C or 4°C (Fig 6A and 6B), suggesting a specificity of transcript binding for RBD1.
We further compared the molecular defects among the 5 RNP mutants after 4 weeks of 4°C treatment. The rbd1 mutants had the most drastic defect in 23S rRNA processing among the 5 RNP mutants. By RNA gel staining, reduction of processed 23S rRNAs compared to the wild type was the most pronounced in the rbd1-1 mutant (Fig 7A). By RNA blotting, the processed 1.1-kb transcript was found to decrease in all of these 5 RNP mutants, with the most severe reduction happening in rbd1-1. The partially processed 2.9- and 2.4-kb transcripts were over-accumulated in these mutants at 4°C except for orrm1 (Fig 7B). Among the 5 RNP mutants, the rbd1-1 mutant exhibited the highest ratio of partially processed transcripts (2.4- and 2.9-kb) over processed transcripts (1.1-kb). Consistent with the observed 23S rRNAs processing defect in the 5 RNP mutants, the accumulation of RbcL protein at 4°C was reduced to different extent, with an approximately 20% to 40% reduction compared to the wild type (Fig 7C). In contrast to the RbcL protein, TOC75, a cytosol synthesize chloroplast outer membrane protein [40–42], showed an increased amount of 20% to 40% at 4°C in the 5 RNP mutants compared to the wild type (Fig 7D). In addition, rbd1-1 showed the highest sensitivity to spectinomycin followed by aml1 among the 5 RNP mutants assayed (Fig 7E). At 3mg/L concentration, all rbd1-1 seedlings turned white and the aml1 mutant was pale yellow, while the other three mutants stayed green similarly to the wild type. These results indicate that loss of RBD1 has a larger effect on 23S rRNA processing compared to other mutants, which may impact chloroplast translation more than the others. It also shows that the ORRM1 mutation has the least impact on rRNA processing.
In this genome-wide chilling sensitivity screen, we identified 54 T-DNA insertion lines that exhibited a chilling sensitive phenotype in Arabidopsis. Although we have not tested if all results from disruption of the gene that the T-DNA is inserted in, we expect that most of them are. Interestingly, a large proportion of these genes are related to chloroplast function, indicating that it is one of the weakest links in chilling tolerance. Earlier studies have implicated a strong association of compromised chloroplast function with chilling tolerance, and this study supports this notion at the genome wide scale. Wild-type Arabidopsis plants usually survive at low temperatures such as 4°C but mutants with compromised chloroplast function are sensitive to chilling similarly to the subtropical and tropical plants. To what extent difference in chilling tolerance among different plants is due to difference in chloroplast function at low temperatures is worth studying in the future. This screen also suggests that chilling tolerance and cold acclimation may not use identical mechanisms in Arabidopsis. Mutants of the known key regulators of the CBF pathway, which are unfortunately not present in the collection, are not known or reported to have chilling sensitive phenotypes. Mutants identified in the chilling sensitive screen here were not isolated as defective in the CBF pathways mutants. Indeed, the rbd1-1 mutant is not defective in cold induction of the CBF pathway, suggesting that the induction of the CBF pathway may not require optimal chloroplast function and chilling sensitivity can result from CBF independent defect. Nevertheless, chilling tolerance can be connected with the CBF pathway. In Arabidopsis, a chilling sensitive mutant crlk1 (calmodulin receptor like kinase 1) is chilling sensitive and is also delayed in CBF induction [43]. In rice, a chilling sensitive mutant cold1 is defective in CBF induction [44]. It is possible that chilling sensitivity can result from defects in multiple processes and some might be shared with the CBF pathway.
The chloroplast localized proteins critical for chilling tolerance are involved in multiple aspects of chloroplast function, including transcript regulation and maturation, chloroplast development, protein transportation and secretion, as well as photosynthesis. For instance, YS1 is required for editing of rpoB transcripts and chloroplast development during early growth [45]. PQL3 is required for NDH activity in photosynthetic electron transport chain [46]. ATTK1B plays an important role in plant growth and development through the nucleotide salvage pathway [47]. Thus, the high proportion of chloroplast related genes in these mutants indicated a key role of chloroplast function in chilling tolerance in Arabidopsis.
We found in this study that RBD1, one of the chloroplast localized proteins identified in the chilling sensitive screen, is involved in 23S rRNA processing. The highly conserved rrn operon on chloroplast genomes encodes the 23S, 16S, 5S and 4.5S rRNAs and three tRNAs [48]. The primary precursor is initially processed to generate tRNAs, precursors of 23S, 16S, 5S and 4.5S rRNA by a series of endo- and exonucleolytic cleavages [49,50]. The partially processed 23S rRNA is then excised at two sites named the hidden breaks [38] (Fig 5B). Several molecules have been found to be involved in this final processing step. The CSP41b endonuclease and the RH39 RNA helicase are shown to be involved in processing the 23S rRNA at the 1st and 2nd hidden breaks respectively [36,38,51]. The RNase E endonuclease as well as the PNPase and RNase R exonucleases may also be involved in the final processing of 23S rRNA because their loss of function mutants accumulate incompletely processed 2.9-kb and 2.4-kb of the 23S rRNA, although their main function is to excise the 23S-4.5S species to generate 23S and 4.5S rRNAs [50,52,53]. Our study identifies RBD1 as a new regulator of 23S rRNA processing. Loss of RBD1 causes over-accumulation of the partially processed 2.9-kb and the 2.4-kb of 23S rRNA and reduction of the fully processed 1.1-kb species while the total amount of 23S rRNA does not change significantly (Fig 5B). The RBD1 protein is localized to nucleoid where the processing happens (Fig 4C), and it binds to the 23S rRNA but not other RNAs tested (Fig 6). This indicates that RBD1 functions in the last step of 23S rRNA processing, and it may facilitate the processing by guiding the endonucleases to the hidden breaks. Further refining its RNA biding sequences and determining its interaction with processing enzymes will reveal more its mode of action. Compared to the other 4 RNP mutants with similar chilling sensitive phenotypes, the rbd1-1 mutant exhibited the strongest 23S rRNA processing defect (Fig 7B). It has additional chloroplast RNA defects, but they did not occur at normal growth temperature and were induced at a much later stage after plants were chilling treated (S3 and S4 Figs). Because RBD1 does not appear to have enzymatic activities, it might interact with an RNase to bring it to the location for processing via its specific binding to the target RNA. Therefore, RBD1 is a positive facilitator of 23S rRNA processing. This function of RBD1 is more important at low temperatures than at normal temperatures. The rbd1 mutants show a slight defect at normal growth temperature but very strong defect at 4°C (Fig 5B and S4A Fig). Interestingly, RBD1 exhibited a higher association with the 23S rRNA at 4°C than at 22°C, suggesting a regulation of its activity by temperature. The enhanced binding of RBD1 to the precursor could compensate for a less efficient binding of the processing enzyme RNase to 23S rRNA, so that chloroplast translation machinery can be efficiently produced at low temperatures as well.
It is still yet to be determined how chloroplast function, especially translation, is the weak link in chilling tolerance. Likely, RNAs in chloroplast form more non- or less- functional secondary structures under chilling temperatures. Such structures of rRNAs may compromise the assembly or function of the translation machinery, while those of mRNAs may reduce their translation efficiency. RNA binding proteins such as RBD1 and other RNPs may reduce the formation of such non- or less-functional structures and thus become more critical under chilling conditions.
Our study also indicates that various processes of RNA metabolism and protein translation in chloroplasts are inter connected. Different primary defects can have ripple effects leading to a similar chilling sensitive phenotype at later stages. For instance, RBD1 is closely associated with processing of the 23S rRNA transcripts. The primary defect of rbd1 mutants is likely the reduced processing and accumulation of mature 23S rRNA products. This will lead to defect of plastid ribosome especially under chilling conditions and subsequently the translation defect in the chloroplasts. Since the core subunits of PEP are synthesized on plastid ribosomes, the rbd1 mutants are expected to have reduced PEP level which will lead to a reduction in PEP-dependent mRNAs. As a compensation, the NEP dependent genes might become overexpressed [54]. Indeed, the chilling-treated rbd1 mutants had reduced transcript levels of the PEP-dependent genes tested such as ndhF, psaA and rbcL, while NEP-dependent genes such as ycf3 had higher expression. rps4, transcribed by both PEP (in green tissue) and NEP (in white tissue) [55], had an overexpression in rbd1 mutants under chilling condition (S3 Fig). The slight deficiency in 16S rRNA processing might be an indirect effect of losing the RBD1 function because abnormal ribosome assembly have been shown to indirectly affect RNA processing [56,57]. Interestingly, the rbd1 mutants had a reduced level of chloroplast synthesized RbcL protein, but increased level of cytosol translated TOC75 protein (Fig 7D), suggesting a compensation at the translation level as well.
Together, this study reveals the importance of chloroplast RNA-binding proteins in chilling tolerance, and further studies will further enhance our understanding of molecular mechanisms of chilling tolerance and its variations in different plant species.
Arabidopsis T-DNA insertion lines were obtained from Arabidopsis Biological Resource Center with the stock numbers CS27941, CS27942 and CS27943. Four seeds of each line were sterilized and sowed on 0.5× MS (Murashige and Skoog, Sigma) solid medium with 1.2% agar and 2% sucrose. Plants were grown on vertical plates under a 16 h-light/d photoperiod for 8 days before being transferred to 4°C for two months. When using soil, plants were grown under a 12h-light photoperiod with light intensity at 100 μmol m-2 sec-1 and relative humidity at 50–70%. For spectinomycin treatment, seeds were sterilized and planted on 0.5×MS medium containing 0.8% agar and 2% sucrose with 0–3 mg/L spectinomycin under the conditions described above.
For complementing the rbd1 mutants, a full-length At1g70200 cDNA was amplified and cloned into PMDC32 [58] to generate the construct of p35S::RBD1. For the promoter reporter construct pRBD1::GUS, a 1.5-kb sequence upstream of the RBD1 translation start site was amplified and cloned into pCAMBIA1300 vector (CAMBIA, http://www.cambia.org). To generate a GFP-tagged RBD1 fusion protein (RBD1:GFP) for transient expression in protoplasts, the coding region of the RBD1 was amplified from the cDNA and cloned into the pSAT6-EGFP-N1 vector [59].
Agrobacterium tumefaciens stains of GV3101 carrying the resulting constructs were used to transform plants by standard floral dipping [60]. Primary transformants were selected on 0.5×MS (Murashige and Skoog, Sigma) medium containing 0.8% agar and 2% sucrose with 25 mg/L hygromycin for selection.
Protoplast isolation and transformation were carried out as previously described [61]. In brief, protoplasts were generated from 14-day-old wild-type Arabidopsis seedlings grown on plates with 12h light/ 12h darkness photoperiod and transformed with the plasmid DNAs. Protoplasts were then analyzed for GFP signals or for RNA co-IP from 12 hours to 48 hours after transformation.
GUS staining was performed as described previously [62]. Briefly, tissues were stained with X-Gluc staining buffer (5 mM potassium ferrocyanide, 5 mM potassium ferricyanide,100 mM sodium phosphate buffer, pH 7.0, and 0.005% Triton X-100 and 2 mM 5-bromo-4-chloro-3-indolyl-beta-D-glucuronic acid) for 1–2 hours at 37°C, followed by incubating in 70% ethanol to remove chlorophyll.
Total RNA was extracted with TRIzol reagent (Invitrogen) according to the manufacturer’s protocol. cDNAs were synthesized from total RNA by using AffinityScript QPCR cDNA Synthesis Kit (Agilent Technologies). Real-time quantitative PCR was performed on the BIO-RAD PCR System using iQSYBR GREEN SuperMix (BIO-RAD). Actin was used as a control.
Total RNA was extracted with TRIzol reagent (Invitrogen) according to the manufacturer’s instructions. For transcript analysis, five youngest visible leaves were harvested for RNA extraction. Ten micrograms of RNA per sample were separated on an agarose gel containing 1.2% formaldehyde and then transferred to uncharged nylon membranes (HybondN; GE Healthcare). The blots were UV cross-linked (150 mJ/cm2) and hybridized with gene specific, 32P labeled, single strand DNA probes.
Arabidopsis leaves were quick frozen in liquid nitrogen and homogenized in extraction buffer (50mM Tris-HCl, 1mM EDTA, 1mM EGTA, 150mM NaCl, 10% Glycerol, 5mM DTT, 0.25% Triton-X 100, 2% PVPP, 1mM PMSF). After centrifugation at 14,000 rpm for 10 min twice, the supernatants were mixed with loading buffer, boiled and loaded onto 12% SDS polyacrylamide gel. The proteins were visualized with Coomassie Blue staining of the gel or Ponceau S staining of the transferred blot.
After transformation, the protoplasts were incubated at 22°C for 8 h before half of each sample was transferred to 4°C and the other half to 22°C for 12 h. Protoplasts were then disrupted in 500 μL protoplast disruption buffer (0.3M sorbitol, 20mM Tricine-KOH (pH 8.4), 10 mM EDTA, 10 mM NaHCO3 and 0.1% BSA) and incubated on the ice for 30 minutes with several inversion, then were centrifuged at 300×g for 2 minutes. Poured off the supernatant, chloroplasts pellet were disrupted in 200 μL chloroplast disruption buffer (2mM DTT, 200 mM KOAC, 30 mM HEPES, pH 8.0, 10 mM MgOAc, and 2mg/ml proteinase inhibitor cocktail) and incubated on the ice for 30 minutes with occasional roughly vortex, then were centrifuged at 16,000×g for 30 minutes at 4°C. The supernatant was diluted with one volume of Co-IP buffer (150 mM NaCl, 20 mM Tris-HCl, 1 mM EDTA, 5 mM MgCl2, 1.1% Triton X-100, 100U/ml RNase inhibitor, 1 mM PMSF and 2mg/ml proteinase inhibitor cocktail) and incubated with 10 μg of GFP antibody (mouse IgG2a, Invitrogen) for 6 h with rotation, followed by 2 h rotation with 50 μL Dynabeads Protein G (Invitrogen) at 4°C. Beads containing the IPed protein and its bound RNAs were collected with a magnet, and supernatants were recovered and pellets were washed three times with Co-IP buffer. After the last wash, pellets were resuspended in 200μL Co-IP buffer. Resuspension was then extracted with TRIzol reagent (Invitrogen) according to the manufacturer’s protocol. Total RNA was treated by DNase I (Promega) to remove DNA contamination, and RT-minus control was performed to confirm complete removal of DNA in the following steps. cDNAs were synthesized from total RNA by using Superscript III reverse transcriptase (Invitrogen). Real-time quantitative PCR was performed on the BIO-RAD PCR System using iQSYBR GREEN SuperMix (BIO-RAD).
|
10.1371/journal.pgen.1003701 | Integration of the Unfolded Protein and Oxidative Stress Responses through SKN-1/Nrf | The Unfolded Protein Response (UPR) maintains homeostasis in the endoplasmic reticulum (ER) and defends against ER stress, an underlying factor in various human diseases. During the UPR, numerous genes are activated that sustain and protect the ER. These responses are known to involve the canonical UPR transcription factors XBP1, ATF4, and ATF6. Here, we show in C. elegans that the conserved stress defense factor SKN-1/Nrf plays a central and essential role in the transcriptional UPR. While SKN-1/Nrf has a well-established function in protection against oxidative and xenobiotic stress, we find that it also mobilizes an overlapping but distinct response to ER stress. SKN-1/Nrf is regulated by the UPR, directly controls UPR signaling and transcription factor genes, binds to common downstream targets with XBP-1 and ATF-6, and is present at the ER. SKN-1/Nrf is also essential for resistance to ER stress, including reductive stress. Remarkably, SKN-1/Nrf-mediated responses to oxidative stress depend upon signaling from the ER. We conclude that SKN-1/Nrf plays a critical role in the UPR, but orchestrates a distinct oxidative stress response that is licensed by ER signaling. Regulatory integration through SKN-1/Nrf may coordinate ER and cytoplasmic homeostasis.
| Proteins that are placed in membranes or secreted are produced in a cellular structure called the endoplasmic reticulum (ER). An accumulation of misfolded proteins in the ER contributes to many disease states, including diabetes and neurodegeneration. The ER protects against a toxic buildup of misfolded proteins by activating the unfolded protein response (UPR), which maintains ER homeostasis by slowing protein synthesis and enhancing ER functions such as protein folding and degradation. Many of these processes are controlled by three canonical ER/UPR gene regulatory factors. Here we identify the gene regulator SKN-1/Nrf as also playing a critical role in the UPR. SKN-1/Nrf is well known for its functions in oxidative stress defense and longevity. We now report that SKN-1/Nrf mobilizes an ER stress gene network that is distinct from its oxidative stress response, and includes regulation of other central UPR factors. Surprisingly, we also find that ER- and UPR-associated mechanisms are needed to “license” SKN-1/Nrf to defend against oxidative stresses. Our findings show that UPR and oxidative stress defense mechanisms are integrated through SKN-1/Nrf, and suggest that this integration may help maintain a healthy balance between ER and cytoplasmic functions, and stress defenses.
| The endoplasmic reticulum (ER) is responsible for multiple functions in protein synthesis and processing, lipid metabolism, xeno/endobiotic detoxification, and Ca2+ storage (reviewed in [1], [2]). The ER forms a continuous structure with the nuclear envelope and maintains extensive contact with mitochondria [3], [4]. Consequently, the ER is well positioned to sense and respond to changes in the cellular environment.
All secretory and membrane-bound proteins are synthesized in the rough ER, a process that is highly regulated so that only properly folded and modified proteins are released to the Golgi [1], [2], [5], [6]. Maturation and folding of these proteins involves glycosylation and formation of appropriate Cys-Cys crosslinks. When its protein folding capacity is exceeded (ER stress), the ER protects itself through the Unfolded Protein Response (UPR) (Figure S1A) [2], [5], [6]. This signaling and transcription program decreases protein translation, expands ER size and folding capacity, and directs misfolded proteins to be degraded in the cytosol. The UPR functions continuously to maintain ER homeostasis, but is amplified and diversified under ER stress conditions [5], [7]–[10]. In response to severe ER stress, the UPR promotes ER absorption through autophagy and ultimately may induce cell death. ER stress and the UPR have been implicated in many human diseases, including diabetes, inflammatory disease, neurodegenerative disease, secretory cell malignancies, and other cancers [6], [11], [12].
The canonical metazoan UPR is orchestrated by three major ER transmembrane signaling proteins (IRE1, PERK, and ATF6), and three bZIP-family transcription factors (XBP1, ATF4, and cleaved ATF6) (Figure S1A) [2], [5], [6]. The most ancient of these transmembrane proteins, IRE1, is a cytoplasmic endoribonuclease and kinase that senses unfolded proteins in the ER. In response to ER stress, the IRE1 RNAse initiates cytoplasmic splicing of the mRNA encoding XBP1, the transcription factor that is most central to the UPR. The IRE1 kinase contributes to ER homeostasis by regulating the IRE-1 endonuclease activity, and transmits signals through JNK, p38, and other pathways. The kinase PERK phosphorylates the translation initiation factor eIF2α, thereby globally decreasing translation. This reduces the ER protein-folding load, but also favors translation of mRNAs that encode protective proteins, including ATF4. ATF6 resides in the ER membrane but is transported to the Golgi and cleaved in response to ER stress. The activation status of these transmembrane proteins is influenced by their interactions with the ER chaperone BiP (HSP-3/-4 in C. elegans).
The ER lumen maintains an oxidative environment, in contrast to the cytoplasm, because the ER enzyme systems that form disulfide bonds generate reactive oxygen species (ROS) [1], [13], [14]. Accordingly, ER stress may eventually lead to cellular oxidative stress and activation of oxidative stress defense genes [15]. Metazoan oxidative and xenobiotic stress responses are orchestrated mainly by the Nrf bZIP-family transcription factors (Nrf1, 2, 3 in mammals). Nrf-family proteins regulate genes involved in various small molecule detoxification processes, including glutathione biosynthesis and conjugation, and have been implicated in longevity assurance in invertebrates and mammals [16]–[21]. These transcription factors have recently been shown to function in proteasome regulation, stem cell maintenance, and metabolism, suggesting that they may control a wider range of processes than previously realized [22]–[26]. It has been reported that mammalian Nrf1 and Nrf3 associate with the ER membrane and nuclear envelope [27]–[30], and that Nrf2 is phosphorylated by PERK [31], [32]. While these last observations are intriguing, it is unknown whether Nrf-family proteins might actually be involved in ER stress defenses, either through mobilizing an oxidative stress response or participating in the UPR itself.
The nematode C. elegans has been a valuable system for investigating how Nrf proteins function and are regulated in vivo, because of its advantages for employing genetics to elucidate regulatory networks, and performing whole-organism analyses of stress resistance and survival. The C. elegans Nrf ortholog SKN-1 plays a critical role in resistance to oxidative and xenobiotic stress, and in various pathways that extend lifespan [16], [17], [19], [23], [33]. Here we describe a comprehensive analysis of whether SKN-1 might be involved in the UPR. We found that under ER stress conditions SKN-1 directly activates many genes involved in ER function, including canonical ER signaling and transcription factors that in turn induce skn-1 transcription. Importantly, this response is distinct from that which SKN-1 mobilizes under oxidative stress conditions. SKN-1 is required for resistance to ER stress, including reductive stress, a surprising finding given the importance of SKN-1 for oxidative stress defense. Unexpectedly, UPR signaling is needed for SKN-1 to mobilize an oxidative stress response, suggesting that the ER has a licensing and possibly sensing role during oxidative and xenobiotic stress responses.
Several observations led us to investigate whether SKN-1/Nrf might be involved in ER stress defenses. Expression profiling that we performed in C. elegans under normal and oxidative stress conditions suggested that SKN-1 regulates a number of genes that are involved in UPR or ER functions [21]. These included atf-5 (UPR transcription factor ATF4), ckb-4 (choline kinase), pcp-2 (prolyl carboxypeptidase), and many genes encoding xenobiotic metabolism enzymes that localize to the smooth ER (Table S1). Moreover, a genome-wide Chromatin Immunoprecipitation (ChIP) analysis of C. elegans L1 stage larvae (MOD-ENCODE) [34] detected binding of transgenically expressed SKN-1 at the predicted regulatory regions of numerous genes involved in UPR- or ER processes, including UPR signaling and transcription (ire-1, xbp-1, pek-1, and atf-6), Ca++ signaling, and protein folding and degradation (Table S1).
To investigate whether SKN-1 might be involved in the UPR, we first used quantitative (q) RT-PCR to investigate whether it is needed for expression of representative ER stress-induced or ER maintenance genes, many of which are predicted to be SKN-1 targets (Table S1). In these initial gene expression studies we induced ER stress by treating C. elegans with the N-linked glycosylation inhibitor tunicamycin (TM), at a concentration that readily induces the UPR but does not cause detectable toxicity (5 µg/ml, Figure S1B) [15]. TM treatment resulted in skn-1-dependent upregulation of numerous canonical or predicted UPR- or ER-related genes (Figures 1A and 1B, Table S1). skn-1 was also required for the basal expression of psd-1, R05G6.7, and cnb-1, even though these genes were not activated by TM (Figures 1A and 1B). TM-induced ER stress also upregulated two direct SKN-1 targets that are involved in glutathione metabolism (gcs-1 and gst-4) [19] in a skn-1–dependent manner, and transgenic reporter analysis detected gcs-1 activation in the intestine, the C. elegans counterpart to the gut, liver, and adipose tissue (Figures 1C and 1D). Importantly, however, ER stress did not activate various other genes that are typically induced by SKN-1 under oxidative stress conditions (Figure S1C). Taken together, the data indicate that SKN-1 mediates a response to ER stress, but also that this response does not correspond simply to its oxidative stress defense function.
To investigate whether SKN-1 activates genes directly during ER stress, we used ChIP to detect endogenous SKN-1 and markers of transcription activity at pcp-2, atf-5, and gst-4, each of which is flanked by SKN-1 binding sites and upregulated by oxidative and ER stress in a skn-1-dependent manner [21] (Figures 1B and 1C). SKN-1 was readily recruited to these genes in response to either TM-induced ER stress or Arsenite (AS)-induced oxidative stress (Figures 2A, 2E, 2I, and S2A-S2C). During transcription, RNA Polymerase II (Pol II) is phosphorylated on Ser 2 of its C-terminal domain (CTD) repeat (P-Ser2) [35]. At each gene we examined, ER stress increased Ser 2 phosphorylation levels (Figures 2B, 2F, and 2J). Also consistent with transcriptional activation, at these loci ER stress increased acetylation of Histone H3, another marker of transcription activity [36], but reduced overall Histone H3 occupancy (Figures 2C, 2D, 2G, 2H, 2K, and 2L). Taken together, our findings suggest that SKN-1 directly activates a major transcriptional response to ER stress.
We next investigated whether SKN-1 might regulate expression of core UPR signaling and transcription factors, as predicted by the MOD-ENCODE data [34]. XBP-1 is central to the UPR, and in mammals it controls transcription of other core UPR genes (atf4/atf-5, and BiP/hsp-4) along with many downstream genes [6], [37]. During the UPR, xbp-1 expression is regulated at the level of transcription, as well as through cytoplasmic splicing of its mRNA by the IRE-1 endoribonuclease (Figure S1A) [5], [6]. The spliced form of the xbp-1 mRNA (xbp-1s) encodes the transcriptionally active form of XBP-1 (XBP-1s). When SKN-1 was lacking, ER stress failed to induce accumulation of each xbp-1 mRNA form and, remarkably, decreased the ratio of xbp-1s to the unspliced xbp-1 form (xbp-1u) (Figures 3A, 3B, and S3A). The xbp-1 locus includes a predicted SKN-1 binding site (not shown), and ChIP results indicated that endogenous SKN-1 accumulates at the xbp-1 site of transcription in response to ER stress (Figure 3C). This evidence that SKN-1 directly regulates xbp-1 could account for the reduction in total xbp-1 mRNA, but not the apparent effect of SKN-1 on xbp-1 splicing. A plausible explanation is that lack of SKN-1 also reduced basal and ER stress-induced expression of ire-1 (Figures 3D and 3E). Moreover, we observed that SKN-1 is recruited to the ire-1 locus in response to ER stress (Figure 3F), consistent with MOD-ENCODE evidence that ire-1 may be a SKN-1 target [34].
SKN-1 was also required for expression of other core UPR genes. Mutation or RNAi knockdown of skn-1 prevented ER stress-induced expression of the unfolded protein chaperone and sensor HSP-4 (BiP) (Figure S1A)(Figures 3G, S3B, and S3C). Binding of SKN-1 at hsp-4 was not detected in the MOD-ENCODE study of L1 larvae [34], but our ChIP evidence indicated that both SKN-1 and XBP-1 bind directly to the hsp-4 locus (Figures S3D and S3E), which includes predicted SKN-1 binding sites (not shown). SKN-1 similarly contributed to expression of the core UPR factors pek-1 and atf-6 (Figures 3D and 3E). Our evidence that SKN-1 is important for transcriptional induction of core UPR signaling and regulatory factors predicts that it should be important for C. elegans survival under ER stress conditions. Treatment with TM at a 7-fold higher concentration (35 µg/ml) than is sufficient to induce the UPR impaired the survival of skn-1 mutants but not wild type animals (Figure 3H and Table S2). We conclude that SKN-1 plays a critical role in the UPR through its direct transcriptional regulation of core UPR factors, along with many downstream genes.
We next examined whether expression of skn-1 itself is increased when the ER becomes stressed, and whether various conditions that cause ER stress affect SKN-1 activity. Treatment with TM increased the levels of multiple mRNA species that encode SKN-1 isoforms (Figure 4A and S4A). In addition, non-lethal treatment with either the Ca++ pump inhibitor thapsigargin (Thap) or the proteasome inhibitor Bortezomib upregulated transcription of skn-1, and various SKN-1-regulated genes (Figures 1, 4A, and S4B–S4C). Finally, knockdown of either the ER chaperone hsp-4 or the UPR transcription factor atf-6 resulted in transcriptional upregulation of skn-1 and many of its ER stress targets in the absence of drug treatment, presumably because of an elevated level of ER stress (Figures 4A, S4D and S4E). We conclude that skn-1 transcription and activity are increased in response to a variety of conditions that are associated with ER stress.
An important hallmark of the UPR is a decrease in the overall levels of translation [5], [6]. This relieves stress on the ER, and allows translation of atf4 and other protective genes to be maintained or even increased. We investigated whether SKN-1 translation is similarly “spared” under ER stress conditions. Supporting this idea, TM treatment increased SKN-1 protein levels, a trend that was observed in Western and IP-Western analyses of whole animals with two specific SKN-1 antibodies (Figures 4B and S4F–S4I). Based upon its size, this approximately 85 kD SKN-1 species is likely to represent SKN-1a, the largest SKN-1 isoform. While this size is larger than the expected SKN-1a MW of 70 kD, SKN-1 is phosphorylated and predicted to be glycosylated, as is characteristic of Nrf1 and Nrf3 (not shown) [17], [28], [38]–[40]. Our finding that SKN-1 protein levels are increased by ER stress is consistent with earlier evidence that SKN-1 translation seemed to be preserved when translation initiation was inhibited [41].
Prolonged ER stress leads to accumulation of reactive oxygen species (ROS) and induction of an oxidative stress response [15], [42], making it important to determine whether ER stress treatments might activate SKN-1 simply through a secondary response to oxidative stress. Arguing against this interpretation, even though SKN-1 is well known to defend against oxidative stress, we found that reductive ER stress also induced a SKN-1-dependent response. The reducing agent dithiothreitol (DTT) initiates the UPR through reduction of Cys-Cys bonds in the ER [43]. DTT treatment resulted in transcriptional induction of skn-1 and many of its target genes, and increased SKN-1 protein levels (Figures 4C and S4J). SKN-1 appeared to be required for its downstream targets to be activated by DTT-induced reductive stress (Fig. S4K), and knockdown of either skn-1 or hsp-4 rendered C. elegans comparably sensitive to reductive stress from DTT (Figure S4L and Table S3). Another way to reduce oxidation in the ER is through inhibiting expression of the oxidase ERO-1, which promotes Cys-Cys crosslinking [43]. ero-1 RNAi decreases ROS levels, initiates the UPR, and extends lifespan [15]. As observed with DTT, ero-1 RNAi transcriptionally activated skn-1 and several of its downstream targets (Figure 4D).
Additional lines of evidence support the idea that SKN-1 acts in the UPR independently of its role in oxidative stress defense. Many genes that are activated by SKN-1 under oxidative stress conditions were not upregulated by ER stress (Figures S1C and S4M). Oxidative stress from AS treatment induced the SKN-1::GFP (green fluorescent protein) fusion to accumulate to high levels in intestinal nuclei, as previously described (Inoue, et al., 2005), but this did not occur in response to ER stress (Figure S4N). Finally, we did not observe increased levels of oxidized proteins under conditions of TM-induced ER stress (Figure S4O). Taken together, the data show that ER stress directs SKN-1 to activate a specific set of its target genes independently of any secondary oxidative stress response.
If ER signaling pathways regulate SKN-1, then key UPR signaling and transcription factors should be required for ER stress to activate SKN-1 and its target genes. Accordingly, RNAi or mutation of ire-1, atf-5, pek-1, or hsp-4 essentially prevented ER stress from inducing transcription of skn-1 and several of its target genes (Figure 5A). Knockdown of xbp-1 under control conditions increased background expression of some SKN-1 isoforms and target genes (skn-1b, pcp-2, gst-4, hsp-4), possibly because ER stress was increased, but also interfered with ER stress-induced activation of several of these genes (skn-1a, pcp-2, gcs-1, hsp-4) (Figure S5A). RNAi against ire-1, which is essential for XBP-1s expression [5], [6], also blocked TM-induced accumulation of SKN-1, Pol II, or P-Ser2 at the gst-4, pcp-2, and atf-5 loci (Figures 5B–5E, S5B and S5C). Knockdown of hsp-4 or pek-1 had a similar effect (Figure S5D–S5G). The evidence indicates that, in general, core UPR factors are required for ER stress to upregulate expression of SKN-1 and its target genes.
The most straightforward mechanism through which ER stress could increase skn-1 transcription is through the direct regulation of skn-1 by one or more of the canonical UPR transcription factors. During the UPR, downstream gene transcription is controlled largely by XBP1 and ATF4, which may regulate each other directly, with ATF-6 playing a more specialized role [8], [15], [37]. The skn-1 locus contains possible XBP-1 and ATF-6/XBP-1 binding elements (not shown), and genome-wide ChIP studies suggest that mammalian Nrf3 may be a direct XBP1 target [37]. We determined that XBP-1 binds within the skn-1 locus in response to ER stress, suggesting direct regulation (Figure 5F), a remarkable parallel to the direct regulation of xbp-1 by SKN-1 (Figure 3C). Moreover, ATF-6 was also recruited to the skn-1 locus in response to ER stress (Figure 5G). In mammals, XBP-1 may regulate its own expression [37]. Our ChIP analysis indicated that SKN-1 also binds to its own locus with ER stress (Figure 5H), suggesting that SKN-1, XBP-1, and ATF-6 together regulate skn-1 transcription. ER stress also resulted in XBP-1 and ATF-6 recruitment to the direct SKN-1 targets pcp-2 and gst-4 (Figures S5H–S5K). Together, the evidence suggests that SKN-1, XBP-1, and ATF-6 may function together to regulate several downstream genes. We conclude that SKN-1 is transcriptionally integrated into the UPR, in which it functions upstream, downstream, and in parallel to the known core UPR transcription factors.
The mammalian SKN-1 orthologs Nrf1 and Nrf3 have been detected in association with the ER (see Introduction), raising the question of whether this might also be true for a proportion of SKN-1. Consistent with this idea, Nrf1 and the SKN-1a isoform each contain a predicted transmembrane domain [27] (Figure S6A). To investigate whether SKN-1 might be present at the ER, we asked whether it might be detected in association with the ER-resident chaperone BiP (HSP-3/-4)(Figure S1A). We performed co-immunoprecipitation (IP) analyses of intact worms that had been crosslinked with formaldehyde as in our ChIP experiments. These conditions capture direct and indirect in vivo interactions that occur within approximately 2 Å, and allow for high-stringency detergent and salt-based washings that minimize non-specific binding [44], [45]. Under both normal and ER stress conditions, association between HSP-4 and SKN-1 was readily detected by high-stringency IP performed in either direction (Figure 6A and 6B). As in Figure 4B, the size of this SKN-1 species suggested that it may correspond to SKN-1a. The data suggest that some SKN-1 may be produced at the ER and might remain associated with this organelle.
Given that BiP has been found in other cellular locations besides the ER [46], we also investigated whether SKN-1 is present in a cellular fraction that is enriched for the ER (Figure S6B). SKN-1 was readily detectable in an ER fraction that included HSP-4, but not the cytoplasmic protein GAPDH (Figures 6C and 6D). The interaction between endogenous SKN-1 and HSP-4 was confirmed within this ER fraction by a co-IP that was performed without crosslinking (Figure 6E). Together, our findings suggest that the association of SKN-1/Nrf proteins with the ER is evolutionarily conserved.
Our finding that UPR factors are required for SKN-1 activity to be increased under ER stress conditions raised a related question: might UPR-related mechanisms also be involved in SKN-1 responses to oxidative stress? Surprisingly, we found that RNAi or mutation of core UPR signaling and transcription factors (atf-5, pek-1, ire-1, hsp-4 and xbp-1) impaired oxidative stress (AS)-induced activation of several SKN-1 target genes, including skn-1 itself (Figures 7A, 7C, and S7A). Similarly, ire-1 RNAi attenuated activation of the gcs-1::GFP reporter in the intestine (Figure S7B). This impairment of the oxidative stress response is particularly striking because ire-1 RNAi actually increased oxidized protein levels, in contrast to the mild AS treatment conditions used for gene expression analyses (Figure S4O).
Importantly, oxidative stress from AS did not simply activate the canonical UPR. Many SKN-1-regulated genes that were induced by oxidative stress were not upregulated by ER stress, and vice-versa (Figures S1C, S4M, and S7C). This shows that SKN-1 mobilizes distinct transcriptional responses to oxidative and ER stress, even if these responses overlap to an extent. Moreover, AS primarily increased accumulation of the unspliced xbp-1 mRNA form (xbp-1u), in striking contrast to the increase in xbp-1s levels that is characteristic of ER stress (Figures 3A and 7C). Treatment with the oxidative stressor tert-butyl hydrogen peroxide (tBOOH) induces a SKN-1-dependent response that overlaps with the AS response, but includes SKN-1-independent activation of many genes that are otherwise SKN-1-dependent [21]. Knockdown of ire-1 or hsp-4 inhibited tBOOH from upregulating skn-1 and some SKN-1 targets (Figure 7B), but did not eliminate activation of other genes (gcs-1, sdz-8, and gst-10; not shown). The data suggest that core UPR factors are needed for SKN-1 to function properly under oxidative stress conditions, in addition to the setting of ER stress.
The extensive regulatory integration that exists among UPR transcription factors, as described by others and in this study (Figures 7A, 7B, and S7A) [8], [15], [37], could explain why multiple UPR-associated signaling and transcription factors are needed for skn-1 expression to be increased in response to oxidative stress. However, we considered that the UPR might also influence SKN-1 regulation at a post-translational level. In the C. elegans intestine SKN-1 is predominantly cytoplasmic under normal conditions, but accumulates in nuclei in response to oxidative stress from AS treatment [38]. This nuclear accumulation was dramatically reduced in animals that had been exposed to ire-1 RNAi (Figure S7D). The presence of SKN-1 in intestinal nuclei is dependent upon its phosphorylation by the p38 kinase, which is activated by oxidative stress [23], [38], [47]. The IRE-1 kinase activity transmits signals through the JNK and p38 MAPK pathways [6], [48]–[50], and we determined that ire-1 knockdown largely prevented the increase in p38 signaling that occurs in response to oxidative stress (Figures 7D and S7D). Taken together, these data suggest that IRE-1 is required for oxidative stress to activate SKN-1 post-translationally.
If UPR signaling and transcription factors are required for SKN-1 to mobilize appropriate oxidative stress responses, then oxidative stress sensitivity should be increased when these canonical UPR factors are lacking. Accordingly, RNAi or mutation of these genes significantly increased sensitivity to oxidative stress from exposure to AS, paraquat, or t-BOOH (Figures 7E, S7E, and S7F; Table S4). We conclude that signaling from the ER is required for SKN-1 to respond to oxidative stress, and therefore that UPR-mediated regulation of SKN-1 plays a central role in the homeostatic integration of ER and oxidative stress responses.
It is well-established that the canonical UPR transcription factors XBP1, ATF4, and ATF6 control overlapping sets of downstream genes and processes [5], [6], but much less is known about how their responses to ER stress might be integrated with other mechanisms that maintain cellular stress defense and homeostasis. We have determined that the oxidative/xenobiotic stress response regulator SKN-1/Nrf functions as a fourth major UPR transcription factor in C. elegans. Without SKN-1, ER stress failed to increase the expression of core UPR signaling and transcription factors, many of which are regulated directly by SKN-1 (ire-1, xbp-1, atf-5, and hsp-4; Figures 1, 2, 3 and S3). It was particularly striking that SKN-1 was disproportionally required for production of spliced xbp-1 mRNA (xbp-1s), presumably because of its importance for IRE-1 expression (Figures 3D–F). SKN-1 was also needed for ER stress to upregulate numerous genes that are known or predicted to be involved in various ER- or UPR-related processes, including ER homeostasis (ero-1, pdi-2), chaperone-mediated protein folding (hsp-3, hsp-4, dnj-28, T05E11.3 (HSP-90/GRP94)), autophagy (lgg-1, lgg-3), calcium homeostasis (sca-1, crt-1), ER membrane integrity (ckb-4), and a pathway that defends against ER stress when the canonical UPR is blocked (abu-8, abu-11 [51]) (Figure 1, 3G and Table S1). Together, our data indicate that SKN-1 regulates transcription of essentially the entire core UPR apparatus and many downstream ER stress defense genes in vivo.
We were surprised to find that SKN-1 was so broadly important for UPR transcription events. A trivial explanation for our findings would be that skn-1 mutants did not need to induce the UPR robustly because they were resistant to ER stress. This explanation was ruled out, however, by our finding that skn-1 mutants are actually sensitized to ER stress from diverse sources (Figures 3H and S4L). Importantly, our ChIP studies and MOD-ENCODE data [34] indicate that SKN-1 controls many core and downstream UPR genes directly by binding to their promoters (Figures 2, 3, and S3E, Table S1). We also found that ER stress induces SKN-1, XBP-1, and ATF-6 to bind promoters directly to regulate many of the same genes, including skn-1 itself (Figures 5, S3, and S5). In addition, under ER stress conditions, UPR signaling increased levels of skn-1 mRNA and protein (Figures 4 and S4), indicating that SKN-1 is controlled by the UPR and is an active participant in this response. Together, our data reveal that a remarkable degree of regulatory and functional integration exists between SKN-1 and the three canonical UPR transcription factors (Figures 7F and S1A).
Although ER stress increases skn-1-dependent transcription and SKN-1 occupancy at several downstream gene promoters, it did not detectably alter the overall levels of SKN-1 in intestinal nuclei, at least as indicated by levels of a transgenic GFP fusion protein (Figure S4N). While this might seem paradoxical, we observed a similar situation with reduced TORC1 signaling [19]. Under conditions of low TORC1 activity SKN-1 target genes were activated in a skn-1-dependent manner, and this was accompanied by increased SKN-1 binding to their promoters, but not by an obvious increase in the bulk levels of SKN-1 in nuclei. Our finding that SKN-1 binds to downstream UPR genes together with other UPR transcription factors suggests a paradigm that could explain this phenomenon. If SKN-1 binds cooperatively with UPR factors or other co-regulators to some of its targets, this could shift the binding equilibrium to allow those targets to be activated by SKN-1 that is already present in the nucleus, without it being necessary to “flood” the nucleus with higher levels of SKN-1. This scheme might be important for fine-tuning of SKN-1 downstream functions, and for allowing SKN-1 to activate different targets in different situations, as we have observed in this study.
In performing these analyses, we were mindful of the concern that the involvement of SKN-1 in the UPR might derive from its possible role in a secondary oxidative stress response. Several lines of evidence argued against this interpretation. For example, the direct involvement of SKN-1 in regulating multiple core UPR signaling and transcription factors during the UPR (Figures 3 and S3) is not consistent with its UPR functions deriving simply from a secondary oxidative stress response. Moreover, under our ER stress conditions SKN-1 was required for accumulation of the spliced form of the xbp-1 mRNA, whereas oxidative stress increased levels of the unspliced xbp-1 message (Figures 3A, 3B, and 7C). It was particularly striking that SKN-1 defended against reductive ER stresses (Figures 4C, 4D, S4J, S4K, and S4L), given the extensively described role of SKN-1/Nrf proteins in oxidative stress responses. These last observations indicated that SKN-1 defends against ER stress per se, and not only against oxidative conditions. Importantly, ER stress and the UPR directed SKN-1 to activate some of its target genes that are induced by oxidative stress, but not others (Figure S1C and S4M). On the other hand, many genes that SKN-1 activated under ER stress conditions were not induced by oxidative stress (Figure S7C). Taken together, the data show that SKN-1 does not simply activate oxidative stress defenses in the context of ER stress, but orchestrates a specific transcriptional ER stress response that is integrated into the broader UPR.
Our finding that SKN-1 mobilizes overlapping but distinct responses to ER and oxidative stress defines a new function for this surprisingly versatile transcription factor. It also supports our model that SKN-1/Nrf proteins do not control the same genes under all circumstances, but instead induce protective responses that are customized to the challenge at hand [19], [26]. The idea that SKN-1 works together with canonical UPR transcription factors at downstream genes may provide a model for understanding how particular SKN-1 functions can be mobilized under different conditions, if these proteins and other SKN-1 “partners” guide its activities.
Consistent with reports that Nrf1 and Nrf3 are present at the ER [27]–[30], we found that some SKN-1 also localizes to the ER. We detected association between SKN-1 and the ER chaperone HSP-3/4 (BiP) in crosslinking analyses of intact animals, the presence of SKN-1 within an ER fraction, and association between SKN-1 and HSP-3/4 within that fraction (Figure 6 and S6). Each of these experiments involved analysis of endogenous proteins. These strategies would have detected either direct or indirect interactions, so they do not demonstrate that SKN-1 binds directly to HSP-3/4 (BiP), but they do show that these proteins reside very close to each other at the ER. Apparently, association between SKN-1/Nrf proteins and the ER is evolutionarily conserved. The example of ATF-6, which is activated through cleavage in the Golgi (Figure S1A), predicts that ER-associated SKN-1 might have a signaling function in which it is cleaved in response to ER stress. However, the relative instability of SKN-1 and the presence of smaller isoforms have so far confounded the resolution of this question (not shown). We recently determined that some SKN-1 also localizes to mitochondria and that SKN-1 can promote a starvation-like state when overexpressed, a function that also appears to be conserved in Nrf proteins [26]. Given the extensive communication between the ER and mitochondria [4], [52], our results suggest that SKN-1/Nrf might respond directly to the status of each of these organelles. Consistent with this notion, SKN-1 is required for expression of the C. elegans ortholog of mitofusin (fzo-1) (Figure 1A), which mediates mitochondrial fusion and mitochondria-ER interactions [4].
Taken together, our findings show that processes controlled by SKN-1/Nrf proteins are critical for ER stress defense and homeostasis, and that SKN-1 is extensively intertwined with the UPR in vivo. While differences could exist between C. elegans and mammals with respect to regulatory networks, the extent of the functional interactions we have observed predicts that mammalian Nrf proteins are likely to play an important role in the UPR that is distinct from their familiar function in oxidative stress responses.
Perhaps our most surprising finding was that core UPR signaling and transcription factors were required for SKN-1 to mount a transcriptional response to oxidative stress (Figures 7 and S7). Cooperative interactions between SKN-1 and UPR transcription factors could account for some of these findings, through their effects on SKN-1 expression, but it was striking that ire-1 was needed for AS to induce SKN-1 nuclear accumulation, a phenomenon that does not occur under ER stress conditions (Figures S4N and S7D). Moreover, ire-1 was required for the AS-induced p38 signal that is needed for SKN-1 to be present in nuclei (Figure 7D). These last findings indicate that IRE-1 affects the oxidative stress response at a step upstream of SKN-1. One speculative possibility for further investigation is that the IRE-1 kinase activity might be needed to initiate the oxidative stress-induced p38 signal. Together, our data show that signaling from the ER is required to “license” the oxidative/xenobiotic stress response, and suggest that the ER might function in effect as a stress sensor. This importance of the UPR for SKN-1 activity may have implications for our understanding of aging and longevity assurance. SKN-1/Nrf not only defends against resistance to various stresses, but is also important in pathways that affect longevity, including insulin-like, TORC1, and TORC2 signaling, and dietary restriction [16], [17], [19], [20]. IRE-1 and XBP-1 have each been implicated in longevity [53], [54], making it important to determine the extent to which these UPR-based mechanisms might influence aging through regulation of SKN-1/Nrf and its functions.
Why would such extensive integration have arisen, in which SKN-1/Nrf is essential for the UPR, and signaling from the ER is needed for SKN-1/Nrf activities that are distinct from the UPR (Figure 7F)? SKN-1/Nrf controls cellular processes that profoundly influence the ER. Its target genes drive synthesis of glutathione, the major redox buffer within the ER, and encode many endobiotic and xenobiotic metabolism enzymes that reside on or within the smooth ER (Table S1) [20], [21], [55]. Under some circumstances SKN-1/Nrf also regulates proteasome expression and activity, and numerous chaperone genes [20], [21], [23]–[25]. One possibility is that the influence of SKN-1 could attune the UPR to events taking place in the cytoplasm. It might be advantageous to mount a robust transcriptional UPR if the cytoplasm is under duress, for example, and to moderate the UPR when cytoplasmic stress is low. Under these conditions, SKN-1 activity would be relatively high and low, respectively. SKN-1 activity is also comparatively low when translation rates are high [19], [23]. If the ER becomes stressed under growth conditions it might be useful to limit the transcriptional UPR initially, because a reduction in translation rates might largely suffice to restore homeostasis. Again, under these conditions low SKN-1 activity could act as a brake on the transcriptional UPR. With respect to the oxidative/xenobiotic stress response, it could be important for the ER to have a “vote” on its intensity, given the profound influence of SKN-1/Nrf on cellular redox status and resources devoted to the ER. It seems likely, therefore, that the ER not only manages its own homeostasis, but through SKN-1/Nrf has a broader impact on cellular stress defense networks that is likely to be critical in their normal and pathological functions.
For each condition studied, RNA was extracted from approximately 100 µl of packed mixed-stage worms that were collected in M9 at the indicated time point. To induce UPR-associated gene expression, at day three of adulthood worms were treated with 5 µg/ml TM (Sigma) for 16 hours [15], or at day four with 5 mM DTT (Sigma) [54] for two hours, 5 µM thapsigargin (Enzo) [56] for two hours, or 5 µM Bortezomib (proteasome inhibitor, LC Labs) for six hours (similar to published C. elegans MG132 proteasome inhibitor treatment [57]). In each case, these treatments were non-lethal. For arsenite (AS) and tBOOH exposure, up to 100 µl of packed worms were collected and nutated in 5 mM AS or 12 mM tBOOH for 1 hour (a non-lethal duration). Each of these treatments was performed in a volume of 1 ml, and was followed by pelleting. RNA was analyzed by qRT-PCR as described, with values normalized to an internal standard curve for each amplicon [19], [44]. The same treatment conditions were used for ChIP experiments.
Expression or nuclear accumulation of transgenic GFP proteins was scored as “low,” “medium,” or “high” essentially as published [19], or were quantified using ImageJ 1.45S.
ChIP was performed essentially as described [19], [44]. 2 ml of packed mixed-stage worms were crosslinked with formaldehyde at room temperature for 20 minutes. After quenching, lysis, and determination of protein concentration, 1 mg/ml samples were frozen as aliquots at −80°C. The resolution of the assay was approximately 250–500 bp [44]. The monoclonal antibody FC4 [58] was used for SKN-1 ChIP experiments, as in previous ChIP analyses [19]. Other antibodies are described in the Supplemental Experimental Procedures. Analyses of intergenic regions and control genes (not shown) indicated that average signals of 14%, 11%, 26%, 4%, 11%, 7%, and 8% represent thresholds for specific presence of SKN-1, Pol II, PSer2, and H3-AcK56, XBP-1, ATF-6, and Histone H3 respectively.
Worms from five confluent 20 cm2 plates were collected in M9 with or without TM treatment (5 µg/ml) for 16 hours, in order to generate 2× 1 ml of packed mixed-stage animals. Worms were sonicated 3× for 20 seconds in homogenization buffer (supplied by IMGENEX kit, supplemented with HDAC inhibitors, protease inhibitors, phosphatase inhibitors, and MG132) with the Branson midiprobe 4900 Sonifer before fractionation with the IMGENEX Endoplasmic Reticulum Enrichment Kit (Cat No. 10088K) [59]. Mitochondrial and ER fractions were washed 3× with 1 ml PBS and resuspended in 400 µl PBS (supplemented with HDAC, protease, and phosphatase inhibitors and MG132). Up to 100 µl of the ER or cytoplasmic fractions were used for each IP.
Controls for a polyclonal rabbit antiserum raised against SKN-1c (JDC7, referred to as pSKN-1) are shown in Figures S4F–S4J. HSP-3/4/BiP was detected with either C-terminal Drosophila Hsc3 [60] (Figures 6A and 6B) or N-terminal human BiP antibody (Sigma et21) [61], [62] (Figures 6C and 6E). Note that both BiP antibodies recognized the same 75 kD band. ATF-6 (Abcam ab11909), Tubulin (Sigma #9026), and GAPDH (Santa Cruz sc25778) antibodies were also used. Phosphorylated p38 was detected using an antibody from Cell Signaling T180/Y182 as described previously [23]. For Western blotting, antibodies were used at the following dilution: 1∶200 FC4 monoclonal αSKN-1, 1∶200 polyclonal αSKN-1, 1∶1000 αPol II, and 1∶1000 for αHsc3. All other antibodies were used at manufacturer's recommended concentrations.
For IPs, the indicated antibodies (50 µl FC4 monoclonal αSKN-1 or polyclonal αSKN-1,10 µl Hsc3 (BiP) or 20 µl BiP (Sigma)) and pre-blocked Salmon Sperm DNA/Protein A beads (Zymed) were added to lysates or samples from the fractionation described above. The final volume was brought to 500 µl in 1× PIC, 1× PMSF, and 1∶1000 MG132 diluted in 1× PBS. Samples were nutated overnight at 4°C and washed three times for 5 minutes at 4°C the next day with NP-40 wash buffer. Beads were spun down at 3000 rpm and resuspended in 4× SDS Laemmli Buffer. Samples were boiled for 15 minutes with 20 µl β-mercaptoethanol and 50 µl 4× SDS Laemmli. Samples were loaded (50 µl each) onto NuPAGE Novex Bis-Tris 10% Gels. Pierce ECL or Femto Western Blotting Substrate was used for detection.
Other methods are available in Text S1 (Supplementary Materials and Methods).
|
10.1371/journal.pntd.0006037 | A multi-country study of the economic burden of dengue fever: Vietnam, Thailand, and Colombia | Dengue fever is a major public health concern in many parts of the tropics and subtropics. The first dengue vaccine has already been licensed in six countries. Given the growing interests in the effective use of the vaccine, it is critical to understand the economic burden of dengue fever to guide decision-makers in setting health policy priorities.
A standardized cost-of-illness study was conducted in three dengue endemic countries: Vietnam, Thailand, and Colombia. In order to capture all costs during the entire period of illness, patients were tested with rapid diagnostic tests on the first day of their clinical visits, and multiple interviews were scheduled until the patients recovered from the current illness. Various cost items were collected such as direct medical and non-medical costs, indirect costs, and non-out-of-pocket costs. In addition, socio-economic factors affecting disease severity were also identified by adopting a logit model. We found that total cost per episode ranges from $141 to $385 for inpatient and from $40 to $158 outpatient, with Colombia having the highest and Thailand having the lowest. The percentage of the private economic burden of dengue fever was highest in the low-income group and lowest in the high-income group. The logit analyses showed that early treatment, higher education, and better knowledge of dengue disease would reduce the probability of developing more severe illness.
The cost of dengue fever is substantial in the three dengue endemic countries. Our study findings can be used to consider accelerated introduction of vaccines into the public and private sector programs and prioritize alternative health interventions among competing health problems. In addition, a community would be better off by propagating the socio-economic factors identified in this study, which may prevent its members from developing severe illness in the long run.
| Dengue fever has been prevalent in South-East Asia and South America. Despite the increase of dengue fever cases, there continues to be a lack of economic assessment partly due to the absence of vaccines until recent times. Many of the previous economic burden studies for dengue fever were not standardized, making them difficult to compare. We implemented the standardized economic burden survey for dengue fever in a multi-country setting: Vietnam, Thailand, and Colombia. We found that the economic burden of dengue fever is substantial in all three dengue endemic countries. Our study also identified socio-economic factors which are related to the probability of experiencing severe illness. The first live attenuated, tetravalent dengue vaccine (CYD-TDV) has been already licensed in some dengue-endemic countries. As three countries will soon face decisions on whether and how to incorporate current and future vaccine candidates within their budget constraints, the updated economic burden estimates can be used to develop sustainable financing plans.
| Dengue fever is a major public health concern in many parts of the tropics and subtropics. Dengue is a vector-borne viral illness and transmitted to humans by two mosquito vectors: Aedes aegypti and Aedes albopictus. There are four serotypes that cause dengue, with a wide clinical spectrum of symptoms. A previous study estimates there are 96 million apparent and 294 million inapparent dengue infections occurring yearly [1].
While many countries in the tropics and subtropics recognize dengue as a public health concern, dengue control efforts were not often considered as a priority due to the absence of specific treatment and costly vector control interventions in developing countries [2]. Having reviewed publicly available literature on the economic burden of dengue, there were a relatively small number of empirical studies. While Suaya et al. covered eight countries with a common method, many of the other existing studies were not standardized in terms of their methodologies, with varying assumptions making them difficult to compare [3–5].
Given that there continues to be a lack of economic assessment of dengue, evidence on the economic burden of dengue fever is urgently needed to guide decision-makers on the expected returns on their investments and to set health policy priorities [3]. The Dengue Vaccine Initiative (DVI) has conducted extensive multidisciplinary dengue-fever studies for decision-makers in three countries: Vietnam, Thailand, and Colombia [6]. The rationale for these studies is to contribute data to inform policymakers about the economic burden of dengue fever and to assess the economic efficiency of vaccine introduction strategies.
In Vietnam, dengue is highly prevalent, and the incidence rate is reported to be 145/100,000 population according to the national surveillance system data from 2010 [7,8]. All four dengue serotypes circulate [9]. Vietnam seems to follow similar trends to Thailand where dengue epidemiology has been studied extensively. Despite mosquito control efforts, dengue fever/dengue hemorrhagic fever (DF/DHF) has steadily increased in both incidence and range of distribution in Thailand, and the incidence rate in 2010 was reported to be 177/100,000 population [8,10]. Dengue has now spread to all provinces, districts, sub-districts and communities, and is seen every year [11].
Meanwhile dengue has been prevalent for the past 20 years in Colombia. In 2010, Colombia experienced a nationwide epidemic with a dengue incidence rate of 685/100,000 population [8,12]. Further, outbreaks have been observed in many parts of the country since 2011. Dengue epidemiology in Colombia is characterized by the circulation of all four dengue serotypes. The country has a well-established national dengue surveillance system.
A recombinant tetravalent vaccine candidate (CYD-TDV) has completed Phase 3 clinical trials in Southeast Asia and South America, and has been already licensed in some dengue-endemic countries. Understanding the benefits of vaccination is necessary in order to facilitate accelerated development and introduction of safe and effective dengue vaccines into public-sector programs of dengue-endemic countries.
These three endemic countries will soon face decisions on whether, and how, to incorporate current and future vaccine candidates within their highly budget-constrained national vaccination programs [8]. As population-wide vaccination campaigns must be carefully determined considering the allocation of scarce resources among competing health problems [13], the updated economic burden estimates of dengue from this study can be used to develop cost-effective vaccination strategies and sustainable financing plans for dengue vaccine introduction in Vietnam, Thailand, and Colombia. Furthermore, taking advantage of the standardized study design in a multi-country setting, this study attempts to explore and understand socio-economic factors affecting the level of disease severity that results in different levels of economic burden.
Table 1 summarizes the study areas for three countries, and the overall study design is shown in Fig 1. The economic burden survey was conducted in locations where DVI’s epidemiologic studies were being carried out at the same time. The health facilities were chosen because they were main hospitals and centers providing care for the defined catchment area population. In Vietnam, Khan Hoa General Hospital is a tertiary care facility with 1,000 beds and the largest hospital in the province. The hospital is staffed with 195 doctors, 59 assistant doctors, 5 pharmacists, and 336 nurses. Bang Phae Community Hospital (BPCH) is a primary healthcare provider in Bang Phae district in Thailand. The facility is a 48-bed medium-sized tertiary care facility which conducts up to 400 outpatient consultations per day. Two health facilities were chosen in Colombia. Clinic of Piedecuesta is a secondary and tertiary care facility with a total of 42 beds. The clinic has 19 general practitioners, 36 specialists and assistants, and 39 nurses. Hospital de Piedecuesta is a 27-bed medium sized primary care facility that conducts up to 76.9% of its consultations for outpatients, with the rest being for the emergency service. Among patients who came to one of the study facilities listed in Table 1, febrile patients experiencing fever for less than 7 days were recruited for rapid tests. The economic burden survey was administered for those who were positive on the test result and who consented to participate in the survey. There were three sections in the cost-of-illness (COI) survey: day 1, day 10~14, and day 28. As part of the study procedure, the febrile patients were tested on the first day of their clinic visit, and COI surveys were administered to patients throughout the duration of their illness. On day 1, dengue-confirmed patients were asked whether they had visited any health facilities prior to the current visit (visit 1). If they had made any visits, interviewers asked how much money they had paid for direct medical and direct non-medical costs such as facility fees, medications, food, lodgings, and transportation, etc. Patients were reminded to come back to the hospital in 10 to 14 days, and, so as to prevent recall bias, they were provided with a diary card to record costs for further treatments, and other indirect cost information, such as substitute labor or caregivers. On the second visit (day 10 to 14), direct medical and non-medical costs spent during the first visit and any additional visits before the second interview were collected. In addition, patients were asked if any indirect costs had occurred such as wage loss, costs for substitute labor and/or caretakers. Once direct and indirect cost information was obtained, interviewers asked patients about socio-economic questions such as patients’ household income, education, general perceptions towards dengue fever, and routine vector control activities, etc. If there were any patients who were still feeling sick at the end of their second visit, a third interview was scheduled for day 28. The third questionnaire covered the period between day 10–14 and day 28. The day 28 survey could be conducted via telephone rather than in-person if necessary. The project-related tests were omitted from the analyses, as these may not have been incurred under routine clinical practice.
Direct cost can be divided into two categories: (1) direct medical costs; and (2) direct non-medical costs. Direct medical costs include consultation, medication, and laboratory test costs. Patients’ out-of-pocket costs for these categories were first assessed through a questionnaire. Respondents were asked whether total treatment cost was covered by the patient (or his/her family members), insurance (private, public), or both. Because payments made by patients may not always cover the entire costs of medical treatments, hospital bill records for patients who were enrolled in the study were accessed to identify treatment costs that were not captured by patients’ out-of-pocket costs. In Colombia and Thailand, the hospital records were computerized and showed different sources of payments such as from private / public insurance, etc. However, no such system was available in Vietnam, thus medical service utilization forms, which recorded the type of services provided and hospital charges for each service, were employed for enrolled dengue fever patients in Vietnam. The data from the private surveys was then linked to the treatment records of patients (hospital charges) to better understand the full spectrum of the costs of dengue and how the economic burden is distributed among the public and private sectors.
Indirect cost consists of three components: patient wage loss due to illness; substitute labor cost; and caretaker’s cost.
While dengue infection usually causes flu-like illness with mild symptoms, the infection occasionally develops into a potentially lethal complication called severe dengue [26]. Severe dengue has become a leading cause of hospitalization and death in some Asian and Latin American countries.
The DVI’s cost-of-illness (COI) surveys include questions not only on direct and indirect cost information of dengue fever, but also on general socio-economic information such as education, routine vector control activities, and general perceptions towards dengue fever. Questions were identically administered at all three sites, and these variables can be compared by disease severity. In this study, severe illness was defined if a patient was either hospitalized or confirmed as having dengue hemorrhagic fever (DHF) based on clinical symptoms. Whether a patient goes through severe illness or not is a dichotomy, and a dichotomous dependent variable can be considered as a function of healthcare-seeking behavior, education, and household perceptions on dengue fever and likelihood, etc. A logit model is suitable for this analysis because such models handle categorical dependent variables using maximum likelihood estimation. The binomial logistic regression estimates the log odds that individuals will be in each of two categories of a dichotomous dependent variable.
ln(π1−π)=∑k=0Kxkβk
where K is the number of independent variables, xik, β are the coefficients, and the dependent variable is the log of the expected probabilities of being in each of the two categories, conditional on the values of the independent variables. Given that the surveys were standardized across the three sites, all samples were combined to achieve a larger sample size for robust model outcomes.
The cost-of-illness survey questionnaires were approved by the ethical review committees in three countries (National Institute of Hygiene and Epidemiology in Vietnam, Faculty of Tropical Medicine, Mahidol University in Thailand, Universidad de Santander in Colombia), as well as Ministry of Health in three countries and Institutional Review Board of the International Vaccine Institute. Written informed consent was obtained prior to conducting interviews, and respondents were informed that they could terminate interviews at any time. If any study participants were minors, their parents or guardians provided consent on behalf of all child participants under the age of 18 years old.
Table 2 shows patient characteristics for each study site by patient type (see S1 Table for additional information). The average number of total sick days ranges from 7 to 10 for inpatients and 6 to 9 for outpatients. While a majority of the patients sought treatment prior to the study enrollment in Vietnam, less than half of the patients did so in Thailand and Colombia. More patients tended to have caregivers than substitute labor in all three countries. In Vietnam, more than 80% of the patients responded that they had caregivers. On average, patients older than 5 years were not completely able to perform their usual activities for 5 to 7 days as inpatients, and 3 to 5 days as outpatients. The average age of the patients was 20–23 years in Vietnam and Colombia, and 15 years in Thailand. The self-reported mean household income per month was $423, $548, and $665 in Vietnam, Thailand, and Colombia, respectively.
Fig 2 summarizes the proportions of patient burden by expenditure, as well as the percentage share of patient’s out-of-pocket (OOP) costs versus non-OOP costs, such as government subsidy or payments from other insurance schemes. For all three countries, indirect cost accounts for the highest proportion of patient’s OOP among the three expenditure types in Fig 2(A): direct medical cost (DMC), direct non-medical cost (DNMC), and indirect cost (IC). While DMC is the second highest burden for patients in Vietnam and Colombia, DMC is only 2% in Thailand. This is mainly because people in Thailand are covered by its universal healthcare system so patients do not have to pay for most of the OOP costs related to direct medical services. This is more apparent in Fig 2(B). About 99% of DMC are covered by the public sector, which is the universal healthcare system in Thailand. The proportion of the non-OOP costs is also high in Colombia, at 81%. In Vietnam, public DMC coverage is less than that of the other two countries, meaning that more patients tend to pay directly for the medical services that they receive.
The economic burden of dengue fever is shown in Table 3. DMC is highest among the three expenditure types, after taking into account both private and public spending. Total cost per inpatient episode is $200, $141, and $385 in Vietnam, Thailand, and Colombia, respectively. In the case of outpatients, total cost per episode ranges from $40 to $158, with Colombia having the highest and Thailand having the lowest. It is interesting to see that IC and DNMC are higher in Vietnam than in Thailand despite the higher average income in Thailand. This is because there was no reported substitute laborer cost in Thailand, and the duration of caregiver’s work was shorter in Thailand than in Vietnam. In addition, while the meal service was included for inpatients in the hospital charge in the health facility of Thailand, all patients had to pay extra for food in Vietnam. Considering the mean duration of the disease, total cost per day was calculated. It appeared that Colombia had again the highest cost per day, followed by Vietnam and Thailand. The societal total costs were also estimated after adjusting for the ratio of cost-to-charge (RCC). While the total cost per episode went up after the adjustment in Vietnam and Thailand, it went down in Colombia. By age group, the total cost per episode is higher for the older age group than the younger age group in Thailand and Colombia, but this is the opposite in Vietnam as shown in Fig 3. This is mainly due to the high number of inpatients than outpatients in the younger age group in Vietnam. The economic burden was further disaggregated by age group, patient type, and cost type in Table 4.
The economic burden of dengue fever would have varying impacts on household equity. For those who reported their monthly income, three income groups were first generated based on percentiles of monthly income: low-income group (income ≤ 25%), middle-income group (25% < income ≤ 75%), and high-income group (income > 75%). The list of household assets in the survey was designed to include various items from essentials to luxury goods reflecting the local context of each country. Given this, the level of the household assets in each income group was identified, and the respondents who did not report their monthly income were categorized into the three income groups based upon their levels of household assets. The mean income value of the group to which these households belonged was assigned. The total OOP for the economic burden of dengue fever was then estimated as a proportion of the self-reported household monthly income by income group. As shown in Fig 4, the percentage of the economic burden of dengue for the low-income group was 36%, 17%, and 45% in Vietnam, Thailand, and Colombia, respectively. The percentage of the economic burden of dengue fever decreased moving towards the higher income groups. Due to the universal healthcare system, the percentage of the DMC burden was less than 1% for all three income groups in Thailand, which reduces the overall OOP burden for patients.
Table 5 shows the regression outputs. While the type 1 model is parsimonious and only includes the variables related to the current illness, the type 2 model was run with all possible covariates including socio-economic variables. As expected, the number of sick days before and after the study enrollment was highly significant and positively related to severe illness. The seeking-early-treatment variable is also statistically significant at the 1% level and decreased the odds of experiencing severe illness for both models. This indicates that patients who sought treatment at the early stage of illness would be less likely to develop severe illness later on. While patients with severe illness tended to have more substitute labor or caretaker(s) than patients with non-severe illness, this variable is marginally significant in type 2. Compared with patients with no education, patients with some or higher education were less likely to have severe illness for the current episode. Age, household income levels, and vector control activities were not statistically significant. In addition, the dengue perception score was developed based upon respondents’ general knowledge of dengue fever, and was categorized into three levels with the lowest score indicating the poor level of knowledge on dengue fever. Having less awareness or lacking general knowledge of dengue fever was positively associated with experiencing severe illness. It was also shown that patients were less likely to develop less severe illness when there were more household members who had been previously infected and recovered. Overall, the type 2 model was preferred over the type 1 model based on the difference of the Akiake Information Criterion (AIC) values [27]. It should be noted that some of the milder cases may never receive medical attention, and this may be related to income and education levels. While this would be difficult to be statistically tested in this study given the absence of healthcare-seeking behavior indicators outside the study facilities, the distributions of non-severe and severe patients appear to be similar over income and education levels, holding other factors constant (see S1 Fig). This partly indicates that there was no serious selection bias in the dataset.
Estimates of the economic burden of dengue fever may vary depending upon a country’s healthcare system and the different methodologies applied. While there were existing economic burden studies, the advantages of the current study are that it relies on the implementation of a standardized survey tool and the inclusion of various expenditure types and age groups in multiple countries. This makes it possible to make a more complete comparison across the three dengue-endemic countries. Using the standardized methodology, this study found that the total cost per dengue episode is $200, $141, $385 for inpatients and $62, $40, $158 for outpatients in Vietnam, Thailand, and Colombia, respectively. Among the various OOP expenditure components, indirect cost was the biggest private burden for patients in all three countries. Non-OOP payments, such as from insurance or nationwide healthcare system, reduce patients’ burden of DMC. The patients’ share of burden of DMC was the least in Thailand, followed by Colombia and Vietnam. In all three countries, the percentage of the private economic burden of dengue fever was highest in the low-income group and lowest in the high-income group.
Though many previous studies were not based on standardized methodologies, making them difficult to compare [3,4], some existing studies that did not use secondary data sources or extrapolation were compared with the current study outcomes (see S4 Table for summary). Tam et al. reported $167.8 per DHF inpatient episode in 2006 in Vietnam [28]. Once the estimate was inflated to 2014 USD using consumer price index, this value was higher than our estimate. However, it should be noted that this previous study only looked at DHF inpatients who were supposed to experience more severe illness. Another study by Harving et al. showed the average total cost of $61.4 per DHF inpatient younger than 15 years in 2005 [29]. Our estimate per inpatient episode for the same age group is $125.2, which is less than Harving et al.’s estimate after inflation adjustment ($147.1), but again Harving et al. also looked only at DHF inpatients. In Thailand, while Anderson et al. and Clark et al. reported the average total cost of $31.8 to $44 per inpatient and $10.2 per outpatient [14,30], Okanurak et al. estimated the total cost per inpatient as $172.9 and $141.5 from two different sites in Thailand [31]. The estimates from Anderson et al. and Clark et al. were lower than our study estimates even after inflation adjustment. One of the main reasons is that the former two studies focused on patients’ OOP costs, whereas our study took into account both patients’ OOP payments and public payments, which would make for a large difference especially in the context of Thailand. On the other hand, Okanurak et al. looked at DHF patients, resulting in higher costs than our study estimates. Suaya et al. also estimated dengue cost of illness in Thailand and reported $573 for inpatients, which is a lot higher than the other study estimates including ours [5]. Rodriguez et al. estimated a total cost of $497.9 per inpatient and $202.3 per outpatient in Colombia. There are similarities between this previous study and our study in terms of the cost items considered and the sub-component estimates. They reported higher indirect cost estimates, making the overall costs higher than our estimates. This could be partly because the previous study included projected lost income due to death, whereas the current study did not have any deaths occurred during the study period.
In addition to the estimation of the economic burden of dengue fever, the logit model was constructed to identify factors explaining variance of disease severity. The logit analyses showed that early treatment, higher education, and better knowledge of dengue disease would be associated with a reduction of the probability of developing more severe illness. This can be partly explained by the fact that secondary dengue infection which would likely cause more severe illness would be inversely correlated with the independent variables. As the disease ranges from flu-like mild symptoms (dengue) to potentially lethal complications (severe dengue), resulting in different levels of economic burden, a community would be better off expanding the factors identified in this study, which may prevent its members from developing severe illness in the long run.
Some areas of uncertainty deserve attention. While the study captured all patients’ OOP costs spent during the entire period of illness by conducting multiple interviews, public DMC payments and RCCs were limited to our study’s health facilities. Because a majority of the costs were incurred during the enrollment visit at the study’s facilities rather than at referral facilities or pharmacies across the study sites, the estimates would not be greatly influenced. However, our estimates are conservative. In Colombia, where two health facilities were involved, the RCC was only obtained from one of the facilities due to logistical issues. Because these study facilities were in the same district, a single RCC was used for all patients. While various cost and charge items were taken into account for the estimation of the RCC, the detailed list of financial information was confidential for some of the health facilities, limiting the way that the estimates were presented. Ideally, our study samples should be more heterogeneous and representative of the entire countries. Given that this study was carried out in locations where epidemiologic studies were being conducted at the same time, care must be taken when generalizing the estimates beyond the study’s communities. Nonetheless, hospital charges may vary in different locations of a country, but within the same country the unit prices for service items may not be as variable as the charges. Assuming the duration of illness is similar within a country, factors which may influence the total cost adjusted by the ratio of cost-to-charge in the same country are patient’s out-of-pocket expenditures for direct non-medical costs and indirect costs. Because these are directly related to the income level in a community, it would be possible to understand the representativeness of our estimates by comparing the average income of the study communities with Gross National Income (GNI) per capita. The study refusal rates among eligible patients were 2.3% and 1.4% in Thailand and Colombia respectively. Due to logistical issues, the refusal rate could not be estimated in Vietnam. While there were no patients who felt sick after the 2nd interview of the current study, Tiga et al. reported that some dengue patients experience persistent symptoms such as asthenia, fatigue, and trouble working [32]. It should be noted that there may be patients who do not generally come to the professional health sector such as our study facilities [33]. While the current study captured any non-medical sector visits such as traditional healers or others before and after the study enrollment, there may have been non-enrolled dengue patients who did not come to any of our study facilities during the study period, and this may have been partially related to patient’s education and income levels.
The burden of dengue has been rising partly due to the absence of population-wide vaccination campaigns and the limited impacts of vector control activities. A previous study assumed that vector control would not be able to achieve permanent reduction of dengue based on the experience of Singapore [13]. The first live attenuated, tetravalent dengue vaccine (CYD-TDV) has been already licensed in some dengue-endemic countries. To maximize the effective use of a dengue vaccine, vaccination strategies should be carefully considered after taking into account various competing health problems at the country level. Identifying effective vaccination strategies requires various sources of information such as accurate incidence rates, the long-term behavior of the vaccine (i.e., waning rates of efficacy, the rate of any adverse effects, etc.) and the price of the vaccine. Along with more details, this study’s estimates of the economic burden of dengue fever can play a critical role in determining cost-effective vaccination strategies, and so facilitate the process of dengue vaccine introduction by guiding decision-makers on the expected societal health benefits through vaccination.
|
10.1371/journal.pntd.0006763 | Rural youths' understanding of gene x environmental contributors to heritable health conditions: The case of podoconiosis in Ethiopia | Assess the feasibility of engaging youth to disseminate accurate information about gene by environmental (GxE) influences on podoconiosis, a neglected tropical lymphedema endemic in southern Ethiopia.
A cross sectional survey was conducted with 377 youth randomly selected from 2 districts of Southern Ethiopia. Measures included GxE knowledge (4 true/false statements), preventive action knowledge (endorse wearing shoes and foot hygiene), causal misconceptions (11 items related to contagion) and confidence to explain GxE (9 disagree/agree statements).
Over half (59%) accurately endorsed joint contributions of gene and environment to podoconiosis and preventive mechanisms (e.g., wearing protective shoes and keeping foot hygiene). Multivariable logistic regression showed that youth with accurate understanding about GxE contributors reported having: some education, friends or kin who were affected by the condition, and prior interactions with health extension workers. Surprisingly, higher accurate GxE knowledge was positively associated with endorsing contagion as a causal factor. Accuracy of GxE and preventive action knowledge were positively associated with youth’s confidence to explain podoconiosis-related information.
Youth have the potential to be competent disseminators of GxE information about podoconiosis. Interventions to foster confidence among youth in social or kin relationships with affected individuals may be most promising. Efforts to challenge youth’s co-existing inaccurate beliefs about contagion could strengthen the link of GxE explanations to preventive actions.
| This study considers the feasibility of engaging rural Ethiopian youth as lay health workers (LHWs) with the objective to improve community understanding of the joint influences of genetics and environment on health. Identifying LHWs to accurately convey contributors to the heritable but preventable neglected tropical disease of podoconiosis provides an optimal context to address this question. Misunderstanding that inherited susceptibility to podoconiosis makes the disease unavoidable has led to numerous negative social consequences (e.g., stigma) and poor uptake of protective footwear. We report data from a pilot study that included a cross-sectional survey of youth ages 15–24 in rural communities with endemic podoconiosis. Results provide preliminary support that a sizeable group of youth hold accurate knowledge about gene x environment influences and self-evaluate as being confident to explain these associations to others. Research to evaluate strategies to engage youth as LHWs and the impact of these approaches on communities’ understanding of the joint influences of genetics and environment in this context is needed. This manuscript fills an important gap in the literature about neglected tropical diseases, as it suggests opportunities to improve the prevention of podoconiosis and reduce misconceptions and stigma through engagement of LHWs in Ethiopia.
| Advances in genomics are increasing scientific understanding that most health conditions worldwide are caused by the joint influence of genetic and environmental (GxE) factors [1]. However, the mechanisms underlying GxE interactions are complex and not well understood by the public [2]. Accordingly, misunderstandings that health conditions with genetic underpinnings cannot be prevented have been well documented in the developing world [3–6]. Global leaders have called for stepping up efforts to increase genomic literacy in low and middle-income countries (LMICs) [7] [8, 9]. Among the challenges to achieving this imperative is that LMICs have limited health service infrastructure, low levels of general literacy and a majority of the populace lives in isolated rural settings. Funding opportunities such as the Human Heredity and Health in Africa (H3Africa) have been initiated to address these challenges [10].
Concurrent with advances in genomics, around the world, cadres of lay health workers (LHWs) have been engaged to expand the reach of preventive health services for a broad array of communicable and noncommunicable health conditions [11, 12]. LHWs carry out a number of functions to promote health including the provision of culturally relevant health information that can be conveyed in everyday community settings [13, 14]. Though it has been suggested that such LHWs might play an important role in promoting community-wide GxE literacy in the developed world to date there has been little consideration of this possibility in LMICs [15, 16] [17].
Efforts directed to disease prevention often target youth as a means to encourage the establishment of healthy behaviors. Considering youth as potential disseminators of information about GxE health influences offers several other advantages. As members of the community, youth are aware of existing interpretations of the causes of health conditions held by their peers and adults in the community [18]. In the context of HIV/AIDs and other STIs, reproductive health, and malaria, studies suggest that youth may more easily process and understand new information [3]. Youth also have been found to have greater accuracy in considering the influence of genes on health conditions than older adults [19]. These information-processing assets have been attributed to youth’s neurodevelopment stage, exposure to schools, youth networks and the media [20]. The feasibility of engaging LHWs to increase GxE literacy in LMICs rests on whether individuals can be identified with the capability and confidence to serve as GxE information disseminators.
Podoconiosis, a non-filarial lymphedema endemic in highland Ethiopia, offers an excellent context to consider the feasibility of engaging youth LHWs for promoting GxE literacy. Globally, it is estimated that four million people in tropical Africa, Central and South America, and Southeast Asia have the condition; approximately 1.5 million people are affected in Ethiopia [4]. The condition develops when genetically susceptible individuals are exposed to irritant particles in volcanic soil via walking and farming barefoot [21–23]. Podoconiosis can be prevented if susceptible individuals begin wearing shoes at an early age and do so consistently [21–23].
Prior health education efforts to promote shoe wearing among at-risk individuals have been based on extensive qualitative and rigorous quantitative evaluation of these interventions show that adults in these communities continue to harbor misconceptions that the condition is not preventable because of its heritability [5, 24, 25] [5, 6] [26]. A number of other misconceptions are common including beliefs that the condition is contagious or influenced by other environmental exposures (e.g., snake bites) [6, 25, 27, 28]. In turn, these inaccurate beliefs are associated with risk behaviors such as consistently walking barefoot [5, 6, 25] and holding stigmatizing attitudes towards patients [5, 6, 29, 30]. Whether engaging youth as disseminators of accurate GxE information could help to eliminate these misconceptions, reduce social stigma towards affected individuals and encourage preventive actions (i.e wearing shoes) is largely unexplored.
In order to consider the feasibility of engaging rural Ethiopian youth to serve as information disseminators in the spirit of LHWs, we posed four research questions: (1) what is the prevalence of youth misconceptions about the causes of podoconiosis?; (2) what factors are associated with accurate understanding of GxE contributions to podoconiosis among youth?; (3) is GxE accuracy associated with correct endorsement of preventive actions?; and (4) is GxE accuracy associated with confidence to explain the causes of podoconiosis?
We conducted a cross-sectional survey in rural communities of the Wolayita zone of Southern Ethiopia with endemic podoconiosis [31]. Data were collected between August and September 2016. Two kebeles or lowest administrative units (Tome Gerira and Sura Koyo) were purposefully selected as neither had participated in prior studies of podoconiosis. Youth defined by United Nations as ages 15 to 24 were eligible to participate in the survey [32]. We used a random sampling strategy with all youth living in the area having an equal chance to be included in the sample. After conducting a census of youth in the participating kebeles 3,542 were identified in the target age range. 52% (n = 1,841) were females. Of youth identified, 11% (n = 383) were from families affected by podoconiosis. The optimal sample size required to detect differences of plus or minus five units in survey responses (with 95% confidence) was 347 [33]. We approached 377 youth in anticipation of 5% refusal to participate found in our prior studies.
Our prior qualitative findings suggested that there was a substantial knowledge disparity regarding podoconiosis causes and preventive actions between youth who were from affected families than youth who were unaffected [28]. Research to date has indicated that affected adults tended to endorse some explanations of the causes of podoconiosis more strongly than others [6]. Thus, we aimed to proportionally represent affected youth in the sample size. To this end, our sample comprised 41 affected youth and 336 unaffected youth randomly selected from the census sampling frame. This ratio of unaffected to affected youth is representative of the population in the study districts.
The questionnaire was initially developed in English, translated into Amharic and Wolayitigna (common dialect in the targeted kebeles), then checked for accuracy and back-translated to English. Interviewers who had previous experience in similar studies were recruited using informal networks. These individuals (N = 10) received three-days of training on the objectives of the study, items included in the survey questionnaire, and how to carry out the survey. The survey was pilot tested in two nearby rural kebels (Sodo Zuria and Damot Woydie woredas) with 30 youth to assess the adequacy of the instruments. These youth were not included in the main survey. Results of the pilot indicated a few limitations such as spelling or grammatical errors that resulted in respondents showing minor hesitation and request for clarification revisions were made.
Interviewers administered the survey to the selected youth at their homes. The expectation was that the interviews would take approximately 45 minutes. Written informed consent was obtained from all participants including thumbprints for those unable to sign. Consent from parents or guardians was obtained for those ages 15 to 18. The respondents were given exercise books and pens as compensation for participation. All aspects of this study were approved by the Addis Ababa University College of Health Science Institutional Review Board.
Selection of survey measures assessed domains of knowledge, and self-efficacy related to conveying GxE information. These constructs of Social Cognitive Theory [34, 35], were based on prior literature on requisite competencies for LHWs to be effective [36] and on our extensive prior qualitative data collected with adults in these communities [28]. Demographic characteristics assessed included: gender, education level, and age. Additional measures of civic engagement and interactions with health extension workers also were assessed.
Three domains of knowledge were assessed:
GxE knowledge was based on four questions that were adapted to the causal factors contributing to podoconiosis [23, 25, 37]. Youth were asked to rate each statements as “true,” “false” or “don’t know” (e.g., “A person can inherit proneness to have their feet be irritated by the soil.”). These measures were dichotomized based on the median score (2.1). Youth who scored three and four were labeled as “mostly accurate” (1) and those who scored less than three were labeled as “mostly inaccurate” (0).
Knowledge about preventive actions was measured by rated agreement with a list of 10 potential preventive actions that included both accurate and inaccurate options (e.g., wearing protective shoes every day–accurate; taking vaccination–inaccurate). Each accurate response was assigned one point. The preventive action score was dichotomized due to skewedness, with zero representing inaccurate understanding (0–3) and one indicating mostly accurate understanding (4–10).
Causal misconceptions were based on 14 questions derived from prior research conducted with adults living in podoconiosis-endemic communities [26]. Misconceptions were assessed with two subscales, a “contagion score” based on responses to 11 questions assessing perception of podoconiosis as a contagious disease. A second subscale was labeled “other misconceptions” and included three questions related to beliefs about bacteria, poor nutrition and evil eye as causes of the disease. Each endorsed misperception was assigned ine point and then summed for a total score.
Confidence to explain (i.e., self-efficacy) was based on nine questions (e.g. “I am confident that I could explain to other people why some individuals develop podoconiosis and others do not”) with three response categories (1 = Disagree, 2 = Undecided, 3 = Agree) [26]. Scores were dichotomized as being above or below the median, where zero represented “less confidence” and one indicated “more confidence”.
Extracurricular civic engagement was based on self-reported involvement in extracurricular activities at school, youth association, Sunday schools and other leadership-oriented experiences.
In Ethiopia more than 30,000 health extension workers provide outreach services and disseminate health information to the general public to encourage health promoting habits (Federal Ministry of Health, 2007). Reported contact with health extension workers (HEWs) was assessed as they may have provided opportunities for youth to increase knowledge of health issues and podoconiosis, specifically. Youth who were visited by HEWs at their residence one year prior to the study were compared to those reporting no contact or visit.
Data were analyzed using SPSS version 20 software. Frequencies and distributions were examined to check for out-of-range values and other errors in the data. After data cleaning, descriptive analyses, bivariate analyses and logistic regressions was performed. Cross tabulation and χ2 tests were performed for associations amongst variables. Level of statistical significance was set at p <0.05 (two-tailed). Variables with significant associations in cross-tabulation were entered into the logistic regression models to test in turn their association with accuracy of GxE knowledge, accuracy of endorsed preventive actions and confidence to explain GxE causes of podoconiosis. We ran Pearson’s correlation to assess the association between accurate GXE knowledge and prevalent misconceptions (i.e., inaccurate causal beliefs). Odds ratios were computed for the likelihood of having mostly accurate (1) and mostly inaccurate (0) understanding about GxE influences and appropriate preventive actions. Odds ratios were also computed for the likelihood of being more confident or less confident to explain podoconiosis-related information.
All 377 who were approached agreed to participate in the survey; 52% were female (See Table 1). The majority of youth participants were in the age range of 15–18 (78.5%). Accordingly, the majority of youth (70%) reported being in 7th to 12th grade education. Just over half reported involvement in some extracurricular civic engagement. Approximately 80% of the youth had contact with HEWs at least once in the year prior to the survey. Eleven percent of participants reported having a blood relative with podoconiosis.
Half of youth believed that podoconiosis is a contagious disease and over one-third also endorsed several causal factors that constitute misconceptions including evil eye, and snake bites. Eleven percent of participants responded correctly to all four of the GxE knowledge questions related to podoconiosis, 34.5% answered three correctly, and 55% answered two or less questions correctly. The highest proportion of correct responses (82%) was agreement with the question “walking barefoot triggers podoconiosis among susceptible individuals” (See Table 2). Additionally, 61% agreed that susceptibility to podoconiosis is passed from generation to generation.
A number of preventive actions were erroneously thought to be protective against podoconiosis, including vaccination (87%) and avoiding walking barefoot in cold weather or dew (87%). Additionally, one-third of youth incorrectly suggested avoiding personal contact with affected people, and nearly three-quarters endorsed avoiding wearing second-hand shoes, both indicators of belief in contagion. However, a majority of the youth also accurately endorsed regular foot washing (86%) and wearing protective shoes everyday (70%).
Due to their high co-prevalence, we tested the association of accurate GxE knowledge with prevalent misconceptions (e.g., endorsement of podoconiosis as a contagious disease). We found a modest positive correlation between GxE knowledge and contagion beliefs (r = 0.23, p<0.05). Youth with more accurate understanding of GxE contributors to podoconiosis also were more likely to harbor “other misconceptions” (r = 0.13, p<0.05).
Participants with 7th to 12th grade education, who reported having affected family members or a friendship relationship with an affected individual were significantly more likely to be accurate in their understanding (correct responses to 3 of 4) of GxE contributors to podoconiosis than their peers (See Table 3). Endorsement of contagion as a cause of podoconiosis and interactions with HEWs also were significantly associated with accurate understanding of GxE contributors.
Variables found to be significant in bivariate analyses were tested in multivariate logistic regression to predict accurate GxE knowledge (See Table 4). Affected status, having friendships with affected individuals, reporting any contact with HEWs, having some formal education and having a high contagion inaccuracy score were significantly associated with increased accuracy of GxE understanding.
Youth who had attended formal schooling from grade 1st to 6th or 7th to 12th education level had 11.8 and 8.7 times, respectively (95% CIs: 2.2–16.6; 1.7–13.4) the odds of having accurate GxE knowledge related to podoconiosis as those without formal schooling. Affected youth had 8.0 times the odds of being accurate in GxE knowledge as unaffected youth (95% CI: 2.9–16.1). Similarly, youth with affected friends had three times the odds of having accurate understanding as those who did not report having friends with podoconiosis (95% CI: 1.3–6.9). Youth who reported contact with HEWs had 3.8 times (95% CI: 2.0–7.2) the odds of knowing about the joint contribution of GxE in the development of podoconiosis as those who did not report contact with HEWs. Youth with stronger contagion beliefs had 1.2 (95% CI 1.0–1.2) times the odds of holding accurate GxE knowledge as those with weaker contagion beliefs.
We tested a logistic model that included the factors found above to be associated with accuracy of understanding in a model for appropriate endorsement of preventive actions (dichotomized as 0 (1–3) = mostly inaccurate and 1(4–10) = mostly accurate). Results showed that youth who reported having a friendship with an affected individual had 2.4 times (95% CI: 1.0–5.4) the odds of of endorsing appropriate preventive actions as those without affected friends (See Table 4). Youth who reported contact with HEWs had 3.8 times (95% CI: 2.0–7.0) the odds of endorsing appropriate preventive mechanisms as those who did not report contact with HEWs. Youth who endorsed contagion as a cause of podoconiosis had lower odds of accurately endorsing preventive actions as those who did not endorse contagion beliefs (OR = 0.92, 95% CI: 0.8–1.0). Counter to our expectation, accurate understanding about GxE contributors to podoconiosis was not associated with endorsement of appropriate preventive actions (e.g., shoe wearing, p = 0.92).
We tested a logistic model that included the factors found above to be associated with appropriate endorsement of preventive action to test their association with confidence to explain the causes of podoconiosis to others (dichotomized as 0 (1–3) = less confidence, 1 (4–10) = more confidence). Results showed that those affected by podoconiosis had lower odds of being confident to explain GxE influences as unaffected youth (OR = .98; 95% CI: 0.1–1). Accuracy of GxE and preventive action knowledge were significantly associated with confidence to explain the causes of podoconiosis (See Table 4). For each unit increase in GxE accuracy and endorsement of preventive action, the odds of being confident increased by a factor of 1.3 (95% CI: 1.0–1.7) and 1.3 (95% CI: 1.1–1.5) respectively, when compared to the reference group.
Our findings show that youth in podoconiosis-endemic communities have similar misconceptions as those reported previously among adults in neighboring communities [4, 6, 25, 38]. Like adults, youth endorsed multiple causes of podoconiosis that included both contagion and environmental exposures (e.g., snake bite, evil eye). However, a sizable proportion of youth responded accurately to questions about the joint contributions of heredity and environment in the development of podoconiosis. It is concerning that youth with the most accurate understanding of GxE contributors to podoconiosis also were most likely to harbor misconceptions about the causes of podoconiosis. Thus, endorsement of GxE contributors did not obviate other erroneous causes such as contagion.
Our previous qualitative work with youth in these communities that aimed to understand youth’s explanatory mental models documented a similar connection between GxE and contagion beliefs [28]. In our previous report we posited that beliefs that susceptibility to podoconiosis passes from generation to generation (i.e., accurate GxE knowledge) co-occurred with observations of families living in close proximity to each other where bodily fluids could be exchanged. Thus, youth perceived a “both/and” connection in which inherited susceptibility and contagion explanations could co-exist.
Generally, health education programs assume that accurate health information can correct inaccurate beliefs. However, our findings suggest that community members can hold numerous, even contradictory, notions of causation simultaneously. The implications of this present challenges for engaging youth as LHWs. Contagion beliefs not only shape priorities given to preventive actions but also fuel stigmatizing beliefs. Endorsement of both GxE and contagion as causes of podoconiosis also may have attenuated the association between related knowledge and endorsement of appropriate preventive behaviors. Thus, any efforts to engage youth as LHAs to promote preventive actions among affected individuals would also need to challenge beliefs about contagion.
A large body of literature has shown that contagion beliefs have deep roots and are very difficult to override [28, 39]. In this study, youth who reported having social ties to affected individuals were more likely to be accurate in their understanding of GxE contributors. Similarly in our prior intervention study, affected adults were less likely to endorse contagion beliefs than unaffected adults [26]. Individuals with social ties to affected individuals, particularly those who have engaged with health extension workers may be more accepting of alternative causal explanations that are linked to preventive actions.
Level of education was also associated with GxE knowledge accuracy. Youth without formal education had lower knowledge about podoconiosis, which is consistent with earlier studies among adults [24, 40, 41]. Lack of schooling limits youth’s access to health information provided in school curriculums and extracurricular activities. Youth engaged in formal education could be optimal disseminators of accurate GxE information. Indeed, prior studies identified some schooling to be a crucial qualification for LHWs in Peru and the Republic of the Congo [42, 43]. However, education level was not associated with accurate knowledge of preventive actions or with confidence to explain in this sample. LHA training targeted to school settings would need to strengthen these skills among youth.
The study further showed that accurate knowledge about GxE contributors of podoconiosis may not necessarily lead to understanding accurate preventive actions. Experts argue that individuals’ understandings of causes of health conditions have strong influences on what they perceive as best preventive actions to take to lower risk of developing diseases [44, 45]. The logical assumption is that provided with accurate causes of health conditions, individuals will be more likely to identify preventive actions to lower the risk of developing ailments. The concern is that holding a number of competing beliefs about the causes of podoconiosis could muddle youth’s understanding about what to do to prevent the conditions.
LHWs training should not assume that improved understanding and acceptance of GxE causal mechanisms will override beliefs about contagion. Curricula specifying the mechanism through which soil exposure (e.g., silica particles) leads to lymphatic inflammation could be included in school science classes. In this environment, classroom exercises could be used to encourage youth to consider these mechanisms and debate the other causal beliefs using peer discussions and co-teaching.
It is heartening that youth with higher GxE accuracy and preventive knowledge also had the highest confidence in their ability to explain the causes of podoconiosis. It has been suggested that effective LHWs have conceptual understanding and confidence to effectively communicate with their constituencies that, in turn, can be capitalized upon to facilitate individual- and community-level health promotion [36]. Similarly, a recent study found confidence to talk about breast and cervical cancer screening information to significant others to be an important individual characteristic for LHWs [36]. However, affected youth had lower self-confidence, despite their greater GxE knowledge compared to unaffected youth. This suggests that confidence-building strategies may need to be different for affected and unaffected youth. Stigma related to being affected by podoconiosis may have lowered self-esteem of these youth and inhibit their confidence to discuss the topic. It also should be noted that further assessment might be needed to identify the factors that motivate young people, both those affected and unaffected, to take part in podoconiosis prevention campaigns.
The study had several noteworthy strengths including a high participation rate, and the survey development was informed by extensive qualitative data collection. However, there were a few limitations. The study was conducted in only two rural communities that may not be generalizable to youth in urban areas or other LMIC settings. The survey focused specifically on factors associated with youth’s accuracy of knowledge of podoconiosis causes, GxE influences and preventive actions with the aim of informing health literacy-building activities. Numerous other factors could influence the viability of youth in these settings serving as LHWs. Indeed, studies show that role-related factors, social network and trust, and other characteristics of target communities can influence motivation and performance of LHWs that were not assessed [11, 36, 46, 47]. Moreover, knowledge improvements unto themselves may be necessary for health promotion but are unlikely to be sufficient to influence behavior change. Our findings of coincident endorsement of causal beliefs related to GxE and contagion may have been influenced by social desirability bias, with participants eager to endorse multiple causes and inherently multiple solutions to the problem.
These results are preliminary but support the pursuit of further research to consider youth assets and deficits relating to their potential role as LHWs. Additionally, extensive qualitative and quantitative studies will be required to guide identification of youth who may be most willing and competent in this role and to evaluate interventions to help youth acquire skills and assess community impact. The potential to avoid disparities in reach of emerging genomic knowledge warrants these efforts.
|
10.1371/journal.ppat.1007249 | A seven-helix protein constitutes stress granules crucial for regulating translation during human-to-mosquito transmission of Plasmodium falciparum | The complex life-cycle of the human malaria parasite Plasmodium falciparum requires a high degree of tight coordination allowing the parasite to adapt to changing environments. One of the major challenges for the parasite is the human-to-mosquito transmission, which starts with the differentiation of blood stage parasites into the transmissible gametocytes, followed by the rapid conversion of the gametocytes into gametes, once they are taken up by the blood-feeding Anopheles vector. In order to pre-adapt to this change of host, the gametocytes store transcripts in stress granules that encode proteins needed for parasite development in the mosquito. Here we report on a novel stress granule component, the seven-helix protein 7-Helix-1. The protein, a homolog of the human stress response regulator LanC-like 2, accumulates in stress granules of female gametocytes and interacts with ribonucleoproteins, such as CITH, DOZI, and PABP1. Malaria parasites lacking 7-Helix-1 are significantly impaired in female gametogenesis and thus transmission to the mosquito. Lack of 7-Helix-1 further leads to a deregulation of components required for protein synthesis. Consistently, inhibitors of translation could mimic the 7-Helix-1 loss-of-function phenotype. 7-Helix-1 forms a complex with the RNA-binding protein Puf2, a translational regulator of the female-specific antigen Pfs25, as well as with pfs25-coding mRNA. In accord, gametocytes deficient of 7-Helix-1 exhibit impaired Pfs25 synthesis. Our data demonstrate that 7-Helix-1 constitutes stress granules crucial for regulating the synthesis of proteins needed for life-cycle progression of Plasmodium in the mosquito vector.
| Plasmodium falciparum is a unicellular parasite responsible for the majority of 440,000 deaths due to malaria every year. Malaria is a vector-borne disease, transmitted by blood-feeding Anopheles mosquitoes. While the high-replicating red blood cell stages are responsible for the symptoms of malaria, the gametocytes, sexual precursors that form in the blood, are crucial for the transmission of the parasite from the human to the mosquito. Once taken up by the mosquito during blood feeding, the gametocytes rapidly convert to gametes, which mark the mosquito-specific phase of the parasite life-cycle. Due to their essential role for malaria transmission, gametocytes represent prime targets for transmission-blocking strategies intended to prevent spread of the disease. In this study, we explored the signaling pathways leading to gametogenesis and identified a hitherto unknown protein, which structurally belongs to the class of seven-helix proteins. The protein, named 7-Helix-1, is specifically expressed in female gametocytes and disruption of its gene leads to impaired gamete formation and reduced transmission of malaria parasites to mosquitoes. Deletion of 7-Helix-1 caused a significant imbalance of parasite components essential for protein synthesis. We thus propose that 7-Helix-1 is needed to coordinate the increased need of proteins at the onset of gametogenesis.
| In eukaryotes, proteins comprising seven helix domains are classified as receptors capable of binding a high variety of ligands. Because many of these receptors transduce signals to heterotrimeric G proteins, they are commonly referred to as G protein-coupled receptors (GPCR) (reviewed in [1]). Alternative names include serpentine receptors (SRs), seven-transmembrane receptors or seven-helix proteins. These proteins are usually integral parts of cell membranes and able to perceive a broad spectrum of stimuli, such as peptides, small molecules or light (reviewed in [1]). Recent studies have further reported on GPCR-like molecules that are not integrated in cell membranes but found in the cytoplasm of a variety of mammalian tissues, such as liver, brain or muscle cells. Due to their homologies with prokaryotic lanthionine cyclases (LanC) of gram-positive bacteria, these GPCR-like proteins were termed LanC-like (LanCL) proteins [2–5].
GPCRs generally transduce the signals via two main pathways, i.e. the cAMP pathway, during which a transmembrane adenylate cyclase is activated, and the phosphatidylinositol pathway, which is mediated by activation of phospholipase PI-PLC, leading to the formation of the second messenger IP3 and subsequent increase in intracellular calcium (reviewed in [6,7]). The physiological importance of the GPCR family is evident from its great expansion in the genomes of complex eukaryotes, with typically around 1,000 different members in mammalian species (reviewed in [1]). Noteworthy, GPCRs are involved in many diseases and are the targets of approximately 40% of all currently used drugs (reviewed in [8]).
Components of GPCR signaling cascades have also been identified in Plasmodium falciparum, the causative agent of malaria tropica, which is responsible for the majority of the ~440,000 deaths due to malaria per year [9]. Various studies demonstrated that the random contact of P. falciparum merozoites with red blood cells (RBCs) leads to PI-PLC activation and consequently to a rise in intracellular calcium, causing the discharge of specialized secretory vesicles, the micronemes. This allows secretion of microneme-resident adhesion proteins, which support RBC binding and invasion by the merozoites (reviewed in [10]).
Calcium-dependent signaling cascades are also involved in gametogenesis, which takes place in the midgut of the Anopheles vector (reviewed in [11,12]). Following uptake of the transmissible gametocytes by a blood-feeding mosquito, a cGMP-dependent signaling pathway induces IP3 synthesis and a rise in cytosolic calcium. The increased calcium levels subsequently activate calcium-dependent protein kinases, which control DNA replication during the formation of male gametes as well as the synthesis of proteins required for life-cycle progression in the mosquito (reviewed in [11–13]). Protein synthesis is particularly important during female gametogenesis, when translational repression of a variety of mRNAs encoding midgut stage-specific proteins, is released. These transcripts are stored in female gametocytes bound to non-translating messenger ribonucleoproteins [14–16]. The complexes of mRNA and ribonucleoproteins resemble stress granules (SGs), i.e. non-membranous cytosolic RNA-protein aggregates, which function in protecting mRNA stalled in translation initiation under cell stress (reviewed in [17–19]). The storage of transcript in female gametocytes is considered a pro-active measure of the parasite to rapidly adapt to the change of host.
While these signaling pathways suggest the involvement of GPCRs in signal perception, neither a heterotrimeric G-protein nor an IP3-responsive calcium channel has been identified until now in Plasmodium [20,21]. Nonetheless, four potential GPCRs comprising typical 7-helix domains were identified by the in silico analysis in the P. falciparum genome ([22]; reviewed in [23]). The proteins were termed SR1, 10, 12, and 25 according to their homologies with other eukaryotic SR family members. A recent study provided a functional analysis for SR25, which acts as a potassium sensor in blood stage parasites and, upon stimulation, activates PI-PLC leading to increased cytosolic calcium levels [24].
In this study, we have screened the genome of P. falciparum for additional GPCR-like proteins and identified a novel protein with typical heptahelical motifs. We show that this protein, termed 7-Helix-1, does not represent a transmembrane GPCR, but is linked to translational repressors in the stress granules of female gametocytes, where it is crucial for the regulation of protein synthesis at the onset of gametogenesis.
In order to identify potential heptahelical proteins of gametocytes we searched the P. falciparum genome for genes with homologies to eukaryotic GPCR. In addition to the genes of the previously described SR1, 10, 12 and 25 [22], we identified a gene encoding a putative GPCR (PF3D7_0525400), henceforth termed 7-Helix-1.
The gene 7-helix-1 encodes a 55.7 kDa protein that comprises seven predicted transmembrane domains (Fig 1A). Proteins homologous to 7-Helix-1 are encoded in the genomes of several other organisms, e.g. the Plasmodium species P. vivax and P. berghei (putative GPCRs comprising a LanCL domain), Homo sapiens (hLanCL1 and hLanCL2), Mus musculus (mLanCL1 and mLanCL2) and Arabidopsis thaliana (LanCL protein GCR2) [25–28]. All of them share seven conserved GXXG motifs [28], with 7-Helix-1 being the only exception, since the protein only shares six of the motifs, while the seventh is changed to GXXA instead (S1 Fig). Furthermore, the repeats include the highly conserved HG motif in repeat four, WCXG in repeat five and CHG in repeat six that are critical for the enzymatic activity of the LanCL proteins [29,30]. The highest homology, i.e. 38%, was found between 7-Helix-1 and hLanCL2, an intracellular abscisic acid receptor of human liver and muscle cells [2,3,5,31,32].
Phylogenetic analysis of 7-Helix-1 and related proteins from other organisms revealed that 7-Helix-1 forms a distinct cluster with the homologous plasmodial proteins (S2A Fig). The 3D structure of 7-Helix-1 was modelled using the I-TASSER server based on the crystal structure of hLanCL1 [27] and showed that 7-Helix-1 comprises a total of 14 α-helices that form two layers of α-helical barrels with seven helices each (S2B Fig). The combined data indicate that 7-Helix-1 is not a typical membrane-spanning GPCR-like receptor, but is rather related to the cytosolic hLanCL proteins.
Diagnostic RT-PCR was employed to investigate expression of 7-helix-1 in the P. falciparum blood stages and demonstrated similar transcript levels in immature (stage II-IV) and mature (stage V) gametocytes as well as gametocytes at 15’ post-activation (p.a.). In contrast, no expression was detected in the asexual blood stages (Fig 1B). Transcript analysis of the housekeeping gene pfaldolase was used as a loading control, and purity of the asexual blood stage and gametocyte samples was demonstrated by amplification of transcripts for the asexual blood stage-specific gene pfama1 (apical membrane antigen 1) [33] and for the gametocyte-specific gene pfccp2 (LCCL-domain protein 2) [34] (Fig 1B).
Mouse antisera were generated against two recombinant 7-Helix-1 peptides (7-Helix-1rp1 and 7-Helix-1rp2; see Fig 1A) to be used for protein expression analysis. Lysates of mixed asexual blood stages, immature gametocytes (stage II-IV), mature gametocytes (stage V) and gametocytes at 30’ p.a. were subjected to Western blot (WB) analysis. Immunoblotting with 7-Helix-1rp2 antisera demonstrated 7-Helix-1 migrating at a molecular weight of ~ 50 kDa in the lysates of immature and mature gametocytes, but detected no signal in asexual blood stage lysates (Fig 1C). Only a weak protein band was detected in the lysates of activated gametocytes. Immunoblotting with mouse antiserum directed against the endoplasmic reticulum-resident protein Pf39 was used as a loading control (Fig 1C).
Indirect immunofluorescence assays (IFAs), using either of the two antisera, revealed that 7-Helix-1 was abundantly expressed in immature and mature gametocytes, where it localized within cytosolic granules (Fig 2A; S3 Fig). Upon activation of the gametocytes, 7-Helix-1 expression decreased within 15 min. In the asexual blood stages, 7-Helix-1 was absent (S4A and S4B Fig). When sera from non-immunized mice (NMS) was used in the IFAs as a negative control, no labelling was detectable in either asexual blood stages or gametocytes (S5 Fig).
We also generated a transfectant line, expressing 7-Helix-1 tagged to hemagglutinin (HA) and streptavidin (Strep), (S6A Fig). Vector integration into the 3´-end of the 7-helix-1 locus was demonstrated by diagnostic PCR (S6B Fig). Subsequent WB, using an anti-HA antibody, demonstrated the expression of the 7-Helix-1 protein fused to HA and Strep (termed 7-Helix-1-HA), migrating at the expected molecular weight of ~60 kDa (Fig 1D). No protein band was detected, though, in lysates of wildtype (WT) gametocytes or in non-infected RBCs, used for negative control. IFAs confirmed that 7-Helix-1-HA is expressed in gametocytes and here accumulates in granular structures, while no labeling was detected in WT gametocytes (S6C Fig).
To verify the cytosolic localization of 7-Helix-1 we used gametocytes of the 7-Helix-1-HA line for subcellular fractioning. Subsequent WB using an anti-HA antibody detected 7-Helix-1-HA in the soluble protein fraction, but not in the integral and insoluble protein fractions (S7A Fig). Immunoblotting with antisera against the cytosolic falcilysin [35], the peripheral PfCCp2 and the parasitophorous vacuolar membrane-resident Pfs16 [36] were used as controls.
7-Helix-1 was found only in a subpopulation of gametocytes, when these were identified via immunolabelling for Pfs230, a plasmalemma-associated protein of the cysteine-rich superfamily, which is present in both male and female gametocytes (reviewed in [37,38]). Counterstaining of gametocytes by immunolabelling of Pfs25, a female-specific EGF domain protein (reviewed in [38]), revealed that 7-Helix-1 is expressed in 94.9 ± 2.43% of Pfs25-positive gametocytes, while only 78.9 ± 1.97% of Pfs230-positive cells were positive for 7-Helix-1 labelling (n = 100, in ten replicates). Pfs25-negative gametocytes (100%) were also negative for 7-Helix-1 labelling (Fig 2B). Taking into account, that the ratio of male:female gametocytes is female-biased with one male for about five female gametocytes (reviewed in [11]), these findings point to a female-specific expression of 7-Helix-1 in P. falciparum and are in accord with a reported female-specific expression by transcriptomics studies [39]. Noteworthy, 7-Helix-1 did not co-localize with the female-specific protein Pfg377, a marker for osmiophilic bodies [40], indicating that 7-Helix-1 is not present in these vesicles (S7B Fig). In conclusion, 7-Helix-1 is specifically expressed in female gametocytes and here localizes in cytoplasmic granules.
For functional analysis, the WT 7-helix-1 locus was disrupted via single cross-over homologous recombination (S8A Fig). Two clonal lines (1D12 and 2E6) were isolated and successful integration of the plasmid into the genome was verified by diagnostic PCR (S8B Fig). Vector integration was further validated by sequencing of the integration locus for clone 2E6 (S9 Fig). The absence of the protein in 7-Helix-1-KO gametocytes of both lines compared to WT gametocytes was subsequently demonstrated by IFA (S8C Fig). Similarly, WB could detect no 50-kDa protein in gametocyte lysates of the two 7-Helix-1-KO lines compared to WT (S8D Fig).
Subsequent phenotype analyses showed that the morphology and intraerythrocytic development of the asexual blood stages of the two 7-Helix-1-KO lines was normal compared to WT (S10A and S10B Fig). Furthermore, morphology and development of gametocytes during maturation from stage II to V was comparable to WT (S11A and S11B Fig). In addition, the percentage of female gametocytes did not differ in the 7-Helix-1-KO compared to the WT (89.3 ± 1.53% and 86.0 ± 3.46%, respectively; two independent experiments performed in triplicate), as shown by immunolabelling with anti-7-Helix-1-rp1 antisera.
When mature gametocytes of the 7-Helix-1-KO line 2E6 were fed to Anopheles stephensi mosquitoes via standard membrane feeding assays (SMFAs), the number of salivary gland sporozoites at day 17 post-infection (p.i.) was reduced to 8.0% and 9.0% compared to the numbers of the WT control in two independent experiments (S1 Table). Further, at day 10 p.i. the numbers of oocysts attached to the mosquito midgut were significantly reduced compared to the WT (Fig 3A). The numbers of retorts and ookinetes were quantified in midgut smears at 24 h p.i. via immunolabelling with rabbit antisera against Pfs28. Quantification revealed a significant reduction of retorts (22.7 ± 3.98%) and ookinetes (17.7 ± 10.88%) when the mosquitoes were fed with the 7-Helix-1-KO line 2E6 compared to WT control (set to 100%) (Fig 3B). Furthermore, we detected a significant decrease in the numbers of zygotes (67.8 ± 3.70%), and macrogametes (69.9 ± 11.21%) formed by the 7-Helix-1-KO line 2E6 compared to WT, as investigated via immunolabelling with anti-Pfs28 and anti-Pfs25 antibody at 4 h and 30 min p.a. in vitro, respectively (Fig 3C and 3D). No significant differences were observed between 7-Helix-1-KO 2E6 and WT gametocytes in the ability to form motile male microgametes (termed exflagellation) upon activation (89.7 ± 25.54%) (Fig 3E).
On the ultrastructural level, non-activated 7-Helix-1-KO 2E6 gametocytes exhibited the typical morphology described for WT gametocytes, having a prominent enveloping erythrocyte membrane (EM) and parasitophorous vacuole membrane (PVM) as well as the inner membrane complex (IMC) (Fig 4A). At 30 min p.a., the WT macrogametes have egressed from the RBC, and at this time point, the PVM has fully ruptured, while the IMC has partially disintegrated. These morphological changes upon gametocyte activation are in accord with previous reports [41]. In 7-Helix-1-KO gametocytes at 30 min p.a., however, approximately 39% of the parasites had not rounded up (30 ultrasections investigated). The PVM and EM were still present in 39% and 69% of the activated 7-Helix-1-KO gametocytes, respectively, while in 46% of the parasites, the IMC was still intact (Fig 4A).
A complementation line was generated by transfecting the 7-Helix-1-KO line 1D12 with an episomal copy of the gene (fused to a GFP-encoding sequence) under the fnpa promotor (a LCCL-domain protein). The presence of the vector in the 7-Helix-1-KO(+) line was confirmed via diagnostic PCR (S12A Fig). Amplification of pfama1 was used as a positive control in the PCRs. Furthermore, WB, using the mouse anti-7-Helix-1rp2 antisera, detected a band at ~ 75 kDa in the 7-Helix-1-KO(+) lysate, corresponding to the 7-Helix-1-GFP fusion protein (S12B Fig). A similar band was detected, when a rabbit anti-GFP antibody was used for immunoblotting. Controls included the detection of 7-Helix-1 in the WT lysate migrating at ~ 50 kDa, while no corresponding band was detected in the 7-Helix-1-KO 1D12 lysate. Immunoblotting with mouse anti-Pf39 antiserum was used as a loading control (S12B Fig). Protein expression and granular localization of the 7-Helix-1-GFP fusion protein in 7-Helix-1-KO(+) gametocytes was further verified via IFA, using either of the two generated anti-7-Helix-1 antisera (S12C and S12D Fig).
The 7-Helix-1-KO(+) line was used to conduct phenotype rescue experiments. While in the 7-Helix-1-KO line 1D12, the numbers of macrogametes (52.2 ± 14.17%) and zygotes (50.2 ± 11.32%) were significantly reduced compared to WT (set to 100%), the 7-Helix-1-KO(+) line exhibited higher numbers of macrogametes (76.1 ± 4.75%) and zygotes (89.2 ± 15.30%). These numbers did not differ significantly from the numbers of macrogametes and zygotes in activated WT parasites (Fig 4B and 4C). Our combined data demonstrate that the loss of 7-Helix-1 results in severely impaired female gametogenesis.
To investigate potential transcriptional changes caused by the lack of 7-Helix-1, we employed microarray analyses. Gametocytes at 30 min p.a. of the WT and the 7-Helix-1-KO line 2E6 were purified, the cDNA was synthesized and analyzed using a P. falciparum DNA Agilent microarray chip containing DNA spots corresponding to the 5,363 coding genes of nucleus, mitochondrion and apicoplast [42,43] (S1 File).
In the 7-Helix-1-KO line 2E6, 198 genes were identified with transcript levels more than 1.5-fold higher compared to WT; 59 of these genes exhibited factors of 2 or higher (S2 File). Gene ontology (GO) enrichment analyses ranked the majority of transcriptionally up-regulated genes to the biological processes of translation and biosynthesis with a high number of gene products being components of ribosomes (S3 File). Further, 122 genes of the 7-Helix-1-KO exhibited more than 1.5-fold lower transcript levels compared to WT. Of these, 34 exhibited factors of 0.5 or lower and included the gene PF3D7_0525400 encoding 7-Helix-1 (down-regulation factor: 0.19). Noteworthy, 60 of the transcriptionally down-regulated genes encrypt non-coding RNAs (S2 File).
For in-depth analyses, we focused on transcriptionally deregulated genes highly expressed in gametocytes and grouped these according to their predicted functions. We identified 86 genes that were transcriptionally up-regulated in activated 7-Helix-1-KO gametocytes (S2 File). The majority of the genes could be assigned to functions in transcription and translation (Fig 5A). Genes assigned to transcription included the RNA-binding proteins ALBA1 and ALBA4 [44,45] as well as PF3D7_0823200, all of which are components of the mRNA-bound proteome of P. falciparum [46]. Out of the 21 genes assigned to translation, 19 coded for ribosomal proteins as part of the 60S or 40S subunits (S2 File). Further, two genes encode for putative elongation initiation factors, i.e. eIF5A (PF3D7_0913200) and SUI1 (PF3D7_1243600). In addition, 53 genes were transcriptionally down-regulated in the activated 7-Helix-1-KO gametocytes, which were mainly associated with translation. Out of 24 genes linked to translation, 10 genes encode tRNAs and 12 genes code for small nucleolar RNAs needed for the assembly of the 60S and 40S ribosomal subunits (S2 File). In addition, a high proportion of genes with deregulated transcription levels were of unknown function (Fig 5A).
In order to validate the microarray data, we randomly chose seven of the identified genes (five of them transcriptionally up-regulated and two down-regulated in microarray analysis) and analyzed differences in transcript expression via real time RT-PCR. For this, gametocytes of the 7-Helix-1-KO line 2E6 and WT gametocytes were purified at 30 min p.a. and total RNA was isolated followed by cDNA synthesis. Real time RT-PCR was performed and the threshold cycle number (Ct) calculated, which was normalized to the Ct of the endogenous control gene encoding P. falciparum seryl tRNA-ligase (PF3D7_0717700) as reference. Fold changes in the 7-Helix-1-KO in relation to the WT were determined. The five genes that were more than 2-fold up-regulated in activated 7-Helix-1-KO gametocytes in microarray analysis also showed transcript up-regulation by 2-fold or more in real time RT-PCR (S13A Fig). Further, in the absence of 7-Helix-1, the two genes that exhibited a more than 2-fold down-regulation in microarray studies also had decreased transcript levels. In addition, potential changes in the transcript expression of SR1, SR10, SR12 and SR25 were investigated in mature and activated gametocytes, however, no significant changes were detected (S13B Fig). In conclusion, the transcriptomics data point to an imbalance of components required for protein synthesis in gametocytes lacking 7-Helix-1.
To further investigate the potential link between 7-Helix-1 and translation during female gametogenesis, we treated mature WT gametocytes with translation inhibitors and investigated their ability to form macrogametes and zygotes. The translation inhibitors used in the experiments included emetine, which blocks translation by binding the 40S ribosomal subunit [47], and cycloheximide, which inhibits ribosome translocation [48]. Antimalarial concentrations were determined by Malstat assay and revealed IC50 values of 35 nM ± 1.4 nM and 435 nM ± 27.4 nM for emetine and cycloheximide, respectively. Chloroquine was used as a positive control in the assays and resulted in growth inhibition with a mean IC50 value of 22 nM ± 2.8 nM.
Mature gametocytes of the WT and the 7-Helix-1-KO line 2E6 were purified, treated with the solvent alone or with the inhibitors at IC50 concentration for 30 min, followed by activation. The numbers of macrogametes and zygotes were determined via IFA at 30 min and 4 h p.a., respectively, by immunolabelling with rabbit anti-Pfs25 antisera. Emetine at IC50 concentrations resulted in significantly decreased numbers of macrogametes (41.8 ± 4.98%) compared to the solvent control (set to 100%) (Fig 5B). These numbers were comparable to the ones of the 7-Helix-1-KO line 2E6, when treated with the solvent control (41.8 ± 3.80%). Similar results were obtained, when the numbers of zygotes were determined at 4 h p.a. following emetine-treatment. Emetine-treated WT parasites formed significantly less zygotes than the solvent-treated control (65.1 ± 11.32%; WT solvent-control set to 100%). Solvent-treated 7-Helix-1-KO 2E6 parasites exhibited zygote numbers (62.4 ± 5.16%) comparable to emetine-treated WT gametocytes (Fig 5B).
Likewise, WT parasites treated with cycloheximide formed significantly reduced numbers of macrogametes (37.9 ± 4.40%) compared to the solvent control (set to 100%), which were comparable to the numbers of macrogametes in solvent control-treated 7-Helix-1-KO 2E6 parasites (43.2 ± 12.56%) (Fig 5B). Cycloheximide-treatment of WT parasites led to reduced zygote numbers (54.7 ± 9.20%) compared to the solvent control (set to 100%). The numbers of zygotes in the solvent-treated 7-Helix-1-KO 2E6 sample were comparable to the numbers of zygotes in cycloheximide-treated WT parasites (38.7 ± 3.95%) (Fig 5B). In conclusion, the chemical-KO phenotype caused by translational inhibitors resembles the one observed for 7-Helix-1-KO parasites.
The granular localization of 7-Helix-1 in female gametocytes and its crucial role during female gametogenesis let us to investigate a potential link between 7-Helix-1 and SGs. Firstly, an mRNA-detecting fluorescence in situ hybridization experiment coupled with an IFA (mRNA-FISH-IFA) was conducted. A biotinylated poly-dT oligonucleotide was used to bind the poly-A-tails of mRNA aggregates. Poly-dT-oligonucleotide binding was demonstrated with fluorescently labeled Strep, 7-Helix-1 was immunolabeled with anti-7-Helix-1rp2 antisera. The IFA demonstrated the presence of mRNA aggregates in most of the 7-Helix-1-positive granules, indicating that 7-Helix-1 accumulates in RNA dense granules (Fig 6A). A Pearson’s correlation coefficient (-1 for perfect negative correlation, 0 for no correlation, +1 for perfect positive correlation) of 0.79 ± 0.079 has been calculated using ImageJ (n = 15). For control, RNA-FISH-IFAs were performed omitting the poly-dT-oligonucleotide, and no mRNA labeling was observed (S14A Fig). Similar mRNA-FISH-IFA experiments were performed, using the 7-Helix-1-HA line, again demonstrating a co-localization of mRNA aggregates and 7-Helix-1, which was detected using anti-HA antibody (Pearson’s correlation coefficient = 0.67 ± 0.127; n = 15; S14B Fig). For control, a P. falciparum line RNF1-HA was used, which expresses the HA-tagged ring finger protein RNF1 in gametocytes and which was generated using a similar strategy. No co-labeling of mRNA aggregates and RNF1 was detected (S14C Fig). When anti-Pfs230 antisera was used as a control in either of the mRNA-FISH-IFA experiments, no co-labeling of the protein with mRNA aggregates was detected (S14D and S14E Fig).
The association of 7-Helix-1 with SGs was subsequently confirmed by SG core fraction enrichment. Mature gametocytes of line 7-Helix-1-HA were treated with sodium arsenite to induce stress in these cells and the SGs were isolated by enrichment. Lysates of total mature gametocytes and of the enriched SG fraction were immunoblotted with anti-HA antibody. A prominent 7-Helix-1 band was observed in the SG fraction (Fig 6B). For positive control, rabbit antisera against plasmodial HSP70-1 was used [49], as heat shock proteins are known components of SGs (reviewed in [17]). Accordingly, prominent HSP70-1-positive protein bands, migrating at the expected molecular weight of 70 kDa, were detected in the gametocyte lysate and the SG fraction. Immunoblotting with anti-Pf39 antisera, used as a negative control, labeled Pf39 in the gametocyte lysate, while only minor labeling was detected in the enriched SG fraction. As an additional negative control, the samples were blotted with mouse antisera directed against the full-length and processed form (126 and 68 kDa) of the M1-family alanyl aminopeptidase PfM1-AP [50] and both protein bands could be detected in the gametocyte lysate, but not in the SG fraction (Fig 6B). The combined data show that 7-Helix-1 is enriched in SGs of gametocytes.
Lastly, the interaction of 7-Helix-1 with the known translational repressors CITH (CAR-I/trailer hitch homolog), PABP1 (polyadenylate-binding protein 1) and DOZI (development of zygote inhibited) as components of the female gametocyte-specific SGs [15,45] was investigated via co-immunoprecipitation assay and IFA. When the 7-Helix-1rp2 antisera was used for precipitation, subsequent immunoblotting demonstrated CITH and PABP1 (39 and 97 kDa, respectively) in the precipitate, using the respective antibodies (Fig 6C). When the anti-DOZI antisera was used for precipitation in 7-Helix-1-HA gametocytes, 7-Helix-1-HA (~60 kDa) was detected in the precipitate using the anti-HA antibody for immunoblotting. No Pf39, used for negative control, could be precipitated in either immunoprecipitation. IFA analysis confirmed a co-localization of 7-Helix-1 with CITH, PABP1 and DOZI in female gametocytes (Fig 6D).
In a last set of experiments, we aimed at demonstrating a potential involvement of 7-Helix-1 in translational repression by studying Pfs25, a protein well known to be translationally repressed in the SGs of female gametocytes [16]. Once translational repression is released at the onset of female gametogenesis, Pfs25 is transported to the macrogamete surface, where it plays a role during fertilization (reviewed in [12,37,38]).
In P. falciparum, the RNA-binding protein Pumilio-2 (Puf2) is involved in translational repression of Pfs25 by binding to the 5´-untranslated region of the pfs25 gene [16,51]. We thus generated mouse antisera directed against Puf2 via immunization of the mice with a bacterially expressed peptide (S15A Fig). IFAs demonstrated the presence of Puf2 in the cytosol of mature gametocytes, but not in asexual blood stage parasites (S15B Fig), confirming previously published Puf2 expression data [16]. Co-labeling with antibodies directed against the female-specific antigen Pfs25 shows that Puf2 is expressed in gametocytes of both sexes (S15C Fig). Puf2 was also detected, when gametocytes of the 7-Helix-1-KO line 2E6 were immunolabeled with anti-Puf2 antisera (S15D Fig), demonstrating that lack of 7-Helix-1 does not result in loss of Puf2.
The Puf2 antibody was then employed in co-immunoprecipitation assays, using the 7-Helix-1-HA line. Subsequent WB demonstrated the presence of HA-tagged 7-Helix-1 in the precipitates, using the anti-HA antibody, while no Pf39 signal was detected, when the anti-Pf39 antisera was used for control (Fig 7A). Additional controls included Puf2-antibody-based precipitation of WT gametocyte lysate, which did not result in any co-precipitation of 7-Helix-1-HA or Pf39 as well as the successful precipitation of Pf39 in WT gametocyte lysates using the corresponding antibody (Fig 7A).
We then analyzed, if pfs25 transcript is also associated with 7-Helix-1. RNA-co-immunoprecipitation (RIP) assays followed by RT-PCR were employed, using the 7-Helix-1-HA line. When antibodies directed against the HA-tag or against CITH for positive control were used in the RIP assays and cDNA was prepared from co-precipitated transcript, subsequent RT-PCRs amplified pfs25 transcript (Fig 7B). Furthermore, transcript for pfs28, encoding the Pfs25 paralog Pfs28 also known to be repressed in translation during gametocyte maturation [51], was amplified. However, no pfccp2 transcript encoding the gametocyte-specific protein PfCCp2 was detected. Preparations lacking reverse transcriptase were used for negative control in the assays, and no PCR products were detected (Fig 7B).
We then explored, if the interaction between 7-Helix-1 and Puf2 affects Pfs25 synthesis. IFAs were employed and demonstrated a reduced immunolabelling for Pfs25 in 7-Helix-1-KO gametocytes at 30 min p.a. compared to WT (Fig 7C). No reduction in immunolabeling, however, was observed, when antisera against the gametocyte-specific protein Pfs230 was used. The fluorescence intensities for Pfs25 and Pfs230 were measured by Image J1.51f in three independent experiments in activated gametocytes at 30 min p.a. of WT and the 7-Helix-1-KO line 2E6. The quantifications revealed significantly lower Pfs25 levels in gametocytes lacking 7-Helix-1 compared to WT, while no difference in the Pfs230 levels in the same gametocytes was detected (Fig 7D). We thus conclude that the interaction of 7-Helix-1 and Puf2 affects Pfs25 synthesis.
The human-to-mosquito transmission of P. falciparum requires a high degree of tight coordination allowing the parasite to rapidly adapt to the changing environment. The parasite prepares for transmission by the formation of intraerythrocytic gametocytes, which immediately convert into male and female gametes, once taken up by the blood-feeding mosquito. In order to pre-adapt to the change of host, female gametocytes store several hundred mRNAs in specific SGs via translational repressors like DOZI, CITH or Puf2. These repressed mRNAs code for proteins required for further development in the mosquito, particularly the ookinete and oocyst stages ([14–16]; reviewed in [12]). With the onset of gametogenesis, the repressed mRNAs become accessible again for ribosomes, hence female gametogenesis requires a rapid increase in translational activities to ensure parasite survival. Despite the crucial function of the gametocyte-specific SGs for parasite transmission and thus spread of the tropical disease, to date not much is known about their compositions and the molecules involved in the translational fine control.
Here we identified the heptahelical protein 7-Helix-1 of P. falciparum as a novel SG component of female gametocytes. As an integral part of these SGs, 7-Helix-1 interacts with translational repressors, like CITH, PABP1, DOZI and Puf2, and is thereby involved in the regulation of translational processes during sexual reproduction. Consequently, gametocytes deficient of 7-Helix-1 are severely impaired in female gametogenesis and thus transmission to the mosquito. Further, 7-Helix-1-KO gametocytes exhibit a transcriptional deregulation of proteins required for translational control. This transcriptional deregulation is particularly evident for ribosomal proteins, but can further be observed for initiation and elongation factors, heat shock proteins, RNA-binding proteins, as well as RNA helicases, all of which represent typical components of SGs in other organisms [52]. Noteworthy in this context, the orthologues of 18 of the transcripts up-regulated in the 7-Helix-1-KO are components of previously identified gametocyte repressome of P. berghei [53] providing a further link between 7-Helix-1 and gametocyte SGs. In addition, non-coding RNAs, like snoRNAs and tRNAs required for ribosomal assembly and functionality, are down-regulated in their transcription levels. Such a deregulation has previously been observed for tRNAs in other organisms and can either be explained by cellular stress [54] or by feedback regulation due to the high numbers of uncharged tRNAs in the cytosol (reviewed in [55–57]).
In silico analyses demonstrated that 7-Helix-1 is homologous to hLanCL2, a heptahelical regulator of glycemic control. Previous studies reported that hLanCL2 does not represent a transmembrane GPCR, but is located in the cytoplasm of a variety of human cells, including liver, immune and muscle cells. Here, hLanCL2 acts as a receptor for the stress hormone abscisic acid (ABA), which in humans regulates both inflammation and glycemia. Binding of ABA by hLanCL2 leads to a G-protein-mediated increase in cytosolic cAMP and calcium [5,32,58–60]. Further hLanCL2 acts as a positive regulator within the Akt/mTORC2 pathway by enabling the insulin-dependent phosphorylation of Akt via an interaction with the mTORC2-complex, thereby contributing to cell survival [61]. Due to the role of ABA in inflammatory diseases and diabetes, hLanCL2 is currently considered a promising therapeutic target [62–64].
Future studies are required to shed light on the molecular function of 7-Helix-1 during protein synthesis control in female gametocytes and to unveil potential activator and effector proteins of this heptahelical stress regulator. Initial clues might be gained from its homology with hLanCL2. While apicomplexan parasites have lost most of the components of the TORC pathway, including the TOR kinase, through genomic reduction [65,66], a recent study in P. falciparum described the RNA polymerase III repressor Maf1 as a plasmodial mTORC1 component. Maf1 regulates transcription of this polymerase under stress conditions and parasites deficient in Maf1 are defective in downregulation of tRNA synthesis [67]. Further, the plasmodial Akt kinase PKB is linked to a Ca2+/calmodulin-dependent signaling cascade relevant for intraerythrocytic survival of the malaria parasite [68–71], providing the presence of further conserved interaction partners of hLanCL2 in Plasmodium.
In conclusion, our combined data demonstrate that 7-Helix-1 constitutes SGs with key functions in translational control during female gametogenesis. Due to its crucial role for human-to-mosquito transmission 7-Helix-1 might represent a novel gametocytocidal drug-target.
Experiments for the generation of antisera in mice were performed according to the German Animal Welfare Act (Tierschutzgesetz) and were approved by the animal welfare committees of the government of the District Council of Cologne, Germany (ref. no. 84–02.05.30.12.097 TVA). The University Hospital Aachen Ethics commission approved all work with human blood, the donors remained anonymous and serum samples were pooled. Human erythrocyte concentrate and serum used in this study were purchased from the Department of Transfusion Medicine (University Hospital Aachen, Germany).
The following PlasmoDB gene identifiers are assigned to the P. falciparum genes and proteins analyzed in this study: AMA1 [PlasmoDB: PF3D7_1133400]; PfCCp2 [PlasmoDB: PF3D7_1455800]; Pfs230 [PlasmoDB: PF3D7_0209000]; EXP1 [PlasmoDB: PF3D7_1121600]; MSP1 [PlasmoDB: PF3D7_0930300]; Pf39 [PlasmoDB: PF3D7_1108600]; 7-Helix-1 [PlasmoDB: PF3D7_0525400]; PfAldolase [PlasmoDB: PF3D7_1444800]; Pfs16 [PlasmoDB: PF3D7_0406200]; Pfs25 [PlasmoDB: PF3D7_1031000]; Pfs28 [PlasmoDB: PF3D7_1030900]; FNPA [PlasmoDB: PF3D7_1451600]; Pfg377 [PlasmoDB: PF3D7_1250100]; RNF1 [PlasmoDB: PF3D7_0314700]; HSP70-1 [PlasmoDB: PF3D7_0831700]; M1-AP [PlasmoDB: PF3D7_1311800]; Falcilysin [PlasmoDB: PF3D7_1360800]; CITH [PlasmoDB: PF3D7_1474900]; DOZI [PlasmoDB: PF3D7_0320800]; PABP1 [PlasmoDB: PF3D7_1224300]; Puf2 [PlasmoDB: PF3D7_0417100]; SR1 [PlasmoDB: PF3D7_1131100]; SR10 [PlasmoDB: PF3D7_1215900]; SR12 [PlasmoDB: PF3D7_0422800]; SR25 [PlasmoDB: PF3D7_0713400].
Prediction of the transmembrane domains and intra-/extracellular regions of 7-Helix-1 was performed using the TMHMM Server v. 2.0 (http://www.cbs.dtu.dk/services/TMHMM/); further domain prediction was performed using the Simple Modular Architecture Research Tool (SMART, http://smart.embl.de/; [72]). The 3D modelling of 7-Helix-1 was conducted using the I-TASSER server (http://zhanglab.ccmb.med.umich.edu/I-TASSER/; [73]). For sequence alignment of homologous proteins and the generation of the phylogenetic tree the software MainWorkbench 7.5 was used. Predictions of gene expression and function were made using the database PlasmoDB (http://plasmoDB.org;[74]).
In this study, the following antisera were used: rabbit polyclonal antisera against MSP-1 (ATCC), EXP1 [75], Pfs230 (Biogenes), Pfs25 (ATCC), Pfs28 (ATCC), PfHSP70-1 [49], PyPABP1 [76], PfCCp2 [77], Pfg377 ([78], kindly provided by Pietro Alano, Istituto Superiore di Sanità, Rome, Italy), and HsDDX6/PfDOZI (kindly provided by Joseph Reese, Penn State University); mouse polyclonal antisera against PfM1-AP [79], Pfs230 [80], PfFalcilysin [79], Pfs16 [81]; kindly provided by Pietro Alano, Istituto Superiore di Sanità, Rome, Italy) and Pf39 [77]; rabbit anti-GFP antibody (Santa Cruz Biotechnology) and rabbit anti-HA antibody (Sigma-Aldrich). The generation of antisera against 7-Helix-1 (anti-7-Helix-1rp1 and 2), CITH and Puf2 is described below.
P. falciparum strain NF54 (WT NF54) was used in this study. Asexual blood stage parasites and gametocytes of WT NF54, the two 7-Helix-1-KO lines 2E6 and 1D12, the complementation line 7-Helix-1-KO(+) and the 7-Helix-1-HA line (for the generation of the 7-Helix-1 mutant lines, see below) were cultivated in vitro in human blood group A+ erythrocytes as previously described [82,83]. All parasite stages were maintained in RPMI1640/HEPES medium (Gibco) supplemented with 10% v/v heat inactivated human A+ serum, 50 μg/ml hypoxanthine (Sigma-Aldrich) and 10 μg/ml gentamicin (Gibco). For cultivation of 7-Helix-1-KO parasites, the selection drug blasticidin (InvivoGen) was added in a final concentration of 5.4 μM; for cultivation of 7-Helix-1-KO(+) and 7-Helix-1-HA parasites, the selection drug WR99210 (Jacobus Pharmaceutical Company) was added in a final concentration of 2.5 nM. All cultures were kept at 37°C in an atmosphere of 5% O2 and 5% CO2 in N2. Gametocytes were enriched via Percoll gradient centrifugation (GE Healthcare Life Sciences) as described previously [84]. Gametogenesis was induced by adding xanthurenic acid in a final concentration of 100 μM dissolved in 1% v/v 0.5 M NH4OH/ddH2O and incubation for 15 min at room temperature (RT).
7-Helix-1-KO: 7-Helix-1-KO parasites were generated via single cross-over homologous recombination using the pCAM-BSD vector [85–88]. A 534-bp gene fragment homologous to a region of the N-terminus of the gene was amplified via PCR using the respective 7-Helix-1-KO primers (for primer sequences, see S2 Table). Ligation of insert and vector backbone was mediated by BamHI and NotI restriction sites (underlined). A NF54 WT culture synchronized for 5% ring stages was loaded with 60 μg of the vector in transfection buffer via electroporation (310 V, 950 μF, 12 ms; Bio-Rad gene-pulser) as described [87]. Starting at 6 h post-transfection, blasticidin (Invivo-Gen) was added to a final concentration of 5.4 μM. A mock control was electroporated using transfection buffer without the disruption vector and cultured in medium without selection drug. After approximately 10 weeks, blasticidin-resistant parasites appeared in culture. To check for successful plasmid integration into the 7-helix-1 gene locus, genomic DNA (gDNA) of transfected parasites was isolated using the NucleoSpin Blood Kit (Macherey-Nagel) following the manufacturer’s protocol and used as template in diagnostic PCR. The following primers were used to confirm vector integration: 7-Helix-1-KO-5’-integration forward primer (1), 7-Helix-1-KO-3’-integration reverse primer (2), pCAM-BSD forward primer (3) and pCAM-BSD reverse primer (4) (for primer sequences, see S2 Table). After confirmation of plasmid integration, a mixed culture with >3% ring stages was diluted and transferred to a 96-well-plate. After three weeks of cultivation, clones were identified via Malstat assay (see below) and subsequent diagnostic PCR was performed to confirm vector integration and the absence of the WT 7-helix-1 gene locus. Two clonal lines, 7-Helix-1-KO 1D12 and 2E6, were isolated.
Anopheles stephensi mosquitoes were maintained as described previously [87]. Briefly, rearing of the mosquitoes was conducted under standard insectary conditions at 26 ± 0.5°C, 80 ± 2% humidity and a 12/12 h light/dark cycle. Egg production was induced by feeding adult female mosquitoes with non-infected human erythrocyte cultures and eggs were collected on filter paper in a beaker containing 0.1% w/v sea salt solution four days after the blood meal. Emerged larvae were reared with a density of 300 larvae/tray (3 l) in 0.1% w/v sea salt solution and fed on fodder pellets. After transformation, pupae were collected and placed in cages for mosquito emergence. Feeding of adult mosquitoes was performed using a cotton pad soaked with 5% w/v sterile saccharose solution supplemented with para-aminobenzoic acid and 40 μg/ml gentamicin [92]. For SMFAs, mature gametocytes of WT NF54 and the 7-Helix-1-KO line 2E6, which were positively tested for exflagellation activity, were enriched and the parasitemia was adjusted to 0.25% by the addition of human erythrocytes in A+ serum. The mosquitoes were allowed to feed on the cell suspension via glass feeders for 20 min [93]. For ookinete quantification, samples of the midgut were prepared at 24 h p.i. and subjected to IFA. Ookinetes were immunolabeled with rabbit anti-Pfs28 antisera and counted in 30 optical fields for four times. To determine the numbers of oocysts, midguts were dissected at day 10 p.i. and stained with 0.2% v/v mercurochrome/PBS. For sporozoite quantification, salivary glands were prepared at day 17 p.f. and parasites were counted using a Neubauer chamber.
Two recombinant 7-Helix-1 proteins, spanning amino acids 234–320 (7-Helix-1rp1) and 1–467 (7-Helix-1rp2) were expressed as fusion proteins with an N-terminal maltose binding protein (MBP)-tag using the pIH902 vector [80] for 7-Helix-1rp1 or with an N-terminal glutathione-S-transferase (GST)-tag using the pGEX-4T-1 vector (Amersham Bioscience) for 7-Helix-1rp2. Gene fragments were amplified from P. falciparum gDNA using the respective primers, resulting in gene fragments of 279 bp and 1426 bp, respectively. Ligation into the expression vectors was mediated by BamHI/PstI restriction sites (underlined) for 7-Helix-1rp1 into pIH902 and by BamHI/XhoI restriction sites (underlined) for 7-Helix-1rp2 into pGEX-4T-1. Recombinant Puf2rp1 protein (amino acids 243–424) was expressed as fusion protein with an N-terminal MBP-tag using the pMALTMc5X vector (New England Biolabs). Amplification of the gene fragment from P. falciparum gDNA resulted in a 579 bp gene fragment. Ligation into the pMALTMc5X expression vector was mediated by XmnI/PstI restriction sites (underlined). Expression of the recombinant fusion proteins was performed in Escherichia coli BL21 (DE3) RIL cells (Stratagene) according to the manufacturer’s protocol. The fusion proteins were purified via affinity chromatography from bacterial extracts using amylose resin (New England Biolabs) for 7-Helix-1rp1 and Puf2, and glutathione-sepharose (GE Healthcare) for 7-Helix-1rp2 according to the manufacturer’s protocols. Successful purification of the proteins was validated by SDS-PAGE. Recombinant P. yoelii CITH protein (amino acid 1–100) was expressed and rabbit antibody was generated using a method described in [76]. Primer sequences used for the expression of recombinant proteins are listed in S2 Table.
Recombinant fusion proteins 7-Helix-1-rp1-MBP, 7-Helix-1-rp2-GST and Puf2-rp1-MBP were purified via affinity chromatography as described above followed by PBS buffer exchange via filter centrifugation using Amicon Ultra 15 (Sigma-Aldrich) according to the manufacturer’s protocol. Six weeks-old female NMRI mice (Charles River Laboratories) were immunized with subcutaneous injections of 100 μg recombinant protein emulsified in Freund’s incomplete adjuvant (Sigma-Aldrich) followed by a boost after 4 weeks with 50 μg recombinant protein. At day 10 after the boost, mice were anesthetized by intraperitoneal injection of a mixture of ketamine and xylazine according to the manufacturer’s protocol (Sigma-Aldrich) followed by the collection of polyclonal immune sera via heart puncture. The immune sera of three mice immunized with the same antigen were pooled; NMS were used as a negative control.
Co-immunoprecipitation assays were performed to precipitate either protein or RNA. Gametocyte lysates of WT NF54 and the 7-Helix-1-HA line were obtained as described above. For the precipitation of proteins, the lysates were initially incubated with 5% v/v pre-immune mouse or rabbit sera and 20 μl of protein G-beads (Roche) for 1 h at 4°C. After centrifugation, the supernatant was incubated for 1 h at 4°C with 5% v/v polyclonal mouse antisera against 7-Helix-1, Puf2 as well as Pf39 [94] used as a negative control, or polyclonal rabbit antisera against CITH, DOZI or HA. A volume of 20 μl protein G-beads was added and incubated overnight at 4°C. The beads were centrifuged, washed with PBS five times and resuspended in an equal volume of loading buffer. The sample was then subjected to WB (see below). For the precipitation of RNA (RIP; [53]), the beads were resuspended in Trizol instead of loading buffer, and RNA isolation and diagnostic RT-PCR were performed as described below.
Either asexual blood stages of synchronized WT NF54 cultures were harvested, immature and mature gametocytes were enriched via Percoll gradient purification, and activated gametocytes were collected at 15 min p.a. Total RNA was isolated using the Trizol reagent (Invitrogen) according to the manufacturer’s protocol or the co-precipitated RNAs from the RIP assays were used [53]. To remove gDNA contamination, samples were treated with RNase-free DNase I (Qiagen), followed by phenol/chloroform extraction and ethanol precipitation. Photometric analysis revealed A260/280 ratios greater than 2.1. The cDNA synthesis was performed with two micrograms of each RNA sample using the SuperScript III First-Strand Synthesis System (Invitrogen) following the manufacturer’s protocol. Transcript for 7-helix-1 (527 bp), pfs25 (459 bp) and pfs28 (494 bp) was amplified in 25–35 cycles using the respective primers (for primer sequences, see S2 Table). To confirm purity of the asexual blood stage and gametocyte samples, transcript amplification of pfama1 (189 bp) [33] and of pfccp2 (198 bp) [34] using the respective primers were performed (for primer sequences, see S2 Table). Amplification of pfaldolase (378 bp) was used as loading control and to investigate potential contamination with gDNA in the negative control lacking reverse transcriptase (S16A Fig). PCR products were separated by 1.2% w/v agarose gel electrophoresis.
Total RNA was isolated from Percoll-purified, mature gametocytes of WT NF54 and the 7-Helix-1-KO line 2E6 as described above. The gametocytes were either non-activated or harvested at 30 min p.a. One μg of each total RNA sample was used for cDNA synthesis using either the SuperScript III or SuperScript IV First-Strand Synthesis System (Invitrogen), following the manufacturer's instructions. The synthesized cDNA was first tested by diagnostic RT-PCR for asexual blood stage contamination using primers specific for the gene encoding the apical membrane antigen AMA-1 [33] and for gametocyte specificity using primers specific for the gene encoding the LCCL-domain protein PfCCp2 [34] (S16B Fig). Controls without reverse transcriptase were used to investigate potential gDNA contamination by using pfaldolase primers (for primer sequences, see S2 Table). Primers for quantitative real time RT-PCR were designed using the Primer 3 software (http://frodo.wi.mit.edu/primer3/) and tested on gDNA in conventional PCR to confirm primer specificity (for primer sequences, see S2 Table). Real time RT-PCR measurements were performed using either the Bio-Rad iQ5 Real-Time Detection System or the StepOnePlus Real-Time PCR System (Applied Biosystems). Reactions were performed in a total volume of 20 μl using the maxima SYBR green qPCR master mix according to manufacturer’s instructions (Thermo Scientific) in triplicate. Controls without template and without reverse transcriptase were included in all real time RT-PCR experiments. Transcript levels were calculated by the 2-ΔCt method [95], the Ct was normalized with the Ct of the gene encoding seryl tRNA-ligase (PF3D7_0717700) as reference [96,97] and fold changes were calculated via the normalized Ct ratio of 7-Helix-1-KO:WT.
Mixed asexual blood stage and gametocyte stage cultures as well as macrogametes and zygotes of WT NF54, the 7-Helix-1-KO lines 2E6 and 1D12 and the 7-Helix-1-KO(+) line collected at 2, 5, 10, 30 min (macrogametes) and 4 h (zygotes) p.a., and ookinetes obtained from mosquito midguts at 24 h p.i. were subjected to IFAs. After the cell monolayers were air-dried on glass slides, they were fixed with 4% w/v paraformaldehyde/PBS (pH 7.4) for 10 min at RT followed by membrane permeabilization with 0.1% v/v Triton X-100/125 mM glycine (Carl Roth)/PBS at RT for 30 min. Blocking of non-specific binding sites was performed using 3% w/v BSA/PBS for 1 h, followed by incubation with polyclonal mouse antisera against 7-Helix-1rp1 (dilution 1:50), 7-Helix-1rp2 (dilution 1:20), Puf2rp1 (dilution 1:20), Pfs230 (dilution 1:200), with polyclonal rabbit antisera against the HA-tag (dilution 1:50) or with NMS (dilution 1:20) for 2 h at 37°C. Binding of primary antibody was detected by incubating the samples with monoclonal Alexa Fluor 488-conjugated goat anti-mouse or anti-rabbit IgG antibodies (dilution 1:1,000; Molecular Probes) for 1 h at RT. The different parasite stages were detected by double-labelling with stage-specific polyclonal rabbit or mouse antisera, i.e. anti-MSP1 antisera (dilution 1:1,000), anti-EXP1 antisera (dilution 1:100), anti-Pfs230 antisera (dilution 1:200), anti-Pfs25 antisera (dilution 1:1,000), anti-Pfs28 antisera (dilution 1:200), anti-CITH antisera (dilution 1:1,000), anti-PABP1 antisera (dilution 1:1,000), anti-DOZI antisera (1:500) and anti-Pfg377 antisera (1:500), followed by incubation with monoclonal Alexa Fluor 594-conjugated goat anti-rabbit or anti-mouse IgG antibodies (dilution 1:1,000; Molecular Probes). Nuclei were highlighted by treatment with Hoechst33342 nuclear stain for 1 min at RT and cells were mounted with anti-fading solution AF2 (Citifluor Ltd) and sealed with nail varnish. For sex-specific quantifications, 100 gametocytes labelling for either Pfs230 or Pfs25 were counted in ten replicates and the percentage of 7-Helix-1-positive cells was calculated. Data analysis was performed using MS Excel 2013. IFAs were analyzed using a Leica DM 5500B fluorescence microscope and digital images were processed using Adobe Photoshop CS5 software.
For subcellular fractioning, ~ 3 x 107 gametocytes of the 7-Helix-1-HA line were purified as described above and liberated from the RBC in 0.03% saponin/PBS for 3 min at 37°C. Cells were pelleted and lysed by resuspension in 100 μl of 5 mM Tris-HCl (pH 8) supplemented with protease inhibitor cocktail (complete EDTA-free, Roche) and 10 min incubation at RT followed by freezing and thawing. Soluble proteins were separated by centrifugation and the pellet was resuspended in 100 μl of 1% Triton X-100 and incubated for 30 min at RT to extract integral proteins, and the final pellet representing the insoluble proteins was resuspended in 100 μl of 0.5 x PBS/4% SDS/ 0.5% Triton X-100. The samples were then subjected to WB (see below).
SG core fraction enrichment was performed following a previously published protocol [98]. Gametocytes of line 7-Helix-1-HA were purified via Percoll gradient as described above and stress was induced by sodium arsenite treatment (0.5 mM) for 1 h at 37°C. Cells were pelleted and washed with RPMI1640. The cell pellet was flash-frozen in liquid N2 and stored at -80°C if necessary. For the enrichment of the SG core fraction, the cell pellet was thawed on ice, resuspended in 100 μl SG lysis buffer (50 mM Tris HCl pH 7.4, 100 mM potassium acetate, 2 mM magnesium acetate, 0.5 mM DTT, 50 μg/ml heparin, 0.5% NP40, 1 complete mini EDTA free protease inhibitor tablet per 50 ml of lysis buffer (Roche), RNasin 1 U/μl (Promega)) and lysed by five passages through a 26G needle. Cell debris was pelleted by centrifugation at 1000g for 5 min at 4°C and the supernatant was transferred to a new microcentrifuge tube. After centrifugation at 18,000g for 20 min at 4°C, the pellet was resuspended in 100 μl SG lysis buffer followed by another centrifugation step at 18,000g for 20 min at 4°C. The pellet was resuspended in 100 μl SG lysis buffer and centrifuged at 850g for 2 min at 4°C. The supernatant represented the SG core enriched fraction and was subjected to WB (see below).
Asexual blood stage parasites of WT NF54, the 7-Helix-1-KO lines 2E6 and 1D12, the complementation line 7-Helix-1-KO(+) and the 7-Helix-1-HA line were harvested from mixed or synchronized cultures, while gametocyte stages were enriched by Percoll purification. Parasites were incubated with 0.05% w/v saponin/PBS for 10 min for erythrocyte lysis. Pelleted parasites were resuspended in lysis buffer (150 mM NaCl, 0.1% v/v Triton X-100, 0.5% w/v sodium deoxycholate, 0.1% w/v SDS, 50 mM Tris-HCl pH 8.0) supplemented with protease inhibitor cocktail (complete EDTA-free, Roche) and incubated for 10 min on ice. Lysed non-infected erythrocytes were used as negative control. 5x SDS-PAGE loading buffer containing 25 mM dithiothreitol was then added to the lysates, heat-denatured for 10 min at 95°C, and separated via SDS-PAGE and transferred to Hybond ECL nitrocellulose membrane (Amersham Biosciences) according to the manufacturer’s protocol. Blocking of non-specific binding sites was performed by incubation with 5% w/v skim milk and 1% w/v BSA in Tris-buffered saline (pH 7.5) overnight at 4°C. For immunodetection, membranes were incubated for 2 h at RT with polyclonal mouse anti-7-Helix-1rp2 antisera (dilution 1:1,000), polyclonal mouse anti-Pf39 antisera (dilution 1:500), polyclonal mouse anti-PfM1-AP antisera (dilution 1:50), polyclonal mouse anti-Falcilysin antisera (dilution 1:50), polyclonal mouse anti-Pfs16 antisera (dilution 1:100), rabbit anti-HSP70-1 antibody (dilution 1:5,000), rabbit anti-HA antibody (dilution 1:200), rabbit anti-CITH antibody (dilution 1:1,000), rabbit anti-PABP1 antibody (dilution 1:1,000), rabbit anti-CCp2 antisera (1:1,000) or rabbit anti-GFP antibody (dilution 1:1,000) in blocking solution or TBS/0.1% Tween-20. Following several washing steps, the membranes were incubated for 1 h at RT with a goat anti-mouse or anti-rabbit alkaline phosphatase-conjugated secondary antibody (dilution 1:10,000; Sigma-Aldrich) and developed in a solution of nitroblue tetrazolium chloride (NBT) and 5-bromo-4-chloro-3-indoxyl phosphate (BCIP; Roche) for 15 min at RT. Blots were scanned and processed using Adobe Photoshop CS5 software. For quantification of band intensities ImageJ 1.51f was used; data analysis was performed using MS Excel 2013.
To determine the antimalarial effect and inhibitory concentrations of the translation inhibitors emetine and cycloheximide (Sigma-Aldrich), a Malstat assay was performed as described previously [99]. P. falciparum WT NF54 cultures synchronized for ring stages were plated in triplicate in 96-well plates (200 μl/well) at a parasitemia of 1% in the presence of inhibitors in concentrations ranging from 500 μM to 0.5 nM. Chloroquine, dissolved in double-distilled water, served as a positive control in the experiments; incubation of parasites with the solvents alone (double-distilled water for emetine, ethanol for cycloheximide) was used as a negative control. After cultivation for 72 h, the parasites were resuspended, and aliquots of 20 μl were removed and added to 100 μl of the Malstat reagent in a 96-well microtiter plate. Parasite lactate dehydrogenase activity was evaluated by adding 20 μl of a mixture of NBT (nitroblue tetrazolium) and diaphorase (1:1; 1 mg/ml stock each) to the Malstat reaction, and measurement of optical densities at 630 nm. Each compound was tested three times, and the IC50 values were calculated from variable-slope sigmoidal dose-response curves using the GraphPad Prism 5 program.
For the comparison of asexual blood stage replication and the development of gametocytes between WT NF54 and the 7-Helix-1-KO lines 1D12 and 2E6, blood stage cultures were synchronized and set to an initial parasitemia of 2% ring stages. To compare asexual blood stage replication, Giemsa-stained thin blood smears were prepared over a time span of 49 h at seven different points (0, 15, 20, 25, 39, 44, and 49 h post-seeding); to compare the development of gametocytes, Giemsa-stained thin blood smears were prepared over a time span of 15 days at six different time points (6, 7, 8, 10, 13, and 15 d post-seeding). The smears were analyzed by light-microscopy at 1,000-fold magnification and 50 parasites were counted four times and grouped according to their developmental stage.
Mature gametocytes of the WT NF54 and the 7-Helix-1-KO line 2E6 were activated in vitro with 100 μM xanthurenic acid. At 15 min p.a. the numbers of exflagellation centers were examined microscopically and counted at 400-fold magnification in 30 optical fields for four times. The relative numbers of exflagellation centers were calculated (WT set to 100%). Four independent experiments were conducted; data analysis was performed using MS Excel 2013.
For the analysis of macrogamete and zygote formation, mature gametocytes of WT NF54, the 7-Helix-1-KO lines 2E6 and 1D12 and the complementation line 7-Helix-1-KO(+) were enriched via Percoll purification and set to a gametocytemia of 2–4%. For inhibitor experiments, the parasites were incubated with emetine or cycloheximide at IC50 concentration (as determined by Malstat assay) for 30 min at 37°C. Incubation with the solvents alone was used as a negative control. The gametocytes were activated and at 30 min p.a. (macrogametes) or 4 h p.a. (zygotes), the samples were subjected to IFA. Immunolabeling was performed as described above and macrogametes and zygotes were labelled with rabbit anti-Pfs25 antisera. Parasites were counted microscopically using a Leica DM 5500B fluorescence microscope with 600-fold magnification. For comparative macrogamete and zygote formation assays, 30 optical fields were counted for three to four times. The relative numbers of macrogametes and zygotes were calculated (WT set to 100%). Three to eight independent experiments were conducted; data analysis was performed using MS Excel 2013 and GraphPad Prism 5. For inhibitor experiments, macrogametes and zygotes per 1,000 RBC were counted for five times. Two independent experiments were performed; data analysis was performed using MS Excel 2013 and GraphPad Prism 5.
Mature gametocytes of NF54 WT and the 7-Helix-1-KO line 2E6 were enriched via Percoll purification and activated as described above. Samples were collected at 0 and 30 min p.a. and fixed with 1% v/v glutaraldehyde and 4% w/v paraformaldehyde/PBS (pH 7.4) overnight at 4°C. Post-fixation of the specimens was performed with 1% v/v osmium tetroxide and 1.5% w/v K3Fe(CN)6 in PBS for 2 h at RT, followed by incubation in 0.5% w/v uranyl acetate for 1 h. For dehydration of the specimens, increasing concentrations of ethanol (70%/80%/95%/100%) were used, followed by an incubation step for 1 h in propylene oxide and another 1 h incubation step in a 1:1 mixture of propylene oxide and Epon (Sigma-Aldrich). Subsequently, specimens were embedded in Epon at 60°C for 48 h. Ultrathin sections were cut with a Leica ultramicrotome Ultracut UCT and post-stained with 1% w/v uranyl acetate for 30 min and 0.2% w/v lead citrate for 15 s. Examination of the sections was performed using a CM100 transmission electron microscope (FEI) and images were recorded digitally with a Quemesa TEM CCD camera and iTEM software (Olympus Soft Imaging Solutions). Alternatively, samples were analysed with a Zeiss EM10 transmission electron microscope and the photographs taken were scanned and processed using Adobe Photoshop CS software.
Mature gametocytes of the WT NF54 and the 7-Helix-1-KO line 2E6 were enriched using Percoll purification and the purity of the gametocyte samples was confirmed via Giemsa smears (S16C Fig). The gametocytes were activated and at 30 min p.a., total RNA was isolated as described above. Quality of RNA samples were assessed using a ND-1000 (NanoDrop Technologies, Thermo Scientific) and by agarose gel electrophoresis. The microarray experiments were carried out as described previously [43,96]. Briefly, first strand amino-allyl cDNA was synthetized using SuperScript II reverse transcriptase (Invitrogen) and then cleaned and concentrated using the Zymo DNA clean and concentrator-5 column (Zymo Research) followed by coupling with Cy5 dye (GE Healthcare). The reference pool consisted of a mixture of RNA from asexual blood stages and gametocytes in which synthesis of first strand amino-allyl cDNA was performed as described above and coupled with Cy3 dye. Equal amounts of Cy5-labelled sample and Cy3-labelled reference pool were subjected to array hybridization for 17 h at 65°C using a P. falciparum DNA Agilent microarray chip (Agilent Technologies, AMADID #037237) containing the 5,363 coding genes [42]. The Agilent G2600D microarray scanner (Agilent Technologies) was used to scan the arrays. Normalized intensities were extracted using the Agilent feature extractor software version 11.5.1.1 and uploaded to the Princeton University Microarray Database (PUMA.princeton.edu) for analysis. After background subtraction, the log2 of the (Cy5/Cy3) intensity ratio was extracted and the transcript abundance of 7-Helix-1-KO samples was compared to that of WT NF54 samples. For the selection of up- and down-regulated genes, a cut-off value of greater than 1.5-fold applied. Data were analysed using MS Excel 2010. The database PlasmoDB (http://plasmodb.org/plasmo; [74]) was used for gene annotation analyses. GO enrichment analysis for deregulated genes was performed using the PlasmoDB GO analysis tool, p-value < 0.05. Genes with high transcription in gametocytes were defined by transcript levels > 50% of peak transcript levels as published previously [100].
The mRNA-FISH-IFA was performed following previously published protocol [101]. WT NF54 or 7-Helix-1-HA gametocytes were liberated from the enveloping RBC by lysis with 0.05% saponin/PBS followed by three washing steps in RPMI1640/HEPES medium (Gibco). Cell monolayers were air-dried on glass slides and fixed with 100% methanol. Permeabilization was performed with 0.1% v/v Triton X-100/125 mM glycine (Carl Roth)/PBS at RT for 10 min. After incubation in 2x SSC for 10 min at RT, hybridization was carried out by incubation of the sample with hybridization buffer (50% deionized formamide/200 μM dextran sulfate (MW = 500,000 g/mol) in 20x SSPE buffer) containing 1 μg biotinylated oligo-dT25 probe (Promega) in a humidity chamber overnight at 37°C. The slides were washed twice with 2x SSC for 30 min and once with 0.5x SSC for 15 min at RT, followed by the addition of Alexa Fluor 594-conjugated streptavidin (dilution 1:500; Thermo Scientific) in 4x SSC for 1 h at RT. Washing of the slides was performed twice with 4x SSC for 10 min and once with 4x SSC with 0.01% Triton X-100 for 10 min at RT. Counterlabelling was performed by incubation with mouse anti-7-Helix-1rp2 antisera (dilution 1:50), rabbit anti-Pfs230 antisera (dilution 1:200) or rabbit anti-HA antibody (Sigma-Aldrich, dilution 1:50) in 1% blocking reagent (Roche) in maleic acid buffer for 1 h at 37°C. Following three washing steps in maleic acid buffer, binding of primary antibody was detected by incubating the samples with Alexa Fluor 488-conjugated goat anti-rabbit or anti-mouse IgG antibodies (dilution 1:1,000; Molecular Probes) diluted in 1% blocking reagent (Roche) in maleic acid buffer for 1 h at RT. The samples were washed with maleic acid buffer for three times and nuclei were highlighted by incubation with Hoechst nuclear stain 33342 for 1 min at RT. Subsequently, cells were mounted with anti-fading solution AF2 (Citifluor Ltd) and sealed with nail varnish. Samples were analyzed using a Leica DM 5500B fluorescence microscope and digital images were processed using Adobe Photoshop CS5 software.
Data are expressed as means ± SD. Statistical differences were determined using two-tailed Mann-Whitney-U test, unpaired two-tailed Student’s t-test or One-Way ANOVA with Post-Hoc Bonferroni Multiple Comparison test, as indicated. P values < 0.05 were considered statistically significant. Significances were calculated using GraphPad Prism and are represented in the figures as follows: ns, p > 0.05; * p < 0.05; ** p < 0.01; *** p < 0.001; **** p < 0.0001.
|
10.1371/journal.pgen.1002326 | Natural Selection Affects Multiple Aspects of Genetic Variation at Putatively Neutral Sites across the Human Genome | A major question in evolutionary biology is how natural selection has shaped patterns of genetic variation across the human genome. Previous work has documented a reduction in genetic diversity in regions of the genome with low recombination rates. However, it is unclear whether other summaries of genetic variation, like allele frequencies, are also correlated with recombination rate and whether these correlations can be explained solely by negative selection against deleterious mutations or whether positive selection acting on favorable alleles is also required. Here we attempt to address these questions by analyzing three different genome-wide resequencing datasets from European individuals. We document several significant correlations between different genomic features. In particular, we find that average minor allele frequency and diversity are reduced in regions of low recombination and that human diversity, human-chimp divergence, and average minor allele frequency are reduced near genes. Population genetic simulations show that either positive natural selection acting on favorable mutations or negative natural selection acting against deleterious mutations can explain these correlations. However, models with strong positive selection on nonsynonymous mutations and little negative selection predict a stronger negative correlation between neutral diversity and nonsynonymous divergence than observed in the actual data, supporting the importance of negative, rather than positive, selection throughout the genome. Further, we show that the widespread presence of weakly deleterious alleles, rather than a small number of strongly positively selected mutations, is responsible for the correlation between neutral genetic diversity and recombination rate. This work suggests that natural selection has affected multiple aspects of linked neutral variation throughout the human genome and that positive selection is not required to explain these observations.
| While researchers have identified candidate genes that have evolved under positive Darwinian natural selection, less is known about how much of the human genome has been affected by natural selection or whether positive selection has had a greater role at shaping patterns of variation across the human genome than negative selection acting against deleterious mutations. To address these questions, we have combined patterns of genetic variation in three genome-wide resequencing datasets with population genetic models of natural selection. We find that genetic diversity and average minor allele frequency are reduced in regions of the genome with low recombination rate. Additionally, genetic diversity, human-chimp divergence, and average minor allele frequency have been reduced near genes. Overall, while we cannot exclude positive selection at a fraction of mutations, models that include many weakly deleterious mutations throughout the human genome better explain multiple aspects of the genome-wide resequencing data. This work points to negative selection as an important force for shaping patterns of variation and suggests that there are many weakly deleterious mutations at both coding and noncoding sites throughout the human genome. Understanding such mutations will be important for learning about human evolution and the genetic basis of common disease.
| A substantial amount of effort in human population genetics has been aimed at understanding how natural selection operates in the human genome. However, we lack a basic understanding of the importance of positive natural selection versus negative selection at shaping overall patterns of genome variation. Thus far, most of the attention has been aimed at locating genes that have been under positive selection [1]–[19]. These studies have identified several hundred candidates throughout the genome that may have been affected by positive natural selection. However, fewer studies have attempted to gauge the prevalence of positive natural selection in the human genome. Those that have attempted have come to very different conclusions. Several studies suggested that positive selection may be common, with around 10% of the genome having been affected by a recent selective sweep [9], [10], [14], [16]. Other studies argued that selective sweeps were less common [20], [21]. Finally, some have estimated that approximately 10%, but perhaps up to 40%, of nonsynonymous human-chimp differences have been fixed by positive natural selection [22], [23]. Thus, there is little consensus regarding the importance of positive natural selection at shaping patterns of variability.
Additionally, the role of negative selection at shaping broad patterns of genetic variation across the genome needs to be clarified. Many studies have suggested that nonsynonymous mutations and mutations in conserved noncoding sequences are weakly deleterious but may persist in the population due to genetic drift and other demographic phenomena [22], [24]–[32]. The effect that these weakly deleterious mutations have on nearby patterns of genetic variation remains unclear. Furthermore, the importance of negative versus positive selection at shaping overall patterns of variation also remains ambiguous.
If natural selection (either positive or negative) is common in the genome, it should affect patterns of genetic variation at linked neutral sites across the genome [33], [34]. Selection may alter genetic variation in different ways. We review these ways, discuss the empirical evidence for these effects, and highlight open questions that our study seeks to address.
First, selection may generate a correlation between levels of neutral diversity and recombination rate [35], [36]. This can occur under models with strong positive selection (selective sweeps) or negative selection acting on many deleterious mutations (background selection). Selective sweeps remove genetic diversity at linked neutral sites [33], [37]. In a region of the genome with a low recombination rate, a large length of sequence will have the same genealogy as the selected site. As such, the selective sweep will remove neutral variation over a larger portion of the sequence in low recombination rate regions than in regions with higher recombination rates. Background selection against deleterious mutations can also generate this correlation [34], [38]–[41]. Chromosomes carrying many deleterious mutations will be rapidly eliminated from the population. Any neutral variation linked to the deleterious mutations will also be eliminated from the population. This model predicts reduced variability in regions of the genome with low recombination rate because, as with the case of a selective sweep, a larger portion of the chromosome will share the same genealogy as the selected site(s) in regions of low recombination rather than in high recombination. Several studies have searched for a correlation between diversity and recombination rate in humans. Early studies based on a small number of genes came to conflicting conclusions. Nachman et al. [42], [43] found a significant correlation between diversity and recombination rate, but found no correlation between divergence and recombination rate, suggesting the effects of natural selection. Hellmann et al. [44], examining a different dataset, found that the correlation between diversity and recombination rate disappeared after correcting for human-chimp divergence. They suggested that recombination may be mutagenic and that the original correlation was driven by co-variation of mutation and recombination rates. Another study found that microsatellite diversity was not correlated with recombination rate [45]. More recent studies on larger datasets have found significant correlations between diversity and recombination rate [46]–[48]. These studies have found that the correlation between human diversity and recombination rate persists after controlling for human-chimp divergence. While this is suggestive of the effects of natural selection, important features of this correlation have yet to be characterized. For example, if natural selection is primarily driving the correlation, the correlation ought to be stronger in genic regions of the genome than in non-genic regions, because functional sites near genes are the most likely targets of selection. This feature has yet to be explored.
Second, natural selection may generate a correlation between the allele frequency distribution and recombination rate. Specifically, models of selective sweeps predict a skew toward an excess of low-frequency single nucleotide polymorphisms (SNPs) near the target of selection [49]–[51]. Following the same logic as above, a larger region of the genome will be affected in areas with lower recombination rates, thus generating a correlation between allele frequency and recombination rate. The effect of background selection on allele frequencies is less clear. Simulation studies have suggested that intermediate strengths of background selection, especially in regions of low recombination, can generate a skew toward an excess of low-frequency SNPs [34], [38], [52]–[58]. Most of the analytical formulae that describe background selection model the process as a reduction in effective population size, which does not predict a skew of the frequency spectrum ([34], [38]–[41], but see Santiago and Caballero [59]). Consequently, it has been argued that the effect of background selection on the frequency spectrum is rather weak, and as such, a skew toward low-frequency SNPs is more indicative of positive, rather than background selection [60]–[66]. It is unclear whether there is a correlation between allele frequency and recombination rate in the human genome, though several small studies have found suggestive evidence [6], [67]. Furthermore, it is unclear which models of selection may be compatible with such a correlation.
Third, if selection is common, it ought to primarily affect patterns of genetic variation near genes because genes are the likely targets of selection. Several studies have found that human-chimp divergence and human diversity were reduced near genes, suggesting the importance of selection at shaping overall patterns of variability throughout the genome [67]–[69]. It is less clear whether there is a skew toward low-frequency alleles near genes.
Fourth, pervasive positive natural selection may generate a negative correlation between nonsynonymous divergence and levels of neutral genetic diversity ([70]–[77] and reviewed in [78]). The reason for this is that selective sweeps acting on amino acid changing mutations generate nonsynonymous fixed differences between species. Regions of the genome that have been affected by these sweeps will likely also have reduced neutral polymorphism, thus generating the negative correlation between these two quantities. It is unclear whether such a correlation can be generated in the absence of positive selection and how strong the correlation might be under various models of positive selection.
Here we further investigate these issues by studying patterns of genetic variation in three different genome-wide genetic variation datasets obtained from resequencing European individuals. We find that levels of diversity are positively correlated with recombination rate and negatively correlated with genic content. Minor allele frequency is also positively correlated with recombination rate and negatively correlated with genic content. Using simulations, we show that these correlations are best explained by a model where many sites are under weak negative selection. Models with numerous selective sweeps on nonsynonymous mutations predict too strong a negative correlation between neutral polymorphism and nonsynonymous divergence. Though not required to explain the data, some smaller fraction of sites may be under positive selection. Overall, this work points to the importance of weak negative selection at shaping patterns of variation throughout the human genome.
We analyzed genomic patterns of polymorphism from three genome resequencing datasets. First, we analyzed low-coverage next-generation sequence data obtained from an exome-capture study of 2,000 Danish individuals. Due to the non-specificity of the exome-capture arrays, portions of the genome outside of the targeted regions were sequenced, but at lower coverage. Given the shallow sequencing depth across most of the genome (roughly 0.1× per individual), it would be impossible to infer genotypes for each individual with any appreciable accuracy. Instead, we implemented a statistical approach to estimate the population allele frequency of a SNP using the counts of different nucleotides at a particular site in the genome (see Materials and Methods for a detailed description). When combining reads across all individuals, approximately 30–40% of the genome had a sequencing depth of at least 100 reads. We estimated the minor allele frequency (MAF) for all of these sites with a depth of at least 100 reads. Those sites with an estimated MAF>5% were considered to be SNPs in this dataset and were used for subsequent analyses. We used this conservative cut-off because of the difficulties in reliably estimating allele frequencies of rare alleles in low-coverage data [79].
In order to verify patterns found in our low-coverage resequencing dataset, we also analyzed two other complementary datasets. One dataset consisted of six European genomes that were sequenced to higher coverage (denoted “higher coverage,” see Materials and Methods for details). The other dataset consisted of five genomes from Utah residents with ancestry from northern and western Europe (abbreviated CEU) and one genome from a Toscan individual sampled from Italy (abbreviated TSI) sequenced to high coverage by Complete Genomics (denoted “CGS,” see Materials and Methods). Summaries of genetic variation were positively correlated across the three datasets (Figure S1 and Figure S2). Due to the stochasticity of the evolutionary process, even with perfect data, patterns of polymorphism will not be perfectly correlated across different datasets.
To analyze correlations between different summaries of polymorphism and other genomic features, we divided the genome into non-overlapping 100 kb windows (see Materials and Methods for further details). Within each window, we tabulated the number of SNPs, average MAF, number of human-chimp differences, GC content, recombination rate (as estimated from the high-resolution deCODE map [80]), fraction of each window where sequencing data was available, and the fraction of the window that overlaps with a RefSeq gene. Since we wanted to examine the indirect effects of natural selection due to linkage, rather than assess the effects of natural selection on the selected sites themselves, all of our analyses removed the roughly 5% of the genome that was most conserved across species (i.e. the phastCons regions [81], see Materials and Methods). These were the regions most likely to be directly under negative selection in the human genome [81]. We then assumed that the remaining sequence that we analyzed was selectively neutral. Because many of the genomic features were correlated with each other (Table S1, Table S2, Table S3), we performed partial correlation analyses to remove the effects of possible confounding variables. The partial correlation can be thought of as the correlation between two variables when one or more other confounding variables are held constant. We used partial correlations, rather than a full multivariate analysis, because the partial correlations have a simpler biological interpretation and have been used in other recent evolutionary studies [82].
We found a strong positive correlation between the number of SNPs in a window and the recombination rate of the window (Spearman's , Table S1) when looking at the low-coverage data. We also observed a strong correlation between the number of human-chimp differences within a window (d) and recombination rate (Spearman's , Table S1). When scaling diversity by divergence (i.e. dividing the number of SNPs per covered base within a window by the number of human-chimp differences) to potentially account for differences in mutation rate across the genome, we still found a strong correlation between scaled SNP diversity (defined here as Snorm) and recombination rate (Spearman's , Table 1, Table S1). In particular, regions of the genome with low rates of recombination (i.e. <0.5 cM/Mb) had especially low levels of polymorphism. The rate of change of Snorm was less dramatic over the rest of the range of recombination rates.
We also found a positive correlation between Snorm and recombination rate when analyzing the higher-coverage and CGS datasets (Spearman's , Table 1, and Table S2 for the higher-coverage data; Spearman's , Table 1, and Table S3 for the CGS data). The correlation was even stronger than that observed in the low-coverage data. We discuss several possible reasons for this difference in the Discussion section. Nevertheless, the fact that we found the correlation in all three datasets strongly argues that it is a true biological correlation and not an artifact due to biases in the low-coverage Danish data. The correlation between Snorm and recombination rate remained significant even after controlling for GC content, d, the number of neutral bases covered by sequencing data, and the fraction of genic bases within a window (Table 1), suggesting that these factors cannot completely explain this correlation. Further, the average number of pairwise differences per window normalized by d was also positively correlated with recombination rate in both datasets (Table S2 and Table S3).
If natural selection is responsible for this correlation between Snorm and recombination rate, it may be stronger in genic regions of the genome than in non-genic regions. The reason for this is that, all else being equal, genic regions will likely experience more natural selection than non-genic regions. Non-genic windows were defined to be those that did not overlap with a RefSeq transcript. Genic windows were those where at least half the window overlapped with a RefSeq transcript.
Indeed, the correlation was significantly stronger in genic windows than in non-genic windows in all three datasets (P<0.0001 by permutation test, Figure 1, Table 2, Figure S3, and Figure S4). This pattern holds even after controlling for confounding variables using a partial correlation analysis. Inspection of the lowess lines in Figure 1A illustrates the differences between the correlation in genic and non-genic regions. In genic regions with low recombination rates (<0.5 cM/Mb), there is a sharp decrease in Snorm. However, non-genic regions with low recombination rates did not show such a pronounced decrease in Snorm (Figure 1A). One concern with these analyses is that the low-coverage dataset was an exome resequencing dataset and the exome-capture process may have resulted in systematic differences between genic and nongenic regions. However, we found the same pattern in the higher-coverage dataset and the CGS dataset, which were not targeted toward genes or exons (Figure S3 and Figure S4). This argues that the differences between genic and non-genic regions were not due to systematic biases in the data, but rather to inherent differences between genic and non-genic regions of the genome.
We then examined the correlation between average MAF within a window and recombination rate (Table 1 and Table S1) in the low-coverage data. We found a weak, but statistically significant, positive correlation between these two variables (Spearman's ). In regions of low recombination, there was a skew toward lower average MAF. The correlation remained significant even after controlling for GC content, d, the number of neutral bases covered by sequencing data, and genic content, suggesting that it cannot be completely explained by these other factors (Spearman's , Table 1). Finally, we also found a positive correlation between average MAF and recombination rate in the higher-coverage and the CGS data (Table 1, Table S2, Table S3), again suggesting that it was not due to biases in estimating SNP frequencies from low-coverage data. A different summary of the frequency spectrum, Tajima's D [49], also showed a correlation with recombination rate (Table S2 and Table S3), indicating that this correlation was not sensitive to the summary of the frequency spectrum employed.
However, no clear pattern emerged when testing whether the correlation between average MAF and recombination rate was stronger in genic versus non-genic regions. For all three datasets, the pairwise correlation between average MAF and recombination rate was higher in genic regions than non-genic regions (P<0.05, by permutation test, Figure 1B, Figure S3B, Figure S4B, Table 2). In the higher-coverage dataset, genic regions showed a stronger correlation between MAF and recombination rate than non-genic regions even after controlling for GC content, d, and the number of bases covered by sequencing data using a partial correlation analysis (P<0.02 by permutation test, Table 2). However, after controlling for the confounding variables, there was little difference in the partial correlation coefficients between genic and non-genic regions in the low-coverage and the CGS datasets (Table 2). Thus, there was no clear evidence suggesting that the correlation between MAF and recombination rate was stronger in genic than non-genic regions of the genome. This may not be surprising because this correlation was quite weak, making it difficult to detect subtle changes in its strength across the genome.
If natural selection affects patterns of genetic variation across the genome, Snorm, average MAF, and d may be reduced in windows of the genome that contain more genic bases. These patterns would be expected if most of the selection in the genome occurs near genes, rather than in intergenic regions.
Indeed, in all three datasets, we found a negative correlation between Snorm and the fraction of bases within a window that overlapped with a RefSeq transcript (Table 1). In other words, windows with a higher genic content tended to have fewer SNPs. These correlations became stronger when controlling for d, recombination rate, the fraction of the window with sequencing coverage, and GC content (Table 1).
There was a weak, but significant, negative correlation between MAF and fraction of bases that overlapped with a RefSeq transcript in all three datasets examined (Table 1). Windows with a higher genic content tended to have lower average MAF than windows with lower genic content. In the low-coverage and higher-coverage datasets, the correlation became stronger when controlling for d, recombination rate, the fraction of the window with sequencing coverage, and GC content (Table 1).
Finally, we found a very strong negative correlation between d and the fraction of genic bases within a window (Spearman's , Table S1, Table S2, Table S3). These results were in agreement with those from a study [67] which found reduced diversity and divergence near genes even after removing the regions of the genome most conserved across species (i.e. the phastCons elements).
We next tested whether there was a correlation between Snorm and the number of nonsynonymous human-chimp differences within a window (DN). A negative correlation between these two variables has been interpreted as evidence of selective sweeps across the genome ([70]–[77] and reviewed in [78]). When tabulating DN, we did not remove sites which were conserved across species. We observed weak negative correlations between Snorm and DN as well as between Snorm and the number of synonymous human-chimp differences (DS) for several of the datasets (Table S4). However, when we normalized DN by the number of nonsynonymous sites per window (the normalized value is called dN) or used a partial correlation analysis to control for the number of nonsynonymous sites per window, none of the datasets showed a significant negative correlation (Table S4). The same was true for synonymous human-chimp differences.
Haddrill et al. [77] suggested that a negative correlation between Snorm and dN may be more apparent in genes with elevated dN. Thus, we also tested for a correlation between Snorm and dN using only the windows in the 90th percentile of dN. In general, the values of Spearman's were more negative in this subset of the data than when analyzing the entire dataset (Table S5). For example, in the CGS data, when controlling for d, GC content, recombination rate, the number of nonsynonymous sites, and the fraction of the window with sequencing coverage. However, Snorm was also negatively correlated with dS in the windows in the 90th percentile of dS (, controlling for d, GC content, recombination rate, the number of synonymous sites, and the fraction of the window with sequencing coverage). The fact dS showed a similar negative correlation with Snorm as dN did, combined with the fact that synonymous sites are usually assumed to be neutrally evolving in humans, suggested that these correlations may have been driven by a neutral process, rather than positive selection. One possibility was that the recent fixations of neutral synonymous or nonsynonymous mutations led to a decrease in neutral diversity, as suggested by earlier theoretical work [83]. As such, regions with high dN (or high dS) would have lower Snorm, generating the negative correlation. Overall, these results suggest that regions of the genome that have more nonsynonymous human-chimp differences do not have lower levels of neutral polymorphism, beyond the reduction in diversity already expected in genic regions of the genome or surrounding neutral fixations.
We next evaluated whether population genetic models including population size changes, recombination rate variation, and natural selection could generate the correlations that we observed in the empirical datasets. We simulated 100 kb regions consisting of exons, introns, and an intergenic sequence (see Materials and Methods, Figure S5). We examined several different models of selection (see Table S6 for the specific parameter values) and examined the correlation between patterns of genetic variation in the neutrally evolving intergenic sequence and other genomic attributes. Because many studies have found that nonsynonymous mutations are weakly deleterious [22], [26], [28], [84], one model included weak negative selection acting only on nonsynonymous sites (shown in purple in Figure 2). It had been suggested that conserved noncoding sites are also likely to be weakly deleterious [25], [27], [31], so another model included negative selection acting on a fraction of intronic sites (shown in blue in Figure 2). In the third model (shown in orange in Figure 2), most mutations at nonsynonymous positions were negatively selected, but a small fraction was positively selected. Finally, the fourth model added weak negative selection at a fraction of intronic sites to a model where most mutations at nonsynonymous positions were negatively selected, but a small fraction was positively selected.
Our simulations confirmed previous predictions that both hitchhiking and background selection [33], [34], [37]–[41] could generate a positive correlation between genetic diversity at linked neutral sites and recombination rate (Figure 2A and Figure S6A). Importantly, these simulations demonstrated that the background selection effect can occur with weak negative selection acting on many sites simultaneously. Models with negative selection acting on noncoding and coding mutations, as well as models with positive selection, could generate positive correlations similar to those in the observed data (red lines in Figure 2A and Figure S6A).
Models of natural selection predicted a positive correlation between average MAF at linked neutral sites and recombination rate (Figure 2B and Figure S6B). The strongest correlations seen for models with only negative selection were for intermediate strengths of selection (e.g. 25% of intronic sites with s = 2.5×10−4). Stronger selection (s = 5×10−3) resulted in a weaker correlation (Table S7). Importantly, models that contained no sites under positive selection predicted a correlation between MAF and recombination rate roughly similar in magnitude to that seen in the observed data (red lines in Figure 2B and Figure S6B). These results suggest that both positive and weak negative selection were capable of affecting allele frequencies at linked neutral sites. Thus, a correlation between allele frequency and recombination rate cannot be taken as unambiguous evidence of positive selection.
In some cases, the correlation coefficients between MAF and recombination rate and diversity and recombination rate were significantly higher than zero under purely neutral models (Figure 2 and Figure S6). We performed coalescent simulations using ms [85] under the standard neutral model with different rates of recombination to further investigate this issue. Not only was the variance of the distribution of diversity (or average MAF) greater in simulations without recombination, but the shape of the distribution changed depending on the recombination rate. For example, in the case of a high recombination rate, the distribution of the number of segregating sites approached a Poisson distribution, and was symmetric about its mean. However, with no recombination, the distribution became less symmetric, with a higher mass below the mean and a longer tail to the right (Figure S7). Thus, the median of the distribution of diversity simulated with no recombination was lower than the median of the distribution with the high recombination rate. As such, a weak positive correlation between recombination rate and diversity may be expected. The same arguments hold for understanding the correlation between MAF and recombination rate (Figure S7) and Tajima's D and recombination rate (Figure S7, see also [63], [86]). Since we used simulations to interpret the correlations observed in the actual data, this effect did not alter our interpretation.
Previous authors ([70]–[77] and reviewed in [78]) had suggested that a negative correlation between neutral polymorphism and nonsynonymous divergence may be a signature of positive selection that cannot be generated by negative selection and/or demographic processes. In our simulations, a model with negative selection acting on noncoding sites, but where a fraction of coding mutations were positively selected showed a negative correlation between Snorm and dN (orange points in Figure 2C and Figure S6C). Models that did not include any positive selection, but included negative selection on a fraction of noncoding sites (blue points in Figure 2C and Figure S6C), showed little correlation between these two variables. Thus, for the models investigated here, the negative correlation was specific to models of positive selection. As such, it may offer a way to distinguish between models of negative and positive selection. However, a significant negative correlation was not always seen in models that included some sites under positive selection (green points in Figure 2C and Figure S6C). Instead, the correlation was influenced by the relative amounts of negative versus positive selection. Negative selection made the correlation more positive, while positive selection made the correlation more negative. The correlation ultimately observed was due to the net effect of both types of selection.
We next used the simulations to evaluate what role positive selection may have played in shaping patterns of variability across the genome. We first examined models with only strong positive selection. A model where 0.5% of nonsynonymous mutations were positively selected (s = 0.625%) could generate the observed correlation between Snorm and recombination rate (black, p+ = 100%, p− = 0% in Figure 3A; p+ denotes the proportion of simulated windows where positive selection could occur). However, this model predicted too strong a negative correlation between Snorm and dN to be compatible with the data (black, p+ = 100%, p− = 0% in Figure 3B). Because several studies have suggested that 0–10% of the genome has been affected by a selective sweep [9], [10], [14], [16], [20], [21], we next examined a model where 5% of the simulated windows included positive selection. A model where the remaining 95% of the windows were neutral does not predict a correlation between Snorm and recombination rate strong enough to match the actual data (black, p+ = 5%, p− = 0% in Figure 3A). This suggests that a small number of positively selected sites by themselves are not sufficient to generate this correlation. Further, this model still predicted a negative correlation between Snorm and dN (black, p+ = 5%, p− = 0% in Figure 3B). However, a model where 5% of the simulated windows included positive selection and the remaining 95% of windows included negative selection on coding and noncoding sites predicted a correlation between Snorm and recombination rate similar to that observed in the actual data (black, p+ = 5%, p− = 95% in Figure 3A). Because adding negative selection resulted in an increase in the strength of this correlation, we concluded that the correlation observed in the data has been primarily driven by negative selection. Also, under this model, the negative correlation between Snorm and dN was very weak and was compatible with that from the actual data (black, p+ = 5%, p− = 95% in Figure 3B), presumably because most of the windows have been subjected to negative selection. A model where the strength of positive selection was weaker showed similar trends (pink points in Figure 3). This analysis indicated that the correlation between neutral diversity and recombination rate was primarily driven by many weakly deleterious polymorphisms across the genome, rather than by a small proportion of strongly positively selected mutations.
Finally, our simulations (Figure 4) suggest that negative or positive selection can generate a strong correlation between neutral human-chimp divergence (d) and recombination rate even when the mutation rate is constant across all simulation replicates. This correlation was likely driven by selection occurring in the ancestral population [67], [87]. Thus, the correlation between d and recombination rate can be readily explained by mechanisms other than recombination itself being mutagenic [44], [46], [88].
We have examined patterns of putatively neutral genetic variation in three genome-wide resequencing datasets to gauge the extent of natural selection throughout the human genome. To the best of our knowledge, this is the first report that the allele frequency spectrum is correlated with recombination rate across the human genome (though suggestive evidence was found in smaller datasets [6], [67]). As discussed below, these correlations are best explained by natural selection affecting linked neutral variation across the human genome, rather than artifacts in the data or other mutational processes. Through the use of population genetic simulations, we have shown that a model with negative selection acting on both coding and noncoding mutations fits the data. While we cannot rule out models that include some positive selection, models with abundant positive selection on nonsynonymous mutations and little negative selection predict too strong a negative correlation between neutral polymorphism and nonsynonymous divergence.
In general, we observed qualitatively similar patterns in all three resequencing datasets. However, several of the correlations between different genomic attributes were stronger in the higher-coverage and CGS data than in the low-coverage data (Table 1 and Table 2). Several characteristics of the datasets may contribute to this difference. For example, the higher-coverage and CGS datasets are likely to be of higher quality than the low-coverage dataset. Additionally, a greater proportion of each window is covered in the higher-coverage and CGS datasets than in the low-coverage dataset (Figure S8). Both of these features lead to estimated correlation coefficients that are lower in the low-coverage data than in the higher-coverage and CGS data. Finally, the higher-coverage and CGS datasets contain a sample of a smaller number of chromosomes than the low-coverage data. Population genetic simulations suggest that some of the correlations are expected to be stronger in smaller samples than in larger samples (compare Figure 2 to Figure S6). Thus, the quantitative differences among the correlation coefficients across the different datasets are not too surprising. Instead, the fact that all three datasets show the same general trends is powerful evidence that the correlations are not technical artifacts specific to any one type of data.
Thus, it is our conclusion that these correlations were, at least in part, driven by natural selection across the human genome. Several lines of evidence support this conclusion. First, the correlations remain significant after filtering repetitive sequence and CpG islands (Table S8), and after controlling for the effects of GC content, suggesting that base composition or mutational patterns associated with base composition are not entirely responsible for the correlations.
Second, we have evaluated whether biased gene conversion, a neutral alternative sometimes invoked to explain signatures of natural selection [89], [90], can generate the correlations we have identified. Our simulations show that neutral models with biased gene conversion cannot generate a correlation between Snorm and recombination rate similar in magnitude to that observed in our datasets (Table S6 and Table S7).
The third line of evidence is that the correlation between neutral polymorphism and recombination rate is stronger in genic regions compared to non-genic regions. Natural selection would predominately occur closer to genes, while mutational effects would be distributed throughout the genome [88]. We have also found that both diversity and minor allele frequency are negatively correlated with genic content, suggesting a difference in patterns of variability between genic and non-genic regions of the genome. As discussed further below, models that include natural selection can readily account for these observed patterns.
We have explored which models of selection can generate the correlations that we observed in the actual data. While we have found population genetic models that qualitatively predict the correlations that we have observed in the data, it is more difficult to translate these models into specific statements about the absolute amount of selection in the genome. For example, many of our simulations of negative selection on noncoding sites assume that 25% of intronic sites were under weak negative selection. This is likely to be a substantial over-estimate of the proportion of sites under negative selection [81], [91]. One explanation for this discrepancy is that, for computational convenience, we simulated 100 kb windows independently of each other, rather than whole chromosomes. In reality, each 100 kb window of the genome is linked to other selected mutations outside of the window that may affect patterns of diversity within the window. In fact, simulations of larger windows (348 kb) provide similar values of Spearman's when only 5% of intronic sties are under negative selection (Table S6 and Table S7). This may explain why the models that fit the data include so many selected sites. Simulating larger regions would only yield more biologically relevant simulations if we were able to simulate the correct magnitude of selection at noncoding sites, as well as the correct spatial distribution of sites under selection across the genome. Though there has been some progress from comparative and population genomic studies [25], [27], [31], [81], [91], further work is needed in this area. Additionally, there are nearly an infinite number of possible models for how selection can operate in the genome. For example, selection coefficients within a given window may be correlated with each other, and windows may not be exchangeable (i.e. each window may have its own distribution of selective effects). Our simulations do not capture these phenomena and instead merely illustrate the types of correlations predicted for very basic models of certain types of selection.
Nevertheless, our simplified models do allow some important qualitative statements regarding the relative importance of negative versus positive selection in the human genome. First, all of the correlations observed in all three datasets can be explained without invoking positive selection. Different models of negative selection can readily account for these correlations (Table S6 and Table S7). Second, based on the lack of a negative correlation between Snorm and dN in any of our datasets (Table S4, Figure 2C, Figure S6C), we can reject models with an abundance of selective sweeps acing on nonsynonymous mutations in the presence of few negatively selected sites (Figure 2, Figure 3, Figure S6). This finding is complementary to what was found in a recent study by Hernandez et al. [21].
However, we cannot rule out the presence of some positively selected mutations in the presence of many negatively selected ones. It is difficult to precisely estimate the fraction of the genome that has been affected by positive selection because such inferences are likely to be highly model-dependent and influenced by many unknown variables. Yet, for the model shown in Figure 3, which fits the actual data (black, p+ = 5%, p− = 95%), 5% of the simulated windows included positively selected mutations. This model predicts that roughly 2.3% of the windows will have at least one positively selected nonsynonymous mutation that fixed in humans within the last Ne generations (here 20,000 generations, or 500,000 years, assuming 25 years per generation). This is likely to be an upper bound on the fraction of the genome subjected to such strong positive selection because a higher fraction would predict a negative correlation between Snorm and dN that is too strong to match the data. However, if the strength of positive selection on individual mutations is weaker, if selection operates on standing variation, predominantly on noncoding mutations, or on multiple mutations simultaneously, then a much greater fraction of the genome could have been subjected to positive selection [20], [92]–[94]. Nonetheless, even if a small fraction of the genome was linked to a selective sweep, this amount of selection is not sufficient to generate the correlation between diversity and recombination rate seen in the actual data (Figure 3A). The widespread presence of weakly deleterious alleles, however, can generate this correlation, even in the presence of some positively selected sites (Figure 3A). Taken together, our results suggest that selective sweeps were not the dominant factor explaining the distribution of variability across the human genome.
The notion that sites under natural selection can affect linked neutral variation in the human genome has several important implications for learning about human history using genetic variation data. Most methods to infer parameters in population genetic models assume that all of the SNPs being analyzed are selectively neutral and are not linked to other sites that are affected by selection [17], [95]–[100]. Many of these methods summarize the genetic variation data by the number or proportion of SNPs at different frequencies in the sample (i.e. the frequency spectrum) and then find the demographic parameters that can generate the observed frequency spectrum. Compared to other regions of the genome, we found an excess of low-frequency SNPs in regions near genes and with low recombination rate. It is unlikely that these regions provide an accurate picture of the selectively neutral frequency spectrum for the population of interest. It is unclear what effect including such regions in demographic studies will have on the final parameter estimates. Further investigation of this topic is warranted. In the meantime, one way of circumventing the potential problem of natural selection confounding studies of demography would be to study regions of the genome far away from genes and with high recombination rate [101].
Finally, our study illustrates the utility of low-coverage sequencing data for population genetic studies. Here we have shown that analyzing the low-coverage data without first inferring individual genotypes provides estimates of allele frequency across the genome that are in broad agreement with estimates made from higher-coverage sequencing of a smaller number of individuals. Another unique feature of the low-coverage dataset was that it was generated as part of an exome-capture experiment [84]. Because the capture process is not completely specific and only enriches for sequences within the targeted regions, portions of the genome outside of the targeted regions were sequenced at a lower rate. Such data from a large number of individuals can be used to study patterns of genetic variation across the non-targeted regions of the genome, provided that one analyzes it using an approach that is appropriate for low-coverage data. Such studies promise to yield new insights in population and medical genetics.
The low-coverage dataset that we used here was an augmented version of the dataset published in Li et al. [84]. The sequencing was performed on 2,000 Danish individuals ascertained from three sources: 1) the population-based Inter99 study [102] (ClinicalTrials.gov ID-no: NCT00289237; n = 887), 2) the ADDITION study [103] (ClinicalTrials.gov ID-no: NCT00237548); n = 354) and 3) the Steno Diabetes Center (n = 759). All participants (mean age of 54.5 years) were of self-reported Danish nationality. All study participants provided written informed consent, and the study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Copenhagen County, Denmark. DNA from these individuals was analyzed in an exome-capture resequencing experiment. Each individual was sequenced separately without any pooling. NimbleGen2.1M HD arrays were used to enrich for exome sequences. These arrays contain probes complementary to exonic DNA fragments. Exonic DNA hybridized to the array while non-exonic DNA was washed away. However, this hybridization process was not perfect, and some non-exonic DNA remained bound to the array and was sequenced. The Illumina Genome Analyzer II was used to perform the sequencing. Further methodological details can be found in Li et al. [84].
The bioinformatic pipeline used for these data is similar to the one previously published [84]. First, reads were aligned to the NCBI human genome reference assembly (build 36.3) using SOAPaligner [104], [105]. Reads that mapped outside of the exome target regions were retained for further analyses, but bases with a Q score <20 were removed. Ideally, since we wish to compare allele frequency estimates for different regions of the genome, we would like to have a similar depth of coverage across the genome. However, depth of coverage varied greatly across the genome with the target regions having very high coverage and the non-target regions having substantially lower coverage. To circumvent this problem, at each position in the genome, we selected a random subset of 100 reads (from the 2,000 individuals) to be used for the frequency estimation process. We chose a cutoff of 100 reads since about 35–40% of the total genome was covered by at least 100 Q >20 bases. Decreasing this cutoff would increase the number of bases that were covered, but it would also make it harder to accurately estimate the frequency of lower-frequency SNPs.
We estimated allele frequencies directly from the read counts without attempting to call SNPs or individual genotypes from these data. For each site in the genome with at least 100 reads, we first estimated the population minor allele frequency (MAF) using the method-of-moments estimator [84]. For sites that had an estimated MAF >1% using , we obtained a more precise estimate of the MAF using the maximum likelihood approach described by Kim et al. [79], [106]. Due to computational constraints on analyzing a dataset of this size, we did not use the genotype likelihood files from soapSNP [107]. Rather, we used the binomial distribution to compute the probability of the read counts for each individual, taking the base-specific sequencing error probabilities into account. We treated the second-most common base at each site as the minor allele. Finally, only sites with estimated MAF >5% were considered as SNPs and were used in subsequent analyses. Given the low depth of coverage (100 reads), it would be difficult to distinguish lower-frequency SNPs from sequencing errors. For example, for a SNP with a MAF of 1%, the less common allele would only be seen approximately one time across all individuals.
We also analyzed a dataset of six European individuals whose genomes were sequenced to higher coverage. This dataset is complementary to the low-coverage dataset because each individual in this dataset was sequenced to higher coverage, coverage was more uniform across the genome, and a higher fraction of bases were covered. But, the sample depth at any particular site in the genome was substantially lower (only 12 chromosomes at most). This dataset included the genomes of James Watson [108], Craig Venter [109], the two parents from a CEU trio (NA12891 and NA12892) that was sequenced to high coverage in pilot 2 of the 1000 Genomes project [69], and two European genomes (NA07022 and NA20431) sequenced by Complete Genomics [110]. Since each individual's genome was sequenced to higher coverage, we treated the called genotypes as though they were the true genotypes throughout subsequent analyses.
For the Venter and Watson genomes, we downloaded SNP genotypes from the “Genome Variants” table of the UCSC browser. Coverage information across these two genomes was obtained from “emf” files from the Ensembl database. Sites with a score of 1 or greater were considered covered. SNPs overlapping regions with a lower score as well as indels and other structural variants were dropped from the analysis. Sites that were covered by reads, but did not have a SNP genotype were considered to be homozygous for the reference genotype.
We downloaded the “.vcf” and “mask” files for the CEU trio of the 1000 Genomes Project. Genotypes for variable positions were obtained from the .vcf files. For the rest of the genome, the individuals were assumed to be homozygous for the reference allele if SNP calling was attempted at the position (i.e. the position had a score of “0” in the mask file). A small number of reported SNPs in the .vcf files that fell in masked positions of the genome were removed from subsequent analyses.
Coverage and SNP genotype information could be directly obtained from the Complete Genomics “variations” files. SNPs and positions that were within 2 bp of indels or structural variants were removed from subsequent analyses.
We intersected the variant genotypes and coverage information from all six genomes and called genotypes for each individual. SNPs with more than two different alleles across all individuals or SNPs where one of the two alleles did not match the reference sequence were removed from subsequent analyses. For sites where one individual had a variant genotype, the genotypes for the other individuals who did not have a variant allele were considered to be homozygous for the reference if they had coverage at that particular site, or were considered to be missing if they did not have any coverage. Subsequent analyses of diversity levels and MAF only used those SNPs and sites that were covered in all six individuals.
We also analyzed six European genomes sequenced by Complete Genomics (CGS). Five of the genomes were from the CEU sample (NA06985, NA06994, NA07357, NA10851, and NA12004) and one was from a TSI individual (NA20502). We used the genotype calls made by CGS that were found in the “masterVarBeta” files. SNPs with more than two different alleles across all individuals, SNPs where one of the two alleles did not match the reference sequence, and sites that were within 2 bp of structural variants called in any one of the individuals were removed from subsequent analyses. Later analyses of diversity levels and MAF only used those SNPs and sites that were covered in all six individuals.
We noted that some windows of the genome appeared to have an unusually high number of SNPs where many individuals were heterozygous (Figure S9). We removed windows which had at least 10 SNPs where the average number of heterozygous genotypes per SNP was greater than 3 (out of 6). This filtering resulted in dropping 3.8% of the windows and appeared to remove the outlier regions (Figure S9).
We divided the genome into non-overlapping 100 kb windows. Windows that were within 10 Mb of an annotated centromere, telomere, or end of a chromosome were omitted from further analyses. For each window, we tabulated several genomic features. First, we obtained the recombination rate for each window using the high-resolution pedigree-based genetic map assembled by deCODE [80]. Second, we tabulated the number of sites within each window where the hg18 base differed from the pantro2 base. This was done using the .axt alignments obtained from the UCSC browser. Importantly, bases in RepeatMasked parts of the genome or where the hg18 or pantro2 alleles were missing were not counted. Since we wanted to examine putatively neutral sites, bases falling in the 17-way phastCons regions were also not counted [81], except when analyzing synonymous and nonsynonymous human-chimp divergence (see below). Third, we tabulated GC content within each window as the fraction of bases where the hg18 sequence was a G or a C. Only those bases that met the inclusion criteria described above were counted in this analysis. Fourth, as a measure of genic content, we tabulated the proportion of bases within each window that overlapped with a RefSeq transcript. We then tabulated the number of SNPs within each window and the number of bases that had sequencing coverage (see above for the criteria used to define covered bases). Importantly, SNPs falling RepeatMasked regions or phastCons regions were dropped from the analysis. Similarly, these bases were not counted as covered bases. The number of SNPs per covered base was used as a summary of diversity within each window. Finally, we summarized the frequency spectrum within each window by the average MAF over all the SNPs within each window.
We tested for correlations between the variables described above using non-parametric correlation tests. Specifically, we tested for pairwise correlations between variables using Spearman's . Since many of the variables were correlated with each other (Table S1), we calculated partial correlations to remove the effects of confounding variables on the variables of interest. Partial correlation statistics were calculated using the pcor function in R [111].
We tested whether the correlations were stronger in genic windows compared to non-genic windows using a permutation test. For each permutation, windows were randomly assigned to a genic and a non-genic group, keeping the number of genic and non-genic windows equal to that in the observed data. We recorded the difference in the correlation coefficient between each permuted genic and permuted non-genic dataset. The P-value for the test was the proportion of 10,000 permuted datasets with differences larger than those seen in the non-permuted data.
To test for a correlation between neutral polymorphism (Snorm) and nonsynonymous divergence, we found the number of nonsynonymous hg18-pantro2 alignment differences in each window (DN). This was done by putting those alignment differences that were not in RepeatMasked sequence and overlapped with an exon in the Consensus Coding Sequence (CCDS) table from the UCSC Table Browser into the SeattleSeq SNP annotation pipeline (http://gvs.gs.washington.edu/SeattleSeqAnnotation/). The human and the chimp bases were used as the two alleles. If multiple CCDS genes overlapped, we selected the longest one and discarded the remainder. We used the Nei-Gobjori [112] approach with the CCDS gene model to count the number of synonymous (LS) and nonsynonymous (LN) sites per window. LN and LS were only counted from the hg18 sequence, rather than averaged between the hg18 and pantro2 sequences. Only those sites that were not Repeat-Masked and were aligned with pantro2 were counted. The number of nonsynonymous differences per nonsynonymous site (dN) was then calculated as DN/LN. Similarly, the number of synonymous differences per synonymous site (dS) was then calculated as DS/LS.
To determine which models of selection could generate the correlations we observed in the resequencing data, we performed forward-in-time population genetic simulations using the program SFS_CODE [113]. Specifically, we simulated 100 kb regions that included exons and introns separated by an intergenic spacer region (Figure S5). We assumed a Jukes-Cantor mutation model [114] with a per-base pair mutation rate of 2.5×10−8.
Figure S10 shows the demographic model used for the simulations. Briefly, we simulated a human population with a chimp outgroup where the chimp population split from the human population 5 million years ago (assuming 25 years per generation). The ancestral human-chimp population was assumed to be of size 20,000 because previous studies have found that the ancestral human-chimp population was likely 2–10-fold larger than the current human effective population size [115]–[119]. At the human-chimp speciation event, the both the chimp and human populations underwent an instantaneous 2-fold contraction to their current sizes. Since our data consisted of European individuals, we also included a bottleneck in the human population with parameters from Lohmueller et al. [120], but using an ancestral population size of 10,000 between the human-chimp split and the more recent bottleneck.
The recombination rates for the simulated windows were chosen to approximately match the distribution of estimated recombination rates of the genic windows from the low-coverage dataset. This was done by assigning each window in the low-coverage data to one of 100 different bins based on its recombination rate. A single recombination rate was chosen for each bin (the mid-point of the bin), and this rate was used to simulate the number of replicates proportional to the number of windows in the actual data falling into the bin. A total of 20,000 simulated windows were generated for each model of selection.
Recombination hotspots were added to each window. Hotspots were assumed to have a width of 2 kb and the inter-hotspot distances were drawn from an exponential distribution with a mean of 20 kb. To specify the intensities of the hotspots in SFS_CODE, one needs to provide the proportion of the total amount of recombination that occurred within each of the hotspots and coldspots. We set the proportion of recombination that occurred in hotspot i to be 0.8xi, where xi was drawn from a Dirichlet distribution with parameter k equal to the number of hotspots within the window, and . This framework allowed hotspots to have different intensities and kept the total proportion of recombination that occurred in hotspots in each window at 80% [121]. A similar approach was used to determine the background recombination rates for each part of the sequence outside of the hotspots, except 0.2 was used instead of 0.8.
We examined several different models of natural selection (Table S6). In most models, nonsynonymous mutations were weakly deleterious with their selection coefficients drawn from a gamma distribution of selective effects, the parameters of which had been estimated from human resequencing data [22]. Some models also included positive selection acting on a fraction of nonsynonymous mutations, or a fraction of intronic mutations that were weakly deleterious.
We then tabulated diversity and divergence summary statistics from the simulations. Importantly, we only analyzed SNPs and human-chimp differences that occurred in the neutral intergenic sequence. For comparison to the low-coverage Danish data, we used the population MAFs from the simulations, counting only those SNPs with MAF >5% as we did in the observed data. From these same simulations, we took a sample of six individuals to analyze and compare to the higher-coverage data. The strength of some correlations may depend on how precisely diversity statistics could be estimated, and these estimates likely depend on the amount of sequence analyzed within each window (or, in other words, the fraction of bases within the 100 kb window that were covered). Therefore, we sampled the amount of intergenic sequence to be analyzed in each simulated window from the empirical distribution of the number of bases covered in each window. This was done separately for the low and higher-coverage datasets because the number of bases covered differed between the two datasets. To compute the number of human-chimp differences from the simulations, we compared the sequence of a single chimp individual to a single human individual. Sites where the two individuals were homozygous for different alleles were counted as differences. Sites where both were homozygous for the same allele were not counted as differences. All other sites (e.g. chimp was heterozygous and human was homozygous, chimp was heterozygous and human was heterozygous, chimp was homozygous and human was heterozygous) were counted as half a difference.
For computational efficiency, we simulated an ancestral population of 500 individuals while keeping the population-scaled mutation and recombination rates and selection coefficients equal to their original values. This approach increased computational efficiency, but should result in the same patterns of variation as larger population sizes since the patterns of variation depend only on the scaled population parameters.
|
10.1371/journal.pgen.1000458 | The RNA Polymerase Dictates ORF1 Requirement and Timing of LINE and SINE Retrotransposition | Mobile elements comprise close to one half of the mass of the human genome. Only LINE-1 (L1), an autonomous non-Long Terminal Repeat (LTR) retrotransposon, and its non-autonomous partners—such as the retropseudogenes, SVA, and the SINE, Alu—are currently active human retroelements. Experimental evidence shows that Alu retrotransposition depends on L1 ORF2 protein, which has led to the presumption that LINEs and SINEs share the same basic insertional mechanism. Our data demonstrate clear differences in the time required to generate insertions between marked Alu and L1 elements. In our tissue culture system, the process of L1 insertion requires close to 48 hours. In contrast to the RNA pol II-driven L1, we find that pol III transcribed elements (Alu, the rodent SINE B2, and the 7SL, U6 and hY sequences) can generate inserts within 24 hours or less. Our analyses demonstrate that the observed retrotransposition timing does not dictate insertion rate and is independent of the type of reporter cassette utilized. The additional time requirement by L1 cannot be directly attributed to differences in transcription, transcript length, splicing processes, ORF2 protein production, or the ability of functional ORF2p to reach the nucleus. However, the insertion rate of a marked Alu transcript drastically drops when driven by an RNA pol II promoter (CMV) and the retrotransposition timing parallels that of L1. Furthermore, the “pol II Alu transcript” behaves like the processed pseudogenes in our retrotransposition assay, requiring supplementation with L1 ORF1p in addition to ORF2p. We postulate that the observed differences in retrotransposition kinetics of these elements are dictated by the type of RNA polymerase generating the transcript. We present a model that highlights the critical differences of LINE and SINE transcripts that likely define their retrotransposition timing.
| SINE retroelement amplification has been extremely successful in the human genome. Although these non-autonomous elements parasitize factors from LINEs, both the human Alu and the cumulative rodent SINEs have generated over one million copies in their respective hosts. Alu-induced mutagenesis is responsible for the majority of the documented instances of human retroelement insertion-induced disease. Our data indicate that SINEs require a shorter period of time to complete insertion than L1s, possibly contributing to the ability of Alu elements to effectively parasitize L1 components. We demonstrate that RNA polymerase changes the timing Alu requires to complete retrotransposition and creates the need for the L1 ORF1protein in addition to ORF2p. We postulate that the way cells manage pol III and pol II (mRNA) transcripts affects the timing of a transcript going through the retrotransposition pathway. We propose a model that highlights some of the critical differences of LINE and SINE transcripts that likely play a crucial role in their retrotransposition process.
| Mobile elements have constantly assaulted genomes, shaping and molding their structure and organization. In particular, mobile elements have flourished in mammals generating between 40–50% of their genomic sequence [1]–[3]. About one third of the human genome can be attributed directly or indirectly to the activity of the non-LTR retroelements also referred to as LINEs (Long INterspersed Elements). LINE-1 (L1) and its non-autonomous partners Alu, SVA, and retropseudogenes continue to amplify in the human genome. L1 and the SINE (Short INterspersed Element), Alu, are by far the most numerous, adding up to 1.5 million copies [1]. Although Alu mobilization depends on L1 proteins [4], they outnumber L1 inserts by 2 to 1. Similarly, the sum of the total copies of all rodent SINEs outnumber L1 copies about 2 to 1 [2],[3]. Alu and the rodent SINE inserts have been more successful than other non-autonomous retroelements, such as the retropseudogenes [5]. Size and sequence composition differences between SINEs and LINEs may allow the mammalian genome to better tolerate SINE insertions, reviewed in [6]. Negative selection has clearly played a role in reducing L1 copy number through ectopic recombination and elimination of many full length and nearly full length L1 inserts [7]. However, processes other than negative selection must influence the observed differences. The updated reports of diseases caused by de novo inserts (where little, or no, selection has occurred) show that Alu inserts outnumber those of L1 by about 2 to 1 [6],[8].
Tissue culture assay systems indicate that L1 retrotransposition rates are consistently higher than those observed for SINEs [4],[9]. This is possibly a reflection of the strong cis-preference contained by L1 [10],[11], while Alu must compete for L1 proteins in trans. How is it that Alu with a lower retrotransposition rate than L1, contributes more de novo disease cases? It is likely that multiple factors are involved, such as the ability to bind SRP9/14 [12],[13].
Retroelements are mobile elements that amplify through an RNA intermediate in a process known as retrotransposition [14]. There are limited data on the details of the mechanism of LINE retrotransposition, and even less for SINE retrotransposition. The process begins with the generation of RNA (Figure 1A). Active L1 elements express two proteins from a bicistronic mRNA: ORF1p[15] and ORF2p (Figure 1B and C). Both L1 proteins are needed for L1 retrotransposition [16]. In contrast to L1, ORF2p expression is sufficient for SINE retrotransposition [4],[9],[17], while ORF1p may enhance the process [17]. ORF1p possesses nucleic acid chaperone activity [18],[19], an essential property for L1 retrotransposition [19],[20]. ORF2p is a multifunctional protein with endonuclease and reverse transcriptase activities [21],[22]. Both proteins are proposed to interact in cis [10],[11] with the L1 RNA to form a cytoplasmic RNP complex interacting with polyribosomes [20],[23]. SINE RNA is predominantly found in the cytoplasm as an RNP complex [12],[24],[25] (Figure 1C) and uses L1 protein(s) in trans for its mobilization. The endonuclease of the L1 ORF2p generates the first nick within the L1 endonuclease recognition sequence generating single stranded DNA that primes the reverse transcription [22],[26]. Both L1 and Alu are proposed to undergo integration through a target-primed reverse transcription (TPRT) reaction [27].
To generate a new insertion, L1 and SINE elements must return to the nucleus either together or independently (Figure 1D). Reported data suggest that retrotransposition-competent L1 RNPs may transit through the nucleolus [28]. The 3′ poly-A stretch or “A-tail” of LINEs, SINEs and processed pseudogenes is required for the priming of reverse transcription (Figure 1E) [4],[29]. Unlike the post-transcriptionally generated A-tail of pol II RNAs (mRNA), SINE A-tails are included within their sequence and play an important role in SINE retrotransposition [30],[31]. The details of the final integration and ligation of the L1 or Alu inserts into the host DNA remain unclear. Recent reports indicate that cellular factors, such as DNA repair enzymes, may aid in the L1 retrotransposition process [32],[33]. The final inserted sequence is typically flanked by direct repeats (Figure 1G). Non-autonomous retrotransposed inserts, such as Alu, SVA, hYs and retropseudogenes share these hallmarks with L1 inserts, strongly suggesting that these elements use the L1 ORF2p endonuclease generated nick for their integration [34]–[36].
To date, all known SINEs are ancestrally derived from RNA pol III transcribed RNA genes, reviewed in [37]. The vast majority are derived from different tRNA genes and only two (Alu and the rodent B1) originated from the 7SL RNA gene, a component of the signal recognition particle (SRP) [38]. Other examples of pol III transcribed repeats include the four hY genes (hY1, hY3, hY4 and hY5) that likely contributed directly or indirectly to the generation of almost 1000 copies in the human genome by retrotransposition [36],[39]. In contrast to SINEs, an internal RNA pol II promoter drives LINE transcription with the unusual ability to start transcription upstream of its location. Like other pol II RNAs, L1 transcription is regulated by different mechanisms, including promoter methylation [40], transcriptional attenuation due to A-richness [41], premature polyadenylation [42], and the generation of different splice variants [6]. Additional studies suggest that at least some portion of the L1 mRNAs are capped [43] and that the capping enhances L1 translation [44].
Previously, an L1 element tagged with a green fluorescent protein (EGFP) retrotransposition cassette was used to detect L1 retrotransposition “near real time” [45]. The earliest detection of an L1 retrotransposition event was 48 h post-transfection. In this manuscript, we evaluate the timing of retrotransposition (defined as the time required for a retroelement from the initial transcription step to complete an insertion) of tagged Alu and L1. We demonstrate that Alu elements only require about half of the amount of time as L1 to generate an insert. Our data demonstrate that the type of RNA polymerase dictates the retrotransposition timing, but does not determine the retrotransposition rate (defined as the number of inserts a given element can generate, i.e. the “efficiency” of an element). After evaluating several potential time limiting steps, we show that the RNA polymerase type is an important early factor contributing to the divergent retrotransposition kinetics between LINEs and SINEs.
Reverse transcriptase (RT) domains of multiple sources can be grouped into a family of shared sequence homology [NCBI cdd pfam00078.12] [46], including the RT of the human immunodeficiency virus and L1 ORF2 protein. Endogenous RT activity is inhibited by two antiretroviral agents nevirapine and efavirenz [47]. L1 retrotransposition in a culture assay system can be suppressed by the addition of a variety of HIV RT inhibitors [48],[49]. This system utilizes a tagged vector designed to allow expression of the reporter gene only when the retroelement goes through its reverse transcriptase-dependent amplification process (Figure 2A). Therefore, only the newly inserted element will express the reporter gene (e.g. neo).
Using the established L1 and Alu retrotransposition tissue culture assays [4],[16], we evaluated the dose of, 2′,3′-didehydro-3′-deoxy-thymidine (d4t) required to abolish retrotransposition of L1 and L1 ORF2p driven Alu without adversely affecting cell growth and viability. Treatment of transiently transfected HeLa cells showed that both L1 and Alu activities presented a d4t activity inhibitory concentration 50 (IC50) of about 2 µM (Figure S1). For our subsequent experiments we utilized d4t treatments at 50 µM (25 fold the IC50) to inhibit SINE and LINE retrotransposition in tissue culture. We selected this dose for its efficient inhibition of retrotransposition and lack of observed negative effects, determined by colony formation of an unrelated plasmid that expresses a functional neomycin resistance gene and integrates into genomic DNA by random integration rather than by an L1-dependent mechanism (data not shown).
We took advantage of the d4t inhibition to determine L1 and Alu retrotransposition kinetics in cultured cells. By treating cells with d4t at different time points after the transient transfection with the vectors expressing the tagged L1 or Alu plus ORF2p, we specifically inhibited the retrotransposition process at designated time periods (shown in Figure 2B). Any detected L1 or Alu inserts are presumed to have completed the insertion process prior to the addition of the d4t, as inhibition of ORF2p RT activity would prevent the generation of the cDNA. Using this approach, we show that L1 inserts are not detected in cultured cells during the first 24 h post-transfection (Figure 2C). Similar results were previously observed using a green fluorescent protein (EGFP)-tagged L1 element [45],[50]. The earliest detection of L1 inserts occurred at 32 h post-transfection (Figure S2). In contrast, we can easily detect Alu inserts 24 h and sometimes as early as 18 h post-transfection (Figure 2C).
Generation of an RNA transcript is an essential first step of the retrotransposition cycle (Figure 1A). Besides serving as a template for protein translation, L1 RNA acts as the insertion template during retrotransposition. Thus, transcriptional limitations or variations can directly impact retrotransposition of L1 elements as well as other retroelements. Previous studies demonstrate that L1 elements generate low amounts of full-length transcripts due to premature polyadenylation [42], transcriptional inefficiency due to A-richness [41], and multiple splicing events [6]. In all these studies, a decrease in the amount of L1 mRNA contributed to reduced retrotransposition and, conversely, the rate increased with higher amounts of full-length L1 RNA [42],[51],[52]. To determine whether L1 RNA transcription and processing contributes to the observed timing difference between L1 and Alu inserts, we performed a time course to evaluate the generation of the spliced RNA product in cells transiently transfected with L1mneo, AluYa5neoTET, and L1neoTET (Figure 3). Because the Alu construct is driven by RNA polymerase III, its tag (neoTET) contains a self splicing intron disrupting the neomycin gene. Therefore, we included an additional L1 construct that contains the exact same self splicing (neoTET) tag present in the Alu vector to control for any potential variations introduced by splicing dynamics. Full-length spliced and unspliced transcripts from Alu and both L1 constructs could be detected as early as 3 hours post-transfection (northern blots shown in Figure S3). The mneo and neoTET tagged L1 constructs exhibited similar kinetics for the spliced transcript (only RNA that will generate G418R colonies when retrotransposed), peaking by 24 h and declining by 72 hours (Figure 3). Splicing efficiency of the RNA produced by different expression vectors was evaluated (Table S1). Equivalent splicing efficiency was observed for the L1 and Alu transcripts sharing the same neomycin cassette (neoTET or mneo). Alu-tag transcripts were only detected in the cytoplasmic fraction at any of the time points evaluated (data not shown), consistent with what has been previously reported for the authentic Alu “untagged” RNA [53]. Despite early L1 mRNA availability, no L1 inserts were observed at the 24 h time point. Spliced Alu transcripts peak around 48 h, declining by 72 h, much like L1 mRNA (Figure 3). However, in contrast to L1, numerous Alu inserts are readily detectable by 24 h. These results demonstrate that the full-length properly spliced L1 RNA is generated in the same time period as the Alu RNA. Thus, it is unlikely that RNA transcription or variation in the type of splicing within the neo cassette account for the observed time difference between the generation of Alu and L1 inserts.
Another difference between the Alu and L1 elements involves the length of the transcript, which could alter the time required by the reverse transcriptase to generate a full-length cDNA. In this assay system full length inserts are not required to generate a G418R colony. In both Alu and L1 assays, inserts are detected with the retrotransposition of the minimal unit of a functional neomycin gene, which is identical in length in both transcripts once the intron is removed. Therefore, the timing differences observed between these two elements should be independent of the transcript length.
We next assessed whether the delay reflects the time required for translation of the L1 proteins and the ability to reach the nucleus (Figure 1B–E). ORF2 protein has been notoriously difficult to observe by conventional techniques, such as western blot analysis [28]. As an alternative, the ORF2p activities can be evaluated.
Because Alu elements require ORF2p for retrotransposition, evaluation of Alu retrotransposition serves as an alternate method to detect ORF2p activity. Therefore, we exploited the trans-complementation assay to monitor the ability of L1 to trans-mobilize Alu, using AluYa5neoTET as a reporter construct. We determined the Alu insertion kinetics in cells cotransfected with the AluYa5neoTET plus the L1 no tag vector. Multiple Alu inserts were detected as early as 24 h post-transfection (Figure 4), corroborating the availability of the ORF2p expressed from the L1 vector in the nucleus by 24 h. Equivalent results were observed when using a blasticidin tagged L1 to drive Alu retrotransposition (data not shown). Under our experimental conditions, endogenous L1 present in HeLa cells does not significantly contribute to the generation of the G418R colonies as the Alu vector was unable to generate any inserts without L1 supplementation at 24, 32 and 48 h post-transfection (vector control, Figure 4). A few solitary colonies (2 and 1) were observed at the 42 and 72 h time points. This observation clearly demonstrates that a full-length L1 vector generates enough ORF2p to reach the nucleus within 24 h and to mobilize a tagged Alu element in our assay system. Our observations are in agreement with previously published data demonstrating that cells transiently transfected with L1 exhibit extensive double strand breaks at 24 h post-transfection [33]. The observed DNA breaks are dependent on the endonuclease activity of the L1 ORF2p. Our data strongly suggest that translation and nuclear localization of ORF2p is unlikely to be the main limiting step for the observed differences between the L1 and Alu time requirements.
In addition, pre-transfection of high amounts of ORF2p or any of the L1 factors (proteins and/or RNPs) in trans did not alter L1 retrotransposition timing (Figure S4). This is not surprising considering that L1 RNA exhibits a strong cis-preference for its own translated proteins for retrotransposition [10],[11]. Pre-transfection with ORF2p showed a few more Alu inserts at early time points (data not shown). However, this slight increase was not statistically significant (Student's paired t-test, p = 0.297).
Transcripts generated from RNA polymerase II and III promoters differ in their capping, 3′ end processing, folding structures, post-transcriptional processing, interaction with translation factors and degradation pathways, reviewed in [54]–[56]. In addition, these two transcriptional complexes can be observed in different spatial locations in the nucleus indicating discrete transcriptional sites [57],[58]. To evaluate the timing of retrotransposition of other pol III-driven genes we generated “tagged” versions of 6 human genes (7SL, U6, hY1, hY3, hY4 and hY5) by cloning the genes with at least 300 bp of their upstream enhancer sequence 5′ of the neoTET cassette (details in materials and methods). Although the “functional” genes are not SINEs per se, we selected these as examples of pol III-driven genes. The human genome contains multiple examples of retrotransposed copies with sequence homology to these genes [36],[59]. Thus, these serve as our best examples of other human pol III-driven constructs. We also included in our analysis the pol III-driven B2 element as a known active rodent SINE [9],[60]. In our d4t-assay system, all tagged pol III-driven elements generated inserts by 24 h post-transfection when supplemented with just L1 ORF2p (Figure 5).
To better understand the RNA polymerase influence on retrotransposition, we also evaluated the time requirement of two pol II-driven (CMV) constructs: ORF1mneo and pol II Alu (Figure 6A). We selected ORF1mneo because it generates a transcript of L1 ORF1, which has previously been used to reflect retropseudogene activity [10]. The ORF1mneo vector can retrotranspose when a source of ORF2p is supplied in trans [10]. The pol II Alu (pCMVYa5mneo) contains an Alu tagged with the “mneo” cassette from the L1-tagged construct [61], which contains pol III terminators (4 Ts) that would generate truncated transcripts if the internal pol III A and B boxes in the Alu sequence are used for transcription. The “normal A-tail” at the end of the Alu sequence and 5′ of the neo cassette (Figure 6A) was not included in order to prevent potential internal priming for TPRT in the cDNA extension step (Figure 1E), which would circumvent inclusion of the neo reporter gene in the retrotransposed copy. Thus, only the Alu body sequence was utilized in the construct. Just like the L1 construct, the A-tail used in the TPRT step is generated from the transcript polyadenylation by the RNA polymerase II from the SV40pA signal at the 3′ end of the neo cassette (Figure 6A). Spliced and unspliced transcripts were detected from both constructs by 24 h (Figure 6B). The tagged ORF1p transcript driven by an ORF2p generated one single insert at 24 hours (Figure 6C), while the total number of colonies generated were 136 and 226 for 48 h and 72 h respectively. It is possible that the endogenous L1 expression in HeLa cells [6] affected the timing. However, our data on Alu retrotransposition indicates that effects from endogenous L1 expression under our experimental conditions are negligible (Figure 4). Most likely, the single G418R colony observed at 24 hours is due to a rare event that escaped d4t inhibition. A quantitative time course evaluation of the spliced RNA product in cells transiently transfected with ORF1mneo and AluYa5mneo further indicates that the availability of spliced product is not limiting retrotransposition timing (Figure 6E).
No pol II-generated Alu inserts were ever observed when supplemented with ORF2p under any conditions tested, representing a rate of less than 1×106 cells/µg of plasmid. However, retrotransposition of the pol II-driven Alu transcript occurred when it was supplemented with both ORF2p and ORF1p expression plasmid (Figure 6D). Under these conditions, G418R colonies were observed at 48 h post-transfection, much like L1 and retropseudogene behavior. No colonies were ever observed at the 24 h time point in 5 independent experiments using triplicates for each time point. Swapping the RNA pol III for an RNA pol II promoter changed the retrotransposition requirements of the tagged Alu to reflect those observed for pseudogenes and LINEs.
Recent data demonstrate that one amino acid substitution in the mouse L1 ORF1 protein dramatically affects retrotransposition rate and the ability to detect new inserts earlier [50]. We evaluated the insertion timing of the most efficient L1 available at the time, the synthetic mouse L1 (L1m syn) previously reported to increase retrotransposition efficiency by more than 200 fold relative to the wildtype L1spa element [52]. Despite having a much higher retrotransposition rate, L1m syn required 48 h to generate inserts even when spliced RNA could be detected as early as 3 hours post-transfection (Figure 7). There were a few (1 to 2) colonies at 24 hours or earlier but these are likely outlier observations as they only represent 0.001 of the total observed G418R colonies. Our data are consistent with the observation that all of the evaluated pol II-driven constructs require 48 h, while all of the pol III-driven constructs generate inserts by 24 h despite their very low retrotransposition rates relative to L1 (Table 1). Because of the large variation in retrotransposition rates, we opted to show the relative number of inserts in the figures for each construct by designating the 48 or 72 hour time point as 100%. While both U6 and Alu tagged transcripts, for example, can generate inserts by 24 hours, their retrotransposition rates (i.e., the actual number of observed inserts) differ dramatically. The same is true for the tagged L1 and ORF1 RNAs.
Throughout mammalian evolution different mobile elements have flourished within genomes. Retroelements such as LINEs and SINEs have been particularly successful, generating more than one third of human sequence mass. Interestingly, the parasitic non-autonomous SINE elements outnumber their autonomous LINE partners in the primate and rodent genomes. The success of SINEs is especially evident when compared to the copy numbers of other non-autonomous elements such as the retropseudogenes.
Our data reveal differences between retropseudogenes, Alu, and L1 retrotransposition. When evaluating Alu and L1 retrotransposition kinetics, the tagged Alu transcript required less time to generate an insert. This timing difference can not be attributed to differences in the time required to generate functional transcripts or availability of L1 proteins. It is clear that full-length functional L1 transcripts can be detected as early as 3 hours post-transfection and are abundant by 24 h post-transfection. In addition, the difference observed between Alu and L1 kinetics could not be attributed to the type of detection cassette system (self splicing or not) or to the differences in the retrotransposition rates. L1 colonies were rarely observed (Figure 7) at time points earlier than 48 h. These few observed G418R colonies possibly represent the rare event that circumvented inhibition by d4t (in one experiment a colony was observed even at the zero time point). In our assay, production of L1 ORF2p is not limiting. Our data demonstrate that enough ORF2p is generated from an L1 construct to drive Alu insertions within 24 hours post-transfection, which indicates that ORF2p is made and readily available for Alu transcript mobilization. However, at this time we do not know if the ORF2p reaches the nucleus as a “free” protein or as part of an RNP with the L1 RNA or Alu RNA. As expected, due to the L1 cis-preference [10], pre-transfections with ORF1p, ORF2p or other L1 components, such as full-length transcripts or RNPs, did not affect the L1 time requirement.
Although unexpected, it is not totally surprising that Alu and L1 present different retrotransposition time requirements. Previous data show that, although Alu and L1 share the same insertion hallmarks, the two elements can exhibit differences in their behavior. For example, of two HeLa “cell lines,” only one supports Alu retrotransposition while both support L1 retrotransposition [62]. In addition, Alu and L1 are selectively inhibited by different APOBEC3 proteins [62]. This corroborates our observations that cellular components differentiate between Alu and L1 retrotransposition.
Our data suggest that the observed time differences are dependent on the type of RNA polymerase generating the transcript. Multiple features that distinguish these two transcript types may collectively or individually contribute to the observed differences in the retrotransposition timing between L1 and Alu elements. RNA capping, association with the translational machinery and ORF1 requirement are plausible factors that could influence SINE and LINE retrotransposition kinetics. As a pol II product, L1 mRNA is likely capped. Experimental evidence indicates that at least part of the L1 mRNA is capped [43] and that capping enhances L1 translation in vitro [44]. In contrast, pol III genes lack the 7-methylguanosine cap and are subjected to different processing in a spatially separate location of the nucleus [57],[58]. L1 mRNA likely interacts with most, if not all, of the pol II protein complexes that assemble with the transcription of generic mRNAs, as evidenced by the premature polyadenylation and splicing of L1 transcripts [6],[42].
Even though both pol II and pol III produced RNAs form complexes with various cellular proteins, the structure and composition of these RNPs varies dramatically. As a rule, pol III transcripts do not code for proteins and therefore interact with the translational machinery in a different manner than mRNA. Most known pol III transcripts fold to form a structured RNA and associate with a variety of proteins to form RNPs. Specifically, Alu interaction with SRP9 and SRP14 [12] is thought to transiently provide proximity to the ribosomal complexes and translating L1 RNA, allowing the Alu transcript to efficiently compete for the L1 factors required for retrotransposition [26]. It is also likely that the ability of the dimeric Alu to bind these proteins contributes to the dramatic difference in retrotransposition rates observed between Alu and other SINEs [9],[13]. In contrast, the polyribosomes and translation machinery assemble with the L1 mRNA in a more stable complex to undergo translation. The cis-preference displayed by L1 [20] suggests that the L1 RNA must dissociate from the cellular translation machinery to form L1 RNPs as an intermediate step in the retrotranspositional process. These L1 complexes are composed of L1 RNA, ORF1p [20], and likely ORF2 protein [11]. All three components are shown to co-purify in the polyribosomal fraction of the cytoplasm [11],[23]. It is plausible that ORF1p directly competes with the cellular translation machinery for access to L1 mRNAs, transitioning the L1 transcript away from the polyribosomal fraction and into the retrotranspositionally competent RNPs. Because of their nature and subcellular localization, SINEs completely avoid these two potentially time consuming steps in their mobilization. Therefore, SINE transcripts may enter their retrotransposition cycle as soon as L1 ORF2p becomes available.
The pol II-driven Alu transcripts that are most likely to associate with the cellular translational machinery, at least transiently, require L1 ORF1 protein in addition to ORF2 protein for retrotransposition in a manner reminiscent of retropseudogenes [29]. The retrotransposition time of the pol II-driven Alu parallels that of L1. At this stage it is unclear what the role of ORF1p is in the trans-mobilization of retropseudogenes or the pol II Alu transcript. However, it is consistent with the above-discussed hypothesis implicating ORF1 protein in removing pol II RNAs from their expected cycle of translation and degradation. Thus, the pol II L1 and the pol III Alu transcript interactions with different cellular components may dictate the timing difference between L1 and Alu RNAs to form their respective retrotranspositionally competent complexes.
The inefficient retrotransposition rate of the pol II-driven Alu construct suggests that the presence of an Alu sequence within an mRNA would not facilitate its retrotransposition by L1 factors. Although there is no available data on the SVA promoter, it is unlikely that the pol III polymerase drives SVA transcription due to the presence of numerous pol III terminators within its sequence. Thus, it is questionable whether the truncated antisense Alu-like sequences present in the SVA element contribute to the L1 trans-complementation of this retroposon as previously suggested [35].
In addition to assisting its own retrotransposition, the cis-preference exhibited by L1 may decrease cell damage by limiting random retrotransposition of cellular mRNA. A previous study demonstrated the co-localization of ORF1p and cellular proteins to stress granules[63]. The authors suggest that the sequestering of ORF1 protein in stress granules for degradation may prevent promiscuous binding of ORF1p to non-L1 mRNAs. Thus, as a side effect of L1 self-preference, retropseudogene formation is less likely [5]. In addition, this “cis-preference” could help the L1 transcript “escape” the ribosomal complex and degradation pathways. Once translation is completed, most transcripts decay by several known mRNA degradation pathways, reviewed in [56]. In contrast, pol III transcripts are meant to perform their function as RNA molecules in the cytoplasm or nucleus before degradation by the exosome [64]. Essentially, the functional molecule of pol III genes is the RNA, while for pol II genes the mRNA is an intermediary prior to the generation of the functional protein. In the case of L1, the ORF1p may play an additional role by protecting the L1 RNA from degradation, increasing the chance of returning to the nucleus where the involvement of ORF1p in the L1 integration process has been previously suggested [18],[23]. Thus, the requirement for both ORF1 and ORF2 proteins could contribute to the longer time needed for L1 transcripts to generate inserts. In addition, it is plausible that interactions with different cellular components during insertion, mediated by ORF1p, may contribute to the timing differences observed.
We postulate that the differences observed in retrotransposition kinetics are dictated by the type of RNA polymerase generating the transcript. We propose an initial model where the cytoplasmic interactions of pol II (L1 and mRNA) and pol III transcripts and pathways influence the amplification kinetics of LINEs and SINEs (Figure 8). Overall, it is evident that the type of RNA polymerase generating the transcript alters the timing of mobile element insertion and remains a critical parameter in the classification of different types of retroelements.
The basic transient L1 [16] or Alu [4] retrotransposition assay was performed as previously described with some minor modifications. Briefly, HeLa cells (ATCC CCL2) were seeded in T-75 flasks at a density of 5×105 cells/flask or in 6 well plates at a density of 2.5–5×104/well. Transient transfections were performed the next day with Lipofectamine Plus following the manufacturer's protocol (Invitrogen), with 3 µg of SINE-neoTET vector plus 1 µg pBud-ORF2opt or 1 µg of L1 no tag. For L1 assays 1 µg of JM101/L1.3 was used. Inhibitory effects on cellular growth or colony formation capabilities by the d4t treatment was evaluated by transfecting cells in parallel with 0.3 µg of a plasmid expressing neomycin resistance (pIRES2-EGFP; BD Biosciences Clontech) as a “toxicity” control. Following removal of transfection cocktail, the cells were treated with the appropriate media containing 400 µg/ml Geneticin/G418 (Fisher Scientific) alone or in combination with 50 µM d4t for selection and/or reverse transcriptase inhibition. After 14 days, cells were fixed and stained for 30 minutes with crystal violet (0.2% crystal violet in 5% acetic acid and 2.5% isopropanol). The inhibitor d4t- (2′,3′-Didehydro-3′-deoxy-thymidine; Sigma-Aldrich) was freshly added to the selection media at the indicated time period. During the inhibitor treatment period all cells in the experiment were refreshed daily for the first week with the appropriate media. The rate of insertion efficiency (retrotransposition rate) was determined as the number of visible G418R-resistant colonies obtained at 72 h after transient transfection of 1×106 seeded HeLa cells with 1 µg of the neo tagged construct.
RNA extraction and poly(A) selection was performed as previously described [42]. Total RNA was extracted using the recommended protocol for TRIzol Reagent (Invitrogen) from two 75 cm2 cell culture flasks at 3, 6, 24, 48, and 72 hours post-transfection. The PolyATract mRNA isolation system III (Promega) was used to select polyadenylated RNA species following the manufacturer's protocol. After separation in a 1% (L1) or a 2% (pol III constructs) agarose-formaldehyde gel, the RNA was transferred to a Hybond-N nylon membrane (Amersham Biosciences). The RNA was cross-linked to the membrane using a UV-light (GS Gene linker, BioRad) and pre-hybridized in 30% formamide, 1× Denhardt's solution, 1% SDS, 1 M NaCl, 100 µg/ml salmon sperm DNA, 100 µg/ml-1 yeast t-RNA at 60°C for at least 6 h. The 3′ region of the neomycin gene was amplified by PCR using the following primers T7neo (−): 5′-TAATACGACTCACTATAAGGACGAGGCAGCG-3′ and Neo northern (+): 5″- GAAGAACTCGTCAAGAAGG-3′. The isolated PCR product was used as a DNA template to generate a 32P-CTP (Amersham Biosciences) labeled single strand-specific RNA probe using the MAXIscript T7 kit (Ambion) following the manufacturer's recommended protocol. We utilized material included in the kit to generate the riboprobe for the β-actin. The radiolabeled probe was purified by filtration through a NucAway Spin column (Ambion). Hybridization with the probe (final concentration of 4–12×106 cpm/ml) was carried out overnight in the pre-hybridization solution at 60°C. Two ten-minute washes were performed at high stringency (0.1×SSC, 0.1%SDS) at 60°C. The results of the northern blot assays were evaluated using a Typhoon Phosphorimager (Amersham Biosciences) and the ImageQuant software.
|
10.1371/journal.pcbi.1005566 | Chemomechanical regulation of myosin Ic cross-bridges: Deducing the elastic properties of an ensemble from single-molecule mechanisms | Myosin Ic is thought to be the principal constituent of the motor that adjusts mechanical responsiveness during adaptation to prolonged stimuli by hair cells, the sensory receptors of the inner ear. In this context myosin molecules operate neither as filaments, as occurs in muscles, nor as single or few molecules, as characterizes intracellular transport. Instead, myosin Ic molecules occur in a complex cluster in which they may exhibit cooperative properties. To better understand the motor’s remarkable function, we introduce a theoretical description of myosin Ic’s chemomechanical cycle based on experimental data from recent single-molecule studies. The cycle consists of distinct chemical states that the myosin molecule stochastically occupies. We explicitly calculate the probabilities of the occupancy of these states and show their dependence on the external force, the availability of actin, and the nucleotide concentrations as required by thermodynamic constraints. This analysis highlights that the strong binding of myosin Ic to actin is dominated by the ADP state for small external forces and by the ATP state for large forces. Our approach shows how specific parameter values of the chemomechanical cycle for myosin Ic result in behaviors distinct from those of other members of the myosin family. Integrating this single-molecule cycle into a simplified ensemble description, we predict that the average number of bound myosin heads is regulated by the external force and nucleotide concentrations. The elastic properties of such an ensemble are determined by the average number of myosin cross-bridges. Changing the binding probabilities and myosin’s stiffness under a constant force results in a mechanical relaxation which is large enough to account for fast adaptation in hair cells.
| Myosin molecules are biological nanomachines that transduce chemical energy into mechanical work and thus produce directed motion in living cells. These molecules proceed through cyclic reactions in which they change their conformational states upon the binding and release of nucleotides while attaching to and detaching from filaments. The myosin family consists of many distinct members with diverse functions such as muscle contraction, cargo transport, cell migration, and sensory adaptation. How these functions emerge from the biophysical properties of the individual molecules is an open question. We present an approach that integrates recent findings from single-molecule experiments into a thermodynamically consistent description of myosin Ic and demonstrate how the specific parameter values of the cycle result in a distinct function. The free variables of our description are the chemical input and external force, both of which are experimentally accessible and define the cellular environment in which these proteins function. We use this description to predict the elastic properties of an ensemble of molecules and discuss the implications for myosin Ic’s function in the inner ear as a tension regulator mediating adaptation, a hallmark of biological sensory systems. In this situation myosin molecules cooperate in an intermediate regime, neither as a large ensemble as in muscle nor as a single or a few molecules as in intracellular transport.
| The myosin family includes at least 20 structurally and functionally distinct classes [1, 2]. Although they all exhibit a common chemomechanical cycle, myosin molecules have remarkably diverse functions-including intracellular transport, force production in muscles, and cellular migration-as well as important roles in sensory systems [3]. To understand the emergence of these different functions, it is necessary to characterize the biophysical details of the chemomechanical cycle for each myosin class.
Myosin molecules transduce chemical energy into mechanical energy through the hydrolysis of adenosine triphosphate (ATP). The hydrolysis reaction and the subsequent release of inorganic phosphate (Pi) and adenosine diphosphate (ADP) induce structural changes that result in a power stroke and generate forces. The biochemical reaction rates and the response to external forces determine the specific function of each myosin [3]. On the basis of their biochemical and mechanical properties, myosins have been classified into four groups: (i) fast movers, (ii) slow but efficient force holders, (iii) strain sensors, and (iv) gates [4]. Although single-molecule experiments and structural studies have vastly advanced our understanding of force-producing molecules, we still lack a consistent description that quantitatively relates cellular functions to the molecular details. One prominent case is myosin Ic, which has been identified as a component of the adaptation motor of the inner ear [5].
Hair cells in the inner ear transduce mechanical stimuli resulting from sound waves or accelerations into electrical signals. On the upper surface of each hair cell stands a hair bundle comprising dozens to hundred of actin-filled protrusions called stereocilia. Cadherin-based tip links connect the tip of each stereocilium to the side of the longest adjacent one. When a mechanical force deflects the bundle, the resultant shearing motion raises the tension in the tip links. This tension increases the open probability of transduction channels and allows ions to diffuse into the stereocilia, depolarizing the hair cell.
To retain sensitivity, a hair cell adapts to a prolonged stimulus by changing the tension in the tip links. This adaptation has a fast component lasting a millisecond or less and a slow component of a few tens of milliseconds, the molecular details of which remain uncertain. To explain slow adaptation, it has been proposed that an ensemble of myosin Ic molecules alternately step up or slide down the actin filaments inside the stereocilia to regulate the tension in the tip links. Sliding of myosin is triggered by a locally elevated Ca2+ concentration. This picture has been quantitatively supported by experimental studies on hair cells and complemented by mathematical descriptions [6–9]. Fast adaptation describes the rapid reclosure of transduction channels after abrupt stimulation of the hair bundle. This process is poorly understood and several possible explanations at a molecular level are debated [6, 10]. One promising mechanism is the release model, in which a component of the transduction apparatus becomes more flexible and abruptly releases some of the tension in the tip links, allowing the channels to close rapidly [11, 12]. Although myosin Ic has been implicated in both slow and fast adaptation and an ensemble of myosin Ic molecules is a good candidate for the element that releases [10], the precise role of myosin Ic in adaptation has yet to be elucidated.
The rapid response of the transduction channels to a displacement of the hair bundle suggests a direct mechanical activation through the transformation of the deflection into a force by a spring [6, 13]. This mechanism underlies the gating-spring hypothesis that is the prevailing explanation for mechanotransduction by hair cells. The elastic property of the gating spring is the most important parameter in setting the precise relation between the deflection of a hair bundle and the open probability of the ion channels. Despite numerous studies of the molecular components of the hair bundle and their biophysical properties, we remain uncertain of the identity of the gating spring [14–18]. Every molecule that lies in series with the tip link could in principle influence the elastic properties, including the ensemble of myosin Ic molecules. These molecules bind and unbind from actin filaments and thereby change the elasticity dynamically. In order to fully explain mechanotransduction by hair cells, it is important to understand how the dynamics of single myosin Ic molecules determines the elastic properties of an ensemble and how it is regulated.
Over the past few years, the biophysical properties of individual myosin Ic molecules have been characterized in optical traps, biochemical assays, and structural studies [19–24]. Like other myosin isoforms, myosin Ic displays catch-bond behavior, a prolonged attachment to an actin filament in response to increased external force [19, 25]. The force-sensitive step in myosin Ic’s cycle is the isomerization following ATP binding, however, and not ADP release as in other slow myosins [19, 20]. To understand how this behavior relates to the molecule’s physiological function, we introduce a consistent mathematical description of myosin Ic’s cross-bridge cycle.
After the introduction of the basic framework by Huxley and Huxley, cross-bridge models have been widely used to describe the dynamics of myosin motors [2, 26–35]. However, these models often assume irreversible transitions at fixed nucleotide concentrations that determine the input of chemical energy. In a seminal work, T. L. Hill showed how to couple a description of an enzymatic cycle to free-energy transduction in a thermodynamically consistent manner, an approach that has been applied to study muscle myosin [36–39]. We build our cross-bridge cycle for myosin Ic on these concepts and furthermore include the catch-bond behavior.
Our description allows a quantitative analysis of the differences between in vitro and in vivo conditions, of Ca2+ regulation, and of cooperativity between force-producing molecules. Here we introduce a thermodynamically consistent description of myosin Ic based on single-molecule data and focus on the responses to external force, to different nucleotide concentrations, and to the availability of actin. We use this description to predict the elastic properties of an ensemble of myosin molecules and highlight the potential implication for the release model of fast adaptation.
As a functional description of myosin Ic we introduce a chemomechanical cycle consisting of five states: one state in which myosin is unbound from actin and four actin-bound states. Because we primarily focus on the force-producing states, we consider only a single, effective unbound state that combines the actin-detached ADP⋅Pi and ATP states. Each of the actin-bound states is associated with the nucleotide occupancy of the binding pocket of the myosin head (Fig 1). Myosin Ic performs its main, 5.8 nm power stroke upon phosphate release; a smaller power stroke of 2 nm follows ADP release. To account for the work done by these power strokes, we include a force dependence of the associated transition rates. We consider an effectively one-dimensional description in which the force acts along the coordinate of the power stroke: a positive force is oriented in a direction opposite to the power stroke. The nucleotide-binding rates depend linearly on the nucleotide concentrations and the actin-binding rates increase linearly with the actin concentration.
By cycling through the five states, myosin performs work whose magnitude is bounded by the free-energy input associated with the nucleotide concentrations. We base our description on the free-energy transduction of enzymes and thus ensure thermodynamic consistency. To incorporate myosin Ic’s unique force sensitivity, we include a simple force dependence of the rate of unbinding from the filament of myosin in the ATP state. Under high forces, we expect myosin Ic to be trapped in the ATP state. Therefore we consider the ADP state (3), the nucleotide-free state (4), and the ATP state (5) as strongly bound. The remaining states are weakly bound or unbound (Fig 1).
Our description, which captures many of the characteristics of myosin Ic, incorporates as free variables the experimentally controllable quantities external force, nucleotide concentrations, and actin concentration. This approach allows us to obtain analytic expressions for quantities that have been measured in experiments, then to use that information to determine the unknown parameter values of the model. An overview of the parameters is given in Table 1. A mathematical description of the cross-bridge cycle and details of the estimation of parameter values are presented in the Methods section.
In a single-molecule experiment using an isometric optical clamp, the lifetime of the myosin Ic-actin bond was measured for different external forces and two sets of nucleotide concentrations [20]. Because a rapid transit into and out of the weakly bound state (2) could not be resolved experimentally, this bound lifetime must be interpreted as the average time tsb that myosin Ic spends in the strongly bound states. We determined an analytic expression for the unbinding rate t sb - 1 from the strongly bound states (Eq 53) as functions of force and nucleotide concentrations and fit this function simultaneously to two sets of experimental data acquired for distinct nucleotide concentrations. This unbinding rate is independent of the transition rate ω15 and of the actin concentration. Both quantities determine how often the molecule binds to the filament rather than how long it remains bound. From the average time that myosin Ic resides in the weakly bound states we estimate the binding rate ω15 for an actin concentration of 100 μM appropriate for the experiments. A detailed explanation for the fitting procedure is given in the Methods section.
Fits of the unbinding rate t sb - 1 from the strongly bound states describe the experimental data well, indicating that our description is able to capture the force sensitivity of myosin Ic (Fig 2a). Although none of the transition rates can account individually for the plateau around zero force, their combined effect in the cycle clearly displays such a behavior, which is characteristic of myosin Ic.
The numerical values obtained in this way for the transition rates ω21 ≃ 164 s−1 and ω 51 0 ≃ 314 s - 1 suggest that in the absence of force, state (2) and state (5) are both configurations from which the myosin head rapidly detaches. The force-distribution factors (δ) indicate that phosphate release is only weakly dependent on force (δ1 ≃ 0.12) and ADP release not at all (δ1 ≃ 0).
The concentrations of nucleotides in cells differ from those in single-molecule experiments. We can use our description to predict the behavior of myosin molecules for different nucleotide concentrations. Although in single-molecule experiments the phosphate concentration usually remains low, the phosphate concentration in vivo is on the order of 1 mM [2]. In cells the ATP concentration is also near 1 mM and the ADP concentration is around 10 μM [2]. In the remainder of this study we refer to these numbers as the physiological nucleotide concentrations. The unbinding rate does not significantly change for higher phosphate concentrations (Fig 2a and 2b). The main reason for this robust behavior is the very low rate constant for phosphate binding (Eq 38). Even for a millimolar phosphate concentration the phosphate-binding rate ω32 is very small compared to the other transition rates in the cycle. In contrast, increasing the ADP concentration decreases the overall binding rate because the molecule spends more time in the ADP state. This effect can be counteracted by an increase in the ATP concentration (Fig 2b).
Using the formulation given in the Methods section with the explicit solutions in Eqs 45–49, we can determine the steady-state probability distribution for the cross-bridge cycle at different nucleotide and actin concentrations (Fig 3). For physiological nucleotide concentrations and 100 μM of actin, myosin is trapped in the ATP state (5) under forces exceeding 2 pN (Fig 3a). Comparing only the strongly bound states, the molecule predominantly occupies the ADP state (3) for forces smaller than 1.5 pN. According to our description, myosin Ic’s cycle through the strongly bound states is limited by ADP release for forces smaller than 1.5 pN and by ATP release for forces larger than 1.5 pN. This result is consistent with experimental findings [19, 20].
In the stereocilium of a hair cell, myosin Ic is thought to extend between the crosslinked actin filaments of the cytoskeleton and the insertional plaque to which the tip link is anchored [5, 6, 40]. To analyze the implications of an environment with a high concentration of actin, we determined the probability distribution for an actin concentration of 10 mM (Fig 3b). Because of the increased binding probability, the unbound state (1) is depopulated. The weakly bound ADP⋅Pi state (2) dominates for forces smaller than 2 pN and the ATP state (5) for larger forces. An increased ADP concentration of 250 μM traps the myosin head in the ADP state for forces smaller than 2 pN and larger than 4 pN (Fig 3c). In the intervening regime the ATP state predominates.
In our stochastic description without irreversible transitions, we define myosin’s effective velocity as the average number of forward power strokes minus the average number of reverse power strokes per time. We refer to this definition as an effective velocity to emphasize that this quantity is neither the gliding velocity of an actin filament nor the ensemble velocity of several myosin Ic heads cooperating to produce a continuous movement. Every time the myosin head traverses the states (2) → (3) → (4) it performs a net power stroke of size Δx1 + Δx2. In contrast, the reverse pathway (4) → (3) → (2) is associated with a reverse power stroke of size −(Δx1 + Δx2). The effective velocity v is accordingly given in terms of the combined local excess fluxes ΔJij (Eq 26) as
v ≡ Δ x 1 Δ J 23 + Δ x 2 Δ J 34 . (1)
An increasing actin concentration enhances the binding of myosin and therefore decreases its cycling time, which leads to a higher effective velocity (Fig 4). The velocity saturates for an actin concentration above 1 mM. For large forces the effective velocity decreases until it becomes negative for forces larger than the stall force. According to our thermodynamic description the stall force
F s = k B T Δ x 1 + Δ x 2 ln [ ATP ] K eq [ ADP ] [ P i ] (2)
arises directly from Δμ = Eme, the equality of the Gibbs free energy for the hydrolysis reaction and the mechanical output. This relation reflects an implicit assumption that all of the chemical energy can be converted into mechanical energy. To account for mechanical inefficiency, the description could be extended with a loss parameter. Because we restrict our analysis to forces smaller than 6 pN, for which power strokes have been observed experimentally, we ignore the precise behavior for larger forces and consider the stall force for myosin Ic as an unknown quantity.
A widely accepted definition of the duty ratio is the fraction of the total duration of an ATPase cycle that myosin spends in the strongly bound states [3, 41–43]. Ignoring the weakly bound, actin-attached states or combining them into other states, the duty ratio is often defined as the fraction of the total cycle time during which myosin is attached to an actin filament [2, 44–46]. Because the initiation of myosin Ic’s power stroke is limited by phosphate release, myosin Ic can bind to actin in the ADP⋅Pi state but detach without proceeding through the cycle if it detaches prior to Pi release. Such an event contributes to the attachment to the filament but not to the time that the molecule spends in the strongly bound states. The time that the molecule spends in the strongly bound states therefore differs from that spent attached to the filament. The probability Psb of occupying the strongly bound states accordingly differs from the probability Pon of being attached to actin. Our complete cycle description allows us to explicitly calculate both probabilities and to compare them. We determine Psb in terms of the fraction of the cycle that the molecule spends in the strongly bound states as
P sb ≡ t sb t sb + t wb = ∑ i = 3 5 P i , (3)
in which tsb is the average time spent in the strongly bound states, twb is the average time spent in the weakly bound and detached states, and Pi is the steady-state probability (Eqs 45–49). Similarly, we obtain Pon from the fraction of the total cycle time during which the myosin molecule is attached to the filament as
P on ≡ t on t on + t off = ∑ i = 2 5 P i , (4)
in which ton is the average time that myosin is attached to the filament, toff the average time that myosin is detached, and Pi is again the steady-state probability (Eqs 45–49). Whereas the former quantity is closely related to the duty ratio, the later quantity is important for estimation of the number of bound molecules in an ensemble.
The probabilities of being attached to actin and of occupying the strongly bound states depend on the ADP concentration, on the available actin, and on the external force (Fig 5). In general, because of the catch-bond behavior an increasing force enhances the probability of attachment to actin. An elevated ADP concentration likewise traps myosin Ic in the strongly bound ADP state and increases both probabilities (Fig 5a and 5c). An increased accessibility of actin enhances the binding of the myosin head, which results in a high-almost unity-probability of being bound to the filament at high actin concentrations (Fig 5d). In contrast, the probability of occupying the strongly bound states saturates at a high actin concentration, for entering these states is limited by phosphate release (Fig 5b).
Although in vestibular hair cells myosin Ic activity is required for fast adaptation, the precise molecular details remain unknown [10]. Here we focus on two aspects that might contribute to the mechanism: the cooperative unbinding of an ensemble of myosin heads under force and a qualitative Ca2+ dependence that changes the binding probability and the elasticity of individual myosin Ic molecules [23, 24]. In particular, we determine how these properties influence the overall elasticity of an ensemble. The myosin heads contribute to the rigidity of the adaptation motor by crosslinking the insertional plaque to the actin cytoskeleton. We think of each myosin head as a linear spring, arranged in parallel to the others, such that the overall stiffness is given by the sum of the actin-attached myosin heads multiplied by the stiffness of each myosin molecule. Because the binding and unbinding of the heads depend on the force and the nucleotide and actin concentrations, these quantities also influence the overall elastic properties of the ensemble. In general the binding process could be very complicated because of the geometry and possible steric interactions between the heads. Furthermore the helical structure of the actin filaments provides binding sites with an appropriate orientation only about every 37 nm [5]. These constraints change the number of myosin molecules that can potentially interact with actin. In our description, the total number of myosin heads is thus an effective number of molecules that can potentially bind to actin.
To estimate the average number of bound myosin molecules in an ensemble, we use the attachment and detachment rates determined from our description of the chemomechanical cycle. We assume that each myosin head can bind to the filament with a binding rate kon and unbind with an unbinding rate koff. Both rates stem directly from our description, kon = ω12 + ω15 and koff from Eq 58. Because of the stochastic binding and unbinding, the number n of bound molecules fluctuates. To describe the system as a Markov chain, we introduce a state space (Fig 6a) associated with the number of bound myosin heads [47]. The effective transition rates between these states are
k on n ≡ ( N - n ) k on , (5)
and
k off n ≡ n k off . (6)
Here kon depends on the actin concentration and koff on the nucleotide concentrations and on the force f per myosin molecule. We assume that an external force F applied to the ensemble is distributed equally among the attached myosin molecules, resulting in the effective force f = F/n per attached head. If one head releases from the filament then the force is redistributed among the remaining bound heads and the force per myosin molecule accordingly increases, which changes the unbinding rate koff. In general this mechanism leads to cooperative effects because the unbinding rate depends on the number n of attached myosin heads. In the case in which the myosin heads act independently, the transition rates of a single head are independent of the number of attached myosin molecules.
We determine the average number of bound myosin molecules from the linear Markov chain as explained in the Methods section,
n = ∑ n = 0 N n 1 + ∑ l = 0 N - 1 ∏ i = 0 l k on i k off i + 1 - 1 ∏ j = 0 n - 1 k on j k off j + 1 . (7)
For the cooperative case in which koff = koff(F/n), we evaluate this equation. In the independent case, in which the unbinding rate koff is independent of the number of bound myosin heads, we can simplify this expression to
n = N 1 + k off / k on = N t on t on + t off = N P on . (8)
Note that Pon = Pon(f) is a function of the force acting on a single myosin head. For the independent case, we estimate this force by f = F/N. However, in this way we underestimate the magnitude of the force per molecule because we expect that N > n. For a better approximation, we distribute the external force between the mean number of bound motors, f = F/〈n〉, an approach that leads to an implicit equation for 〈n〉 that is not easy to solve. For physiological nucleotide concentrations and for 100 μM actin, we notice in Fig 5d that 〈Pon〉 ≈ 0.5. Using this value, we estimate that in a group of 30 molecules about 〈 n ˜ 〉 ≃ 15 of them are bound on average. We then approximate the average force on a myosin molecule as f = F / 〈 n ˜ 〉 for the independent case. Note that in the independent case the force per myosin head does not depend on the number of bound heads, in contrast to the cooperative case. The mean number of bound myosin heads is influenced by the cooperative release of the molecules and the three approaches are different for intermediate forces (Fig 6b).
We calculate the average number of bound myosin heads as a function of force for different total numbers of myosin molecules (Fig 7a). In small ensembles, the force per head is higher and therefore more heads are bound as a result of the catch-bond behavior. Increasing the concentration of available actin causes more myosin heads to attach to the filament (Fig 7b).
To validate our effective description, we compare our analytic results to Monte Carlo simulations as detailed in the Methods. In these simulations, each myosin head is represented as a spring that is attached to a rigid common structure. At each time step of the simulation the extensions of all springs are calculated by solving Newton’s law of force balance. In this way, we obtain for each myosin head a force that determines the transition rates of the chemomechanical cycle of that molecule. There are important differences from the analytic approach. Whereas in the simulation a myosin head proceeds stochastically through the five-state chemomechanical cycle, the heads only bind and unbind in the analytic description. As a consequence the myosin molecules step stochastically and exert fluctuating forces on each other, which in turn influences their dynamics. In our analytic model, the myosin heads are only indirectly coupled through the number of bound motors and not through an elastic interaction. The simulations show reasonable agreement with the analytic results (Figs 6b and 7). An increased coupling stiffness increases the forces between the myosin heads, which in turn result in a longer attachment because of the catch-bond behavior (Fig 6b). Especially for a high actin concentration, the agreement between the simulations and the analytic description is very good. In the following, we will focus on this particular case and therefore consider only the analytic description.
These results show that the average number of bound myosin heads depends on the external force, the total number of myosin molecules, the actin concentration, and-not shown here-the nucleotide concentrations. We expect that the mechanical properties of a cellular structure including myosin Ic molecules also depend on these quantities.
To investigate the elastic properties of an ensemble of myosin Ic heads, we determine the force-extension relation
F = n κ x , (9)
in which κ is the spring constant of a single myosin head. The underlying assumption of this approach is a linear force-extension relation of the individual myosin heads, for which we take the value of κ = 500 μN/m [21]. Applying forces below 20 pN to the ensemble leads to an extension smaller than 5 nm (Fig 8a). A reduced total number N of myosin molecules increases the extension because the force per myosin head is larger and stretches it farther.
To test whether a mechanical release of myosin Ic molecules is related to fast adaptation, we investigate two qualitative effects of Ca2+. First, Ca2+ could decrease the binding probabilities of the myosin head to actin [23]. Second, it could change the stiffness of myosin by initiating the dissociation of one or more calmodulin molecules from the light chains, allowing the myosin molecules to attain a more flexible conformation [24].
We next consider the mechanical release owing to Ca2+ binding,
Δ x ≡ F 1 κ Ca 2 + n Ca 2 + - 1 κ n . (10)
We first study the effect of a reduced binding probability on the mean number of bound myosin molecules and maintain their stiffness before and after Ca2+ binding, κ Ca 2 + = κ = 500 μ N/m. We reduce the binding probability by the factor β and determine the resulting release for N = 10 or N = 20 myosin molecules (Fig 8b). A large decrease of the binding probability leads to fewer bound molecules and a larger release. The release for a group of 10 myosin molecules exceeds that for an ensemble of 20 molecules: the force on each individual myosin head is higher and stretches the molecule farther. However, the overall distance for forces smaller than 20 pN is still less than 20 nm. When we add to the 100-fold decrease of the binding probability a tenfold decrease of myosin’s elasticity and determine the resulting release for different total numbers of myosin molecules (Fig 8c), the displacement is of the order of several tens of nanometers and becomes almost insensitive to force for a group of 50 myosins.
An important goal of biology is understanding how the structures and interactions of molecules result in measurable functions of cells and organisms. By combining findings on different spatial scales in a consistent manner, mathematical descriptions help us understand how physiologically relevant function is determined by the interplay of molecular components. We have constructed a quantitative description of myosin Ic’s chemomechanical cycle and studied the resulting properties at both a single-molecule and an ensemble level, which allows us to discuss important implications on the physiological function of hair cells at the whole-cell level.
On the single-molecule level, it is important to understand how different members of the large myosin family display distinct biophysical properties despite a common general structure of the chemomechanical cycle. To describe myosin Ic, we constructed such a cycle and chose as control parameter the nucleotide concentrations and the external force, both of which are experimentally accessible and biologically relevant.
Our simplified, one-cycle description reproduces many of the characteristic features of myosin Ic, especially the force-dependent exit from the strongly bound states. The probabilities of occupying the different states indicate that myosin Ic’s strongly bound states are dominated by the ADP state for forces below 1.5 pN and by the ATP state for larger forces (Fig 3). Although this behavior is in contrast to previous models in which the ADP state is the only force-sensitive state, it is nevertheless consistent with the role of myosin Ic in adaptation [5, 21]. Increasing the ADP concentration traps the myosin heads in the ADP state, bound to actin filaments (Figs 3 and 2b). This effect can be reversed by increasing the ATP concentration (Fig 2b). Such a behavior accords with recordings of transduction currents in hair cells isolated from the bullfrog: changing nucleotide concentrations alters the relative occupancy of the states in the cross-bridge cycle and thus the number of bound myosin molecules, which in turn controls the tension on the mechanically sensitive ion channels. Indeed, in the presence of an ADP analog, adaptation disappears and the tension on the channels increases. Both effects can be reversed by increasing the concentration of ATP [48]. This qualitative agreement constitutes direct evidence that the model, although constructed from single-molecule measurements in vitro, captures important aspects of the behavior of living cells.
Our description suggests a low effective velocity for myosin Ic. Although velocities of only tens of nanometers per second have been reported from motility experiments in vitro [10, 49–51], larger values have been discussed [5]. In motility assays, multiple myosin molecules work together to create motion. How the velocity measured in motility experiments is related to the effective rate of a cross-bridge cycle and to other biophysical parameters of the molecules is an open question [52–56]. However, our stochastic simulations suggest that 200 elastically coupled myosin Ic molecules, each described by the five-state chemomechanical cycle, display a motility rate of 25 nm⋅s-1 which is in good agreement with the experimental values of 16–22 nm⋅s-1 [51]. In these experiments the myosin molecules where coupled through a membrane. Greater speeds of 60 nm⋅s-1 have been reported in gliding assays, but the data were acquired at a temperature of 37°C [22], whereas the numerical values of the biochemical rates of our model stemmed from experiments conducted at 20°C. We conclude that our description of myosin Ic constrained by single-molecule data accords with the experimental data on a larger scale.
Speeds of tens of nanometers per second are too low to be consistent with rates estimated for the adaptation motor in the inner ear, which has been associated with the function of myosin Ic [5, 10]. Depending on the species, the velocity of the adaptation motor ranges from several hundred to a few thousand nanometers per second [5, 57, 58]. The discrepancy between the velocities in vivo and in vitro might stem from several factors. It is still unknown to what extent these rates relate to the speed of myosin Ic molecules and to relaxations of other elastic elements. It has been suggested that the recoil of an elastic element located parallel to the myosin heads, the extent spring, contributes to the dynamics [59, 60]. Furthermore the reaction rates of the myosin cycle could be different in vivo and in vitro. In particular, the complex composition of the cytosol and molecular modifications could lead to differences in the energy barriers between the states [61].
Another possibility is that myosin Ic, which has been shown to constitute the adaptation motor of young mice [10], might be replaced during subsequent development by the closely related paralog myosin Ih, which has been identified as a hair-bundle protein [62]. Myosin Ih’s molecular properties have yet to be characterized and it might operate more swiftly.
In hair cells the deflection by a stimulus is communicated to the transduction channel by an elastic element, the gating spring. Of uncertain origin, this elasticity displays complex behavior with implications for sensory coding [63]. Our model suggests that a cluster of myosin Ic molecules contributes to this elasticity and additionally provide the regulatory function to explain fast adaptation. If we assume that Ca2+ reduces the binding probability of myosin by a hundredfold and its stiffness by tenfold, the resultant release on the order of 40 nm accords with measurements from frog hair bundles [10]. For displacements exceeding 400 nm the extent of fast adaptation is independent of the stimulus [10]. We speculate that the insensitivity of the release to the external force for an ensemble of 50 myosin molecules is related to this experimental observation (Fig 8c). Although biochemical studies have suggested that stereocilia contain around 100–200 myosin Ic molecules, the number of actively engaged molecules in an adaptation motor is probably lower [5].
Although myosin Ic’s cycle is slow, binding of Ca2+ could rapidly change the relative occupancy of specific states. Under force, most of the myosin heads are trapped in the ATP state. The binding of Ca2+ to a myosin Ic molecule triggers the release of calmodulin from the IQ domains and increases the molecule’s flexibility, as recently shown in a structural study [24, 64]. A sudden increase of flexibility would release the myosin head from any force until all elastic elements have relaxed to a new equilibrium state. In the absence of force, our description predicts a transition rate for unbinding from the ATP state as large as ω 51 0 ≃ 314 s - 1. This value is so great that the head would unbind immediately, probably before the forces could be redistributed among the bound myosin molecules. Although the load-free biochemical rates have been reported to be rather insensitive to Ca2+, this mechanism might explain a possible Ca2+-induced unbinding of myosin molecules from the actin filament under force [23]. Such a fast disengagement of the myosin molecules is necessary for the adaptation motor to slide down the stereociliary actin and thus to relax the tension in the tip links in order to accomplish adaptation to an abrupt stimulus.
Because the transduction channels have been localized at the lower end of the tip links, Ca2+ regulation of the adaptation motor is effective only at the next lower insertional plaque [65]. Because the tallest stereocilia lack transduction channels through which Ca2+ could enter, the forces between different rows of stereocilia are differently regulated. Inner hair cells consisting of three rows of stereocilia might therefore display less Ca2+-regulated slow adaptation [14]. How much of the hair cell’s function is impeded by the reduced regulation is an open question. Hair-bundle models based on detailed descriptions of the relevant molecular mechanisms, such as myosin Ic’s chemomechanical cycle, could provide more insight.
Biochemical experiments and studies of single-molecule motility are ordinarily conducted under chemostatic conditions in which energy sources such as ATP and products such as ADP and Pi are maintained at nearly constant concentrations. In the present study, however, we have endeavored in two ways to model the behavior of myosin Ic under more lifelike conditions. First, we have imposed a thermodynamic constraint that requires the modeled reaction cycles to respect energy balance. And second, we have examined an extensive range of concentrations for the relevant nucleotides and their products. A typical stereocilium, which is about 3 μm in length and 0.2 μm in diameter, has a volume of only 100 aL. Even a substance found at a high concentration in the cytoplasm, such as ATP at 1 mM, can be depleted rapidly in such a small volume. When transduction channels open, for example, the plasma-membrane Ca2+ ATPase in a stereocilium confronts a flood of Ca2+ that could exhaust the available ATP in only milliseconds! It is thus important to understand the operation of myosin Ic-based motors under realistic and potentially fluctuating conditions.
A final feature of the adaptation motors that remains to be investigated is the noise associated with their activity. By pulling directly on a tip link, each motor influences the opening and closing of the transduction channel or channels at the link’s opposite end. In conjunction with thermal bombardment of the bundle as a whole and stochastic clattering of the transduction channels, the adaptation motors in a hair bundle thus contribute to the mechanical noise that interferes with the detection of faint sounds and weak accelerations [66]. It will be interesting to learn whether the activation mechanism of the myosin molecules in adaptation motors or perhaps their cooperative behavior has been optimized to mitigate this source of noise.
Our study has provided new insights into biological mechanisms. The chemomechanical cycle suggests that the force-dependent unbinding rate is rather robust even under physiological nucleotide concentrations. Although the force-sensitive state is the ATP-bound state, an increased ADP concentration reduces the unbinding rate and the myosin Ic molecules are strongly bound to actin filaments. The elastic properties of an ensemble of myosin Ic molecules can be regulated by an external force and by the actin and nucleotide concentrations. Although the reaction rates of actin-bound myosin Ic are largely insensitive to Ca2+ [23], we have shown that in an ensemble of myosin Ic molecules a possible reduction of the binding rate and elasticity could nevertheless account for fast adaptation by hair cells.
The cross-bridge cycle of a myosin molecule consists of distinct states associated with different biochemical compositions and molecular conformations. The transitions between these states involve myosin’s head binding to and unbinding from the actin filament, nucleotide binding and release, and conformational changes. We simplify the cross-bridge cycle and describe myosin’s dynamics with one state (1) in which the head is detached from actin and four states (2)−(5), in which the head is attached (Fig 1). The four actin-bound states correspond to distinct occupancies of the nucleotide-binding pocket: in state (2) ADP and Pi are bound, whereas in state (3) only ADP is bound. State (4) is the nucleotide-free state and state (5) refers to the ATP-bound state.
We represent the cross-bridge cycle as a time-continuous Markov process for which we must specify the transition rates between the states. Although all transition rates could be force- and nucleotide-dependent, it is reasonable to assume that the main effect of the nucleotide concentrations is exerted on the nucleotide-binding rates. Before introducing those transition rates, we discuss the force dependencies of the mechanical transitions.
The transition rates associated with a mechanical power stroke decrease with an increase in the opposing force. Experimental data indicate that myosin Ic performs its power stroke in two steps: the lever arm is remodeled by a distance of Δx1 ≃ 5.8 nm upon phosphate release and then by Δx2 ≃ 2 nm upon ADP release [20]. Assuming local equilibrium, we associate the ratio of the forward and backward transition rates for the power stroke upon phosphate release with a Boltzmann factor as
ω 23 ω 32 = exp ( - ( Δ G 23 + F Δ x 1 ) / k B T ) . (11)
Here ΔG23 is the Gibbs free-energy difference between the states, FΔx1 is the mechanical work performed by the power stroke of distance Δx1 against the opposing load force F, and kB T is the Boltzmann constant times the temperature [67]. The equation above relies on an assumption of local equilibrium that does not indicate how the individual transition rates depend on the force. Therefore, we use the following general forms for the individual transition rates:
ω 23 ≡ ω 23 0 exp ( - δ 1 F Δ x 1 / k B T ) , (12) ω 32 ≡ ω 32 0 exp ( ( 1 - δ 1 ) F Δ x 1 / k B T ) , (13)
in which we introduce the force-free rate constants ω 23 0, ω 32 0 and the force-distribution factor δ1 ∈ [0, 1]. The restriction on the numerical values for the force-distribution factor is a consequence of the assumption that a force opposing the power stroke diminishes the corresponding transition rate [67, 68]. For an effective description based on a projection of a high-dimensional free-energy landscape on to a single reaction coordinate, the force-distribution factor is not restricted [69]. Using the same argument as for the release of phosphate, the general forms of the forward and backward transition rates associated with the power stroke upon ADP release are
ω 34 ≡ ω 34 0 exp ( - δ 2 F Δ x 2 / k B T ) , (14) ω 43 ≡ ω 43 0 exp ( ( 1 - δ 2 ) F Δ x 2 / k B T ) . (15)
To account for the force-dependent behavior of myosin Ic, we must include the force sensitivity of the isomerization following ATP binding [19]. In our simplified state space this sensitivity effectively changes the unbinding rate ω51 from the ATP state. We therefore introduce a force-dependent factor g(F) that modifies the unbinding rate
ω 51 ≡ g ( F ) ω 51 0 . (16)
We require that for zero force g(F = 0) = 1 and for large force g(F ≫ 1) = ωoff saturates and use
g ( F ) ≡ 2 ( 1 - ω off ) 1 + exp ( ξ F / k B T ) + ω off , (17)
in which ξ is a characteristic length scale.
In general the binding interface between the head of a molecular motor and its filament is more complicated than the idealized receptor-ligand bond considered by Bell [70]. The bond interface consists of multiple partial charges that lead to complex unbinding pathways through multiple states in the free-energy landscape [71–74]. To capture the characteristic behavior, we use the force factor of Eq 17 that has been used previously to describe the chemomechanical cycle of kinesin-1 and myosin V [75–79].
In the following we give an intuitive justification of the force factor given in Eq 17. In our description the ATP state (5) comprises several sub-states including the binding and isomerization of ATP. The unbinding rate ω51, which must be considered as an effective rate for proceeding through all the sub-states, therefore includes the force dependence of the isomerization step. As a first approximation, we consider that isomerization is not associated with a conformational change that would result in a displacement of an applied load. The free-energy between the state (A) before isomerization and the state (B) after isomerization accordingly does not depend on the applied force. We assume that an applied force increases the free-energy barrier between those two states without changing the difference of the energy between the states. Motivated by Kramers rate theory [80], we use the force dependence ω A B = ω A B 0 exp ( - F ξ / k B T ) for the forward transition rate and the same force dependence for the reverse transition rate ω B A = ω B A 0 exp ( - F ξ / k B T ). We consider the main forward pathway through these sub-states,
→ ( A ) ⇌ ω A B ω B A ( B ) → ω B , (18)
which implies the effective transition rate
ω eff = ω B 1 + ω B A 0 ω A B 0 + ω B ω A B 0 e F ξ / kB T . (19)
This approach leads to a force dependence similar to that in Eq 17. Using the same argument for the reverse pathway, we find that the transition rate ω54 has a similar force dependence. As described later in Eq 43, the force dependence of the transition rate ω54 is imposed naturally by thermodynamic consistency.
To capture the dependence on the nucleotide concentrations, we consider the nucleotide-binding steps as first-order reactions that are independent of force, leading to
ω 32 0 ≡ ω ^ 32 0 [ P i ] , (20) ω 43 0 ≡ ω ^ 43 0 [ ADP ] , (21) ω 45 ≡ ω ^ 45 0 [ ATP ] . (22)
Note that the units of the rate constants with a caret are M−1s−1. Such a linear dependence of the transition rates on the reactants is motivated by macroscopic chemical-reaction laws and is widely used to describe chemomechanical cycles [2, 67]. In a similar way, we assume a linear dependence of the actin-binding rates on the actin concentration,
ω 12 ≡ ω ^ 12 [ actin ] , (23) ω 15 ≡ ω ^ 15 [ actin ] . (24)
The remaining transition rate ω54 is determined by a balance condition obtained from thermodynamic consistency.
We have specified above the general forms of the transition rates of our theoretical description of myosin Ic. We next introduce the dynamics and the thermodynamic constraints. We consider the stochastic dynamics of the myosin head as a continuous-time Markov process [81]. The probability Pi(t) of finding myosin in state (i) therefore evolves in time t according to the master equation
d d t P i ( t ) = - ∑ j Δ J i j ( t ) , (25)
with a local net flux between the states (i) and (j) given by
Δ J i j ( t ) ≡ P i ( t ) ω i j - P j ( t ) ω j i . (26)
A thermodynamically consistent description, which ensures that myosin does not produce more mechanical energy than the chemical energy provided by the nucleotide concentrations, implies a relation between the mechanical energy and the chemical energy. This relation provides a constraint on the transition rates of the cycle that is obtained by incorporating free-energy transduction [37]. We can express the change in the Gibbs free energy of the hydrolysis reaction for a dilute solution by
Δ μ ≡ k B T ln [ ATP ] [ ADP ] [ P i ] K eq , (27)
in which Keq is the equilibrium constant for the reaction, here with the numerical value Keq ≃ 4.9 ⋅ 105 M [2]. Note that at the equilibrium concentration of the nucleotides the change in the free energy Δμ vanishes.
The mechanical energy, which is the work done by the protein against an external force F, is given by
E me ≡ ( Δ x 1 + Δ x 2 ) F . (28)
This mechanical energy is produced when the protein passes through a forward cross-bridge cycle, which in our description represents directed transitions through the states (1), (2), (3), (4), (5), and finally (1). In contrast, the backward cycle is associated with a path traversed in the opposite direction. Thermodynamically consistent coupling of the energy conversion by the protein to the hydrolysis reaction then imposes the constraint
ω 12 ω 23 ω 34 ω 45 ω 51 ω 21 ω 32 ω 43 ω 54 ω 15 = exp ( ( Δ μ - E me ) / k B T ) , (29)
which can further be related to the entropy production [37, 77]. This equation has an intuitive interpretation [37]: the left side is the ratio of the average number of complete forward cycles to the average number of complete backward cycles, whereas the right side is the exponential of the difference between the chemical input energy and the mechanical output energy. At equilibrium this difference vanishes and the right side is equal to one, which requires the completion of identical numbers of forward and backward cycles. This constraint ensures that the average net cycling of the protein is thermodynamically consistent with the energy input. In our simple approach each hydrolysis reaction produces a power stroke, meaning that there are no futile cycles and therefore the chemistry is tightly coupled to the mechanics.
We incorporate into our description of myosin Ic as many experimental data as possible. Some transition rates and parameter values have been reported [19, 20]; an overview of these is given in Table 1. The force-dependent lifetime of an actin-myosin bond has been determined with an optical trap [20]. Because of the finite time resolution of the experimental apparatus, it is reasonable to assume that the strongly bound states rather than the weakly bound ones dominate the lifetime. We accordingly interpret the reported lifetime as the time that myosin is attached to actin in the strongly bound states. As a consequence, the reported duty ratio r ≃ 0.11 and force-free binding time in the strongly bound states tsb ≃ 0.213 s provide an estimate of the time that the myosin molecule resides in the weakly bound states,
t wb = t sb 1 - r r ≃ 1 . 72 s . (30)
The rate constant, ω 23 0 for phosphate release has been reported for human myosin-IC [82] as
ω 23 0 ≃ 1 . 5 s - 1 . (31)
Because of the opaque nomenclature of myosin-I isoforms, human myosin-IC is instead myosin Ie in the nomenclature of the Human Genome Organization [83]. This value is reasonable if the rate-limiting step is phosphate release and the order of magnitude is in agreement with the considerations for myosin Ic given in [19].
We next discuss the binding probability and the free-energy difference associated with the power stroke. Because there are to our knowledge no direct measurements for myosin Ic, we use values reported for myosin II. To estimate the probability π2 that the head binds in state (2), we refer to the cycle for rabbit skeletal muscle [2]. The reported numbers suggest a probability
π 2 ≃ 0 . 998 , (32)
which implies that myosin starts its cycle predominantly in state (2). Using the definition of this binding probability
π 2 = ω 12 ω 12 + ω 15 , (33)
we can relate the binding rates to each other as
ω 12 = π 2 ω 15 1 - π 2 . (34)
Because both transition rates depend linearly on the actin concentration (Eqs 23 and 24), the actin concentration cancels and
ω ^ 12 = π 2 ω ^ 15 1 - π 2 . (35)
Several studies suggest that a large free-energy difference is associated with the main power stroke [2, 28, 32–34, 38]. Using the value ΔG23 ≃ −15 kBT inferred from fitting a model of the myosin II cycle to experimental data acquired with frog muscle [28], we obtain the ratio
ω 23 0 ω ^ 32 0 [ P i ] = exp ( - Δ G 23 / k B T ) , (36)
which provides the phosphate-binding rate constant
ω ^ 32 0 = ω 23 0 [ P i ] exp ( Δ G 23 / k B T ) . (37)
Assuming a phosphate concentration of [Pi] = 1 mM in frog muscle and using Eq 31, we determine that
ω ^ 32 0 ≃ 4 . 5 · 10 - 10 s - 1 μ M - 1 . (38)
A complementary approach is to use the values for the rabbit muscle cycle [2]; we then obtain ω ^ 32 0 ≃ 4 . 6 · 10 - 9 s - 1 μ M - 1, a value one order of magnitude larger. Because both rate constants are very small compared to the other transition rates of our description, they make no significant difference in our results.
In a one-cycle description, the effects of changing the ATP and ADP concentrations are tightly coupled and determined by the magnitudes of the rate constants. If we assume a very fast and irreversible unbinding from the ATP state, ω51 ≫ 1 and ω15 = 0, and an irreversible Pi release, ω32 = 0, the average time in the strongly bound states reads
t sb = ω 34 + ω 43 + ω 45 ω 34 ω 45 . (39)
In the force-free case the ADP and ATP binding rates are given in Eqs 21 and 22. Using the equilibrium binding constant KADP to estimate the rate constant for ADP binding as
ω ^ 43 0 = ω 34 0 K ADP , (40)
we rewrite Eq 39 as
t sb = ω 34 0 ( 1 + [ ADP ] / K ADP ) + ω ^ 45 0 [ ATP ] ω 34 0 ω ^ 45 0 [ ATP ] . (41)
The two rate constants and the equilibrium constant have been determined experimentally as ω 34 0 ≃ 3 . 9 s - 1, ω ^ 45 0 ≃ 0 . 26 s - 1 μ M - 1 and KADP ≃ 0.22 μM [20]. Because of the low ATP-binding rate constant and the small equilibrium constant KADP for ADP release, the lifetime of the strongly bound states is very sensitive to elevated ADP concentrations. Because the experimental findings differ, we resolve this problem in our one-cycle description by using the higher value of KADP ≃ 1.8 μM for the equilibrium constant for ADP release [22, 23]. Although another possibility would be to introduce a multi-cycle description, that strategy increases complexity and the number of unknown parameters.
To satisfy thermodynamic consistency, we express the transition rate for ATP release in terms of all the other transition rates as given by the balance condition of Eq 29,
ω 54 = ω 12 ω 23 ω 34 ω 45 ω 51 ω 21 ω 32 ω 43 ω 15 exp ( - ( Δ μ - ( Δ x 1 + Δ x 2 ) F ) / k B T ) . (42)
This transition rate is dependent on force in the same way as ω51 but independent of the nucleotide concentrations, as can be concluded by applying eqs (12)–(16), (20)–(22) and (27), resulting in
ω 54 = ω 12 ω 23 0 ω 34 0 ω ^ 45 0 g ( F ) ω 51 0 ω 21 ω ^ 32 0 ω ^ 43 0 ω 15 K eq ≡ ω 54 0 g ( F ) . (43)
We are left with seven unknown parameter values: the unbinding rates ω51 and ω21, the binding rate ω15, the force distribution factors δ1 and δ2, the characteristic length ξ, and the offset rate ωoff. To estimate these values, we use analytic expressions of the average time spent in the weakly bound states and the effective unbinding rate from the strongly bound states and fit these functions to the experimental data. We derive these analytic expressions in the following section.
The probability Pi of finding myosin in one of the states of the cycle is given by the solution to the steady-state master equation
0 = - ω 12 - ω 15 ω 21 0 0 ω 51 ω 12 - ω 21 - ω 23 ω 32 0 0 0 ω 23 - ω 32 - ω 34 ω 43 0 0 0 ω 34 - ω 43 - ω 45 ω 54 ω 15 0 0 ω 45 - ω 54 - ω 51 P 1 P 2 P 3 P 4 P 5 , (44)
and read
P 1 ≡ ( ω 23 ω 34 ω 45 ω 51 + ω 21 ( ω 34 ω 45 ω 51 + ω 32 ( ω 43 + ω 45 ) ω 51 + ω 32 ω 43 ω 54 ) / N , (45) P 2 ≡ ( ω 12 ( ω 34 ω 45 + ω 32 ( ω 43 + ω 45 ) ) ω 51 + ( ω 12 + ω 15 ) ω 32 ω 43 ω 54 ) / N , (46) P 3 ≡ ( ω 15 ( ω 21 + ω 23 ) ω 43 ω 54 + ω 12 ω 23 ( ω 45 ω 51 + ω 43 ( ω 51 + ω 54 ) ) ) / N , (47) P 4 ≡ ( ω 15 ( ω 23 ω 34 + ω 21 ( ω 32 + ω 34 ) ) ω 54 + ω 12 ω 23 ω 34 ( ω 51 + ω 54 ) ) / N , (48) P 5 ≡ ( ω 12 ω 23 ω 34 ω 45 + ω 15 ( ω 21 ω 32 ω 43 + ω 23 ω 34 ω 45 + ω 21 ( ω 32 + ω 34 ) ω 45 ) ) / N , (49)
in which N is determined by the normalization ∑i Pi = 1.
To determine the effective unbinding rate from the strongly bound states, we calculate the average attachment time in the strongly bound states using a framework introduced by Hill [84, 85]. We promote the detached state (1) and weakly bound state (2) to absorbing states by setting the transition rates ω15, ω23, ω12, and ω21 to zero (Fig 9a). The associated effective unbinding rate then becomes the inverse of the average time to absorption for the appropriate initial condition. The basic idea is to use an ensemble average instead of a time average. The dynamics of the correct ensemble is described by a closed diagram in which the absorbing state is eliminated by redirecting the transitions into that state to the starting states weighted with the appropriate starting probabilities [86]. For example, the transition from state (3) to state (2) is redirected with the weight 1 − π3 to state (5) and with weight π3 to state (3). The latter transition, a self loop, cancels in a master equation and can therefore be disregarded. This procedure creates a closed diagram (Fig 9b) for which the steady-state probability pi of being in state (i) is determined from the master equation
0 = - ( 1 - π 3 ) ω 32 - ω 34 ω 43 π 3 ω 51 ω 34 - ω 43 - ω 45 ω 54 ( 1 - π 3 ) ω 32 ω 45 - ω 54 - π 3 ω 51 p 3 p 4 p 5 , (50)
together with the normalization condition ∑pi = 1. This probability distribution provides the average state occupancy before absorption. The average rate of arrivals at either of the absorbing states is therefore given by the probability current, which is identical to the effective unbinding rate from the strongly bound states
t sb - 1 ≡ ω 32 p 3 + ω 51 p 5 . (51)
The probability of starting in state (3) and not in state (5) is given by the relative probability current into state (3) of the complete cycle as
π 3 ≡ ω 23 P 2 ω 23 P 2 + ω 15 P 1 , (52)
in which P1 and P2 are given in Eqs 45 and 46, respectively. For the sake of completeness we give the rather cumbersome expression for the effective unbinding rate from the strongly bound states,
t sb − 1 = ( ( ( ω 15 ω 21 + ω 12 ω 23 ) ω 32 ω 43 + ( ω 12 ω 23 ( ω 32 + ω 34 ) + ω 15 ( ω 21 ω 32 + ( ω 21 + ω 23 ) ω 34 ) ) ω 45 ) ω 51 + ( ω 15 ω 21 + ( ω 12 + ω 15 ) ω 23 ) ω 32 ω 43 ω 54 ) / H , (53)
in which
H ≡ ω 15 ( ω 21 ( ω 32 ( ω 43 + ω 45 ) + ω 34 ω 45 ) + ω 23 ω 34 ω 45 + ( ω 23 ( ω 34 + ω 43 ) + ω 21 ( ω 32 + ω 34 + ω 43 ) ) ω 54 ) + ω 12 ω 23 ( ( ω 43 + ω 45 ) ω 51 + ω 43 ω 54 + ω 34 ( ω 45 + ω 51 + ω 54 ) ) . (54)
The time that myosin spends in the weakly bound states can be obtained from Eq 3 as
t wb = t sb 1 ∑ i = 3 5 P i - 1 , (55)
in which Pi are the steady-state probabilities given in Eqs 47–49.
To determine the time during which myosin is attached to the filament, we promote state (1) to an absorbing state. The corresponding closed diagram is obtained by redirecting the transition from state (2) to state (1) with weight 1 − π2 to state (5) and the transition from state (5) to state (1) with weight π2 to state (2). The steady-state probability si of being in state (i) for this closed diagram is the solution of the master equation
0 = - ω 23 - ( 1 - π 2 ) ω 21 ω 32 0 π 2 ω 51 ω 23 - ω 32 - ω 34 ω 43 0 0 ω 34 - ω 43 - ω 45 ω 54 ( 1 - π 2 ) ω 21 0 ω 45 - ω 54 - π 2 ω 51 s 2 s 3 s 4 s 5 , (56)
together with the normalization condition ∑si = 1. This probability distribution gives the probability current into state (1), which is identical to the unbinding rate from the actin filament
k off ≡ ω 21 s 2 + ω 51 s 5 . (57)
Again, for completeness, we give the unwieldy expression for the unbinding rate from the filament,
k off = ( ω 23 ω 34 ω 45 + ω 21 ( ω 34 ω 45 + ω 32 ( ω 43 + ω 45 ) ) ) ω 51 + ω 21 ω 32 ω 43 ω 54 A + B + C , (58)
in which
A ≡ π 2 ( ω 34 ω 45 + ω 32 ( ω 43 + ω 45 ) ) ω 51 + ω 32 ω 43 ω 54 , (59) B ≡ ( 1 - π 2 ) ω 21 ( ω 34 ω 45 + ( ω 34 + ω 43 ) ω 54 + ω 32 ( ω 43 + ω 45 + ω 54 ) ) , (60) C ≡ ω 23 ( π 2 ( ω 43 + ω 45 ) ω 51 + ω 43 ω 54 + ω 34 ( ω 45 + π 2 ω 51 + ω 54 ) ) , (61) π 2 ≡ ω 12 / ( ω 12 + ω 15 ) . (62)
Note that π2 is independent of the time toff during which myosin is detached from the filament, and therefore independent of the actin concentration. As a consequence the unbinding rate koff is also independent of the actin concentration.
To describe the ensemble of myosins as a Markov chain, we introduce a state space (Fig 6a) associated with the number of bound myosin heads [47]. Assuming that the heads bind and unbind independently of one another, transitions between these states can be expressed in terms of the individual binding and unbinding rates kon and koff. A transition from state (n), in which n myosins are bound, to state (n + 1) is associated with the binding rate
k on n ≡ ( N - n ) k on . (63)
The reverse transition is described by the unbinding rate
k off n ≡ n k off . (64)
We next incorporate an external force F into the description. As a first approximation, we assume that the myosin heads share this load equally, resulting in an effective force F/n exerted on each of the bound myosin heads and in a modified transition rate
k off n ≡ n k off ( F / n ) . (65)
With this specific choice of the transition rates we determine the probability Sn of being in state (n) from a master equation. The solution for such a finite linear Markov chain is a standard result in stochastic dynamics and can be obtained recursively or by standard methods [47, 81]. The probability of being in state (n) is
S n = S 0 ∏ i = 0 n - 1 k on i k off i + 1 , (66)
in which S0 is determined from the normalization ∑Si = 1 as
S 0 = 1 + ∑ n = 0 N - 1 ∏ i = 0 n k on i k off i + 1 - 1 . (67)
From this probability distribution we obtain the average number of bound myosin heads as
n = ∑ n = 0 N n S n . (68)
We describe the dynamics of a myosin head with the five state chemomechanical cycle that we developed in this study. Each myosin head is coupled with a spring to a rigid common structure. The extension of this spring determines the force that is exerted on the myosin molecule and thus all transition rates of its cycle. We assume that a myosin head binds without tension to the actin filament and when it proceeds through its cycle the power-stroke transitions stretches the spring. At each time step we determine the extensions of all myosin springs from Newton’s law of force balance and adjust all transition rates accordingly. For the Monte Carlo simulations we use a Gillespie algorithm which is a standard method to simulate multi-component stochastic reactions and used for elastically coupled motor molecules [35, 78, 79, 88]. After disregarding the transient behavior of our simulations, we determine the average number of myosins bound to actin from a time average. We consider all myosin molecules in state (2) to (5) bound to actin.
|
10.1371/journal.pcbi.1000104 | Circadian Phase Resetting via Single and Multiple Control Targets | Circadian entrainment is necessary for rhythmic physiological functions to be appropriately timed over the 24-hour day. Disruption of circadian rhythms has been associated with sleep and neuro-behavioral impairments as well as cancer. To date, light is widely accepted to be the most powerful circadian synchronizer, motivating its use as a key control input for phase resetting. Through sensitivity analysis, we identify additional control targets whose individual and simultaneous manipulation (via a model predictive control algorithm) out-perform the open-loop light-based phase recovery dynamics by nearly 3-fold. We further demonstrate the robustness of phase resetting by synchronizing short- and long-period mutant phenotypes to the 24-hour environment; the control algorithm is robust in the presence of model mismatch. These studies prove the efficacy and immediate application of model predictive control in experimental studies and medicine. In particular, maintaining proper circadian regulation may significantly decrease the chance of acquiring chronic illness.
| The robust timing, or phase, of the circadian clock is critical in directing and synchronizing molecular, cellular, and organismal behaviors. The clock's failure to maintain precision and adaption is associated with sleeping disorders, depression, and cancer. To better study and control the timing of circadian rhythms, we make use of systems theoretic tools such as sensitivity analysis and model predictive control (MPC). Sensitivity analysis is used to identify key driving mechanisms without having to fully understand or investigate the detailed mechanistic interconnections of the large complex circadian network. Contrary to intuition, sensitivity analysis of the circadian model highlights several non-photic control inputs (such as transcriptional regulation) that outperform light-based circadian phase resetting – light is known to accelerate protein degradation. Aside from targeting individual parameters as control inputs, our methods identify combinations of control targets that may further the efficiency of entrainment. We compare the phase resetting performance of our MPC algorithm among cases involving individual and multiple simultaneous control targets (in wild-type simulations). We then tailor the algorithm to correct continuously the phase mismatch that occurs in short and long period mutant phenotypes. Through use of the presented tools, our algorithm is robust in the presence of model mismatch and outperforms the natural in silico sun-cycle–based phase recovery strategy by nearly 3-fold.
| Control theoretic tools have been used to model mRNA transcriptional/translational regulatory feedback mechanisms [1], to analyze nonlinear phenomena [2],[3], and to control complex biological behavior [4],[5]. In our research, we couple systems theoretic tools (such as sensitivity analysis) with model predictive control, to better address phase resetting properties of nonlinear biological oscillators. Our work aims to alleviate circadian-related disorders (such as jet lag and advanced/delayed sleep phase syndromes) by investigating the phase resetting properties of an example circadian mathematical model. More specifically, we manipulate multiple control inputs (or target parameters) to drive the dynamic behavior of the system.
Many researchers have shown that the systematic application of light pulses may reset the phase of circadian clocks. This light pulse (input) to induced phase-shift (output) mapping is most notably characterized by the phase response curve (PRC). Daan and Pittendrigh studied the PRC to establish a relationship among circadian behavior (nocturnal vs. diurnal activity), free-running period, and maximum phase advance/delay [6]. The free-running period of an organism reflects its circadian behavior without the influence of entrainment factors such as environmental light∶dark cycles. The free-running period of nocturnal animals, for instance, is often less than 24 hours such that dusk triggers a phase delay and the onset of activity. Conversely, diurnal animals often exhibit free-running periods greater than 24 hours such that dawn triggers a phase advance and the onset of activity [6]. Other researchers have made use of PRCs to establish light as a means to accelerate circadian entrainment [7], or as a means to start, stop, and reset the phase of simplified circadian models [8]–[11].
In a previous study, we develop a closed-loop nonlinear model predictive control (MPC) algorithm that minimizes the phase difference between a reference and a controlled system (each modeled as a single deterministic oscillator) through the systematic application of continuous light. Through use of MPC, circadian phase is recovered in almost half the time required by the natural open-loop sun cycles [4]. Next, we investigated how the MPC algorithm's tuning parameters might affect the model's phase resetting dynamics [12]. Here, we make use of sensitivity analysis to identify additional control inputs (or drug targets) that, when used by the MPC algorithm, outperform light-based circadian phase resetting. The target identification of single and multiple control inputs, coupled with the analysis of their respective performance, parallels efforts in the pharmaceutical industry to yield the greatest behavioral response with respect to the smallest system perturbation. In other words, our methodology may be used to identify optimal (and arguably non-intuitive) drug targets for therapy.
To establish an upper bound relating to the time required to recover phase differences, we begin by evaluating the open-loop control algorithm in the Open-Loop Phase Recovery section. The identification and manipulation of a set of single, dual, and triple control inputs are then used to minimize phase recovery dynamics of a wild-type circadian system (as described in the Single, Dual, and Triple Target Phase Resetting sections, respectively). This case is most similar to resetting a healthy organism's phase when subject to an environmental disturbance such as jet lag. In the Short and Long Period Mutants section, we further investigate how MPC may be used to alleviate chronic circadian disorders. More specifically, we apply the algorithm to circadian oscillator models that exhibit either short or long-period mutant phenotypes. Results suggest that organisms with such syndromes may track regular 24 hour rhythms through the systematic application of light. Our findings support this unique application of systematic drug target identification coupled with model predictive control for use in medicine and pharmacology (see the Discussion section). In the Methods section, we describe the employed model predictive control algorithm and the state-based sensitivity analysis used to identify single and multiple parametric control inputs.
A 10-state, 38-parameter Drosophila melanogaster (fruit fly) circadian model serves as the example system. This stable nonlinear limit cycle oscillator consists of two coupled negative feedback loops that characterize the transcriptional regulation of period and timeless mRNA and protein dynamics [13]. per and tim genes are transcribed in the nucleus, after which their mRNAs are transported into the cytosol where they serve as a template for protein synthesis. The doubly phosphorylated proteins form a heterodimer, PER-TIM, that enters the nucleus and inhibits gene expression, closing the feedback loop. Researchers find that environmental light increases the rate of TIM protein degradation: in this model, light targets the system by magnifying νdT, the doubly phosphorylated TIM protein degradation rate [13].
The phase response of this model as a function of light is shown via the dash-dotted line in Figure 1. This curve maps the circadian time of the entraining stimulus (light pulses) against the resulting change in phase of an organism kept in a free-running environment. The circadian time index repeats every 24 hours with CT0 defining the commencement of dawn and CT12 that of dusk. It is important to note that the magnitude of light-induced phase changes (the quantitative dynamics of the PRC) may vary with respect to the intensity of light. While this model does not account for the complexity of the real network that, for instance, includes additional positive feedback loops [14],[15], it has been experimentally validated [15] and is widely employed as a reference model [3],[16].
Due to the inherent nonlinear phase response of circadian rhythms when subject to environmental/parametric perturbations, phase recovery dynamics are characterized as a function of the initial condition (IC, the circadian time at which control or entrainment begins), and initial phase difference (IP, the amount of circadian time to be recovered). To establish a phase resetting set point or upper bound (the maximum amount of time required to recover a given phase difference), we evaluate the open-loop control algorithm, where environmental light∶dark cycles serve as the only mechanism for phase re-entrainment. The phase recovery surface (Figure 2) displays the time required for the open-loop case to recover from any possible initial condition and initial phase difference. The asymmetry of the surface may be attributed to the nonlinear effects of light, as characterized by the PRC. The input (light) to output (induced phase shift) mapping of the PRC is seldom symmetric. In Drosophila melanogaster, a 15 minute pulse of light has shown to induce up to 3.6 hours of phase advance and 4.2 hours of phase delay [13]. Recent studies suggest that the change in phase is less sensitive to the duration of the light, and more sensitive to its time-profile [17]. Phase recovery times (for both open and closed-loop simulations) are evaluated with respect to initial conditions and phase differences discretized at 3 hour intervals. Thus, given the integers i,j ∈ [0,7], IC = 3i and IP = 3j.
The open-loop entrainment strategy requires at most 183 hours to reset the observed states of the controlled system (cumulative protein complex concentrations) to within 15% of the reference trajectories. Mandating the convergence of state trajectories is a tighter constraint than mandating only phase trajectories, since it incorporates amplitude characteristics. The algorithm, however, may be tuned to consider only strict phase measures. The maximum open-loop recovery time refers to a 9 hour initial phase difference whose control action begins at an initial condition of 15 hours. The initial condition, or start of entrainment, is described with respect to circadian time (CT). Interestingly, there is a stark difference between resetting a 3 to 6 hour initial phase difference versus an 18 to 21 hour initial phase difference (a −6 to −3 hour phase difference). In the former, phase recovers in over 100 hours; in the latter, phase recovers in fewer than 60 hours. Additionally, the open-loop algorithm recovers 9 hour phase differences in a fraction of the time required to correct for smaller phase difference. These properties may be attributed to the nature of the phase response curve and are discussed further in the Discussion section. Experimental studies in mammalian SCN cells support this asymmetry: Reddy et al. show that circadian clock resetting from a 6 hour phase advance (IP6) is accompanied by dissociation of cellular gene expression and may take up to 1 week to recover [18]. Conversely, resetting a 6 hour phase delay (IP18) is accompanied by coordinated gene expression and requires only 2 days of recovery [18]. Our simulations support these experimental conclusions as the cumulative protein concentrations in the former case diverge and require several days to converge to the nominal trajectory. In the latter, cumulative protein concentrations oscillate with smaller amplitude until they converge to the nominal trajectory within a couple days. An example of the corresponding simulations is presented in Figure S1.
The MPC algorithm (described in the Model Predictive Control section) minimizes the normalized difference between the cumulative protein complex concentration over a prediction horizon of 48 hours, by admitting control action during the first 8 hours of the simulated trajectory. This control action is multiplicative, allowing the algorithm to increase/decrease the nominal parameter by a factor of 2. The control profile defined within the move horizon is updated every 2 hours. Through use of MPC, the re-synchronization rate of the controlled system is increased nearly 3-fold through the control of light, or νdT. Although light serves as a powerful control input, we show that the manipulation of parameters such as transcription and mRNA degradation rates (νs and νm, respectively) may provide more immediate phase resetting. Since we make use of the symmetric version of the mathematical model [13], we do not differentiate between per or tim specific functions. Instead, we assume that the isolated control of νsP is equivalent to the isolated control of νsT, for instance.
Parametric sensitivity analysis quantifies the relative change of system behavior with respect to an isolated parametric perturbation. A large sensitivity to a parameter, for instance, suggests that the system's performance is subject to greater change with small variations in the given parameter. We make use of the Fisher Information Matrix (FIM) to evaluate the effect of parametric perturbations on the circadian system's state trajectories [19]. Investigation of the diagonal values, off diagonal values, and singular value decomposition of the FIM points out the relative order, or rank, of parametric sensitivity measures. This relative ordering highlights sets of control inputs whose manipulation may further reduce phase recovery times. The three greatest diagonal values, for instance, identify the most prominent individual control targets (ranked from most to least sensitive);
Recall that νdT is the target parameter of environmental light in Drosophila. Interestingly, the rate of mRNA transcription is the target of environmental light in Mus (via per genes) [20],[21] and Neurospora (via frq genes) [22]. Furthermore, in our previous studies of Mus and Drosophila circadian networks, mRNA transcription rates were among the most sensitive parameters with respect to both the state- and phase-based sensitivity analysis of two independent network representations [2].
The greatest off diagonal values identify the most prominent pairs of control targets (ranked accordingly);
Since the manipulation of more than 1 parameter voids the symmetry argument, we target tim specific parameters in the implementation of multiple control targets.
The greatest input directions of the singular value decomposition identify the most prominent set of three control targets (ranked accordingly);
We investigate the phase recovery dynamics corresponding to four independent isolated control inputs with respect to the initial condition and initial phase difference (Figure 3). Results show that control targets identified via sensitivity analysis (Figure 3(A)–3(C)) serve as more effective re-entrainment factors than light (Figure 3(D)). More specifically, the maximum recovery time corresponding to a control input of νs is 44 hours (at IC9 and IP12/IP15), νm is 50 hours (at IC21 and IP15), ks is 59 hours (at IC12 and IP15), and νd (the light target) is 60 hours (at IC12 and IP15, or IC9 and IP12). The control profiles and state response dynamics relating to the phase recovery of IC9 and IP12 are provided in Figure S2 and Figure S3.
There is a subtle similarity among the single-input phase recovery data; namely, the sudden drop in recovery time with respect to the initial condition for initial phase differences of 0 to 15 hours. We attribute this steep recovery gradient to the PRC as it depicts a greater region of phase delay than it does a phase advance. For this reason, it is more beneficial if the organism delays its phase to recover from a 12 hour initial phase difference. Furthermore, recall that a phase delay is incurred if the organism is to receive a photic input in the late evening hours. Hence, recovering from a phase difference via a set of delaying control inputs is most efficient if control action begins around the late subjective evening. Thus, if we observe phase resetting behavior corresponding to a small phase difference (such that the subjective day of the controlled system and reference are similar), we expect it to have the shortest recovery time near an initial condition of 12 hours, or dusk (Figure 3(D)). Interestingly, each of the control inputs exhibits this property. We attribute this similarity to the unique PRC of each control input (Figure 1).
Mutant phenotypes of the circadian oscillator represent cases in which nominal light∶dark cycles are unable to maintain synchrony. For this reason, the MPC tuning parameters must be re-evaluated according to this phase resetting problem. In wild-type, for instance, we can afford to be more aggressive with control penalties since nominal light∶dark cycles (or, no control) will eventually entrain the system. In mutants, the weights used to penalize the state error and control inputs prove to be more influential since nominal light∶dark cycles will not entrain the system. Therefore, we set both the move and prediction horizon to 24 hours and reduce the penalty of state error and control to ones. To counter the computational expense incurred with a longer move horizon, we set the time step to 4 hours. Through MPC, we identify a more suitable light∶dark cycle that synchronizes organisms exhibiting abnormally short and long free-running periods (22 and 27 hours, respectively, as shown in Figure 5). Determining the complete range of entrainment (which is likely wider than the 22 to 27 hour period) is non-trivial. In a previous study, we found that (i) the predicted range of entrainment may be very sensitive to the employed performance metric, and (ii) the control/light input strength may also play a dominant role in defining the bounds of this range [23].
Given that the PRC characterizing the behavior of Drosophila melanogaster consists of phase delays during the late subjective evening, we expect short-period mutant phenotypes to require bright light after subjective dusk. Similarly, we expect long-period mutant phenotypes to require bright light in the early subjective morning to advance the cycle. Our results confirm this hypothesis. In Figure 2, we demonstrate how bright light, admitted during the environmental night, resets the phase of short-period mutants such that it matches that of its environment. Given that the controlled system is 2 hours short, the occurrence of light during the night overlaps with the advance region of the system's PRC. Similarly, the onset of bright light at dawn overlaps with the delay region of long-period mutant PRCs (Figure 2). Our ability to maintain appropriate phase relationships between mutant phenotypes (models characterized by non-nominal parameters) and the environment (the nominal case) further proves the robustness of the algorithm despite model mismatch.
As implied by the PRC (Figure 1), a 3 hour phase difference may be recovered immediately through admission of light at CT15. Hence, for open-loop control action to be most effective, environmental daylight should occur during the controlled system's subjective night (at CT15). In cases with small initial phase difference (such that the subject's internal time is nearly equal to environmental time), however, daylight begins entrainment once the subjective day is around CT12, by inducing small phase delays. This delay reduces the overlap between environmental daylight and the subjective night since re-entrainment of the initial phase difference began before subjective night. The opposite occurs with small negative phase differences, where an 18 hour (or −6 hour) phase difference may be recovered via a light pulse admitted at CT21. In this case, environmental daylight affects the controlled system at the start of day while it has not yet begun entrainment, maximizing the phase advancing effect of light. For this reason, open-loop entrainment via phase advances requires less recovery time despite the fact that a single pulse of light may induce a greater phase delay than advance.
More generally, we find that any given initial phase difference is more readily recovered if open-loop entrainment begins between CT0 and CT9; the rate of re-entrainment depends on the initial condition. To correct initial phase differences of 0 to 9 hours (by inducing a phase delay), daylight is most effective at the end of the day, suggesting greater performance if the algorithm were to begin control action around CT6. To correct initial phase differences of 0 to −6 hours (by inducing phase advances), daylight is most effective at the start of the day, suggesting greater performance if the algorithm were to begin around CT0. In the former case, daylight overlaps with the delay region of the subject's PRC, while in the latter it overlaps with the advance region. Resetting an initial condition of 12 to 15 hours, however, presents an interesting control dilemma as environmental daylight may induce both a phase delay and phase advance. For this reason, the open-loop control algorithm requires several days to correct for such phase differences. If light were accessible to entrain the system continuously throughout the day and night (in other words, if we were to close the loop), phase recovery dynamics would be less extreme since phase resetting would rely less on the initial condition.
Additional phase resetting properties may be inferred through investigation of the simulated PRCs. For instance, in the single input case, νs and ks exhibit similar recovery dynamics with the exception that νs is more effective at resetting initial phase differences of 15 to 21 hours. This quality may be associated with the fact that manipulating ks exhibits a strikingly similar phase response as νs where their input to output mapping is shifted by about 5 hours (Figure 1). This similarity may be attributed to the fact ks and νs are directly involved with the irreversible production, and transcriptional/translational regulation, of clock-specific genes/proteins. Additionally, the “active” region of the νs and νm PRCs are wider than those of ks and νd (or, their dead zones are shorter than those of ks and νd), suggesting that their perturbation-induced phase shifts are accessible throughout a greater portion of the circadian day.
Of the single control input results, the manipulation of νs, identified as the most sensitive parameter, provides the shortest phase recovery times. Despite these results, νd or light-based control is most efficient. In Figure 6, we relate the cumulative control input (a unitless measure that integrates the multiplicative control target action) to the convergence of phase via the PER-TIM complex state error. The data shown reflects the recovery of an initial phase difference of 15 hours from IC12. Analyzing this relationship may provide a basis from which the pharmaceutical industry might select one drug over another. If two different drug targets demonstrate similar response, the one that requires the least number of doses should be admitted, minimizing cost and the potential for drug related side-effects. Moreover, if the symptoms of illness are more severe than the potential for side effect, the drug that minimizes the state error may be preferred over others. The assessment of system convergence and the corresponding admitted control is key to the identification and application of control targets.
In our modern “24/7” work world, social and commercial pressures often oppose our natural circadian timekeeping, causing a source of circadian stress that may lead to chronic illnesses such as cardiovascular disease and cancer [24]. Numerous studies seem to show the effect of circadian rhythms on processes such as cell proliferation and apoptosis that eventually lead to proper growth control [25]–[27]. For instance, components of the cell cycle that dictate the G1-S and G2-M transition phase have been associated with circadian transcriptional regulation [28],[29]. Also in certain conditions, cancer can be a direct consequence of the absence of the circadian regulation [25],[26],[30]. A review of circadian related clinical disorders describes how mutations in some clock genes are associated with alcoholism, sleeping disorders, hypertension, and morbidity [24],[31]. Most commonly, poor circadian regulation leads to advanced sleep phase syndrome, delayed sleep phase syndrome, non-24-hour sleep-wake syndrome, and irregular sleep-wake pattern [32]. In each of these cases, poor circadian phase resetting may be achieved through the systematic admission of controlled light pulses.
Assuming we have access to drugs that specifically target circadian genes, we can identify the targets whose manipulation yields the most effective and immediate response through investigation of each control's phase dynamics (as shown in Figure 1). Or, it is possible to minimize the use of control and choose targets that require the least number of doses. We may also tailor the MPC algorithm to correct phase more readily through simultaneous manipulation of multiple control targets. Even further, we may reduce the computational expense by enumerating the control solutions over a grid in the solution space (light magnitude as a function of time), and choosing the optimal control sequence via an exhaustive search. The algorithm approaches a globally optimal solution as the total possible quantization steps of the control input increases. We tested the efficacy of the algorithm with respect to a quantization of 2, 4, 8, and 16 steps [12]. Results suggest that the shorter recovery time associated with the finer-grid enumeration may not outweigh the increase in computation time. Therefore, we may dramatically reduce computational expense by investigating control solutions for as few as 2 possible control values.
Our methods show great promise for use in the pharmaceutical industry as our theoretical phase entrainment of mutant phenotypes demonstrates the robustness of the algorithm in the presence of model mismatch. This robustness alleviates concerns in the pharmaceutical industry to tailor mathematical representations of bio-chemical pathways to individual people.
The study of controlled light pulses as a means of correcting phase is a common area of interest. Studies have shown that humans are much more sensitive to light than initially suspected since room light can significantly reset the phase of the human circadian clock [33],[34]. Furthermore, the admission of morning light has been considered as an antidepressant by realigning the internal clock with the environment [35].
Additional studies suggest that the human circadian clock mechanism functions similarly to those of other mammals [34]. This similarity may be attributed to shape/amplitude characteristics of their respective phase response curves. Humans show phase-delay shifts of up to 3.6 hours and phase-advance shifts of up to 2.01 hours (with respect to a 6.7 hour pulse of bright light) [36], which is both quantitatively and qualitatively similar to other mammalian species. This parallel motivates the experimental application of controlled light pulses for phase resetting in mammals. We have taken this first step by assessing the efficacy and computational utility of model predictive control as applied to a detailed 71-state Mus musculus circadian model [37]. Furthermore, melatonin has proven to be a key circadian phase resetting agent for totally blind people who cannot synchronize to environmental day∶night cycles (or do so at an abnormal time) [35]. Therefore, melatonin may be used individually (in cases to treat the totally blind), or in combination with light to provide more effective phase resetting.
Therapies designed to alleviate circadian load would have an important impact on morbidity and mortality across the developed world. Aside from correcting mutant phenotypes, phase resetting would increase performance in many healthy, or wild-type, cases such as frequent flyers avoiding jet-lag or astronauts maintaining a rigorous schedule during space exploration [17]. The real-time application of the proposed algorithm, however, may be a major issue; in practice, it will not be feasible to collect the corresponding protein concentration data at the molecular level. However, behavioral and/or physiological parameters that are controlled by (and correlated with) the circadian clock's dynamics are easily accessible. Such data may include actograms such as wheel running data for rodents [6]. Hence, a missing link in the current work concerns the development of corresponding (non-linear) state estimators for reconstructing the molecular dynamics. Given the discrete nature of MPC (sampling every 4 hours), the proposed strategy is feasible in practice since sampling rates of such physiological circadian markers may be much higher.
A 10-state Drosophila melanogaster circadian limit cycle oscillator serves as the model system. This model consists of two coupled auto-regulatory transcription/translation negative feedback loops that characterize period and timeless gene and protein dynamics [13]. As demonstrated in previous work, the MPC algorithm may be applied to any stable limit cycle oscillator, including a more complex Mus musculus model [4]. Thus, we describe the example system as a general set of nonlinear ordinary differential equations with time t, n-length state vector x(t), environmental light input l(t), additional control inputs u(t), and system dynamics f(x(t), l(t), u(t)):Given that both environmental light and additional control variables may be modeled as multiplicative inputs, the nominal wild-type (sun-cycle entrained) case requires u(t) = 1, while l(t) oscillates as a square wave with a frequency of 24 hours, between values 1 and 2. For consistency, the natural sun-cycle environment (or reference) is characterized by the nominal Drosophila melanogaster model and denoted by r(t). This reference is pre-entrained to normal 24 hour light∶dark cycles and is not subject to additional control inputs.
Model predictive control [38] is used to increase the re-synchronization or entrainment rate of circadian oscillators through the systematic application of specified control inputs. The algorithm follows a sample and hold strategy, updating the prediction and control input every ts = 2 hours, where the discrete time index , such that a function g(kts) = g[k]. For simplicity, we refer to k as being equivalent to and ignore its rounding component. The manipulated control variable, u[k], optimizes an open-loop performance objective on a time interval extending from the current time to the current time plus a prediction horizon of P = 48 hours, where . This horizon allows the algorithm to take control action at the current time in response to a forecasted error. The move horizon, M = 12 hours, limits the number of control inputs within the prediction horizon such that u[k] spans a time interval . Beyond hours of simulation, the predictive model defaults to u[k] = 1. Future behaviors for a variety of control inputs are computed according to the mathematical model of the system [13].
The efficacy of the algorithm was evaluated with respect to a sample and hold time interval of 1, 2, and 3 hours (reflecting a move horizon of 3, 6, and 9 hours, respectively). Although shorter light pulses offer a more dynamic manipulated variable profile, it shortens the move horizon and may reduce the utility of model predictive control. Conversely, a longer pulse may reduce the possible control profiles since extended exposure to light leads to arrhythmic behavior [39]. Thus, we set the sampling rate to 2 hours.
The fitness function penalizes the normalized predicted state error between the reference and controlled trajectories, ē[k], and the net control, ū[k], over the prediction horizon. The system output used to evaluate circadian performance (or, phase entrainment) is the trajectory defined by the total period and timeless protein complex concentrations. This state error, e[k], is normalized with respect to the nominal amplitude of oscillation while the time dependent control input, u[k], is normalized with respect to the nominal set of values, 1:where the state dynamics r[k] characterize the nominal reference. Note that the vector e[k] is , while the matrix ū[k] is m×c ( and c denotes the number of control inputs).
To avoid penalizing transient effects, the state error is weighted uniformly over the move horizon (reflected in the first m diagonal values of the p×p matrix Q), and with increasing weight of slope 2 over the prediction horizon (reflected in the p−m to p diagonal values of Q). The cost of applying a light input is weighted uniformly with a magnitude of 100 as reflected in the diagonal values of the m×m matrix R. We can afford to be conservative with the cost of control in the wild-type case, since we can ensure that the lack of control (the open-loop algorithm) will eventually entrain the system. The values contained in R will be re-evaluated when the algorithm is designed to entrain mutant phenotype models. The performance of an m-length control input is measured byOnly the first move of the lowest cost control sequence evaluated at time k, , is implemented. Therefore, the sequence of actually implemented control moves may differ significantly from the sequence of control moves calculated at a particular time step. This discrepancy disappears as the prediction and move horizons near infinity. Feedback is incorporated by using the next measurement to update the optimization problem. Once the controlled state trajectories converge to within 15% of the reference state trajectories, the system is considered to have recovered its phase in Tr = mink[|e[k]|∞≤0.15] hours. At this point, the algorithm defaults to no control since nominal light∶dark cycles will keep the system synchronized to the new environment. Optimization of the phase synchronizing control sequences is completed through use of a genetic algorithm [40]–[42].
Parametric sensitivity analysis quantifies the relative change of system behavior with respect to an isolated parametric perturbation. Parametric state sensitivity analysis assigns a value to each system parameter that defines how its perturbation affects state dynamics: . This tool is often used to identify the robustness and fragility tradeoffs of regulatory structures [3], and may be tailored to evaluate specific output performance such as period, amplitude, or phase characteristics [2].
Assuming the model has n states and ρ parameters, the FIM is a ρ×ρ matrix describing how any two parametric perturbations might affect state dynamics. More notably, the diagonal values of the FIM describe how any single parameter may affect state dynamics. As a result, we sort the values of the FIM from greatest to least magnitude and choose the top three individual parameters (reflected by the sorted diagonal values) and top three pairs of parameters (reflected by the sorted off-diagonals) whose perturbations yield the greatest change in output.
We further analyze the FIM via the singular value decomposition [43]. Assuming FIM = F, it may be decomposed as F = UΣVT, where Σ is an n by p diagonal matrix of non-negative singular values, σ, n is the number of states, and p is the total number of system parameters. Matrices U and V contain the eigenvectors of FFT and FTF, respectively. U, Σ, and V are ordered according to the magnitude of the singular values. Thus, the first column vector of U (and V) represents the output (and input) direction with largest amplification. The next most important direction is associated with the second column vector, and so forth. We determine the top three parameters associated with the three greatest input directions in ν1 and ν2 as ideal inputs for studying the multiple control input strategy. |
10.1371/journal.pntd.0006281 | Human T-Lymphotropic Virus type 1c subtype proviral loads, chronic lung disease and survival in a prospective cohort of Indigenous Australians | The Human T-Lymphotropic Virus type 1c subtype (HTLV-1c) is highly endemic to central Australia where the most frequent complication of HTLV-1 infection in Indigenous Australians is bronchiectasis. We carried out a prospective study to quantify the prognosis of HTLV-1c infection and chronic lung disease and the risk of death according to the HTLV-1c proviral load (pVL).
840 Indigenous adults (discharge diagnosis of bronchiectasis, 154) were recruited to a hospital-based prospective cohort. Baseline HTLV-1c pVL were determined and the results of chest computed tomography and clinical details reviewed. The odds of an association between HTLV-1 infection and bronchiectasis or bronchitis/bronchiolitis were calculated, and the impact of HTLV-1c pVL on the risk of death was measured.
Radiologically defined bronchiectasis and bronchitis/bronchiolitis were significantly more common among HTLV-1-infected subjects (adjusted odds ratio = 2.9; 95% CI, 2.0, 4.3). Median HTLV-1c pVL for subjects with airways inflammation was 16-fold higher than that of asymptomatic subjects. There were 151 deaths during 2,140 person-years of follow-up (maximum follow-up 8.13 years). Mortality rates were higher among subjects with HTLV-1c pVL ≥1000 copies per 105 peripheral blood leukocytes (log-rank χ2 (2df) = 6.63, p = 0.036) compared to those with lower HTLV-1c pVL or uninfected subjects. Excess mortality was largely due to bronchiectasis-related deaths (adjusted HR 4.31; 95% CI, 1.78, 10.42 versus uninfected).
Higher HTLV-1c pVL was strongly associated with radiologically defined airways inflammation and with death due to complications of bronchiectasis. An increased risk of death due to an HTLV-1 associated inflammatory disease has not been demonstrated previously. Our findings indicate that mortality associated with HTLV-1c infection may be higher than has been previously appreciated. Further prospective studies are needed to determine whether these results can be generalized to other HTLV-1 endemic areas.
| The Human T-Lymphotropic Virus type 1 (HTLV-1) infects up to 20 million people worldwide who predominantly reside in resource-limited areas. The virus is associated with a haematological malignancy (adult T-cell leukaemia/lymphoma, ATL), and inflammatory diseases involving organ systems including the spinal cord, eyes and lungs. Determining the outcomes of infection in most HTLV-1 endemic areas is extremely difficult; however, the virus is highly endemic to central Australia where the Indigenous population has access to sophisticated medical facilities. We prospectively followed a large hospital-based cohort of Indigenous Australian adults that was well characterized with regard to base-line comorbid conditions, HTLV-1 serostatus and HTLV-1 proviral load (pVL). A higher baseline HTLV-1 pVL was strongly associated with an increased risk of airway inflammation (bronchitis/bronchiolitis and bronchiectasis) and death, which most often resulted from complications of bronchiectasis. Increased mortality due to an HTLV-1-associated inflammatory condition has not been demonstrated previously. The morbidity and mortality associated with HTLV-1 infection may therefore be substantially higher than has been assumed from an analysis of cohorts of subjects with adult T-cell leukaemia or HTLV-1-associated myelopathy. These findings have important implications for epidemiological research and for determining health care priorities in resource-limited settings.
| The Human T-Lymphotropic Virus type 1 (HTLV-1) is an oncogenic retrovirus that preferentially infects CD4+ T cells[1]. Worldwide, HTLV-1 infects as many as 20 million people who predominantly dwell in areas of high endemicity in south-western Japan and developing countries of the Caribbean basin, South America and sub-Saharan Africa[2]. An endemic focus is present in central Australia[3] where more than 40% of Indigenous adults are HTLV-1c-infected in some remote communities[4].
Clinically significant sequelae of HTLV-1 infection include a haematological malignancy, Adult T cell Leukemia/Lymphoma (ATL), and inflammatory diseases, such as HTLV-1 associated myelopathy/tropical spastic paraparesis (HAM/TSP)[1]. In Japan and the Caribbean, life-time risks of HAM/TSP and ATL range between 0.3–4% and 1–5%, respectively[1]. Bronchiectasis is the most common clinical manifestation of HTLV-1 infection in Indigenous Australians, amongst whom the adult prevalence of this condition is the highest reported worldwide (>1%)[5,6]. Chest computed tomography has also revealed bronchiectasis in Japanese adults infected with HTLV-1; however, the most frequently reported radiological pattern of HTLV-1 associated pulmonary disease in this population is bronchitis/bronchiolitis[7,8], which has not been described in Indigenous Australians.
In endemic areas in Japan and Africa, HTLV-1 seropositivity is associated with increased mortality[9–12], which has been attributed to non-neoplastic conditions[9,10]. The interpretation of these studies is limited by their inability to control for clinically defined comorbid conditions that might independently increase mortality[9,11,12] [10]. For example, HTLV-1 seropositivity had no effect on mortality in a large hospital-based cohort of Indigenous Australian adults after adjusting for other medical conditions[13]. Given the close association between the number of HTLV-1-infected cells in peripheral blood (the HTLV-1 proviral load, pVL) and serious HTLV-1 associated complications[1,14], any influence of HTLV-1 infection on mortality might be revealed by stratifying outcomes according to HTLV-1 pVL. In central Australia, Indigenous adults with higher HTLV-1c pVL have more extensive, radiologically defined pulmonary injury[6] and are more likely to present with life-threatening bacterial infections[15]. A single, small study in Guinea-Bissau, where causes of death could not be ascertained, found that mortality increased with HTLV-1 pVL[16]. The present study was therefore commenced to quantify the prognosis of HTLV-1c infection and chronic lung disease and the risk of death according to the HTLV-1c pVL in a hospital-based cohort of Indigenous adults who were well characterized with regard to comorbid conditions and for whom causes of death could be accurately determined in nearly all cases.
Alice Springs Hospital (ASH) is the only medical facility serving central Australia, an area of >1,000,000 km2. Critically ill patients are transferred by air to ASH, which has sophisticated diagnostic capabilities.
All Indigenous patients aged >15 years with a discharge diagnosis of bronchiectasis, 1st June 2008 to 31st December 2013, were identified from the ASH patient management database, which coordinates all in-patient and out-patient hospital activities. Indigenous status was determined from self-reported data obtained at admission, as recorded in the patient information database. Potential subjects were offered enrolment when next admitted for >48 hours. Among 165 eligible cases, 154 were recruited (eleven subjects left hospital before recruitment was possible). Written reports for chest high-resolution computed tomography (cHRCT) were reviewed for all subjects, confirming bronchiectasis in 104 cases and bronchitis with or without bronchiolitis in 33 cases (bronchitis alone, 20; bronchitis and bronchiolitis, 12; bronchiolitis alone, 1)(Fig 1). Patients with chronic pulmonary disease were treated according to local guidelines which includes antibiotic therapy for infective exacerbations [17]. A further 686 Indigenous patients aged >15 years who were admitted for >48 hours were prospectively recruited during the same period. These control subjects had no evidence of lower respiratory tract infection at the time of recruitment, no recorded discharge diagnosis of bronchiectasis, and no clinical or radiological evidence of bronchiectasis, Research team members who were unaware of HTLV-1 serostatus were responsible for recruitment (Fig 1).
Demographic and clinical details were extracted from medical records at the time of recruitment using a standardized data-collection form. HTLV-1 associated conditions were identified from medical records at baseline and study end. No control patient developed chronic pulmonary disease during the study period. Mortality data was obtained at study end from the ASH patient management database, and the cause of death was determined from death certificates held in Registries in the Northern Territory of Australia and South Australia. Death certificates were not available for four subjects who died in remote communities in Western Australia, for whom a cause of death was sought from the responsible remote clinic.
Bronchitis was diagnosed where cHRCT revealed bronchial wall thickening or dilatation not fulfilling criteria for bronchiectasis, and bronchiolitis where cHRCT revealed multiple centrilobular nodules or a ‘tree-in-bud’ pattern[8]. Chronic obstructive pulmonary disease (COPD) required a clinical diagnosis in the medical record and appropriate chest X-ray findings. Emphysema without bronchial wall injury or bronchiolitis was recorded for 18 subjects with COPD examined by cHRCT. Chest HRCT was not performed on 12 subjects who did not meet criteria for such imaging[17]. No subject with symptoms consistent with HAM/TSP received lumbar puncture; the diagnosis was therefore considered ‘probable’ in all cases. Asymptomatic HTLV-1-infected subjects were those without radiological evidence of airway inflammation or recognized HTLV-1 associated conditions [1]. Residence >80 km from the township of Alice Springs was defined as remote.
The study was approved by the Central Australian Human Research Ethics Committee. All patients, and their parents/guardians if aged <18 years, gave written informed consent in primary languages.
Whole blood samples were collected from each participant at the time of recruitment. Peripheral blood buffy coats (PBBC) were prepared, and plasma and PBBCs were stored at ASH at -80° C until transfer to the National Serology Reference Laboratory, Melbourne. Samples were screened for antibodies to HTLV-1 using both an enzyme immunoassay (Murex HTLV-I + II, DiaSorin, Italy) and a particle agglutination assay (Serodia HTLV-1, Fujirebio, Tokyo, Japan). Any sample reactive on either screening assay was tested by Western blot (HTLV-I/II Blot2.4, MP Biomedicals Asia Pacific Pte. Ltd., Singapore) and HTLV-1c PCR. Primers and fluorescently labelled hydrolysis probes were designed to target a highly conserved 88 bp fragment of the gag gene in the p19 coding region of the Australo-Melanesian HTLV-1 subtype C[18] and multiplexed with primers and probes to the albumin gene[19]. SP cells were used to generate a standard curve from which HTLV-1 pVL (copies per 105 peripheral blood leukocytes; PBL) was calculated. Samples and standards were extracted using the Qiagen QIA blood Mini Extraction kit and the extracts amplified on a Stratagene Mx3000p Real Time PCR Instrument (Integrated Sciences). The extract (5 μL) was added to 20 μL of Master mix containing 2 x Brilliant Multiplex QPCR Master Mix (Agilent Technologies) 0.3 μM of each primer (Gene works) and 0.16 μM of each probe (Sigma-Aldrich) and amplified at 95°C for 10 minutes, 45 cycles at 95°C for 30 seconds, 65°C for 60 seconds and 72°C for 60 seconds.
The clonality of HTLV-1-infected PBLs was determined by high-throughput sequencing of PBBC cell genomic DNA. The oligoclonality index (OCI) was calculated as previously described[20], and then adjusted to limit underestimation of the OCI due to the small observed number of proviruses[21]. The OCI provides a measure of the non-uniformity of the clone abundance distribution of the infected cell population: OCI = 1 indicates perfect monoclonality (only one clone constitutes the total proviral load); OCI = 0 indicates perfect polyclonality (all clones have the same abundance)[20]. Samples were selected for clonality analysis if subjects had HTLV-1 pVL >100 copies per 105 PBL, were HBsAg negative and strongyloides seronegative. Although 53 subjects met these criteria, technical difficulties prevented analysis for nine subjects (inadequate number of unique integration sites to accurately determine OCI, 7; unable to sequence integration site, 2). The OCI was therefore compared between 29 asymptomatic and 15 symptomatic subjects (bronchiectasis, 10; bronchitis/bronchiolitis, 3; uveitis, 2).
All analysis was performed using Stata version 14.2 (StataCorp, College Station, USA). HTLV-1 pVL was log-transformed and also categorized as low if <1000 and high if ≥1000 per 105 PBL, a cut-off that has been associated with an increased risk of HAM/TSP[22]. Differences between subjects who were HTLV-1 uninfected, those with low HTLV-1 pVL, and those with high HTLV-1 pVL were assessed using ANOVA for continuous variables and chi-squared tests for categorical variables. For statistical purposes, causes of death were grouped into six non-overlapping categories: bronchiectasis, sepsis, cardiovascular disease, malignancy, chronic kidney disease and chronic liver disease. We used survival analysis to determine the association between HTLV-1 pVL and both overall and cause-specific mortality. Subjects were followed until either date of death or 30th March 2015. The association between overall mortality and HTLV-1 pVL was assessed using log-rank tests and Kaplan-Meier curves in univariate analysis and using Cox regression for multivariate analysis. Associations with cause-specific mortality were assessed using competing risks analysis with all causes except the specific cause of interest treated as a competing risk. Where HTLV-1 pVL was treated as a categorical variable we also tested for a trend by creating a continuous variable with value zero for those uninfected, and with the median value of HTLV-1 pVL for those in the low and high pVL categories. Predictors of chronic airways inflammation were assessed using multivariate binary logistic regression. A 2-sided Type 1 error rate of p<0.05 was regarded as indicating statistical significance in each analysis.
Demographic and clinical characteristics are presented in Table 1. Among 840 subjects recruited to the study, 307 (36.5%) were HTLV-1 infected (HTLV-1 Western blot positive, 268; Western blot indeterminate/HTLV-1c PCR positive, 39).
Radiologically defined airways inflammation was more common among HTLV-1c-infected subjects (Table 1). Bronchiectasis was confirmed in 59/307 (19.2%) HTLV-1c-infected subjects and 45/533 (8.4%) HTLV-1c uninfected subjects. Similarly, cHRCT revealed bronchitis/bronchiolitis in 21/307 (6.8%) HTLV-1c-infected subjects and 12/533 (2.3%) who were HTLV-1c uninfected (Table 1).
Compared to the median HTLV-1c pVL of asymptomatic subjects (n = 208, 30.4 copies per 105 PBL (min, 0.01; max, 18600), those for subjects with bronchiectasis (n = 59, 494 copies per 105 PBL; min, 0.01; max, 87900) (p = 0.001) and bronchitis/bronchiolitis (n = 21, 486 copies per 105 PBL; min, 0.01; max, 70200)(p = 0.042) were 16-fold higher,(Fig 2). Median HTLV-1c pVL of subjects with any airways inflammation (bronchiectasis and bronchitis/bronchiolitis; 490 copies per 105 PBL; min, 0.01; max 87900) was 16-fold higher than that of asymptomatic subjects (p<0.001). Few other recognised causes of bronchiectasis were found (Table 2).
In a multivariate model that controlled for demographic factors, smoking and harmful alcohol consumption, HTLV-1 infection increased the risk of any airways inflammation 2.9-fold (p<0.001) (Table 3), while HTLV-1c pVL ≥1000 copies per 105 PBL increased the risk 2.2-fold among HTLV-1-infected subjects (p = 0.006) (Table 4).
Other HTLV-1 associated conditions included infective dermatitis (4), crusted scabies (4), probable HAM/TSP (2) and uveitis (2). Four subjects with HTLV-1-associated bronchiectasis had other sequelae of HTLV-1 infection (infective dermatitis, 2; HAM/TSP, 1; uveitis, 1). There was no difference in strongyloides seropositivity between groups (Table 1), nor was there any difference in log-transformed HTLV-1c pVL using linear regression (p = 0.409); median (IQR) HTLV-1c pVL copies per 105 PBL for subjects who were strongyloides seronegative (133; IQR 2.1, 1387), seropositive (61; IQR 1, 1039) and those with equivocal strongyloides serological results (190; IQR 2.7, 2473).
Although the risk of malignancy was not increased among HTLV-1c infected subjects, one subject (HTLV-1c pVL, 5500 copies per 105 PBL) developed ATL during follow-up, one developed penile cancer (HTLV-1c pVL, 70200 copies per 105 PBL) and another metastatic anal cancer (HTLV-1c pVL, 28000 copies per 105 PBL).
The median OCI did not differ between asymptomatic (0.401; IQR 0.350, 0.485) and symptomatic groups (0.411; IQR 0.355, 0.552) (p = 0.51) or when symptomatic subjects with uveitis were excluded from the analysis (symptomatic group median OCI 0.429; IQR 0.369, 0.552) (p = 0.341). Median log10 HTLV-1c pVL copies per 105 PBL in subjects selected for clonality analysis did not differ between asymptomatic subjects (0.947; IQR 0.342, 1.964) and those with airways inflammation (1.50; IQR 0.669, 4.074)(p = 0.20).
During 2140 person-years of follow-up, 155 deaths were recorded (HTLV-1 uninfected, 85; HTLV-1 infected, 70). Non-bronchiectasis causes of death were infections (37)(lower respiratory tract infections, 15), cardiovascular disease (37), malignancy (15), end-stage kidney disease (ESKD) (15), chronic liver disease (10), intracerebral haemorrhage (4), primary pulmonary hypertension (2) and amyloidosis (1).
Subjects with high HTLV-1c pVL were more likely to die during the study period (Fig 3). Mortality rates for high HTLV-1c pVL, low HTLV-1c pVL and HTLV-1 uninfected subjects were 28.4% (27/95), 20.2% (43/212) and 15.9% (85/533), respectively (p = 0.011). The unadjusted HR for death among subjects with low and high HTLV-1c pVL were 1.24 (95% CI, 0.85, 1.81) and 1.75 (95% CI, 1.13, 2.670), respectively (p = 0.021 for trend). The statistical significance was diminished after adjusting for age, gender, place of residence and harmful alcohol consumption (p = 0.084). The effect of HTLV-1c pVL on overall mortality was lost in a multivariate model that included bronchiectasis (aHR, 1.045; 95% CI, 0.658–1.660) (Table 5). Other predictors of death were age at test, male gender and comorbid conditions (Table 5).
In a large hospital-based cohort of Indigenous Australian adults, a higher baseline HTLV-1c pVL prospectively predicted a bronchiectasis-related death, which occurred at a mean age of only 49.5 years. In addition to confirming a previously reported association between HTLV-1c infection and bronchiectasis[5], the present study also revealed an association with bronchitis/bronchiolitis. Airways inflammation was strongly associated with higher HTLV-1c pVL[6,15]. The median HTLV-1c pVL of subjects with radiologically defined airways inflammation was 16-fold higher than that for asymptomatic HTLV-1-infected subjects, and risk of airway inflammation increased three-fold among subjects with higher HTLV-1c pVL in an adjusted model.
HTLV-1 associated inflammatory diseases are thought to result from a genetically determined, inefficient cytotoxic T lymphocyte response, permitting widespread dissemination of the virus in a large number of HTLV-1-infected T-cell clones, which is reflected in a high HTLV-1 pVL[23]. Organ infiltration by HTLV-1-infected lymphocytes then leads to high local HTLV-1 antigen levels, provoking an immune response and tissue injury following the release of pro-inflammatory cytokines and chemokines[23]. Although this has been best studied for the prototypical HTLV-1 associated disease, HAM/TSP, HTLV-1 infection is also associated with inflammation in other organs[14] including the lungs [24]. Consistent with the presumed mechanism of pathogenesis of HAM/TSP [23], pulmonary involvement is associated with infiltration of HTLV-1-infected lymphocytes[25,26], increased tax/rex mRNA[27] expression, and an inflammatory cytokine milieu in bronchoalveolar lavage fluid[27]. In large Japanese case series, cHRCT was abnormal in 30–61% of HTLV-1 infected subjects of whom 23.6–29.5% had a bronchitis/bronchiolitis pattern of disease and 15.6–22.5% had frank bronchiectasis[7,8]. The pathological correlate of these observations is lymphocyte infiltration in bronchiole walls[7]. Persistent HTLV-1-mediated airways inflammation may therefore lead to progressive bronchial wall dilatation, and bronchiectasis. High rates of bronchiectasis among HTLV-1-infected Japanese adults[7,8], and associations between HAM/TSP and bronchiectasis in UK [28] and Brazilian cohorts [29] suggest that HTLV-1-associated bronchiectasis affects individuals of diverse genetic backgrounds infected with HTLV-1 strains other than HTLV-1c. Although HTLV-1 associated pulmonary disease in Japan is thought to be largely sub-clinical[14], published clinical details are limited and prospective survival studies have not been performed.
Consistent with other HTLV-1 associated inflammatory diseases[1,14], the median baseline HTLV-1c pVL of subjects with airways inflammation was substantially higher than that of asymptomatic subjects. For example, the median HTLV-1c pVL in subjects with HAM/TSP is between 7-fold and 16-fold greater than that of asymptomatic subjects[22][28][30–32]. Among subjects with HAM/TSP, higher HTLV-1 pVL correlates with more severe motor weakness[31] and more rapid neurological progression[32]. We previously demonstrated that HTLV-1-infected Indigenous adults have more diffuse bronchiectasis[5] and that a higher HTLV-1c pVL correlates with more extensive pulmonary injury[6]. In contrast to HTLV-1-mediated inflammation in other tissues, pulmonary parenchymal injury can result in directly life-threatening complications, including respiratory failure[5]. Among subjects with airways disease in whom HTLV-1 oligoclonality could be studied, there was no difference in the median OCI when compared to that of asymptomatic subjects. This suggests that higher HTLV-1c pVL were due to an increased number of infected clones rather than clonal expansion, which is consistent with the conclusion previously reported for subjects with HAM/TSP[20]
Increased mortality due to a specific HTLV-1-associated inflammatory disease has not been prospectively demonstrated previously. However, an excess mortality that is not attributable to ATL or currently recognized HTLV-1-associated inflammatory diseases has been reported in other endemic areas. Adjusted hazard ratios of death are 1.3[10] to 1.77–1.87[9] in Japanese outpatient cohorts, and 3.8 and 2.3 for young and middle-aged adults, respectively, in a community-based cohort in Guinea-Bissau[11,12]. In a study that included only 48 HTLV-1-infected subjects in Guinea-Bissau, mortality was associated with higher HTLV-1 pVL[16]. In Japan, excess mortality was attributed to non-neoplastic diseases, most commonly unspecified kidney and cardiac conditions[9,10]. Although the ASH cohort included subjects with established ESKD and heart disease, HTLV-1 infection was only associated with bronchiectasis-related deaths, and this effect was only revealed after stratifying by HTLV-1c pVL. The difference between studies in the clinical conditions associated with excess mortality may reflect the high burden of illness in our hospital-based cohort, the inability to control for comorbid conditions in other studies, and differences in the social circumstances of the various study populations.
The strengths of this prospective cohort study include the recruitment of nearly all eligible subjects with a discharge diagnosis of bronchiectasis, the use of cHRCT for diagnosis and the blinding of ASH researchers to the HTLV-1 serostatus of subjects and of those who performed HTLV-1 studies to their clinical state. Nevertheless, some design limitations must be recognized. First, investigations to exclude other causes of airways disease could not be performed in all cases. However, consistent with previous studies[5,6], a specific aetiology was rarely found among >70% of subjects who were screened for conditions generally associated with bronchiectasis[17]. Although an effect of childhood respiratory infections cannot be excluded in the present study, we previously found HTLV-1 infection to be the major predictor of adult bronchiectasis in a case-control study that controlled for such infections[6]. Second, only subjects with a discharge diagnosis of bronchiectasis were specifically targeted for recruitment. Twelve subjects with COPD (HTLV-1 infected, 4; HTLV-1 uninfected, 8) were incidentally recruited but not examined by cHRCT because they did not clinically warrant further imaging[17]. The contribution of HTLV-1c infection to less severe respiratory disease than that associated with a discharge diagnosis of bronchiectasis, and the validity of our conclusions in a community setting, require further study. Finally, the absence of an association between strongyloides seropositivity and higher HTLV-1c pVL may be due to the fact that strongyloides serology was assayed in subjects without symptomatic strongyloidiasis.
In summary, HTLV-1c infection and higher HTLV-1c pVL were strongly linked to airways inflammation in a hospital-based cohort of Indigenous Australian adults. Furthermore, higher baseline HTLV-1c pVL prospectively predicted death due to bronchiectasis, which may result from more extensive disease[5,6], predisposing to life-threatening complications[5] among subjects who are unable to control HTLV-1 replication[6]. Elucidating the causes of higher mortality among people infected with HTLV-1 is relevant to an estimated 20 million people living with HTLV-1 infection in resource-poor areas[2] where the impact of HTLV-1 infection has been little studied.
|
10.1371/journal.pcbi.1004901 | Using Combined Diagnostic Test Results to Hindcast Trends of Infection from Cross-Sectional Data | Infectious disease surveillance is key to limiting the consequences from infectious pathogens and maintaining animal and public health. Following the detection of a disease outbreak, a response in proportion to the severity of the outbreak is required. It is thus critical to obtain accurate information concerning the origin of the outbreak and its forward trajectory. However, there is often a lack of situational awareness that may lead to over- or under-reaction. There is a widening range of tests available for detecting pathogens, with typically different temporal characteristics, e.g. in terms of when peak test response occurs relative to time of exposure. We have developed a statistical framework that combines response level data from multiple diagnostic tests and is able to ‘hindcast’ (infer the historical trend of) an infectious disease epidemic. Assuming diagnostic test data from a cross-sectional sample of individuals infected with a pathogen during an outbreak, we use a Bayesian Markov Chain Monte Carlo (MCMC) approach to estimate time of exposure, and the overall epidemic trend in the population prior to the time of sampling. We evaluate the performance of this statistical framework on simulated data from epidemic trend curves and show that we can recover the parameter values of those trends. We also apply the framework to epidemic trend curves taken from two historical outbreaks: a bluetongue outbreak in cattle, and a whooping cough outbreak in humans. Together, these results show that hindcasting can estimate the time since infection for individuals and provide accurate estimates of epidemic trends, and can be used to distinguish whether an outbreak is increasing or past its peak. We conclude that if temporal characteristics of diagnostics are known, it is possible to recover epidemic trends of both human and animal pathogens from cross-sectional data collected at a single point in time.
| We have developed a Bayesian approach that can estimate the historic trend of incidence from cross-sectional samples, without relying on ongoing surveillance. This could be used to evaluate changing disease trends, or to inform outbreak responses. We combine two or more diagnostic tests to estimate the time since infection for the individual, and the historic incidence trend in the population as a whole. We evaluate this procedure by applying it to simulated data from synthetic epidemics. Further, we evaluate its real-world applicability by applying it to two scenarios modelled after the UK 2007 bluetongue epidemic, and a small outbreak of whooping cough in Wisconsin, USA. We were able to recover the epidemic trends under a range of conditions using sample sizes of 30–100 individuals. In the scenarios modelled after real-world epidemics, the hindcasted epidemic curves would have provided valuable information about the distribution of infections. The described approach is generic, and applicable to a wide range of human, livestock and wildlife diseases. It can estimate trends in settings for which this is not possible using current methods, including for diseases or regions lacking in surveillance; recover the pattern of spread during the initial “silent” phase once an outbreak is detected; and can be used track emerging infections. Being able to estimate the past trends of diseases from single cross-sectional studies has far-reaching consequences for the design and practice of disease surveillance in all contexts.
| Infectious disease surveillance is the first line of detection and defence against infectious pathogens and therefore crucial to maintaining animal and public health. However, the current state of disease surveillance has been characterised as deficient in terms of both coverage and reporting speed for both humans [1] and animals [2,3]. The challenge is to use the data generated by this often sparse and biased surveillance to decide on an appropriate response to disease outbreaks. This is dependent on the extent of situational awareness, which can be defined as “Knowledge and understanding of the current situation which promotes timely, relevant, and accurate assessment … in order to facilitate decision making.” (taken from [4], p 171). Such situational awareness is necessary in order to balance the social and economic consequences of the adopted control strategy with the social and economic risks posed by the outbreak [5].
Limited situational awareness can have substantial negative impact. In the case of the pandemic H1N1 flu in 2009, early analyses mistakenly assumed that the epidemic had been only recently introduced, causing substantial overestimates of the basic reproduction ratio [6] and case fatality rates [7] that suggested a far greater risk to human life than was actually the case, leading to a more resource-intensive response than was necessary [8]. The more complex settings typical of livestock and particularly wildlife systems tend to result in the available surveillance data being sparser still for animal diseases [9].
Adding missing information on the time of exposure of detected cases would allow for a better awareness of the early development of an epidemic and would help inform evaluations of the potential risks posed by an outbreak, leading to a more proportionate response than would be the case when waiting for the epidemic trends to be revealed by subsequent real-time monitoring. In the current study, we introduce a novel statistical approach to infer the timing of exposure events for individuals by combining knowledge of the dynamic characteristics of multiple diagnostic tests. This approach could be integrated into any model of a disease epidemic to replace missing information on case exposure times. In this paper, we demonstrate its usefulness by recovering population-level trends of exposure from cross-sectional data collected from a single point in time. Here we refer to the process of recovering such trends as “hindcasting”, following terminology established in other papers [10–12] for reconstructing historical trends from currently available data.
Disease surveillance has been described [13] as improving the situational awareness in relation to a disease outbreak on three levels: Perception, Comprehension, and Projection. Perception refers to the collection of data that allows us to monitor disease; Comprehension to extracting information from this raw data that places the current disease situation in a context that allows us to understand its characteristics; and Projection to statistical models as well as more holistic approaches that aim to describe what is likely to happen in the future. Research focused on improving the collection of surveillance data [14–16], on risk-based surveillance [17,18], or the extensive literature focusing on the early detection of statistical deviations in surveillance data to outbreaks [19–21], can be seen as improving the Perception stage. Approaches such as phylodynamics contribute to the Comprehension stage by modelling the genetic change of the pathogen, e.g. using this to estimate the epidemiological parameters governing an outbreak such as the recent Ebola outbreak [22–24]. Models that use current information to predict the future [25–27] instead focus on improving the situational awareness at the Projection stage. From this perspective, hindcasting contributes to the Comprehension stage by leveraging quantitative diagnostic test results (using the statistical methods described in this paper) to add a temporal dimension to data for which the times of exposure of cases are missing, thus improving the understanding of unfolding epidemics.
Several papers have recovered limited historical characteristics of epidemics from cross-sectional data using a single diagnostic test, e.g. an antibody test. For example, Giorgi et al. estimated the time of the start of an HIV outbreak under assumptions of exponential growth of viral load [28]. Others have exploited information on diagnostic test kinetics, i.e., the pattern of diagnostic test values during the course of infection, to estimate average incidence rates. Examples include the use of antibody test kinetics to estimate sero-incidence rates for influenza [29], Salmonella in cattle [30] and Salmonella in humans [31]. One challenge in these kinds of studies is that the relationship between the magnitude of signals from diagnostic tests and time since exposure is usually not monotonic; they tend to increase and then decrease. This means that the inverse problem of estimating time since exposure given a test value is non-unique, and although this can be framed as a statistical problem the resulting inference is highly uncertain [28,32], limiting what can be estimated from test data. However, there are often several diagnostic tests available that target different aspects of the multi-faceted dynamic interaction between host and pathogen, and thus exhibit different test kinetics [33]. That is, the profile of test responses, as a function of time since exposure, will differ depending on the underlying diagnostic used and the immune-pathogenesis of the disease. Thus, in principle we can generate a unique signal for a given time since exposure by combining results of diagnostic tests that respond on different time scales. Here, we exploit this fact to develop a more robust statistical approach for analysing cross-sectional field data from multiple diagnostic tests. To do so we make use of empirical infection models that characterise test kinetics to infer the time since exposure for each individual. While there is considerable uncertainty in the estimated exposure time for each individual, the combined estimates from multiple individuals can be used to describe the overall population-level distribution of infection times and estimate the shape of the overall epidemic trend with a high level of confidence.
A detailed description of the hindcasting framework and implementation of the evaluation scenarios can be found in the methods section. We demonstrate the hindcasting of epidemic trends by applying the framework developed here to case studies of real outbreaks of two contrasting diseases, whooping cough in humans and bluetongue in cattle (see Fig 1). For each disease, we investigate two scenarios representing detection during either the increasing or the decreasing phase of the epidemic. We conclude that when combined with knowledge of the temporal characteristics of two (or more) appropriate diagnostic tests, our methods allow historical epidemic trends to be recovered from cross-sectional sample data. Moreover, for the example diseases considered suitable diagnostic tests and data describing their temporal characteristics already exist.
We evaluated the hindcasting framework by applying it to simulated data sets, and comparing the recovered trend with the known underlying distribution. We first generated collections of exposure times from lognormal probability distributions with different sets of parameters. Using published test kinetics for whooping cough, we then generated diagnostic test data based on these infection times and an assumed cross-sectional sampling time. The hindcasting framework was applied to these generated test data, and the estimated posterior distribution for the epidemic trends was compared to the known simulated epidemic trend. In order to explore the real-world applicability of this approach, we also simulated diagnostic test data using published test kinetics and published distributions of case reporting times, from an outbreak of bluetongue in the UK in 2007, and from an outbreak of whooping cough in Wisconsin in 2003.
We generate exposure times from four different lognormal distributions, each representing a different epidemic scenario as follows:
Epi1∼logN(log(μ)=log(2),log(σ)=log(5))
Epi2∼logN(log(μ)=log(4),log(σ)=log(5))
Epi3∼logN(log(μ)=log(20),log(σ)=log(2))
Epi4∼logN(log(μ)=log(50),log(σ)=log(2))
where the above notation means that the exposure times in each epidemic are drawn from the corresponding log-normal distribution. These represents epidemics peaking 2, 4, 20, and 50 days before the time of sampling, with the relative standard deviation chosen to provide more and more gradual increasing trends.
We found that we could reliably recover the epidemic trends when using sample sizes of 30 or more, and with levels of test variability less than 1.5, and that the estimated trend showed better fit when the peak was less recent than if it had just occurred (likely due to a difficulty in resolving very rapid dynamics relative to diagnostic test characteristics).
Fig 2 shows the hindcasted trends for 320 simulations that were conducted with a sample size of between 30 and 100, with a test variability of 1.3, evenly split across the four parameterizations. As can be seen, these trends all manage to adequately capture the timing and duration of the true epidemic, with a clear separation between the estimates for different sets of true parameter values.
Turning to summary statistics of the epidemic fit across these sets of posterior mean trends, the median R2 (and root mean squared error of prediction—RMSEP) for the Epi1 parameterisation was 0.71, with a 95% inter-quantile range (IQR) of 0.19–0.97 (RMSEP of 0.054[0.021–0.109]), a median of 0.85 with a 95% IQR of 0.48–0.99 (RMSEP of 0.019[0.005–0.041]) for the Epi2 parameterisation, a median of 0.96 with a 95%IQR of 0.64–0.998 (RMSEP of 0.005[0.01–0.020]) for the Epi3 parameterisation, and a median R2 of 0.97 with a 95% IQR of 0.69–0.999 (RMSEP 0.002[0–0.007]) for the Epi4 parameterisation.
Fig 3 shows the relationship between sample size and estimation performance. As can be seen, increasing the sample size improved the performance as measured with R2 for all of the parameterizations except Epi1 (fitting Epi1 was limited by the time resolution of the diagnostic test kinetics used). The posterior credible intervals for the parameters of the epidemic also shrunk in width, as would be expected. The performance when hindcasting using sample sizes of 10 was not very reliable; however, for sample sizes of 30 or more, the recovered trends reliably represented the true trend, with R2 values of 0.75 or more for all parameterizations except Epi1.
In order to evaluate the robustness of the hindcasting framework, we explored a range of testing errors from 1.1 up to 2.0 (multiplicative standard deviation). We found that that the width of the credible intervals increased moderately with increasing variability, but that the recovered parameters exhibited similar levels of bias regardless of the level of test variability. This was true even for testing errors of as high as 2.0, far beyond the reported variability of the examined diagnostic tests for bluetongue and whooping cough. (See S1 Text for full plots regarding the relationship between test variability and performance.)
Finally, we investigated the effect of violating the assumption of conditional independence of antibody and nucleic acid tests. Changing the amount of correlation from 0 up to 1 in 0.25-unit intervals showed no detectable difference in the results, whether measured with R2, RMSEP, or parameter estimates. (See S1 Text for related figures.)
We also applied the hindcasting framework to two case studies based on a recorded outbreak of whooping cough in humans, and a bluetongue outbreak in cattle (see Fig 1). See the methods section for details. For each outbreak we simulated two scenarios, firstly where a random subset of all individuals exposed thus far was sampled and tested at a single time, midway through the outbreak (increasing epidemic trend/early detection), and in a second scenario where a random subset of all exposed cases were sampled and tested at a time point at the end of the outbreak (decreasing epidemic trend/late detection). We assumed that no information about the time since exposure was available, nor any other information about the epidemic trend. Based on published temporal characteristics of real diagnostics, test results were then simulated for these samples (see Methods and Fig 4). For each disease (whooping cough and bluetongue) and each scenario (increasing and decreasing outbreaks) the hindcasting framework was applied to the corresponding test results to assess performance in recovering early increasing phases and late decreasing phases of outbreaks.
The results show that the recovered epidemic trends provided a representative picture for both increasing and decreasing scenarios, in both whooping cough and bluetongue outbreaks (Fig 1). For the increasing whooping cough epidemic, when assuming a sample of all 122 cases that had occurred between the start of the epidemic up to week 25, the R2 between underlying case counts (smoothed by a 7-day moving average) and the estimated epidemic trends was 0.74, with a 95% confidence interval of [0.69–0.78]. The corresponding RMSEP was 0.0017[0.0015–0.0018] When sampling 230 cases from the full whooping cough epidemic up until week 36, after it had declined, the curve fit was somewhat better, with R2 of 0.82[0.68–0.94] (RMSEP 0.0013[0.0008–0.0017]).
The results from hindcasting the bluetongue outbreak indicated that when assuming that a sample of the 26 animals had occurred during the increasing phase, the fitted curve was nearly perfect, with an R2 of 0.9[0.86–0.92] (RMSEP 0.0019[0.0018–0.002]). However, for the corresponding decreasing scenario, assuming a sample of the 61 animal cases that had occurred up to week seven, the hindcast trend could not fully capture the erratic nature of the underlying case count data, as indicated by R2 values of 0.21[0.15–0.27]). The trend did indicate an elevated incidence over the stretch of time when the majority of cases occurred, thus capturing the approximate time that had elapsed between the start of the epidemic and the time of sampling. This aspect is also captured by the RMSEP, which is more sensitive to shifts in locations, and less sensitive to upward- and downward trends, and which slightly improved to 0.0016[0.0015–0.0016].
When reducing the sample size, the hindcasting technique was still able to recover both increasing and decreasing phase for the whooping cough scenarios. The good fit was maintained with sample sizes as low as 20 individuals, with R2 values of 0.77[0.27–0.83] (RMSEP 0.0064[0.0055–0.0124]), for the increasing and 0.67[0.09–0.86] (RMSEP 0.0038[0.0022–0.0069]) for the decreasing scenario. The performance was also maintained for the increasing bluetongue scenario, also assuming 20 samples, with an R2 of 0.91[0.87–0.93]) (RMSEP 0.042[0.034–0.051]). However, the full bluetongue scenario performed substantially worse with the reduced sample size with an R2 of 0.12[0.07–0.36] (RMSEP 0.014[0.013–0.015]).
We further investigated how the hindcasting framework would be affected by different combinations of test kinetics. Fig 5 shows the mean likelihood surface of true vs. posterior times of exposure when using antibody-based tests, nucleic-based tests, or a combination, for the whooping cough and bluetongue exemplar diseases. For the hindcasting framework the ideal combination of diagnostic tests would have a likelihood surface with a single narrow diagonal representing a maximum likelihood coinciding with the true exposure time given observed data. A likelihood surface with a more diffuse diagonal implies a wider posterior distribution. A likelihood surface where there is an “X” of high-likelihood regions implies that the true exposure times are not uniquely identifiable.
Each pixel represents the average likelihood of 10 observations from the distribution of test measurements at a time since exposure given by the X axis, calculated at a time of exposure given by the Y axis. Areas in dark red indicate regions of higher likelihood. The times shown are times since exposure, with low numbers indicating more recent exposure, where the test response is changing rapidly. Looking at Fig 5a, during the first 20 days the probable exposure times (red pixels), given the data, are centred on the diagonal (i.e. the true exposure times) with a narrow band of high-probability red pixels. For times since exposure of greater than 20 days, when the kinetics of the antibody test are developing at a slower pace, the diagonal of red pixels becomes more diffuse, indicating a greater variation around the true times since exposure. Furthermore, we can see that there are two different diagonals crossing at 25 days. This corresponds to the peak of the diagnostic response curve, with the two diagonals indicating the possibility that a given test result could have been the result of testing an individual during either the increasing or the decreasing phase of the response curve. Estimation of the time since exposure is more precise when the true time since exposure corresponds to phases where the response is changing rapidly, and is more difficult to infer when the test response levels out (Fig 5b and 5d). For diagnostic tests with a peaking response, estimating the time of exposure can be precise but not unique, with two different regions of probable exposure times for a given test response (Fig 5e).
To further evaluate the gain from utilizing two different tests, we ran simulations against the four parameterizations of a lognormal epidemic mentioned above, using two diagnostic tests and starting with respectively whooping cough and bluetongue test kinetics. The kinetics were then modified to be increasingly similar to each other (full details and results in S1 Text). We found that for the bluetongue scenarios, the level of similarity of the tests did not seem to noticeably affect the accuracy (as measured with R2/RMSEP) of the estimated trends, while for the whooping cough scenarios when recovering Epi3 and Epi4 scenarios, performance degraded gradually, with a very low R2 when using two identical tests. bluetongue test kinetics, the MCMC sampler converged well for the unmodified test configuration (as measured using Gelmans R statistic), but the convergence behaviour then degraded as the tests became increasingly similar. It completely failed to converge in the limit when the two tests were identical (either identical NA tests, or identical antibody based tests). For the whooping cough scenarios the sampler converged for all combinations of diagnostic tests, no matter how similar.
We have shown how to recover epidemic trends of both human and animal pathogens from cross-sectional data collected at a single point in time. We were able to recover this temporal information using a novel statistical framework which combines paired diagnostic test measurements made on collected samples with known temporal kinetics of diagnostics test measurements over the course of infection.
The inferential framework introduced here allow us to extract rich temporal information from collected diagnostic samples. Here we focused on purely cross-sectional samples, but the methods are applicable to longitudinal data and data sets combing both longitudinal and cross-sectional samples. We were able to estimate the trends of both increasing epidemics and decreasing epidemics, as well as estimate the approximate pace of increase or decrease. Such information would improve situational awareness during outbreaks, enabling appropriate management decisions to be implemented immediately when an outbreak has been detected, without the need to observe its subsequent spread to estimate the trend.
The implementation of the framework used in this paper combines surveillance data with information on the test kinetics using a simplified model. For example, individual variation in the test response is modelled as variation around a common mean test curve, rather than as variation in the shape of the curve itself. Variations in the two tests are considered independent, and the error distribution is assumed to be log normal. This limits the pattern and range of variation our model can capture, but facilitates model specification and estimation. We also assume that the test variability is known. While this is partly owing to technical limitations (models with unknown variance parameters tended to converge to degenerate solutions by maximising the variance), it is a realistic assumption since the reliance on test kinetics require that the diagnostic test has been studied in depth. More detailed modelling of the individual and population level processes (including the effect of various covariates such as age or gender) in order to tailor the model to a particular disease is entirely consistent with the statistical framework introduced and would increase the real-world validity and predictive power beyond what has been demonstrated here.
The current implementation of the framework does not include a sampling process component, and so the generated posterior distribution does not currently take sampling uncertainty into account. If the samples are randomly drawn from a larger population of infected individuals, the estimated trend will be an unbiased estimator of the wider population trend. A potential avenue for future research would be to integrate hindcasting into a wider framework describing the sampling process in detail; such an approach might also allow for simultaneously estimating potential sampling bias.
We make use of the lognormal distribution as a parsimonious parameterization of the epidemic trend. This is suitable for epidemics where a single peak is expected, allowing fast model fitting whilst capturing the time span and general direction of the trend. The trade-off is that more complex aspects of trends in the epidemic are omitted. A second limitation is that the lognormal distribution requires the trend to decline to zero after any peak. Should either of these limitations pose a problem, more complex parameterizations of the epidemic trend—with multiple peaks and stages, or even compartmental SEIR-type models—could be used, though such models are likely to come at substantially higher computational cost.
The hindcasting framework introduced here estimates epidemic trends by combining observed data with information on how test responses develop after exposure. Woolhouse and Matthews [37] give an extensive overview of studies that incorporate different data sources to recover the underlying dynamics of disease spread [38,39], and argue that the future of disease analysis lies in models taking account of a wider range of inputs, such as diagnostic test performance, disease pathogenesis, or transmission mechanics, in addition to regular surveillance data. Our methodology improves on earlier studies incorporating test kinetics [29–31] in three ways: by incorporating information from more than one diagnostic test; by considering their joint kinetic pattern; and by modelling non-constant incidence. It could be further extended to model other aspects of the disease system such as population demography, contact networks, or the spatial distribution of cases.
The hindcasting approach make use of knowledge of the within-host development of test markers. Phylodynamics, on the other hand, leverages information about how the genetics of the pathogen change as it spreads through the population to estimate between-host transmission events, and use this to e.g. reconstruct the transmission network of outbreaks [40] and to inform future control measures and forecasts of outbreak trajectories such as the 2015 Ebola outbreak [41], and the 2009 H1N1 influenza outbreak [42]. However, phylodynamics requires sequenced samples of genetic material, and that the pathogen of interest is mutating quickly enough that the dynamics of the epidemic can be resolved. In contrast, the hindcasting approach relies on test kinetics and measures within-host times since infection. Recent papers [43,44] discuss ways to integrate epidemiological and genetic information when modelling disease epidemics; given the complementary nature of phylodynamics and hindcasting, a natural future step would be to combine the two sources of information into a single framework.
Since hindcasting exploits knowledge of the host-pathogen interaction, it relies on previously conducted longitudinal studies of such interactions, and requires that the test response after initial pathogen exposure has been described. Our results demonstrate one of the many ways in which experimental infection studies can provide substantial additional benefits to disease control and research. Currently, only a fraction of pathogen tests have published information on how time since exposure affects test response; this has limited which pathogens we could usefully simulate. Similarly, data sets of infectious diseases often only record whether a test has been positive or not. Presentation of the underlying continuous test response is rare—and it is rarer still to find such results for paired diagnostic tests.
It is hoped that the method introduced here can give some further motivation to record continuous test responses from more than one diagnostic test, and that it can also serve as an argument for conducting further studies on test kinetics. The results regarding the impact of combining diagnostic tests indicate that combining diagnostic tests increases the robustness of the hindcasting procedure. The bluetongue scenarios exhibited severely degraded convergence behaviour with two identical tests, while the whooping cough scenarios converged, but produced posterior estimates that did not match the true trend.
Further research is required to characterise how combinations of tests interact to affect hindcasting. In general, however, the diagnostic tests used in disease surveillance should be chosen to complement each other. By combining early responding tests with later responders, it becomes possible to create a joint test signature that combines the best features of both tests. In terms of the method presented here, the best combination of tests is likely that which provides a unique and precise signature along the timeline of infection for an individual.
The hindcasting framework described here makes use of diagnostic test information, that has up until now been under-utilized, to improve situational awareness during an outbreak.
This approach could also be used to improve the detection of new outbreaks by extracting more information from existing surveillance data and thus make outbreak detection algorithms [20,21] more sensitive. It could similarly be used to provide additional sources of information when estimating epidemiological parameters and trends, and thus improve the accuracy of forecasting models.
We have described a new framework for hindcasting the temporal patterns of epidemics, using two example host-pathogen systems and the pairing of antibody tests with pathogen load. The framework demonstrates the potential to utilise the information inherent in the increasing variety of diagnostic tests. We were able to estimate both increasing and declining epidemic trends under the assumption that all individuals were being tested at a single point in time, implying its usefulness for cross-sectional surveillance data as well as in less restrictive settings. Recovering temporal incidence trends using multiple tests on cross-sectional field data has the potential to be of considerable value, and is a key determinant of introducing proportionate responses to ongoing disease outbreaks.
Our method assumes test data yik from multiple disease diagnostics indexed by k = 1, …, K on individuals i = 1, …, N. We assume that each individual is tested at some time ti, after having been exposed to the pathogen at some earlier time ei. We further assume that these individuals are chosen in an unbiased, random manner from a larger population. Each diagnostic test is assumed to return a value in the form of a continuous ‘level’, which might, for example, be the highest dilution at which antibodies are detected in a serological test. Without loss of generality we assume that these levels are scaled to the unit interval [0,1].
Initial exposure to a pathogen is the start of a complex dynamical process within the host. We conceptualize such internal host-pathogen interactions as a multivariate process that depends on the time since initial exposure. Each diagnostic test is assumed to target the state of a different component of this process so that each test k carried out at time ti on individual i can be modelled as a latent variable lik(ti, ei) = lik(di), with each test having differing but correlated response patterns over the time since initial exposure di = ti−ei. We model these latent variables using results from experimental infection studies for a given host-pathogen system, where the length of time since initial exposure di is known.
The known data, across all individuals in the sample, comprises a set of test results denoted by Y = {yik} with sampling times T = {ti}. Our aim is to infer the unknown set of exposure times E = {ei}, using information on the behaviour of the latent processes L = L(T, E) = {lik(ti, ei)} generating the test results. Note that when describing these sets the limits of each index k = 1, …, K and i = 1, …, N are implicit.
Under our statistical model we assume that the sampling times T are precisely known whereas the quantities Y, L and E are assumed to be subject to uncertainty and variation. There are thus three components to the statistical model: a latent process model P(L|T, E, θL) describing uncertainty and variation in the host-pathogen interaction process within the host in terms of the time since initial exposure; a testing or observation model P(Y|L, θY) describing the distribution of results from tests carried out on the hosts conditional on the internal latent process; and an epidemic trend model P(E|T, θe), describing the historical development of the epidemic in terms of the distribution of exposure times in the sampled host population, at the time of sampling. We discuss specific implementations of each of these components in the examples described below.
Combining the three parts of the model, we write the full data likelihood given an observed data set {Y, T}as
P(Y,E,L|T,θ)=P(Y|L,θY)P(L|T,E,θL)P(E|T,θE),
where θ = {θY, θL, θE}. Thus the likelihood combines models for testing with those for within and between host pathogen interactions.
According to Bayes’ theorem, the so-called posterior distribution for the unknown parameters is proportional to the data likelihood and prior P(θ). We can express this relationship for the parameters of interest, the latent process L, the exposure times E and the parameters θ, given the observed test data Y and sampling times T, by the equation
P(L,E,θ|Y,T)=P(Y,E,L|T,θ)P(θ)P(Y,T)
Within the Bayesian framework all inference is based on the posterior. The prior P(θ) can result from previous measurements or expert opinion, and represents knowledge about the values of parameters before we see the data used in the likelihood.
In what follows, we will make the simplifying assumption that the latent process L is modelled by a known deterministic function of T and E, and represents the expected value of the test results given the times since exposure. This means that the term P(L|T, E, θL) drops out of the likelihood which then simplifies to P(Y, E|T, θ) = P(Y|L(T, E), θY)P(E|θE), and the posterior becomes
P(E,θ|Y,T)= P(Y,E|T,θ)P(θ)P(Y,T)
Note that under this notation any parameters defining the deterministic latent process L(T, E) = {lnk(tn, en)} are suppressed since they are not inferred i.e. θ = {θY, θE}.
In both cases above the normalisation factor P(Y, T) is typically unknown and computationally expensive to calculate. However, standard Markov Chain Monte Carlo (MCMC) methods circumvent this problem and are able to generate samples from the posterior even though the normalisation is unknown. The results presented in this paper are generated from an MCMC sampler implemented with a Metropolis-Hastings algorithm in JAGS [45] using Gibbs sampling [46].
Whooping cough is a human disease caused by the bacteria Bordetella pertussis, causing prolonged spasmodic coughing. Despite widespread vaccination coverage there has been a resurgence of cases in several countries. In the Netherlands there has been a steady increase in the incidence since 1996; and in California, USA, in 2011, there was a widespread outbreak with 9000 cases and ten deaths [47]. The reasons for such resurgence is currently a matter of scientific debate; some hypotheses include antigenic drift of the bacterium [48,49], asymptomatic transmission of B. pertussis by vaccinated individuals [50], or the resurgence being the consequence of changing vaccines and vaccination schedules [51]. Here we make use of data describing a county-wide outbreak of whooping cough primarily among adolescents and adults in Fond du Lac County, Wisconsin, USA in 2003–2004, [52]. After an early cluster of cases in a high school in early May 2003, there was a large outbreak of whooping cough throughout the county starting from October. After some time, this outbreak was contained, and the final cases occurred in February 2004. The upper part of Fig 1 shows interpolated case counts per 48-hour period over the duration of the outbreak.
Bluetongue virus (BTV) is a midge-borne virus that can infect ruminants such as sheep, cattle, deer, and camelids, causing bluetongue disease with symptoms such as internal haemorrhages, swelling of the tongue, lesions in the mouth, and in some species death (most notably in naïve sheep and white-tailed deer). Bluetongue infections can have severe economic consequences for livestock farming, both due to loss of productivity, and because of the severe control measures needed to prevent spread [53]. In 2006, bluetongue emerged throughout northern Europe, with recorded outbreaks in the Netherlands, Belgium, Germany, and Luxembourg. In 2007, the UK had its first recorded outbreak [54]. The first infections occurred sometime in early August 2007 [54] when midges introduced the pathogen to the British Isles, but the first case was not detected until September. The lower part of Fig 1 shows the case count per day, with numbers interpolated from the published weekly data [54].
In order to assess our methodology, we consider two scenarios for each pathogen outbreak. In the “increasing” scenarios we assume the epidemic is recognised early and explore test results from samples taken at a time early on in the outbreak (when the outbreak is increasing, see e.g. Fig 1). In contrast in the “decreasing” scenarios we use test results assumed to be obtained from individuals exposed during the entire outbreak, with samples collected at a relatively late stage in the outbreak (i.e. when it is in decline). The goal was to see how well hindcasting could distinguish between increasing scenarios and scenarios where the epidemic is in decline. We were also interested to see if it was possible to estimate the approximate time span of the epidemics.
The results of diagnostic testing are characterised in terms of an underlying mean trend and a model which accounts for variation around this reflected measurement error, and within and between individual variability in test response.
Given simulated times of exposure, we then simulated test results, based on the elapsed time between the time of exposure in the outbreak and the assumed time of sampling, using published kinetics of real-time PCR analysis and quantitative ELISA for B. pertussis [34,56], to inform a latent process P(L|T, E, θL). Specifically these were the kinetics of ELISA IgG B. pertussis antitoxin [35] for antibody test response ab(d)as a function of time since exposure d, and real-time PCR measurement of persistence over time of B. pertussis DNA in nasopharyngeal secretions [34] (see Fig 2) for the pathogen load DNA(d). As noted earlier formally, we defined the deterministic function L(di) = (DNA(di), ab(di)) by fitting interpolated curves to the published data on DNA and antibody levels using LOESS [55].
The distribution P(Yi|L(di)) of test measurements was modelled as a lognormal distribution conditional on the state of the latent process: let yi = (yNA, yab)i represent a bivariate measurement of nucleic acid and antibody levels on individual i, and define the distribution P(Yi|L(di))= lN(L(di),Σ2)), where Σ2 is a diagonal covariance matrix, reflecting the assumption of no correlation between test results when conditioned on the time since exposure. The variance for each test (i.e. the diagonal elements of Σ2) was assumed to be known. Antibodies as well as levels of pathogens in a host often follow log-normal distributions, as has been rigorously argued [57]; the suitability of using the lognormal distribution for modelling a wide range of biological phenomena has also been described more recently [58].
We modelled the test behaviour based on published data [36], and assumed lognormal distributions for the epidemic trend, as well as for the variance of the diagnostic tests (test kinetics shown in Fig 4). Specifically, we based the behaviour of the latent process P(L|T,E, θL) on a study of experimental infection of European red deer with BTV serotype 1 and 8 that described the dynamics of BTV serotype 1 viral load (vl) as measured with RT-PCR, and antibody levels (ab) as measured with ELISA. As above, we define the latent process describing antibody concentration and viral load as a deterministic bivariate function of the duration d elapsed since exposure as L = {l(di)} ≡ (vl(di), ab(di)), which does not vary between individuals. We estimate L by fitting smooth and interpolated curves to the experimental study data on viral load and antibody levels independently and take the values of these curves at each exposure time d to define the values of the deterministic functions, vl(d), ab(d). A smoothing spline algorithm [55] was used as a nonparametric fitting method. Conditional on the time since exposure, the observed test values yi = (yvl, yab)i were modelled as a bivariate log-normal distribution with mean equal to the deterministic latent process = {l(di)} = (vl(di), ab(di)). For individual i, this can be formally written as P(yi|l(di))=lN(l(di),Σ2), where lN indicates a bivariate lognormal probability function, and Σ2 is the covariance matrix. We assumed that the variation in observed antibody levels and viral loads to be independent so that the covariance matrix Σ2 is diagonal, with variance components σ12, σ22. The variance for each test (i.e. the diagonal elements of Σ2) was assumed known.
The third and final part of the model, the distribution of times since exposure P(E|T, θe), was modelled as a lognormal distribution P(E|T, θe={μ,σ})=lN(μ, σ).
In this case, we exploit the ability of the lognormal to model extreme skewness to capture both increasing and decreasing epidemics using only two parameters. Note that the implementation of the lognormal distribution as an epidemic trend must be conducted in such a way as to allow the sampler to jump between tails of the distribution for the individual times since infection. Details for how to do this can also be found in the supplementary information.
We followed the recommendations of Gelman et al. [59] and used vague priors for the parameters. Such priors incorporate information about what parameter values are nonsensical in a given problem setting, but without using any previously collected data. The means for the lognormal distribution describing the epidemic trends were themselves given lognormal priors. These priors were set to indicate the timescale of relevance for the epidemics in question. This translated to setting the prior means for the for the increasing whooping cough to 100 days, and prior means for the decreasing whooping cough scenario to 200 days. The corresponding values for bluetongue were 10 days and 100 days, respectively. The standard deviations for the prior distributions were chosen as log(10) corresponding to 99% confidence intervals of (mean/100,mean*100). This standard deviation was chosen to model that any peak time more than a factor 100 different from the time scale of interest was nonsensical.
The standard deviation of the lognormal distributions for the epidemic trend was given a vague prior parametrized as a folded, non-standardized t-distribution with five degrees of freedom, a standard deviation of log(100), and a mean of 0, indicating in the spirit of vague priors that a spread of on the order of more than 100 days in the past was not sensible.
In the process of developing this work, we also explored the use of uninformative priors with nearly flat distributions, such as the standard gamma distribution, and uniform priors with very wide support; however, these were found to lead to very slow mixing and a high rate of convergence failure of the MCMC algorithm. Changing the specific values of the priors did not influence the posterior estimates noticeably. See the supplementary information for MCMC traceplots, details of convergence evaluation and sensitivity analysis of the priors.
|
10.1371/journal.ppat.1000338 | The Tetraspanin Protein CD37 Regulates IgA Responses and Anti-Fungal Immunity | Immunoglobulin A (IgA) secretion by plasma cells in the immune system is critical for protecting the host from environmental and microbial infections. However, the molecular mechanisms underlying the generation of IgA+ plasma cells remain poorly understood. Here, we report that the B cell–expressed tetraspanin CD37 inhibits IgA immune responses in vivo. CD37-deficient (CD37−/−) mice exhibit a 15-fold increased level of IgA in serum and significantly elevated numbers of IgA+ plasma cells in spleen, mucosal-associated lymphoid tissue, as well as bone marrow. Analyses of bone marrow chimeric mice revealed that CD37–deficiency on B cells was directly responsible for the increased IgA production. We identified high local interleukin-6 (IL-6) production in germinal centers of CD37−/− mice after immunization. Notably, neutralizing IL-6 in vivo reversed the increased IgA response in CD37−/− mice. To demonstrate the importance of CD37—which can associate with the pattern-recognition receptor dectin-1—in immunity to infection, CD37−/− mice were exposed to Candida albicans. We report that CD37−/− mice are evidently better protected from infection than wild-type (WT) mice, which was accompanied by increased IL-6 levels and C. albicans–specific IgA antibodies. Importantly, adoptive transfer of CD37−/− serum mediated protection in WT mice and the underlying mechanism involved direct neutralization of fungal cells by IgA. Taken together, tetraspanin protein CD37 inhibits IgA responses and regulates the anti-fungal immune response.
| Antibody, or immunoglobulin (Ig), production by plasma cells in the immune system is important for protecting the host from microbial infections. IgA is the most abundant antibody isotype produced in the body. However, the molecular mechanisms underlying the generation of IgA–producing plasma cells remain poorly understood. We now report that the B cell–expressed protein CD37 regulates IgA immune responses, both in steady-state conditions and during infection. We found highly increased levels of IgA in serum and elevated numbers of IgA+ plasma cells in lymphoid tissue of mice that are deficient for CD37 (CD37−/− mice). To demonstrate the importance of CD37 in immunity to infection, CD37−/− mice were exposed to the fungus Candida albicans. C. albicans can cause systemic infection with high mortality in immunocompromised patients. We demonstrate that CD37−/− mice are evidently better protected from infection than wild-type mice, which was dependent on C. albicans–specific IgA antibodies. The underlying mechanism involved direct neutralization of fungal cells by IgA. In summary, the B cell protein CD37 inhibits IgA responses and anti-fungal immunity. This study may contribute to the development of novel immunotherapeutic approaches for invasive fungal disease.
| Plasma cells (non-dividing antibody-secreting cells, ASC) are terminally differentiated B cells that are central to humoral immunity. Both mucosal and systemic immune responses to infection can induce IgA+ plasma cell formation, resulting in production of secretory and serum IgA respectively. In both pathways, isotype class-switching in B cells is tightly regulated by the cytokine-milieu. While it is believed that interleukin-6 (IL-6), originally described as a B-cell stimulating factor [1], stimulates plasma cell differentiation and promotes IgA responses in vivo [2], the molecular mechanisms underlying the development of IgA-secreting plasma cells remain poorly understood. In particular, detailed knowledge about proteins in the B cell membrane that control this process is lacking.
IgA-secreting plasma cells are predominantly found in the mucosal-associated lymphoid system (MALT) and bone marrow. During the mucosal immune response, intestinal dendritic cells present mucosal antigen to CD4+ T cells in Peyer's patches, or present antigen directly to B1 cells in a T cell-independent manner [3],[4]. This leads to the local production of secretory IgA, which is important in neutralizing intestinal microbes and controlling gut homeostasis. Secondly, systemic immune responses to T cell-dependent antigens lead to the development of germinal centers (GC) in spleen and peripheral lymph nodes that are the origin of long-lived plasma cells that produce high affinity IgG and IgA antibodies in serum [5]–[7]. Recent studies have underlined the importance of local production of B cell stimuli (cytokines, TLR ligands) in inducing IgA production during the humoral immune response [3],[8], including IL-6 that promotes the generation of IgA-secreting plasma cells [2],[3].
Tetraspanins, or transmembrane-four superfamily proteins, are implicated in organizing (immuno)-receptors, integrins, and signaling molecules into functional membrane complexes (tetraspanin microdomains) [9]–[12]. Consequently, tetraspanins are important in fundamental cellular processes including migration, proliferation, differentiation, and –when deregulated- cancer [13]. CD37 is expressed exclusively on cells of the immune system, in contrast to most other tetraspanins. CD37-deficient mice display defects in various arms of the immune system, including impaired T cell-dependent IgG responses, T cell hyperproliferation, and increased antigen-presenting capacity by dendritic cells [14]–[16]. We have recently demonstrated that the C-type lectin dectin-1 interacts with tetraspanin CD37 [17]. Dectin-1 recognizes β-glucans that are found in the cell wall of fungi, and is involved in cytokine production and killing of fungal pathogens including Candida albicans [18],[19].
In this study, we provide novel insights into the mechanisms underlying IgA production during the humoral immune response. We demonstrate that the B cell-expressed tetraspanin CD37 inhibits the formation of IgA-secreting plasma cells in vivo that is critically dependent on IL-6. Moreover, CD37-deficient mice are protected against C. albicans infection, which was dependent on fungal-specific IgA antibodies. Taken together, tetraspanin protein CD37 inhibits IgA responses both in steady state conditions and during infection. This is the first demonstration that tetraspanins control the immune-mediated defense against fungal pathogens.
CD37 expression was determined in different mouse and human leukocyte populations at the mRNA and protein level, respectively. In both mice and human, B cells were found to be the major CD37 expressing cells (Figure 1A and 1B). In lymphoid tissue, CD37 expression was highest in the B cell follicle area as expected (Figure 1B, data shown for spleen). This high CD37 expression on B cells prompted us to study humoral immunity in CD37-deficient mice in detail. When analyzing the basal antibody levels in sera of naïve mice, we observed that basal levels of serum IgA in CD37−/− mice were increased more than 15-fold compared to C57BL/6J wild-type (WT) mice (Figure 1C). At the same time, IgG1 levels were decreased 2-fold and serum IgM levels were unaltered compared to WT mice. Next, we investigated antibody responses after immunization with the T cell-dependent antigen (NP-KLH) in CD37−/− and WT mice. Sera were analyzed for the presence of antibodies reactive with NP3-BSA (high affinity anti-NP antibody) and NP20-BSA (total anti-NP antibody) [20]. Again, CD37−/− mice developed high titers of anti-NP IgA, with >10-fold increase of high affinity IgA compared to controls 21 days after immunization (Figure 1D).
We wanted to investigate whether CD37-deficiency on the B cell population was indeed responsible for the high IgA response. WT mice were sublethally irradiated and reconstituted with bone marrow; 80% from μMT mice (B cell-deficient mice), and 20% from WT or CD37−/− mice as a source of B cells. This created WT mice and mice with a CD37-deficient B cell compartment (referred to as chimeric CD37−/− mice). The percentages of CD19-, CD3-, GR-1-, and F4/80-expressing cells assessed in peripheral blood revealed equally efficient reconstitution in WT and chimeric CD37−/− mice (data not shown). Next, NP-specific antibody production was analyzed 35 days after immunization with NP-KLH. The results clearly demonstrate that chimeric mice with a CD37−/− B cell repertoire also produced high titers of NP-specific IgA comparable to intact CD37−/− mice after immunization (Figure 1D, right). Thus, we can conclude that the elevated IgA response in CD37−/− mice is a B cell-intrinsic defect.
The high IgA titers in serum of CD37−/− mice suggest an enhanced generation of IgA-secreting plasma cells. To determine if this was the case, spleens, bone marrow, and mucosal-associated lymphoid tissue from immunized mice were examined for the frequency of NP-specific IgA-ASC using ELISPOT assays. In line with IgA levels in serum, the numbers of IgA-ASC were increased in spleen of CD37−/− mice compared to control WT mice at day 14 following immunization (Figure 2A). This observation was confirmed by immunohistochemical analysis (Figure 2B). High numbers of IgA+ cells were present in GC and white pulp area in spleens of immunized CD37−/− mice, whereas in spleens of WT mice IgA+ cells were very rarely detected. CD138 stainings confirmed that the IgA+ cells in CD37−/− spleens were indeed plasma cells (Figure 2C). Since IgA-ASC are known to migrate preferentially to the MALT using specific gut-homing receptors [21], Peyer's patches and mesenteric lymph nodes of immunized CD37−/− mice were analyzed for the presence of NP-specific IgA-ASC. Both the morphology and organization of mesenteric lymph nodes (Figure 2D) and Peyer's patches (Figure 2E) of CD37−/− mice were comparable to WT mice, although CD37−/− Peyer's patches were slightly increased in size. The percentage of IgA+ NP-specific plasma cells was increased in MALT of immunized CD37−/− mice compared to WT controls (Figure 2F), demonstrating that IgA-ASC generated in CD37−/− spleens preferentially homed to the MALT rather than to the bone marrow. IgG1+ NP-specific plasma cells were hardly detected in MALT of CD37−/− and WT mice as expected. Taken together, CD37−/− mice possess increased generation of antigen-specific IgA+ plasma cells after immunization.
IL-6 is a cytokine that has been implicated in promoting the generation of IgA-secreting plasma cells [2],[3]. Moreover, we have recently shown that signals transduced through the CD37 molecular partner, dectin-1, lead to an elevated production of IL-6 by CD37−/− cells [17]. Consequently, we examined whether a dysregulation of IL-6 production in immunized CD37−/− mice may underlie the excess production of IgA. We first examined the expression of IL-6 in the germinal centers (GC) of immunized mice by immunohistochemistry. We readily detected IL-6 expression in the GC area of spleens of immunized CD37−/− (Figure 3A). In contrast, non-immunized mice (WT and CD37−/− mice) as well as immunized WT mice had non-detectable levels of IL-6 in spleens. IgA production by peritoneal B1 cells during T cell-independent responses is not dependent on IL-6 [22], which correlates with the normal IgA response in CD37−/− mice after T cell-independent immunization (data not shown).
Next, the effect of IL-6 on IgA production during ex vivo restimulation experiments was analyzed. Splenocytes of immunized WT and CD37−/− mice were stimulated with NP-KLH in vitro in the absence or presence of neutralizing IL-6 antibodies. Figure 3B shows increased IgA production by CD37−/− cultures compared to WT cells as expected. Blocking IL-6 resulted in substantially reduced IgA production by CD37−/− cells, which supported our hypothesis that the mechanism underlying the elevated IgA responses in CD37−/− mice is controlled at the level of IL-6. WT and CD37−/− cultures produced 1900 vs. 5500 pg/ml IgA respectively, which decreased to 500 vs. 2000 pg/ml in the presence of neutralizing IL-6 antibodies. We also established that purified CD37−/− splenic B cells were capable of autocrine IL-6 production upon restimulation in vitro using intracellular cytokine stainings (data not shown).
To prove that increased IgA production in CD37−/− mice was indeed dependent on IL-6 in vivo, we neutralized IL-6 in WT and CD37−/− mice during immunizations using blocking antibodies. Serum was analyzed for the presence of high affinity (anti-NP3) IgA antibodies at different days after immunization. Evidently, IL-6 neutralization reversed the production of high IgA levels in CD37−/− serum (Figure 3C, left), demonstrating the importance of IL-6 in IgA+ plasma cell development in CD37−/− mice. Only 25% of the CD37−/− mice produced NP-specific IgA titers after treatment with anti–IL-6 antibodies, compared to 100% of isotype control-treated CD37−/− mice 14 days after immunization (Figure 3C, right). IgA formation by WT mice (with or without anti–IL-6) was below the level of detection (not shown). Our findings are in accordance with observations made in IL-6-deficient mice that exhibit impaired antibody production during systemic immune responses [23]–[25]. Mucosal IgA responses were reported to be significantly impaired or unaltered in IL-6-deficient mice dependent on the model of immunization used [2],[23],[25],[26]. Moreover, IL-6 polymorphisms were recently reported to be associated with IgA-deficiency in patients [27]. Since IL-6 is crucial for the induction of already-committed (surface IgA+) B cells to become IgA-secreting plasma cells [2]), our data suggest that CD37 is directly implicated in the inhibition of the late-stage development of IgA+ plasma cells, i.e. at the level of terminal B cell differentiation. Taken together, we established that high IL-6 levels directly relate to the increased IgA production in CD37−/− mice.
Given that IgA contributes to protecting the host from microbial infections, we hypothesized that CD37-deficiency may have important implications for the outcome of infectious diseases. We have recently demonstrated that CD37 interacts with the C-type lectin dectin-1 [17], a β-glucan receptor that is required for effective immunity to fungal infections [28],[29]. Therefore, the CD37-deficient immune response during Candida albicans infection was explored. C. albicans normally colonizes the mucosa without causing disease, but can cause systemic infection with high mortality in immunocompromised patients [30],[31]. In particular, the incidence of invasive C. albicans infections is high among cancer patients [32]–[34].
CD37−/− and WT mice were infected with C. albicans and IL-6 production by CD37−/− and WT splenocytes was studied upon restimulation with fungal antigens. CD37−/− splenocytes produced increased levels of IL-6 compared to WT cells upon exposure to either live or heat-killed C. albicans, or the fungal cell wall extract zymosan both 3 and 7 days after infection (Figure 4A). Blocking dectin-1 with antibody 2A11 inhibited IL-6 production by both WT and CD37−/− splenocytes stimulated with C. albicans or the dectin-1 ligand curdlan (Figure S1), showing that IL-6 production is dependent on dectin-1. As such, CD37 controls dectin-1-mediated IL-6 production, possibly by recruiting dectin-1 into tetraspanin microdomains that may alter signal transduction pathways and subsequent cytokine profiles. In line with our findings, IL-6-deficient mice are more susceptible to C. albicans and A. fumigatus infection, which is related to decreased neutrophil effector activity, impaired Th1-mediated immune responses [25], and defective Th17 responses [35]. Studying Th2/Th1/Th17 cytokine production by CD37−/− splenocytes revealed that IL-10 production was comparable between CD37−/− and WT splenocytes, and γIFN production was low but increased by CD37−/− cells 3 days after infection (Figure 4A). The role of IL-6 in inducing Th17 responses is well established in mice. Accordingly, we observed significantly increased IL-17 production by CD37−/− splenocytes upon C. albicans stimulation (Figure 4A). Th17 cells have been implicated as an important effector mechanism against C. albicans infection [36],[37], although IL-17 may also impair anti-fungal immunity under certain conditions [38],[39].
Next, the development of fungal-specific IgA antibodies in CD37−/− and WT mice exposed to C. albicans was analyzed. WT mice did not generate a detectable IgA response after C. albicans after infection. In contrast, all CD37−/− mice exhibited high titers of IgA antibodies specific for C. albicans and zymosan in their serum (Figure 4B). Candida-specific IgG was not detected in serum of WT and CD37−/− mice 7 days after infection (not shown). Finally, we investigated the role of tetraspanin CD37 in the outcome of infectious disease. WT and CD37−/− mice were systemically infected with C. albicans yeasts, and kidneys were analyzed for the outgrowth of viable C. albicans after 1 and 7 days. CD37−/− mice exhibited significantly decreased susceptibility to infection and reduced fungal outgrowth in their kidneys –the main target organ for C. albicans - 7 days after infection when compared to WT mice (Figure 4C, left). Histology revealed major infection areas with abscesses and hyphal infiltration in WT kidneys. In contrast, morphology of kidneys of CD37−/− mice looked normal with no mycelial structures and presence of leukocyte infiltrates (Figure 4C, right). Next, we evaluated survival of WT and CD37−/− mice after lethal fungal infection. The absence of CD37 resulted in prolonged survival and decreased mortality after lethal candidiasis (Figure 4D), further emphasizing the importance of CD37 in regulating anti-fungal immunity.
In order to get more insight into the underlying mechanism, CD37−/− serum was analyzed for fungal opsonization and/or neutralizing capacity. IgA-mediated phagocytosis through Fc alpha receptor CD89 is well-established in humans [40],[41], in whom serum IgA is mostly monomeric. In mice, serum IgA is predominantly dimeric/polymeric and it is currently unknown whether the murine Fcα/μ receptor is effective in clearing IgA-opsonized pathogens [42]. We observed similar uptake of C. albicans by GR-1+ neutrophils in the presence of WT or CD37−/− serum of infected mice (Figure 5A). Serum heat-inactivation inhibited C. albicans uptake, demonstrating that the phagocytosis was mainly complement-mediated and not dependent on IgA. Also, the increase in GR-1-expressing cells in blood after C. albicans infection was comparable between WT and CD37−/− mice (Figure S2). Thus, these data demonstrate that IgA in CD37−/− serum does not increase C. albicans phagocytosis by neutrophils.
Next, we investigated whether serum IgA of CD37−/− mice could directly neutralize fungal cells. C. albicans was grown in the absence or presence of WT and CD37−/− serum of infected mice and viability was assessed by classical plate assays. CD37−/− serum significantly inhibited C. albicans growth in contrast to WT serum (Figure 5B). To demonstrate that the growth inhibition was due to fungal-specific IgA, CD37−/− serum was depleted for IgA. Serum deprived from IgA abrogated the effect, and C. albicans growth was comparable between WT and CD37−/− serum. From these findings we conclude that IgA in CD37−/− serum directly neutralizes C. albicans and may thereby prevent dissemination in vivo.
To test this hypothesis, WT mice were infected and treated with pooled CD37−/− serum containing high levels of IgA anti-C. albicans. We observed that WT mice were protected from fungal disease by CD37−/− serum treatment (Figure 5C). Importantly, the resistance was abrogated when IgA was depleted from CD37−/− serum, demonstrating that the passive transfer of resistance to candidiasis in CD37−/− mice is dependent on IgA. Although the elevated IL-6-IL-17 response may also contribute to increased anti-fungal effector activity in CD37−/− mice (Figure 4A), our passive transfer data clearly show that C. albicans-specific IgA alone is sufficient to confer protection against invasive fungal disease. Whilst several studies have alluded to a role for the development of protective anti-fungal antibodies in experimental candidiasis [43],[44], this is—to the best of our knowledge—the first description of an important role for IgA in protection against systemic candidiasis. These studies also highlight the non-redundant role of the tetraspanin CD37, as CD37-deficiency stimulates anti-fungal immunity even in the presence of other tetraspanins. The results of this study may stimulate the development of novel immunotherapeutic strategies for invasive fungal disease that are urgently warranted [33].
In summary, we reveal a unique role for the tetraspanin CD37 in the humoral immune response; CD37 is a negative regulator of IgA+ plasma cell generation in vivo and this role is critically dependent on IL-6. Moreover, this is the first demonstration that tetraspanins may control the immune-mediated defense against fungal pathogens. This study increases our understanding of molecular mechanisms underlying the formation of IgA-secreting plasma cells and may contribute to better insight into anti-fungal immunity.
Murine CD37 expression was analyzed at the mRNA level by quantitative RT–PCR due to unavailability of anti-mouse CD37 antibodies. Total RNA was extracted from WT and CD37−/− leukocyte populations (purified by MACS sorting, Miltenyi Biotec) using Trizol (Invitrogen) and transcribed into cDNA using random hexamer primers (Amersham) and Superscript II RT (Invitrogen). Brain tissue was taken as negative control. Real-time RT–PCR was performed in ABI Prism Sequence Detection system 7000. cDNA was amplified using SYBR Green PCR mastermix (Applied Biosystems) with mouse CD37 primers (5′GTCCTTTGTGGGTTTGTCCTT and 5′GAGACAGCGCAGCTCCTTTAG). The amount of CD37 mRNA expression was normalized to the housekeeping gene PBGD.
Peripheral blood lymphocytes were stained with anti-CD37 (clone WR17, Serotec) and processed for FACS analysis as previously described [17]. CD37 staining in tissue was performed on frozen sections (6 µm) of spleen and lymph node that were fixed in acetone for 10 minutes at −20°C. Slides were blocked (5% horse serum) and stained with anti-CD37 or mIgG2a isotype control. Horse anti-mouse-biotin (Molecular Probes) was used as secondary Ab, and staining was revealed using SA-alkaline phosphatase labeling kit (Vector Laboratories) with Fast Red substrate. Slides were counterstained with Meyer's heamatoxylin.
CD37−/− mice were generated by homologous recombination [14] and backcrossed 10 times to the C57BL/6J background. Age and sex-matched C57BL/6J wild type (WT) mice were obtained from the Walter and Eliza Hall Institute (Melbourne, Victoria, Australia) and Charles River (France). CD37−/− mice were bred at the Burnet Institute Austin Campus (Heidelberg, Victoria, Australia) and the Central Animal Laboratory (Radboud University Nijmegen, The Netherlands). μMT (B cell-deficient) mice [45] were bred at the Walter and Eliza Hall Institute. Mice were used at 8–12 weeks of age. Animal studies were approved by animal ethics Committee of the Austin&Repatriation Medical Centre, and the Nijmegen Animal Experiments Committee.
WT and CD37−/− mice (n = 6) were injected intraperitoneally with 100 µg of NP (4-hydroxy-3-nitrophenylacetate) coupled to Keyhole Limpet Hemocyanin (KLH) (NP17∶KLH, conjugation ratio, 17∶1) precipitated in alum. Mice were bled and sera collected at days 7, 14, 21 and 35 post-immunization. Lymphoid organs (spleens, bone marrow, mesenteric lymph nodes, and Peyer's patches) were frozen for immunohistochemistry or processed for FACS analysis.
Basal immunoglobulin levels in serum of non-immunized mice (n = 9) were determined by ELISA using goat anti-mouse IgG/IgA/IgM (BD-Pharmingen). Ig standards were purchased from BD. Levels of NP-specific antibodies in serum after immunization were assayed by ELISA using plates coated with 20 µg/mL of either NP20-BSA or NP3-BSA to detect total and high affinity anti-NP antibody, respectively. The frequency of total and high affinity NP-specific ASC in spleen and bone marrow was determined by ELISPOT as described [20],[46], again using NP20-BSA and NP3-BSA for antibody capture.
Spleens, mesenteric lymph nodes and Peyer's patches taken from mice before and after immunization were embedded in OCT. Frozen sections (6 µm) were fixed in acetone for 10 min. at −20 C. Slides were blocked (5% goat serum) and stained with anti-IgA, anti-IgG1, anti-IL6, anti-CD138, anti-CD45R (B220) (all from BD-Pharmingen), or isotype controls. Goat anti-rat-biotin (Molecular Probes) was used as secondary Ab, and staining was revealed using SA-alkaline phosphatase labeling kit (Vector Laboratories) with Fast Red substrate. Slides were counterstained with Meyer's heamatoxylin.
Bone marrow chimeras were generated as previously described [47]. Briefly, 107 (CD37−/− or WT) and μMT (CD37+) bone marrow cells were injected intravenously into sublethally irradiated C57Bl/6J recipients at a ratio of 4 parts μMT to 1 part WT or CD37−/− bone marrow as a source for B cells. This created WT mice and mice with a CD37-deficient B cell compartment (referred to as chimeric CD37−/− mice). Six weeks after reconstitution, heparinized blood was obtained from chimeric mice and reconstitution of all major leukocyte populations was monitored by flow cytometry, using the antibodies against: PE-conjugated 1D3-PE (CD19, BD-Pharmingen), PE-conjugated RB6-8C5-PE (GR-1, BD-Pharmingen), Cy5-conjugated KT3.1 (CD3, made in house), PE-conjugated F4/80 (BD-Pharmingen), and FITC-conjugated M1/70 (CD11b, made in house). 2 weeks thereafter, mice (n = 7) were immunized with NP-KLH as described above. Sera were collected 35 days after immunization and analyzed for presence of NP-specific antibodies by ELISA.
Splenocytes were prepared from WT and CD37−/− mice 14 days after NP-KLH immunization. Cells were stimulated with NP-KLH (1 µg/ml) in the absence or presence of 5 µg/ml (azide- and LPS-free) neutralizing anti–IL-6 MP5-20F3 (Biolegend). Supernatants were collected after 48 h, and IgA production was determined by ELISA.
WT and CD37−/− mice (n = 6) were immunized with NP-KLH as described above. Mice received 2 injections of neutralizing (azide- and LPS-free) anti–IL-6 MP5-20F3 or isotype control; 60 µg intraperitoneally at day 0, and 60 µg intravenously 7 days after immunization. Mice were bled and sera collected at days 7, 14, 21 and 35 post-immunization. Lymphoid organs (spleens, bone marrow, mesenteric lymph nodes, and Peyer's patches) were frozen in OCT for immunohistochemistry or processed for FACS analysis.
C. albicans ATCC MYA-3573 (UC820), a well-described clinical strain [48], was injected intravenously (1×105 colony-forming units (CFU), in 100 µL sterile pyrogen-free saline) in CD37−/− and WT mice. Blood (50 µl) was collected at different time points after infection and stained with GR-1-PE to identify neutrophils by flow cytometry. Subgroups of five or six animals were sacrificed on day 1 or 7, and kidneys were removed aseptically, weighed and homogenized in sterile saline in a tissue grinder. The number of viable Candida cells in the tissues was determined by plating serial dilutions on Sabouraud dextrose agar plates. Colony-forming units were counted after 24 h of incubation at 37°C, and expressed as log CFUg−1 tissue. Alternatively, kidneys were were embedded in OCT, and frozen sections (6 µm) were stained with periodic acid-Schiff to identify C. albicans by histology. In adoptive transfer experiments, WT mice (n = 5) received 1×105 CFU C. albicans intravenously (in 100 µL sterile pyrogen-free saline) with 20% pooled serum of infected CD37−/− mice that was either untreated or depleted for IgA (as described below). Colony-forming units were determined in kidneys 7 days after infection. For survival experiments, CD37−/− and WT mice (n = 6) were injected intravenously with lethal dose of 5×105 CFU C. albicans and mice were assessed twice daily.
To assess cytokine production, spleens were removed 3 or 7 days after infection and 5×106 splenocytes were stimulated with 1×107 live or heat-killed C. albicans yeasts (E∶T ratio 2∶1), or 1 mg/ml zymosan (Sigma). Measurement of IL-6, IL-10, IL-17 and γIFN was performed in supernatants collected after 48 h of incubation at 37°C (5% CO2) in 24-well plates by commercial ELISA assays (Biosource, Camarillo, CA; detection limit 16 pg/ml), according to the instructions of the manufacturer. In dectin-1 blocking studies, antibody 2A11 (10 µg/ml) was added during stimulations as described [17]. Curdlan is a specific dectin-1 ligand with no TLR2- or TLR4-stimulating properties [49].
Serum of CD37−/− and WT mice (n = 5) was collected 7 days after C. albicans infection, and different dilutions were incubated with 1×105 Candida or 1×105 zymosan particles for 30 min. at 4°C. After washing twice in saline, Candida particles were incubated with biotinylated anti-mouse IgA (BD-Pharmingen) followed by SA-CY5 (Invitrogen), and analyzed by flow cytometry. Serum of non-infected mice was used as control.
Streptavidin-coated 2 µm microspheres (Polysciences, PA) were incubated with biotinylated anti-mouse IgA (BD Biosciences) according to the instructions of the manufacturer. Sera of WT and CD37−/− mice (n = 4) were collected 7 days after C. albicans infection, pooled, and incubated with 100×106 anti-IgA coated microspheres (1 h at RT, shaking).
C. albicans phagocytosis was performed as described [41]. Briefly, FITC-labeled C. albicans particles were incubated with blood leukocytes (ratio 4∶1) in the absence or presence of serum from infected WT or CD37−/− mice for various time points at 37°C. As control, sera were heat-inactivated at 56°C for 30 min. to inactivate complement. GR-1-PE was used to stain neutrophils and phagocytosis was quantified by flow cytometry. In addition, uptake was analyzed by fluorescence light microscopy. For neutralization experiments, C. albicans was cultured overnight at 37°C in Sabouraud dextrose medium (Oxoid, UK), washed, and 1×105 yeasts were incubated in the absence or presence of 20% serum (from WT or CD37−/− mice; collected 7 days after infection) or 20% IgA-depleted serum. Viability was assessed after shaking for 5 h at 37°C by plating serial dilutions on Sabouraud dextrose agar plates.
Data are presented as mean±SEM. Statistical differences were determined using the unpaired Student's t-test. Significance was accepted at the p<0.05 level.
Human CD37, GeneID 951; Murine CD37, GeneID 12493, MGI: 88330; Murine IL-6, GeneID: 16193, MGI: 96559; Murine IgA, GeneID: 238447, MGI: 96444.
|
10.1371/journal.pgen.1006830 | Cleavage of the SUN-domain protein Mps3 at its N-terminus regulates centrosome disjunction in budding yeast meiosis | Centrosomes organize microtubules and are essential for spindle formation and chromosome segregation during cell division. Duplicated centrosomes are physically linked, but how this linkage is dissolved remains unclear. Yeast centrosomes are tethered by a nuclear-envelope-attached structure called the half-bridge, whose components have mammalian homologues. We report here that cleavage of the half-bridge protein Mps3 promotes accurate centrosome disjunction in budding yeast. Mps3 is a single-pass SUN-domain protein anchored at the inner nuclear membrane and concentrated at the nuclear side of the half-bridge. Using the unique feature in yeast meiosis that centrosomes are linked for hours before their separation, we have revealed that Mps3 is cleaved at its nucleus-localized N-terminal domain, the process of which is regulated by its phosphorylation at serine 70. Cleavage of Mps3 takes place at the yeast centrosome and requires proteasome activity. We show that noncleavable Mps3 (Mps3-nc) inhibits centrosome separation during yeast meiosis. In addition, overexpression of mps3-nc in vegetative yeast cells also inhibits centrosome separation and is lethal. Our findings provide a genetic mechanism for the regulation of SUN-domain protein-mediated activities, including centrosome separation, by irreversible protein cleavage at the nuclear periphery.
| The nucleus, where the eukaryotic chromosomes are stored, is enclosed by a double-membrane structure called the nuclear envelope. Located at the inner nuclear membrane, a class of highly conserved proteins called SUN-domain proteins regulates a range of nuclear activities at the nuclear periphery, including tethering the centrosomes to the nuclear membranes. Defective SUN-domain proteins have been linked to certain types of muscular dystrophy and premature aging in humans. In this study, we use budding yeast as a model and show that its sole SUN-domain protein, Mps3, a key component of the yeast centrosome, is phosphorylated and then subject to proteolytic cleavage in order to properly split the centrosomes before spindle formation and chromosome segregation. Understanding how SUN-domain proteins are posttranslationally modified can shed light on the regulation of centrosome separation; it also provides insight into the homeostasis of SUN-domain proteins that are associated with many human diseases of the nuclear envelope.
| Centrosomes nucleate microtubules and form a bipolar spindle that separates chromosomes during cell division. Like DNA replication, centrosome duplication occurs only once per cell cycle. Duplicated centrosomes are tethered, and their timely separation ensures accurate chromosome segregation. Supernumerary centrosomes and the subsequent formation of aberrant spindles can lead to aneuploidy in humans [1, 2]. At the core of the animal centrosome lies a pair of centrioles, whose cleavage by the cysteine protease, separase, necessitates their disengagement [3], although the substrate of the separase at the centriole remains controversial. In contrast, how the centrosome linkage is dissolved is less clear. In particular whether irreversible protein cleavage is required for centrosome separation, also called centrosome disjunction in animal cells, is not known.
In yeast, the centrosome is often referred to as the spindle pole body (SPB), which shares structural components with and is functionally equivalent to the human centrosome [4, 5]. One unique feature in budding yeast is that the SPB is embedded in the nuclear envelope. Duplicated SPBs are tethered by a nuclear membrane-anchored structure called the half-bridge [6], two of which form the full bridge (Fig 1A). There are four known half-bridge components; three of them, Sfi1 (called hSfi1 in human) [5], Cdc31 (called centrin in animals)[7] and Kar1[8], localize to the cytoplasmic side of the bridge, whereas Mps3 localizes to the nuclear side [9, 10]. Currently unknown is whether protein cleavage is required for yeast centrosome separation. If so, what is the nature of the protease that cleaves the yeast centrosomes?
During vegetative growth, the yeast SPB is duplicated at late G1 phase or early S phase of the cell cycle, but duplicated SPBs separate immediately in order to form a bipolar spindle even when DNA is still being replicated [4, 11]. SPB separation requires kinase activities from the cell-cycle-dependent kinase Cdk1 and the polo-like kinase Cdc5 [12–14]. Recent studies have shown that Sfi1 is a target of both Cdk1 and Cdc5, and phosphorylation of Sfi1 plays a critical role in both SPB duplication and separation [15, 16]. The current model for SPB linkage posits that at the cytoplasmic side of the nuclear envelope, Sfi1 forms protein dimers that span the entire bridge to mediate SPB cohesion [17, 18]. Together with the actions from Cdc31 and Kar1, Sfi1 tethers duplicated SPBs to generate a side-by-side SPB configuration (Fig 1A). Phosphorylation of Sfi1 at its C-terminal domain plays a role in SPB separation [15, 16, 19]. But how Mps3 forms the half-bridge at the nuclear side of the membrane and what its contribution is in SPB separation both are unclear.
In contrast to their separation within minutes after duplication in mitosis, duplicated SPBs are linked for hours during the meiotic G2 phase (often called prophase I) when recombination takes place [20, 21]. The telomere-associated protein Ndj1 also localizes to the SPB in an Mps3-dependent manner and protects SPBs from premature separation during meiosis [22]. Ndj1 binds to the N-terminal region of Mps3 [22, 23], indicating that Mps3 is the target of Ndj1 at the SPB.
Mps3 belongs to the SUN-domain protein family, in which the SUN domain is typically located at the C-terminus of the protein and is found in the lumen of the nuclear envelope [9, 24]. The SUN domain of Mps3 is required for targeting Mps3 to the SPB and for inserting the newly duplicated SPB into the nuclear membranes [25, 26]. In contrast, the N-terminal region of Mps3, which is present in the nucleoplasm, appears dispensable in mitosis, because deletion of the first 64 amino acids of Mps3 does not interfere with cell growth [10]. However, during meiosis the N-terminal domain is required for Ndj1 binding and regulates telomere movement at the nuclear periphery [22, 23].
Here we investigate how Mps3, in particular its N-terminal domain, regulates SPB cohesion. We have determined that Mps3 is proteolytically cleaved at its N-terminal domain during yeast meiosis. Noncleavable Mps3 inhibits SPB separation, demonstrating that cleavage of Mps3 is a critical event in the process of SPB disengagement and spindle formation in budding yeast meiosis.
We have shown previously that Mps3 is an abundant but unstable protein during yeast meiosis [22]. Using a C-terminally HA-tagged allele of MPS3, which was incorporated at the endogenous MPS3 locus and served as the only functional copy of MPS3 in the yeast genome, we found that the level of Mps3 protein peaked 6 h after meiotic induction (Fig 1B), the time of which corresponded to meiosis I in the SK1 yeast genetic background. The level of the full-length Mps3 then decreased as yeast cells exited meiosis I, with the concomitant appearance of a prominent C-terminal fragment, about 12 kDa smaller than the full length Mps3 (Fig 1B). Note that this C-terminal fragment of Mps3 persisted toward the end of meiosis and appeared to have a longer half-life than the full length Mps3 (Fig 1B and see below). These findings indicate that the Mps3 protein is subject to modification during yeast meiosis.
To determine the localization of Mps3 to the SPB, we used an N-terminally tagged GFP-MPS3 allele, which also was the only functional copy of the MPS3 gene in these cells. By fluorescence microscopy of live meiotic yeast cells, we found that Mps3 localized to the nuclear periphery, but it was concentrated at the SPB (Fig 1C and 1D). Line scan of the fluorescence intensities of Mps3 and the SPB marker Tub4, which was fused to RFP, revealed colocalization of Mps3 with Tub4 (Fig 1D). Crucially, ~10 min before SPB separation at the transition from prophase I to metaphase I when the Tub4 signal was stretched to form an axial line, the GFP-Mps3 signal remained focused and was found at the center of that of Tub4 (Fig 1D), demonstrating Mps3’s localization to the SPB half-bridge area as shown previously in mitosis [9, 27]. By fluorescence microscopy, we further observed that Mps3 colocalized with another half-bridge component, Kar1, during yeast meiosis (Fig 1E). Together, these observations support the notion that Mps3 preferentially localizes to the half-bridge area of the SPB.
We noticed that the fluorescence intensity of GFP-Mps3 decreased at the SPB prior to SPB separation at metaphase I (Fig 1D and 1E). This was in contrast to the fluorescence of Kar1, which was fused with the RFP at its N-terminus and reached its highest intensity at the time of SPB separation (Fig 1E). To produce the Mps3 protein more abundantly during meiosis for western blot analysis of its N-terminus, we used the meiosis-specific DMC1 promoter to induce MPS3 expression (Fig 2B and 2C). Note that in these cells, the endogenous copy of MPS3 was present, because MPS3 is an essential gene for cell viability. The level of the full-length Mps3 remained relatively stable when yeast cells were arrested at prophase I by means of deletion of NDT80 (Fig 2A and 2B), which encodes a meiosis-specific transcription factor that is required for mid and late meiotic gene expression[28]. In contrast, the level of Mps3 fell precipitously when cells transited from prophase I to metaphase I as shown in Cdc20-depleted (PCLB2-CDC20) cells (Fig 2B), indicating that degradation of Mps3 occurs at the time of SPB separation.
We seek to understand how Mps3 is degraded. By western blotting, we observed the formation of an N-terminal fragment of Mps3 at the molecular weight of ~12 kDa, excluding the GFP tag, in meiotic cells (Fig 2B and 2C and S1A Fig). The abundance of this fragment peaked 6 h after the induction of meiosis in the wild-type background (S1A Fig), the time which corresponds to meiosis I SPB segregation [22]. In contrast, the N-terminal fragment was present at a much lower level, and its appearance was delayed in ndt80Δ cells arrested at prophase I (Fig 2B). To determine the biological relevance of the formation of this N-terminal fragment, we staged yeast cells at prophase I using the ndt80Δ allele and simultaneously depleted meiotic Ipl1p (PCLB2-IPL1), because in double ipl1 ndt80Δ mutant cells, SPBs can separate prematurely at prophase I without the activation of the mitotic cyclin-Cdk1 [21, 29, 30]. Crucially, in the absence of Ipl1, the N-terminal fragment of Mps3 became abundant at prophase I (Fig 2C). These results support the idea that Mps3 is cleaved at its N-terminus, the process of which appears to be correlated with SPB separation during meiosis I in budding yeast.
To test the hypothesis that Mps3 was cleaved prior to SPB separation, first we deleted the N-terminal 64 amino acids from Mps3 because removal of these residues abolishes Ndj1’s binding to Mps3 [22, 23]. The size of the N-terminal fragment became correspondingly smaller without the first 64 residues (Fig 2C), suggesting a site-specific cleavage event occurred beyond position 64. Next, on the basis of the molecular weight of the cleaved product, we replaced amino acids 65 to 93 of Mps3 with the TEV protease recognition sequence to generate mps3-nc, which effectively abolished the formation of the N-terminal fragment (Fig 2A and 2C and see below). Amino acid 93 ends at the acidic motif of Mps3 that has been described previously [25]. The above findings are consistent with the notion that Mps3 is cleaved specifically at its N-terminal domain, which binds to Ndj1 and is located in the nucleoplasma.
To determine whether Mps3 was cleaved at the SPB, we constructed a GFP-MPS3-RFP double-tagged allele (Fig 2A), which was under the control of the endogenous MPS3 promoter and the only functional copy of MPS3 in these cells. By time-lapse fluorescence microscopy, we observed cells that were undergoing prophase I to metaphase I transition, at which point Mps3 formed a distinctive focus at the SPB (Fig 2D). The ratio of the fluorescence intensity of RFP, fused to the C-terminus of Mps3, over that of GFP, fused to the N-terminus, increased up to 50% prior to SPB separation (Fig 2D), indicating that roughly a third of the SPB-associated Mps3 is cleaved at its N-terminal domain. This estimate likely underrepresents the real rate of N-terminal cleavage, because reloading of the full length Mps3 at the SPB appeared active before SPB reduplication at interphase II (Fig 2D and see below). At the end of meiosis II, majority of Mps3 was cleaved at the N-terminal domain, leaving only the C-terminal RFP-fused Mps3 fragment visible (Fig 2E). Of note, the cleaved C-terminal fragment of Mps3, which is expected to retain its transmembrane domain and therefore remains anchored at the nuclear periphery, has a longer half life than the cleaved N-terminal fragment, making it possible for us to directly visualize Mps3 cleavage in live meiotic yeast cells by fluorescence microscopy. The above results also indicate that cleavage of Mps3 appears to take place throughout meiosis, and therefore meiosis II cells can serve as a reliable cytological readout of Mps3 cleavage.
As a control, we swapped the tags to construct an RFP-MPS3-GFP allele, which was also under the control of its endogenous promoter and the only functional copy of MPS3 in these cells (S2 Fig). For the same reason, the full length of the Mps3 protein appeared continuously produced, then reloaded at the SPB during meiosis I. We therefore focused on cells in meiosis II when Mps3 was no longer produced, and found that in these cells, the GFP signal but not the RFP was retained much longer at the SPBs (S2 Fig). Combining the above biochemical and cytological findings, we conclude that the N-terminal domain of Mps3 is cleaved at the SPB during yeast meiosis.
To determine whether membrane-bound Mps3, in addition to SPB-associated Mps3, was also cleaved at the nuclear periphery, we used the meiosis-specific DMC1 promoter to express GFP-MPS3-RFP more abundantly (Fig 3A). Overproduced Mps3 protein localized robustly both to the SPB and to the nuclear periphery in meiosis I (Fig 3), demonstrating cytologically that the heterologous PDMC1-MPS3 construct overexpressed MPS3 during meiosis, as shown by western blotting (Fig 2B and 2C). To observe Mps3 cleavage, we focused again on cells in meiosis II when both the expression of PDMC1-MPS3 and the production of Mps3 protein were terminated, permitting us to follow the fate of existing GFP-Mps3-RFP. Using time-lapse microscopy, we found that the RFP signal was retained much longer at the nuclear periphery at the end of meiosis (Fig 3A and 3B), indicating that the GFP fused to the N-terminus was removed earlier than the RFP fused to the C-terminus of Mps3. Therefore, Mps3 proteins located at the inner nuclear membrane are also cleaved during meiosis, just as are the ones associated with the SPB. Taking these observations together, we conclude that Mps3 is cleaved at its N-terminal domain, and that its cleavage can take place both at the SPB and at the nuclear periphery.
Using a protein mass spectrometry-based approach [31], we determined that Mps3 is phosphorylated at serine 70 during meiosis (Fig 4A). As we have shown previously, we used TAP tagged Spc97, a component of the γ-tubulin ring complex of the yeast SPB, to purify SPBs from yeast cells that were induced to undergo synchronous meiosis [22]; the enriched SPBs were then subject to SPB phosphoproteome (see materials and methods). By protein mass spectrometry, 38% of the Mps3 protein sequence was covered with a total of 39 spectra, of which 4 showed Mps3 phosphorylation at S70 (Fig 4A shows one of them). Because the Mps3 protein analyzed was copurified with Spc97-TAP, we conclude that SPB-associated Mps3 is phosphorylated at S70.
To start to determine how Mps3-S70 is phosphorylated during yeast meiosis, we raised a phosphospecific antibody against the phosphorylated form of Mps3-S70 (Fig 4B). Alkaline phosphatase treatment of affinity-purified Mps3-TAP extracted from meiotic yeast cells removed Mps3-S70 phosphorylation (Fig 4B), demonstrating the specificity of our antibody and further confirming the phosphorylation of Mps3 at S70 (Fig 4B). To determine the timing of Mps3-S70 phosphorylation, we performed western blotting on protein extracts from progressing meiotic yeast cells, of which the endogenous Mps3 was tagged with the V5 epitope at its C-terminus (Fig 1B). As shown in Fig 4C, Mps3-S70 was highly phosphorylated 6 h after induction of meiosis, the time which roughly corresponds to prophase I to metaphase I transition and SPB separation. We also observed Mps3-S70 phosphorylation in cells that were arrested at prophase I by ndt80Δ (S3 Fig), indicating that phosphorylation of Mps3 takes place at meiotic prophase I.
To determine whether S70 phosphorylation regulates Mps3 cleavage, we generated alleles of mps3-S70A, which abolished phosphorylation at position 70, and mps3-S70D, which mimicked phosphorylation at this position (Fig 2A). Both the mps3-S70A and mps3-S70D constructs were under the control of the DMC1 promoter for expression during meiosis (Fig 4C). As shown in Fig 4D, the level of the cleaved N-terminal product was dramatically reduced in mps3-S70A cells, compared to that of mps3-S70D cells. Therefore phosphorylation of Mps3 at S70 increases the efficiency of its cleavage during meiosis.
In our search for the protease that is responsible for Mps3 cleavage, we hypothesized that the proteasome regulates Mps3 cleavage because previous work has shown the specific regulation of proteasome activity during yeast meiosis [32, 33]. To inactivate the proteasome, we induced yeast cells to undergo synchronous meiosis and 3 h later added MG132, a potent inhibitor of the proteasome, to the yeast culture medium (Fig 5A diagram). In normally progressing meiotic cells, the cleaved Mps3 N-terminal fragment appeared most abundantly 6 h after the induction of meiosis (Fig 5A). In contrast, addition of MG132 greatly inhibited the production of the Mps3 N-terminal fragment, indicating that Mps3 cleavage is regulated by the proteasome activity (Fig 5A and 5B). Of note, in cells treated with MG132, Mps3 protein degradation was also inhibited, resulting in the increased level of the full length Mps3 (Fig 5A). Using an alternative genetic approach to inactivate the proteasome, we deleted PRE9, which encodes the nonessential α3 subunit of the 20S proteasome [34]. In the absence of α3, the α4 subunit can take over α3’s place to form a functional proteasome [34], but the activity of this α3-less proteasome is impaired at an elevated temperature [35]. In wild-type cells that were induced to undergo meiosis at 33°C, we found that Mps3 was cleaved just as those undergoing meiosis at 30°C (Fig 2B and S1A Fig), the optimal temperature for yeast sporulation. In contrast, cleavage of Mps3 was inhibited in pre9Δ cells as determined by western blotting (S1A Fig), further demonstrating that the proteasome activity is required for Mps3 cleavage.
By fluorescence microscopy, we found that in both MG132-treated and pre9Δ cells, noncleaved Mps3 with intact N-terminus remained at the nuclear periphery and the SPBs at the end of meiosis, just as it was found in mps3-nc cells (Fig 5C and S1B Fig), demonstrating that Mps3 cleavage was inhibited in these cells. Taking these findings together, we conclude that the proteasome activity regulates Mps3 cleavage during yeast meiosis.
Because the timing of Mps3 cleavage correlates to that of SPB separation in meiosis I, we hypothesized that cleavage of Mps3 was necessary for SPB separation. Previously we have shown that depletion of Ipl1 during meiosis can lead to premature SPB separation at prophase I; we therefore used the ipl1 ndt80Δ strain as a tool to assay SPB separation (Fig 6A). Because MPS3 is an essential gene, we overexpressed GFP-tagged MPS3 alleles specifically in meiosis with the heterologous PDMC1-GFP-MPS3 expression constructs (Fig 6A and 6B). As such, any mps3 mutant phenotype observed would be dominant negative. When the wild-type GFP-MPS3 allele was expressed, about 58% of the cells separated SPBs 10 h after the initiation of meiosis; among them, 9% had completed the second round of SPB duplication and separation (Fig 6B). Abolishing S70 phosphorylation (mps3-S70A) drastically inhibited SPB separation, whereas the mps3-S70D mutation retained the wild-type level of separation (Fig 6B). As expected, eliminating Mps3 cleavage in mps3-nc cells severely inhibited SPB separation (Fig 6B). Crucially introduction of the TEV protease into the mps3-nc cells restored SPB separation to the wild-type level (Fig 6B). Therefore both phosphorylation and cleavage of Mps3 are important for SPB separation in ipl1 ndt80Δ cells during meiosis.
If cleavage of Mps3 facilitates SPB separation, we expect that overexpression of Mps3(Δ1–93), which lacks the N-terminal domain, would lead to a higher level of SPB separation in the ipl1 ndt80Δ strain at prophase I. Indeed, 95% of mps3(Δ1–93) cells separated their SPBs prematurely, and more than 66% of them completed the second round of SPB duplication and separation in the ipl1 ndt80Δ background (Fig 6B). Because the Mps3(Δ1–93) protein lacked the cleavage site, N-terminal GFP tagged Mps3(Δ1–93) was found at the SPB and the nuclear membranes at the end of meiosis (Fig 5C). These findings support the notion that removal of the N-terminal domain from Mps3 promotes SPB separation.
To further determine the effect of Mps3 cleavage on SPB separation and cell cycle progression in meiosis, we generated an mps3-nc mutant allele that was under the control of its endogenous promoter and served as the only copy of the MPS3 gene. In the absence of Mps3 cleavage, yeast cells largely retained viability during vegetative growth and were able to be induced to undergo meiosis (Fig 6C). Although early meiotic progression appeared to be on schedule in mps3-nc mutant cells, 35% of the mutant cells aborted meiosis without SPB separation and spore formation after 12 h in the sporulation medium, whereas more than 90% wild type cells formed tetrads or dyads at that time (Fig 6D). Importantly, among the mutant cells that underwent second round of SPB separation, 46% of them showed unequal SPB separation on the basis of the fluorescence intensity of Tub4-RFP and GFP-Mps3 (Fig 6C, arrows), indicating that noncleavable Mps3 inhibits proper SPB disjunction during meiosis. These observations demonstrate an essential role of Mps3 cleavage in SPB separation and cell cycle progression during yeast meiosis.
To determine whether posttranslational modification of Mps3 was necessary for SPB separation during mitosis, we expressed MPS3 mutants in vegetative yeast cells from the S288C background using a galactose inducible promoter (Fig 7 and S4 Fig). Overexpression of mps3-nc was lethal to yeast cells (Fig 7A), but this lethal phenotype could be suppressed by pom152Δ (Fig 7B), indicating that the lethality caused by mps3-nc is due to the mitotic defects occurring at the SPB, because pom152Δ restores proper SPB function when Mps3 is absent [36]. To pinpoint the role of Mps3 in SPB separation, we synchronized yeast cells at G1, induced MPS3 expression, and then released these cells from G1 arrest to analyze SPB duplication and separation (Fig 7C). Galactose-induced Mps3, including Mps3-nc, first appeared at the SPB (Fig 7C). Notably, Mps3-nc started to form protein aggregates at the nuclear periphery about 80 min after the activation of the galactose promoter (Fig 7C). On the basis of the fluorescence intensity of Tub4-RFP, SPB duplication appeared comparable between wild-type and mps3-nc cells (Fig 7D), and wild-type and mutant proteins were produced in a similar time frame (Fig 7E). However, cells expressing mps3-nc were essentially blocked at anaphase with large buds (Fig 7F). In large-budded cells that overexpressed mps3-nc, 40% failed to separate their SPBs, and if they were separated, 49% showed unequal SPB separation (Fig 7G), demonstrating that mps3-nc cells either failed to separate or misseparated their SPBs. Failure to properly partition SPBs was also observed in cells overexpressing mps3-S70A (S4 Fig). We note that, under the control of its endogenous promoter, mps3-nc cells are viable, albeit with a slight growth defect. Together, our findings indicate that posttranslational modification of Mps3 is also critical for proper SPB separation in mitosis.
The topology of Mps3 resembles a single-pass cell-surface receptor, which forms protein oligomers, typically dimers, upon ligand binding [37]. To determine whether Mps3 forms protein oligomers, we constructed an MPS3-TAP/MPS3-3HA strain, induced cells to undergo synchronous meiosis and performed TAP protein affinity purification to determine the physical interaction between Mps3-TAP and Mps3-3HA (Fig 8A–8C). Note that the tagged alleles of MPS3 were incorporated at the endogenous MPS3 locus and served as the only functional copies of MPS3. By western blotting, we found that Mps3-3HA was copurified with Mps3-TAP (Fig 8A), indicating that Mps3 can form homo oligomeric protein complexes in vivo. Crucially, Mps3-TAP appeared to interact more strongly with the full-length Mps3-3HA, but minimally with the N-terminally cleaved form of Mps3-3HA (Fig 8A and 8B), suggesting that without its N-terminal domain, Mps3 oligomer formation is weakened. This finding is consistent with the observation that mps3(Δ1–93) cells, which lacked the N-terminus domain of Mps3, separated their SPBs prematurely at prophase I (Fig 6B). In addition, we performed sucrose gradient fractionation of affinity purified Mps3 protein complexes and observed that Mps3-TAP*, of which the TAP tag was essentially removed enzymatically (see materials and methods), and Mps3-3HA peaked in the same sucrose fractions (Fig 8B and 8C), supporting the idea that Mps3 forms protein oligomers in vivo.
In this report, we have shown for the first time that a SUN-domain protein is cleaved at the nuclear membranes. Mps3 localizes to the inner nuclear membrane and to the nuclear side of the SPB half-bridge to mediate SPB cohesion, and its cleavage at the N-terminal domain plays a role in meiotic SPB separation. Posttranslational modifications, including protein phosphorylation [31], have been found to take place extensively at the yeast SPB, but irreversible protein cleavage at the half-bridge was not known previously. Our finding of Mps3 cleavage at the SPB and the nuclear membranes therefore provides insight into the mechanism of half-bridge disassembly, which is critical for proper centrosome separation and chromosome segregation.
Three lines of evidence we have obtained demonstrate that Mps3 is cleaved at its N-terminal domain, which is positioned in the nucleoplasm. First, our biochemical and genetic analysis reveals that Mps3 is cleaved at the acidic motif located at its N-terminal domain. Second, phosphorylation of serine 70, which is adjacent to the acidic motif, promotes Mps3 cleavage, perhaps by increasing the acidity or by another means of modifying the local environment that contains the cleavage site. Finally, our cytological observations unambiguously show that Mps3 can be cleaved at both the SPB and the nuclear periphery, providing strong evidence that the protease responsible for Mps3 cleavage is located at or adjacent to the inner nuclear membrane.
We propose that cleavage of Mps3, which removes the binding site for its nuclear ligands such as Ndj1, regulates Mps3 oligomer disassembly and half-bridge breakage (Fig 8D). Similar to a single-pass cell-surface receptor, upon which ligand binding on one side of the membrane often triggers conformational changes of the peptide domain that is located on the opposite side of the membrane [37], Mps3, when cleaved from its N-terminal domain, appears to have a reduced affinity for oligomer formation (this study). Our findings therefore indicate that the nucleoplasm-localized N-terminal domain of Mps3 has a profound impact on the conformation of its C-terminus, which is positioned in the lumen of the nuclear envelope and contains a coil-coiled segment and the SUN-domain, both of which have been implicated in protein trimerization [38–40]. Our findings also suggest that the bridge which tethers duplicated SPBs is disassembled, at least partially, at the point of SPB separation. Phosphorylation at the cytoplasmic side of the half-bridge, which is a reversible process, also regulates SPB separation [15, 16]. Therefore, posttranslational modifications of half-bridge components on both sides of the nuclear envelope appear to work either simultaneously or redundantly to ensure the timely separation of yeast centrosomes. Although the cyclin-dependent kinase Cdk1 and the polo kinase Cdc5 both have been implicated in phosphorylating the half-bridge component Sfi1 [15, 16], the nature of the kinase responsible for Mps3-S70 phosphorylation remains to be determined. We also note that in the absence of Mps3 cleavage, to some degree duplicated SPBs can be separated albeit with many cells showing unequal partitioning of SPB components. Our observations therefore indicate that the cytoskeleton, for example microtubule-based, forces could overwhelm the structural integrity of the half-bridge during SPB separation.
Because cleavage of Mps3 depends on the proteasome activity, we speculate that the proteasome, which is highly concentrated in the yeast nucleus [41], acts as the protease responsible for Mps3 cleavage. The proteasome is capable of mediating site-directed peptide cleavages after either the acid or basic amino acid residues [42]. Alternatively, proteasome activity is required for activating the yet unidentified protease that is directly responsible for Mps3 cleavage. In either scenario, the proteasome would play a crucial role in modifying Mps3 located at both the SPB half-bridge and the nuclear membranes. Noncleavable Mps3 accumulates at the nuclear periphery (this study), indicating that in addition to Mps3 cleavage, the proteasome activity also regulates Mps3 protein homeostasis at the nuclear membranes.
Mps3 is highly expressed in meiosis and regulates SPB cohesion (this study) and telomere attachment to the nuclear periphery [23], which raises an intriguing question: Is Mps3 cleavage specific to yeast meiosis? Our preliminary study of Mps3 in vegetative yeast cells using the inducible galactose promoter failed to detect its N-terminal cleaved fragment (Fig 7E), owing to a much short half-life of the cleaved products during mitosis. Alternatively, perhaps a specialized modification of Mps3 occurs in yeast meiosis, at which point SPBs remains tethered and telomeres are clustered for an extended period of time. Indeed, the meiosis-specific telomere-associated protein Ndj1 binds specifically to the N-terminus of Mps3 [22, 23], and ectopic expression of NDJ1 in vegetative yeast cells is lethal [22]. Upon the transition of prophase I to metaphase I during meiosis, Ndj1 is degraded [22] and concomitantly Mps3 is cleaved, which leads to irreversible structural modifications at the SPB half-bridge and the telomeres. Therefore, at the nuclear periphery a unique signal transduction cascade appears to be in place during meiosis to regulate Mps3 phosphorylation and protein cleavage. (Mps3 cleavage and cell cycle progression).
Although SUN-domain proteins vary at their N-termini [24], Mps3 shares a high degree of conservation with SUN1 from mammals, in particular at the N-terminus [9, 26]. Phosphorylation at the N-terminal domain has been previously reported for SUN1 from mammals and SUN-1 in C. elegans [43, 44]. Intriguingly, at the meiotic G2, SUN-1 shows decreased affinity to an antibody raised specifically against its N-terminus, a finding which has been interpreted as epitope masking [43]. Alternatively, SUN-1 could be cleaved at its N-terminus, similar to our observation of Mps3 cleavage during yeast meiosis. Before their separation, animal centrosomes are also attached to, although not embedded in, the nuclear envelope in a SUN-domain-protein-dependent manner [45]. Therefore, lessons learned from yeast centrosome separation mediated by irreversible protein cleavage have implications for understanding the regulation of centrosome disjunction and other SUN-domain protein-mediated nuclear activities in all eukaryotes.
Yeast strains and plasmids used in this study are listed in S1 Table and S2 Table. We used a PCR-based method [46] to introduce protein tags to the C-terminus of Mps3, including alleles of MPS3-V5, MPS3-3HA, MPS3-TAP, MPS3-GFP and MPS3-RFP. Using a similar strategy, we constructed NUP49-RFP. Nup49 is a core component of the nuclear pore complex and serves as a marker for the nuclear envelope (Fig 1C). The alleles of TUB4-RFP and SPC97-TAP have been reported previously [22]. These tagged alleles are the only functional copies of the corresponding genes in the yeast genome. A similar PCR-based method was used to replace the PRE9 open reading frame with the KanMX cassette. Correct transformations were confirmed by colony-based diagnostic PCR. PCR primers are listed in S3 Table. The following mutant alleles have been reported previously: ndt80Δ [28], PCLB2-CDC20 [47], PCLB2-IPL1 [21], PCLB2-MPS3 [22], pom152Δ [36] and pdr5Δ.
We used the GAL1 promoter to induce expression of MPS3 and its mutant alleles with galactose, and the DMC1 promoter to express MPS3 and its mutant alleles in meiosis. Using pRS305 as the cloning vector, we constructed alleles of PGAL1-GFP-MPS3 (pHG323) and PDMC1-GFP-MPS3 (pHG350). Other MPS3 alleles were derivatives of these two. These heterologous MPS3 constructs were linearized by StuI and incorporated at the endogenous MPS3 locus by yeast transformation. We used a PCR-based mutagenesis method to generate point mutations of mps3-S70A and mps3-S70D, deletion mutations of mps3(Δ1–64) and mps3(Δ1–93), and a sequence replacement mutation (mps3-nc), of which amino acids 65 to 93 of Mps3 were replaced by the Tobacco Etch Virus (TEV) recognition sequence ENLYFQG (Fig 2A). The expression of the TEV protease was under the control of the DMC1 promoter (pHG394). We cloned a fragment of 1 kb DNA sequence in front of the MPS3 open reading frame to serve as the MPS3 promoter and constructed alleles of PMPS3-GFP-MPS3 (pHG454) and PMPS3-RFP-MPS3 (pHG468) using plasmid pRS306 as the backbone. With a similar approach, we constructed alleles of PKAR1-GFP-KAR1 (pHG465) and PKAR1-RFP-KAR1 (pHG501). About 1 kb DNA sequence upstream of the KAR1 open reading frame was cloned to serve as the KAR1 promoter. Plasmids were linearized before their transformation to yeast cells.
To construct the double-tagged alleles of PMPS3-GFP-MPS3-RFP and PMPS3-RFP-MPS3-GFP (Figs 2 and 3), plasmids pHG454 and pHG468 were linearized with the restriction enzyme AflII, which cuts at the MPS3 promoter, and then were transformed into MPS3-RFP and MPS3-GFP yeast cells, respectively. Positive yeast transformants were patched onto the 5’-fluoroorotic acid medium to select recombinants that had excised the untagged copy of MPS3. The double-tagged alleles of MPS3, which served as the only functional copy of MPS3 in the yeast genome, were determined by microscopy and colony PCR. A similar approach was used to construct GFP-MPS3 and GFP-mps3-nc alleles under the control of the endogenous MPS3 promoter (Fig 6C and 6D). Plasmids pHG454 and pHG459 were linearized with AflII, and yeast transformants were selected on 5’-fluoroorotic acid plates for recombinants, which were then confirmed by microscopy and colony PCR.
In cells where the heterologous MPS3 construct was under the control of either the DMC1 or GAL1 promoter, the wild-type copy of MPS3 was retained because MPS3 is an essential gene. As such, the phenotypes of mps3-nc, mps3-S70A and mps3(Δ1–93) shown in Figs 6, 7 and S4 are dominant negative.
Yeast cells were grown in YPD (1% yeast extract, 2% peptone and 2% dextrose) [48] at 30°C. To induce synchronous meiosis, YPD cultures were diluted with YPA (1% yeast extract, 2% peptone and 2% potassium acetate) to reach OD (optical density, λ = 600nm) of 0.2 and incubated at 30°C for about 14 hours to reach a final OD of 1.6. Then, yeast cells were washed once in water and resuspended in 2% potassium acetate, the point of which was defined as time zero after the induction of meiosis. Cell aliquots were withdrawn at indicated times for microscopy and protein extraction. The pre9Δ cells were grown in the YPA media at 25°C and induced to undergo meiosis at 33°C to inactive the proteasome (S1 Fig). To chemically inactivate the proteasome activity (Fig 5), yeast cells with pdr5Δ were induced to undergo synchronous meiosis for 3 hours; the culture was then split into three fractions: untreated, DMSO only and DMSO with MG132 (50 μM final concentration). Cells aliquots were withdrawn at indicated times, and protein extracts were prepared for western blotting.
For mitotic experiments shown in Figs 7 and S4, synthetic complete (SC) medium [48] with 2% raffinose was used. To arrest yeast cells at G1, we grew them in the raffinose medium to reach OD of 0.4, then added 10 μg/ml alpha factor. Galactose (2% final concentration) was added to the culture medium to induce the expression of the GAL1 promoter 30 minutes before the removal of the alpha factor, upon which cells resumed mitosis. To determine cell viability (Fig 7A), yeast cells were grown overnight in YPD liquid medium to reach saturation, 10 fold diluted, spotted onto SC plates with either 2% dextrose or 2% galactose, and then incubated at 30°C for about two days.
Time-lapse live-cell fluorescence microscopy was carried out on a DeltaVision imaging system (GE Healthcare Life Sciences) at 30°C. We used a 60x (NA = 1.40) objective lens on an inverted microscope (IX-71, Olympus). Microscopic images were acquired with a CoolSNAP HQ2 CCD camera (Photometrics). Pixel size was set at 0.10700 μm. Time intervals were set at 2 to 5 minutes, and optical sections at 12, each with 0.25–0.5μm thickness. Ultra-high signal-to-background hard coated custom filter sets were used. For GFP, the excitation spectrum was at 470/40 nm, emission spectrum at 525/50 nm; for RFP, excitation was at 572/35, and emission at 632/60 nm. During microscopy, yeast cells were grown on agarose pads as described previously [21]. To minimize photo toxicity to the cells and photo bleaching to fluorophores, we used neutral density filters to limit excitation light to 32% or less of the normal output from the fluorescence illuminator. Exposure time was set at 0.1 second or less.
For time-course experiments shown in Figs 6, 7 and S4, at least two independent experiments were performed, and more than 100 cells were analyzed for each strain at each time point.
Acquired microscopy images were deconvolved using the SoftWorx package (GE Healthcare Life Sciences). Projected images are used for image display. Display of single optical sections is specified in figure legend (Fig 3). We used the SoftWorx measuring tools to determine the fluorescence intensity of line scans. Pixel intensities were plotted along the line as shown in Figs 1D, 1E and 3B. To determine the Tub4-RFP intensity at the SPB (Figs 7D and S4B), we defined a 6 x 6 pixel area as the position of SPB. The mean background intensity was subtracted from the region of interest to yield the net intensity of Tub4-RFP.
We used Spc97-TAP and Mps3-TAP for affinity purification of the SPB and Mps3 as reported previously [22, 49]. Briefly, yeast cells were induced to undergo synchronous meiosis for 6 hours; about 10g of cells were harvested and ground in the presence of liquid nitrogen. TAP tagged proteins were then enriched using the epoxy-activated M-270 Dynabeads (Thermo Fisher Scientific, Cat#14305D) that were cross-linked to rabbit IgG (Sigma-Aldrich, Cat#I5006).
Purified SPB proteins were precipitated and digested with trypsin. The resulting peptide mixture was processed through a TiO2-based phosphopeptide enrichment [50]. The resulting enriched phosphorylated peptide samples were directly pressure-loaded onto a C18 column (in-house packed, 15 cm x100 um, 5 μ Gemini C18 resin (Phenomenex)) and were then analyzed by online nanoflow liquid chromatography tandem mass spectrometry (LC-MS/MS) on an Agilent 1200 quaternary HPLC system (Agilent, Palo Alto, CA) connected to an LTQ Orbitrap Velos mass spectrometer (Thermo Fisher Scientific) through an in-house built nanoelectrospray ion source with a linear gradient of 5% to 40% acetonitrile in 0.1% formic acid in 150 min with a flow rate of ~300 nL/min (through split). MS instrument method consisted of one FT full-scan mass analysis (300–1600 m/z, 120,000 resolving power at m/z = 400) followed by 20 data-dependent collision-induced dissociation (CID) MS/MS spectra (normalized collision energy at 35%) with dynamic exclusion for 120 s. Application of mass spectrometer scan functions and HPLC solvent gradients were controlled by the Xcalibur data system (Thermo Fisher Scientific).
Protein identification was carried out with Integrated Proteomics Pipeline—IP2 (Integrated Proteomics Applications, Inc., San Diego, CA. http://www.integratedproteomics.com/). Briefly, MS/MS spectra were extracted using RawXtract (version 1.9.9) [51] and searched with ProLuCID algorithm [52] against a Saccharomyces cerevisiae database concatenated to a decoy database in which the sequence for each entry in the original database was reversed [53]. A static modification (+ 57.02146) on cysteine was added into the search due to alkylation of cysteine residues. For phosphopeptide analysis, the ProLuCID search was performed using differential modification of serine and threonine due to phosphorylation (+79.9663). The precursor mass tolerance was set at 30 ppm and fragment mass tolerance was set at 600 ppm. The enzyme specificity was semi-tryptic, with number of missed cleavages at 2. ProLuCID search results were then assembled and filtered using the DTASelect (version 2.0) program [54]. For protein identification, the protein false positive rate was kept below one percent and the average mass deviation was less than 5 ppm. For phosphopeptide identification, only modified peptides were considered and the peptide false positive rate was set at less than one percent.
Using a TAP-purification method similar to that described above, we purified Mps3-TAP by affinity purification. To remove the phosphate group from Mps3, purified Mps3-TAP was treated with alkaline phosphatase from calf intestine (CIP, NEB M0290L) at 37°C for 30 min.
Affinity purified Mps3 protein complexes were subject to sucrose gradient fractionation. First, Mps3-TAP was removed from the magnetic beads with TEV protease treatment (0.5mg/ml final concentration) at 16°C for 2 h. Then, cleaved Mps3-TAP* protein samples (note that the protein A peptide domain was now removed from Mps3-TAP) were laid on top of a 7%-47% 12 ml linear sucrose gradient (Gradient Master, Biocomp Instruments) and centrifuged at 39,000 rpm for 18 h at 4°C with the SW41 rotor on a Beckman Ultra 100 centrifuge. Finally, we used an 18G needle to collect 1 ml fractions, the top 9 fractions were shown in Fig 8B. Presence of Mps3 was determined by western blotting.
Yeast aliquots were withdrawn at indicated times for protein extraction with the trichloroacetic acid (TCA) method as described previously [55]. Briefly, 7 to10 ml of yeast cells were collected, resuspended in 2.5% ice cold TCA, and then incubated at 4°C for 10 minutes. Cell pellets were stored at -80°C, and proteins were extracted by bead beating with a mini bead-beater homogenizer for 1 minute at 4°C. For the mitotic experiments shown in Figs 7 and S4, 2 ml yeast cells were precipitated in the presence of 20 mM NaOH. GFP-tagged proteins were detected by an anti-GFP mouse monoclonal antibody (1:10,000 dilution, Thermo Fisher Scientific, cat#GF28R). HA-tagged proteins were detected by an anti-HA mouse monoclonal antibody (1:10,00 dilution, 12CA5, Sigma). Mps3-V5 was detected by an anti-V5 antibody (1:5,000, Thermo Fisher Scientific, cat#R960-25). We used an anti-TAP antibody (1:10,000, Thermo Fisher Scientific, cat#CAB1001) to determine the presence of Mps3-TAP. This antibody was raised against the C-terminus of the TAP construct after removing the protein A domain by the TEV protease and therefore recognizes TEV-cleaved Mps3-TAP* (Fig 8B). The following synthetic peptide Ac-DKEND(pS)AYEMFKC-amide was used to generate the phosphospecific antibodies against the phosphorylated Mps3-S70 (New England Peptide Inc., Gardner, MA 01440, USA). Raised polyclonal antibodies were affinity purified before use (Fig 4B). A home-made Tub2 antibody [55] was used to detect the level of β-tubulin, and the level of Pgk1 was probed by a Pgk1 antibody (Thermo Fisher Scientific, cat#PA5-28612). The levels of β-tubulin and Pgk1 served as loading controls. Horseradish peroxidase-conjugated secondary antibodies, goat anti-mouse and goat anti-rabbit (Bio-Rad, cat#1706516 and 1705046), were used to probe the proteins of interest by an enhanced chemiluminescence kit (Bio-Rad, cat#1705060).
|
10.1371/journal.pgen.1003007 | Regulation of ATG4B Stability by RNF5 Limits Basal Levels of Autophagy and Influences Susceptibility to Bacterial Infection | Autophagy is the mechanism by which cytoplasmic components and organelles are degraded by the lysosomal machinery in response to diverse stimuli including nutrient deprivation, intracellular pathogens, and multiple forms of cellular stress. Here, we show that the membrane-associated E3 ligase RNF5 regulates basal levels of autophagy by controlling the stability of a select pool of the cysteine protease ATG4B. RNF5 controls the membranal fraction of ATG4B and limits LC3 (ATG8) processing, which is required for phagophore and autophagosome formation. The association of ATG4B with—and regulation of its ubiquitination and stability by—RNF5 is seen primarily under normal growth conditions. Processing of LC3 forms, appearance of LC3-positive puncta, and p62 expression are higher in RNF5−/− MEF. RNF5 mutant, which retains its E3 ligase activity but does not associate with ATG4B, no longer affects LC3 puncta. Further, increased puncta seen in RNF5−/− using WT but not LC3 mutant, which bypasses ATG4B processing, substantiates the role of RNF5 in early phases of LC3 processing and autophagy. Similarly, RNF-5 inactivation in Caenorhabditis elegans increases the level of LGG-1/LC3::GFP puncta. RNF5−/− mice are more resistant to group A Streptococcus infection, associated with increased autophagosomes and more efficient bacterial clearance by RNF5−/− macrophages. Collectively, the RNF5-mediated control of membranalATG4B reveals a novel layer in the regulation of LC3 processing and autophagy.
| Autophagy is an intracellular catabolic process by which a cell's own components are degraded through the lysosomal machinery. Autophagy is implicated in various cellular processes such as growth and development, cancer, and inflammation. Using biochemistry, cell biology, and genetic models, we identify a ubiquitin ligase that limits autophagy in the absence of an inducing stimulus (e.g. starvation). The control of basal autophagy is mediated by the ubiquitin ligase RNF5 through its regulation of the membrane-associated ATG4B protease. Using RNF5 mutant mice we demonstrate the implications of this regulation for host defense mechanisms that limit intracellular infection by bacterial pathogens.
| Autophagy is an intracellular catabolic process by which cellular components are degraded through the lysosomal machinery. Conserved from yeast to humans, autophagy is fundamental to eukaryotic cell homeostasis [1], [2]. Autophagy functions in diverse cellular processes such as growth and development, cancer, and inflammation [3]–[5], and is implicated in both cell survival and death, depending on the cell type and stress conditions. Accordingly, autophagy has been associated not only with disease progression but also with its prevention [6], [7]. Interestingly, while certain viruses and bacteria can subvert and manipulate autophagic pathways during establishment of infection, autophagy plays a protective role against intracellular replication of several pathogens including group A Streptococcus (GAS) [8], [9]. Given the broad importance of autophagy in cell biology, it is of great interest to define the mechanisms underlying its control under normal and stress-related conditions.
Autophagy takes place through a series of steps that include initiation, elongation, and formation of autophagosomes, followed by fusion with lysosomes, and finally maturation and degradation of the autolysosome [10], [11]. Each step in this process involves a number of autophagy (ATG)-specific proteins that control a highly coordinated cascade of events culminating in autolysosome formation [12]. Among these, ATG7 and ATG3 conjugate mammalian LC3 homologues to phosphatidylethanolamine (PE), and ATG7 and ATG10 conjugate ATG12 to ATG5 [13], [14]. The cysteine protease ATG4 contributes to this chain of events by cleaving the LC3 C-terminal domain to generate LC3-I [15]. Consequently, LC3-I is converted by ATG7 and ATG3 to LC3-II, which is essential for phagophore and autophagosome formation [16]–[18]. ATG4 also plays a role in the final step of autophagy by deconjugating LC3-II, enabling LC3 to be released from autolysosomal membranes and recycled [19]–[21].
Four mammalian homologues of yeast ATG4 have been identified: ATG4A, ATG4B, ATG4C, and ATG4D [22]. ATG4B has broad specificity for the mammalian ATG8 homologues GATE-16, GABARAP, and LC3, whereas ATG4C and ATG4D show minimal activities toward LC3 substrates [23], [24]. Following cleavage by caspase, ATG4D stimulates GABARAP-L1 processing and autophagosome formation [25]. Studies of ATG4 gene knockout mice have revealed some differential functions of the isoforms; while ATG4C−/− mice exhibit marked changes in autophagic activity following prolonged starvation [26], ATG4B−/− mice show a clear reduction of basal- and starvation-induced autophagy in all tissues, associated with impaired proteolytic cleavage of LC3 orthologs [27].
The availability of LC3 is regulated co-translationally, suggesting that ATG4B is not a limiting factor in the control of LC3 processing and the early stages of autophagy [16]. Nonetheless, accumulative evidence suggests that ATG4B is regulated in a manner that has concomitant effects on LC3 processing. For example, upregulating ATG4 by Egr1 or ARH1 is associated with increased LC3 processing and autophagy in lung tissues and ovarian cancer [28], [29]. Moreover, disruption of ATG4B inhibits processing of LC3 paralogues and autophagy [27], [30], [31]. Collectively, these observations identify a critical role for ATG4B in control of autophagy. While growing evidence suggests that LC3 processing is induced prior to the formation of the pre-initiation ATG1/13 complex, the mechanisms controlling basal and induced levels of LC3 are largely unknown. Here, we have investigated the relationship between ATG4B activity and LC3 processing, and demonstrate that the ubiquitin ligase RNF5 controls the stability, and hence availability, of a membrane-localized fraction of ATG4B. Accordingly, RNF5 limits ATG4B processing of membranal LC3 with concomitant effects on autophagy.
RNF5 (also named RMA1), an 18-kDa RING finger E3 ligase, has been implicated in C. elegans muscle structure through its control of UNC-95, a LIM domain-containing protein involved in maintenance of dense bodies, the muscle attachment sites [32] and in the molting process [33]. In mammalian cells, RNF5 regulates cell motility by ubiquitinating the focal adhesion protein paxillin [34]. Deregulated expression of RNF5 in the muscle of RNF5 transgenic mice results in the formation of inclusion body myositis, an endoplasmic reticulum (ER) stress-associated muscular disorder [35]. A link to ER stress was also demonstrated through RNF5's role in ER-associated protein degradation (ERAD), where it contributes to clearance of misfolded proteins [36], [37]. RNF5 promotes tumor cell resistance to cytoskeletal-targeting anticancer agents, and its expression is inversely correlated with survival in breast cancer and melanoma patients [38]. Lastly, RNF5 contributes to the cellular response to infection by controlling the stability of the mitochondrial proteins MITA and MAVS, which are involved in antiviral innate immune signaling [39], [40]. RNF5-dependent ubiquitination of SOP-A affects Salmonella trafficking from endosomes/vacuoles to the cytosol [41].
Given the multiple effects of RNF5 on ER stress, innate immunity, and bacterial infection, processes that are each influenced by autophagy, we have examined the possibility that RNF5 may play a role in the control of autophagy. In the present study, we demonstrate that RNF5 negatively regulates basal levels of autophagy by mediating the ubiquitination and degradation of a membranal-associated pool of ATG4B. Macrophages from RNF5 knockout (KO) mice exhibit more efficient processing of GAS, and RNF5 KO mice are less susceptible to lethal infectious challenge by this leading bacterial pathogen.
Given the various effects of RNF5 on ER stress and innate immune pathways, processes that are influenced by autophagy, we examined the possibility that RNF5 may play a direct role in the control of autophagy. A cDNA library screen using a yeast-based functional assay for ATG4B inhibitors identified RNF5 as a candidate regulator (1 in 12 hits among 2×105 colonies; Figure S1). To confirm a possible interaction of RNF5 with ATG4B, we examined the association between exogenous and endogenously expressed proteins. Immunoprecipitation of exogenously expressed WT or activity-dead mutant (C74A) forms of ATG4B confirmed their association with exogenously expressed RNF5 (Figure S2a). ATG4B also associated with the RING mutant (RM) form of RNF5, indicating that the ubiquitin ligase activity of RNF5 was not required for this interaction. However, RNF5 lacking its membrane-spanning C-terminal domain (ΔCT) was no longer able to interact with ATG4B (Figure 1A), suggesting that the interaction takes place within the membranal domain. The latter is consistent with earlier studies showing that RNF5 interactions with, and effects upon, its substrates require membrane anchoring [36].
Given the role of ATG4B in autophagy, we monitored the RNF5–ATG4B interaction following exposure of cells to Hank's balanced salt solution (HBSS), a commonly used method to induce starvation with concomitant autophagy. Surprisingly, the interaction between ectopically expressed RNF5 and ATG4B in 293T cells was transiently reduced within 2–4 h following HBSS treatment, then returned to initial levels as early as 6 h after treatment (Figure S2B). This pattern was confirmed in HeLa cells, where the interaction between endogenous RNF5 and ATG4B was present prior to HBSS treatment, decreased within 2–4 h of autophagy induction then resumed between 6–16 h post-treatment (Figure 1B). These observations suggest that an interaction between RNF5 and ATG4B exists under basal conditions, but is attenuated following exposure of cells to autophagy stimuli. This implies that RNF5 may have a negative regulatory role to limit autophagy under normal growth conditions.
RNF5 and ATG4B contain multiple cysteines that have been implicated in their respective activities; Cys26 and Cys30 in RNF5 RING domain, among a total of seven cysteines, and Cys74 and Cys78 within the catalytic domain of ATG4B, among a total of 12 cysteines [32], [34], [42]. Therefore, we assessed whether reducing conditions could affect the interaction between RNF5 and ATG4B. Treatment of cells with the thiol-reducing agents dithiothreitol (DTT), glutathione (GSH) or H2O2 attenuated the ATG4B–RNF5 interaction (Figure S2C). These observations suggest that reducing conditions affects RNF5 interaction with ATG4B, as one mechanism underlying RNF5 association with and control of ATG4B stability, consistent with the reported role of reducing agents and ROS on the control of ATG4B and autophagy [42].
To determine whether the RNF5–ATG4B interaction has implications for ATG4B processing of LC3, we manipulated RNF5 expression in 293T cells and then assessed the interaction between ATG4B and LC3. Inhibition of RNF5 expression with established shRNA [35], [38] increased the level of ATG4B–LC3 interaction, while conversely RNF5 overexpression reduced the interaction, compared to cells transfected with control vectors (Figure S2D). These data indicate that the RNF5–ATG4B interaction affects ATG4b association with LC3, which is expected to affect LC3 processing.
Among the different ATG4 members, ATG4B was found to exhibit the highest affinity for RNF5 (Figure S3A). Among other members of the ATG family, including ATG3, ATG5, and ATG7, only ATG4B exhibited an appreciable association with RNF5 (Figure S3B). Consistent with these observations, there was no significant change in the levels of ATG3, ATG5, or ATG7 in MEFs prepared from RNF5-deficient animals (Figure S3C). Mapping RNF5–ATG4B interaction domains identified both N- and C-terminal domains (amino acids 61–126 and 320–393) of ATG4B (Figure S3D) and C-terminal domain of RNF5 (aa121–180; Figure S3E) as those required for the interaction between these two proteins.
Having identified the interaction between RNF5 and ATG4B, we next examined whether this interaction affects ATG4B stability, given the E3 ubiquitin ligase activity of RNF5. To determine if RNF5 ubiquitinates ATG4B constructs for the two proteins were co-expressed with HA-tagged ubiquitin in HeLa cells. ATG4B ubiquitination was induced in the presence of the WT but not the RING mutant form of RNF5 (Figure 1C). To confirm that ATG4B was directly ubiquitinated by RNF5, we performed in vitro ATG4B ubiquitination in the presence of RNF5. Notably, poly-ubiquitinated ATG4B was observed in the presence of E1 and E2 enzymes, ubiquitin, and RNF5 (Figure 1D). These results suggest that RNF5 directly induces ATG4B ubiquitination.
We next assessed whether RNF5-mediated ubiquitination affects ATG4B stability. We initially assessed the steady-state levels of ATG4B in cells in which RNF5 expression was increased or inhibited. The steady-state level of ATG4B was higher in RNF5−/− MEFs compared to WT cells (Figure S2D). The half-life of ATG4B was measured in cycloheximide (CHX) chase experiments comparing RNF5 KO to WT MEFs. Consistent with results obtained under steady-state conditions, the half-life of ATG4B increased from 1.75 to 3.5 h in RNF5 KO cells that were maintained under normal growth conditions. Following exposure to HBSS, the half-life of ATG4B was prolonged further in the absence of RNF5, albeit to a lesser degree (from 3 h to just over 4 h) (Figure 1E), consistent with the pattern of RNF5–ATG4B association (Figure 1B, Figure S2B). Similarly, shRNA inhibition of RNF5 expression in WT MEFs efficiently prolonged ATG4B half-life, whereas overexpression of RNF5 reduced ATG4B stability (data not shown). In agreement, reduced ATG4B protein, seen upon ectopic expression of RNF5, was attenuated in the presence of the proteasome inhibitor MG132 (Figure S4A). Notably, addition of MG132 increased the level of ATG4B expression in both RNF5 WT and KO MEF cells, although the degree of increase was more pronounced in the WT cells (Figure S4B). The latter suggests that other ubiquitin ligases may also contribute to the regulation of ATG4B stability. Consistent with these observations, a similar degree of ATG4B ubiquitination was seen in MG132 treated WT and RNF5 KO cells, although without MG132 WT cells exhibited higher degree of ATG4B ubiquitination compared with the RNF5 KO cells (Figure S4C). These results establish that RNF5 regulates ATG4B stability through its ubiquitination and proteasome-dependent degradation.
Having shown that RNF5 regulates ATG4B stability, and having obtained initial evidence that this affects ATG4B–LC3 interactions, we next explored the implications of these observations on ATG4B activity and autophagy. An established fluorogenic assay that measures ATG4B cleavage of an in vitro-synthesized pro-LC3 substrate was employed [43]. Using immunopurified ATG4B proteins prepared from control or shRNF5-expressing cells, we found that the in vitro activity of ATG4B was 40–60% higher in cells with reduced RNF5 expression (Figure S5A). To investigate the role of RNF5 in LC3 processing, we detected intracellular proteolysis of LC3 based on non-conventional secretion of Gaussia luciferase. In this system, LC3 is anchored in the cell by fusion to β-actin and the cleaved product is secreted. Thus, ATG4B activity can be monitored by quantifying extracellular luciferase activity resulting from the proteolytic cleavage of the LC3 site between Gaussia luciferase and β-actin [44], [45]. Corroborating our in vitro findings, intracellular proteolysis of LC3 measured by the luciferase assay increased in shRNF5-expressing HeLa cells, and conversely, was attenuated following ectopic expression of RNF5 (Figure S5B). These results indicate that RNF5 expression regulates ATG4B activity and LC3 cleavage.
To pursue this finding further, we monitored the effect of RNF5 on the appearance of different intracellular LC3 forms by western blotting. We monitored changes in the processing of LC3 in WT and RNF5 KO MEFs prior to, or 1, 2, and 4 h after treatment with HBSS (Figure 2A). Loss of RNF5 expression enriched in expression of the matured LC3 form, LC3-II, with the greatest difference being observed at the earlier time points (Figure 2A; control and 1 h, compared with 2 or 4 h). The latter is consistent with the finding that RNF5–ATG4B association is seen prior and at early time points following autophagy stimuli (Figure 1B). Neither GABARAP nor NBR1 expression were altered under the same conditions (Figure S6A). Notably, following LC3-II or LC3 decoration at later time points (6 and 8 h) did not reveal changes (not shown), probably since RNF5 affects to lesser degree the course of autophagy, once triggered. Inhibition of RNF5 led to the appearance of the processed LC3-II form in cells growing under normal conditions, and has increased the relative amount of LC3-II forms in cells that were subjected to starvation or after induction of ER stress with tunicamycin (Figure 2B). These data demonstrate that disruption of RNF5 expression affects LC3 processing under normal growth conditions, while also contributing to Atg4B availability following induction of autophagy. The level of LC3-II in both WT and RNF5 KO cells was elevated following treatment with the lysosomal inhibitors E64A and pepstatin A (Figure S6B). Similarly, degradation of the autophagy marker p62 was elevated upon inhibition or depletion of RNF5 expression (Figure 2A, 2B). Collectively, these data point to a role for RNF5 as a negative regulator of autophagy.
The appearance of LC3-II has been associated with autophagosome formation, which can be visualized by the appearance of LC3-positive puncta [15], [17]. We therefore monitored the effect of RNF5 on the accumulation of endogenous LC3-positive puncta. Compared with WT cells, RNF5−/− MEFs maintained under normal growth conditions contained higher levels of LC3-positive puncta, consistent with the appearance of the LC3-II form. This difference was not apparent in cells subjected to starvation or ER stress stimuli (Figure 2C). These results substantiate the notion that RNF5 negatively regulates ATG4B availability with concomitant effect on LC3 processing and autophagy under normal growth conditions. Correspondingly, inhibition of RNF5 leads to increased basal levels of autophagy.
To determine whether this change in LC3 processing can be associated with the effect of RNF5 on ATG4B, we monitored changes in LC3 puncta in WT or ATG4B−/− MEFs in the presence of shRNF5. As shown in Figure 3A, inhibition of RNF5 expression increased the number of LC3 puncta (three times) in WT, but not in ATG4B−/− cells. Support for ATG4B-dependent effect of RNF5 on LC3 processing comes from the use of RNF5 that is truncated in its C-terminal transmembrane domain (ΔCT), a mutant that retains its E3 ligase activity, but no longer interacts with ATG4B (Figure 1A, Figure S3C). Ectopic expression of RNF5 ΔCT (Figure S6C, S6D) in RNF5−/− MEF did not reduce the number of LC3 puncta, as seen with the WT RNF5 (Figure S6C). These findings substantiate that the negative regulation of LC3 puncta by RNF5 occurs via its regulation of ATG4B.
We next assessed whether processing of LC3 by ATG4B can promote autophagosome formation and maturation in the absence of RNF5 expression. We monitored the formation of autophagosomes and autolysosomes using the mRFP-GFP-LC3 construct [46]. Notably, we also tested a corresponding mRFP-GFP-LC3 mutant in which the C-terminal TFG residue has been exposed such that its sequence mimics pre-cleaved LC3-I, and thus is no longer subject to cleavage by ATG4B [15]. Because GFP is more sensitive than RFP to the acidic environment of lysosomes, the tandem RFP-GFP-LC3 protein will label autophagosomes as yellow (GFP/RFP) puncta, and autolysosomes as red (RFP) puncta. Thus, quantification of yellow and red puncta allows the determination of the effect on autophagosomes versus autolysosomes, respectively. In ATG4B WT-expressing cells, the LC3 TFG construct contained 2-fold more LC3 puncta (both yellow and red) than the WT LC3 construct, indicating that pro-LC3 cleavage by ATG4B promotes the formation of autophagosomes and autolysosomes (Figure S7A). As expected, cells containing the WT LC3 construct formed lower number of puncta in ATG4B−/− cells than in ATG4B WT cells (Figure S7A). However, the pre-cleaved LC3 TFG mutants still formed, albeit 2-fold fewer, puncta in the ATG4B−/− cells compared to WT cells (Figure S7A), suggesting ATG4B contributes to additional steps in the autophagy process.
Expression of WT LC3 led to higher levels of both yellow and red puncta in RNF5−/− versus WT cells under normal (Figure 3B), but not starvation (Figure S7B) conditions, suggesting that absence of RNF5 promotes the formation of both autophagosomes and autolysosomes under basal conditions. In clear contrast, WT and RNF5−/− cells expressing the pre-cleaved LC3 mutant formed similar numbers of red LC3 puncta (autolysosome), consistent with ATG4B-dependent conversion of pro-LC3 to LC3-I. Taken together, these results suggest that disruption of RNF5 results in elevated autophagosome formation through upregulation of ATG4B, which in turn increases LC3 processing to promote formation and maturation of autophagosomes.
We next asked if RNF5 also regulates autophagy in C. elegans. To accomplish this, we used rnf-5(tm794) deletion mutant that encodes a 36 amino-acid protein lacking the RNF5 RING domain, and thus predicted to be inactive [34]. rnf-5(tm794) L3 larvae expressing the autophagy marker GFP::LGG-1 had 2.6-fold more puncta in their seam cells compared to larvae on the WT background (Figure 4A, p<0.0001). This increase in GFP::LGG-1 puncta was also detected in worms treated with rnf-5(RNAi) (Figure 4B, 2.4-fold, p<0.0001). Conversely, animals expressing RNF-5 under the heat shock promoter [34] that were grown at 25°C (Figure 4C, conditions that increase the number of puncta in WT worms) resulted in decreased number of GFP::LGG-1 puncta in the seam cells of animals with elevated RNF-5 levels compared to the control animals at 25°C (2.2-fold; p<0.0001). These data suggest that RNF5 is a negative regulator of autophagy in C. elegans. Notably, degree of ER stress, measured via the hsp-4::gfp transcriptional reporter [47], confirmed that depletion of rnf-5does not exhibit increased hsp-4 transcription (Figure S8).
Because RNF5 is a membrane-bound E3 ubiquitin ligase, and since it requires its membrane-anchor to affect ATG4B and autophagy we assessed whether RNF5 activity towards ATG4B may preferentially take place at membranal domains. The autophagy markers LC3 and the ER membrane marker Sec61 were used to determine degree of RNF5-colocalization with LC3 at the membranal fraction. Analysis was performed in both MEFs and HeLa cells grown under normal conditions. In HeLa cells we found that 74% of the LC3/RNF5 colocalized spots were within the sec61b structures, whereas in the MEFs this percentage was found to be 83% (Figure 5A). This data suggest that RNF5 effect on LC3 processing does take place within the ER domain.
Biochemical analysis of subcellular fractions, enabling enrichment of membrane domains confirmed that under normal growth conditions RNF5 colocalizes with the ER marker calnexin (Figure 5B), but not with ATG9 (not shown), a marker for the trans-Golgi network and late endosomes, which is redistributed in phagophores/autophagosomes under starvation conditions [48]. Both LC3-I and LC3-II co-fractionated with RNF5 and calnexin under normal growth conditions (Figure 5B, upper panel fractions 7–12). Following HBSS treatment, level of ATG4B increases within the membranal fractions, and LC3B forms are found at earlier fractions, compared with normal growth conditions (Figure 5B, lower panel, fractions 5–7). These results demonstrate that RNF5 colocalizes with components of autophagic machinery within the membranal fractions. To further assess whether RNF5 regulates ATG4B at membranal domains we monitored membrane fractions for the presence of ATG4B during autophagy. Notably, ATG4B and ATG5 were found in the membrane-enriched fractions and their abundance in these fractions increased 2–6 h after starvation (Figure 5C). This pattern of ATG4B localization is consistent with the finding that RNF5 dissociates from ATG4B 2–4 h after HBSS treatment (Figure 1B). ATG4B levels within membrane-enriched fractions were increased in cells that expressed RNF5-targeted shRNA or that were subjected to HBSS treatment (Figure 5D). Collectively, these observations suggest that RNF5 regulates basal autophagy via control of ATG4B at the membranal ER compartment.
Given the demonstration that RNF5 controls basal autophagy activity, and the fact that autophagy levels are associated with altered susceptibility to intracellular bacterial replication [8], we next asked if RNF5 WT and KO mice exhibit differences in their response to bacterial challenge. In these studies, we employed the leading human pathogen GAS, previously identified to be susceptible to autophagy-mediated intracellular clearance [9]. Bone marrow-derived macrophages isolated from RNF5−/− mice demonstrated increased levels of LC3, ATG4B, and decreased level of p62 compared to RNF5+/+ macrophages (Figure 6A), suggesting they had a higher basal level of autophagy. Furthermore, levels of ATG4B and LC3 were elevated in mitochondrial and ER-enriched fractions of RNF−/− macrophages compared to RNF5+/+ macrophages (Figure 6B). These observations are consistent with membrane anchoring of RNF5 and suggest that RNF5 limits ATG4B-dependent LC3 processing in both mitochondrial and ER membranes. Autophagy-deficient (ATG5−/−) cells have been shown to be more susceptible to GAS intracellular proliferation [9]. In agreement with those findings, we found higher numbers of LC3 puncta in in RNF5−/− macrophages infected with GAS than RNF5+/+ macrophages (Figure 6C, 6D), pointing to a more efficient bacterial processing by autophagy in the absence of RNF5. Ultrastructural analysis of GAS-infected macrophages using EM confirmed that higher numbers of bacteria were engulfed in RNF5−/− macrophages than in WT cells (Figure 6E).
GAS intracellular survival was significantly reduced in RNF5−/− macrophages compared to WT controls (Figure 7A). However, the enhanced intracellular bacterial clearance by RNF5−/− macrophages was abolished when cells were transfected with ATG4B shRNA (Figure 7B). These results specifically link the enhanced bacterial clearance in the RNF5−/− macrophages to increased basal autophagy in these cells. Using a murine model of systemic GAS infection, at a challenge dose that produced 75% mortality in WT mice, almost all RNF5−/− mice survived (Figure 7C). The bacterial loads in mouse organs during the early stages of infection were higher in RNF5+/+ mice compared to RNF5−/− mice (Figure 7D). Taken together, our data show that the negative regulation of autophagy by RNF5 influences the outcome of GAS infection, and demonstrate for the first time a protective role for autophagy in an in vivo model of this important human infectious disease.
Our findings demonstrate that the ubiquitin ligase RNF5 functions to limit basal levels of autophagy by regulating ATG4B stability. Our conclusion is supported by two genetic models—C. elegans and mice—in which inactivation of RNF5 resulted in higher levels of autophagosomes decorated by LC3 puncta. These observations were substantiated by studies with cultured cells in which manipulation of RNF5 altered ATG4B expression and concomitantly affected LC3-II formation and association with autophagosomes. Notably, control of ATG4B by RNF5 primarily occurs prior to and following autophagy, as reflected in the decreased ATG4B–RNF5 interaction between 2 and 6 h after stimulation of autophagy. Correspondingly, effects of RNF5 on ATG4B and autophagy were seen, albeit to lesser degree, following autophagy stimulation. These data indicate that RNF5 limits basal levels of autophagy in the absence of stimulatory factors. Significantly, the regulation of ATG4B by RNF5 primarily takes place at the membranal domains, thereby restricting the effect to the relatively small pool of ATG4B that reaches the membrane domain. This finding suggests that formation of phagophores from cellular membranes may be primarily regulated by a membranal pool of ATG4B, through the membrane-anchored E3 ligase RNF5. Thus, the current study describes a previously unknown aspect of ATG4B regulation, through its availability for autophagosome formation, and thus for autophagy. Current dogma suggests that cleavage of pro-LC3 by ATG4B is regulated co-translationally, and thus, LC3-I is not the limiting factor in the early stages of autophagy. Our findings uncover an important regulatory role for a fraction of ATG4B that must be located at membrane domains and is important for modification of LC3 and subsequent phagophore formation. The importance of RNF5 in this process is supported by the use of mRFP-GFP-LC3 construct containing a truncated ATG4B recognition sequence, circumventing the need to be modified by ATG4B. This construct was no longer subject to regulation by RNF5, demonstrating a pivotal role for RNF5 in ATG4B function. Consistent with this finding, LC3 localization within autophagosomes and autolysosomes was inversely correlated with the level of RNF5 expression. LC3 colocalized with RNF5 within the ER membrane, unless autophagy or proteasomes were inhibited, enabling colocalization.
The implications of this newly discovered layer in the control of autophagy are reflected in the decreased susceptibility of RNF5-deficient mice to bacterial infection with GAS. The higher basal levels of autophagy in these mice allow more efficient clearance of invading bacteria, reducing both the load and physiological consequences of infection, as reflected in the improved survival of RNF5−/− mice following infectious challenge.
Our findings raise the question as to which physiological cues might regulate the RNF5–ATG4B association/dissociation. It is likely they will involve post-translational modifications of one or both binding partners. Initial data supports the role of ROS in the association between RNF5 and ATG4B (both proteins contain cysteine-rich domains that are required for their function), which is consistent with a reported role for ROS in autophagy [42]. As a membrane-anchored protein, RNF5 must retain its membrane localization to associate with and affect the function of ATG4B, consistent with our earlier findings for other RNF5 substrates [32], [35], [36]. While the extreme carboxyl terminal domain of RNF5 contains its membrane-spanning region and is required for its anchor to the membrane [32], [35], the central to the carboxyl terminal domains was found to be required for its association with ATG4B. Control of ATG4B by RNF5 is likely limited to the small pool of ATG4B localized at these membranes, consistent with phagophore formation from cellular membranes [49], [50]. However, based on the experiments using the LC3-GFP-RFP our data also point to a role for ATG4B in autophagy, other than its role in LC3 processing. This is in agreement with the emerging concept that certain stimuli can elicit autophagy independently of ATG4B activity [51].
Our findings also add to the existing links between the ubiquitin–proteasome system and autophagy. First, the ubiquitin ligase Parkin is linked to mitophagy as well as autophagy, primarily through its function in mitochondrial membrane organization [52]. However, Parkin substrates that directly contribute to autophagic activity have yet to be identified. Second, both p62 and Nbr1 contain ubiquitin-binding domains (UBAs), which are implicated in recruitment of ubiquitin chain-conjugated misfolded proteins to autophagic vesicles [53]. Third, thapsigargin, a model compound often used to induce ER stress, was recently shown to control autophagosome–lysosome fusion, a critical step in the late phases of autophagy, thus revealing an additional link between ER stress and autophagy [54]. Given the number of ubiquitin ligases involved in the control of the ER stress response in general, and ERAD in particular, one would expect that some of these ligases would also affect distinct phases of the autophagy process.
It is clear there is a growing link between autophagy and the clearance of invading bacteria. Initially, genetic inactivation of essential Atg genes in C. elegans and Dictyostelium was shown to alter the fate of invading bacteria [55]. In vitro studies have also shown that inactivation of autophagy enhances Salmonella replication in macrophages [56]. Consistent with these studies, we found that macrophages from RNF5−/− mice contain a greater number of autophagosomes surrounding bacterial pathogens than do cells from WT mice. Of note, it has previously been shown that mice defective in intestinal expression of the Atg16L1 or Atg5 genes have defects in gut secretion of antimicrobial peptides [57]. Thus, in addition to restricting the multiplication of intracellular bacteria, future studies could explore whether RNF5 affects the production and secretion of antimicrobial peptides to influence intestinal microbial colonization
In conclusion, our data establishes the presence of a new layer in the control of autophagy through limiting basal levels of autophagy. This control is mediated by the ubiquitin ligase RNF5 through its regulation of the membranal ATG4B protease. We further demonstrate the implications of this regulation for host defense mechanisms that limit intracellular infection by bacterial pathogens, suggesting the possible development of RNF5 inhibitors as new means for treatment of select pathogens and possibly means to overcome their commonly experienced resistance.
All animal work has been conducted according to relevant national and international guidelines in accordance with recommendations of the Weatherall report and approved by the IACUC committee at SBMRI and UCSD.
HeLa cells and PC3 cells were cultured in RPMI 1640 medium with 10% heat-inactivated fetal bovine serum (FBS) and antibiotics. 293T cells were cultured in Dulbecco's modified Eagle's medium supplemented with 10% FBS and antibiotics. RNF5 WT and KO MEFs were prepared from 13.5-day mice embryos following the standard protocol, and were cultured in Dulbecco's modified Eagle's medium supplemented with 10% FBS and antibiotics. RNF5 KO and WT mice were described previously [35] and bred and maintained by the animal facility in our institute (AUF08-008). ATG4B WT and KO MEFs were a gift from Dr. Carlos López-Otín (Universidad de Oviedo, Oviedo, Spain).
The plasmids expressing Myc-RNF5, Flag-RNF5, Flag-RING domain mutant (RM), Flag-C-terminal transmembrane domain-deleted mutant (dCT), and shRNF5 have been described previously [36], [38]. The reporters of intracellular proteolysis based on non-conventional secretion of Gaussia luciferase Actin-DN and Actin-LC3-DN were a gift from Dr. Brian Seed (Massachusetts General Hospital, Boston, MA). Flag-ATG4B and ATG4BC74A were previously reported [43]. Tandem mRFP-GFP-LC3–expressing plasmid ptfLC3 was purchased from Addgene. mRFP-GFP-LC3 TFG was generated by inserting a stop-codon after the C-terminal TFG residue using Quikchange mutagenesis. Lentivirus-based shRNAs of ATG4B and ATG5 were purchased from Sigma validated MISSION shRNA, and prepared following the standard procedure.
Rabbit polyclonal anti-ATG4B, -ATG5, -ATG7, -LC3B, and anti-Flag M2 antibodies were purchased from Sigma. Rabbit anti-LC3 antibody was purchased from MBL. Anti-Ub, anti-p62, anti-GFP, anti-HA, anti-myc antibodies, and protein A/G affinity gels were purchased from Santa Cruz Biotechnology. Tunicamycin were purchased from Calbiochem. The anti-RNF5 antibody was previously described [35], [38].
Briefly, cells were lysed with lysis buffer (50 mM Tris-HCl, pH 7.4, 150 mM NaCl, 1% NP-40, 10% glycerol, 40 mM β-glycerophosphate, 30 mM sodium fluoride, 5 mM EDTA, 1× protease inhibitor cocktail (Roche), and 1 mM phenylmethylsulfonyl fluoride). For immunoprecipitation, the cell lysates were incubated with antibody for 1 h and then incubated with protein A/G affinity gels for 4 h or overnight at 4°C. After washing three times with lysis buffer, the protein-bead complexes were resolved by SDS-PAGE (20 µg protein/lane), transferred to nitrocellulose membranes, and subjected to western blotting. The membranes were blocked with 5% dried milk in phosphate-buffered saline (PBS) plus 0.2% Tween 20 and then incubated with primary antibodies for 2 h at room temperature or overnight at 4°C. Anti-rabbit, anti-goat, or anti-mouse IgG antibodies conjugated to horseradish peroxidase (Pierce) or fluorescent dyes (Invitrogen) were used as the secondary antibodies. Blots were visualized by direct imaging on a Licor Odyssey or with an enhanced chemiluminescence system (Pierce).
In vitro and in vivo ubiquitination were described previously [36]. Briefly, for in vitro ubiquitination, His6-tagged ATG4B protein was expressed in E. coli strain BL21 and purified using nickel-agarose beads. The bead-bound ATG4B was resuspended in buffer (50 mM Tris-HCl, pH 8.0, 5 mM MgCl2, 2 mM ATP, 0.5 mM DTT, 2 mM NaF) prior to addition of ubiquitin and ubiquitin system enzymes. The final concentrations were as follows: 200 ng/µL ubiquitin, 2 ng/µL E1, 20 ng/µL E2, and 200 ng/µL substrate in a total volume of 25 µL. The reactions were incubated for 60 min at 37°C, then bead-bound material was washed three times with washing buffer (50 mM Tris-HCl, pH 8.0, 150 mM NaCl, 0.2% Triton X-100, 0.1% SDS) prior to analysis by western blotting. For in vivo ubiquitination, cells were transfected with the indicated plasmids together with a plasmid expressing HA-tagged ubiquitin. Harvested cells were lysed in 2% SDS in TBS (10 mM Tris-HCl, pH 8.0) at 95°C for 10 min. The lysates were diluted 20-fold with TBS containing 0.2% Triton X-100 and 2 mM EDTA to give a final SDS concentration of 0.1%, then incubated on a shaker at 4°C for 30 min. The lysates were incubated for 30 min at 4°C with protein A/G beads and clarified by centrifugation for 10 min (14,000 rpm) at 4°C. For immunoprecipitation, the lysate was incubated with anti-Flag antibody at 4°C for 1 h before protein A/G beads were added and incubated for a further 2 h. Beads were washed three times with TBS containing 0.2% Triton X-100 and 0.1% SDS. Proteins were resolved by SDS-PAGE, transferred, and immunoblotted with the indicated antibodies.
Cells were fixed with 4% formaldehyde in PBS for 15 min, permeabilized with 0.2% Triton X-100 in PBS for 10 min, and blocked with 2% bovine serum albumin in PBS for 30 min. Cells were then incubated with primary antibodies overnight at 4°C. After washing three times with PBS containing 0.1% Triton X-100, cells were incubated with Alexa 594-, 488- or 350-labeled anti-rabbit or anti-mouse IgG antibodies (Invitrogen) for 1 h. Cells were counterstained with DAPI and washed twice, then mounted in antifade agent on glass slides and visualized with an inverted fluorescence microscope (Olympus).
Briefly, HEK293 or HeLa cells were seeded in 24-well plates and co-transfected with 100 ng of the appropriate Renilla luciferase reporter actin-DN or actin-LC3DN plasmids, 600 ng myc-RNF5 or shRNF5 expression plasmids, or with control plasmids. Transfections were performed using Lipofectamine 2000 (Invitrogen) according to the standard protocol. After 48 h, the medium was refreshed and then collected at different time points. The Renilla luciferase assays were carried out according to the manufacturer's instructions for the dual luciferase reporter assay system (Promega).
GAS strain NZ131 (serotype M49) was grown in Todd-Hewitt broth (THB, Difco, Detroit, MI) at 37°C to mid-logarithmic phase (OD600 = 0.4). For in vivo experiments, GAS were resuspended in PBS+5% mucin to an inoculum of 2×107 CFU and injected intraperitoneally into WT or RNF5−/− mice. Three days post-infection, half of the mice were sacrificed and organs were collected and homogenized. Surviving bacteria in the organs were enumerated on Todd-Hewitt agar (THA) plates. The mortality of the remaining mice was monitored for an additional 7 days (10 days total). Mouse infection experiments were performed twice. For in vitro experiments, bone marrow cells were collected from age-matched WT and RNF5−/− mice and cultured in RPMI supplemented with 20% FBS and 30% L-929 cell conditioned medium for 7 days. Adherent macrophages were collected and seeded at 4×105 cells per well in RPMI supplemented with 10% FBS and 10 ng/mL mouse macrophage colony-stimulating factor (Pepro-Tech, Rocky Hill, NJ) in 24-well plates for 1 day prior to bacterial infection or transfection with shRNA. Bone marrow-derived macrophages were then transfected with scrambled ATG4B or ATG5 shRNAs and bacterial infection was assessed 4 days later, as previously described [9], [58]. Intracellular survival was calculated as the percentage of bacteria remaining after antibiotics were removed from the culture medium.
To analyze the number of LC3 puncta in macrophages incubated with GAS, BM-derived macrophages were seeded on glass coverslips (5×104/well in a 24-well plate) and infected on the following day with GAS expressing GFP (MOI = 10) for 3.5 h before washing and fixation in 3% paraformaldehyde (PFA). Cells were processed for immunofluorescence microscopy using affinity-purified rabbit anti-LC3 IgG, followed by goat anti-rabbit Alexa-488 F(ab′)2 and DAPI to stain the bacterial and cellular DNA. Images were taken on an Olympus TH4–100 fluorescence microscope using a 60× oil immersion lens (1.42 NA). Five Z-planes per field were captured and LC3 was quantified using Slidebook v.4.1 software.
BM-derived macrophages were infected with GAS (MOI = 10) for 2.5 h. Cells were subsequently fixed with 4% PFA, 1.5% glutaraldehyde, and 5% sucrose in 0.1 M cacodylate buffer for 2 h at room temperature, postfixed in 1% OsO4 in 0.1 M cacodylate buffer for 1 h at room temperature, and then embedded as pelleted cells in LX-112 (Ladd Research, Williston, VT) as described previously [59]. Sections were stained in uranyl acetate and lead citrate and observed with an electron microscope (JEOL 1200 EX-II).
|
10.1371/journal.pcbi.1005957 | Modeling the interactions of sense and antisense Period transcripts in the mammalian circadian clock network | In recent years, it has become increasingly apparent that antisense transcription plays an important role in the regulation of gene expression. The circadian clock is no exception: an antisense transcript of the mammalian core-clock gene PERIOD2 (PER2), which we shall refer to as Per2AS RNA, oscillates with a circadian period and a nearly 12 h phase shift from the peak expression of Per2 mRNA. In this paper, we ask whether Per2AS plays a regulatory role in the mammalian circadian clock by studying in silico the potential effects of interactions between Per2 and Per2AS RNAs on circadian rhythms. Based on the antiphasic expression pattern, we consider two hypotheses about how Per2 and Per2AS mutually interfere with each other's expression. In our pre-transcriptional model, the transcription of Per2AS RNA from the non-coding strand represses the transcription of Per2 mRNA from the coding strand and vice versa. In our post-transcriptional model, Per2 and Per2AS transcripts form a double-stranded RNA duplex, which is rapidly degraded. To study these two possible mechanisms, we have added terms describing our alternative hypotheses to a published mathematical model of the molecular regulatory network of the mammalian circadian clock. Our pre-transcriptional model predicts that transcriptional interference between Per2 and Per2AS can generate alternative modes of circadian oscillations, which we characterize in terms of the amplitude and phase of oscillation of core clock genes. In our post-transcriptional model, Per2/Per2AS duplex formation dampens the circadian rhythm. In a model that combines pre- and post-transcriptional controls, the period, amplitude and phase of circadian proteins exhibit non-monotonic dependencies on the rate of expression of Per2AS. All three models provide potential explanations of the observed antiphasic, circadian oscillations of Per2 and Per2AS RNAs. They make discordant predictions that can be tested experimentally in order to distinguish among these alternative hypotheses.
| A better understanding of the molecular mechanisms underlying circadian rhythms will undoubtedly improve the treatment of human health problems related to circadian dysrhythmias. However, the inventory of genes and genetic interactions in the circadian clock is still incomplete. Important players may yet be unknown or under-appreciated. For example, in mouse liver, the core clock gene PER2 is transcribed into both a Per2 mRNA molecule (a ‘sense’ transcript) and an antisense RNA transcript (Per2AS). Because it is important to know how interactions between Per2 and Per2AS may affect circadian gene expression, we have carried out a mathematical modeling study of two possible mechanisms for these interactions. In the pre-transcriptional model, Per2 mRNA interferes with the transcription of Per2AS RNA and vice versa. In the post-transcriptional model, Per2 and Per2AS molecules form double-stranded RNA duplexes, which are rapidly degraded by RNases. We find that the pre-transcriptional model gives a more robust account of the circadian, antiphasic oscillations of Per2 and Per2AS transcripts in mouse liver. The model makes an unexpected prediction that co-overexpression of the ROR gene and Per2AS sequences can generate a new mode of circadian oscillations not seen in contemporary models of circadian rhythms and not yet looked for experimentally.
| Messenger RNAs, which encode proteins, are transcribed in the 5'-to-3' direction from one strand (the sense strand) of a structural gene, under the control of an upstream promoter region. For some genes, an ‘antisense’ RNA molecule is transcribed from the opposite strand, driven by an alternative promoter which often lies in an intron of the sense transcript [1, 2]. Antisense transcripts are rarely translated into proteins; their primary effects are in regulating the expression of a ‘target’ transcript [3–6]. Because of their complementary sequences, the natural target of an antisense transcript is typically its sense counterpart and vice versa. Interactions between these transcripts are possible not only post-transcriptionally [7–9] but also during the transcription process [10–12]. Difficulties in simultaneously transcribing RNAs from both strands of the same genomic locus, termed transcriptional interference, can mutually repress the expression of both sense and antisense transcripts [13].
Recently Koike et al. [14] reported that an antisense transcript of PER2, a key core-clock gene, displays oscillatory dynamics. The maximum level of the antisense transcript, Per2AS, was about 5% of Per2’s maximum level, and the two transcripts were expressed in antiphase, i.e., the peak of Per2AS expression was displaced about 12 h from the peak of Per2 mRNA. From previous studies of the regulation of gene expression by antisense transcripts in other organisms, it is known that antisense expression can effectively control expression of sense mRNAs; for example, by a tunable, bistable switch [13, 15, 16]. To date the potential regulatory roles of antisense transcripts in a system with oscillatory dynamics have not been studied systematically. Therefore, a natural question is to what extent the rhythms in the mammalian circadian clock can be affected by Per2AS expression.
In this work, we study, by numerical simulation and bifurcation analysis [17], the effects of sense-antisense interactions in a mathematical model of the mammalian circadian network proposed by Relogio et al. [18]. Relogio’s model is based on two, synergistic feedback loops: the classic, negative feedback loop involving CLOCK/BMAL1 and PER/CRYPTOCHROME (CRY), and the alternative, mixed feedback loop involving BMAL1, REV-ERB (REV) and ROR. We supplement Relogio’s model with an additional, double-negative feedback loop between Per2 and Per2AS RNA species. (Simulations of the original Relogio model agree with many previously reported experimental observations [2, 19–21], and we are careful to retain these successful features of the published model.) By incorporating new terms and variables into Relogio’s model, we study, in silico, the effects of two different hypotheses concerning Per2-Per2AS interactions. In our first model, called the pre-transcriptional model, we assume that Per2 and Per2AS mutually repress each other’s production during the process of transcription. This hypothesis is motivated by recent observations of circadian rhythmicity in Neurospora, where it was shown that sense and antisense transcripts of the FREQUENCY (FRQ) gene control the circadian rhythm by transcriptional interference [22]. Our second model, the post-transcriptional model, is based on the assumption that fully transcribed Per2 and Per2AS form double-stranded duplex RNAs, which are degraded by RNases, similar to siRNA- or miRNA-mediated RNA degradation mechanisms [23]. After considering these two models separately, we study a third model that combines pre- and post-transcriptional interactions.
In our simulations of these three modified Relogio-models, the dynamics of Per2 and Per2AS are consistent with the fundamental observation of Koike et al. that the RNAs oscillate with ~24 h period and in antiphase to each other. Our pre-transcriptional model shows that the interference of Per2AS on the transcription of Per2 and vice versa can generate new modes of oscillations (both circadian and non-circadian) in the network, because of the way the double-negative feedback loop between Per2 and Per2AS interacts with the synergistic feedback loops in the original Relogio model. In contrast, the post-transcriptional model shows that circadian rhythms can be destroyed by Per2AS overexpression, because duplex formation rapidly suppresses the expression of Per2 mRNA. A characteristic feature of the pre- and post-transcriptional models is that the period of the oscillation is sensitive to the interactions of Per2 and Per2AS. The combined pre/post-transcriptional model shows that if Per2AS is involved in two different levels of Per2 regulation, then the period of the oscillation, as a function of Per2AS overexpression, can be restricted to a narrow interval.
Fig 1A presents a schematic diagram of the circadian clock network in mammalian cells, as originally proposed by Relogio et al. [18]. CLOCK/BMAL1 up-regulates the expression of the core clock genes, PER, CRY, REV, and ROR. Newly synthesized PER and CRY proteins form multimeric complexes in the cytoplasm, and these complexes enter the nucleus, in both phosphorylated and unphosphorylated forms of PER. The PER/CRY complex inhibits CLOCK/BMAL1-activated transcription, by creating a delayed negative-feedback loop in the transcription-translation process. The PER/CRY complex is degraded during the night, releasing its inhibitory effect on CLOCK/BMAL1, to allow a fresh restart of the transcription processes [18].
ROR and REV proteins in the nucleus bind to the promoter region of the BMAL1 gene, thereby modulating the expression of Bmal1 mRNA. ROR is an activator and REV an inhibitor of BMAL1 expression [24]. Previously, ROR and REV genes were often considered as auxiliary elements in the network, whose primary roles were to fine-tune the expression of BMAL1 and add robustness to the rhythmic dynamics [25, 26]. However, in the model of Relogio et al., the effects of REV and ROR on BMAL1 expression form independent loops that can generate sustained oscillations autonomously, even if the PER and CRY genes are expressed constitutively. Some experimental evidence suggests that the feedback loops through REV and ROR are critical for maintaining circadian oscillations; for instance, when REV or ROR is overexpressed or both REV-ERBα and REV-ERBβ are knocked-out, circadian rhythmicity can be lost [18, 19, 27].
From the schematic diagram in Fig 1A, Relogio et al. derived a system of ordinary differential equations (ODEs) that represent the temporal dynamics of these circadian genes and proteins. Other groups have presented alternative mathematical models of mammalian circadian rhythms [28–31], but the Relogio model is most fitting for our purposes in this paper. In contrast to other models that focus on the negative feedback loop, in which PER/CRY inhibits CLOCK/BMAL1, the Relogio model considers the mammalian circadian clock as a network of synergistic and interlocked feedback loops whereby, in addition to PER/CRY inhibition of CLOCK/BMAL1, REV and ROR control the expression of BMAL1, as inhibitor and activator, respectively (see Fig 1A).
The Relogio model [18] consists of 19 ODEs with 76 parameters (rate constants for the constituent biochemical reactions in the network). With an appropriate choice of these parameter values, the model generates simulations in agreement with many well-established experimental properties of circadian rhythms in mammalian cells. For this reason, we have chosen the Relogio model for studying the effects of Per2 sense-antisense interactions. Our strategy is to incorporate into the model new variables and reaction rates that represent potential interactions of sense-antisense RNAs (Per2 and Per2AS), while keeping the modified model as close as possible to the original Relogio ODEs, and keeping the parameter values as close as possible to the ‘wild-type’ (WT) values in reference [18].
Previously, it was shown by Xue et al. in Neurospora crassa [22] that coupled transcription of the key circadian gene FRQ and its antisense partner QRF directly modulates the circadian rhythm, as a consequence of mutually inhibitory interactions between frq and qrf RNAs. Following this lead, we hypothesize that the interactions of Per2 and Per2AS may also modulate circadian rhythmicity in mammalian cells, by forming a double-negative feedback loop. In Fig 1A, we indicate the mutually inhibitory interactions between Per and PerAS RNAs by the red lines in a small blue box.
At present, there are no experimental data about the exact molecular mechanisms by which Per2 and Per2AS interact in the circadian network. Therefore, our strategy is to propose reasonable hypotheses for the interaction and to study the consequences of these interactions in silico. We propose two simple, feasible mechanisms for sense-antisense interactions, which function either before or after the transcriptional process is complete (Fig 1B). Our aim is not to prove that one or other of these hypotheses is correct, but rather to study the potential effects of sense-antisense interactions on circadian rhythms of the core-clock network, in terms of modulating the period, amplitude, and phases of oscillations.
Numerical simulations were carried out in Mathematica, and bifurcation diagrams were calculated using AUTO [17]. In some circumstances, parameter values in the models were fitted to experimental data using the ensemble method [32] described in Suppl. S4 Text.
The differential equations of the pre-transcriptional model (i.e., Relogio’s differential equations supplemented with Eqs (1A) and (1B)) are provided in Suppl. S2 Text. The parameter values proposed by Relogio et al. [18] are listed in Suppl. S1 Table, where they are called ‘WT’ values. Suppl. S1 Table also lists proposed values for the parameters μ, λ, KS and KAS that characterize the mutual interference between Per2 and Per2AS.
In the post-transcriptional model (the Relogio model modified by Eq (2)), we assume that the physical interaction (duplex formation) between sense-antisense transcripts causes mutual degradation of both RNAs. In this case, the amount of Per2AS in a cell is especially important, and this amount is determined by the parameter λ0, which represents constitutive transcription of Per2AS from both the endogenous PER2AS sequence and from exogenous Per2AS sequences carried on a plasmid. In Fig 7 we show how the period, amplitudes and phases of the rhythm depend on the value of λ0. In these simulations, unless otherwise specified, all parameters of the Relogio model are fixed at their WT values, and the additional parameters in Eq (2) are fixed at kassn = 0.1, kdiss = 0.1, and ddup = 0.1.
We assume that the contribution to λ0 from the endogenous gene is small (say, 0.1 < λ0 < 1) compared to the contribution due to plasmid copies of Per2AS sequences (say, λ0 > 1). At λ0 = 0.2 (representative of endogenous synthesis only), the period of the oscillations in the post-transcriptional model is ~23.5 h, the maximum level of Per2AS is about 5% of the maximum level of Per2, and Per2 and Per2AS oscillate out-of-phase, i.e. |ϕPer2 − ϕPer2AS| ≈ 12 h (see Suppl. S7 Fig). In other words, at these parameter values, the post-transcriptional model exhibits oscillations that fit reasonably well the time-courses of Per2 and Per2AS oscillations observed by Koike et al., shown by the black circles in Fig 2.
Fig 7 shows how the properties of circadian rhythms change in the post-transcriptional model with increasing rates of synthesis of exogenous Per2AS (parameter λ0). The period of oscillations increases modestly with increasing λ0 (Fig 7A). The amplitude of Per2 oscillations drops with increasing λ0, because of the duplex formation, whereas the amplitudes of oscillation of other core-clock genes increase (presumably CLOCK/BMAL1 is less strongly repressed by PER/CRY) (Fig 7B). With Per2AS phase set at 0 hours, we plot in Fig 7C the changes in the phases of oscillation of core clock genes. Blue and black lines in Fig 7C show that the phases of Per2 and Cry mRNAs are most sensitive to the increase of Per2AS level.
In Suppl. S8A Fig we plot a two-parameter bifurcation diagram on the parameter plane (λ0, kassn). Although the oscillatory domain is very large in this diagram, the region where the post-transcriptional model oscillates with circadian properties is restricted; the black symbols mark the region where following conditions are fulfilled:
23h<T<25h,(AmaxPer2−AminPer2)>0.5,11h<|ϕPer2−ϕPer2AS|<13h.
(4)
In Suppl. S8B Fig we plot a two-parameter bifurcation diagram on the parameter plane (ddup, kassn), while fixing λ0 = 10 and kdiss = 0.1. The non-oscillatory domain in the middle of the diagram separates a region of circadian oscillations (22 h < T < 25 h) at the bottom of the diagram from a region of slow oscillations (T > 50 h) at the top. The region of this diagram where conditions in Eq (4) are fulfilled (marked by small red symbols) is quite restricted: 0.08 < ddup < 0.27 and 0 < kassn < 0.2. The blue symbols in Suppl. S8B Fig mark the region where the first and second conditions of Eq (4) are fulfilled, but the oscillations of Per2 and Per2AS are not strictly antiphasic, i.e., 9 h < |ϕPer2 − ϕPer2AS| < 15 h.
In Suppl. S9 Fig, we plot the time-courses of oscillations at three locations in Suppl. S8B Fig. Suppl. S9A–S9C Fig show the case: kassn = 1, ddup = 0.1 for two values of λ0. When λ0 = 1, the dynamics of Per2 is reminiscent of WT dynamics in the Relogio model, but when λ0 = 10, the amplitude of Per2 oscillations has become very small. For the case kassn = 1, ddup = 0.2 (Suppl. S9D–S9F Fig), the amplitudes of oscillations at λ0 = 10 are larger, but the waveform has become distinctly non-harmonic. For the case kassn = 5, ddup = 0.2 (Suppl. S9G–S9I Fig), the amplitudes of oscillations at λ0 = 10 are quite large, the waveforms are very non-haromonic, and the period (~30 h) is non-circadian.
Our explorations of the post-transcriptional model show that it can be parameterized to fit the observations in Koike et al. [14], but the range of suitable parameter values is restricted. If any of the parameters ddup, kassn, or kdiss in Eq (2) deviate too much from the preferred values, the oscillations may no longer fulfill the requirements in Eq (4). Especially if kdiss > kassn, the peak amplitudes of Per2 and Per2AS become quickly non-antiphasic. Therefore, we conclude that the formation of Per2/Per2AS duplex RNA tends to destroy circadian rhythms over a wide range of values of the parameters ddup, kassn, and kdiss.
Eq (3) details how we modified the Relogio model to include both pre- and post-transcriptional interactions of Per2 and Per2AS. Fig 8 shows how the period, amplitude, and phases of circadian oscillations change with increasing λ1 for fixed λ0 = 0, μ = 1 and kassn = 0.1. The combined model is consistent with circadian, antiphasic oscillations of Per2 and Per2AS (see Suppl. S10 Fig). Unlike simulations of the pre- or post-transcriptional model, shown in Figs 4 and 7, the period, amplitude and phases of oscillation in the combined model are distinctly non-monotonic in dependence on λ1.
On Fig 9 we continue the limit cycle oscillations of period T = 23.5 h on the parameter plane (μ, λ) for three different values of the rate constant for duplex formation, kassn. Notice that, compared to the case kassn = 0 (i.e., no duplex formation), the locus of 23.5-hour rhythms does not change much for kassn = 0.05, but it is radically different for kassn = 0.1, intersecting the line μ = 1 twice, at λ ≈ 1 and λ ≈ 25. Therefore, as Figs 8 and 9 show, the combined pre/post-transcriptional model can restrict the period of oscillations within tighter bounds of μ. The reason is that, unlike the pre- or post-transcriptional model for which Per2-Per2AS interactions directly modulate only a single process of gene regulation, in the combined model two different gene-regulatory processes are simultaneously modulated. As a result, due presumably to counter-balancing effects, the period of oscillations can be restricted to a narrow interval.
A better understanding of the molecular mechanisms underlying mammalian circadian rhythms will undoubtedly inform our efforts to improve human health and deal with modern societal problems such as shiftwork and jetlag. However, the inventory of genes and genetic interactions in the mammalian circadian-clock network is still incomplete. Important players may be yet unknown or under-appreciated [34, 35]. For example, recent experimental data about oscillations of an antisense RNA transcript in the circadian rhythm in mouse liver [14, 36, 37] suggest a possible antagonistic relationship between a core-clock mRNA, Per2, and its natural antisense partner, Per2AS. Because antisense transcripts can be fundamental regulators of gene expression, the interactions between Per2 and Per2AS may be important factors for controlling circadian rhythms [1]. To date, the molecular mechanisms of Per2-Per2AS interactions are unknown. In this work, we propose two realistic mechanisms for these interactions and study their effects in silico by incorporating Per2-Per2AS interactions into a well-documented mathematical model [18] of mammalian circadian rhythms. In the first hypothesis, Per2 mRNA molecules interfere with the transcription of Per2AS molecules and vice versa. In the second hypothesis, mature Per2 and Per2AS molecules form double-stranded RNA duplexes, which are rapidly degraded by RNases.
Simulations and analysis of our pre-transcriptional model (the first hypothesis) show that mutual transcriptional interference can generate emergent oscillations in the clock network. That is to say, Per2-Per2AS interactions can generate new modes of circadian oscillations not seen in the original model [18]. For example (Fig 5; purple curve), our model predicts that Per2AS overexpression restores circadian rhythms to ROR-overexpressing cells by rebalancing the positive and negative interactions exerted on BMAL1 expression by ROR and REV (Fig 1).
According to our post-transcriptional model (the second hypothesis), circadian oscillations are expected to be eradicated by an increasing rate of Per2AS expression, which is to be expected if Per2AS forms unstable duplex molecules with Per2 mRNA. For both the pre- and post-transcriptional models and for a combined pre/post-model, we have computed how the period of oscillation and the amplitudes and phases of core clock gene oscillations will vary with the rate of synthesis of Per2AS transcripts (see Figs 4, 7 and 8). By altering the rate of expression of Per2AS transcripts, these predicted dependencies of period, amplitudes, and phases can be tested experimentally. Comparison between such experimental results and mathematical predictions can evaluate the accuracy and predictive power of the three alternative models of sense-antisense interactions. In this way, experimental interrogation, in combination with mathematical simulations, can shed light on the mechanisms of sense-antisense interactions in the mammalian circadian rhythm, and a more realistic mathematical model can be developed.
Of the three models we have studied (pre-, post-, and combined pre/post-transcriptional models), the pre-transcriptional model is the most likely, in our opinion, because it provides the most robust account of the observed, circadian, antiphasic oscillations of Per2 and Per2AS RNAs [14], in the context of all the other experimental data that went into the development and parameterization of the circadian-rhythm model of Relogio et al. [18]. Furthermore, the pre-transcriptional model makes the counterintuitive prediction that Per2AS overexpression can restore circadian rhythms to cells that are overexpressing ROR. This striking prediction of the model can be tested in a suitably designed mutant strain of mouse liver cells that overexpress both Per2AS RNA and Ror mRNA.
Our study of sense-antisense interactions has been made in the context of a specific mathematical model of mammalian circadian rhythms [18], but we suspect that our results are generic, in the sense that similar results will be found if our hypotheses are tested in different models of the circadian clock [29–31, 38]. As an example, we studied the effects of Per2 and Per2AS interactions in the Mirsky et al. model [30] of mammalian circadian rhythms. The three main differences between the Relogio and Mirsky models are that (a) Mirsky’s model includes paralogs of Per and Cry (i.e., Per1 and Per2, Cry1 and Cry2), (b) the two models make different assumptions about how PER/CRY interferes with CLOCK/BMAL-induced gene expression, and (c) Rev and Ror play less prominent roles in the generation of rhythmic dynamics in Mirsky’s model relative to Relogio’s model. Suppl. S11 Fig shows how period, amplitudes, and phases of oscillations change in the Mirsky et al. model [30] with increasing rate of Per2AS transcription. Notice the similarity between Fig 4 and Suppl. S11 Fig, despite the fact that Mirsky’s model distinguishes between Per1 and Per2 transcripts and proteins. Suppl. S12A Fig shows that Per1 oscillations are indirectly affected by the double negative feedback interactions of Per2 and Per2AS, but the amplitude changes of Per1 and Per2 are uncorrelated. Suppl. S12B Fig shows that Cry1 and Cry2 oscillations also respond to Per2AS interference, and that their amplitudes are anti-correlated with each other. The generic effects of Per2-Per2AS interactions in different models are due, presumably, to generic, network-level consequences of a double-negative feedback loop embedded in the delayed negative-feedback that generates circadian rhythms.
Obviously, depending on the choice of a base model, of the mathematical representations of our hypotheses, and of parameter values, a rich repertoire of interesting dynamics are possible in a mathematical model involving many feedback loops that can generate independent oscillations [39–41]. For example, in a recent paper El-Athman et al. [42] have combined the Relogio-2011 model of the mammalian circadian clock with a model of mammalian cell-cycle controls and shown that knocking out the tumor suppressors that bridge the two systems induces notable phase shifts in the expression of circadian clock genes. Interesting research directions in the future would be a) whether these phase shifts can be controlled by antisense transcripts of Per2, and b) whether the positive regulation of the tumor protein p53 by Per2, as reported by Gotoh et al. [43], can induce predictable amplitude and phase modulations in the oscillations of cell cycle elements. Finally, we hope that the modeling results reported here, suggesting that Per2-Per2AS interactions may have profound effects on circadian rhythmicity, may stimulate new experiments about the roles of this sense-antisense pair of RNAs in the mammalian circadian-clock network.
|
10.1371/journal.pgen.1003609 | Re-Ranking Sequencing Variants in the Post-GWAS Era for Accurate Causal Variant Identification | Next generation sequencing has dramatically increased our ability to localize disease-causing variants by providing base-pair level information at costs increasingly feasible for the large sample sizes required to detect complex-trait associations. Yet, identification of causal variants within an established region of association remains a challenge. Counter-intuitively, certain factors that increase power to detect an associated region can decrease power to localize the causal variant. First, combining GWAS with imputation or low coverage sequencing to achieve the large sample sizes required for high power can have the unintended effect of producing differential genotyping error among SNPs. This tends to bias the relative evidence for association toward better genotyped SNPs. Second, re-use of GWAS data for fine-mapping exploits previous findings to ensure genome-wide significance in GWAS-associated regions. However, using GWAS findings to inform fine-mapping analysis can bias evidence away from the causal SNP toward the tag SNP and SNPs in high LD with the tag. Together these factors can reduce power to localize the causal SNP by more than half. Other strategies commonly employed to increase power to detect association, namely increasing sample size and using higher density genotyping arrays, can, in certain common scenarios, actually exacerbate these effects and further decrease power to localize causal variants. We develop a re-ranking procedure that accounts for these adverse effects and substantially improves the accuracy of causal SNP identification, often doubling the probability that the causal SNP is top-ranked. Application to the NCI BPC3 aggressive prostate cancer GWAS with imputation meta-analysis identified a new top SNP at 2 of 3 associated loci and several additional possible causal SNPs at these loci that may have otherwise been overlooked. This method is simple to implement using R scripts provided on the author's website.
| As next-generation sequencing (NGS) costs continue to fall and genome-wide association study (GWAS) platform coverage improves, the human genetics community is positioned to identify potentially causal variants. However, current NGS or imputation-based studies of either the whole genome or regions previously identified by GWAS have not yet been very successful in identifying causal variants. A major hurdle is the development of methods to distinguish disease-causing variants from their highly-correlated proxies within an associated region. We show that various common factors, such as differential sequencing or imputation accuracy rates and linkage disequilibrium patterns, with or without GWAS-informed region selection, can substantially decrease the probability of identifying the correct causal SNP, often by more than half. We then describe a novel and easy-to-implement re-ranking procedure that can double the probability that the causal SNP is top-ranked in many settings. Application to the NCI Breast and Prostate Cancer (BPC3) Cohort Consortium aggressive prostate cancer data identified new top SNPs within two associated loci previously established via GWAS, as well as several additional possible causal SNPs that had been previously overlooked.
| The challenges of precise identification of disease-causing variants underlying GWAS signals have recently received much attention [1]–[3]. For post-GWAS statistical analysis that aims to accurately identify potentially causal variants, a major hurdle is the development of methods to distinguish disease-causing variants from their highly-correlated proxies. While GWAS-era statistical methods focused on identifying associated regions via tag SNPs at the coarse scale of GWAS arrays, next generation sequencing (NGS) technology offers the capability to not only detect associated regions, but to distinguish the causal SNPs within these associated regions. Here we make a distinction between ranking SNPs across the genome to identify an associated region, and ranking to pinpoint the potential causal variant within an associated region. Identifying an associated region requires that trait-associated SNPs be ranked above null SNPs, while identifying the causal variant requires that, among associated SNPs, associations due to causality are ranked above indirect associations due to other factors, e.g. linkage disequilibrium (LD). GWAS and imputation studies typically report the top-ranked SNP for each associated locus, and follow-up studies typically attempt replication for these top-ranked SNPs (for further discussion of ranking see Text S1).
Zaitlen et al [4] proposed a measure of performance for sequencing and fine mapping analysis, their localization success rate metric is the probability that the causal SNP has the top-ranked test statistic within an associated region. When multiple SNPs are in high LD, the localization success rate drops dramatically [5]. Udler et al (2010) investigated the difficulty in overcoming the stochastic effect of high LD among causal and non-causal SNPs [5]. The sample size required to distinguish the causal SNP can be 1 to 4 times the size required to detect the association at genome-wide significance. Zaitlen et al [4] showed that this problem could be overcome through joint analysis of samples from carefully selected populations with differing LD structure. Although candidate causal SNPs will require further bioinformatic or functional study to ultimately delineate potential causal mechanisms, optimized study design and analysis can point to the best possible candidate causal SNP(s) and help develop testable hypotheses about biological mechanisms.
Studies of complex traits now underway are leveraging the cost efficiency of integrating GWAS, low- and high-coverage sequencing, and imputation to achieve sample sizes in the tens of thousands [6], [7]. For example, the Genetics of Type 2 Diabetes (GoT2D) study is combining low and high-coverage sequencing with 2.5M-SNP GWAS genotyping and imputation to achieve a total sample size of over 28,000 [8]. Sequencing the GWAS sample exploits the GWAS findings to ensure that an association signal is present at the genome-wide level and eliminates the cost of recruiting new individuals. Analysis of sequenced and imputed SNPs (post-GWAS data) can thus be informed by previous GWAS results, allowing a prioritized use of post-GWAS data in fine-mapping regions surrounding significant GWAS tag SNPs [9]–[11]. Selection of associated regions for further studies can also be based on combined GWAS and post-GWAS criteria [12], [13]. For example, the WTCCC [13] required a marginally significant (p-value<10−4) GWAS SNP to support the evidence at a genome-wide significant imputed SNP. However, these strategies lead to two important issues that have received little attention in the context of causal SNP identification: (1) the effect of the re-use of successful GWAS data and (2) the effect of genotyping error rates that differ between sequenced or imputed SNPs.
The re-use of GWAS data that had contributed to the identification of an associated region for post-GWAS analysis can adversely affect accurate causal SNP identification. For example, the simulation study of Wiltshire et al [14] showed that when a significant GWAS tag SNP is followed up by sequencing in the same sample, the tag SNP is in fact ranked higher than the true causal SNP 30% to 63% of the time, depending on the genetic model and effect size. When a GWAS tag SNP is selected based on small p-value, the magnitude of the association at the tag tends to be over-estimated; this form of selection bias is also known as the winner's curse [15]–[19]. To a variable extent, depending on the LD pattern, this selection bias is carried over from the GWAS tag to post-GWAS sequenced or imputed SNPs [20]. While this earlier work empirically demonstrated the effect of selection for a significant GWAS tag SNP on the causal SNP, no work to date explores whether it also affects the rank of the causal SNP among all neighboring SNPs within an associated region, and if so how to correct for the bias.
High error rates and differences in error rates, due to differences in coverage, read length and depth, minor allele frequency (MAF), GC content, local sequence structure, and other sequence-specific factors, are common to NGS SNPs and are well-recognized obstacles to analysis [21]–[29]. Error rates for low-read-depth sequencing studies are estimated to be 1%–3% [22], [30], [31], and as little as 1% error can produce a large loss in power [27]. The strategy of low-coverage sequencing in a portion of GWAS samples has been used to discover sequencing variants and build a reference panel to drive imputation in the remaining samples, but the genotyping accuracy can be worse than if all individuals were sequenced [25]. The choice of lower-coverage design is also motivated by reports that low-coverage sequencing in a large sample, alone or combined with GWAS and imputation data, can achieve superior power to detect associations compared to high-coverage sequencing in a small sample with similar cost [25], [29], [32], [33]. However, whether the localization success rate of the causal variants responsible for these associations is similarly high has not yet been examined. High error rates that differ among SNPs also occur in high-coverage sequencing; for example, within targeted high-coverage regions, highly repetitive elements can be difficult to capture resulting in low accuracy for some SNPs [34].
Differential genotyping accuracy between studies has been shown to reduce power of meta-analysis in the imputation setting [35], and differential accuracy between cases and controls has been shown to cause confounding and elevated type I error [36], [37]. Accounting for differential genotyping accuracy in the association test can recover some of the lost power and reduce type I error [35], [36]. However, whether it affects our ability to distinguish causal SNPs from correlated SNPs, and how best to account for the effect of differential genotyping accuracy jointly for all SNPs (GWAS tagged, imputed or sequenced) is an open question.
In this report, we first demonstrate that:
We develop an analytic description of how these factors influence the probability of localization success and evaluate this probability for a range of plausible parameter values. We then show how to properly adjust for the adverse effects of these factors with a re-ranking procedure. We evaluate the performance of the method with extensive simulation studies under a wide range of realistic scenarios, and we demonstrate the practical use of re-ranking with an application to the NCBI BPC3 aggressive prostate cancer GWAS with imputation [38].
Suppose that M sequenced (or imputed) SNPs, Si, i = 1, …, M, in the region surrounding a significant GWAS tag SNP G are ranked by the magnitude of their association statistics in order to identify the causal SNP C. Table 1 provides the notation for the various parameters and statistics used throughout the report. Briefly, TSi is the Wald test statistic at a sequenced SNP Si; is the sample Pearson correlation coefficient between the GWAS/imputed/sequenced genotypes (most likely or fractional allele dosage) for SNPs G and Si (r2 is the well-known pair-wise correlation measure of LD between two SNPs); is the estimated correlation between the true genotype and the called genotype for a sequenced SNP Si (we use correlation as a measure of genotyping accuracy because of its simple interpretation in terms of power and genotyping quality; this quantity is provided by both MACH [24] and BEAGLE [39] software); and are proportions of samples with non-missing genotypes (termed call rates) at SNPs G and Si, respectively, and is the joint call rate, the proportion of samples with non-missing genotypes at both SNPs, and is the call rate at the causal SNP.
Let be an estimate of the selection bias in genetic effect estimation at the tag SNP G (described further below), that is the excess in the expected value of the test statistic at the tag SNP G induced by selection based on its small p-value (or high rank). We call this phenomenon the selection effect (ΔG is zero if the region was not selected via a tag SNP that achieved the given significance or ranking criterion in the same sample). Our proposed re-ranking statistic for a sequenced SNP Si is(1)Equation (1) depends on the selection effect , the tagging effect , the genotyping accuracy effect and scaling factors that depend on the call rates . Justification for Equation (1) now follows in the remainder of this section. (Full details are provided in Text S2.)
Without loss of generality, let >0 be the genetic effect (e.g. the log odds ratio or the regression coefficient in the model relating the phenotype and genotype) at the causal SNP C which could be: one of the sequenced or imputed SNPs Si, i = 1,…, M; the GWAS tag SNP G although this is unlikely; or neither if the genomic coverage was incomplete. Let the tag SNP G be coded such that the coded allele is positively correlated with the causal allele. Let be the genetic effect estimate and be the estimated standard deviation (SD) of the estimate from n observations. We assume that the distribution of the Wald test statistic at the causal SNP, is approximately normal, , where . The following also applies to test statistics that are asymptotically equivalent to the Wald test statistic.
Let be the difference between the observed test statistic and its expected value,(2)Here is the correlation between the genotypes of the causal C and the tag SNP G. (We assume that the tag is coded so that it is positively correlated with the risk allele of the causal SNP.) The value of is unobserved and needed only in the theoretical formation of the problem not in the practical implementation, which we discuss later. The selection effect is most pronounced when there is low power at the tag SNP. (For discussion of this point see Text S3).
The conditional distribution of the test statistic TSi at the sequenced SNP Si, conditional on the value of the observed test statistic at the tag SNP G, is(3)Derivation of this distribution is detailed in Text S2. The first term, , is the unconditional expected association signal at the sequencing SNP; the second term, , is the distortion due to the tag SNP selection propagated through correlation. Therefore, ΔG, the selection effect at the GWAS tag SNP G carries through to each sequenced SNP Si in proportion to the correlation between G and Si. The combination of attenuation due to LD and upward selection bias at the tag, ΔG, distorts the association evidence so that SNPs in high LD with the tag are more likely to be top-ranked. We call this phenomenon the tagging effect, and use an estimate to remove bias from the conditional expected value of in (3).
Third, differential call rates among SNPs (, and ) and estimated genotyping accuracy ( is the estimated and is the actual correlation between the called genotype and true genotype) of sequenced or imputed SNP Si appear in both the numerator and denominator of Equation (1). In the numerator, the tagging bias, , is scaled by a factor of because correlation between the test statistics depends on the individual and joint call rates at the two SNPs (see Text S2 for derivation). The bias-corrected statistic in the numerator is scaled by because(4)where is the correlation between the genotype of the causal SNP and the called or estimated genotype of the sequenced SNP (in contrast to , for the true genotype of the sequenced SNP). Assuming the probability of genotyping error is independent of the actual genotype, then . It is clear that, without correction, smaller ρSi (higher genotyping error) and smaller (higher missing data rate) tend to lower the probability that SNP Si would be top-ranked. We call this phenomenon the genotyping accuracy effect.
To conceptually demonstrate the joint effects of selection, tagging and genotyping accuracy on the localization success rate (the probability that the causal SNP is topped ranked within an associated region), we first consider the simplified case of 2 SNPs, one causal (from sequencing or imputation) and one tag (from GWAS) with correlation between the two SNPs ranging from r = 0.2 to 1 (from almost no LD to perfect LD). The inclusion of low LD value is motivated by the fact that correlation between the causal SNP and the best tag is often lower than expected. The coverage of GWAS platforms tends to be overestimated for both sequenced and imputed SNPs (see Text S4 for further discussion of this point). We assume that the MAFs of both SNPs are 0.12, the causal SNP has an additive odds ratio (OR) of 1.25, and selection at the tag SNP, if present, is based on its association test p-value<0.05 in a sample of 1000 cases and 1000 controls. Localization success rates (before applying the proposed re-ranking procedure) for all figures were computed based on Equations (2)–(3) and the equation in Text S3 and by numerically integrating over the following bivariate normal density function,(5)
Analytical evaluations of Equation (5) were used to generate Figures 1–3, which give insight into the relative influence of the tagging, selection, genotyping accuracy and sample size effects outlined in the Introduction and explicitly defined in Materials and Methods. We find similar patterns of influence for a rare SNP (MAF = 0.02, OR = 1.5; Figures S2, S4 and S6) and a higher frequency SNP (MAF = 0.25, OR = 1.25; Figures S3, S5 and S7), and when the number of non-causal SNPs increases (Figures S8, S9, S10).
The above analytical results demonstrate the need to correct for the joint effects of selection, tagging and genotyping accuracy on the localization success rate. The practical implementation of the proposed re-ranking statistic in Equation (1) is as follows. The estimated selection bias at the tag SNP G can be obtained using BR-squared that provides Bias-Reduced estimates via Bootstrap Resampling at the genome-wide level [40], [41]. (The original program, designed to provide estimates for the genetic effect β, has been modified slightly to provide estimates for the test statistic T; see software documentation on author's website for details.) The bootstrap estimator can be applied whether the region of interest was selected by rank or by p-value threshold. Unlike the threshold-based likelihood and Bayesian methods [42]–[46], the genome-wide bootstrap method incorporates information across the entire GWAS in order to account for the effects of LD and rank on the bias at each SNP. The values of the individual and joint call rates are available from the dataset, and genotype correlation can be estimated from the sample. Correlation between the actual and estimated genotypes at a sequenced SNP can be obtained from the mean posterior genotype (e.g. MACH ratio of variances estimate, [24]) or from the full genotype posterior probabilities (e.g. BEAGLE allelic r2 estimate [39]). An R script that implements Equation (1) is available. The R script calls the BR2 software (http://www.utstat.toronto.edu/sun/Software/BR2/), which provides the essential quantity of if the original GWAS dataset was used for fine-mapping.
We conducted extensive simulation studies to empirically evaluate the performance of the re-ranking method under five general scenarios (Table 2):
The parameter values in Table 2 were chosen to best reflect realistic scenarios. For example, in order to address realistic tagging, we examined the Affymetrix 5.0 chip and identified the SNP that best captured each significant WTCCC T1D GWAS SNP. The correlation between the two SNPs ranges from r = 0.79 to 1. For the range of genotyping accuracy, we note that in practice, the average sequencing ρ can vary substantially from study to study. For example, for low-coverage studies, it can vary from 0.63 to 0.99 depending on the coverage, MAF and sample size [25]. When low-coverage sequencing (4×) and imputation are combined, the average ρ can range from 0.89 to 0.99 depending on the reference panel size [24]. Sequencing ρ also depends on MAF; the same error rate in a lower MAF SNP results in a smaller ρ.
Even when the average ρ is high, SNP-level ρ can vary widely within a single study. Browning and Browning [39] found that imputation with a phased reference panel of 60 Hapmap CEU samples yielded a median ρ of 0.95, however individual ρ was less than 0.77 for 20% of the SNPs. We show that coverage rates can also vary widely between SNPs (Figure S1) by examining the 1000 Genomes low-coverage whole-genome pilot data from chromosome 1 in the CHB and JPT samples (Figure S1; October 2010 release; 1000 Genomes Project, 2010). We mimicked this variability in our simulations by randomly assigning each SNP in each dataset an error rate that ranged from zero to twice the overall average error rate. No random error however was introduced into the genotypes of the tag SNP (ρG = 1), because GWAS genotyping has been estimated to be over 99.8% accurate [13], [47]. In order to ensure realistic correlation structure among post-GWAS sequencing/imputation SNPs, we examined all SNPs in the regions surrounding the WTCCC T1D significant SNPs using the HapMap3 dataset. The average correlation between adjacent SNPs in these regions was approximately 0.975.
One of the main findings of the simulation study is that GWAS-based region selection or moderate genotyping error can substantially reduce the probability of correctly identifying the causal SNP (Tables 3–4 and Tables S1, S2), consistent with that of the analytical study. For example, results detailed in Table S1 demonstrate that the combined tagging and genotyping accuracy effect can reduce the localization success rate by over 30%.
The simulation study also shows that the proposed re-ranking procedure can recover much of this lost power to identify the causal SNP, increasing the localization success rates by 1.5- to 3-fold in many cases (Table 3). When genotyping accuracy is high, the power lost due to tagging is small and so re-ranking tends to have little effect.
For studies using GWAS-based selection (scenario 1), the adverse effects of tagging and genotyping accuracy on localization success rate are strongest when the causal SNP is well tagged (larger r) and less accurately sequenced/imputed (smaller ρ) (Tables 3, 4 and S1). High-density GWAS followed up with low-coverage sequencing would fall into this category. Well-tagged causal SNPs tend to suffer from lower localization success rates because the perfectly genotyped tag often captures the association better than the imperfectly sequenced or imputed causal SNP. Re-ranking corrects this problem, so that the localization success rate does not depend on how well the causal SNP is tagged, except when the tag SNP is in fact the causal SNP. In this case, the tagging and genotyping accuracy effects actually increase the localization success rate. After re-ranking, the localization success rate is similar to levels seen when the tag is not causal. We consider this a minor tradeoff, because the causal SNP is unlikely to be found among the GWAS SNPs for a number of reasons: GWAS SNPs are typically selected independent of the phenotype of interest and post-GWAS SNPs tend to greatly outnumber GWAS SNPs.
When the discovery sample is also used for fine-mapping, but significance is not required at the GWAS-tag SNP (scenario 2), the genotyping accuracy effect alone could still considerably reduce power to identify the causal variant (Table 3). When an independent sample is used for fine-mapping (scenario 3, Table 3), localization success rates are very similar to those seen in scenario 2. In both cases, the re-ranking method improves the probability of correctly identifying the causal SNP. The improvement is most pronounced (2- to 4-fold improvement) when genotyping accuracy is low. When there is more than one causal variant (scenario 4, Table 3), we find that re-ranking effectively increases localization success rates for both causal SNPs. Imperfect call rates affect localization success rate in a similar manner to imperfect genotyping accuracy (scenario 5, Table 4). Equation (4) implies that a call rate of 0.80 should affect the distribution of the causal SNP test statistic in the same manner as a sequencing accuracy ρ of 0.89, and this is borne out in our simulations. The re-ranking procedure corrects for both missing data and genotyping error to the same degree.
In some cases, investigators are more interested in delimiting a set of best candidate causal SNPs instead of a single top SNP. In the supplementary material, we include additional simulation results for this scenario. We define an alternative localization success rate metric as the probability that the causal SNP is in the top 10% of SNPs by rank (Table S2). Briefly, we examine the probability that the causal SNP is among the top 5 SNPs when there are 50 total SNPs (ranked by test statistic or re-ranking statistic). Without re-ranking, the probability that the causal SNP is in the top 10% of SNPs over the region is moderate. Re-ranking provides an improvement up to 1.8-fold.
Machiela et al [28] used the August 2010 release of the 1000 Genomes Project European-ancestry (EUR) panel to impute 11.6 million variants in 2,782 aggressive prostate cancer cases and 4,458 controls. These subjects were genotyped as part of the NCI Breast and Prostate Cancer (BPC3) Cohort Consortium aggressive prostate cancer GWAS [48], [49]; genotyping platforms varied across the seven BPC3 studies, although all used versions of the Illumina HumanHap arrays and most used the Illumina HumanHap 610 Quad array. The correlation between imputed genotype dosage and genotypes thus varied across studies. Imputation and association analyses using imputed genotype dosages were conducted separately for each study, and the association results were combined via fixed-effect meta-analysis. For each imputed SNP, studies with imputation r2<0.8 were excluded from the meta-analysis test statistic, leaving a total of 5.8 million GWAS and imputed SNPs.
Fine-mapping in the meta-analysis context ranks SNPs by the meta-analysis test statistic. Re-ranking requires that we compute the correlation between the meta-analysis test statistic on the Z-score scale (i.e. normally distributed test statistic) with and without accounting for genotyping error. Assume Zj is the normally distributed test statistic for study j, and wj is the weight for study j, the meta-analysis test statistic used for the standard naïve ranking isIf is an estimate of pair-wise correlation between the actual and imputed genotypes in study j (e.g. the square root of allelic-r2 [39], or ratio of variances r2 [24]), it follows that the estimated correlation between the meta-analysis test statistic computed with perfectly genotyped SNPs (Zact) and the meta-analysis test statistic computed with the observed imperfectly genotyped SNPs (Zobs) isThe re-ranking statistic in the meta-analysis case iswhere is the meta-analysis test statistic Z scaled for variance of 1.
Machiela et al [38] reported five statistically independent associated regions within the 8q24.21 locus and one for each of 11q13.3 and 17q24.3. We selected all SNPs in LD (r2>0.2) with the index SNP from each region for analyses (Figures 4 and 5, and Figures S11, S12, S13). In the application, we first ranked SNPs using the naïve test statistics [38]; and excluded any SNP with MAF <0.01; but unlike Machiela et al [38] we did not exclude any studies. Machiela et al selected significant regions by examining all imputed and genotyped SNPs at once and so we corrected for the imputation accuracy effect only (i.e. ).
Re-ranking identifies new top SNPs for 2 of the 3 associated loci: 8q24.21 and 17q24.3 (Figures 4 and 5 respectively). In addition to the most significant region at 8q24.21 (Figure 4), re-ranking also identifies a new top SNP for the third most significant region (Figure S11). For both regions re-ranking also identifies SNPs that may have otherwise been missed due to imperfect imputation. After re-ranking, 2 SNPs in the most significant region at the 8q24.21 locus (Figure 4) and 8 SNPs at the 17q24.3 locus (Figure 5) move from the lower ranks into the top 10 percent. On the other hand, SNPs in the top 10% are moved down by only a few ranks. In this way, re-ranking keeps highly significant SNPs identified by the naïve ranking and adds a few SNPs that would have otherwise been missed. When the top test statistics are of similar size, re-ranking may identify a new top SNP. When most SNPs are well-genotyped, re-ranking makes only subtle changes (Figure S11, S12 and S13).
There is one poorly imputed SNP at 17q24.3 (rs1014000, r2 = 0.20) that moves from the naïve rank of 245 to the new rank of 16 after adjustment. This SNP's apparent association is largely driven by data from a single study: the naïve rank in the EPIC study is 10. When we remove this study from the meta-analysis, the naïve rank is 306 and the adjusted rank is 119. No other SNP in the top 10% is this drastically affected when the EPIC study is removed from the analysis. In the meta-analysis context, we recommend examining top SNPs for heterogeneity among studies when re-ranking produces dramatically different results.
Overall, we observed that the tagging and genotyping accuracy effects are non-trivial sources of bias that could obscure association evidence at the causal SNP. The proposed re-ranking procedure is simple to implement and can substantially increase the probability of identifying the causal SNP. For low-coverage sequencing, we recommend the re-ranking method to improve causal SNP identification. For imputation and high-coverage sequencing, we recommend that unfiltered SNPs in associated regions be examined to see if correlation varies across SNPs and if so, we recommend adjustment with the re-ranking method. Large changes in rank should be carefully examined for underlying issues such as heterogeneity among meta-analysis studies or differential accuracy between cases and controls, and procedures to correct for these issues should be incorporated.
Re-ranking is most beneficial when genotyping accuracy is moderate to low, that is, the average correlation between the actual and estimated genotypes of post-GWAS (sequenced or imputed) SNPs is less than 0.97. A large number of post-GWAS SNPs in a study may appear to be significant, but when not all were directly genotyped with high accuracy, re-ranking can help select the most probable causal SNPs for follow-up. High density genotyping followed by low-coverage sequencing in the same sample can produce misleading results, as demonstrated by our simulations, so we do not recommend this design for identifying causal variants. Our re-ranking method tends to down-rank the tag SNP. If the tag SNP is suspected to be causal (e.g. based on prior study), we recommend examining the rank of the tag SNP using both the naïve and re-ranked methods when selecting SNPs for further study. Several imputation and sequencing software packages provide accurate estimates of ρ or quantities from which ρ can be computed [24], [39]. Re-ranking depends on accurate estimates of ρ. Recalibration of sequencing quality scores can greatly improve accuracy and so we recommend this step prior to re-ranking [27].
Re-ranking is especially important when study-specific factors exacerbate the effects of GWAS-based selection and genotyping error. Such factors include: high genetic diversity which makes sequencing reads difficult to align [27]; low LD among SNPs or lack of population-specific reference panels which makes some populations particularly difficult to impute (e.g. some African populations [50]); and imputation error which can be as high as 10% for these populations. Low MAF SNPs tend to suffer from both low power (which exacerbates the tagging effect) and high genotyping error. Re-ranking can be applied to rare and low MAF SNPs with allele counts large enough for test statistics to reach asymptotic normality. Very low (1×−2×) and extremely low (0.1×−0.5×) read depth sequencing has received recent attention as a way to maximize cost efficiency and make use of off-target sequencing data [29], [32]. Error rates for such regions would be both very high and highly variable among SNPs and so re-ranking to account for errors in the estimated genotypes would be crucial. When genotyping accuracy is extremely poor, the re-ranking method may not be able to sufficiently improve the localization success rate to ensure useful results. We recommend that investigators consider the accuracy thresholds recommended by the genotype calling or imputation algorithm they are using before re-ranking is applied.
We emphasize that re-ranking improves the localization success rate when applied to SNPs under the alternative, i.e. SNPs that are themselves causal or in LD with a causal SNP. Including null SNPs in the re-ranking procedure increases the number of SNPs the causal must out-compete, and so we recommend that only SNPs suspected to be under the alternative be included. In our application we included all SNPs that had squared pairwise correlation (r2) with the index SNP (most significant SNP in the region) greater than 0.2.
Existing methods that incorporate genotype uncertainty into tests for association to reduce power lost due to genotyping error or missing data [e.g. 51]–[54] do not completely recover lost power, and so the genotyping accuracy effect will remain. The simplest way to deal with genotype uncertainty in a test is to use the expected additive genotype (i.e. the posterior mean or dosage) in the standard linear or logistic regression. In this case, the re-ranking method can be applied using the allele dosages in place of called genotypes as described above. Guan and Stephens [55] compared several frequentist and Bayesian methods that incorporate genotype uncertainty into tests for association. The re-ranking procedure could be extended to any case where the correlation between test statistics or Bayes factors can be worked out.
We expect that re-ranking will play an important role as sequencing costs fall and GWAS platform coverage increases. Ultra-high density GWAS platforms are more likely to include tag SNPs in very high correlation with the causal SNP, which increases power to detect indirect association at the tag SNP. However, without re-ranking, strong tagging also decreases power to correctly identify the causal SNP in subsequent low-coverage sequencing. Advances in GWAS and sequencing platforms will allow researchers to drill down into lower MAFs and smaller effect sizes. Both low MAF and small effect size yield lower power, which exacerbates upward bias at the tag [20] and, therefore, the adverse tagging effect. Low MAF SNPs tend to suffer from higher error rates, which exacerbates the genotyping accuracy effect. Association study sample sizes will therefore need to continue to increase, so even as sequencing costs fall, it is anticipated that low-coverage will continue to be the most cost-effective design for many studies, despite the high genotyping error rates [27]. In conclusion, we anticipate that re-ranking to correct for the adverse effects of selection, tagging and differential genotyping accuracy rates among SNPs will continue to be important in candidate causal SNP identification for some time.
|
10.1371/journal.pntd.0003995 | CD4+CD25hiFOXP3+ Regulatory T Cells and Cytokine Responses in Human Schistosomiasis before and after Treatment with Praziquantel | Chronic schistosomiasis is associated with T cell hypo-responsiveness and immunoregulatory mechanisms, including induction of regulatory T cells (Tregs). However, little is known about Treg functional capacity during human Schistosoma haematobium infection.
CD4+CD25hiFOXP3+ cells were characterized by flow cytometry and their function assessed by analysing total and Treg-depleted PBMC responses to schistosomal adult worm antigen (AWA), soluable egg antigen (SEA) and Bacillus Calmette-Guérin (BCG) in S. haematobium-infected Gabonese children before and 6 weeks after anthelmintic treatment. Cytokines responses (IFN-γ, IL-5, IL-10, IL-13, IL-17 and TNF) were integrated using Principal Component Analysis (PCA). Proliferation was measured by CFSE.
S. haematobium infection was associated with increased Treg frequencies, which decreased post-treatment. Cytokine responses clustered into two principal components reflecting regulatory and Th2-polarized (PC1) and pro-inflammatory and Th1-polarized (PC2) cytokine responses; both components increased post-treatment. Treg depletion resulted in increased PC1 and PC2 at both time-points. Proliferation on the other hand, showed no significant difference from pre- to post-treatment. Treg depletion resulted mostly in increased proliferative responses at the pre-treatment time-point only.
Schistosoma-associated CD4+CD25hiFOXP3+Tregs exert a suppressive effect on both proliferation and cytokine production. Although Treg frequency decreases after praziquantel treatment, their suppressive capacity remains unaltered when considering cytokine production whereas their influence on proliferation weakens with treatment.
| Schistosomiasis, a parasitic worm infection, affects over 240 million people worldwide, especially children in sub-Saharan Africa. It is associated with immune hypo-responsiveness which results in an inability of the immune system to eliminate parasites. Animal models suggest that helminths induce regulatory T cells (Treg) which suppress effector cells and dampen anti-parasite activity as part of the parasites’ own strategy for survival in the human host. However, little is known about the functional capacity of Tregs during human Schistosoma haematobium infection and their interaction with adaptive responses. We designed a longitudinal study addressing the question of how anti-parasite treatment influences effector T cell activity and Treg function in peripheral blood of schoolchildren living in an S. haematobium endemic area in Lambaréné, Gabon. Our findings show that schistosome infection is associated with increased Treg frequency and that Tregs exert a suppressive effect on immune cell function in terms of both proliferation and cytokine production. Although Treg frequency decreases after anti-schistosome treatment, their suppressive capacity remains unaltered for cytokine production but their influence on proliferation weakens with treatment. By understanding how immune system is prevented from killing parasites, we hope to offer a novel route for intervention to achieve an immunological cure.
| The immune system has evolved several regulatory mechanisms to maintain immune homeostasis, prevent autoimmunity and restrain inflammation [1–3]. Many pathogens have developed mechanisms to manipulate the regulatory network of the host to their advantage, thereby generating conditions that ensure their survival for a prolonged period of time. In particular FOXP3+ regulatory T (Treg) cells have been shown to play a major role in the control of various parasitic infections suppressing local tissue damage and pathology that would result from otherwise over-reactivity. However, enhanced Treg cell activity may also allow the long-term survival of the parasite as the host is hampered from fighting the intruding pathogen effectively [4].
Schistosomiasis is a helminth infection affecting over 240 million people worldwide, especially children [5]. When chronic in nature it has been shown to be associated with general T cell hypo-responsiveness—evident from down-modulated antigen-specific Th1 and Th2 cell responses [6,7]]. This might result from mechanisms involving peripheral anergy and suppression triggered by regulatory cells, such as Tregs [8]. For example, in experimental murine models, it was observed that the presence of regulatory T cells was associated with suppressed development of pathology [9], and down-modulated Th1 and Th2 responses [10,11], promoting parasite survival within the host [12,13]. Evidence for Treg activity in human helminth infections has been provided by the detection of T cells with a regulatory phenotype in patients with lymphatic filariasis [14], onchocerciasis [15,16] and schistosomiasis [17,18].
Effective chemotherapy with praziquantel has been shown to result in elevated antigen-specific proliferation and cytokine production, in particular interleukin (IL)-4, IL-5, and interferon (IFN)-γ [6,19–22]. Although, the frequency of Tregs, defined phenotypically as CD4+CD25hi without considering FOXP3 as a marker, decreased substantially after treatment with praziquantel [18], their functional activity has not been studied before.
The aim of this study was to assess whether S. haematobium infection in Gabonese children was associated with induction of regulatory T cells and to evaluate Treg activity during infection. To this end immune responses were evaluated before and 6 weeks after praziquantel treatment. Moreover, the functional activity of Tregs was assessed by comparing responses before and after their removal from peripheral blood mononuclear cells by in vitro magnetic depletion.
Ethical approval for the study was obtained from the Comité d’Ethique Régional Indépendant de Lambaréné. A signed informed consent form was obtained from parents or legal guardians of all children participating in the study.
The longitudinal study was conducted at Centre de Recherches Médicales de Lambaréné (CERMEL; formerly Medical Research Unit (MRU) of the Albert Schweitzer hospital). Participants were schoolchildren from a rural area (PK15) highly endemic for S. haematobium approximately 15 km south of Lambaréné in the province of Moyen-Ogooué, Gabon [22,23]. Children were excluded from the study if they received anthelminthic treatment within the previous six weeks or had fever or other symptoms of acute illness. Fifty-two schoolchildren were recruited to participate in the study and of those twenty-eight (mean age: 10.3y (SD: 2.2); sex ratio: 15f/13m; median egg count/10ml urine: 72.5 (IQR: 24.5–296.3) pre-treatment and 0 (IQR: 0–1.5) at post-treatment; haemoglobin level: 11.1 g/dL (SD: 1.0) pre-treatment and 11.1 g/dL (SD: 1.0) post-treatment; and white blood cell level: 8.8 x103/mm3 (SD: 3.2) pre-treatment and 9.4 x103/mm3 (SD: 2.9)) were included in the cellular analyses. Of the 24 schoolchildren that were not included in the final analysis, 4 did not return for post-treatment visit, 5 had less than 90% clearance of Schistosoma eggs at 2nd treatment and in 8 donors at pre-treatment and 11 donors at post-treatment Treg depletion failed. There was no significant difference between the schoolchildren that were included in the final analysis and those that were not. To compare the frequency of CD25hiFOXP3+ Treg cells between S. haematobium infected and uninfected schoolchildren an additional 10 S. haematobium-infected participants (mean age: 12.5y (SD: 1.5); sex ratio: 9f/1m; median egg count/10ml urine: 19.5 (IQR: 3.3–216.5); haemoglobin level: 11.4 g/dL (SD: 0.31); and white blood cell level: 8.8 x103/mm3 (SD: 0.6)) were recruited from the same rural area and seven uninfected subjects (mean age: 12.9y (SD: 2.6); sex ratio: 4f/3m; median egg count/10ml urine: 0; haemoglobin level: 10.9 g/dL (SD: 0.7); and white blood cell level: 6.3 x103/mm3 (SD: 0.7)) were recruited from semi-urban Lambaréné. In addition to their positivity for S. haematobium infection, rural children also had significantly higher white blood cell levels (p<0.05) compared to children from semi-urban Lambaréné. When comparing the 2 cohorts, while the additional cohort of children was a little older (12.7y vs 10.3y; p<0.05), there was no significant difference in haemoglobin and white blood cell levels nor in infection intensity between the 2 groups.
A midstream urine sample was collected during the day and 10 ml were passed through a 12.0 μm polyamide N filter (Millipore) for the detection of S. haematobium eggs by microscopy. Children were classified as infected if at least one S. haematobium egg was detected in the urine.
An initial treatment with praziquantel (40 mg/kg) was administered to Schistosoma-infected children at inclusion, and repeated after three weeks, in order to ensure clearance of parasites. Six weeks after initial treatment the efficacy of praziquantel was assessed by measuring egg load in urine. Donors were excluded from analysis, if their reduction in egg load was less than 90% following second treatment. All subjects who were egg-positive after the second treatment were given a third dose of treatment.
Peripheral blood mononuclear cells (PBMCs) were purified from heparinized venous blood (7-10ml) by Ficoll-Hypaque centrifugation (Amersham Biosciences, Netherlands). Depletion of CD25hi T cells was performed using a suboptimal concentration of CD25 microbeads (Miltenyi Biotec, Bergisch Gladbach, Germany) according to the manufacturer’s instructions. This method has been shown by us in other studies to successfully deplete FOXP3 Tregs [24]. Similar results were obtained in Gabon as shown in S1A Fig.
To analyse proliferation green-fluorescent dye carboxyfluorescein succinimidyl ester (CFSE; Sigma-Aldrich, CA, USA) was used; CFSE divides over daughter cells upon cell division and can be tracked by decreasing fluorescence intensity. CD25hi-depleted and total PBMCs were stained with 2μM CFSE for 15 minutes at room temperature prior to culture. After labelling, cells were cultured in RPMI 1640 (Gibco, Invitrogen®, Carlsbad, CA, USA), supplemented with 10% fetal bovine serum (FBS; Greiner Bio-One GmbH, Frickenhausen, Germany), 100 U/ml penicillin (Astellas, Tokyo, Japan), 10 μg/ml streptomycin, 1 mM pyruvate and 2 mM L-glutamine (all from Sigma-Aldrich, CA, USA). Cells were stimulated in 96-well round bottom plates (Nunc, Roskilde, Denmark) with medium, 10 μg/ml AWA, 10 μg/ml SEA, or 10 μg/ml BCG (Bacille Calmette-Guérin; SSI, Copenhagen, Denmark) and incubated in 5% CO2 at 37.5°C. After four days, supernatants were collected and stored at -80°C, while cells were harvested, fixed with 2% formaldehyde (Sigma-Aldrich, CA, USA) and, subsequently, frozen in RPMI 1640 medium supplemented with 20% FCS and 10% dimethyl sulfoxide (DMSO; Merck KGaA, Darmstadt, Germany) and stored at -80°C.
After thawing, CFSE-labelled cells were incubated with CD4-PE (SK3; BD Bioscience, San Diego, CA, USA) and CD25-APC (M-A251; BD Bioscience, San Diego, CA, USA), acquired on a FACSCalibur flow cytometer (BD Biosciences, San Diego, CA, USA) and data were analysed in a FlowJo Proliferation application (Tree Star Inc., Ashland, OR, US) by calculation of the fraction of CD4+CD25hi cells that had divided from the starting population (division index). To assess Treg depletion, ex-vivo PBMC were fixed with the FOXP3 fixation/permeabilization kit (eBisocience, San Diego, CA, USA) and frozen in RPMI 1640 medium supplemented with 20% FBS and 10% DMSO and stored at -80°C. For immunophenotyping isolated PBMCs were stained with CD4-PE (SK3; BD Bioscience, San Diego, CA, USA), CD25-APC (M-A251; BD Bioscience, San Diego, CA, USA) and FOXP3-PE (PCH101; eBioscience, San Diego, CA, USA), acquired on a FACSCalibur flow cytometer (BD Biosciences, San Diego, CA, USA) and data were analysed in a FlowJo software (Tree Star Inc., Ashland, OR, US). To assess the frequency of CD25hiFOXP3+ Tregs, ex-vivo PBMC were fixed with the FOXP3 fixation/permeabilization kit (eBisocience, San Diego, CA, USA) and frozen in RPMI 1640 medium supplemented with 20% FBS and 10% DMSO and stored at -80°C. For immunophenotyping isolated PBMCs were stained with CD4-PE/Cy7 (SK3; BD Biosciences, San Diego, CA, USA), CD25-PE (2A3; BD Biosciences, San Diego, CA, USA) and FOXP3-APC (PCH101; eBioscience, San Diego, CA, USA), cells were acquired on FACSCanto II flow cytometer (BD Biosciences, San Diego, CA, USA) and analysed in FlowJo software (Tree Star Inc., Ashland, OR, US) using Boolean combination gates.
Cytokines were measured from supernatants using Luminex 100 IS System (Invitrogen, Carlsbad, CA, USA) and commercially available beads and standards from BioSource (Bleiswijk, Netherlands) for interferon-gamma (IFN-γ), interleukin-5 (IL-5), IL-10, IL-13 and IL-17 and tumor necrosis factor (TNF). Beads were titrated for optimal dilution and used according to manufacturer’s instructions.
Data analysis was performed using IBM SPSS Statistics version 20 for Windows (IBM Corp., Armonk, USA).
Differences between groups were determined by the Fisher’s exact test for sex, by Mann-Whitney U test for S. haematobium infection intensity and by the independent student’s T test for age and haematological parameters.
Cytokine concentrations in response to stimulation were corrected for spontaneous cytokine production by subtracting responses of unstimulated medium wells to obtain net cytokine responses, with negatives values set to half of the lowest value detected per given cytokine.
To avoid type I and type II errors in multiple testing, immunological parameters were reduced by principal-components analysis (PCA). First, R v2.15.1 Development Core Team software (R Foundation for Statistical Computing, Vienna, Austria, 2012, http://www.R-project.org) was used to estimate Box-Cox transformation parameter for each cytokine to increase normality of the data. Principal Component Analysis with Varimax rotation was used on all data points simultaneously (i.e. stimuli AWA/SEA/BCG; total and Treg-depleted PBMC; pre- and post-treatment time-points) to reduce the data into a smaller number of uncorrelated variables. Rotation converged in 3 iterations and principal components (PC) with eigenvalues greater than one were selected. Differences in PC scores between pre- and post-treatment and Treg-depleted and total PBMC were tested with the Wilcoxon matched pairs test. For all tests, statistical significance was considered at the 5% level.
To investigate whether S. haematobium infection affects the frequency of peripheral blood Tregs we compared circulating CD4+CD25hiFOXP3+ Tregs from infected and uninfected children by flow cytometry. Gating strategy for identification of CD4+CD25hiFOXP3+ Tregs is shown in Fig 1A. We found that frequencies of FOXP3+ Tregs were significantly higher in infected children compared to uninfected children (Fig 1B). Importantly, six weeks after praziquantel treatment Treg frequencies were reduced by half to frequencies comparable to the uninfected control group. Over the same six weeks period, there was also a slight but consistent decrease in the Treg frequencies in the uninfected group.
Next, we assessed the effect of anthelmintic treatment on cell proliferation and cytokine production in response to stimulation with schistosome-specific antigens SEA and AWA and a non-specific antigen BCG. Proliferation was determined by calculating the division index on the basis of the dilution of CFSE in CD4+CD25hi T cells. There were no significant differences in proliferation between pre-treatment and six weeks post-treatment responses (medium p = 0.397, AWA p = 0.188, SEA = 0.454 and BCG = 0.271) (Fig 2). Cytokine production on the other hand significantly changed between pre-treatment and 6 weeks post-treatment; raw cytokine values are shown in S1 Table. We applied Principal Component Analysis (PCA) in order to provide a more global assessment of the effect of schistosome infection on responses to not only SEA and AWA stimulation but also to the non-Ag specific stimulant BCG. Two distinct principal components were identified (Fig 3 and Table 1) which captured 73.7% of variance in our dataset: principal component 1 (PC1) which reflects regulatory and Th2-polarized cytokine responses due to its positive loading with IL-5, IL-10 and IL-13 responses (and accounted for 40% of the total variance in the data); and principal component 2 (PC2) which reflects pro-inflammatory and Th1-polarized cytokine responses due to its positive loading with IFN-γ, IL-17 and TNF (and accounted for 33.7% of total variance in the data. We saw a significant increase in both PC1 and PC2 following treatment compared to baseline values (Table 2).
To study the suppressive effect of Tregs on proliferation and cytokine responses, CD4+CD25hi T cells were depleted from PBMC by magnetic beads. The CD4+CD25hi population decreased by 45%, p = 0.0073 of which a representative example is shown in S1B Fig.
Depletion of Tregs at pre-treatment resulted in enhanced spontaneous proliferation (medium condition) as well as in enhanced proliferation to specific schistosomal antigens AWA and SEA and to vaccine antigen BCG (Fig 4A). At 6 weeks after anthelmintic treatment Treg depletion resulted in significant increase in proliferation in response to AWA only (Fig 4A). A typical plot of CFSE staining showing the effect induced by depletion of Tregs (Fig 4B).
Next, we investigated the capacity of Tregs to suppress cytokine responses by evaluating the effect of Treg depletion on principal component 1 (IL-5, IL-10 and IL-13) and principal component 2 (IFN-γ, IL-17 and TNF). We found that Treg depletion at pre-treatment resulted in increased values of both PC1 and PC2 in the infected individuals, and similarly at post-treatment, the depletion of Tregs resulted in an increase in the values of both PC1 and PC2 in the now infection free schoolchildren (Table 3).
Down-regulation of immune responses has been attributed to a strong immunomodulatory network of regulatory cells induced by schistosomes [25,26]. Here we provide evidence that human Schistosoma infection is associated with increased FOXP3+ regulatory T cells that play a significant role in controlling Th1 and Th2 responses. This finding is consistent with reports of increased numbers of FOXP3+ Tregs in peripheral blood from children 8–13 years old with active schistosomiasis [27] as well as other helminth infections including lymphatic filariasis [14,28]. Moreover, the reduction in the number of FOXP3+ Treg after treatment is in line with reports showing that drug induced clearance of Shistosoma parasites reduces Treg numbers defined only as CD4+CD25hi [18].
A much smaller in magnitude, yet statistically significant decrease was observed in the frequency of Tregs in the uninfected control group, which indicates that there were additional factors that affected the measured frequency of the CD4+CD25hiFOXP3+ cells; this could include technical or environmental changes such as seasonal effects, that might be associated with longitudinal studies. However, as both sampling time-points occurred during the long rainy season changes in Treg frequency in the control group are less likely to reflect seasonal changes and other factors such as exposure to concomitant infections could play a role.
In order to obtain a global assessment of the effect of S. haematobium infection on Th1, Th2, regulatory and pro-inflammatory cytokine responses we applied PCA analysis. This allowed us to summarize the various responses into two principal components [29]. Principal component 1 (PC1) reflected regulatory and Th2-polarized cytokine responses due to its positive loading with IL-5, IL-10 and IL-13, responses commonly associated with chronic schistosome infections. Principal component 2 (PC2) reflected pro-inflammatory and Th1-polarized cytokine responses due to its positive loading with IFN-γ, IL-17 and TNF, responses more commonly associated with acute schistosome infection or bacterial infections such as tuberculosis. We show that S. haematobium infection is associated with hypo-responsiveness as demonstrated by increases in cytokine production represented by both PC1 and PC2 following treatment of schistosomiasis with praziquantel. T cell division was also assessed, but despite the consistently higher proliferation to all stimuli tested, at post treatment, the change was not statistically significant. These data indicate that the increased frequency of CD4+CD25hiFOXP3+ Tregs during schistosome infection may be associated with poor cytokine responsiveness.
To assess the functional capacity of the regulatory T cells, a field applicable method was used which consists of the depletion of regulatory T cells from PBMC to assess their effect on cytokine production or proliferation. The data show that depletion of Tregs is associated with increased cytokine production, of both PC1 and PC2 which means that both Th2/regulatory and Th1/pro-inflammatory cytokine production improves. This is the case at both pre-treatment and post-treatment time-points, although the increase appears to be stronger at pre-treatment. Altogether, this would suggest that even though regulatory T cell numbers change with infection, their functional capacity to supress cytokine production remains.
Furthermore we evaluated the effect of Treg depletion on cell proliferation. While cytokine responses were similarly affected at both pre-and post-treatment, proliferative responses were predominantly affected by Treg depletion in infected individuals at pre-treatment only. These data could be explained if the ability of Tregs to suppress proliferation would be distinct from suppressive mechanisms required for inhibiting cytokine production, for example Tregs only supress PBMC proliferation in a higher Treg:responder ratio as seen during active infection. The removal of S. haematobium infection could then affect the Treg function partially. Tregs are thought to exert their function via a number of different mechanisms including IL-10 and/or TGF-β production, IL-2 consumption, or cell-cell contact where inhibitory molecules such as CTLA-4 and PD-1 are key [30,31]. A recent study suggests a unique role for microRNAs in suppression of T cell proliferation by Tregs [32]. Future studies are needed to delineate how Tregs exert their suppressive role during the course of schistosome infection and furthermore the difference in mechanism between suppression of proliferation versus effector cytokine production. Additional alternative mechanism, such as T cell anergy due to increased expression of the E3 ubiquitin ligase GRAIL (gene related to anergy in lymphocytes) which has been shown in a mouse model to be linked to Th2 cell hypo-responsiveness could also play a role and should also be investigated in future studies [33]. The role of inflammation in the induction and maintenance of Treg cells should likewise be considered as Shistosoma infection is a chronic inflammatory disorder and Tregs protect the human host against excessive inflammation, thus increased Treg numbers during schistosomisis may be a responses against inflammation rather than directly induced by the parasites [34].
Alternatively, recently described regulatory CD8+ T cells which likewise produce IL-10, may also in part contribute to the differences observed [35,36]. CD25hi cell depletion will in addition to depleting CD4+CD25hiFOXP3+ T cells also deplete the CD8+CD25hiFOXP3+ T cell population and therefore future studies are needed to re-assess the relative contributions of these different subsets. Moreover studies with more extensive panels of markers associated with suppressive T cell functions are necessary as FOXP3 expression has been shown to be transiently up-regulated on activated CD4+ T cells [37].
Finally, the concomitant role of the different regulatory cells, including in addition to Tregs, regulatory B cells [38] and regulatory monocytes/macrophages [39,40] and their relative contribution to the suppressive activities observed need to be further investigated.
In summary, this study shows that infection with S. haematobium is associated with alterations of the frequency and activity of CD4+CD25hiFOXP3+ regulatory T cells and that these in turn affect proliferation and global cytokine responses. These data indicate that the functional activity of regulatory T cells needs to be taken into consideration when studies consider co-infections, treatment or vaccine responses in areas where helminths are prevalent.
|
10.1371/journal.pgen.1006835 | The yeast protein kinase Sch9 adjusts V-ATPase assembly/disassembly to control pH homeostasis and longevity in response to glucose availability | The conserved protein kinase Sch9 is a central player in the nutrient-induced signaling network in yeast, although only few of its direct substrates are known. We now provide evidence that Sch9 controls the vacuolar proton pump (V-ATPase) to maintain cellular pH homeostasis and ageing. A synthetic sick phenotype arises when deletion of SCH9 is combined with a dysfunctional V-ATPase, and the lack of Sch9 has a significant impact on cytosolic pH (pHc) homeostasis. Sch9 physically interacts with, and influences glucose-dependent assembly/disassembly of the V-ATPase, thereby integrating input from TORC1. Moreover, we show that the role of Sch9 in regulating ageing is tightly connected with V-ATPase activity and vacuolar acidity. As both Sch9 and the V-ATPase are highly conserved in higher eukaryotes, it will be interesting to further clarify their cooperative action on the cellular processes that influence growth and ageing.
| The evolutionary conserved TOR complex 1 controls growth in response to the quality and quantity of nutrients such as carbon and amino acids. The protein kinase Sch9 is the main TORC1 effector in yeast. However, only few of its direct targets are known. In this study, we performed a genome-wide screening looking for mutants which require Sch9 function for their survival and growth. In this way, we identified multiple components of the highly conserved vacuolar proton pump (V-ATPase) which mediates the luminal acidification of multiple biosynthetic and endocytic organelles. Besides a genetic interaction, we found Sch9 also physically interacts with the V-ATPase to regulate its assembly state in response to glucose availability and TORC1 activity. Moreover, the interaction with the V-ATPase has consequences for ageing as it allowed Sch9 to control vacuolar pH and thereby trigger either lifespan extension or lifespan shortening. Hence, our results provide insights into the signaling mechanism coupling glucose availability, TORC1 signaling, pH homeostasis and longevity. As both Sch9 and the V-ATPase are highly conserved and implicated in various pathologies, these results offer fertile ground for further research in higher eukaryotes.
| In Saccharomyces cerevisiae, Sch9 is part of the highly conserved TORC1 pathway which plays a central role in the nutrient-induced signaling network, thereby affecting many aspects of yeast physiology such as stress resistance, longevity and cell growth [1–3]. The rapamycin-sensitive TORC1 mediates these effects mainly via two key branches. In the first branch, activated TORC1 phosphorylates and inhibits Tap42, which in turn controls the activity of type 2A and type 2A-like protein phosphatases [4, 5]. In the second branch, TORC1 contributes to Sch9 activation by phosphorylating multiple residues in its C-terminus [6]. In addition to TORC1-mediated activation, Sch9 can be phosphorylated by the sphingolipid-dependent Pkh1-3 kinases on the conserved PDK1 site, and this phosphorylation is indispensable for its function [6, 7]. Although Sch9 is a downstream effector of the TORC1 complex, the protein kinase can also function independently of TORC1 [3, 8, 9]. Moreover, it has been proposed that Snf1, the orthologue of mammalian AMP kinase, also modulates Sch9 activity by phosphorylation [10]. As three different kinases modulate Sch9 activity in response to diverse stimuli, Sch9 appears to act as a major integrator that regulates various aspects of yeast physiology. A prime example of this is the control of lifespan by Sch9. Indeed, both tor1Δ and sch9Δ strains display increased lifespan as compared to the WT strain [2, 11], and downregulation of nutrient signaling via the TORC1-Sch9 branch seems to be part of an evolutionary conserved mechanism that extends lifespan across a wide range of eukaryotic species [12, 13].
The V-ATPase is a highly conserved proton pump that mediates the luminal acidification of multiple organelles of the biosynthetic and endocytic pathway, thereby regulating numerous cellular processes including vesicle trafficking, autophagy, pH- and ion- homeostasis (reviewed in [14–16]). These V-ATPases are multi-subunit protein complexes consisting of a membrane-embedded V0 sector containing the proton translocation pore, and an attached peripheral cytosolic V1 sector responsible for ATP hydrolysis to fuel proton transport. Although higher eukaryotes often exhibit tissue- and/or organelle-specific expression of multiple isoforms of one subunit, in yeast only the V0 sector subunit a is encoded by organelle specific homologues: VPH1 encodes the isoform localized at the vacuole, while STV1 encodes the isoform that cycles between the Golgi apparatus and endosomes [17].
In both yeast and higher eukaryotes, V-ATPase activity is tightly regulated by reversible assembly of the V1 and V0 sector [14, 18–20]. Although the exact molecular mechanism governing glucose-dependent reversible assembly is still a matter of debate, recent reports shed some light on the signaling mechanisms by which yeast might control this assembly process. Addition of glucose to carbon starved cells triggers an increase in cytosolic pH (pHc), possibly through a rise in ATP levels. The V-ATPase responds to pHc by assembling and transducing the cellular signal through distinct GTPases, ultimately leading to enhanced Ras/PKA and TORC1 activity. As a result, cells adapt growth in response to carbon source availability [21, 22]. Besides fermentable sugars, additional signals, such as osmotic stress and high extracellular pH, also influence V0-V1 assembly levels [22–24]. Interestingly, increasing evidence suggests that V-ATPase activity is required for regulating cell survival in both yeast and higher eukaryotes [25–29].
To better understand the mechanisms by which Sch9 regulates cell physiology in yeast, we performed a genome-wide synthetic genetic array (SGA) screening and identified mutants that require Sch9 function for their growth and survival. Gene ontology (GO) analysis revealed the V-ATPase as one of the most significant hits. Further analysis showed that Sch9 physically interacts with the V-ATPase, thereby influencing V-ATPase assembly/disassembly in response to glucose availability, while receiving input from TORC1. Importantly, we show that the interaction of Sch9 with the V-ATPase is required to allow proper control of both pHc and vacuolar pH (pHv), and we found that particularly pHv is important to determine longevity.
As only a few direct substrates are known for Sch9 [30–32], we performed a genome-wide SGA analysis to discover additional functions and targets. To this end, the sch9Δ mutant was mated with the library of non-essential deletion strains and double deletion mutants were scored for a synthetic growth phenotype. GO analysis on the identified genes (S1 Table) revealed several functional classes for which a role of Sch9 had already been established, such as transcription, protein synthesis, mitochondrial function and cellular amino acid biosynthesis [3, 32, 33], validating the approach. Interestingly, our screening also identified numerous new genes displaying a genetic interaction with SCH9 as only for a small subset of these genes (± 15%) a putative interaction with SCH9 has been predicted by BioGRiD (S1 Table). Two GO categories that were significantly enriched in our screening, but not represented in BioGRID are connected to vesicular trafficking and V-ATPase function (S2 Table and Fig 1).
We first investigated whether the synthetic interaction between SCH9 and genes encoding proteins involved in vacuolar protein sorting arises from a defect in vesicular trafficking in the sch9Δ mutant. To this end, we monitored the localization of several marker proteins known to be sorted to the plasma membrane or vacuole by one of different trafficking routes (S1 Fig) [34]. Fig 2 (Fig 2A and Fig 2B) shows that the soluble hydrolase carboxypeptidase Y (CPY) was correctly sorted to the vacuolar lumen in the sch9Δ mutant, where it was processed to its mature form. Interestingly, deletion of SCH9 results in increased abundance of CPY (Fig 2A and S2A Fig), and, in contrast to vacuolar protein sorting mutants [35], CPY was not secreted from the cell in the sch9Δ strain (S2B Fig), additionally confirming the CPY pathway is not impaired. The CPY receptor Vps10, which is recycled from the late endosome (LE) to the trans-Golgi network (TGN) after CPY dissociation [36], localizes to punctate endosome and Golgi compartments in both WT and sch9Δ mutant cells, suggesting that retrograde transport from LE to TGN is not compromised in the sch9Δ strain (S2C Fig). Mutants affecting early endosome function mislocalize the v-SNARE protein Snc1. In agreement with previously data [37], most of the GFP-Snc1 fluorescence was present at the plasma membrane in both WT and sch9Δ cells and only a fraction of the protein was observed in internal cellular structures (S2D Fig). This result indicates that the secretion of proteins to the cell surface, as well as the recycling of proteins from the cell surface to the TGN is not impaired in the sch9Δ mutant. Additionally, the successful staining of internal membranes of sch9Δ cells with FM4-64 indicates that SCH9 deletion does not impair endocytosis nor impact vacuolar morphology in exponentially growing cells (Fig 2B and 2D).
Proteins are also transported to the vacuole by non-endosomal routes. Indeed, the alkaline phosphatase (ALP) route directly transports proteins from the TGN to the vacuole [38]. Moreover, the selective cytoplasm-to-vacuole-targeting (Cvt) and non-specific autophagy pathways deliver cytosolic proteins to the vacuole [39, 40]. As can be seen in Fig 2 (Fig 2C and 2E), the vacuolar hydrolases ALP, encoded by PHO8, and Ape1 are processed to their active form in sch9Δ cells, suggesting that Sch9 does not play an essential role in the ALP or Cvt pathway, respectively. In line with this, Sch9 did not influence GFP-Pho8 localization to the vacuolar membrane (Fig 2D). Previous studies have implicated TORC1, Sch9 and PKA in negatively regulating autophagy [8, 41]. We used the Pho8Δ60 and GFP-Atg8 processing assays to monitor the delivery and lysis of autophagic bodies in the vacuolar lumen [42]. Although we observed a small increase in basal and starvation induced autophagic flux in the sch9Δ mutant (Fig 2F and S2E Fig), this small difference is unlikely to explain the observed synthetic growth defects when combining the deletion of SCH9 with that of genes encoding proteins involved in vesicular trafficking. Moreover, none of the ATG genes were retrieved in our SGA screening. Collectively, our results show that Sch9 does not directly control vesicular trafficking pathways for protein transport to the vacuole, the plasma membrane, or for endocytosis.
Another GO category that was significantly enriched in our screening is connected to the vacuolar proton pump (Fig 1). Since the V-ATPase is a key regulator of pH homeostasis [43] and since the majority of Sch9 localizes to the vacuolar membrane, where also the V-ATPase is found [17, 44], we investigated whether the deletion of SCH9 affected pH. As a first indication, we measured the ability to acidify the extracellular medium upon re-addition of glucose to glucose-starved cells. Consistent with a potential role in maintaining pH homeostasis, we found that cells lacking SCH9 displayed a reduced glucose-activated proton export similar as been reported for cells lacking a functional V-ATPase (Fig 3A) [43]. Next, we monitored the effects on pHc. We could not observe a difference in pHc between WT and sch9Δ cells during fermentative growth (Fig 3B), although sch9Δ cells maintained their neutral pHc longer due to slower growth and, consequently, later depletion of glucose. However, after the diauxic shift, the pHc of sch9Δ cells dropped below that of WT (Fig 3B). Similarly, a more acidic pHc for the sch9Δ mutant was observed when cells were starved for carbon (Fig 3C and S3A Fig). In line with previous data [45], subsequent re-addition of glucose to WT starved cells resulted in a rapid acidification of the cytosol followed by an alkalization to neutral pH within 3 minutes after the glucose pulse. In contrast, the sch9Δ mutant displayed a retarded recovery of pHc after glucose re-addition as it reached neutral pH values only 5 minutes after the pulse (Fig 3C). Whereas the rapid acidification step is believed to be caused by initiation of glycolysis, subsequent alkalization of the cytosol is the result of coordinated activation of the plasma membrane H+-ATPase Pma1 and the vacuolar V-ATPase. Hence, similar to sch9Δ cells, the pHc of mutants with a dysfunctional V-ATPase also recovers more slowly after glucose deprivation [43]. Thus, the data described above are all consistent with a functional link between Sch9 and the V-ATPase.
Because Sch9 is a known effector of the nitrogen-responsive TORC1 complex, we asked whether nitrogen starvation, similarly to glucose starvation, may influence pHc in a Sch9-dependent manner. Interestingly, pHc remained unchanged following nitrogen starvation in both WT and sch9Δ cells (Fig 3D and S3B Fig).
Because of the high-throughput nature of our SGA screen, we decided to manually validate the genetic interaction between Sch9 and the V-ATPase using tetrad dissection. Remarkably, all mutants that combined the deletion of SCH9 with the deletion of a vma subunit displayed a synthetic sick phenotype (Fig 4A and S4 Fig). A quantitative measure of the synthetic fitness defect was obtained by measuring colony sizes using Image J (Fig 4B and S5A Fig). Consistent with the genetic interaction, we observed a severe deteriorated growth phenotype of these mutant strains on fully supplemented medium, even when this medium was buffered at pH 5 to fully support growth of the V-ATPase mutants (Fig 4C and 4D, S5B–S5E Fig). Interestingly, our analysis confirmed the partial functional redundancy of the V0 subunit a isoforms [17], as it required the simultaneous deletion of VPH1 and STV1 to observe a synthetic growth phenotype in combination with the deletion of SCH9 (Fig 4 and S5 Fig). Taken together, we established a negative genetic interaction between SCH9 and the V-ATPase complex and show that this is a general phenomenon that cannot be attributed to a specific V-ATPase subunit or sector.
To gain insight into the molecular mechanisms underlying the genetic interaction, we sought processes that are regulated by both Sch9 and the vacuolar proton pump. To this end, we assayed growth on media known to impair growth of either the sch9Δ or V-ATPase deficient mutants [9, 14]. Similar growth defects were observed for both sch9Δ and various V-ATPase mutants when they were exposed to osmotic stress or grown on media with high calcium concentrations. In addition, both types of mutants were unable to grow on non-fermentable carbon sources and on YPD medium containing 60 mM CaCl2 buffered at pH 7.5 (Table 1 and S6 Fig). Interestingly, as both Sch9 and V-ATPase activity are required for tolerance to rapamycin and manganese, two substances known to affect TORC1 signaling [6, 46], our results point towards a functional relationship between TORC1 and the V-ATPase. Unlike V-ATPase deficient mutants, however, the sch9Δ mutant is tolerant to zinc and high extracellular pH, but the additional deletion of SCH9 does not restore growth of the V-ATPase deficient mutants on these two media (Table 1). On the other hand, the deletion of SCH9 appears to cause enhanced sensitivity of the stv1Δ or vph1Δ strains to osmotic stress and elevated extracellular calcium, and to partially impair growth on non-fermentable carbon sources. In general, it seems that phenotypes which rely on the H+ pumping ability of the V-ATPase are only mildly influenced by the deletion of SCH9, while mostly the phenotypes that are not readily attributable to the vacuole acidifying function of the V-ATPase, i.e. the non-canonical functions, appear to be affected by loss of Sch9. Taken together, these results show that the sch9Δ mutant has a partial vma- phenotype and suggest that Sch9 may somehow regulate the V-ATPase.
One of the best studied phenotypes associated with the loss of Sch9 is the extension of chronological lifespan (CLS) [2, 12, 13]. Although V-ATPase activity has been implicated in regulating yeast ageing [25, 29], not much is known about the effect on CLS when the V-ATPase function is abrogated in S. cerevisiae. Hence, we determined the CLS of mutants lacking SCH9 and/or the V-ATPase subunit encoded by VMA2. To this end, cells were grown in complete synthetic medium with 2% glucose and viability of cells was measured 8 days after they entered stationary phase. In line with previously published data [2, 11], we observed an increased lifespan for sch9Δ cells as compared to WT. In contrast, the vma2Δ mutant displayed a significantly reduced viability (Fig 5A and S7A Fig). Interestingly, the CLS decreased dramatically in the vma2Δsch9Δ mutant even when compared to vma2Δ mutant alone. This suggests that when V-ATPase activity is compromised, Sch9 is important for maintaining cell viability. Very similar results were obtained when we extended our analysis to other V0 and V1 subunits (S8A and S8B Fig). Thus, the role of Sch9 in lifespan determination is highly dependent on the presence of a functional V-ATPase.
As especially superoxide anions are detrimental to survival and as both Sch9 and V-ATPase activity have been implicated in oxidative stress resistance [11, 26], we assessed the levels of endogenous superoxide anions using dihydroethidium (DHE) during chronological ageing. We found that in stationary phase cells the level of this reactive oxygen species (ROS) was elevated in the vma2Δ mutant when compared to WT cells and this difference became more pronounced as cells aged (Fig 5B and S7B Fig). In sch9Δ cells, the level of ROS also increased during ageing, but here the amount of ROS was always significantly lower than in WT cells. In agreement with the reduction in CLS, the additional deletion of SCH9 in the vma2Δ mutant resulted in a striking increase in cells showing DHE staining, indicating that superoxide-induced oxidative stress may represent one of the main factors contributing to the rapid ageing of this mutant. Indeed, only a minor fraction of the vma2Δsch9Δ cells stained with DHE did not accumulate SYTOXgreen at day 8 in stationary phase (S8C Fig). Moreover, the fraction of cells displaying SYTOXgreen staining but no accumulation of superoxide anions was negligible in all investigated strains (< 0.2%).
It has been shown that extracellular acidification is an important extrinsic factor affecting CLS [47–50]. Hence, we also determined CLS of the WT and mutant strains in complete synthetic medium buffered to pH 5.5. As shown, buffering of the medium indeed reduced mortality significantly as the WT and single sch9Δ and vma2Δ strains maintained their viability during the first week of chronological ageing, while there was only a small drop in survival for the vma2Δsch9Δ mutant (Fig 5C and S7C Fig). Again, an inverse correlation was seen with the ROS levels measured under the same conditions (Fig 5D and S7D Fig). However, when cells were allowed to age for a longer period, it was once more evident that also in buffered medium Sch9 promoted ageing in case of a functional V-ATPase, while it supported survival when V-ATPase function was compromised (Fig 5E). Of note, we also measured the pH of the buffered medium in the aged cultures. We noticed enhanced acidification of the culture medium when cells were lacking Sch9, while the medium pH of aged WT and vma2Δ cells was higher and similar. Thus, the pH of the medium cannot explain why the role of Sch9 in regulating longevity switched from pro-ageing to pro-survival upon impairment of the V-ATPase (Fig 5F).
The amino acid composition of the growth medium can affect yeast CLS [51, 52], with methionine availability having a highly significant impact [53]. Indeed, several studies demonstrated that genetic or dietary restriction of methionine promotes longevity [54, 55]. Because our strains are all in the BY4741 background and thus contain a deletion of MET15, and since we retrieved MET6 and MET22 from the genome-wide SGA screening (Fig 1 and S1 Table), we wondered whether methionine availability would differentially influence the CLS of our strains. To this end, strains were aged in non-buffered synthetic medium containing different concentrations of methionine. For WT and sch9Δ mutant cells, lifespan decreased with increasing methionine supply, while for cells with a dysfunctional V-ATPase, i.e. the vma2Δ and vma2Δsch9Δ cells, survival was better in medium containing the standard 20 mg/L methionine than in medium containing lower or higher concentrations of the amino acid (Fig 5G and S7E Fig). These data confirm that CLS depends on methionine availability and is determined in part by V-ATPase function [55]. Nonetheless, independent of methionine availability, the deletion of SCH9 still extended longevity, while it reduced longevity when combined with disruption of the V-ATPase activity. Again, a tight correlation between survival and superoxide levels could be observed for all methionine concentrations tested (Fig 5H and S7F Fig).
Since our data indicate that Sch9 impacts on pH homeostasis and since several studies have linked yeast ageing to V-ATPase activity and vacuolar acidification [25, 29, 56], we also measured pHv in WT and mutant strains to determine whether this could explain the apparent differential roles of Sch9 in regulating CLS. We used the pH-sensitive fluorescent dye BCECF-AM and performed measurements in cells growing exponentially on glucose, as well as in glucose-starved cells. As compared to WT cells, the sole deletion of SCH9 was associated with a significant drop in pHv in both conditions, indicative for enhanced V-ATPase activity (Fig 5I). Remarkably, when the SCH9 deletion was introduced in the strain lacking Vma2, it caused an increase in pHv. This effect was only apparent in cells growing on glucose, suggesting that especially under these conditions Sch9 controls pHv also independently of the V-ATPase. Because the changes in pHv correlated well with the CLS profiles for the strains studied, our data are in line with a model in which vacuolar acidity dictates cellular longevity.
Both Vph1-containing V-ATPase complexes and Sch9 are known to locate at the vacuolar membrane during fermentative growth. Hence, we reasoned they might physically interact. Prior to studying this interaction, we identified conditions that influenced the intracellular localization of Sch9 or the assembly state of the V-ATPase. In agreement with previous work [33], GFP-Sch9 was found to be enriched at the vacuolar membrane in exponentially growing cells, but a significant portion dissociated from the vacuolar membrane upon glucose starvation (Fig 6A). In contrast, the protein kinase remained stably associated with the vacuolar membrane upon nitrogen starvation and rapamycin treatment (Fig 6A), as well as in V-ATPase deficient mutants (Fig 6B), indicating that the intracellular localization of Sch9 is specifically regulated by C-source availability. Concerning V-ATPase assembly, we found that the V-ATPase was fully assembled in both WT and sch9Δ cells during exponential growth, as well as upon nitrogen starvation and rapamycin treatment (Fig 6C and 6E, S9A Fig). However, when sch9Δ cells were subjected to glucose starvation a significant fraction of Vma5-RFP remained localized with Vph1-GFP at the vacuolar membrane, as indicated by fluorescence intensity profile plots (Fig 6D and 6F, S9B Fig) and by the Pearson’s coefficient (R) (S9C Fig). Thus, the disassembly of the V-ATPase is apparently hampered in sch9Δ cells, which may explain why the cells have a lower pHv as described above.
Since microscopic analyses did not allow quantifying an absolute value of assembled V-ATPase complexes, we further studied the V-ATPase assembly state and the putative interaction of Sch9 with the V-ATPase by co-immunoprecipitation (co-IP). Accordingly, Fig 7A shows that in exponentially growing cells a strong interaction between the V1 and V0 sectors, and between the V-ATPase and HA6-Sch9 could be observed. Upon glucose depletion, both interactions weakened considerably, but were rapidly restored by re-supplementation of glucose. In contrast, nitrogen deprivation did not impact on either interaction (Fig 7B); at most there was a slight decrease in the interaction between Vma1 and Sch9. As this interaction was not strengthened by the subsequent supplementation of nitrogen, the minor decrease cannot be attributed to a starvation effect. Concerning V-ATPase assembly levels in WT and sch9Δ cells, the results in Fig 7C indicate that in exponentially growing cells, both the deletion of SCH9, or the treatment of WT cells with rapamycin, significantly increases V-ATPase assembly as compared to untreated WT cells. Because the effect of rapamycin is comparable to that triggered by the deletion of SCH9, the data suggest that Sch9 functions downstream of TORC1 to modulate V-ATPase assembly (Table 2). In line with our microscopy data, the absence of Sch9 also significantly lowered the amount of V-ATPase that disassembled upon glucose-starvation. This effect is only partially mimicked when WT cells are treated with rapamycin, suggesting that Sch9 may facilitate glucose starvation-induced V-ATPase disassembly to some extent independently of TORC1. Importantly, glucose depletion still triggered V-ATPase disassembly in the absence of the Sch9, indicative that the role of this kinase is only modulatory.
To confirm our data, we repeated the co-IP experiments using strains expressing different Sch9 mutants. First, we monitored the effect of glucose availability on V-ATPase assembly in a strain expressing the analog-sensitive sch9as allele [33], the activity of which can be blocked using the ATP-analog 1-NM-PP1. As compared to our previous data with WT cells, even without inhibitor more V-ATPase remained assembled in the sch9as strain upon glucose-starvation, but consistent with the data obtained for the sch9Δ mutant, more V-ATPase assembly was observed when Sch9 activity was blocked by 1-NM-PP1, and this independently of whether the cells were exponentially growing or glucose starved (S9D Fig). Next, we tested the requirement of TORC1-dependent Sch9 phosphorylation by comparing rapamycin-induced V-ATPase assembly in WT and sch9Δ mutant strains complemented with either wild-type Sch9, the Sch95A or the phosphomimetic Sch92D3E alleles [6]. As expected, expression of the wild-type Sch9 allele in the sch9Δ mutant restored the V-ATPase assembly state to WT levels and rendered it again sensitive to rapamycin (Fig 7D). In contrast, rapamycin did not significantly affect V-ATPase assembly upon expression of the Sch9 phospho-mutants. While maximal assembly was obtained in the presence of Sch95A, similar as seen for the empty vector control, an intermediate level was found in case of the Sch92D3E allele. The latter indicates that this TORC1-independent Sch92D3E allele may not be fully functional in downstream signaling as noted before [57]. Nonetheless, when combined, our results suggest that Sch9 integrates input from TORC1 to influence the glucose-dependent assembly state of the V-ATPase.
By conducting a genome-wide SGA screening, we defined a global SCH9 genetic interaction network in yeast and, as such, identified numerous new genes that may function as Sch9 effector or act in pathways connected to Sch9. Among these hits were several genes involved in modulating vacuolar biogenesis and function. However, we could not find evidence that Sch9 is involved in regulating trafficking routes that deliver material for vacuolar biogenesis or degradation. Instead, we found this protein kinase to influence pHc, pHv and extracellular acidification. We also demonstrated that Sch9 interacts with V-ATPase subunits to modulate the assembly state of the latter in function of nutrient availability, thereby integrating input from TORC1 and a yet undefined glucose-dependent sensor.
Our findings are consistent with observations made in a recent study that coupled vacuolar biogenesis and functioning to cell growth and cell cycle progression [58]. This study demonstrated that Vph1 and components of the TORC1 complex are delivered to the vacuolar membrane early during vacuole biogenesis, while Sch9 is only recruited at a later stage. Moreover, this study reported that Sch9, along with TORC1, signals the cell-cycle machinery that a functional vacuole is present. Interestingly, Sch9 is thereby not only activated by TORC1-dependent phosphorylation, but also by additional signals that require a functional vacuole [58]. That Sch9 is involved in determining cell growth and division was already known for some time, but it remained mainly connected to a dynamic network that couples nutrient availability with ribosome biosynthesis [33, 59].
Consistent with its prominent role in nutrient storage, the vacuole has emerged as central player regulating nutrient signaling pathways in both yeast and mammals [22, 60], though only few studies have begun to unravel the underlying molecular basis. For instance, Young et al. (2010) conducted a genome-wide screening to identify inositol auxotrophy mutants. The authors concluded that the drop in pHc triggered by glucose starvation releases the transcriptional repressor Opi1 from a lipid-sensor complex in the ER, which then translocates to the nucleus to repress the Ino2/4 transcription factors and as such many phospholipid metabolic genes [61]. Another genome-wide screening examined pHc and cell division rate during fermentative growth [62]. This screening retrieved several mutants that could be classified in different categories depending on their pHc-growth rate relationship. Both screenings not only identified various players and potential sensors involved in pHc signaling, but also provided additional links with Sch9. Indeed, more than 20% of the genes retrieved by each screening overlapped with our SGA screening (S10 Fig and S1 Table). As such, Sch9 seems to emerge as a key player connecting inositol and lipid metabolism with pHc, V-ATPase and growth. Whether this connection relates to the sphingolipid-dependent function of Sch9 [7, 63] or the PI(3,5)P2-dependent activation of the V-ATPase [24] and vacuolar recruitment of Sch9 [64] needs to be investigated in more detail, but most likely additional mechanisms are at play. We also compared the data from our SGA screening with those obtained by the Cardenas group who performed a genetic screening for synthetic interactions with TOR1 [65]. Albeit the latter allowed to link Tor1 signaling with vacuolar functions, the overlap between both screenings was mainly restricted to genes encoding cytoplasmic and mitochondrial ribosomal proteins (S1 Table). This may indicate that Sch9 does signal also independently of TORC1, as suggested before [3, 8, 9].
One well-established phenotype of the sch9Δ strain is its increased survival during stationary phase [2, 12], which is partly due to increased respiration and expression of mitochondrial oxidative phosphorylation subunits [11, 66]. Consistently, CLS extension by deletion of SCH9 can be blocked and even reversed to lifespan shortening by the additional deletion of respiratory genes, by introducing the SCH9 deletion in rho0 strains that lack functional mitochondria or by pregrowing cells under a different nutritional regime [66–68]. We now report that Sch9 can either extend or shorten CLS depending on the presence of a functional V-ATPase and the vacuolar acidity. This is in line with a previous study that connected CLS to the V-ATPase and autophagy-dependent vacuolar acidification under conditions of methionine restriction [55, 56]. Another study demonstrated that CLS extension by methionine restriction requires activation of the retrograde response pathway to regulate nuclear gene expression in function of mitochondrial activity [54]. Interestingly, our SGA-screening confirmed a genetic link between SCH9 with RTG2, encoding a sensor for mitochondrial dysfunction, RTG3, encoding a key mediator of retrograde signaling, and with the methionine metabolism genes MET6 and MET22. Hence, the question arises whether both vacuolar acidification and mitochondrial functioning are part of the same regulatory scenario determining longevity. At least for the control of replicative lifespan this seems to be the case. Here, vacuolar acidification is required to maintain proper pH-dependent vacuolar amino acid storage and this prevents age-induced mitochondrial dysfunction [25]. Furthermore, a systematic gene deletion analysis confirmed that several vacuolar mutants, including V-ATPase mutants, affect mitochondrial functions and display similar phenotypes as mitochondrial petite mutants [69].
Given the previously published link between Sch9 and respiratory capacity [11, 66] and the data presented in this paper connecting Sch9 with the V-ATPase and vacuolar pH, it is well possible that Sch9 is part of a system that monitors vacuolar and mitochondrial function in order to sustain growth and lifespan. Such monitoring system may be quite complex as evidenced by a study that demonstrated additive effects of methionine, glutamic acid and glucose availability on yeast longevity [53]. Interestingly this study also attributed a role to Sch9 in the sensing of methionine and glucose, while implicating Gcn2, a conserved protein kinase that links amino acid sensing with global protein synthesis, in the sensing of glutamic acid [53]. As Vam6/Vps39 regulates the formation of the vacuolar-mitochondrial contact sites involved in lipid transfer [70, 71], and contributes to the activation of TORC1 [44], it could play an important regulatory role in this monitoring system.
In contrast to Sch9, the localization of Tor1 in yeast is not regulated by glucose availability [72]. However, it has been shown that the essential TORC1 subunit Kog1 is transiently sequestered in cytoplasmic stress granules upon heat stress [73] and in cytoplasmic foci in a Snf1-dependent manner upon glucose starvation [72]. As such, the movement of Kog1 in and out of these so-called Kog1-bodies determines the formation of TORC1. Once reconstituted at the vacuolar membrane, the activity of TORC1 is controlled via Vps-C complexes and the amino acid sensing EGO complex (EGOC), the yeast functional counterpart of mammalian Rag-Ragulator [74–76]. Besides being a downstream effector of TORC1, a role of Sch9 in the control of EGOC has not been reported to our knowledge. However, such a role can be suspected given the data we now present on the contribution of Sch9 in regulating V-ATPase assembly and thereby vacuolar acidity. The latter is important for the vacuolar degradative capacity and the pH-dependent storage of amino acids in the vacuolar lumen [25]. In this context, it has been suggested that the Rag GTPases of EGOC could, in addition to mediating cytoplasmic amino acid signals [77], also sense the vacuolar amino acid load through amino acid transport across the vacuolar membrane [74–76]. In addition, and similar to the mammalian system, the Rag GTPase Gtr1 was found to interact with the V-ATPase raising the possibility that the proton pump itself could be involved in the activation of the GTPase [22, 60]. As depicted in Fig 8, the consequence of the above is that Sch9 can be part of a feedback loop that keeps V-ATPase and TORC1 activity in balance during growth. The idea of an Sch9-dependent feedback control of TORC1 is not new, because it was already proposed in a previous study that analyzed the effectors by which TORC1 controls the transcription of ribosomal protein and ribosome biogenesis genes [57]. Moreover, both TOR complexes have already been proposed to function in feedback loops to maintain cellular homeostasis [78].
Besides connections with TORC1, the V-ATPase was also identified as signaling intermediate linking C-source availability and pHc with activation of PKA via the GTPase Arf1 [21, 22]. Thus, by regulating V-ATPase assembly, Sch9 would also be upstream of the Ras/PKA pathway. It is known for some time that Sch9 is implicated in the Ras/PKA pathway [79], but only recently it was reported that Sch9 indirectly regulates the phosphorylation of the PKA regulatory subunit Bcy1 via the MAP kinase Mpk1 in a TORC1 dependent manner [80]. Whether this means that Arf1 signals through Mpk1 to control PKA activity needs to be investigated. Importantly, it is known that enhanced PKA activity prevents glucose-induced disassembly of the V-ATPase [81] and therefore also this signaling route is most likely subjected to feedback control.
Several mechanisms have been proposed to control V-ATPase assembly [15, 82]. Of these, the Vph1-specific interactor RAVE (Regulator of ATPase of Vacuoles and Endosomes) might be a good candidate to mediate Sch9 control on V-ATPase assembly, especially since our SGA screening retrieved the gene encoding the central RAVE component Rav1 [83]. However, another possible scenario comes from the observation that during exponential growth on glucose the deletion of SCH9 was associated with enhanced vacuolar acidification, but when the disruption of SCH9 was introduced in the strain lacking an active V-ATPase vacuolar alkalization was observed. The reason for this is not known, but it may indicate that Sch9 affects proton exchange by vacuolar antiporters like the Na+/H+ and K+/H+ exchangers Nhx1 and Vnx1, both known to play a role in pH homeostasis [84, 85]. If Sch9 would control such antiporters, it would provide a mechanism by which the kinase regulates V-ATPase assembly in response to glucose. Indeed, pH-dependent alterations in the N-terminal cytoplasmic domain of Vph1 has been proposed as possible mechanism governing glucose-dependent reversible assembly of the V-ATPase [86]. This model would explain the observed synthetic genetic interaction between SCH9 and the genes encoding the V-ATPase. Moreover, it is also in line with the observation that mammalian PKB co-localizes with, phosphorylates and inhibits the cardiac sarcolemmal Na+/H+ exchanger Nhe1 following intracellular acidosis [87]. This raises again the question whether Sch9 is the genuine orthologue of the mammalian orthologue of PKB/Akt [88, 89] or whether it combines this role with an S6K function in yeast [6].
S. cerevisiae strains and plasmids are listed in S4 and S5 Tables, respectively. All strains used in this study are derived from the BY4741 series. Cells were grown at 30°C in YPD (1% yeast extract, 2% bactopeptone, 2% glucose) or synthetic defined medium (Formedium; 0.5% ammonium sulfate, 0.17% yeast nitrogen base, amino acids, 2% glucose). When indicated, the culture medium was buffered to the specified pH with either MES, sodium citrate or MOPS buffer. For nitrogen starvation experiments, cells were made prototrophic by introducing auxotrophy complementing plasmid(s). For short-term nutrient deprivation, exponentially growing cells were washed with and further grown in starvation medium. For re-stimulation, cells were supplemented with a final concentration of 2% glucose or 0.2% glutamine. For serial dilution growth assays, stationary phase cells were diluted to an OD600nm of 1 and 10-fold serial dilutions were spotted. For growth curve analysis, cultures were grown for 48h and diluted to the same density. OD600nm was measured every two hours in a Multiscan GO Microplate Spectrophotometer (Thermo Scientific).
Diploids were generated by crossing the sch9Δ mutant (JW 04 039) with single deletion strains derived from the BY4741 Yeast Knock-out Collection (EUROSCARF) and incubated at room temperature on solid sporulation medium (1% potassium acetate, 1.5% ager) for 5–6 days. A small amount of sporulated cells was resuspended in water containing 0.02 mg/ml lyticase and incubated for 10–15 min at room temperature. Next, tetrads were dissected on a YPD plate using a micromanipulator (Singer instruments). After 3–5 at 30°C, the genotypes of the germinated spores were analyzed based on the segregation of the genetic markers, and/or by PCR analysis. Only deletion mutants with a BY4741 genotype (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) were used in subsequent experiments.
The S. cerevisiae deletion strain collection, constructed in the BY4741 background, was obtained from EUROSCARF (Frankfurt, Germany). SGA screening was performed as described previously [90], using the sch9::NATMX4 strain as bait. Colony sizes of single and double mutants were scored by visual inspection. Double mutant strains that displayed an aggravated colony size as compared to the respective single deletion strain and the single sch9Δ mutant were retained as potential candidates for a synthetic genetic interaction with SCH9.
Several interesting candidate genes involved in protein sorting and vacuolar function were manually confirmed using tetrad dissection (see above). After three to four days at 30°C, colony sizes resulting from individual spores were measured using ImageJ software (NIH). Briefly, average colony surface areas were determined using the particle analyzer function for at least seven independent colonies for each genotype. Mutants were categorized as having a negative genetic interaction with SCH9 when the relative colony size of the double mutant was less than the product of the relative colony sizes of the corresponding single mutants. To quantify the extent of co-localization, the Pearson correlation coefficient was generated using the ImageJ plugin JACoP. Quantifications were performed on different independent experiments, with at least 30 cells analyzed in total. Fluorescence intensity profile plots were created using the Plot Profile function of ImageJ.
Precultures grown overnight in YPD or minimal medium lacking uracil (Sch9 mutants) were inoculated in fresh YPD medium buffered at pH 5 with 50 mM MES and grown till exponential phase. Next, half of the culture was treated with 200 nM rapamycin (Sigma-Aldrich) for 30 min and subsequently starved for glucose in the presence of rapamycin. The untreated half was further grown for 30 min and subsequently starved for glucose. For V-ATPase assembly analysis by sch9as inhibition, precultures were diluted in YPD medium (pH 5, 50 mM MES) with or without 300 nm 1-NM-PP1 (Merck-Millipore) and grown for 6 hours, after which cultures were starved for glucose in the absence or presence of the inhibitor. After treatment and/or starvation, cells were collected by centrifugation at 4°C, washed with ice-cold PBS, snap frozen in liquid nitrogen and stored at -80°C. Protein extraction was performed by bead beating in buffer A (40 mM Hepes-NaOH [pH 7.5], 120 mM NaCl, 1 mM EDTA, 0.3% CHAPS, 50 mM NaF, 10 mM β-glycerophosphate) supplemented with complete protease inhibitor tablets without EDTA (Roche). Extracts were cleared by centrifugation and incubated overnight at 4°C with anti-Vma1 or anti-Vph1 (Abcam, 8B1 and 10D7). Next, samples were incubated with magnetic anti-IgG beads (Invitrogen) for 30 min and beads were washed four times with buffer A. Proteins were eluted by boiling in SDS-sample buffer and subjected to immunoblotting. Relative V-ATPase assembly was quantified by calculating the ratio of Vph1 IPed with anti-Vma1 (assembled V0V1 complexes) vs Vph1 levels IPed with both antibodies (free V0 and assembled V0V1) [20].
Cells were pregrown to stationary phase in non-buffered fully supplemented medium containing 2% glucose. Next, stationary phase cells were diluted to an OD600nm of 0.1 in fresh medium and grown for 48h, which was set as time zero. For standard CLS experiments, cells were aged in non-buffered fully supplemented medium (~ pH 5.5) containing 2% glucose. Buffered CLS experiments were performed in fully supplemented medium buffered at pH 5.5 with 100 mM MES. For CLS experiments with varying methionine concentrations, methionine was added at the indicated concentration to minimal medium lacking methionine. Of note, a standard concentration of 20 mg/L methionine was used throughout all other experiments. At various time points, cell death was measured by flow cytometry (Guava easyCyte 8HT, Merck Millipore) using SYTOXgreen or propidium iodide (Molecular Probes). Alternatively, cell survival was measured by clonogenictiy. To this end, the amount of cells/μl was determined by flow cytometry and 250 cells were plated on YPD agar plates. Subsequently colony forming units were counted and values are displayed as percentage of viable cells. ROS levels were measured via DHE staining and subsequent flow cytometry analysis. Collected flow cytometry data were processed and quantified with FlowJo software.
Cytosolic pH was measured in prototrophic yeast cells expressing pHluorin [45] grown in low fluorescence medium (loflo; Formedium) containing 2% glucose, buffered at pH 5 with 25 mM sodium citrate. Fluorescence emission was recorded at 510 nm using a FLUOstar OPTIMA microplate reader (BMG labtech) with excitation at 390 nm and 470 nm. For starvation experiments, log phase cells were washed twice with starvation medium buffered at pH 5 and fluorescence was measured every 5 min at 30°C for 1h. For pulse experiments, 2% glucose or 0.2% glutamine was administered to starved cells. For growth curve analysis, stationary phase cultures were re-inoculated at the same density in fresh loflo medium and monitored every hour for pHc and OD600nm. Calibration was performed by incubating digitonin permeabilized cells in citric acid/Na2HPO4 buffers of different pH values.
Vacuolar pH was measured as described previously with minor modifications [43]. Briefly, yeast cells were grown to log phase in fully supplemented loflo medium containing 2% glucose, buffered at pH 5 with 50 mM MES and labelled with 50 μM BCECF-AM (Thermo Scientific). Next, cells were washed twice in growth medium with or without 2% glucose and fluorescence was recorded for 30 min at 30°C using a Fluoroskan Ascent FL Microplate Fluorometer and Luminometer (Thermo Scientific). Fluorescence emission was recorded at 538 nm after excitation at 440 nm and 485 nm. Calibration curves were constructed for each strain in every experiment using the calibration mixture described by Brett et al. [84], except that the ionophores (monensin and nigericin) were omitted.
The acidification of the culture medium was monitored using the pH indicator bromocresol green sodium salt (BCG; Sigma-Aldrich). For glucose-induced acidification of the medium, cells were grown to exponentially phase in fully supplemented medium containing 2% glucose, buffered at pH 5. Next, cells were washed with glucose starvation medium and resuspended at an OD600nm of 0.1 in starvation medium containing 0.01% BCG. The absorbance of the medium (595 nm) was monitored for 1h in a Multiscan GO Microplate Spectrophotometer. Medium acidification was initiated by the addition of 2% glucose and changes in absorbance over time were recorded. For the measurement of growth media pH of ageing cultures, cells were pelleted and BCG was added to the supernatants at a final concentration of 0.01%. A calibration curve was used to convert the measured absorbance to pH values.
Autophagy was monitored using the Pho8Δ60 and GFP-Atg8 processing assay as described previously [91]. Concerning the Pho8Δ60 assay, the generation of p-nitrophenol from p-nitrophenyl phosphate (Sigma-Aldrich, N9389)] was monitored in pho8Δ and pho8Δsch9Δ mutant strains harboring the PHO8Δ60 gene by measuring absorbance at 405 nm using a Beckman DTX880 plate reader (Molecular Devices). Specific activities were calculated as nmol p-nitrophenol/min/mg protein. Data are the mean of at least four independent transformants. Concerning the GFP-Atg8 assay, TCA protein extracts were prepared from WT and sch9Δ strains harboring the GFP-Atg8-expressing plasmid and equal amounts of proteins were resolved on a 10% SDS-PAGE. Blots were probed with anti-GFP (Roche Diagnostics).
The cytoplasm-to-vacuole pathway was monitored using the prApe1 processing assay as described previously [39]. Briefly, WT and sch9Δ cells were harvested, proteins precipitated using the TCA method and equal amounts were loaded on a 10% SDS-PAGE. After western blotting, membranes were probed with anti-Ape1 (kindly provided by Dr. Klionsky).
Missorting of CPY was measured by a colony overlay assay as described previously [35]. Briefly, cells were grown to stationary phase and spotted on fully supplemented medium at the indicated OD600nm. Plates were placed at 30°C for 4-6h and overlaid with a nitrocellulose membrane. After ± 24h of growth at 30°C, the membrane was washed several times with distilled H2O and TBS buffer containing 0.1% Tween-20, and subjected to immunoblotting with anti-CPY (Molecular probes, 10A5).
V-ATPase assembly was investigated by co-localization of pRS315-Vph1-GFP (gift from Robert C. Piper) with Vma5-RFP. To this end, we constructed WT and sch9Δ strains expressing a chromosomally encoded, RFP-tagged version of Vma5 by crossing sch9::NATMX4 with RD157 [21]. Correct localization of Sch9 in V-ATPase deficient mutants was examined by co-transforming cells with pRS415-GFP-Sch9 and pRS316-mCherry-Pho8 [92]. Plasmids expressing fusion proteins that served as marker for vesicular compartments were generous gifts (S2 Table). All images were generated using a Leica DM 4000B fluorescence microscope (Leica Microsystems) equipped with a Leica DFC 300G camera.
Protein extraction and western blot analysis were performed as described previously [63]. Protein concentrations were determined using the Bradford method (Bio-Rad) or the Pierce 660 nm protein assay (Thermo Scientific). Equal amounts of protein were mixed with SDS-sample buffer and resolved on a SDS-PAGE gel. Either anti-ADH2 (Merck Millipore, AB15002) or anti-PGK1 (Molecular probes, 22C5) were used as loading controls. The ECL method was used for detection and blots were visualized using a UVP Biospectrum Multispectral Imaging System. Signals were quantified by densitometry using UVP VisionWorks LS software (VWR).
Unless stated otherwise, the results shown are mean values and standard deviations displayed as error bars. For other experiments, representative results are shown. The appropriate statistical tests were performed using GraphPad Prism. Significances: * p < 0.05, ** p < 0.01, *** p < 0.001.
|
10.1371/journal.pcbi.1003931 | Pattern of Tick Aggregation on Mice: Larger Than Expected Distribution Tail Enhances the Spread of Tick-Borne Pathogens | The spread of tick-borne pathogens represents an important threat to human and animal health in many parts of Eurasia. Here, we analysed a 9-year time series of Ixodes ricinus ticks feeding on Apodemus flavicollis mice (main reservoir-competent host for tick-borne encephalitis, TBE) sampled in Trentino (Northern Italy). The tail of the distribution of the number of ticks per host was fitted by three theoretical distributions: Negative Binomial (NB), Poisson-LogNormal (PoiLN), and Power-Law (PL). The fit with theoretical distributions indicated that the tail of the tick infestation pattern on mice is better described by the PL distribution. Moreover, we found that the tail of the distribution significantly changes with seasonal variations in host abundance. In order to investigate the effect of different tails of tick distribution on the invasion of a non-systemically transmitted pathogen, we simulated the transmission of a TBE-like virus between susceptible and infective ticks using a stochastic model. Model simulations indicated different outcomes of disease spreading when considering different distribution laws of ticks among hosts. Specifically, we found that the epidemic threshold and the prevalence equilibria obtained in epidemiological simulations with PL distribution are a good approximation of those observed in simulations feed by the empirical distribution. Moreover, we also found that the epidemic threshold for disease invasion was lower when considering the seasonal variation of tick aggregation.
| Our work analyses a 9-year time series of tick co-feeding patterns on Yellow-necked mice. Our data shows a strong heterogeneity, where most mice are parasitised by a small number of ticks while few host a much larger number. We describe the number of ticks per host by the commonly used Negative Binomial model, by the Poisson-LogNormal model, and we propose the Power Law model as an alternative. In our data, the last model seems to better describe the strong heterogeneity. In order to understand the epidemiological consequences, we use a computational model to reproduce a peculiar way of transmission, observed in some cases in nature, where uninfected ticks acquire an infection by feeding on a host where infected ticks are present, without any remarkable epidemiological involvement of the host itself. In particular, we are interested in determining the conditions leading to pathogen spread. We observe that the effective transmission of this infection in nature is highly dependent on the capability of the implemented model to describe the tick burden. In addition, we also consider seasonal changes in tick aggregation on mice, showing its influence on the spread of the infection.
| Several ecological studies have shown that the distribution of ticks on their hosts is often highly aggregated, with a large number of hosts harbouring few parasites and a small number harbouring a large number of them ([1]–[5]; other interesting references could be found in [6]). In addition, the distribution of tick development stages is coincident, rather than independent [7]. Specifically, those hosts feeding larval tick stages were simultaneously feeding the greatest number of nymphs. As a result, about of all hosts feed of both larvae and nymphs and the number of larvae feeding alongside nymphs is twice as many as it would be if the distributions were independent [7], [8]. The aggregation of parasites on hosts bears important implications for vector-borne disease dynamics, since the small fraction of hosts supporting the bulk of the vector population is also responsible for the majority of the pathogen transmission [9].
The transmission of tick-borne diseases is characterised by an intricate set of ecological and epidemiological relationships between pathogen, tick vector, vertebrate hosts and humans that largely determine their temporal and spatial dynamics [10]. Tick-borne disease dynamics feature several complexities, due to the presence of a number of heterogeneities in the system coupled with non-linear phenomena operating in the transmission processes between ticks, host and pathogen [11]. The transmission of pathogens from one tick to another, a pre-requisite for the establishment of cycles of infection, may occur via three different pathways depending on the pathogen (see [12] for a comprehensive review). First, adult female ticks may transmit the pathogen to eggs trans-ovarially. Second, ticks may infect a host during their blood meal, leading to a systemic infection in the host; ticks might then acquire the infection by feeding on an infected host, maintaining the infection trans-stadially. Third, ticks may become infected by co-feeding with infected ticks on the same host. Co-feeding transmission is also called non-systemic as it does not require the host to have a systemic infection, since pathogens are transmitted from one tick to another as they feed in close proximity. Vertebrate hosts may vary in their competency to support systemic and co-feeding transmission [13]. Tick-borne pathogens differ also for the mechanisms which they use to persist in nature. For instance, Rickettsia spp., the pathogen agents causing Rocky Mountain Spotted Fever, are maintained by systemic and trans-ovarial transmission in Dermacentor variabili and andersoni [14] while it has been observed that Borrelia Burgdoferi s.l. spirochaetes persist in nature by taking advantage of all three routes of transmission in I. ricinus, [13], [15].
In the case of the tick-borne encephalitis virus (TBEv), which is an increasing public health concern in Europe [16]–[18], trans-ovarial transmission seems to be relatively rare and its contribution is generally thought to be negligible [19]. On the other hand, both systemic and non-systemic transmission can take place on reservoir-competent rodent hosts. However, due to the very short duration of the TBEv infection in rodents, [20], the systemic route would only allow infection of a very limited number of ticks. Indeed, non-systemic transmission through co-feeding ticks is a more efficient transmission route for TBE [8], [20]. Different studies have shown that TBEv would not become established in competent hosts, such as rodents, without the amplification of the overall transmission efficiency provided by co-feeding transmission (see for instance [20]–[23]). The aggregation pattern of ticks on hosts therefore plays a more important role in the transmission of TBEv than in other tick-borne pathogens, such as Borrelia burgdorferi sensu lato and Anaplasma phagocytophilum, where other efficient routes of transmission have been observed.
Tick aggregation on hosts and correlation of tick stages facilitate co-feeding transmission and thus significantly increase the basic reproductive number, , of the pathogen, with direct implications for its persistence [23], [24]. Using different levels of aggregation (from independent to coincident aggregated distribution), Harrison and collaborators [23] showed that values of increase with progressive levels of aggregation, making it more likely for tick-borne pathogens to become established and persist. In addition, the authors of the cited works evinced that when ticks followed a coincident aggregated distribution, the increase of was greater than in the case of independent aggregated distributions.
The degree of aggregation of ticks can be measured in a number of ways. Since the appearance of influential works by Randolph [25] and Shaw et al. ([6] and [26]) the negative binomial (NB) distribution has been extensively used to describe tick aggregation on hosts (see e.g. [23], [27], [28]). Alternatively, other works suggested that different distributions characterised by larger tails than NB (i.e., predicting more rodents with very large tick burden than expected with NB), can be effective in describing tick aggregations. Specifically, a Poisson-LogNormal (PoiLN) mixed model has been successfully used to describe tick distribution on red grouse chicks [29], while Bisanzio and collaborators [30] showed the first evidence that the distribution heterogeneity of ticks on hosts seemed to be better described by a power-law (PL) than a negative binomial distribution. A suitable description of the distribution tail might have important consequences on the dynamics of the pathogen spreading process. Modelling the spread of vector-borne diseases through bipartite networks [30] showed that the extreme aggregation of ticks on hosts has dramatic consequences on the behaviour of the epidemic threshold.
In the current study we used an extensive data set of Ixodes ricinus ticks feeding on mice (a total of 4722 parasitised hosts collected in 9 years) to detect the best fit for the distribution of tick burden on mice by testing the performance of NB and PoiLN versus PL distribution, with particular interest in the shape of the distribution tail which is crucial to suitably describe the fraction of co-feeding ticks necessary for TBEv transmission. Then, we used a stochastic model to simulate the effect of fitting different tick distributions on the infection dynamics of a tick-borne pathogen. Specifically, we investigated the spread of a non-systemically transmitted pathogen (e.g. TBEv) by modelling the pathogen transmission between susceptible and infective ticks, considering only co-feeding transmission and distributing ticks on mice under the hypotheses of NB, PoiLN, and PL distributions. Finally, we investigated the seasonal variations in the pattern of tick burden distribution on mice and its implication on TBE-like infection dynamics.
All animal handling procedures and ethical issues were approved by the Provincial Wildlife Management Committee (renewed authorisation n. 595 issued on 04.05.2011)
Rodent tick burden data was collected by trapping mice using capture-marking-recapture techniques during 2000–2008. The study area was a mixed broadleaf woodland [7], [21], located in Valle dei Laghi within the Autonomous Province of Trento, in the north-eastern Italian Alps (grid reference 1652050E 5093750N, altitude 750–800 m a.s.l.). In the year 2000, mice were monitored in nine selected areas through placement of 8×8 trapping grids with a 15-m inter-trap interval. In 2001 and 2002 the number of trapping grids was reduced to eight, while from 2003 onward their number was further reduced to four.
In summary, the trapping effort consisted of twice-daily trap sessions with at least one capture, resulting in a total number of Apodemus flavicollis captured with at least one tick attached. For each captured rodent the number and life stage of feeding ticks was carefully assessed and registered, without removal [7], [21]. A total number of ticks were counted of which were larvae, were nymphs, and were adults. The number of ticks [nymphs] per rodent was between 1 [0] and 111 [15] with a median number of ticks per rodent equals to 8. Detailed data, on a yearly scale, are reported in Table 1, while the fraction of nymphs observed in different year and grids is reported in Table 2. In Figure 1 the number of captured Apodemus flavicollis per trapping session is shown for the whole nine year period and for different grids (from A to I).
In order to explore the impact of different parasite aggregation distributions on the spread of a TBEv-like pathogen where the main transmission route is through co-feeding, we performed extensive numerical simulations informed by the data about tick aggregation on mice. In this setting, tick larvae were not infective (transovaric transmission has been indicated as negligible [38]), adults only rarely feed on mice (on our data set adults ticks are about of the total number of ticks feeding on mice), and the only transmission link that we considered was the co-feeding between infective nymphs and larvae. Therefore, the only actors in our model were nymphs and larvae feeding on hosts. Moreover, Rosà and collaborators suggested in a recent work devoted to the same geographical area [21] that the larvae that feed in one year generally quest and feed as nymphs in the following year. Therefore, by adapting the Susceptible-Infected-Susceptible (SIS) model [39] to our purpose we assumed that nymphs are categorised as infective or not, that feeding larvae are susceptible and that some of them could eventually be infected by co-feeding with infective nymphs before moulting (thus becoming infective nymphs at time t+1). At each iteration t, with t being a discrete number between and and year, we assigned a number of ticks to each of the mice by drawing a sample from the considered distribution q. Then, on each mouse we said that of ticks feeding on it, were nymphs and the other larvae (with ). These nymphs were larvae in the previous year and were possibly infected. Then, defining as the prevalence among larvae after feeding at time , we assumed that the prevalence at time t among nymphs was . Thus, the number of infective nymphs on a mouse that at time t was parasitised by ticks was . Then, on each of the mice the co-feeding transmission between larvae and infective nymphs could occur with probability and we updated accordingly to the fraction of larvae infected (i.e. the fraction of infective nymphs at next time step). The following meta-code summarises the epidemiological dynamic
1. for t between and :
(a) for each mouse i, with i between 1 and
(b) is updated as the fraction of larvae infected
(c) if is equal to zero we stop the loop
It is worth stressing that in the previous meta-code we did not consider ticks recovering from the infection, since we assumed that a feeding infective nymph at time t will exit the infectious dynamics by moulting to the adult stage or dying.
We also modified the previous dynamics to deal with different distributions in tick aggregation as a function of seasonality. At each year t, we classified mice as observed during the mice peak activity ( = mice, with ) and observed out of the peak (). Therefore, we assigned the number of ticks feeding on mice according to the respectively aggregated distributions qIN and qOUT. Moreover, since the larvae obtaining a blood meal at year t will be nymphs at year without any other involvement in the epidemic spreading at year t, [21], these modifications to the meta-code are sufficient to suitably describe the seasonal variation in the epidemic process. More explicitly, the epidemic dynamic in the presence of seasonality in tick aggregation may be described by the following meta-code:
1. for t between and :
The probability distribution of tick burden on mice was skewed and showed a heavy tail. The best fit of the NB distribution was obtained on the largest available subsets of data, i.e. with , see left panel of Figure 3. In this setting, the MLE method estimated ( confidence intervals (CI) ) and (95%CI = ). However, the GOF of the NB distribution was very low for any value of , see central panel of Figure 3, thus giving evidence for rejecting the hypothesis of the NB functional form. Similarly, the best fit of PoiLN distribution was achieved on the largest subsets of data, (, see left panel of Figure 3). In this case the estimated parameters were (CI = ) and (CI = ). The GOF of the PoiLN, central panel of Figure 3, suggested that PoiLN was acceptable only for . However, for , the KS statistic displayed values that were too large to consider the PoiLN distribution appropriate for describing real data.
On the other hand, by fitting the tail of the distribution to a PL distribution, we found that the best fit was obtained for (with a standard deviation of 5.83), see left panel of Figure 3. This value is matched with an estimated scaling parameter (with standard deviation = 0.41). The GOF test (p-value larger than 0.1) suggested that the optimum PL fit on the tail of the distribution should not be ruled out, and that the result holds for every PL fit with see center panel of Figure 3. Finally, the LLR test highlighted that the PL fitting is to be preferred () to the NB in describing the tail of the distribution for a large range of lower bounds, , see right panel of Figure 3. Similarly, the PL is to be preferred to the PoiLN for . Moreover, it is worth to stress that for values above (55) the sign of the LLR test still indicates the PL fit as the preferred one compared to the NB (and PoiLN), although the indication loses statistical significance due to the scarcity of available data.
In Figure 4 we show the complementary cumulative probability distribution of the best fits resulting from for NB and PoiLN distributions and for PL distribution against field data of the number of ticks per mouse. From this plot we noticed that above a certain number of ticks per mouse NB [PoiLN] under-estimates [over-estimates] the tail of the distribution (indeed both fits were statistically evaluated as very poor). At the same time, in agreement with statistical results summarised in Figure 3, we noticed that the PL fit in Figure 4 more appropriately describes the right tail of the data distribution.
The number of mice captured in different years and grids showed strong seasonal patterns as reported in Figure 1. For each grid and each year we defined two separate periods depending on the mice abundance as defined in section “Data Analysis” and sketched in Figure 2. Imposing a threshold , for each year and grid we identified a time window of high mice abundance. With we found significant evidence that the distribution of ticks on mice within the abundance peak was different from that observed outside. Indeed, the fraction of the KS measures calculated on the synthetic samples lower than the real-data KS statistic was almost , thus indicating very low confidence in obtaining the same measurement by chance. The same statistical evidence was also obtained by using different time window thresholds (such as, and 0.6).
On the data sets classified as inside (IN) and outside (OUT) the time window of mice abundance peak, we fitted for different time-window lengths () the parameters and for NB distribution (Figure 5, left panels) and and for PL distribution (Figure 5, right panels). We observed a larger PL scaling parameter inside the mice abundance peak than outside (two-sample t-test output: for t-statistic = , df, ) indicating a larger heterogeneity in tick burden outside the abundance peak time. Moreover the GOF test indicated a rejection of the NB fit in both sets (IN and OUT) with . On the other hand, the GOF test with showed that the PL model cannot be ruled out in both sets (p-value>0.1) and the LLR test indicated that the PL fitting outperforms the NB model (p-value<0.05) in the estimates both inside and outside the peak time window.
The distribution of larvae and nymphs on mice are coincident rather than independent, and indeed the same most infested hosts feed both of the nymphs and of the larvae. Moreover, Spearman's correlation coefficient measured on the number of larvae and nymphs on mice was positive (0.24) and the probability that this coefficient was detected by chance was very low (the empirical value was the largest if compared to those evaluated in reshuffled samples). In addition, the mean number of larvae co-feeding with a nymph is about which is almost double the value that would be seen if the distributions were independent (mean equal to 12).
To start, we simulated the non-systemic disease spreading of a TBE-like pathogen with a fraction of nymphs among ticks equals to 2%, close to the one observed in our real data (cfr. Table 2), 5%, and 10%, as in literature [8], [40]. We consider the empirical distribution observed on the entire data set. We fixed the number of hosts to which, together with the considered distribution, resulted in a number of vectors pairs equal to . In our simulations, we explored the effects of β, the infection probability, on the observed prevalence at the final time step, , with . (We observed that was larger enough to allow the prevalence to converge toward an endemic pseudo-equilibrium or the disease-free equilibrium). For each we allowed simulations to run starting from an initial prevalence of . In Figure 6 we plotted the prevalences (median value, interquartile intervals and the CI) observed at equilibrium as a function of the transmission probabilities, β. Results showed that the larger the fraction of nymphs among ticks feeding on mice, the larger the probability of pathogen invasion and the infection prevalence.
Then, we explored the effects of different tick burden distributions on the spread of infection. To this end we considered four distributions: PL, NB, PoiLN, and the empirical distribution on the entire data set (aggregated on capture sessions and grids). For synthetic distributions we considered the actual observed distribution below the estimated , while we used the best fit of synthetic distributions to describe values greater than . Again, we fixed the number of hosts to . It is worth stressing that in the synthetic samples generated from these distributions we observed some features similar to those observed in real-data. For instance, the number of nymphs was positively associated with that of larvae and more particularly a nymph co-fed with a mean number of larvae similar to that observed in reality (for PL the mean number was 23, for NB 20, and for PoiLN 27).
Results, plotted in Figure 7 for and in Text S2 for and , corroborated the hypothesis that the transmission probability needed for the pathogen to become endemic is driven by the shape of the tail of the distributions. In particular, we noticed that for the PoiLN distribution (the one with larger fitted tail) the epidemic threshold is the lowest, while for the NB distribution (the one with smaller fitted tail) the infection probability needed for invasion is the highest. Not surprisingly, the PL, which has the best performances in fitting the tail of the empirical distribution, is the one for which the prevalences at equilibria better resemble those observed in simulations using the empirical distribution. We also performed some sensitivity analysis on parameter distributions, further highlighting that the larger the tail of the distribution, the lower the epidemic threshold (see Text S1). In addition, sensitivity analysis on the fraction of nymphs (f) showed that does not qualitatively influence the epidemic behaviour (see Text S2).
Furthermore, we investigated the effect of differences in the distribution of the tick burden as a function of the abundance of mice on the spreading of a non-systemic infectious disease. To this end, we fixed , as measured in the dataset, and as qIN we considered a PL with exponent as estimated with . In a similar way, we assumed as qOUT a PL distribution with exponent . For both qIN and qOUT we further set . Results are summarised in Figure 8, from which it could be inferred that the epidemic outcome was strongly influenced by the different distributions of feeding ticks according to mice abundance. We consistently observed that the transmission probability needed for the pathogen to effectively spread was smaller when the time windows identified by mice abundance are considered.
Tick aggregation on hosts is the result of several complex interactions of biotic and abiotic factors, such as host exposure and susceptibility to ticks, ticks' phenology and host behaviour, environmental factors, availability of resources, and others [27], [41]. Historically, the NB distribution has been preferred to the Poisson distribution to describe parasite heterogeneity across hosts because it suitably reproduces overdispersed observations. It has also been widely used in empirical [6], [25], [26], [28] and theoretical studies [23], [24], [42]. However, fat tailed distributions other than the NB one can also adequately reproduce tick aggregation, as shown by Elston et al. [29] and Bisanzio and collaborators [30].
Through the use of an extensive data set of feeding Ixodes ricinus ticks on mice, we showed that a PL distribution is better able to describe the right tail of the tick distribution on hosts than a NB or a PoiLN distribution (see Figure 3 and 4). This finding may have relevant epidemiological consequences, since it is well documented that the heterogeneity of contact distributions among individuals has large impacts on pathogen spread and persistence [43]–[49]. In fact, it has been demonstrated [50] that the minimum transmission probability for a pathogen to spread on a network, the so-called epidemic threshold, is driven by the first and the second moment of this distribution. In particular, Pastor-Satorras et al. [50] demonstrated that the larger the heterogeneity, the lower the epidemic threshold for the pathogen to spread, with an interesting behaviour in infinite size network showing a zero epidemic threshold [46]. Thus, the epidemiological inferences on the spread of a pathogen are highly influenced by the characterisation of the connectivity distribution and in particular by the distribution tail (i.e. the heterogeneity). Our results corroborate those findings and generalise them in a different framework and for more complex transmission routes, i.e. a vector-host network for non-systemically transmitted diseases. In particular, we found that the tail of the distribution of the number of ticks per rodent highly influences pathogen spreading (see Figure 7 and Text S1). Furthermore, it is worth remarking that although the tail of the distribution as defined here represents about of the entire data set, our simulation findings suggest that this small part of the distribution is crucial for pathogen invasion.
We also confirm that the probability of pathogen invasion and the infection prevalence are strongly influenced by the fraction of nymphs on the total feeding ticks on mice (Figure 6 and Text S2). The co-occurrence of larvae and nymphs on competent hosts is in fact essential for the horizontal transmission of non-systemic transmitted tick-borne pathogens, such as TBE, and it has been documented, both empirically and theoretically, that it could be a key factor in creating TBE hotspots, [51], [52].
Our conclusions confirm previous findings showing that the distribution of ticks on rodents may significantly affect the spread of infections [27], [30], [53], especially for non-viraemic transmitted diseases such as TBE [7], [23], [24]. Under the hypothesis of a NB distribution of ticks across hosts, both Rosà et al. [24] and Harrison and collaborators [23] showed that highly coincident and aggregated distributions favour the establishment of TBEv. However, highly heterogeneous degree distributions do not necessary imply a higher spread of disease. Indeed, Piccardi et al. [54] showed that scale-free networks can be much less efficient than homogeneous networks in favouring the disease spread in the case of a nonlinear force of infection.
The correct description of tick aggregation on hosts could dramatically affect disease control strategies: for instance, Perkins [7] emphasised that an optimised control effort targeted on highly parasitised mice, also identified as sexually mature males of high body mass, could significantly lower the transmission potential. On the other hand, Brunner and colleagues [27] observed that the identification of individuals which fed a disproportionate number of ticks (and that can therefore act as superspreaders) can be challenging, since simple covariates such as sex, age or mass do not entirely explain the differences in parasite burden.
In order to fully understand the different tick attachment behaviours on hosts, we identified different time windows related to rodent seasonal dynamics. Using this approach we found that the distribution of ticks on mice may vary across the season, with higher aggregation heterogeneity in periods of low rodent abundance and lower aggregation heterogeneity during the peak of host abundance (see Figure 5). We also showed that seasonal aggregation patterns, characterised by larger tails in time periods of low host abundance, enhance the spread of non-viraemic transmitted diseases (see Figure 8). Shaw and collaborators [26] observed significant variations in the degree of aggregation between host subsets – stratified by sex, age, space or time of sampling – in several host-parasite systems. In agreement with our results (lower aggregation in period of high mice abundances as shown by estimated exponents of PL), they found that aggregation in copepod (Lepeophtheirus pectoralis) infesting plaice (Pleuronectus platessa) decreases during summer months. They mainly ascribed the observed variation to significant differences in mean parasite burden among months. On the other hand, we did not find significant differences in tick burden inside and outside the window of high rodent abundance. Specifically, in the case of , the average number of ticks per host were inside the window of high rodent abundance and outside and the differences between inside and outside are not statistically significant (permutation tests, p>0.05). However, the second moment of the number of ticks per host drastically changed between high and low abundance periods, driving the difference in the aggregation distributions observed in the two time windows. Seasonal variations in resource availability and host abundance can have a significant effect on the space used by mice. Males and females tend to respond to these changes in different ways, since space use for females is driven largely by food availability, whereas the distribution of males is related primarily to mating opportunities. Yellow-necked mouse (A. flavicollis) females exhibited reduced spatial exclusivity and larger home ranges during lower food availability while males varied their spatial distribution accordingly by also expanding their home ranges [55]. An inverse relationship between population density and home range sizes has also been observed in wood mice (Apodemus sylvaticus) [56]. Consequently, in periods of low rodent abundance more mobile rodents, especially males, are more likely to hit a patch of larval ticks. As a result, these individuals would harbour a large amount of ticks and increase the aggregation of tick distribution among the rodent population. On the other hand, tick density is usually lower in periods of low rodent abundance, and the average tick burden would decrease for the rest of the population, especially females, balancing the overall tick burden. On the contrary, during times of high abundance mice move less and ticks would be distributed more evenly among the rodent population resulting in the observation of a lower aggregation in tick distribution during the peak of rodent abundance.
Our primary goal was to help understand the role of tick aggregation across mice on the spread of non-viraemic transmitted diseases through a simple and general transmission model. Other works – such as [24], [42], [52], [57] – described in very fine detail the transmission of vector-borne diseases, introducing different transmission routes, tick stages and alternative hosts in the epidemic model. For instance, Norman and colleagues [57] demonstrated through an epidemiological model that non-viraemic transmission could have non-negligible effects on the persistence of a disease like the Louping ill. Here, considering the non-systemic transmission only, we explored the effect of using different theoretical functional forms to describe the tick burden on hosts. By estimating parameters of the burden distributions on a very detailed data set, we defined a simple and transparent transmission model that explicitly takes into account the real contact pattern of vectors and hosts in the description of a non-systematically transmitted vector-borne disease. In this way we were able to emphasise that, while the NB and PoiLN models can sufficiently fit the whole real distribution, the PL model represents a better fit for the distribution tail. Furthermore, the vector perspective approach used in our model gives better insights into the dynamics of non-systemic transmitted pathogens respect to host perspective models that were more commonly and widely used in this context [24], [42], [52], [57]. In addition, epidemiological simulations parameterised by the fitted tick burden distributions highlighted the epidemiological consequences of describing tick aggregation on hosts trough distributions with different tails, showing that the shape of the tail distribution has a non-negligible influence on pathogen persistence. Future works will be devoted to extend the present findings to more complex transmission dynamics (e.g. including viraemic or transovaric transmission), in order to assess the effect of a PL decay of the distribution for a wider range of vector-borne diseases.
|
10.1371/journal.pgen.1005461 | Silencing of X-Linked MicroRNAs by Meiotic Sex Chromosome Inactivation | During the pachytene stage of meiosis in male mammals, the X and Y chromosomes are transcriptionally silenced by Meiotic Sex Chromosome Inactivation (MSCI). MSCI is conserved in therian mammals and is essential for normal male fertility. Transcriptomics approaches have demonstrated that in mice, most or all protein-coding genes on the X chromosome are subject to MSCI. However, it is unclear whether X-linked non-coding RNAs behave in a similar manner. The X chromosome is enriched in microRNA (miRNA) genes, with many exhibiting testis-biased expression. Importantly, high expression levels of X-linked miRNAs (X-miRNAs) have been reported in pachytene spermatocytes, indicating that these genes may escape MSCI, and perhaps play a role in the XY-silencing process. Here we use RNA FISH to examine X-miRNA expression in the male germ line. We find that, like protein-coding X-genes, X-miRNAs are expressed prior to prophase I and are thereafter silenced during pachynema. X-miRNA silencing does not occur in mouse models with defective MSCI. Furthermore, X-miRNAs are expressed at pachynema when present as autosomally integrated transgenes. Thus, we conclude that silencing of X-miRNAs during pachynema in wild type males is MSCI-dependent. Importantly, misexpression of X-miRNAs during pachynema causes spermatogenic defects. We propose that MSCI represents a chromosomal mechanism by which X-miRNAs, and other potential X-encoded repressors, can be silenced, thereby regulating genes with critical late spermatogenic functions.
| During male germ cell formation, the X and the Y chromosomes are inactivated. This process is conserved and it is essential for germ cell generation. It is believed that X/Y silencing affects all protein-coding genes, but the status of miRNAs and other non-coding genes needs further investigation. MicroRNAs from the X-chromosome (X-miRNAs) have been reported as potential silencing escapers, and they have been proposed to play a role in the inactivation mechanism itself. By looking at the individual cell level, we show unambiguously that X-miRNAs are subject to X/Y silencing, a finding that contradicts the current literature. Moreover, we generated mouse mutants in which we forced expression of X-miRNAs during X/Y silencing, and this lead to germ cell death. We propose that X/Y silencing can influence transcription of essential germ cell genes by regulating X-repressors.
| Meiotic sex chromosome inactivation (MSCI) describes the transcriptional silencing of the unsynapsed X and Y chromosomes at the onset of pachynema in mammalian male germ cells [1–5]. Inactivation of the sex chromosome results in the formation of a heterochromatic domain called the sex body [6]. MSCI is one example of a general mechanism, meiotic silencing, which inactivates any chromosome that is unsynapsed during male or female meiosis [7,8]. MSCI imposes a repressive chromatin signature on the X and Y chromosomes that is retained later, during spermiogenesis [9–12]. MSCI and its maintenance are regulated by a broad array of DNA double-strand break (DSB) repair proteins and chromatin modifications [2,5,13]. Male mice with chromosome abnormalities, e.g. XYY, or targeted mutations in meiotic synapsis or recombination genes, e.g. Spo11-/- and Brca1-/- frequently exhibit defective MSCI, and this results in misexpression of toxic sex-linked genes and midpachytene arrest [14–18].
Microarray [9,19], RNA-sequencing [20] and RNA FISH [16] studies have concluded that in mice MSCI is robust, and no example of an X-linked protein-coding gene that is actively transcribed during pachynema has yet been identified. This situation contrasts with that later in spermatogenesis, when expression of some X-linked genes from the repressed X chromosome is facilitated by various mechanisms including gene amplification [16] and establishment of active chromatin marks by the ubiquitin ligase RNF8 [21]. However, the activity of X-derived non-coding RNAs, and especially small RNAs, during pachynema is less well understood.
Interestingly, the X mouse chromosome is enriched in miRNA-encoding genes, and many of these are expressed in a testis-biased manner [22–24]. Song et al. conducted an extensive study of X- miRNA expression patterns, in which miRNA levels were evaluated by RT-qPCR in purified spermatogenic cell populations. 86% of X-linked miRNA transcripts were detected at high levels in pachytene spermatocytes, and it was been suggested that these genes escape MSCI [25]. High pachytene levels of X-miRNA transcripts have been confirmed by RT-qPCR [22,25], RNA-sequencing [23,26,27] and in situ hybridisation [25,27] approaches. Non-coding RNAs have a prominent role in gene silencing, e.g. in X chromosome inactivation [28], and repression of transposable elements and centromeric repeats [29], and it is therefore possible that X-linked miRNAs contribute to the process of MSCI itself. However, definitive proof that X-miRNA genes escape MSCI requires that nascent precursors of miRNAs, so-called pri-miRNAs, are generated during pachynema, and ideally, that these can be visualised as nascent transcripts originating from the otherwise inactive X chromosome, e.g. by techniques such as RNA FISH. We therefore sought to reappraise X-miRNA expression in the male germ line focusing on nascent transcripts.
In order to establish whether X-miRNAs are subject to MSCI, we examined their expression during mouse spermatogenesis using RNA FISH (Fig 1A–1G). We focused on spermatogonia, the early diploid germ cell progenitors in which the X chromosome is active, and pachytene spermatocytes, in which MSCI has taken place. There are currently 167 annotated miRNA genes on the X chromosome (source: miRBase version 21), the majority of which fall into clusters. We focused on six clusters, located at different sites on the X chromosome, and expressed in the testis (S1 Fig) [23]. Together these comprise 78 X-miRNAs, and 83% of them have been reported to escape MSCI [25]. We used a combination of antibody staining for the MSCI marker phosphorylated histone H2AFX (γH2AFX) [30], as well as DAPI nuclear staining, in order to accurately substage germ cells.
X-miRNA clusters 1 and 6 reside within introns of the genes Clcn5 and Ftx, respectively, and are transcribed from the same strand as the host genes (Fig 1A and 1F). At miRNA cluster 1 and cluster 6 loci, no putative promoter other than the host gene promoters can be detected upstream the miRNA genes, as assessed by H3K4me3 signal (S2 Fig), and expression of the miRNAs was shown to be dependent on the transcription of Clcn5 and Ftx parental RNAs [31,32]. Clcn5 and Ftx primary transcripts therefore represent the X-miRNA precursor transcripts. We used fluorescently-labelled, denatured fosmid DNA probes spanning the intronic Clcn5 and Ftx X-miRNA containing regions in order to detect X-miRNA precursor transcripts (X-pri-miRNA). Cluster 1 and 6 pri-miRNA FISH signals were observed in spermatogonia (53% and 75% expressing, n = 32 and 51 cells, respectively). However, no pri-miRNA expression could be detected in pachytene spermatocytes (0% expressing, n = 100 cells; Fig 1A, 1F and 1G).
Next, we used fosmid probes to examine expression of the remaining X-miRNA clusters 2, 3, 4 and 5. X-miRNAs located within these clusters do not lie within host genes. For all four clusters, we observed pri-miRNA FISH signals in spermatogonia (23%, 8%, 26% and 14% expressing, n = 43, 46, 42 and 41 cells, respectively; Fig 1B–1E and 1G). In contrast, RNA FISH signals were not observed in pachytene spermatocytes for any of the four clusters (0% expressing for each cluster, n = 100 cells each in each case; Fig 1B–1E and 1G).
The fosmids that we used for our RNA FISH experiments have an average size of 39kb. These probes will detect X-miRNA transcription, but could potentially also detect unannotated transcripts residing in the same locus. To exclude this possibility, we carried out two experiments. For cluster 5, we used recombineering to excise a 7kb segment containing the X-miRNA genes from the fosmid probe. When the resulting, modified fosmid was used for RNA FISH, no signals were observed in spermatogonia (0% displaying signals, n = 41 cells; S3 Fig). Secondly, we designed an RNA FISH protocol to assess transcription of specific X-pri-miRNAs. In this approach, we used ~40 nucleotide-long probes matching sequences present in the pri-miRNA, but not the pre-miRNA or the mature miRNA, at the base of the miRNA-containing stem-loop sequence (S4 Fig). We targeted the X-miRNA miR-465, present in six copies in cluster 4 (Fig 1D). Pri-miRNA signals were observed in spermatogonia but not in pachytene cells (0% expressing, n = 53 cells; S4 Fig). Thus, in conclusion, we observed transcription of all six X-miRNA clusters (total 78 X-miRNAs) in spermatogonia. However, we could not detect expression for any of these X-miRNAs during pachynema.
The absence of cluster 1 to 6 X-pri-miRNA FISH signals in pachytene spermatocytes suggests that these genes are subject to MSCI. To test this possibility, we repeated our RNA FISH analysis on a mouse model in which MSCI is defective. In Spo11 null male mice, a domain of γH2AFX is formed at pachynema, but this rarely encompasses the X and Y chromosomes, and it is therefore termed the “pseudo sex body” [15,33]. The failure to execute H2AFX phosphorylation on the XY bivalent causes misexpression of sex-linked genes during pachynema in this mutant [16]. We performed pri-miRNA FISH for four representative X-miRNA clusters: 1, 3, 4 and 6 (Fig 2). Pachytene spermatocytes were identified in Spo11 null males by the presence of the γH2AFX-labelled pseudo sex body.
Notably, in Spo11 null pachytene spermatocytes we observed pri-miRNA FISH signals for all four gene clusters studied. Expression was observed in 14%, 96%, 98% and 28% of pachytene cells for clusters 1, 3, 4 and 6, respectively (n = 100 cells in each case; Fig 2A). We subsequently repeated the analysis of cluster 4 X-miRNAs using our oligonucleotide RNA FISH approach that specifically detects pri-miRNAs for the six copies of miR-465. We observed FISH pri-miRNA signals in 93% of Spo11 null pachytene cells (n = 69 cells; Fig 2B). In addition, we performed miR-465-specific pri-miRNA FISH in a second MSCI mutant, the Brca1 Δ11/Δ11 model [17,18]. We detected misexpression of this miRNA in 87% of pachytene spermatocytes. (n = 15 cells; Fig 2B). We conclude that defective MSCI leads to X-miRNA misexpression during pachynema.
Although MSCI is defective in most Spo11-/- and Brca1 Δ11/Δ11 pachytene cells, in both models domains of γH2AFX are occasionally seen covering sub-regions of the X chromosome [16,17]. We predicted that in these rare spermatocytes, X-miRNAs encompassed within γH2AFX regions should be normally silenced. This proved to be the case: in the few Spo11 null and Brca1 pachytene cells in which the miR-465 locus, identified using DNA FISH, lay within a γH2AFX domain (Spo11 null: n = 2 out of 54 cells; Brca1Δ11/Δ11: n = 2 out of 15 cells), no pri-miR-465 expression could be observed (Fig 2C). Thus, the expression status of these X-miRNAs is tightly linked to the presence of the meiotic silencing marker γH2AFX. We conclude that in wild type pachytene spermatocytes, the X-linked miRNAs studied herein are silenced during pachynema as a result of MSCI.
To corroborate our X-miRNA FISH data, we next compared expression levels of individual pri-miRNAs in wild type and Spo11 null sibling testes by RT-qPCR at 15.5days post-partum (dpp; Fig 2D). At this age, most spermatocytes are in pachynema, and genes subject to MSCI are expected to be overexpressed in Spo11 null relative to wild type males. We examined transcript levels for a number of X-linked and autosomal pri-miRNAs, and expressed these as a Spo11 null / wild type ratio. Experiments were performed in triplicate, each time using a different Spo11 null and wild type sibling. In each case, Spo11 null / wild type ratios for autosomal miRNAs averaged ca. 1, indicating no difference in pachytene expression levels between the two genotypes (Fig 2D). Conversely, the ratio for X-linked pri-miRNAs significantly exceeded one (p = 3.361e-08; Fig 2D), thereby confirming that X-linked miRNAs are upregulated in the absence of MSCI.
Finally, we used transgenesis to further investigate whether silencing of X-miRNAs in pachynema is due to MSCI. Previous experiments have demonstrated that X-genes present as transgenes on autosomes continue to be expressed during pachynema [34]. This is because unlike the X chromosome, autosomes are synapsed during pachynema and therefore escape the effects of meiotic silencing. To establish whether pachytene silencing of X-linked miRNA genes was due to their location on the X chromosome, we generated a single copy transgenic line in which X-linked miRNA gene clusters 3 and 4 were located together on an autosome by random BAC integration (X-miRBAC line 1; Fig 3A). We chose a BAC that includes a region of local H3K4me3 enrichment upstream of the miRNA gene cluster, indicative of a putative promoter (Fig 3A). Using pri-miRNA microarrays, we confirmed that cluster 3 / 4 X-miRNAs were overexpressed in X-miRBAC line 1 testes relative to non-transgenic siblings (Fig 3B). We then performed pri-miRNA FISH in pachytene spermatocytes from X-miRBAC line 1 transgenics using BAC probes covering clusters 3 and 4. We observed expression of cluster 3/4 miRNAs from both the X chromosome and the autosomal transgene prior to pachynema (Fig 3C). However, during pachynema, while the X-located 3/4 miRNAs were silenced, those located on the transgene continued to express (100%, n = 50). We observed the same results using our miR-465-specific pri-miRNA FISH protocol on X-miRBAC line 1 (S5 Fig) (100%, n = 9). Thus, silencing of X-integrated miRNAs during pachynema is due to MSCI.
Defects in MSCI cause pachytene arrest, due to misexpression of toxic sex-linked genes, e.g. the Y chromosome genes Zfy1 and Zfy2 [14]. Our analyses indicated that X-miRNAs are subject to MSCI. We therefore wondered whether misexpression of these genes during pachynema would give rise to spermatogenic defects. Interestingly, in our X-miRBAC line 1 males, which carry the autosomally-integrated single copy X-miRNA 3 and 4 cluster transgene, we observed reduced testis weights relative to non-transgenic brothers from as early as five weeks post-partum (Fig 4A). Importantly, histological and TUNEL analysis of X-miRBAC line 1 testis sections revealed spermatogenic defects, principally germ cell apoptosis at stage IV, corresponding to midpachynema, and stage XII, corresponding to the meiotic divisions (Fig 4B, S7 Fig).
In order to exclude the possibility that the spermatogenic defects observed in X-mirBAC line 1 males resulted from a transgene integration effects, we subsequently generated two more cluster 3/4 X-miRNA autosomal transgenic lines, with three (X-miRBAC line 2) and eleven (X-miRBAC line 3) transgene copies (S6 Fig). X-miRBAC line 2 showed predominant apoptosis at stage XII (Fig 4C, S7 Fig), while X-miRBAC line 3 exhibited marked apoptosis at mid and late pachynema (Fig 4D, S7 Fig). We conclude that inappropriate expression of X-miRNAs from the X-linked clusters 3 and 4 at pachynema induces spermatogenic defects.
MSCI is a robust silencing process, affecting most or all protein-coding genes on the mouse X chromosome. However, it is unclear whether silencing also affects X-linked miRNAs. Here, using RNA FISH and other transcriptional assays, we find that X-linked miRNA genes are expressed before prophase I but are silent during pachynema. We therefore conclude that X-miRNAs are subject to MSCI.
It is important to highlight that we did not study all miRNAs on the X chromosome. It is therefore formally possible that X-miRNAs omitted in our analyses behave differently with respect to MSCI. We find this unlikely, because we chose miRNAs from multiple, distinct clusters on the X chromosome, and we included many X-miRNAs that were previously reported to escape silencing [25]. A priori, one could also argue that our inability to detect X-miRNA FISH signals during pachynema is because our RNA FISH experiments lack the sensitivity required to detect gene expression during this stage of prophase I, rather than because these genes are subject to MSCI. We doubt that this is the case, because we were able to detect expression of these species during pachynema both in MSCI mutants, and in mice carrying autosomally-located X-miRNA transgenes. Finally, our conclusion that X-miRNAs are subject to MSCI was corroborated by RT-qPCR analysis in MSCI mutants versus controls. Taken together, our data support a model in which X-miRNAs behave like X-linked protein-coding genes with respect to X silencing.
How can our findings accommodate earlier work? Several independent reports have documented high levels of X-miRNA expression in pachytene spermatocytes [22,23,25,26], and there can be little doubt that this contrasts with the generally low level expression detected for protein-coding X-genes. However, in most existing studies X-miRNAs were assayed at the level of mature miRNAs. Notably, with the exception of some rare cases [35,36], miRNAs have an unusually long half-life, on average 5 days, which exceeds that of protein-coding RNAs by ten-fold [37]. The abundant expression of miRNAs during pachynema might therefore be due to their high transcript stability, rather than ongoing generation of nascent miRNA precursors.
Interestingly, our work shows not only that X-miRNAs are subject to MSCI, but also that failure to silence them can result in spermatogenic defects which are manifest in the case of the cluster 3/4 X-miRNAs as arrest predominantly at stages IV and XII. Thus, X-miRNAs join the Y-encoded Zfy1/2 genes as being male “pachytene-lethal” genes. Our findings show that MSCI must be extensive, silencing genes not only on the Y chromosome but also on the X chromosome. Given that miRNAs act as gene repressors, the phenotypes resulting from their ongoing expression in our miRNA transgenic males are presumably due to inappropriate target downregulation as a consequence of miRNA overexpression, or to an inability to appropriately upregulate target genes with meiotic and/or post-meiotic functions. In this model, MSCI could function as a chromosome-based mechanism for regulating expression of repressors. From a broader perspective, the large-scale silencing of genes across the X chromosome by MSCI is likely to influence myriad transcriptomic networks within germ cells. As a consequence, MSCI could regulate multiple facets of the mammalian germ cell development program.
All animal procedures were in accordance with the United Kingdom Animal Scientific Procedures Act 1986 and were subject to local ethical review.
All mice were maintained on an MF1 background. The miRBAC transgenic lines were produced by microinjection of purified BAC BMQ-333E20 into fertilized eggs from CBA/Ca x C57Bl/10 F1s. Spo11 null and Brca1Δ11/Δ11 mice have been described previously [18,38].
Browser tracks were generated in R using the Givz package. The UCSC mouse genome assembly mm10 was used as a basis for analysis. Annotation of transcripts was obtained from the UCSC knownGene database. Genomic coordinates of fosmid and BAC probes were obtained from the CHORI BACPAC resource center.
A small RNA testis library was downloaded from GEO under accession identifier GSE40499 (Meunier et al.). Adapter sequences were removed from the reads (ATCTCGTATGCCGTCTTCTGCTTG), and 15 to 23nt-long reads were selected for analysis. Reads were aligned to the mouse mm10 genome using Bowtie (version 1.6.0) with the parameters -m 50—best—strata -v 2. MiRNA gene coordinates were obtained from miRBAse (version 21). MiRNA duplicates sharing a copy on the X chromosome and a copy on an autosome were removed from the analysis. MiRNA read counts were generated in R using the QuasR package as documented in the reference manual [39]. Read counts are expressed as read counts per million reads mapping to miRNA genes.
RNA and DNA FISH was carried out with digoxigenin- and biotin-labelled probes respectively, using fosmid and BAC genomic clones (cluster 1: WI1-603H11; cluster 2: WI1-1995I23; cluster 3: WI1-1646F11; cluster 4: WI1-2045C16; cluster 5: WI1-2828J23; cluster 6: WI1-2859G17; miRBAC: BMQ-333E20). The technique was described previously [40]. For miR-465-specific pri-miRNA FISH, a mix of nine amino-allyl-modified oligonucleotides labeled with fluorolink Cy3 were used as probes (S1 Table). We used the anti-γH2AFX antibody (Upstate, 16–193; dilution 1/100) for immunofluorescence post-RNA FISH.
Total RNA was extracted from frozen testis tissues with Trizol (Invitrogen), treated with DNAse I (Invitrogen), and reverse transcribed with random hexamers (Invitrogen) and Superscript II (Invitrogen) according to the manufacturer’s instructions. Quantitative PCR was performed with pre-designed Taqman pri-miRNA and U6 snRNA Taqman assays according to the manufacturer's instructions. Relative expression was calculated with the ΔCt method using U6 as a normaliser. Pri-miRNA expression ratio for individual pri-miRNAs are provided in S2 Table.
A library of H3K4me3 ChIP in adult germline precursor cells and corresponding input controls were downloaded from GEO under accession identifier GSE49624 [41]. Reads were aligned to the mouse mm10 genome using Bowtie (version 1.6.0) with the parameters Bowtie -m 1—best –strata.
Affymetrix Mouse miRNA 2.0 microarrays were performed to measure miRNA expression in testis of three wild type and two homozygous transgenic siblings at 37dpp. Total RNA was extracted with Trizol (Invitrogen), treated with DNAse I (Invitrogen), and column-purified (Ambion). Microarray hybridizations were performed according to the manufacturer's instructions. Microarray signal intensity was extracted and normalized using Affymetrix' miRNA QC Tool using default parameters. Statistical analyses were performed using R and the limma package. Fold-changes and FDR-adjusted p-values were computed by fitting a linear model for each microRNA. Standard errors were smoothed using empirical Bayes (eBayes function of the limma package).
The transgene copy number of the miRNA BAC transgenic line was estimated by qPCR on genomic DNA, with a technique adapted from the one described in. The data was normalised with Atr PCR for ΔCt calculations, and quantification of Jun copy number was used as a quality control. Primer sequences are provided in S3 Table.
Recombination-mediated genetic engineering of fosmid WI1-2828J23 was performed to delete a 7kb fragment encompassing miRNA genes miR-201 and miR-547 using standard procedures. The primers used for recombineering are
|
10.1371/journal.ppat.1002360 | Sap Transporter Mediated Import and Subsequent Degradation of Antimicrobial Peptides in Haemophilus | Antimicrobial peptides (AMPs) contribute to host innate immune defense and are a critical component to control bacterial infection. Nontypeable Haemophilus influenzae (NTHI) is a commensal inhabitant of the human nasopharyngeal mucosa, yet is commonly associated with opportunistic infections of the upper and lower respiratory tracts. An important aspect of NTHI virulence is the ability to avert bactericidal effects of host-derived antimicrobial peptides (AMPs). The Sap (sensitivity to antimicrobial peptides) ABC transporter equips NTHI to resist AMPs, although the mechanism of this resistance has remained undefined. We previously determined that the periplasmic binding protein SapA bound AMPs and was required for NTHI virulence in vivo. We now demonstrate, by antibody-mediated neutralization of AMP in vivo, that SapA functions to directly counter AMP lethality during NTHI infection. We hypothesized that SapA would deliver AMPs to the Sap inner membrane complex for transport into the bacterial cytoplasm. We observed that AMPs localize to the bacterial cytoplasm of the parental NTHI strain and were susceptible to cytoplasmic peptidase activity. In striking contrast, AMPs accumulated in the periplasm of bacteria lacking a functional Sap permease complex. These data support a mechanism of Sap mediated import of AMPs, a novel strategy to reduce periplasmic and inner membrane accumulation of these host defense peptides.
| The opportunistic pathogen Haemophilus influenzae is a normal inhabitant of the human nasopharynx, and is commonly implicated in respiratory tract infections, particularly of the middle ear (otitis media), sinuses, and lung (pneumonia, chronic obstructive pulmonary disease and cystic fibrosis). We have identified a multifunctional bacterial uptake system that is required for critical mechanisms of bacterial survival in the host. This Sap transporter system recognizes and transports host immune defense molecules and is involved in uptake of an iron-containing nutrient (heme) that is host-limited, yet required for bacterial growth and survival. We propose that bacteria utilize this, and likely other similar transport systems, for numerous functions that are important for bacterial survival in the host, including host immune evasion and metabolism. Our findings significantly advance our understanding of how single bacterial protein systems co-operate and coordinate multiple functions to equip bacteria to survive and cause disease in the hostile host environment. Our long-range goal is to block this uptake system thereby starving the bacterium of essential nutrients and also promoting clearance by the host immune response. Removal of this important bacterial survival mechanism will thwart the ability for Haemophilus to survive as a pathogen and thus decrease the incidence of disease development.
| Host-derived antimicrobial peptides (AMPs) are typically amphipathic, cationic innate immune defense molecules that target bacterial membranes, disrupt transmembrane potential and trigger cytoplasmic leakage resulting in bacterial cell death [1], [2]. Defensins (α- and β-) and cathelicidin (hCAP-18/LL37) molecules are primarily abundant in neutrophils (α-defensins and cathelicidin), respiratory epithelium (β-defensins 1–3 and cathelicidin), and are secreted by lung and trachea epithelia (cathelicidin) [3]–[6]. As a first line of innate defense, AMPs serve to limit bacterial colonization of mucosal surfaces [7]–[11]. Bacteria therefore adapt to resist AMP lethality through a series of countermeasures: remodeling the bacterial outer membrane surface to dampen charge and alter hydrophobicity [1], [12]–[14], export of AMPs via multiple transferable resistance (MTR)-mediated efflux pumps [15], secretion of exoproteases for AMP degradation [16], secretion of bacterial molecules to suppress host innate defense [17], [18], and release of proteins that function to adsorb extracellular AMPs [19].
Nontypeable Haemophilus influenzae (NTHI) is a commensal of the human nasopharnyx, yet causes opportunistic diseases such as conjunctivitis, sinusitis, exacerbations of chronic obstructive pulmonary disease, complications of cystic fibrosis and chronic and acute otitis media [20]–[25]. During the transition from a commensal to pathogen, NTHI must acquire nutrients and defend against host innate immune defense strategies including increased production of AMPs in response to infection. NTHI outer membrane remodeling provides a first line of defense against cationic AMPs. Lysenko and colleagues demonstrated that the presence of phosphorylcholine, a phase variable modification of NTHI lipooligosaccharide, alters outer membrane hydrophobicity that confers resistance to the cathelicidin LL-37 [26]. Additionally, HtrB is required for hexaacylation of NTHI lipid A, thus mutants lacking htrB are unable to fully acylate their lipid A rendering NTHI susceptible to AMP mediated killing [27].
NTHI lack the other described resistance mechanisms such as AMP efflux or exoprotease activity. The sap (sensitivity to antimicrobial peptides) operon encodes an inner membrane ABC-transporter, previously shown to play a crucial role in defense against AMPs [28]–[36]. Previously, we demonstrated that NTHI SapA, the periplasmic substrate binding protein of the Sap transporter, binds AMPs [37]. NTHI strains deficient in either SapA or the SapD ATPase are susceptible to killing by recombinant chinchilla β-defensin-1, an orthologue of human β-defensin-3, sharing 77% amino acid identity [37], [38]. Moreover, SapA and SapD are required for virulence in a mammalian host [37], [38]. Expression of the sap operon is up-regulated in vivo during NTHI-induced otitis media and in response to AMP exposure in vitro [37], [38]. The mechanism by which the Sap transporter complex confers AMP resistance remains unknown.
Here, we demonstrated that NTHI lacking the ligand binding protein SapA were directly susceptible to AMP exposure in the mammalian host, as neutralization of cBD-1 in vivo reversed attenuation and clearance of the sapA mutant strain. Further, we describe a novel mechanism of AMP import into the bacterial cytoplasm, which is dependent upon Sap transporter function. The Sap-permease was required for cytoplasmic localization and in a Sap-permease deletion strain, AMPs accumulated in the bacterial periplasm. Since AMPs were susceptible to intracytoplasmic peptidase activity, we hypothesize that AMP import, coupled with cytoplasmic degradation, decreases AMP accumulation in the bacterial periplasm and protects the bacterium from subsequent perturbation of the bacterial cytoplasmic membrane, a counter strategy to evade innate immune defense and ultimately benefit the bacterium nutritionally.
Previously, we demonstrated that the SapA periplasmic binding protein is required for NTHI to persist and cause disease in the middle ear of a mammalian model of otitis media [38]. In this co-infection model of middle ear disease (equal amounts of the wild type NTHI strain 86-028NP and the isogenic sapA mutant strain were introduced directly into the middle ear) we observed significant attenuation of the sapA mutant strain, which was unable to compete for survival and cleared from the middle ear. We subsequently demonstrated that recombinant SapA binds the AMP molecule, chinchilla β-defensin 1 (cBD-1) in vitro [37], and further, binds numerous AMPs, including the human β-defensin 3 (hBD-3) and cathelicidin (LL37) peptides [39]. These data support a mechanism whereby SapA confers protection from AMP lethality in vivo during disease progression. To directly assess this hypothesis, we sought to neutralize activity of cBD-1 in vivo and monitor the consequence of this treatment on survival of the SapA-deficient strain during a co-infection model of otitis media. We predicted that SapA was required by NTHI to counter the defensin bactericidal activity in vivo, and that neutralization of cBD-1 would reverse attenuation and clearance of the SapA-deficient strain in this environment.
Altered expression of AMPs can impact the ability of bacteria to colonize a host [11], [40]. We previously demonstrated that β-defensin expression controls NTHI bacterial colonization of the nasopharyngeal mucosal surface in a chinchilla model of NTHI-mediated disease [41]. Neutralization of available native cBD-1 via passive inhalation of affinity-purified antibody assessed the direct contribution of AMPs in mediating NTHI colonization levels. Here, we further developed and utilized this methodology to assess the direct contribution of AMPs in mediating attenuation of sapA mutant infection in the middle ear. We delivered affinity-purified anti-recombinant cBD-1 [(r)cBD-1] polyclonal antibody (or pre-immune serum as a negative control) to the chinchilla middle ear cavity (n = 5 per cohort) via the superior bullae, to inhibit the activity of native cBD-1. Twenty minutes after pre-treatment, we challenged chinchillas transbullarly with NTHI in a co-infection model, then measured the relative concentration of wild type and sapA mutant present in middle ear effusions over a 14 day period. A second dose of neutralizing antibody (or pre-immune serum) was delivered 24 hours after bacterial challenge.
We demonstrated that animals receiving anti-(r)cBD-1 neutralizing antibody remained colonized with the sapA mutant strain (Figure 1, panel B), which was sufficiently able to compete (in many cases out compete) with the wild type strain as shown in competitive index calculations (Figure 1, panel C). In contrast, the SapA-deficient strain is attenuated in animals receiving pre-immune serum (Figure 1, panel A). Consistent with our previous findings, we demonstrated that the sapA mutant was unable to persist over the course of infection (Figure 1, Panel A), and was significantly impaired (p<0.01) in its ability to infect the middle ear compared to that of the wild type strain 14 days after infection. Additionally, as previously described, anti-(r)cBD-1 antibody or the control pre-immune serum was not bactericidal against NTHI at concentrations used for AMP neutralization [41]. These data demonstrate that SapA functions to counter AMP lethality during NTHI infection in vivo. Since SapA binds and delivers periplasmic substrate to the inner membrane Sap transporter complex, we were interested to determine whether AMP molecules are transported to the cytoplasm of intact cells and whether this transport was dependent upon Sap permease activity.
We predicted that AMP substrates, bound by the SapA periplasmic binding protein, would be transported into the bacterial cytoplasm. Thus, we sought to obtain cytoplasm-enriched fractions of NTHI by bacterial fractionation and monitor the presence of AMPs in these fractions following exposure to sub-lethal concentrations of hBD3 or LL37. Traditional methods previously used to fractionate Escherichia coli were not successful with the prototypic NTHI clinical isolate strain, 86-028NP. Therefore, we utilized membrane destabilization and differential centrifugation (see Materials and Methods) and obtained both periplasm and cytoplasm-enriched fractions from lysed NTHI cells. The protein profiles of these two fractions were unique when visualized by two dimensional gel electrophoresis (Figure 2A) and SDS-PAGE analysis (Figure 2B). We confirmed fractionation by immunodetection of the periplasmic enzyme β-lactamase and the cytoplasmic protein SapD (Figure 2C). β-lactamase was localized to periplasm-enriched fractions and the cytoplasmic protein SapD was only detected in the cytoplasm-enriched fraction (not present in the periplasm). Importantly, we demonstrated that hBD-3 and LL37 co-fractionated with cytoplasm-enriched fractions of NTHI, demonstrating that AMPs, even at sublethal concentrations, gained access to the bacterial cytoplasm of intact NTHI (Figure 2D).
We hypothesized that AMP substrates bound by SapA are delivered to the inner membrane Sap transporter complex for energy dependent transport into the bacterial cytoplasm. Indeed, we demonstrated that AMPs localized to the bacterial cytoplasm of intact cells within 30 minutes of exposure to sublethal AMP concentrations. Thus, we sought to characterize the kinetics of localization to the cytoplasm. NTHI were incubated with AMPs for 30 minutes (pulse period), cells were washed to removed unbound AMPs and then incubated in buffer alone (without AMPs) for 0, 2, or 4 hours (chase period). Incubation of AMP-exposed NTHI in buffer, during this chase period, did not affect NTHI viability during the time course (data not shown). At each time point, cytoplasm-enriched fractions were monitored for the presence of AMP molecules by immunoblot (Figure 3). We observed maximal import of hBD3 at the earliest time point tested (Figure 3A and 3C). Interestingly we observed a decrease in the detection of hBD3 over the remainder of the chase period (from 0 to 4 hours, Figure 3A and 3C). Import of hBD-3 was dependent upon SapA production as hBD-3 did not localize to the cytoplasm in a SapA-deficient strain (data not shown). Although similarly detected in the bacterial cytoplasm following the 30 minute pulse period (0 hr), LL-37 peptide appeared to accumulate during the first 2 hours of the chase period and decrease only slightly when incubated an additional 2 hours (4 hr, Figure 3A and 3C). These data suggest differential kinetics of AMP accumulation within NTHI, likely related to AMP structure and charge. We were unable to follow the fate of LL37 beyond the 4 hour chase period due to loss of NTHI viability.
The absence of known MTR in NTHI [15], combined with our observation of decreasing AMP levels in the cytoplasm, led us to hypothesize that the AMPs are degraded in the cytoplasm. To determine susceptibility of AMPs to peptidase activity, we incubated cytoplasm-enriched fractions of NTHI with hBD-3 or LL37 for 0, 2, 5, or 13 hours in the presence or absence of a protease inhibitor cocktail mix (−/+ inhibitor , Figure 4). At each time point, samples were collected and monitored for AMP by immunoblot (Figure 4). Similar to our observations in intact cells, we observed nearly complete loss of hBD-3 detection within 5 hours whereas LL37 loss was not observed to the same extent. After overnight incubation (T13) however, LL-37 detection was dramatically reduced. In both cases, AMPs were susceptible to cytoplasmic fraction peptidase activity that was blocked in the presence of protease inhibitors (Figure 4, + inhibitor). Collectively, these data suggest that AMP molecules are susceptible to cytoplasmic protease activity, and are likely beneficial to NTHI for metabolic purposes.
The Sap transporter belongs to a family of ABC transporters that recognize and transport substrates that are typically small and cationic in nature [29]. We previously demonstrated that AMPs can displace additional substrates previously bound by SapA, and that mutations in Sap transporter proteins alter NTHI susceptibility to AMP exposure [37]–[39]. We hypothesized that the Sap transporter may mediate the transfer of AMPs across the cytoplasmic membrane, as a mechanism of AMP transport into the cell. We therefore generated a deletion mutant lacking the permease components of the Sap transporter (SapB and SapC) and determined the subcellular localization of LL-37 and hBD3. NTHI (parent and sapBC permease mutant) were exposed to sub-lethal concentrations of either LL-37 or hBD3 for 30 minutes and processed to assess AMP localization by transmission electron microscopy, to visualize AMP localization in intact cells. We observed that both LL-37 (Figure 5, top panels) and hBD3 (Figure 5, bottom panels) localized to the periplasm and the cytoplasm of the parental cells. Accumulation of AMPs was not observed in either the bacterial outer or inner membranes under these conditions. In contrast, neither LL-37 nor hBD3 were observed in the cytoplasm of the sapBC permease-deficient cells, yet a striking accumulation of AMPs was observed in the periplasm and cytoplasmic membrane. These data suggest that subcellular localization of AMP molecules to the bacterial cytoplasm is dependent upon Sap-mediated transport. Consistent with this observation, sapBC permease-deficient cells are more sensitive to AMP exposure (data not shown). This targeted transport of AMPs, to the cytoplasm for degradation, may serve to reduce periplasmic accumulation of AMPs, and thus protect the cytoplasmic membrane from AMP association and lethality. Thus, Sap-mediated transport of AMP to the cytoplasm for degradation serves as a newly described subversion mechanism that would provide a nutritional benefit to the bacterium thus coupling innate immune resistance to metabolic activity.
The ability of NTHI to persist in both the commensal and pathogenic lifestyles requires mechanisms to avert the host innate immune response. The epithelium in the upper airway produces AMPs that assist in the control of NTHI [41]–[43]. This led us to investigate the mechanisms used by NTHI to evade killing by AMPs. Modifications of the outer membrane primarily serve as a first line of defense against AMP lethality including addition of phosphorylcholine to lipooligosaccharide to repel positively charged AMPs [26] and lipid A acylation to increase NTHI hydrophobicity, thus altering membrane permeability [27]. Although effective as a first line defense strategy, AMP accumulation and disruption of the outer membrane, periplasmic localization and permeabilization of the cytoplasmic membrane is microbicidal. Many bacteria possess additional mechanisms to detect and resist AMP insult shown to be regulated through two-component regulatory signaling [44], [45], release of extracellular peptidases [16], secreted proteins to bind AMPs [19], and AMP efflux pumps [15], [46], [47]. NTHI however, lacks two-component regulatory systems that sense AMP exposure, and additional mechanisms of AMP resistance have not been identified.
We previously demonstrated that the Sap transporter conferred NTHI resistance to AMPs. The periplasmic binding protein, SapA, binds AMPs [37], contributes to AMP resistance [38] and is also upregulated in response to AMP exposure during otitis media. This, taken with our data presented here, suggests that resistance to AMP lethality, conferred via Sap mediated transport of AMPs, is important for NTHI pathogenesis. Since we observed marked attenuation of Sap transporter mutants in vivo [37], [38], we predicted that neutralization of AMPs would rescue this attenuated phenotype, supporting a critical role for Sap-dependent AMP import in NTHI pathogenesis. In support of this, in vivo neutralization of the hBD3 orthologue, chinchilla β-defensin-1, which is highly expressed in the chinchilla upper respiratory tract [48], results in increased colonization by NTHI [41]. These data suggest that AMPs play an important role in limiting infection in a mammalian host. Indeed, alterations in AMP production can affect the ability of other microorganisms to colonize a host [11], [40], [49]. The ability of NTHI to exploit mechanisms to resist AMP lethality may equip transition to that of opportunistic pathogen and ability to cause disease. We demonstrated here that neutralization of β-defensin activity in the middle ear of a mammalian model of otitis media rescues the attenuated phenotype of the SapA-deficient strain, such that the mutant strain is no longer cleared from this environment (Figure 1). This is the first direct evidence that SapA contributes to bacterial virulence by conferring protection from AMP lethality in vivo. Having established this, we monitored the consequence of SapA-AMP interaction in intact bacterial cells. We describe a novel mechanism of inner membrane transport of AMP molecules, via the Sap ABC transporter, to the bacterial cytoplasm resulting in AMP degradation (Figure 6). This direct mechanism of AMP influx serves to decrease both periplasmic and inner membrane accumulation of AMPs and protect NTHI from AMP bactericidal effects. Current models predict that AMP-induced microbicidal activity results from transmembrane pore formation subsequent to AMP accumulation at the cell surface [1]. However, we observed by TEM analysis, that AMPs localized to the bacterial cytoplasm, even when NTHI were exposed to sublethal concentrations that do not appear to accumulate in the bacterial membranes, suggesting a mechanism whereby bacteria can directly import AMP molecules. Since accumulation of AMPs in the bacterial cytoplasm would likely be detrimental, we hypothesized that AMPs are subsequently targeted for cytoplasmic degradation. Indeed, we observed that hBD3, localized to the bacterial cytoplasm, was susceptible to cytoplasmic degradation, as the ability to detect hBD3 decreased over time (Figure 3A, C). Although we did not observe an appreciable decrease in LL-37 detection following uptake into intact cells, we demonstrated that LL37, like hBD-3, was susceptible to peptidase activity in cytoplasm-enriched fractions (Figure 4). These data suggest differential kinetics of uptake and degradation, likely dependent upon AMP structure and charge. In support of differential kinetic uptake of AMP molecules, we previously demonstrated that LL-37 and hBD3 differ in their abilities to bind and displace the iron-containing compound heme, also shown to bind SapA [39]. Consistent with these studies, cytoplasmic accumulation of LL-37 appeared delayed, relative to hBD3 (Figure 3C). Although we have not, as of yet, identified the binding sites for these SapA ligands, our data indicate that hBD3, which is highly cationic, may preferentially bind SapA based upon charge (unpublished). Additionally, it has been shown that the cathelicidin LL-37 contains proline-rich sequences that resist degradation by serine proteases [50] and NTHI lack PgtE and OmpT homologues, shown to mediate degradation of alpha-helical peptides [51]–[54], which may explain our observation that LL-37 appears more resistant to proteolysis within the bacterial cytoplasm. NTHI encodes homologues of bacterial cytoplasmic proteases, ClpX and Lon proteases, shown to contribute to AMP resistance in other microorganisms [55], [56]. The contribution of these proteases to AMP degradation in NTHI is currently unknown, but remains the focus of future work.
An alternative hypothesis is that an MTR drug efflux pump functions to export AMPs that have accumulated in the bacterial inner membrane and in the cytoplasm. It was recently shown that the MtrC periplasmic membrane fusion protein conferred resistance to LL-37 and β-defensins in Haemophilus ducreyi [47]. Interestingly, the H. ducreyi Sap transporter does not mediate β-defensin resistance [36], [47]. Although NTHI strain 86-028NP contains week homologs of the MTR system, our evidence suggests that this system is not functional or does not function to confer resistance to defensin or cathelicidin peptides, a well described function of the NTHI Sap transporter [37]–[39]. Further, we demonstrated that in vivo neutralization of native cBD-1 was sufficient to rescue competitive growth of the sapA mutant, suggesting that SapA was sufficient to mediate resistance to this defensin molecule in the host. Finally, a functional MTR system would be expected to export AMP molecules thus avoiding periplasmic, membrane and cytoplasmic accumulation of AMPs. We observed, in the absence of a functional Sap permease complex, AMP accumulation (Figure 5) and increased sensitivity to AMP-mediated killing (data not shown). We would expect, based upon these published reports that a functional MTR system would not result in this phenotype.
Development of antibiotic resistant strains secondary to conventional antibiotic use has prompted interest in the use of AMPs as therapeutic alternatives to treat bacterial infections [57]. The use of AMPs in therapeutic regimens may be hindered due to bacterial subversion mechanisms of AMP lethality [54]. Since the Sap transporter functions to protect against the accumulation of AMPs in the bacterial periplasm, and subsequent interaction with the cytoplasmic membrane, a mechanism to block transporter function may provide novel therapeutic supplements for NTHI infections. Delivery of small molecule inhibitors to block substrate binding and transport would confound an important AMP resistance mechanism in NTHI, rendering NTHI susceptible to innate immune attack while preserving the normal flora that are often disrupted by conventional antibiotic use.
The described novel mechanism of AMP import and subsequent degradation expands our understanding of host-pathogen interactions, particularly those that mediate resistance to key components of the host innate immune response. Future studies to identify AMP specific intracellular peptidases, AMP degradation products, and metabolic consequences of targeted peptide degradation, are needed to better understand bacterial survival strategies in the hostile host environment. Further, a better understanding of bacterial mechanisms used to transition from commensal to that of an opportunistic pathogen will better equip the design of therapeutics to combat disease.
All animal experiments were carried out in strict accordance with the accredited conditions in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee (Welfare Assurance Number A3544-01) at The Research Institute at Nationwide Children's Hospital, AR08-00027.
All experimental procedures were performed under xylazine and ketamine anesthesia, and all efforts were made to minimize suffering.
Healthy adult chinchillas (Chinchilla lanigera), purchased from Rauscher's Chinchilla Ranch (LaRue, OH), were used for these studies, after allowing them to acclimate to the vivarium for 7 to 10 days. Chinchillas were anesthetized with xylazine (2 mg/kg, Fort Dodge Animal Health, Fort Dodge, IA) and ketamine (10 mg/kg), Phoenix Scientific Inc., St. Joseph, MO), and nasopharyngeal lavage fluids were obtained by passive inhalation of 500 µl of pyrogen-free sterile saline into one naris with recovery of lavage fluid from the contralateral naris as liquid was exhaled. Middle ear fluids were recovered via epitympanic tap through the superior bullae, and directly obtained from the inferior bullae behind the tympanic membrane. All animal experiments were performed using accredited conditions for animal welfare approved by the Institutional Animal Care and Use Committee (Welfare Assurance Number A3544-01) at The Research Institute at Nationwide Children's Hospital, AR08-00027.
Nontypeable Haemophilus influenzae strain 86-028NP is a minimally passaged clinical isolate obtained at Nationwide Children's Hospital in Columbus, Ohio. This prototypic wild type strain has been sequenced and extensively characterized in chinchilla models of otitis media [58], [59]. The parental NTHI strain 86-028NP:: rpsL is a streptomycin resistant strain constructed as previously described [60]. The sapBC permease deletion mutant was constructed as an unmarked, non-polar deletion mutation of the sapB and sapC genes by recombineering strategy as previously described [39]. Bacterial strains were grown on chocolate II agar (Becton Dickinson, Sparks, MD) or in brain heart infusion broth supplemented with 2 µg heme/mL and 1 µg NAD/mL (sBHI). Bacteria were cultured from overnight growth on chocolate II agar, resuspended in sBHI to OD490 = 0.65 (equivalent to 1×108 CFU/ml), diluted 1:5 into fresh sBHI medium and grown to mid-log phase (3 hours) at 37°C, 5% CO2, static.
In vivo neutralization of chinchilla β-defensin 1 (cBD-1) was performed as previously described with the following modifications [41]. We prepared and purified (r)cBD-1 based upon previously published methods [41], [48]. A HiTrap protein G HP column (GE healthcare, Pittsburgh, PA) was used to affinity purify total IgG from rabbit anti-(r)cBD-1 and the cognate pre-immune serum. One milliliter of serum was dialyzed [3.5 kDa molecular weight cutoff (MWCO), EMD Chemicals Inc., San Diego, CA] at 4°C against 20 mM sodium phosphate buffer, pH 7.0 and applied to the affinity column. Immunoglobulins were eluted from the affinity matrix with 0.1 M glycine-HCl, pH 2.7 in one milliliter fractions into eppendorf tubes that contained 200 µl of 1.0 M Tris-HCl, pH 9.0, to neutralize the acidic elution conditions. Samples that contained the greatest amount of protein (fractions 1 and 2) were pooled and dialyzed overnight at 4°C against sterile saline. Protein concentrations of the anti-(r)cBD-1 and the pre-immune serum were determined using Coomassie Plus Protein Assay Reagent (Thermo Scientific, Rockford, IL). We confirmed that anti-(r)cBD-1 antiserum, and not the pre-immune antiserum, bound purified (r)cBD-1 by immunoblot (data not shown). Five adult chinchillas were administered 90 µg anti-cBD-1 or pre-immune rabbit immunoglobulin, in a total volume of 200 µl. Antisera were administered by passive inhalation of droplets of the solution delivered to the nares or direct transbullar inoculation of the middle ears of anesthetized chinchillas. Animals were then placed in a prone position for 20 minutes prior to intranasal challenge with approximately 5.0×107 cfu NTHI wild type strain 86-028NP mixed with 5.0×107 cfu sapA mutant (co infection) in a 100 µl volume or transbullarly with 2.5×103 cfu NTHI wild type strain 86-028NP mixed with 2.5×103 cfu sapA mutant in a 200 µl volume. Nasopharyngeal lavage and epitympanic taps were performed 1, 2, 4, 7, 10, 14 days after NTHI co-challenge, and bacterial counts were determined by dilution plating of bacteria on chocolate agar.
Parental NTHI strain 86-028NP::rpsL was grown to mid-log phase in brain heart infusion broth supplemented with 2 µg heme/ml and 2 µg NAD/ml (sBHI). Cells were pelleted by centrifugation and resuspended in 1X phosphate buffered saline containing 2 mg Polymyxin B sulfate (Sigma-Aldrich, St. Louis, MO)/ml, 0.05% glycerol, to bind lipooligosacharide and generate spheroplasts, thereby releasing periplasmic contents to the supernatant. Following incubation for 1 hour at 37°C on a rotating platform (100 rpm), spheroplasts were pelleted by centrifugation and resuspended in 1 ml 10 mM HEPES pH 7.4 (Sigma-Aldrich, St. Louis, MO), 0.05% glycerol and subsequently lysed by freeze-thaw method (10 cycles, −78°C to 37°C). Lysed spheroplasts were then incubated at room temperature, on an orbital rocker, with 1 µl Benzonase Nuclease (Novagen, Darmstadt, Germany) and 5 µl 1 M MgCl2 (Ambicon, Alamo, CA)) to a final concentration of 5 µM. Bacterial membranes were pelleted by ultracentrifugation (40,000 rpm for 1 hour, 4°C) and cytoplasmic proteins were removed in the supernatant. Cytoplasm and periplasm-enriched fractions were observed by two dimensional gel electrophoresis (Bio-Rad, Hercules, CA) of enriched fractions, stained with SYPRO Ruby Protein Gel Stain (Invitrogen, Carlsbad, CA) and imaged at 535 nm (Kodak Image Station 2000, New York, NY). To ensure successful NTHI fractionation, cytoplasm and periplasm-enriched fractions were analyzed by silver stain (Pierce, Rockford, IL) following SDS-PAGE separation. In addition, NTHI were exposed to a sub-lethal concentration of 15 µg ampicillin/mL to induce expression and trafficking of β-lactamase to the NTHI periplasm. Periplasm and cytoplasm-enriched fractions were separated by SDS-PAGE and transferred to nitrocellulose (Bio-Rad, Hercules, CA). Membranes were then blocked in 3% nonfat dried milk and probed for the periplasmic protein, β-lactamase, by incubating with mouse anti- β-lactamase for 45 minutes or the cytoplasmic protein, SapD, by incubating with rabbit anti-SapD overnight. Membranes were then washed, incubated with goat anti-mouse IgG (H+L) HRP conjugate (Invitrogen, Carlsbad, CA) for 30 minutes or goat anti-rabbit IgG (H+L) HRP conjugate (Invitrogen, Carlsbad, CA) for 1 hour, washed, and peroxidase activity was detected using Amersham ECL Western Blotting Detection Reagents (GE Healthcare, Little Chalfont, Buckinghamshire, UK).
Parental NTHI strain 86-028NP::rpsL were grown to mid-log phase in sBHI, normalized for cell number, and pelleted. Pellets were resuspended in 10 mM sodium phosphate buffer pH 7.4, supplemented with 2% sBHI. Cultures were incubated with sublethal concentrations of LL-37 or hBD3 as previously described, for a pulse period of 30 minutes. Cells were then pelleted by centrifugation, washed, and resuspended in 10 mM sodium phosphate buffer pH 7.4, supplemented with 2% sBHI. Cultures were incubated at 37°C, 5% CO2, static for a chase period of 0, 2, or 4 hours. Samples were transfered to ice and fractionated to obtain periplasm and cytoplasm-enriched fractions as described above. Enriched fractions were lyophilized (Labconco, Kansas City, MO) overnight and subsequently resuspended in 10 mM HEPES pH 7.4, 0.05% glycerol, then normalized for equal amounts of protein using Coomassie Plus Protein Assay Reagent (Thermo Scientific, Rockford, IL) prior to separation by SDS-PAGE (16.5% Ready Gel Tris-Tricine Precast Gels, Bio-Rad, Hercules, CA). Equal detection of NTHI lipopetide [61] by silver stain (Pierce, Rockford, IL) confirmed normalization of samples. Samples containing LL-37 or hBD3 were transferred to PVDF (Bio-Rad, Hercules, CA) or nitrocellulose (Bio-Rad, Hercules, CA), and blocked in 3% skim milk or 5% BSA, respectively. Membranes were incubated with rabbit anti-LL-37 (Phoenix Pharmaceuticals, Burlingame, CA) or goat anti-hBD3 (Leinco Technologies, St. Louis, MO) overnight at 4°C, washed, and incubated with goat anti-rabbit IgG (H+L) HRP conjugate (Invitrogen, Carlsbad, CA) or rabbit anti-goat IgG (H+L) HRP conjugate (Invitrogen, Carlsbad, CA), respectively. Membranes were washed and peroxidase activity was detected using SuperSignal West Femto Chemiluminescent Substrate (Thermo Scientific, Rockford, IL).
NTHI cells were fractionated as described above. Cytoplasm-enriched fractions were concentrated in Amicon centrifuge filters (3.0kDa MWCO) and normalized for equal amounts of protein using Coomassie Plus Protein Assay Reagent (Thermo Scientific, Rockford, IL). Cytoplasmic peptidase activity was monitored by incubating 8 µg of cytoplasm-enriched fractions with LL37 (40 ng) or hBD-3 (20 ng) in siliconized glass vials and incubated at 37°C for 2, 5.5 and 13 hours. In parallel, a protease inhibitor cocktail set was included (Calbiochem, LaJolla, CA). Following incubation, samples were separated on a SDS-PAGE gel (16.5% Ready Gel Tris-Tricine Precast Gels, Bio-Rad, Hercules, CA). Samples containing LL-37 or hBD3 were transferred to PVDF (Bio-Rad, Hercules, CA) or nitrocellulose (Bio-Rad, Hercules, CA), and blocked in 3% skim milk or 5% BSA, respectively. Membranes were incubated with rabbit anti-LL-37 (Phoenix Pharmaceuticals, Burlingame, CA) or goat anti-hBD3 (Leinco Technologies, St. Louis, MO) overnight at 4°C, washed, and incubated with goat anti-rabbit IgG (H+L) HRP conjugate (Invitrogen, Carlsbad, CA) or rabbit anti-goat IgG (H+L) HRP conjugate (Invitrogen, Carlsbad, CA), respectively. Membranes were washed and peroxidase activity was detected using SuperSignal West Femto Chemiluminescent Substrate (Thermo Scientific, Rockford, IL).
Parental strains NTHI 86-028NP::rpsL and the isogenic sapBC mutant strain were grown to mid-log phase in sBHI, normalized for cell number, and pelleted. Pellets were resuspended in 10 mM sodium phosphate buffer pH 7.4, supplemented with 2% sBHI. Human LL-37 cathelicidin (Phoenix Pharmaceuticals, Burlingame, CA) or human beta defensin 3 (PeproTech, Rocky Hill, NJ) were added to cultures at a final concentration of 0.25 µg LL-37/ml or 1 µg hBD3/ml. Cultures containing AMPs or cells alone were incubated at 37°C, 5% CO2, static for 30 minutes. Samples were transfered to ice and cells were pelleted by centrifugation, washed in 100 mM PIPES pH 7.0, and then resuspended in 100 mM PIPES. For immunolocalization of AMPs at the ultrastructural level, bacteria were fixed in 4% paraformaldehyde/0.05% glutaraldehyde (Polysciences Inc., Warrington, PA) in 100 mM PIPES/0.5 mM MgCl2, pH 7.2 (or PBS) for 1 hr at 4°C. Samples were then embedded in 10% gelatin and infiltrated overnight with 2.3 M sucrose/20% polyvinyl pyrrolidone in PIPES/MgCl2 at 4°C. Samples were trimmed, frozen in liquid nitrogen, and sectioned with a Leica Ultracut UCT cryo-ultramicrotome (Leica Microsystems Inc., Bannockburn, IL). 50 nm sections were blocked with 5% FBS/5% NGS for 30 min and subsequently incubated with rabbit anti-hBD3 (Leinco Technologies, Inc.) or rabbit anti-CAP-18 (LL-37) antibody for 1 hr at room temperature. Sections were then washed in block buffer and probed with 18 nm colloidal gold-conjugated anti-rabbit IgG (H+L) (Jackson ImmunoResearch Laboratories, Inc., West Grove, PA) for 1 hr at room temperature. Sections were washed in PIPES buffer followed by a water rinse, and stained with 0.3% uranyl acetate/2% methyl cellulose. Samples were viewed with a JEOL 1200EX transmission electron microscope (JEOL USA Inc., Peabody, MA). All labeling experiments were conducted in parallel with controls omitting the primary antibody. These controls were consistently negative at the concentration of colloidal gold conjugated secondary antibodies used in these studies.
Proteins discussed in this manuscript are listed followed by their corresponding UniProtKB (Universal Protein Knowledgebase) number. These included: SapA (Q4QL73_HAE18), SapB (Q4QL74_HAE18), SapC (Q4QL75_HAE18), SapD (Q4QL76_HAE18), SapF (Q4QL77_HAE18), OmpT (OMPT_ECOLI), PgtE (PGTE_SALTY), ClpX (Q4QMK4_ HAE18), Lon (Q4QN81_ HAE18), HtrB (Q4QKN9_ HAE18), β-Lactamase (A720C2_ HAE18).
|
10.1371/journal.ppat.1000999 | A Metazoan/Plant-like Capping Enzyme and Cap Modified Nucleotides in the Unicellular Eukaryote Trichomonas vaginalis | The cap structure of eukaryotic messenger RNAs is initially elaborated through three enzymatic reactions: hydrolysis of the 5′-triphosphate, transfer of guanosine through a 5′-5′ triphosphate linkage and N7-methylation of the guanine cap. Three distinctive enzymes catalyze each reaction in various microbial eukaryotes, whereas the first two enzymes are fused into a single polypeptide in metazoans and plants. In addition to the guanosine cap, adjacent nucleotides are 2′-O-ribose methylated in metazoa and plants, but not in yeast. Analyses of various cap structures have suggested a linear phylogenetic trend of complexity. These findings have led to a model in which plants and metazoa evolved a two-component capping apparatus and modification of adjacent nucleotides while many microbial eukaryotes maintained the three-component system and did not develop modification of adjacent nucleotides. Here, we have characterized a bifunctional capping enzyme in the divergent microbial eukaryote Trichomonas vaginalis using biochemical and phylogenetic analyses. This unicellular parasite was found to harbor a metazoan/plant-like capping apparatus that is represented by a two-domain polypeptide containing a C-terminus guanylyltransferase and a cysteinyl phosphatase triphosphatase, distinct from its counterpart in other microbial eukaryotes. In addition, T. vaginalis mRNAs contain a cap 1 structure represented by m7GpppAmpUp or m7GpppCmpUp; a feature typical of metazoan and plant mRNAs but absent in yeast mRNAs. Phylogenetic and biochemical analyses of the origin of the T. vaginalis capping enzyme suggests a complex evolutionary model where differential gene loss and/or acquisition occurred in the development of the RNA capping apparatus and cap modified nucleotides during eukaryote diversification.
| The protozoan parasite Trichomonas vaginalis is the cause of the most common non-viral sexually transmitted disease worldwide. Evolutionary analyses place Trichomonas in a super group called the Excavata, which includes the kinetoplastids and is highly divergent from fungi, metazoa and plants. Despite the vast evolutionary distances that separate these different eukaryotic lineages, a simplified view of eukaryotic evolution based on the complexity of nucleotide modifications at the 5′ end of mRNAs and the distribution of different types of enzymatic apparatus that confer these modifications has been proposed. Our analyses of the T. vaginalis capping enzyme challenges this view and provides the first example of a two-component capping apparatus typically found in metazoa and plants in a protozoan. The 5′-end nucleotide structure of T. vaginalis mRNAs is also shown to contain additional modified nucleotides, similar to that observed for metazoan and plant mRNAs and unlike that found in most eukaryotic microbes and fungi. Evolutionary analyses of the T. vaginalis capping enzyme indicates that this multicellular type capping apparatus may have come into existence earlier than previously thought.
| The 5′ cap is a unique feature of eukaryotic messenger RNAs (mRNA) and eukaryotic viruses not found on eubacterial and archaeal RNAs [1]. The addition of a m7G cap structure, or the cap 0 nucleotide, occurs co-transcriptionally via three consecutive reactions executed by the capping enzymatic apparatus: (i) hydrolysis of the 5′-triphosphate of nascent pre-mRNAs to a diphosphate by RNA 5′ triphosphatase (TPase), (ii) capping of the diphosphate end with GMP by the RNA guanylyltransferase (GTase) and (iii) methylation of the GpppN cap by RNA guanine-7- methyltransferase (MTase). The RNA cap is involved in multiple cellular functions including splicing, nucleocytoplasmic export, mRNA turnover and translation initiation [2], [3].
In addition to the N7-methylated guanosine cap, the 1st and 2nd adjacent nucleotides may be 2′O-ribose methylated after transcription [4] forming cap 1 and cap 2 structures, respectively. In contrast to the cap 0, the role of cap 1 and cap 2 modified nucleotides is unclear. Their presence may reveal a phylogenetic trend with increasing levels of complexity in multicellular eukaryotes [4], [5]. mRNAs in multiple species of yeast have been shown to contain only a cap 0 whereas the adjacent 5′ nucleotides of most multicellular eukaryotic mRNAs are further modified to form cap 1 and/or cap 2 structures [4], [6]. Thus far trypanosomes are the only protists to have their mRNA cap structures examined and these were found to contain a hypermodified cap 4 structure m7Gpppm26AmpAmpCmpm3Um [7]. The unconventional method of providing mRNAs with cap modification via trans-splicing of a splice leader RNA at their 5′ ends confers this unique cap structure to trypanosome mRNAs [8], [9]. Recently, enzymes involved in cap 4 formation have been identified [10]–[12], although the role of these different cap nucleotide modifications remains to be elucidated [11], [13].
The N7-methylated guanosine cap (cap 0) is a structural consensus in all eukaryotic mRNAs unlike the different levels of modification of subsequent nucleotides observed in different organisms. The type and structural organization of the guanosine cap enzymatic apparatus that confers the cap 0 to the 5′end of mRNAs has been examined in a wide variety of eukaryotes [14]–[18], including the divergent protists Giardia [19] and trypanosomes [20]–[22]. An interesting evolutionary scenario has emerged [1], which contradicts [23] current models derived from multiple eukaryotic phylogenetic analyses [24]. This scenario dictates a multicellular (eg. metazoa and plants) versus a microbial eukaryotic pattern (e.g. fungi and microbial eukaryotes). The majority of investigated microbial eukaryotic species possess a three-component capping system (TPase, GTase and MTase) while metazoan and plants encode a two-component system with the fusion of TPase and GTase polypeptides and a separate MTase [1].
In contrast to structural conservation of GTases, the TPases appear to have evolved from different protein ancestors [1]. TPases found in multicellular eukaryotes contain the cysteinyl-phosphatase superfamily motif HCXXXXXR(S/T) named phosphate-binding loop or ‘p-loop’ (TPasePL). This enzyme catalyses a two-step phosphoryl transfer in which the conserved cysteine attacks the γ-phosphorus of the 5′-triphosphate on the nascent RNA to form a covalent protein-cysteinyl-S-phosphate intermediate producing a 5′-diphosphate RNA product [17], [25]. The enzyme-bound phosphate is then hydrolyzed and liberated as inorganic phosphate. In contrast, all investigated microbial eukaryotic TPases to date share structural organization with metal-dependent phosphohydrolases (TPaseMDP) and have different structural configurations and enzymatic characteristics [26]. A current model of the mRNA capping system evolution in eukaryotes suggests that a ‘transitional state organism’ would contain both types of TPases [1]. The fusion of the TPasePL-GTase was followed by a secondary loss or complete divergence of the TPaseMDP, and the last common ancestor (LCA) of plants and metazoans would carry the fused TPasePL-GTase version only. Although this model is supported by the distribution of these enzymes in several microbial eukaryotes [1], and is consistent with eukaryote phylogenies based on few genes supporting a metazoan-plant relationship [23], it is incongruent with phylogenomic and multiple single gene phylogenies [24], [27]. However this mRNA capping centric view of eukaryote phylogeny does not preclude the occurrence of differential gene loss/gain during eukaryotic evolution.
Here, we have characterized the guanosine cap enzymatic apparatus of Trichomonas vaginalis, a divergent microbial eukaryote, that is a member of the Parabasalia and the super-group Excavata [24], [28]. In contrast to microbial eukaryotes, including other members of the Excavata, the diplomonad G. lamblia [19] and kinetoplastids [20], [22] T. vaginalis has a dual TPasePL-GTase capping enzyme (TvCE) resembling the metazoan-plant type, as previously predicted from genome analyses [29]. Moreover, we have demonstrated that T. vaginalis mRNAs contain a complex cap structure with a canonical m7G and adjacent modified nucleotides. Phylogenetic analyses of the GTase domain only and TPasePL-GTase alignments are consistent with a common origin of the T. vaginalis and metazoan-plant enzymes, which suggest that the TPasePL-GTase system is likely to be more ancient then previously thought and that complex scenarios of independent gene loss and/or gain events across various eukaryotic lineages may have taken place.
We have demonstrated previously that T. vaginalis mRNAs have a 5′-end protection that can be removed by pyrophosphatase treatment, and that these mRNAs partially precipitate with a monoclonal anti-trimethylguanosine antibody [30]. Therefore, similar to other eukaryotes this divergent protist must harbor an enzymatic capping apparatus. The T. vaginalis genome database (www.trichdb.org) [31] was screened using BLASTp analysis using available homologs for capping enzymes found in protists, yeast, plants and metazoans. As we previously reported based on analyses of the T. vaginalis genome [29] the same putative T. vaginalis capping enzyme gene (locus tag TVAG_187730, RefSeq accession XP_001327945.1, named here TvCE) is identified in searches conducted using the GTase enzyme found in many microbial eukaryotes and those using the fused TPasePL-GTase from plants and metazoans. No genes were identified in searches using TPaseMDP sequences. Both TPasePL-GTase functional domains in TvCE were conserved relative to the human capping enzyme in protein domain analyses as those from other metazoans, plants, green algae and a choanoflagellate, the latter being a member of the Choanomonada: microbial eukaryotes closely related to metazoan [24], [28] (Figure 1 & Figure S1). TvCE shares 30% identity and 47% similarity to the human capping enzyme and also has its TPasePL-domain fused to the N-terminus of a GTase domain (Figure S1). Strikingly, the TvCE cysteinyl-phosphatase superfamily motif HCXXXXXR(S/T) is 100% identical to the human sequence (Figure 1A). Furthermore, TvCE contained the six conserved peptide motifs (I, III, IIIa, IV, V and VI) at the C-terminus that comprise the active site for GTP binding and nucleotidyl transfer of GTase [32] (Figure 1B). Individual BLASTp searches with either the entire TvCE protein, the N-terminal domain encompassing the TPasePL domain (residues 1–258) or the two C-terminal domains (GTase, residues 259–441 and 442–561) identified by protein domain analyses, recovered as top hits animal sequences (see Data S1). These data indicate that T. vaginalis may have a two-component capping system similar to metazoan and plants.
To examine the enzymatic activity of recombinant TvCE we attempted to express either the full-length protein or each domain separately. We found that the single GTase domain was insoluble and that the TPase, although soluble, was inactive. Therefore, the full-length recombinant TvCE was used to analyse enzyme activity. The capping enzyme TPase specifically hydrolyses the γ-phosphate from the 5′-terminus of RNAs allowing its activity to be monitored by either the production of inorganic phosphate or ATP-ADP conversion. The TPase activity of recombinant TvCE was tested in the absence of metals at varying pH and the protein was found to be active in pH ranging from 4.5 to 6.5 (Figure 2A). Addition of 1–5 mM MgCl2 inhibited TvCE TPase activity and addition of EDTA could reverse this inhibition (Figure 2B). Activity in an acidic pH range and inhibition by MgCl2 are typical of a cysteinyl-phosphatase TPase and are not observed in classic microbial eukaryotic TPases [1]. In mammalian cysteinyl-phosphatase TPase, a transient phosphocysteine-enzyme intermediate can be trapped using a short incubation time at low temperature and acidic pH [17]. We asked whether TvCE also exhibits this property and found that the phospho-TvCE intermediate was detected in the presence of [γ32P]ATP, but not in the presence of [α32P]ATP as predicted, and within a restricted acidic pH (3.5–4.0) (Figure 2C). Phospholabeling of TvCE was lost after treatment with iodine but not hydroxylamine, supporting the presence of a thiophosphate linkage predicted to be at the C126 within the ‘P-loop’ (Figure 2D, left panel). To determine whether C126 is involved in this linkage, TvCE was subjected to specific amino acid mutations and tested for the ability to form a thiophosphate linkage. As shown in Figure 2D (right panel), serine substitution of either C126 or H125 abrogates formation of phospho-TvCE. When enzymatic affinity was compared across substrate concentration, we observed that P-loop mutants (H125S and C126S) and the double mutant (C126S and R526A) were inert to release γ-phosphate from ATP (Figure 3A), which is in agreement with the inability of these mutants to form a phosphocysteine-enzyme intermediate (Figure 2D). However, a mutation in the GTase domain alone (R526A) did not significantly affect TPase activity (Figure 3A). The utilization of ATP and GTP by TvCE as a function of nucleotide concentration is similar (Figure 3B). A relatively low conversion of the substrate is observed which may indicate that TvCE has either a low turnover rate or is partially inactivated during the 2-step purification. Together, these data indicate that C126 and H125 coordinate the cleavage and release of the γ-phosphate and demonstrate that TvCE has characteristics of a typical cysteinyl-phosphatase TPase [1].
The capping enzyme GTase uses a ‘ping-pong’ reaction mechanism for nucleotidyl transfer through a covalent enzyme-(lysyl-N)-GMP intermediate [33]. This allows GTase activity to be detected by 32P transfer from [α-32P]GTP to the enzyme [14], [16], [18], [34]. To determine whether TvCE also uses this reaction mechanism, we incubated recombinant TvCE with [α-32P]GTP in the presence or absence of divalent cations. The formation of the SDS-stable 32P-labeled enzyme was then evaluated by SDS-PAGE (Figure 4). TvCE GTase activity was detected by [α-32P]GTP labeling in a broad pH window (data not shown) and required the presence of either MnCl2 or MgCl2. Calcium did not support activity (Figure 4A). Metal dependence, specifically Mn2+ and Mg2+, is a typical characteristic of an RNA capping GTase.
To determine the specificity of TvCE GTase, competition reactions were performed using increasing concentrations of cold NTPs. The recombinant TvCE GTase displayed an absolute specificity for a GTP substrate since ATP, CTP or UTP were not found to inhibit this enzyme (Figure 4B). Guanylation of TvCE is dependent on nucleotide concentration (Figure 4C). In accordance with previous reported structure/function studies [18], the single arginine substitution at GTase motif VI (R526A) abolished TvCE activity (Figure 4C). Interestingly, we also found that mutations in the TPase domain (H125S and C126S) significantly affected GTase activity, suggesting that these residues may exert a cis-structural effect on the GTase domain (Figure 4C). Using size exclusion chromatography, we found both TPase and GTase activities in a single discrete peak corresponding to ∼71 kDa indicating that TvCE is monomeric (Figure S2).
In addition to using nucleotide substrates to assess the activity of TvCE we have also evaluated its ability to transfer GMP from GTP to the 5′ triphosphate end of an in vitro transcribed RNA, using the full-length enzyme with both active domains. As shown in Figure 5, TvCE is capable of labeling an RNA substrate provided this molecule harbors a 5′ triphosphate (lane 2). When the substrate is dephosphorylated by alkaline phosphatase digestion prior to incubation with TvCE, no transfer of [α-32P]GTP to the substrate is achieved (lane 3). In order for the GTase to transfer [α-32P]GMP from [α-32P]GTP to the 5′-end of the RNA substrate, the γ-phosphate must be first removed by the TPase (lane 4). When TPase or GTase activity is dependent on a TvCE mutant that is inactive in the corresponding domain, no labeling of the substrate is observed (lanes 5 & 8). On the other hand, when TPase or GTase activity is dependent on a TvCE mutant that inactivates the opposite domain, labeling of the substrate is achieved (lanes 6 & 7). These data are consistent with those characterizing individual domains using nucleotide substrates and TvCE mutants (Figure 3A and 4C) and demonstrate that capping of an RNA substrate strictly depends on the activity of both domains (Figure 5).
The observed similarities between metazoa and T. vaginalis RNA capping enzymatic activities prompted us to investigate the mRNA cap structure in this organism. The nucleotides adjacent to the m7G-cap structure are methylated to different extents in eukaryotes. Metazoans can have 2′-O-ribose methylations of the first and second transcribed nucleotides forming the cap structure m7GpppNmpNmpNp, where the first transcribed nucleotide is an adenosine. However, yeast is found to have no modifications beyond the 5′ m7G cap [4], [6]. We developed a multi-step protocol to purify mRNA suitable for structural analysis of nucleotides (see Materials and Methods). None of the steps alone, including consecutive passages over poly-dT chromatography columns, was sufficient to remove abundant RNA species such as tRNA and rRNA (Figure S3). As these contaminants contain hypermodified nucleotides in relatively high abundance that interfered with our analysis, we found it necessary to also immunoprecipitate RNAs with anti-TMG (which cross-reacts with the 7-methyl guanosine cap of mRNAs) to remove uncapped RNA contaminants (Figure S3). As a result, in vivo 32P-labeled mRNA obtained by this purification protocol was shown to be free of contaminating rRNA and tRNA by gel electrophoresis (data not shown) and structural analysis of nucleotides after P1 digestion (Figure 6A and Figure S3). The abundance of adenosine and uridine, relative to guanosine and cytosine, in this heterogeneous mRNA population (Figure 6A) is consistent with the reported ∼65% AT content of T. vaginalis genes [29]. Digestion of the purified mRNA sample with Tobacco Acid Pyrophosphatase (TAP), which specifically hydrolyzes the phosphoric acid anhydride bonds in the triphosphate bridge of a 5′-end cap structure, released m7GMP (Figure 6A). The identity of this modified nucleotide was confirmed by demonstrating that it can be converted to m2,2,7GMP using Schizosaccharomyces pombe trimethylguanosine synthase or SpTgs (Figure 6A) [35], [36].
Next, a sample of unlabeled mRNA was used for analysis of the first transcribed nucleotide (position +1). RNAs were subjected to TAP or a TAP-mock treatment, alkaline phosphatase digestion and then 5′-end labeled using T4 polynucleotide kinase. These were then digested completely by nuclease P1, and nucleotides were resolved on 2D-TLC plates. Appearance of distinct spots presence only in the TAP treated sample should reveal whether the first transcribed mRNA nucleotide (position +1), which is protected by an m7G cap, is modified (Figure 6B). Our result demonstrated that T. vaginalis mRNAs have a typical metazoan 2′-O-ribose methylated cap 1 nucleotide. The cap 1 nucleotide is either an adenosine (80%) or a cytosine (20%) based on a comparison of the Am/Cm ratio (Figure 6B). This is in agreement with ∼75% of T. vaginalis protein-coding genes being preceded by the conserved initiator element (Inr) TCAT/CT/A that dictates transcription initiation at the underlined adenosine [29], [37]. The four unmodified nucleotides (A, C, G and U) observed in both samples likely result from partially degraded RNAs that would not require TAP treatment to be 5′-end labeled or the presence of intact RNA contaminants that lack a cap.
There is a strong bias for a uridine at positions +2 and +3 in most T. vaginalis mRNAs [37], [38]. To demonstrate whether the uridine +2 is 2′-O-ribose methylated forming a cap 2 structure, in vivo labeled mRNA was digested with RNase T2 prior to anti-TMG precipitation, the last step of the mRNA purification protocol (Figure 7A). 2′-O-ribose methylation of this nucleotide would render the adjacent 3′-5′ phosphodiester linkage resistant to RNase T2 as this enzyme cleaves RNA via 2′-3′phosphate cyclization. These samples were then immunoprecipitated with anti-TMG agarose beads and 3′-5′ phosphodiester linkages were cleaved by on-bead RNase P1 treatment, to restrict analysis to the cap 2 modified nucleotide, if present, and its adjacent cap 1 nucleotide. Released nucleotides were then analyzed by 2D-TLC. If nucleotide +2 is not 2′-O-ribose methylated only the cap 1 unmodified nucleotide, mostly uridine, would be released, compared to the release of both 2′-O-ribose methylated nucleotides (position +2) and the unmodified nucleotide (position +3) if a cap 2 modified nucleotide is present (Figure 7A). As a result, no modified nucleotides were detected indicating the absence of a cap 2 structure in T. vaginalis mRNAs (Figure 7B). As predicted the most abundant unmodified nucleotide detected was uridine. Although these data cannot exclude the presence of a cap 2 structure in a small subset of mRNAs, if present these are highly underrepresented.
The unexpected finding of a metazoan/plant-like capping apparatus and a cap 1 modified nucleotide on T. vaginalis mRNAs led us to investigate the phylogeny of TvCE. A global taxa sampling of animal, fungal, plant, microbial and iridovirus GTases as well as a subset dataset including exclusively the TPasePL-GTase structural organization were aligned and subjected to protein maximum likelihood phylogenetic analyses (Figure 8 & Figure 9). Although generally poorly resolved, the global GTase phylogeny recovered the T. vaginalis sequence with modest support value (64% bootstrap proportion, BP but increasing to 70% and 77% when one or two of the most divergent sequences were removed) in a clan of exclusively TPasePL-GTase configured sequences, consistent with a common origin of their structural organization (Figure 8). The TPasePL-GTase maximum likelihood tree revealed that TvCE does not cluster with proteins from animals (Figure 9) and hence, at face value, does not support lateral gene transfer (LGT) between T. vaginalis or a Parabasalid ancestor and their animal hosts. These data are consistent with an ‘Unikonts’/‘Bikonts’ split [39] with on one hand plants and green algae forming a clan with T. vaginalis supported with 58% BP, raising to 77% in the absence of the iridovirus sequence (Figure 9), and the metazoan and choanoflagellates forming the other clan in line with phylogenomic data [24]. As the position of the eukaryotic root is currently unknown [39] it is not clear how the iridovirus sequence relates to the eukaryotic sequences; its divergent sequence makes inference of its position in both analyses tentative. Consistent with potential long-branch attraction (LBA) issues for this sequence, its removal in the TPasePL-GTase alignment improved the BP of several nodes including the one for the T. vaginalis-plant/green algae clan and the metazoan clan (Figure 9). Recoding the 20 amino acids into four categories, to allow optimization of the rate matrix and reduce composition heterogeneity between sequences and mitigate potential LBA artefacts, recovered a similar tree with reduced BP for the split between the TvCE-plant/algae clan and the metazoan clan in maximum likelihood analyses. Together our data favor a tree topology consistent with phylogenomics data [24] and cluster TvCE with plant and green algae sequences; however, LGT from a metazoan donor to a parabasalid cannot be strictly rejected due to lack of strong signal.
This is the first report of the functional characterization of a microbial eukaryote harboring a typical metazoan/plant-like capping apparatus containing a fused dual-functional TPasePL-GTase (TvCE). In addition to this unusual capping apparatus in T. vaginalis, mRNAs from this unicellular eukaryote were found to contain a cap 1 structure, another feature of metazoan mRNAs. The GTase domain of TvCE selectively binds GMP. The transfer of GMP to the 5′ triphosphate end of the RNA is strictly dependent on the prior removal of the γ-phosphate by the TPase activity of TvCE. The TPase domain of TvCE has typical features of a metazoan cysteinyl-phosphatase enzyme; it is active in the absence of metals as a bifunctional monomeric enzyme, prefers an acidic pH, and forms a phospho-enzyme through a cysteine-thiophosphate linkage within a conserved ‘P-loop’ active site which is necessary for phosphatase activity. An apparent cis-structural effect in TvCE was observed between the two active domains as a single amino acid change in the GTase domain results in a detectable reduction of TPase activity. We also observed that purification of the individual TvCE TPase domain resulted in an inactive enzyme. Together, these data suggest that the GTase domain can affect the activity of its neighboring TPase domain.
Analyses of the cap structure of mRNAs in this divergent microbial eukaryote revealed the presence of a canonical m7G cap 0 nucleotide, consistent with the presence of two conserved RNA (guanine-7) MTase genes in the T. vaginalis genome (http://trichdb.org) that could convert the guanosine at the 5′ end of the mRNA by TvCE to an m7G. We previously showed that T. vaginalis has an atypical trimethylguanosine synthase (TgS) that produces m2,7G from m7G RNA substrates [36]. These observations led us to speculate that T. vaginalis mRNAs might contain m2,7G caps. Contrary to this prediction, the analyses presented here demonstrate the presence of a canonical m7G cap nucleotide on T. vaginalis mRNAs.
A general comparison of cap structures among different eukaryotes suggested that complexity increases following a phylogenetic trend across the evolution of eukaryotes [5]. This is illustrated by the presence of a complex cap 2 structure on mRNAs of metazoans but only cap 0 on budding yeast mRNAs. Prior to the studies reported here, the unusual cap 4 in trypanosomatids [7]–[11] appeared to be the exception to this rule. T. vaginalis mRNAs which contain a conserved cap 1 structure composed primarily of m7GpppAmpUp or m7GpppCmpUp are now also exceptions. No cap 2 structure was detected in T. vaginalis mRNA and a strong bias to uridines at position +2 was observed, making this cap structure less complex than metazoans. It should also be noted however, that the use of a heterogeneous population of mRNA may have obscured the detection of a cap 2 nucleotide in a smaller subpopulation of mRNAs. Similarly, although we consider it unlikely, we cannot strictly preclude that immunoprecipitation of mRNAs using the anti-TMG antibody led to the exclusion of a subset of mRNAs with a cap 2 nucleotide or an alternative cap structure that are not efficiently bound by this antibody.
The observation that ∼80% of cap 1 nucleotides in T. vaginalis mRNAs are adenosines (A) while ∼20% are cytosines (C) indicates that transcription initiation by RNA polymerase II in this organism can occur at cytosine, as well as adenosine. In turn this suggests the presence of either unknown variants of the initiator (Inr) motif that surrounds the start site of transcription of T. vaginalis mRNAs [37], [38] or unrelated motifs that can direct transcription to initiate at a cytosine. This is consistent with our previous prediction that only ∼75% of T. vaginalis genes appear to use a classic Inr leading to transcription initiation at an A [29].
The structural organization of the mRNA caps enzyme machinery in eukaryotes has been considered a marker for inference of eukaryote phylogeny [1]. Fungi and most sampled microbial eukaryotes have separate TPaseMDP and GPase capping enzymes with the TPaseMDP structurally and mechanistically distinct from the TPasePL that is fused to a GTase in metazoan and plants. Thus, the acquisition of the metazoan-type dual-function enzyme after the divergence of unicellular and multicellular eukaryotes has been proposed [1]. However recent genome samplings have broadened the taxonomic diversity of TPasePL-GTase encoding taxa, which now includes green algae, a choanoflagellate and T. vaginalis. These additional data and our structural, functional and phylogenetic analyses of TvCE complicates the earlier simple dichotomy observed between metazoan-plants and microbial eukaryotes. Indeed, the proponent of this dichotomy acknowledged that “The scheme is certainly oversimplified…” due to poor taxa sampling [1].
Differential gene losses could explain the unusual presence of a metazoan/plant-like capping apparatus in T. vaginalis, green algae and choanoflagellates. The LCA of microbial eukaryotes, plants and animals may have contained the fused TPasePL-GTase and independent gene losses subsequently occurred in most currently sampled microbial eukaryotes, leaving T. vaginalis, green algae and choanoflagellates as rare microbial eukaryotes carrying this prototype. Notably the green algae and choanoflagellate TPasePL-GTase sequences were recovered in the expected clans as defined by phylogenomics [24], pushing the acquisition of this configuration among a likely microbial ancestor deeper in eukaryotic evolution. Alternatively, the LCA of plants/green algae and animals/choanoflagellates had the split system seen today in many microbial eukaryotes and subsequently lost this and acquired a new fused version. Similar but independent events would than be invoked for the acquisition of TvCE, leading to an enzyme with little sequence relatedness to its counterpart in plants and animals. However the global GTase phylogeny suggests that the GTase from the TPasePL-GTase fusion was shared by the LCA between T. vaginalis, plants/algae and choanoflagellate/animals; as expected if the gene was present in the LCA of all eukaryotes. Alternatively, the GTase for the taxa with TPasePL-GTase configurations was independently acquired by different lineages from similar sources with the same structural organization and/or T. vaginalis acquired TvCE by LGT. Existing phylogenies do not provide positive evidence for these scenarios. As Parabasalia and Diplomonads appear to be closely related within the excavates [24] an independent origin scenario of the TPasePL-GTase fusion could have been suggested if the TvCE GTase formed a clan with the G. lamblia GTase, however this was not found in our maximum likelihood phylogenetic analyses (Figure 8, Figure S4). These data clearly reinforce the importance of further genome sampling among various microbial eukaryotes and viruses before evolutionary hypotheses based on the mRNA capping system can be appropriately assessed and contrasted to the current hypothesis of eukaryote phylogeny [39]. Phylogenomic data obtained so far consistently provide evidence for at least six major eukaryotic lineages [24], [27] that do not match those suggested by the distribution of the mRNA capping enzymes [1]. Our phylogenetic analyses do not support LGT as the source of TvCE in T. vaginalis reducing the conflict between the distribution of the mRNA capping machinery and the six major lineages recovered by phylogenomics [39]. In contrast, relationships among microbial eukaryote GTases functioning with TPaseMDP seem more complex and in conflict with phylogenomic data in terms of major group relationships. For instance non-monophyly is observed for microsporidial GTases, which do not cluster with the Fungi, and the red algae cluster with the Amoebozoa. Could these be explained by LGT events as described for the acquisition of a trifunctional capping enzyme in a mimivirus, thought to be derived from its amoeba host [40]? Likewise the iridovirus TPasePL-GTase may have been acquired from an animal host. The evolution of capping enzymes in eukaryotes appears to have proceeded via multiple events that led to the independent loss and/or gain of genes in different lineages. Polarization of such events will require denser sampling of mRNA capping genes and additional robust independent phylogenetic analyses.
The ORF encoding TvCE (TVAG_187730) was cloned into Escherichia coli expression vector pET 200D- (Invitrogen), which adds a Histidine-tag at the N-terminus. TvCE was cloned either as a full length protein or as two separate domains. The full-length protein has a predicted molecular weight of 69 kDa (not including the histidine tag) and a pI of 6.9. A MUSCLE alignment of the protein predicted the TPase domain to reside between amino acids 1 and 254 and the GTase domain between amino acids 255 and 561. pET-200D constructs containing either the full-length or the TPase or GTase domain were transfected into E. coli strain BL21 as provided by the manufacturer (Invitrogen). 250 ml of bacteria cultures were grown to OD600 0.4–0.5 and 3% (v/v) ethanol and 0.2 mM IPTG was added to induce protein expression. Incubation was continued for 18–20h at 18°C, shaking at 180 rpm. Induced cells were centrifuged and resuspended in 50 mM NaPO4 pH 6.0, 1 mg/ml of lysozyme and 1× Halt protease inhibitor cocktail (ThermoScientific). Lysis was achieved by sonication on ice and cell debris was removed by spinning samples at 16,000× g for 30 min. The presence of soluble expressed recombinant protein was evaluated by SDS-PAGE before loading it on a pre-equilibrated 5 ml Mono-S column (GE Healthcare). The column was washed with 10 volumes loading buffer minus lysozyme. Proteins were eluted with a 5 ml step gradient of NaCl (100, 200, 300, 400 and 500 mM) in the same buffer. Eluted fractions shown to contain rTvCE by SDS-PAGE analysis were pooled together and imidazole was added to 40 mM. The proteins were further fractionated using Ni chromatography (HisTrap; GE Healthcare), as recommended by the manufacturer. After the two-step purification, the final purified rTvCE (50–150 ug of protein) was dialyzed against 50 mM Tris pH 7.4, 100 mM NaCl, 2 mM DTT and 10% glycerol. PCR mutagenesis was carried out as described [41] and proteins were isolated using the 2-step purification scheme described above.
To determine optimal pH of TvCE, 20 µl reactions containing 25 nM of full-length recombinant TvCE (flr-TvCE), 5 mM DTT, 160 nM [γ-32P]ATP were adjusted to 50 mM Tris-acetate (pH 7.0 and below) or 50 mM Tris-HCl (pH 7.5 and above). To evaluate metal dependence of TvCE TPase, 20 µl reactions containing 50 mM Tris-acetate pH 5.5, 350 nM of flr-TvCE, 5 mM DTT and 15 nM [γ-32P]ATP were performed in the presence of 0, 1, 2.5 and 5 mM MgCl2. These reactions were assays in either the presence of 0 or 20 mM EDTA for comparison. TvCE TPase nucleotide dependence was measured in 20 µl reactions containing 50 mM Tris-acetate pH 5.5, 5 mM DTT, a range of 2.5–160 nM [γ-32P]ATP or [γ-32P]GTP and 25 nM of flr-TvCE. For detection of phosphatase activity of TvCE, reactions were incubated at 37°C for 30 min, and 32Pi was detected on TLC plates after autoradiography as described [17]. Products were sliced from the TLC plastic plates for quantification by liquid scintillation. A mock reaction (minus enzyme) was done in parallel to account for spontaneous radiolysis of the substrate. TvCE TPase was tested for the formation of a covalent protein-cysteinyl-S-phosphate intermediate [17], [25]. To test the hypothesis that such an intermediate can be formed under acidic pH, a 10 µl reaction containing 50 mM Tris-acetate pH 3.0 to 5.0, 5 mM DTT, 160 nM [γ-32P]ATP and 25 nM of flr-TvCE was performed at 25°C for 15 seconds. One µl of this reaction was loaded on SDS-PAGE for Coomassie Blue staining and autoradiography. This analysis was carried out for flr-TvCE mutants, except that molar concentration of [γ-32P]ATP was increased to 330 nM, unincorporated nucleotides were removed by G-50 microcolumns (Amersham) and all sample contents were analyzed by SDS-PAGE. To confirm that TvCE is phosphor-labeled through formation of a thiophosphate linkage, chemical stability analysis was performed as previously described [17]. For this analysis, we compared thiophosphate linkage stability when phosphor-TvCE was treated with H2O, 100 mM NH2OH or 10 mM iodine.
To evaluate metal specificity of TvCE GTPase, 10 µl reactions containing 50 mM Tris-acetate pH 7.0, 5 mM DTT, 20 nM [α-32P]GTP and 40 nM of full-length recombinant TvCE (flr-TvCE) were performed, varying the concentrations of either MgCl2, MnCl2 or CaCl2 from 0–10 mM. To evaluate substrate specificity of TvCE GTPase, a cold competition experiment was designed. 10 µl reactions containing 50 mM Tris-acetate pH 7.0, 5 mM DTT, 2.5 mM MgCl2, 100 nM [α-32P]GTP, 40 nM of recombinant TvCE were performed in the presence of 0–1 µM cold nucleotide competitor ATP, CTP, GTP or UTP. TvCE GTase dependence on nucleotide concentration was measured in 20 µl reactions containing 50 mM Tris-acetate pH 7.0, 5 mM DTT, a range of 0–16 nM [α-32P]GTP and 3 nM flr-TvCE. All reactions were incubated at 37°C for 30 min. and GTase activity was detected by the formation of the covalent enzyme-GMP intermediate [18]. The reaction product was detected on SDS-PAGE and autoradiography. The phosphor-labeled enzyme was sliced from gels and quantified by liquid scintillation.
A fragment containing 352 bp of T. vaginalis ß-tubulin was in vitro transcribed by T7 RNA polymerase and quantified as previously described [36], and used as a substrate for the full length recombinant TvCE in a two-step reaction. As a negative control, one RNA sample was dephosphorylated by alkaline phosphatase treatment (Apex, Epicentre) prior to TvCE RNA capping activity and purified by phenol/chloroform extraction and ethanol precipitation. The first step of the reaction, the removal of the γ-phosphate, was tested by incubating 250 ng of the RNA in a 50 µl reaction containing 50 mM Tris-acetate pH 5.5, 5 mM DTT, and 200 nM recombinant TvCE at 37°C for 30 min. The RNA was then purified by phenol/chloroform extraction and ethanol precipitation. Next, the RNA was resuspended in a 50 µl reaction containing 50 mM Tris-acetate pH 7.0, 5 mM DTT, 2.5 mM MgCl2, 90 nM [α-32P]GTP and 200 nM recombinant TvCE and incubated at 37°C for 30 min. The RNA was then purified as described above, split into equal part and analyzed in 7% polyacrylamide gels under denaturing conditions (Tris-Borate EDTA buffer and 8M urea). One sample was stained by ethidium bromide and the other was dried and exposed to X-ray film.
T. vaginalis strain T1, grown in TYM complete media [42] was subjected to in vivo labeling of RNAs. To achieve 12–18% total 32P incorporation, 1–5×108 parasites were starved in the absence of phosphate-, serum-free DMEM and 1 mCi of phosphorus-32 radionuclide for 8–9 hours at 37°C. Cultures were mixed by inversion every hour during incubation. RNA was extracted and size-fractionated from 32P-labeled and unlabeled T. vaginalis cultures using the mirVana PARIS Kit (Ambion). To determine the nucleotide structure of the cap of T. vaginalis mRNAs, a protocol to obtain a population of mRNAs free of the hypermodified ribosomal and transfer RNAs was developed. The high relative abundance of hypermodified nucleotides present in these RNA species interfered with detection of modified nucleotides found specifically on mRNAs in the presence of minor contamination by rRNA or tRNA. First the large-size RNA fraction (>200 nt) was isolated from cells to minimize contamination with small rRNA and tRNAs. The RNA is then passed by two consecutive rounds of poly dT purification (Promega). The eluted RNA was then concentrated by ethanol precipitation, followed by Terminator Exonuclease (Epicentre) digestion to degrade RNAs that contain a 5′-monophosphate (ex. rRNAs). The large-size RNAs were again isolated, and purified from partially degraded rRNAs and nucleotides. As a final step, the mRNAs were purified by immunoprecipitation using the mouse monoclonal anti-TMG (anti-trimethylguanosine agarose conjugate, CalBiochem) as described [30], taking advantage of the cross-reactivity of this antibody with the 7-methyl guanosine cap of mRNAs. The purity of the mRNA preparation protocol was assessed by electrophoresis and two-dimension thin layer chromatography (2D-TLC) analysis of in vivo labeled mRNAs after autoradiography. 2D-TLC was carried out using both combinations of organic solvents A, B and C, as previously described [36], [43]. To evaluate the presence of a 5′-end guanosine cap linked by a triphosphate bridge to the RNA, in vivo labeled mRNAs (∼105 cpm) were digested with TAP (Epicentre) and analyzed by 2D-TLC. To confirm the identity of a possible m7G cap structure in this highly purified fraction of T. vaginalis mRNA, the sample was treated with S. pombe Tgs prior to TAP treatment. This enzyme converts m7G to m2,2,7G which can resolved by 2D-TLC [36]. To identify a possible nucleotide modification at position +1, mRNA was 5′-end labeled. The mRNA was digested or mock-digested with TAP, dephosphorylated by Alkaline Phosphatase (APex, Epicentre) and heat-inactivated, and labeled by PNK with [γ-32P]ATP. Between enzymatic treatments, RNA was purified by phenol/chloroform extraction and ethanol precipitation. TAP-digested or mock-digested samples were comparatively analyzed by 2D-TLC. Spots were quantified by liquid scintillation. To identify a possible nucleotide modification at position +2, we utilized in vivo labeled mRNAs. The protocol necessary to obtain mRNA free of detectable hypermodified nucleotides from ribosomal and transfer RNA, as described above, was followed except that before anti-TMG immunoprecipitation the RNA sample was concentrated by ethanol precipitation and digested with RNase T2 in a 20 µl reaction volume. Six µl of this reaction was then anti-TMG immunoprecipitated in a 0.6 ml end volume. After washes, the RNase T2-digested m7G capped mRNAs bound to the anti-TMG agarose beads were mildly digested with nuclease P1 in a 25 µl reaction volume for 2 h at 37°C under agitation. The anti-TMG beads were then recentrifuged, the supernatant was recovered and P1-digestion continued at 50°C for 16 h to completion. Five µl of this reaction were loaded on 2D-TLC plates for analysis. For 2D-TLC comparative maps, radiolabeled m7G, m2,7G and m2,2,7G standards were produced as previously reported [36]. 2′-O-ribose methylated nucleotide standards were produced by P1-digestion of a AmCmGmUm oligomer (Sigma-Proligo), and all were compared to previous reported maps [36], [43].
Sequences were extracted from protein databases following BLASTp searches at NCBI using default settings except for the sequence of the red algae Cyanidioschyzon merolae GTase which was obtained from the KEGG database. Selected sequences aligned for comparison and phylogenetic inferences are described in Data S1. Proteins structural organizations were investigated with SMART searching both SM and Pfam profiles [44]. For phylogenetics reference protein sequences were chosen from BLAST hit lists to maximize taxa diversity and sequences aligned with ClustalW [45]. SEAVIEW 4.0 [46] was used to visually check the alignment features and sites used for phylogenetic analyses were selected with the mask option. Sites with more then five indels were deleted as where divergent sites. The best-fitting models for the protein alignments were identified with Prottest 2.2 [47], which was invariably the LG rate matrix [48] with a gamma rate model “G” for across site rate variation. PhyML [49] was used within SEAVIEW to perform maximum likelihood phylogenetic inferences. The alpha shape parameter of the gamma rate model (four categories) was estimated using the BioNJ distance tree used as the starting tree in conjunction of both NNI and TPR branch swapping moves for further optimization. The optimal alpha shape parameter for site rate variation was then fixed and used for bootstrap analyses (100 replicates). In order to mitigate potential issues of composition bias and long branch attractions, the protein alignment were recoded for the TPasePL-GTase protein alignment, as described [24] with the 20 amino acids reduced to four categories implied by the JTT rate matrix ([A,N,G,T,P,S], [R,D,E,Q,K], [E,L,M,F,V] and [HWYC]). Following removal of invariant sites in the recoded alignment (410 sites were reduced to 316 sites), PhyML maximum likelihood analyses (GTR rate matrix with gamma model, both estimated) were performed and recovered similar trees as those based on the protein alignments. All alignments are available upon request.
|
10.1371/journal.pcbi.1000716 | Emergence of Spatial Structure in Cell Groups and the Evolution of Cooperation | On its own, a single cell cannot exert more than a microscopic influence on its immediate surroundings. However, via strength in numbers and the expression of cooperative phenotypes, such cells can enormously impact their environments. Simple cooperative phenotypes appear to abound in the microbial world, but explaining their evolution is challenging because they are often subject to exploitation by rapidly growing, non-cooperative cell lines. Population spatial structure may be critical for this problem because it influences the extent of interaction between cooperative and non-cooperative individuals. It is difficult for cooperative cells to succeed in competition if they become mixed with non-cooperative cells, which can exploit the public good without themselves paying a cost. However, if cooperative cells are segregated in space and preferentially interact with each other, they may prevail. Here we use a multi-agent computational model to study the origin of spatial structure within growing cell groups. Our simulations reveal that the spatial distribution of genetic lineages within these groups is linked to a small number of physical and biological parameters, including cell growth rate, nutrient availability, and nutrient diffusivity. Realistic changes in these parameters qualitatively alter the emergent structure of cell groups, and thereby determine whether cells with cooperative phenotypes can locally and globally outcompete exploitative cells. We argue that cooperative and exploitative cell lineages will spontaneously segregate in space under a wide range of conditions and, therefore, that cellular cooperation may evolve more readily than naively expected.
| Cooperation is a fundamental and widespread phenomenon in nature, yet explaining the evolution of cooperation is difficult. Natural selection typically favors individuals that maximize their own reproduction, so how is it that many diverse organisms, from bacteria to humans, have evolved to help others at a cost to themselves? Research has shown that cooperation can most readily evolve when cooperative individuals preferentially help each other, but this leaves open another critical question: How do cooperators achieve selective interaction with one another? We focus on this question in the context of unicellular organisms, such as bacteria, which exhibit simple forms of cooperation that play roles in nutrient acquisition and pathogenesis. We use a realistic simulation framework to model large cell groups, and observe that cell lines can spontaneously segregate from each other in space as the group expands. Finally, we demonstrate that lineage segregation allows cooperative cell types to preferentially benefit each other, thereby favoring the evolution of cooperation.
| Many cell phenotypes alter the growth and division of nearby cells by changing local resource availability [1]–[4]. Some of these phenotypes promote the survival and reproduction of others, and thus qualify as a simple form of cooperation. A cell may be considered cooperative, for example, if it secretes enzymes that free nutrients which neighboring cells can use. The efficiency with which a cell group processes environmental resources or exploits a host often depends on such publicly beneficial cell phenotypes. For instance, many microbial infections and cancerous tumors derive their pathogenicity in part from the cooperative secretion of digestive enzymes by their constituent cells [5]–[8].
How cooperative cell phenotypes evolve therefore presents an important question, one that is particularly challenging because any genetic variants that exploit others' cooperation – without themselves paying a cost – can potentially invade and increase in frequency. In light of this problem, social evolution theory has been developed to understand the evolutionary trajectories of cooperative traits [9], but this framework has only recently been applied to unicellular systems [4], [10]–[12]. The critical prediction is that preferential interaction among genetically related individuals increases the propensity for cooperative phenotypes to evolve.
Variation among individual cells is a common feature of many cell groups: microbial biofilms are often composed of multiple strains or species [13],[14], and cancerous tumors can consist of many different genetic lineages [15],[16]. The majority of work on cooperative cell phenotypes assumes relatively well mixed interactions among different genetic variants in standing or shaken liquid culture [17]–[21]. This kind of environment does not reflect the natural condition of most cell groups, in which cells are typically constrained in space and influence each other in a distance-dependent manner. These spatial relationships may be paramount to understanding the evolution of cellular cooperation [22]. When different cell lineages are segregated in space, those expressing cooperative phenotypes are more likely to benefit others of their own kind [23]–[25]. When different cell lineages are mixed together, on the other hand, cells that exploit the resources of others can thrive [17]–[20].
Local populations of bacterial and cancer cells are often established by groups of progenitors that proliferate into larger clusters. Experiments with bacterial colonies on agar have revealed that expanding cell groups can segregate into sectors that are each dominated by a single genetic lineage [26],[27]. This observation has been used predominantly to motivate new population genetic models [28]–[30]. When only cells on the periphery of an expanding group can access nutrients and reproduce, the group's effective population size is reduced. As a result, neutral or even mildly deleterious alleles can spread by genetic drift along the advancing front. Because they are constrained in space, genetic lineages that manage to proliferate along the population's leading edge become physically separated into zones composed of clonal or closely related individuals.
By promoting interaction between individuals of the same genotype, the spontaneous segregation of different genetic lineages in space may also influence social evolution within cell groups [23],[24]. In the present paper, we use a generalized mechanistic model to define the physical and biological factors that govern cell group spatial structure, and we explore the potential connection between genetic drift along the fronts of expanding cell groups and the evolution of social phenotypes.
To study how the collective structure of cell groups arises from the activity of many individual cells, we used a computational model that employs mechanistic descriptions of solute diffusion and cell growth [31]–[33]. Our framework is derived from the latest generation of agent-based models that have been developed over the last decade using biochemical engineering principles (Methods, Supporting Information, Table S1). The model's underlying assumptions are described and justified in detail elsewhere [33]–[36], and empirical tests have demonstrated the framework's ability to make accurate predictions for real biological systems [37],[38].
Briefly, each cell is implemented as a circular agent in explicit two-dimensional space, and each simulation is set on one of two possible conditions. The first consists of cells growing on a flat surface with growth substrate (nutrients) diffusing from above. The second condition represents a cell cluster immersed in a resource pool, such that substrate diffuses into the cluster from all directions. The transport of all solutes occurs exclusively through diffusion. Each cell grows according to a Michaelis-Menten function of substrate concentration in its local environment and divides once it reaches a maximum radius (Methods, Supporting Information, Table S2). Cells move passively due to the forces exerted between neighboring individuals as they grow and divide.
We began with simulations in which the environment surrounding cell groups was altered by increasing or decreasing growth substrate concentration. These in silico experiments were initiated with equal numbers of randomly distributed red and blue cells, which did not differ in any way other than their color. The two neutral color markers were used to judge whether cell lineages remain randomly mixed or become spatially segregated as cell groups expand. Environmental substrate availability was decreased from saturating to sparse across multiple simulations, and we observed three different regimes in cell group structure:
Our next goal was to describe why environmental substrate concentration affects lineage assortment in expanding cell groups. Under limited growth substrate availability, the majority of cell growth and division occurs along a group's advancing front in an active layer whose depth depends on substrate penetration (Figure 3). Previous work has hinted that active layer depth is a critical factor influencing cell group surface structure [39],[40], and we therefore hypothesized that it is not substrate concentration in particular, but more generally the depth of a cell group's active layer that controls cell lineage segregation. Because segregation increased as growth substrate supply decreased in our preliminary simulations, we predicted that thinner active layers would lead to stronger lineage segregation in expanding cell groups.
Active layer depth is not solely a function of bulk growth substrate concentration. For example, higher substrate diffusivity increases active layer depth by allowing substrate to enter further into the cell group before being depleted. Faster cell growth rates, on the other hand, decrease active layer depth by raising the rate of substrate consumption at the cell group's outer surface. If we are correct that active layer depth is the underlying determinant of lineage segregation, all of the physical and biological factors that control active layer depth should also influence lineage segregation in cell groups.
Using an analytical technique from chemical engineering (Methods), we combined the factors that influence active layer depth into a dimensionless number, δ, which has the following form for our system:(1)
Here, Gbulk is the bulk liquid concentration of growth substrate, DG is the growth substrate diffusion coefficient, Y is the yield with which cells convert substrate to biomass, μmax is the maximum specific cell growth rate, ρ is the cell biomass density, and h is the height of the diffusion boundary layer (Figure 3). The smaller the value of δ, the thinner the cell group's active layer.
We performed three new sets of simulations to test the hypothesis that active layer depth controls cell lineage segregation. Within each set, we varied active layer depth (δ) by altering only one parameter from Equation 1: maximum cell growth rate (μmax), bulk growth substrate concentration (Gbulk), or growth substrate diffusivity (DG). At the end of each simulation, we calculated the segregation index. Our hypothesis makes two key predictions: 1) cell lineage segregation should be inversely related to δ, a proxy for active layer depth. 2) The relationship between cell lineage segregation and δ should be independent of which parameter from Equation 1 is altered.
The results are shown in Figure 4 and support both predictions. Lineage segregation within cell groups declines with increasing δ, regardless of how δ is altered. Using the dimensionless number δ renders our results independent of the exact values of Gbulk, DG, Y, μmax, ρ, and h used to run simulations. It is the relative magnitudes of these parameters in combination that ultimately matter.
How does active layer depth influence cell lineage segregation? When growth substrate penetrates through most of a cell group before being depleted, all cells grow and divide, pushing each other into a homogeneous mixture. As active layer depth decreases below the total thickness of a cell group, however, cells that happen to fall below a critical distance from the group's front can no longer contribute to population expansion. Decreasing active layer depth thus reduces the cell group's effective population size, rendering it more susceptible to genetic drift along its advancing front. Because the cells are constrained in space, reductions in genetic diversity along the group's leading edge lead to localized clusters of individuals that all descend from a common progenitor [30]. This phenomenon – often referred to as sectoring or gene surfing [28]–[30] – has been observed in agar colonies of Paenibacillus dendritiformis [26], Escherichia coli and Saccharomyces cerevisiae [27].
Reducing active layer depth even further yields an additional qualitative shift in cell group structure: the expanding population becomes sensitive to small irregularities along its leading edge. Cells in the peaks of surface irregularities retain access to substrate and grow into tower projections, while cells in the troughs of surface irregularities lose access to substrate and cease growing. This process is related to viscous fingering at the interface of two fluids [39],[41], and it is known to generate rough surface structure along the leading edges of growing biofilms, bacterial colonies on agar [34],[35],[40], and moving fronts in general [36]. From a biological perspective, our analysis predicts that such surface roughness is accompanied by abrupt genetic lineage segregation along the front of an expanding population.
The spatial assortment of cell lineages is potentially critical for traits that affect the reproduction of other individuals in the population. It is increasingly recognized that cells express many such social phenotypes [4],[12], which are often involved in nutrient acquisition and pathogenesis [42]–[45]. A common example is the secretion of extracellular enzymes or nutrient-chelating molecules. Cells that synthesize these substances must forgo a fraction of their reproductive capacity [17]–[19], but if enough cells participate, all can gain a net benefit (to the detriment of their host, in the case of pathogens).
In many cases the evolution of simple cooperative phenotypes depends on three factors: 1) c, the cost incurred by cooperative individuals 2) b, the benefit gained by the receivers of cooperative behavior, and 3) r, the correlation between genotypes of givers and receivers of cooperation. Cooperation is predicted to evolve when rb>c, a condition known as Hamilton's Rule [9]. The cost and benefit factors are measured in terms of reproductive fitness. When cooperation is genetically determined, relatedness may be thought of as the degree to which the benefits of cooperation are preferentially distributed to other cooperative individuals.
The segregation index depicted in Figures 2 and 4 is equivalent to a form of the relatedness coefficient in Hamilton's Rule: both measure the degree of biased interaction among relatives (here, physical proximity amounts to biased interaction). As such, our segregation index forms a bridge between social evolution theory and the emergence of lineage segregation in cell groups, allowing us to extend our prediction from the previous section. Because thin active layer conditions generate lineage segregation, we predict that decreasing active layer depth will promote interaction among clonemates (increasing r in Hamilton's Rule) and favor the evolution of cooperation [9],[12],[23]. Positive spatial assortment of related cells does not guarantee that cooperation will be favored, however, as the same segregation that allows cooperators to preferentially interact also increases the strength of competition between them [24].
We tested our prediction by implementing a cooperative phenotype in our model framework and competing cooperative cells against exploitative cells that devote all resources to growth. Cooperative individuals secrete a diffusible compound that benefits all other cells in the local area (we will refer to the compound as an extracellular enzyme). Local availability of the secreted enzyme increases cell growth rate by a fold factor B, but only after the enzyme's concentration passes a threshold value, τ. Cooperative cells constitutively secrete the enzyme and incur a fold decrease in growth rate of C x RE, where C is a cost scaling factor and RE is the enzyme production rate. In our main analysis, B = 3, C = 0.3, and RE ranges from 0 to 2. We derived these values from experimental data on elastase, a secreted enzyme and virulence factor of the bacterial pathogen Pseudomonas aeruginosa [19],[46].
We asked whether a cooperative cell line, which pays a cost to produce a diffusible, publicly beneficial compound, could outcompete an exploitative cell line that invests all of its resources into growth. Each competition simulation began with a randomly distributed 1∶1 mixed monolayer of the two cell types, and cell groups were grown to a maximum height of 100 µm. We then calculated the evolutionary fitness of the cooperative cell line, relative to that of the exploitative cell line (Methods). This competition pairing was repeated over a range of extracellular enzyme production rates on the part of cooperative cells. The higher the enzyme production rate, the more rapidly cells accrue its benefit, but the larger the cost suffered by cooperative cells. Finally, all competition pairings were repeated across three active layer depth conditions (δ = 10, 2, 1), representing the three cell group structure regimes described in Figure 1.
Figure 5 summarizes the results of our competition simulations. When active layers are thick (δ = 10), leading to well mixed cell lineages, the extracellular enzyme is homogenously distributed through cell groups. The non-cooperative cell line is therefore able to consistently exploit and outcompete the cooperative cell line (Figure 5A). This result is consistent with numerous observations that exploitative mutants outcompete enzyme-secreting bacteria when they are inoculated together in liquid culture, in which cell lineages largely remain mixed [17]–[20].
When active layer depth is decreased (δ = 2), there is a narrow range of extracellular enzyme production rates at which cooperative cells outcompete exploitative cells (Figure 5B). The critical difference is that cooperative cells and exploitative cells no longer remain well mixed; rather, they segregate into clonal regions. As a result, the benefit of extracellular enzyme released by cooperative cells accrues asymmetrically to other cooperative cells. The range of enzyme production rates at which cooperative cells prevail is narrow, however, because the benefits of lineage segregation (increasing r in Hamilton's Rule) can be outweighed by the cost of higher extracellular enzyme production (increasing c in Hamilton's Rule).
Further decreasing active layer depth (δ = 1) leads to the growth of spatially isolated, clonal cell towers. Under these conditions, the benefits of a cooperative secreted enzyme are distributed even more asymmetrically to other cooperative cells. Consistent with our predictions, this allows cooperative cells to outcompete exploitative cells over a larger range of enzyme production rates (Figure 5C). We also noted the sizable variation between simulation runs when δ = 1, particularly if extracellular enzyme production rates were low (Figure 5C, enzyme production rate = 0, 0.25, 0.5). This variation reflects a founder effect; it manifests most strongly when there is no or little difference between the competitive abilities of cooperative and exploitative cell lines, rendering the outcome of each simulation subject to chance events that determine which cells seed the few tower structures that emerge from an expanding cell group.
Our results show that thin active layer conditions allow cells expressing cooperative phenotypes to outcompete exploitative cells within a single cell group. To better account for the long-term evolution of a metapopulation comprising many cell groups, we performed an invasion analysis to determine whether a novel cooperative mutant can spread through a metapopulation otherwise containing only exploitative cells (Supporting Information, Text S1). We also examined the reciprocal case to determine if a rare exploitative mutant can invade a metapopulation otherwise containing only cooperative cells [32],[33]. We found that cooperation can invade under a large swath of parameter space, but only under thin active layer conditions that promote lineage segregation can cooperative cells eliminate exploitative cell types on a metapopulation scale (Supporting Information, Figure S2).
The results of both our local competition and invasion analyses are robust to the cost/benefit ratio of cooperation, with one partial exception when cells invest very heavily into an expensive cooperative phenotype (Supporting Information, Figure S3).
Our study indicates that an order of magnitude change in nutrient availability, nutrient diffusivity, cell metabolic efficiency, cell growth rate, or biomass density can shift cell groups from a regime of lineage mixing to a regime of pronounced lineage segregation. The number δ defined in Equation 1 relates these parameters to the depth of a cell group's active layer, which governs how cell lineages become spatially assorted over time. Thick active layers promote lineage mixing, while decreasing active layer depth generates increasingly strong lineage segregation. Cell lineage segregation, in turn, favors the evolution of cooperative phenotypes.
Previous work performed with bacteria in liquid planktonic culture has concluded that cooperative cell phenotypes cannot be selectively favored within a single population also containing exploitative cells [17],[19],[20]. Our study shows that this conclusion will not always hold because cooperative cells can spontaneously segregate from exploitative cells when they are constrained in space. Our results also imply that, given realistic parameters for a cooperative cell phenotype, the benefits of preferential interaction between cooperators can outweigh the costs of increased competition between related cells that are clustered together in space [24].
Like all models, ours uses simplifying assumptions. We deliberately omit some physical processes, such as shear stress, that may be applied to cell groups in the real world [47]. Our simulations also do not consider active cell motility, which in reality could influence cell group structure and evolution. We have additionally assumed that cell phenotypes of interest, such as extracellular enzyme secretion, are expressed constitutively or not at all. In nature, the expression of many social phenotypes is adjusted in response to environmental cues [48]–[50]. Though these simplifications should be assessed theoretically and empirically, they were critical in allowing us to identify basic physical and biological parameters that control cell group structure and evolution.
In summary, our model suggests that clusters of genetically related cells can emerge quite easily in spatially constrained cell groups, even when cells possess no mechanism for actively gathering with clonemates. Lineage segregation allows cooperative cells to outcompete exploitative cells, and accordingly we predict that localized cooperation will evolve more readily in cell groups than suggested by models and experiments that only consider liquid environments.
We simulate cell groups using an individual-based model described in detail previously [31]. Simulation parameters are listed in Table S1 (Supporting Information). Cell growth is a function of the local microenvironment, namely the concentrations of solutes such as growth substrate (G) and extracellular enzyme (E) (Supporting Information, Table S2). The uptake of growth substrate by each cell is considered when calculating the spatial gradients of substrate concentration. We achieve this by solving a reaction-diffusion equation, where r is a growth rate expression:(2)
Following the common assumption that reaction-diffusion is much faster than cell growth and division [31], our simulations proceed according to the following steps:
The individual-based simulation framework was written in the Java programming language, and its related numerical methods are detailed elsewhere [31]. Briefly, they include the Euler method to grow cells at each iteration, a hard-sphere collision detection method to identify pushing events between neighboring agents, and the FAS multigrid to solve reaction-diffusion equations to steady state [51]. The 3D images in Figure S1 where rendered using POV-Ray. All other figures were prepared using Matlab (the Mathworks, Inc.). The computations in this paper were run on the Odyssey cluster supported by the Harvard University FAS Research Computing Group.
To obtain the segregation index for a cell group at a single point in time, we first identify every actively growing cell. These M cells are indexed by Ai: A1, A2, …, AM. To measure segregation with respect to a single focal cell Ai, we identify all other individuals within a distance of 10 cell lengths. The N cells in this neighborhood are indexed by aj: a1, a2, …, aN.
We define a genetic identity function, g(aj):(3)and a metabolic activity function, m(aj):(4)where [G] is the local concentration of growth substrate, and KG is the half-saturation constant for cell growth rate.
Segregation with respect to a focal cell, s(Ai), is calculated as the mean product of the g and m functions for every cell in its neighborhood:(5)
Finally, we define the segregation index for the entire cell group as the mean value of s(Ai) across all metabolically active cells:(6)
Our segregation index measures the degree to which co-localized, metabolically active cells are clonally related to each other. The index is equal to a form of the relatedness coefficient from social evolution theory under the following assumptions: 1) A cell expressing the cooperative phenotype equally benefits all other individuals within a 10 cell-length radius; 2) Each cell within range of receiving cooperative benefits makes a contribution to mean relatedness proportional to its growth rate; 3) Cell groups are seeded randomly from a large population pool.
The dimensionless number, δ, is a proxy for the depth to which growth substrate penetrates into a cell group before being depleted by cell metabolic activity. δ is derived by non-dimensionalizing Equation 2. We normalize growth substrate concentration by its bulk liquid concentration, , and local biomass by cell biomass density, x = X/ρ. We then normalize the space coordinates by the height of the boundary layer, h. The steady state, dimensionless version of Equation 2 becomes:(7)
Note that the factor multiplying the Laplacian of , , is the square of δ as defined in the main text. δ is also the inverse of the Thiele modulus [52], a number commonly used in chemical engineering to quantify the activity of solid catalysts.
We calculate the competitive fitness of each cell line as the mean number of rounds of cell division per unit time that each achieves over the course of a simulation:(8)where NS,t is the number of cells of strain S present within the cell group at time t. The relative fitness of a strain S1 in local competition with another strain S2 is defined as: .
|
10.1371/journal.pbio.1002180 | A Sensitive and Specific Neural Signature for Picture-Induced Negative Affect | Neuroimaging has identified many correlates of emotion but has not yet yielded brain representations predictive of the intensity of emotional experiences in individuals. We used machine learning to identify a sensitive and specific signature of emotional responses to aversive images. This signature predicted the intensity of negative emotion in individual participants in cross validation (n =121) and test (n = 61) samples (high–low emotion = 93.5% accuracy). It was unresponsive to physical pain (emotion–pain = 92% discriminative accuracy), demonstrating that it is not a representation of generalized arousal or salience. The signature was comprised of mesoscale patterns spanning multiple cortical and subcortical systems, with no single system necessary or sufficient for predicting experience. Furthermore, it was not reducible to activity in traditional “emotion-related” regions (e.g., amygdala, insula) or resting-state networks (e.g., “salience,” “default mode”). Overall, this work identifies differentiable neural components of negative emotion and pain, providing a basis for new, brain-based taxonomies of affective processes.
| Emotions are an important aspect of human experience and behavior; yet, we do not have a clear understanding of how they are processed in the brain. We have identified a neural signature of negative emotion—a neural activation pattern distributed across the brain that accurately predicts how negative a person will feel after viewing an aversive image. This pattern encompasses multiple brain subnetworks in the cortex and subcortex. This neural activation pattern dramatically outperforms other brain indicators of emotion based on activation in individual regions (e.g., amygdala, insula, and anterior cingulate) as well as networks of regions (e.g., limbic and “salience” networks). In addition, no single subnetwork is necessary or sufficient for accurately determining the intensity and type of affective response. Finally, this pattern appears to be specific to picture-induced negative affect, as it did not respond to at least one other aversive experience: painful heat. Together, these results provide a neurophysiological marker for feelings induced by a widely used probe of negative affect and suggest that brain imaging has the potential to accurately uncover how someone is feeling based purely on measures of brain activity.
| Emotions are a class of psychological states comprised of physiological responses, expressive behavior, and subjective experiences that are central to our daily lives and to multiple forms of psychopathology [1] and chronic medical diseases [2]. Emotional information organizes physiological, cognitive, and motor systems into adaptive [3], organism-wide responses to events and situations relevant for survival and well-being [4–6]. These responses allow us to pursue resources and avoid harm [7], translate cognitive goals into motivated behavior [8], and navigate the social world [9,10]. Conversely, emotional dysregulation is at the heart of many brain- and body-related disorders (e.g., mood, anxiety, personality, cardiovascular, and substance use disorders) and likely cuts across traditional diagnostic boundaries [11]. Thus, understanding the neurobiological mechanisms that generate and mitigate negative emotional experience is paramount to understanding both human flourishing and dysfunction.
The importance of understanding the “emotional brain” has motivated hundreds of neuroimaging studies in healthy humans [12,13] and those suffering from psychopathology [14–16]. The promise of these studies for basic research is that they will permit a brain-based taxonomy of emotional processes, avoiding the sole reliance on psychological categories [17,18], while the hope for clinical development is to provide transdiagnostic markers for psychopathology that can identify functional brain dysregulation [19] and physical health risk [2,20], predict treatment response [21,22], and guide new, brain-based treatments [23,24].
In spite of this promise, fundamental requirements must be met before neuroimaging findings can be considered brain representations of emotion that are useful for translational purposes [25]. Previous work has identified many brain correlates of emotional versus nonemotional stimuli [12] and physiological responses [26,27] but has yet to uncover brain signatures diagnostic of an individual’s emotional experience. For example, the amygdala, dorsal anterior cingulate (dACC), anterior insula (aINS), and other regions reliably respond to aversive stimuli [28], and functional alterations in these regions are considered prominent features of anxiety disorders [14,29]. However, activation in these regions does not imply an emotional experience. Amygdala activation can occur in the absence of emotional experience [30] and does not appear to be involved in all aversive experiences [31]. In addition, the dACC and aINS are among the most frequently activated regions in the brain across all types of emotional and nonemotional states [28] and have recently been conceptualized as network “hubs” that may be integrating cognitive, emotional, and motivational information [32,33].
One factor that contributes to this limitation is that the vast majority of studies focus on comparing types of stimuli [12], e.g., “negative” versus “neutral” images, rather than finer grained differences in reported experience [34]. While these emotion-related comparisons are assumed to reflect “affective processing,” confounds with attention, salience, and other processes may render many findings superfluous to emotional experience.
Thus, there is a pressing need for neural signatures that are optimized to predict emotional experiences and functional outcomes. These indicators should: (1) specify a precise set of brain voxels that can be tested in new individuals and prospectively applied to new samples and (2) be sensitive and specific to a class of affective experiences (e.g., negative emotion and not other states such as attention or arousal) [35].
Machine learning provides a new toolbox of algorithms suited for developing sensitive and specific signatures of psychological processes [36–39], particularly when those signatures involve measures across multiple neural systems, as is likely to be the case with emotional experience [12,18,40]. Standard neuroimaging methods generally preclude estimation and optimization of the strength of the brain experience correspondence [28,41–43], but cross validated machine learning analyses can identify whether brain effects are of sufficient magnitude (e.g., sensitive enough) and specific enough to have translational utility. These techniques have recently shown great promise in identifying patterns that discriminate among types of affective experiences from brain [35,44–46] and physiology [47], discriminating patient from control groups [19,48], and predicting treatment response [49].
Here, we use machine learning in a large sample (n = 183) to identify the brain systems that predict the intensity of negative affective experiences elicited by viewing images from the International Affective Picture System (IAPS) [50], which is among the most robust methods of eliciting brief affective experiences (d = 0.81) [51]. In spite of the widespread use of IAPS images in basic and clinical research (e.g., it is the primary affective task in the human connectome project [52]), the brain mechanisms that underlie the genesis of the negative experiences they evoke have not been clearly identified. In addition, it is unclear (a) whether it is possible to identify a pattern that strongly predicts emotional experience prospectively in out-of-sample individuals, (b) which brain systems are involved (cortical, subcortical, or both), and (c) whether brain activity that tracks negative affect is specific for negative affect, or whether it codes for “salience,” arousal, or more general features of stimulus processing. Answers to all of these questions are critical for continued progress in both basic affective and clinical sciences.
We address each of these questions by developing a multivariate pattern that predicts negative emotion and assess its sensitivity and specificity relative to pain—another type of arousing, salient, negative experience. Finally, to examine the distributed versus localized nature of the signature, we examined the subsystems necessary and sufficient for accurately predicting negative emotional experience.
We used Least Absolute Shrinkage and Selection Operator and Principle Components Regression (LASSO-PCR) [35,53] to identify a distributed Picture Induced Negative Emotion Signature (PINES) that monotonically increased with increasing affective ratings in leave-one-subject-out cross validated analyses (n = 121). To apply the model to data from individual test subjects in both cross validation (n = 121) and separate hold-out test datasets (n = 61), we calculated the pattern response—the dot product of the PINES weight map and the test image—for individual subjects’ activation maps for each of 5 levels of reported negative emotion (see Fig 1). The resulting continuous values reflect the predicted intensity of negative emotion for a given activation map. We used these values to classify which of two conditions elicited a stronger negative emotion for an individual (a “forced-choice” test) [35], providing accuracy estimates (Fig 1E). We also used similar classification tests, described below, to evaluate the sensitivity and specificity of PINES responses to negative emotion versus pain. We focus primarily on results for the test sample, as it was completely independent of all model-training procedures and provides the strongest evidence for generalizability [54].
The PINES accurately predicted ratings of negative emotional experience in both cross validation and hold-out test datasets (Fig 2). For individual participants in the cross validation sample, the average root mean squared error (RMSE) was 1.23 ± 0.06 (standard error; SE) rating units, and the average within-subject correlation between predicted and actual ratings was r = 0.85 ± 0.02). Accuracy was comparable in the test sample (RMSE = 0.99 ± 0.07, r = 0.92 ± 0.01). The PINES accurately classified highly aversive (rating 5) versus nonaversive (rating 1) pictures with 100% forced-choice accuracy in both cross validation and test samples (Fig 2B). Classification accuracy was also high in both the highly aversive range (rating of 5 versus 3: forced-choice = 91%; test sample) and the moderately aversive range (rating of 3 versus 1: 100%; test sample) (See S1 Table). We also assessed single-interval classification based on a single image rather than a relative comparison (Table 1), which were only slightly less accurate (Table 1). Comparisons with Support Vector Regression (SVR), another popular algorithm, indicate that these results appear to be robust to the choice of algorithm and, to a large extent, the amount of data used in the training procedure (see S1 Methods).
The PINES pattern included reliable predictive weights across a number of cortical and subcortical regions (Fig 2A). Positive weights (greater activity predicts more negative emotion) were found in many regions typically associated with negative emotion [12,40], including the amygdala, periaqueductal gray (PAG), aINS, dorsomedial prefrontal cortex (dmPFC), ventral occipital cortex, presupplementary motor area (preSMA), ventromedial temporal lobe (mTL), and posterior cingulate cortex (PCC). Negative weights were found in the bilateral parahippocampal gyrus, right superior temporal gyrus, left temporal parietal junction (TPJ), right caudate, and occipital and somatomotor cortices. These regions likely comprise multiple functional systems, as we describe in more detail below. Though the PINES comprises nonzero predictive weights across the brain (see S1 Fig), supplementary analyses indicated that a sparse pattern thresholded at p < .001, as shown in Fig 2 (1.6% of in-brain voxels), was sufficient to predict emotional experience with comparable sensitivity to the full model (see S1 Methods and S5 Fig).
Affect systems may be organized by valence so that a brain signature for negative affect may be found across stimulus modalities and contexts, or in a modality-specific manner, such that there is not one “negative affect system” but many. Testing these hypotheses requires comparing multiple types of negative affect across modalities. Here, we assessed the generalizability and specificity of the PINES response across IAPS pictures and somatic pain, which is a negative, arousing experience arising from a different modality.
We employed two types of analyses to examine the PINES specificity. First, we compared the spatial topography of the PINES to another pattern map, the Neurologic Pain Signature (NPS), which shows high sensitivity and specificity to somatic pain across multiple studies [35]. The PINES and NPS maps were almost completely uncorrelated (robust ranked spatial correlation, ρ^ = −0.01; Fig 4). Several regions showed positive weights in both maps, including the anterior cingulate (ACC), insula, and amygdala. As shown in Fig 5C, however, the weight patterns within these regions were also uncorrelated (bilateral ACC, ρ^ = 0.04, insula, ρ^ = −0.05), though weights in the amygdala were modestly correlated (ρ^ = 0.21).
Second, we assessed the specificity of the pattern responses in the test IAPS (n = 61) and thermal pain (n = 28) [56] datasets. The PINES accurately predicted negative affect in the IAPS dataset (n = 61) but showed no response to increasing pain intensity in the pain dataset (Fig 4). Conversely, the NPS responded robustly to increasing pain but showed no response to increasing negative affect in the IAPS dataset. To further assess sensitivity and specificity, we examined how well responses in each pattern could discriminate (a) high pain versus high negative affect, (b) high versus low pain, and (c) high versus low negative affect (Table 1). Because this involves comparing responses from two separate, imbalanced test sets (n = 61 versus n = 28), the analyses described below employ single interval classification, in which individual images are tested for suprathreshold responses independently (as compared to relative within-subject differences in forced-choice classification). The threshold was determined by finding the point that minimized signal detection response bias (see Methods for details), and we report balanced emotion classification accuracy (chance = 50%), sensitivity, and specificity (See S2 Table for equivalent forced-choice analyses).
Another question is whether the precise pattern of activity specified in the PINES uniquely captures negative affect, or whether regions and networks previously used in the literature are sufficient. In order to fully appreciate the sensitivity and specificity of the PINES, it is necessary to compare it to the standard univariate approach, which typically examines average activation within ROIs compared to baseline activity. In this analysis, we examined the average response to emotion and pain stimuli within anatomical ROIs and canonical networks defined in large-scale resting-state studies [57].
Defining a brain pattern sensitive and specific to a type of negative emotion is a critical first step towards developing meaningful models of brain representations of emotion. Here, the development of the PINES affords the opportunity to characterize the basis of this pattern representation within and across brain networks. Constructionist theories of emotion [12,18] predict that negative affect is created by interactions among discrete subnetworks that span multiple brain systems, whereas more traditional modular views predict that one system may be sufficient. We tested whether the PINES might be composed of multiple distinct subnetworks and whether responses in multiple subnetworks are necessary for predicting emotional responses. If so, the negative affect captured by the PINES might be considered a truly multisystem distributed process.
For this analysis, we calculated pattern responses within each of the largest regions in the PINES (p < .001, k = 10 voxels; see S1 Methods) for every individual trial within each participant and used a robust clustering algorithm to group the PINES regions into separate networks based on similar patterns of trial-by-trial covariation (see Methods). The best solution contained nine separate clusters, which provides a descriptive characterization of the subnetworks that comprise the PINES (Fig 6, S3 Table) that is broadly consistent with constructionist accounts of emotion [12] and previous meta-analyses of emotion-related networks [17]. These subnetworks included (a) two networks encompassing different parts of the visual cortex (e.g., lateral occipital cortex [LOC] and occipital pole) consistent with the visual modality of the stimuli, (b) a left amygdala-right aINS-right putamen network, which has been implicated in multiple forms of arousal and salience, (c) a network that includes bilateral posterior parahippocampi and the precuneus, which are broadly involved in memory and other forms of contextual processing, and (d) a network that includes parts of the dmPFC and PCC that are likely involved in social cognition but are distinct from more executive processes [64,65]. An additional network that includes the right somatosensory cortex and contralateral cerebellum may be involved in preparing for the rating action but may also play a more fundamental role in the emotion generation process [66].
For neuroimaging to be useful in translational applications (e.g., psychiatry, neurology, etc.), sensitive and specific brain signatures must be developed that can be applied to individual people to yield information about their emotional experiences, neuropathology, or treatment prognosis [25]. Thus far, the neuroscience of emotion has yielded many important results but no such indicators for emotional experiences. Signatures that are sensitive and specific for particular affective processes are presumably much closer to brain representations of emotional experience, which can then be interrogated to better understand the mechanisms and typology of emotion at the neurophysiological level.
The goals of the present study were to: (a) develop a brain signature capable of reliably and accurately predicting the intensity of negative emotional responses to evocative images, (b) characterize the signature’s performance in generalizing across individual participants and images, (c) examine its specificity related to another negative and arousing affective experience (pain), and (d) explore the structure of the subnetworks necessary and sufficient to predict negative emotional experience.
We used cross validated machine learning analyses to identify a distributed pattern of activity predictive of emotional experiences, which we term PINES. The PINES fulfills the basic criteria for a brain signature of negative affect. It accurately predicted monotonic increases in negative affect ratings in 93.5% of individual test participants (n = 61; single interval). In forced-choice tests, it correctly identified which of two sets of images was rated as more negative in 90%–100% of individuals, as long as the images differed by two or more subjective rating points (on a five point scale). This demonstrates sensitivity to negative affect across the full range of the measurement scale.
PINES responses were also surprisingly specific to negative emotion. The PINES did not respond to increased levels of physical pain, another type of arousing, aversive, salient experience. Conversely, the NPS [35]—a signature previously found to be sensitive and specific to physical pain—responded strongly to physical pain but not to increasing levels of picture-induced emotional intensity. This double dissociation implies that neither pattern is driven by general arousal, salience, or negative affect. Though the PINES and NPS independently tracked the intensity of negative affect elicited in visual and somatosensory modalities, respectively, we do not believe they are dissociable based simply on differences in sensory processing for two reasons: (1) the PINES was just as accurate in predicting negative emotion without the occipital lobe, and when subnetworks associated with modality-specific processes were removed (e.g., visual, somatosensory, etc.) and (2) the local PINES and NPS patterns within traditional “affect” regions, such as the ACC and insula, were uncorrelated. This is consistent with our previous work demonstrating that pain is distinct from other emotional processes based on distributed spatial topography both across brain regions [28] and within local regions [35,67].
Further analyses explored the nature of emotion-predictive brain representations. The PINES was comprised of multiple separable subnetworks. Each network independently contributed to the prediction of participants’ negative emotion ratings controlling for other brain regions, and no single network was necessary or sufficient for predicting emotional experience. This pattern of results suggests that the PINES is a distributed pattern that encompasses a number of functional systems, and that multiple systems are required to capture negative affective experience.
These results have theoretical implications for the neurobiology of emotion in terms of both the diversity of processes underlying affective experiences and how they are represented in the brain. Emotions are often defined as a composite of multiple intrinsically inter-related processes (e.g., autonomic arousal, expressive behavior, action tendencies, interoception, and conscious experiences). Theories differ widely on how these processes combine to give rise to emotional experience [1], but most major theories suggest that cognitive, sensory, motor, motivational, and interoceptive processes are critical ingredients of emotional experience. For example, appraisal theories view emotion as a dynamically unfolding process and emphasize the role of appraisals [8,68,69], embodied affect theories emphasize interoceptive and somatomotor representations [70], and constructionist theories view emotions as being constructed from all of these component processes [12,18].
In spite of this richness, since MacLean [71], theories of the emotional brain have treated emotion as a singular faculty that is localizable to a specific system. Often, this view has translated into “structure-centric” theories of emotional experience; e.g., the amygdala is critical for fear [72], the ACC for pain affect [73], and the insula for disgust [74]. In other cases, this view translates into circumscribed pathways or networks for “core affect” [17] and emotional awareness [75].
It remains unclear how far the structure-centric view can take us in understanding the brain bases of emotional experience. The regions most strongly identified with emotion are also intimately involved in a wide array of cognitive functions such as attention, error monitoring, associative learning, and executive control [33]. Recent connectivity [63] and functional diversity analyses [32] suggest that these regions are not solely processing affective signals but rather represent functional “hubs” for integrating many types of information.
As the limitations of the structure-centric view are increasingly widely recognized [12,33], researchers have moved towards the identification of intrinsically connected networks conserved both at rest and during active tasks [76]. These networks have been labeled with process-general names including the “salience network” [63], “default mode” network [77], and others, and a modern incarnation of the “emotional brain” theory suggests that the basis of emotional experience is encapsulated in one or a few of these networks such as the “limbic” network named after MacLean’s original formulation.
Our results corroborate the view that structure-centric—and even network-centric—models of emotion are limited and provide an alternative model for the brain representation of emotional experience. In this study, we targeted the conscious experience component, which is the defining feature of subjective distress and suffering. None of the anatomical regions identified in previous literature (e.g., amygdala, ACC, insula) predicted the intensity of emotional experience or discriminated emotion from pain in this study. This suggests that the effects identified in previous work using traditional statistical parametric mapping approaches are small and unlikely to serve as effective signatures of the type or magnitude of an emotional experience in an individual person.
Furthermore, activity in predefined networks was insufficient to capture negative emotion ratings, demonstrating that the pattern we identified using targeted machine-learning analysis is not reducible to these more process-general networks. The fact that networks and regions defined a priori, even from very large resting-state samples [57], were insufficient to capture emotional experience here has broad implications for the study of emotion and attempts to identify biomarkers for mental health disorders going forward [21,25,49].
Finally, our clustering analysis of the PINES map indicated that multiple, separable subnetworks distributed widely throughout the brain made independent contributions to predicting emotional experience. Importantly, no single subnetwork appeared to be necessary or sufficient in characterizing the emotional experience, as the accuracy in predicting the magnitude or type of experience did not significantly decrease when any given network was omitted. This pattern is consistent with both appraisal [68,69] and constructionist theories of emotion [12,78], which posit that emotional experiences result from interactions between core affect, sensory, memory, motor, and cognitive systems [40].
Overall, these results provide an important step towards identifying emotion-related patterns that can serve as indicators for components of emotional experience. Such signatures can be used as intermediate phenotypes for genetic or risk-stratification studies, and they may provide objective neurobiological measures that can supplement self-report. The identification of intermediate brain-based phenotypes is critical, as self-reported emotion can be affected by many independent processes [8,18,68]—e.g., core experience, self-reflection, decision-making heuristics, and communicative intentions—which have different implications for understanding what exactly treatments that modulate emotion are measuring and which processes are affected by interventions.
We close with several key points and future directions. Importantly, the PINES is not necessarily a biomarker of negative emotion in general. We have demonstrated that it is a signature for the type of affect induced by aversive IAPS images, but its transferability to other emotional states (e.g., emotion induced by recall, rejection, positive emotion, or stress) remains to be tested. Such tests are a long-term program of future research that must span many studies and papers. We still know very little about the underlying structure of affect and which types of emotional responses can be cross predicted by the same brain markers. It is possible that the PINES captures some types of negative emotion and not others, and findings to this effect will help us move beyond the categories proscribed in our language to develop a more nuanced, brain-based view of affective processes [7,17].
In addition, testing the specificity and transfer of the PINES across many different kinds of affect is a key to developing more robust and specific markers. The PINES can undoubtedly be improved. For example, with further development and testing, it may be differentiated into markers for more specific types of emotional experiences (e.g., emotion categories like fear, disgust, etc. or canonical affect-inducing appraisals). In addition to types of affect, the PINES can be tested for responses across patient groups (e.g., schizophrenia, depression, or anxiety) and treatments thought to affect emotion (e.g., self-regulation, drug treatment, psychotherapy, etc.). This study provides a foundation and a benchmark for such future developments.
All participants provided written informed consent, and experimental procedures were approved by the Institutional Review Board of the University of Pittsburgh for the IAPS study and the University of Colorado, Boulder for the pain study.
|
10.1371/journal.pntd.0001212 | Polyethyleneimine (PEI) Mediated siRNA Gene Silencing in the Schistosoma mansoni Snail Host, Biomphalaria glabrata | An in vivo, non-invasive technique for gene silencing by RNA interference (RNAi) in the snail, Biomphalaria glabrata, has been developed using cationic polymer polyethyleneimine (PEI) mediated delivery of long double-stranded (ds) and small interfering (si) RNA. Cellular delivery was evaluated and optimized by using a ‘mock’ fluorescent siRNA. Subsequently, we used the method to suppress expression of Cathepsin B (CathB) with either the corresponding siRNA or dsRNA of this transcript. In addition, the knockdown of peroxiredoxin (Prx) at both RNA and protein levels was achieved with the PEI-mediated soaking method. B. glabrata is an important snail host for the transmission of the parasitic digenean platyhelminth, Schistosoma mansoni that causes schistosomiasis in the neotropics. Progress is being made to realize the genome sequence of the snail and to uncover gene expression profiles and cellular pathways that enable the snail to either prevent or sustain an infection. Using PEI complexes, a convenient soaking method has been developed, enabling functional gene knockdown studies with either dsRNA or siRNA. The protocol developed offers a first whole organism method for host-parasite gene function studies needed to identify key mechanisms required for parasite development in the snail host, which ultimately are needed as points for disrupting this parasite mediated disease.
| Freshwater snails are important in the transmission of schistosomiasis. As part of an integral control effort to combat the spread of schistosomiasis new intervention tools are being sought. One method is to interrupt the transmission of the causative schistosome parasite during the intra-molluscan phase of its development. Gene-silencing technology involving the use of dsRNA have used an injection route to disrupt gene translation in the Schistosoma mansoni snail host, Biomphalaria glabrata in an effort to investigate how inhibition of various transcripts can affect the dynamics of the snail/parasite interaction. These studies have been helpful in showing us that a gene-silencing pathway that uses dsRNA indeed exists in snails but the injection method previously utilized is impractical, especially when working with juvenile snails. To make the use of gene silencing technology more widely applicable to functional gene studies in snails, we have developed a more convenient soaking method that uses a cationic carrier polyethylene amine (PEI) to deliver dsRNA or siRNA into juvenile snails. Using this method we show the successful knockdown at both RNA and protein levels of the B. glabrata peroxiredoxin (Prx) gene. The method was also evaluated for silencing the Cathepsin B (CathB) gene in the snail.
| Biomphalaria glabrata is an intermediate snail host that transmits the digenean platyhelminth parasite, Schistosoma mansoni, in the Western Hemisphere. This snail host is easily maintained in the laboratory outside of its natural environment and, therefore, serves as a useful model organism for conducting studies aimed at unraveling the complex biology that underlies the snail's relationship with the schistosome parasite.
Schistosomiasis, the disease caused by the parasite, is prevalent in several countries of the developing world where it is estimated that at least 200 million people are chronically infected [1], [2]. Significant progress is being made towards the identification of genes that govern the snail/schistosome interaction. Accordingly, several genes that are either up or down regulated in the snail, early after parasite infection have now been identified [3], [4], [5]. Among these are genes involved in innate defense and stress response. A genome project for B. glabrata is also near completion [6]. It is hoped that all these advances will lead eventually to the development of novel tools for halting infection at the snail stage of the parasite's life cycle. For this disease transmission blocking strategy to come to fruition, however, we need a better understanding of what genes/cellular pathways in the snail host can be interfered with to bring about subsequent disruption of the parasite's development.
To investigate what gene expression and/or molecular pathways are involved in the snail host/parasite relationship, either enabling or disabling a viable schistosome infection, the technology of RNA interference (RNAi) to specifically silence gene expression in the snail host should help to uncover genes/pathways (in the snail host) that are essential for schistosome development. Fundamentally, it is also possible to envision that this technology might help us to identify conserved molecular pathways that are utilized by the parasite for its survival in both snail and definitive hosts, providing us with an alternative approach towards the identification of new targets for either drug or vaccine development.
All previous studies that have reported successful gene -silencing by RNAi technology in mollusks have been accomplished by an injection approach. For instance, in 2006 by Jiang et al. [7] were able to knockdown the expression of the snail defense lectin gene, FREP 2 by directly injecting the corresponding dsRNA of this molecule into the snail hemolymph. Similarly, in another pulmonate gastropod, Lymnaea stagnalis, the injection of dsRNA corresponding to the transcript for nitric oxide synthase in this pond snail was shown to cause the reduction in expression of this gene, affecting the feeding behavior of treated snails [8]. Both these studies paved the way in showing that a post-transcriptional gene silencing (PTGS) mechanism, mediated by RNA interference, was indeed operational in freshwater snails, as had been extensively shown in other organisms. Recently, the specific knockdown of the B. glabrata ortholog of Macrophage Migration Inhibitory Factor (MIF) was demonstrated at the protein level by injecting the corresponding dsRNA of this molecule into the snail, making this the first time that RNAi technology has been shown to suppress protein function in this snail [9]. In the very few RNAi gene-silencing studies that have been performed in mollusks only one thus far has used siRNA, not dsRNA, for mediating the suppression of specific gene expression. Thus in this recent study, Hannington et al. [4] were able to show the knockdown of the protein expression of FREP 3 with a concomitant increase in snail susceptibility, demonstrating the functional role of this gene in snail innate immunity.
Since the discovery was made several years ago of the existence of a dsRNA mediated PTGS pathway in the cell, the knockdown of specific genes, using either their corresponding dsRNA or siRNA, to study gene-function has grown exponentially. In schistosomes, for example, the technique has now been used widely to demonstrate the importance of several key genes whose function enables optimum development of larval and adult worms [10], [11]. Furthermore, key parasite enzymes belonging to this gene-silencing network are being cloned and characterized [12]. Contrary to these significant milestones that have been achieved in the parasite, in the snail host, however, virtually no information exists on how this PTGS pathway operates to regulate gene expression. One exception to this paucity of data is the recent identification and mapping, by fluorescent in situ hybridization (FISH), of the B. glabrata homolog of P-element induced wimpy testis, piwi, a protein that is involved in siRNA mediated silencing of mobile elements in germ cells [13].
Polyethylenimine (PEI) is a cationic polymer that has been widely used as a carrier for the delivery of DNA, dsRNA and proteins into cells [14], [15], [16]. Because the snail secretes mucus, a complex high molecular weight protein glycoconjugate substance that is used for locomotion, as well as forming a natural protective barrier around the organism [17], it is difficult to deliver nucleic acids in solution directly into mollusks. In order to deliver dsRNA through the negatively charged mucus into the snail, we used PEI as a carrier. Linear (jPEI) and non-linear (branched) PEIs were tested for delivery. The decision to evaluate PEI as a possible carrier for the delivery of dsRNA, or siRNA, into the snail was based on prior experience of using this cationic matrix for separating small oligonucleotides [18]. Furthermore, the synthetic PEI polymer has been shown to be a good non-viral delivery system for studying therapeutic siRNA mediated gene silencing in several animal models. The advantage of using PEI is that the complex that is formed with siRNA is a nanoparticle that is protected from degradation (reviewed by Gunther et al) [19]. Initially, the PEI delivery method for the snail was optimized with ‘mock’ fluorescent siRNA. Then the selected method was used to evaluate specific knockdown, in a ‘proof of principle’ experiment, of transcripts encoding B. glabrata homologs of Cathepsin B (CathB) [20] and peroxiredoxin (Prx) [21]. Using siRNA and dsRNA corresponding to these genes, in combination with linear (jet) PEI, we were able to show specific knockdown of these transcripts in the snail host.
Juvenile B. glabrata snails of the NMRI stock (2–3 mm in diameter) were used for the study. The snails were maintained in de-chlorinated tap water at room temperature and fed on Romaine lettuce as previously described [22], [23]. Before either siRNA or dsRNA delivery, snails were kept overnight in sterile H2O without feeding.
RNA was isolated from the whole snail as previously described [24]. To prepare dsRNA for the Prx (Acc. No. FJ176942) and CathB (Acc.no. EU035711) transcripts, we designed primers corresponding to the transcript sequence that additionally contained sequences of T7 polymerase in cis. Forward and reverse primer sequences were as follows: (Prx forward primer; spanned position 162–173) 5′-taatacgactcaccCGTCATTTCTAA-3′, (Prx reverse primer; spanned position 539–550) 5′-taatacgactcaccCTTTGGGGTCAA-3′, (CathB forward primer; spanned position 381–392) 5′-taatacgactcaccATGACTGACAGA-3′ and (CathB reverse primer; spanned position 851–862) 5′-taatacgactcaccTTCACTGCGTGT-3′. In this study dsRNA (Myo dsRNA) for myoglobin (Acc.no. U89283) was also prepared and used as mock dsRNA control. The myoglobin primers utilized for dsRNA preparation were as follows: forward primer (spanned position 1229–1240) 5′-taatacgactcaccGGCAAAAAGAAC-3′ and reverse primer (spanned position 3112–3123) 5′-taatacgactcaccAAGAGCACTTTC-3′. The primers (forward and reverse) were designed to encompass two contiguous exons of the myoglobin transcript [25]. To prepare DNA templates for dsRNA synthesis, cDNA made by using total RNA from the whole snail, as previously described [26] was amplified by PCR for 30 cycles with the following conditions: denaturation at 95°C for 30 seconds, annealing at 55°C for 30 seconds, and extension at 72°C for 1 minute, for 30 cycles. Amplicons of the expected size (397 bp) encompassing positions 161 bp to 558 bp of the BS-90 stock Prx sequence (accession no. FJ176942, [21]) were purified by using either ‘gene clean’ (MPBIO, Solon, OH) or by using a Sephacryl S-200 microspin column (GE Healthcare, UK) according to the manufacturer's instructions. Purified templates were examined qualitatively by agarose gel electrophoresis in TBE buffer as previously described [24] and quantified by a NanoDrop 1000 Spectrophotometer (Thermo Scientific). Templates (cDNA) used for dsRNA synthesis were stored at −20°C in sterile H2O until they were required. Amplicons synthesized for Cath B and Myo templates were purified, examined quantitatively and re-suspended as described above.
To prepare dsRNA, we used the purified amplified templates from above at a final concentration 20 ng/µl. After denaturing the template for 5 min at 65°C followed by snap -freezing on ice, the following was added to the template on ice in a final volume of 50 µl, 5× transcription buffer (Promega, WI), 2.5 µM rNTPs, 100 mM DTT, 2 µl (80 units) Rnasin (Promega, WI), and 1 µl (19 units) T7 polymerase. Samples were incubated for 2 hrs at 37°C before adding RNAase free DNAase (4 units, RQI, Promega WI), and incubating for a further hour to remove the DNA template. Reactions were inactivated with 1 µl 0.5 M EDTA (pH 8.0), phenol: isoamyl alcohol/chloroform extracted and precipitated with wheat germ transfer RNA (overnight at −20°C) in 2.5× volumes of Ethanol. The dsRNA was recovered by centrifugation at 10,000×g for 15 min at 4°C, and the pellet was washed once in 75% ethanol then air-dried. The dried pellet was re-suspended in sterile de-ionized distilled (dd) H20. The solubilized dsRNA was examined, qualitatively by agarose gel electrophoresis, purified by ‘gene-clean’ (MPBIO, Solon, OH) according to the manufacturer's instructions, quantified by using a Nanodrop spectrophotometer, and stored at −70°C until required.
Two small inhibitory RNAs (siRNAs) for the CathB transcript; CathB1 and CathB2 were designed Silencer Select Custom Designed siRNAs (http://www5.appliedbiosystems.com/tools/sirna/). Both the CathB siRNAs utilized were synthesized by Applied Biosystems. CathB1 and CathB2 siRNAs spanned the CathB gene coding DNA position 94–112 (propeptide C1 region) targeting the sequence 5′-GCGATGCAGAGATCTTCTA′3 and position 850–868 (catalytic region) targeting the sequence 5′-GACACGCAGTGAAGATCAT-3′, respectively. In order to reduce off-target effects, the selected target sequences showed no match with any other B. glabrata and S. mansoni sequences as assessed by Silencer Select Custom Designed siRNAs using their ‘BLAST and Filter’ function. In addition, Silencer Negative control siRNA#1 was also provided by the manufacturer (Ambion, ABI) and used as negative control. Sequence information for accompanying ‘mock’ siRNAs was not provided by the manufacturer. However, according to the manufacturer's instructions, these ‘mock’ siRNAs do not target any gene product. Silencer Negative Control siRNAs are validated for use in human, mouse, and rat cells, and have been functionally tested for producing minimal effects on cell proliferation and viability. Supplied with the ‘Cath B gene specific’ siRNAs, the ‘mock’ siRNA was utilized in all assays to determine the efficiency of the CathB siRNA/PEI transfection, and to rule out non-specific effects of siRNA delivery into the snails. ‘Fluorescence siRNA’ sold as ‘block it’ Alexa 555 (Invitrogen, Life Technologies) was used in this study. According to the manufacturer, the Alexa 555 siRNA sequence is not homologous to any known gene and uptake of this fluorescent siRNA was assessed qualitatively by fluorescent microscopy according to the manufacturer's instructions. Each snail treated with the fluorescent labeled-siRNA was kept in the dark throughout the experimental procedure. Although siRNA and dsRNA are, technically both double-stranded RNA, the current convention is to classify them according to the length of nucleotide sequence, with longer oligonucleotides (>21–23 nucleotides) referred to as dsRNA and shorter ones known as siRNA ‘small interfering RNA’.
The linear PEI commercially sold as Jet PEI (jPEI molecular weight 22 kDa) was purchased from PolyPlus (France) and branched PEI (bPEI) with average molecular weight of 25 kDa was purchased from Sigma-Aldrich (MO, USA) Both reagents were kept at 4°C until required. For RNA delivery, we combined the dsRNA (120 ng) with PEI (93.8 ng of linear or branched) to obtain the complex at a PEI nitrogen/Nucleic acid phosphate (N/P) ratio of 6 [27]. A residue weight of 43 for PEI and 330 for siRNA was used for calculating the required amounts of PEI and siRNA or dsRNA.
The required amount of siRNA (775 ng) or dsRNA (120 ng) was diluted to 250 µl of sterile ddH2O in a microcentrifuge tube. The corresponding amount of PEI was also diluted to 250 µl in sterile ddH2O in a separate tube. This PEI solution was added to the tube containing the dsRNA (or siRNA) followed by immediate vortexing for 10 sec. The mixture (PEI plus dsRNA) was allowed to incubate to form nanoparticles at room temperature for 30 min before placing washed snails into the mixture. After placing the snails in the tubes the samples were mixed gently and the caps were closed. Using a hypodermic needle, holes were punched into the caps, and samples were kept at room temperature until the end of the experiment (routinely 72 hrs but also longer for up to 4 days). Snails (6–8 snails for each experiment) were incubated individually in the mixture, and all experimental sets were done in duplicate. Six separate biological replicates using dsRNA/PEI and 5 with siRNA/PEI complexes were performed in total. Because the snails had a tendency to initially crawl out of the PEI/dsRNA mixture, tubes were checked periodically to make sure that snails remained in the solution. Of the two carriers, snails were observed to crawl less frequently out of the linear than branched PEI, consequently we used jPEI for all subsequent specific gene-silencing experiments conducted herein. For each experiment, controls were set up as follows: normal untreated snails, control snails incubated in either dsRNA or siRNA alone, mock Myo dsRNA, mock siRNA (negative control siRNA supplied by the manufacturer), mock siRNA/PEI, and PEI only control. At the end of the experiment, snails were removed from the mixture, washed in sterile H2O and either processed immediately for RNA or protein isolation. Snails for RNA isolation were stored either in the RNA isolating reagent and stored at −20°C until required.
Snails (2 to 3 mm in diameter) incubated as described above in either linear or branched PEI (n = 117) for up to 4 days were monitored daily for mortality compared to normal untreated snails (n = 29).
To determine the uptake of PEI/fluorescent siRNA nanoparticles by microscopy, the outer shell of each snail was removed and the remaining whole body was rinsed twice in sterile water before viewing under the light microscope (magnification ×10). The same image was then examined for fluorescence (green filter). Images were captured using the advanced SPOT camera imaging software (Diagnostic Instruments Inc. Sterling Heights, MI). Fluorescent signal in tissues was detected qualitatively (as presence or absence of fluorescence) from the snail tissue samples.
To monitor the expression of Prx or CathB genes after dsRNA/siRNA treatment in individual snails, total RNA was extracted from the whole snail using RNAzol RT (Molecular Research Center, Inc) following the manufacturer's instructions. With this method, all contaminating residual DNA was eliminated as previously described [26]. Quantitative real-time PCR (qPCR) was performed using Brilliant II SYBR green QPCR master mix, in a one-step reaction (Stratagene, Agilent) according to the manufacturer's instructions and run by using the ABI7300 Real Time PCR System (Applied Biosystems). Twenty-five microliters of each qPCR mixture contained 80 ng RNA, 12.5 µl Brilliant SYBR green PCR Master Mix, 200 nM of each gene specific primer and 1 µl of Blocking Reverse transcriptase. Primer concentrations for each assay were determined after optimization. In order to avoid the possibility of falsely amplifying any of the original dsRNA/siRNA that was applied, specific primers (forward and reverse) were designed outside the region used to synthesize the original interfering dsRNA product. For Prx, primers were: Prx-F 5′-ATGGCATCCTCTCTGCAAACCGGG-3′, Prx-R 5′-TTAGAGTTCATCGTTAGATTGC-3′, CathB-F 5′-AGCAACACCATTCCACATC-3-′ and CathB-R 5′-ATAGCCTCCGTTACATCC-3′. To assess the degree of knockdown achieved, primers (F 5′-GTCTCCCACACTGTACCTATC-3′, R 5′-CGGTCTGCATCTCGTTTT -3′) for the housekeeping gene actin; Acc.no CO501282 [28] were used as the endogenous standard for normalization. Additional controls for verifying the specificity of gene knockdown included; 1) dsRNA or siRNA alone, 2) PEI alone and 3) Myo dsRNA 4) buffer alone. Three technical replicates of qPCR were performed with two internal controls to assess both potential genomic DNA contamination (no reverse transcriptase added) and purity of the reagents used (no template added). As indicated above, the knockdown experiment was repeated multiple times (5×) for CathB siRNA, and (6×) for CathB dsRNA and Prx dsRNA (n = 5–6) as independent biological replicates, and by two different investigators. The difference in gene transcript levels were calculated by the ΔΔCt method [29] using the actin housekeeping gene to normalize the quantification of targets. For graphical representation, the ΔΔCt values were normalized to controls and expressed as the percentage difference.
Since previous studies have shown that the hepatopancreas is the major location of Prx in B. glabrata, soluble protein extracts from individual juvenile snails were prepared from this tissue [21]. Using a sharp scalpel blade, under a dissecting microscope, half of the hepatopancreas was immediately homogenized on ice in sterile PBS (pH 7.5) containing a cocktail of protease and phosphotase inhibitors (Sigma-Aldrich). Soluble protein was isolated from the homogenate as previously described [21]. The remaining dissected tissue (from each snail) was snap frozen in liquid nitrogen and processed for RNA isolation as described above. To determine whether the PEI mediated Prx dsRNA-soaking method would successfully produce concomitant knockdown of corresponding protein in the snail, changes occurring in the amount of Prx between either experimental or control snails were measured quantitatively by ELISA titre using previously described mouse polyclonal antibody prepared against B. glabrata recombinant Prx [21]. Since only limited amount of protein material was available from the multiple number of tissue samples processed, we found using ELISA instead of Western Blots [30] to determine the degree of knockdown at the protein level to be more practical, less labor intensive, and also cost-effective. For ELISA assays, 100 µl per well of 10 µg/ml of soluble protein in coating buffer (100 mM bicarbonate/carbonate buffer, pH 9.6) was coated in a MaxiSorpflat-bottom 96 well plate (Nunc) at 37°C overnight, then unbound (excess) proteins were washed out 3 times with washing buffer (0.01% gelatin, 0.05% BSA, 0.05% Tween 20 in 100 mM phosphate buffer saline [PBS], pH. 7.2). The remaining binding sites (that would yield false positive results) were blocked by using 1% BSA in PBS at 37°C for 30 min, then washed (3×) in washing buffer as described above. Subsequently, 10-fold dilution (ELISA titres varied from 100–204,800) of antibody either polyclonal antibody of Prx or CathB (100 µl/well) was added into each well and incubated at 37°C for 1 hr. After washing, each well was incubated with donkey anti-rabbit IgG (H&L) HRP (Promega, MA) for Cath B or goat anti-mouse IgG-HRP (Jackson ImmunoResearch Laboratory, PA) for Prx according to the manufacturer's instructions. The colorimetric enzymatic assay was determined by incubation at RT for 30 min with 3-ethylbenzthiazoline-6-sulfonate by ABTS Microwell Peroxidase Substrate System (Kirkegaard&Perry Laboratories, MD), 100 µl/well. The reaction was stopped with 100 µl of 1% SDS solution, and absorbance at 405 nm was read by using a microplate reader (BioRad Laboratories). All assays were done in triplicate and included the following controls: 1) protein extract incubated with only buffer (no antibody negative control), 2) buffer only incubated with antibody (no antigen negative control), Prx recombinant protein incubated with antibody (positive control). The negative control using the protein extract without antibody was to eliminate the possibility that false positive results might occur due to endogenous peroxidase activity present in the extract that might interfere with the assay. These data are expressed as end point titre, with the titre being defined as the highest dilution that yielded an optical density reading greater than twice the background values. The titres were calculated after subtracting the mean absorbance of triplicate wells lacking antigen protein from the absorbance of triplicate wells containing antigen at each antibody dilution. In cases where snails were treated with CathB siRNA or CathB dsRNA commercial rabbit anti-Cathepsin B polyclonal antibody [CathB Ab] (MyBioSource, CA) and anti-Sm31 polyclonal antibody (anti-S.mansoni CathB) were used to determine changes in this protein in treated compared to untreated snails.
To determine whether PEI can mediate the uptake of siRNA into the snail, we tested separately, linear PEI (jPEI) and branched PEI (bPEI) with fluorescent Alexa 555 tagged siRNA, as described in Materials and Methods. Figure 1 shows images of the snail's hepatopancreas and ovotestis regions visualized either under light (A and C), or fluorescent microscopy (B and D) after snails have been incubated for either 24 or 72 hrs at room temperature in fluorescent Alexa siRNA complexed jPEI (Alexa555 siRNA/jPEI nanoparticles). Images of snails incubated in either fluorescent Alexa siRNA alone or untreated normal snails (hepatopancreas and ovotestis regions) also viewed under either light (A and C) or by fluorescent microscopy (B and D) are shown in Figure 2.
In Figure 1, by using fluorescent microscopy (panels B and D), we detected dramatic uptake of fluorescent siRNA/jPEI into the hepatopancreas at 24 and 72 hrs, as shown by the intense fluorescence (red) in this tissue. By comparing the same image, but this time under light microscopy (panels A and C) at 24 and 72 hrs, respectively, it was clear that there was comparatively less fluorescence in the ovotestis. Very little fluorescence was also detected in the head-foot region (data not shown). Results showing low fluorescence in the ovotestis, indicating less delivery of siRNA into this tissue compared to the hepatopancreas were from multiple (×10) experiments and are consistent with other studies. Thus, in linear PEI-mediated delivery of DNA in the mouse, Zou et al. [31] detected uptake of the complex essentially in the liver and lungs but not in other organs.
Figure 2 shows images of the hepatopancreas and ovotestis regions of snails that were treated for 72 hrs with the fluorescent siRNA alone (control without PEI) viewed either by light (panel A) or fluorescence microscopy (panel B). Results showed that in the absence of the PEI carrier, only weak fluorescence was detected in the hepatopancreas of this snail. Images in panels C and D of Figure 2 correspond to the hepatopancreas and ovotestis regions of the untreated snail viewed either by light (panel C) or fluorescent (panel D) microscopy. The absence of fluorescence in the control snail in addition to the weak fluorescence detected in snails incubated only with the naked fluorescent siRNA, without PEI (Fig. 2, panel B), supports the conclusion that this carrier mediates the uptake of fluorescence siRNA into the body of the snail by soaking. We have as yet no explanation for how this uptake penetrates the snail's mucus barrier. Additionally, we cannot rule out the possibility that some of the PEI/siRNA nanoparticles might be ingested/swallowed by the snail. Similar intense fluorescence accumulation in the hepatopancreas relative to other tissues was observed by using bPEI instead of jPEI for the delivery (data not shown).
To determine if the delivery protocol developed with fluorescent siRNA (Alexa 555)/PEI nanoparticles can be utilized to target the specific knockdown of a known snail gene we obtained synthetic siRNA designed from different regions of the CathB transcript. We chose siRNA corresponding to conserved regions denoting the propeptide C1 region (siCathB1) and known consensus active site (siCathB2) of this enzyme. PEI complexes of different Cath B siRNA or Cath B dsRNA were tested in a ‘proof of principle’ study, for knockdown of the corresponding CathB transcript. Results showed a knockdown of the CathB transcript; 45.9% and 55.61% suppression (statistically significant differences between the normal snail and other control groups were shown by One-Way ANOVA analysis [P value<0.01]) in snails soaked in the CathB1 siRNA/PEI (Fig. 3a) and CathB dsRNA/PEI (Fig.3b) complexes, respectively. In snails soaked in the siCathB2/PEI complex, however, knockdown was found to be less than when the siCathB1/PEI complex was delivered into the snail. This result was unexpected, especially since the commercially purchased siCathB2 encompasses the catalytic region of the enzyme. Other siRNAs designed specifically from the active site of the snail enzyme sequence will be tested in future studies to resolve this unexpected result. In the other snails, no knockdown effect was observed in those treated either with siCathB1 alone (without PEI) or those treated only with PEI. Likewise, the CathB transcript in snails soaked in mock siRNA/PEI remained at about the same level (99.3%) as the level detected in the normal snail (100%). Similarly, as shown in Figure 3b, control snails soaked in naked CathB dsRNA, Myo dsRNA/PEI complexes, and the PEI carrier alone produced no significant knockdown of the CathB transcript.
To further evaluate the usefulness of the PEI-complex whole organism soaking method, studies were performed with dsRNA corresponding to another previously characterized snail gene, Prx [21]. As shown in Figure 4a, soaking the snails in the B. glabrata Prx-dsRNA/PEI complex produced a significant knockdown (70% suppression) of the Prx transcript. In contrast, there was no reduction of the transcript in snails that were soaked in mock Myo dsRNA/PEI, Prx-dsRNA, and PEI alone. The level of transcript in these samples remained almost unchanged, and was comparable to the basal level of the Prx transcript in the normal snail.
To determine whether the above delivery of Prx-dsRNA/PEI complexes would result in the successful suppression of the Prx protein as well, by using ELISA (Fig. 4b), we detected dramatic knockdown of the protein in snails soaked in this complex. The protein suppression observed in Prx dsRNA/PEI treated snails analyzed by one-way ANOVA was statistically significant (P value<0.05). Suppression at the protein level was between end-point ELISA titre 800–1,600, compared to either the normal or control groups (end-point ELISA titer between 25,600–51,200). End point titre observed in snails treated with the Prx-dsRNA/PEI complex, reflected a significant reduction in the Prx protein in these snails compared to control snails that were soaked in mock myo-dsRNA/PEI, naked Prx dsRNA, and PEI. Similar ELISA experiments conducted using either commercially available, or S. mansoni anti-CathB antibodies showed no cross reaction between the snail CathB homolog and both these reagents. Therefore, investigations to assess the possible knockdown of Cath B transcript, at the protein level, will have to wait for the future generation of specific antibodies for the snail CathB enzyme. In addition, future studies will be done to examine suppression of the enzyme activities of Prx and CathB in order to evaluate whether the demonstrated PEI mediated knockdown of these transcripts translates into disrupting the function of these enzymes as well.
In preliminary studies, the PEI mediated delivery of fluoroscent siRNA into pre-patent snails enabled the detection of fluorescent labeled sporocysts in the infected snails (data not shown). We have as yet, however, no evidence that our PEI delivery protocol will provide a useful avenue for targeting specific sporocyst (parasite) genes for silencing. Initial attempts to use the PEI delivery protocol for siRNA delivery in miracidia have so far been unsuccessful.
In the present study we found that soaking the snails for 1 to 4 days in either dsRNA/PEI or siRNA/PEI resulted in gene suppression within the first 24 hours that extended to 4 days post- treatment (Fig. 5). Thus, in Figures 5a and 5b, we detected significant knockdown of CathB and Prx after 2 and 3 days of soaking in dsRNA/PEI nanoparticles. Similarly, optimum knockdown of CathB occurred following 2 and 3 days of soaking in the siRNACathB/PEI complex (Fig. 5c). These results also showed that after 4 days of soaking, in either dsRNA/PEI or siRNA/PEI, a recovery from the treatment began to occur, indicating, therefore, that the time period for achieving optimum knockdown (using the conditions we employed) occurs 72 hrs after snails have been soaked in the complexes. Furthermore, we detected no mortality under the conditions utilized even though the snails remained in PEI for as long as 4 days. Thus, results in Figure 6 shows the 100% survival of snails (n = 117) treated with PEI versus untreated snails (n = 29), indicating that the cationic polymer (both linear and branched) was non-toxic at the concentration utilized in the protocol.
Comparable uptake of siRNA/PEI into adult snails was not examined as part of this work because our current research is primarily focused on elucidating the molecular basis of the more vulnerable juvenile snail host's relationship with the parasite. Since our previous attempts to introduce dsRNA into juvenile snails via an injection route produced either inconsistent results or killed the snails, the pressing need for us, therefore, was to develop a reliable and safe gene-silencing tool specifically for snails at this young age and not for adult snails.
By complexing either siRNA or dsRNA to the cationic PEI polymers we have shown in the present study that it is possible to obtain significant knockdown of specific snail transcripts, at both RNA and protein levels, by whole organism soaking. As stated above, most previous studies using RNAi technology in snails have used an injecting route for delivery [7], [8]. The injection method can, however, be cumbersome for an inexperienced investigator, making a straightforward dsRNA delivery method for snails (especially juveniles) until now very challenging and impractical. For this reason, while functional gene analysis with RNAi technology is now used routinely in several organisms, including schistosomes [11] very few molecular functional studies using RNAi have been reported for molluscs. With several genome projects for mollusks now underway, including the B. glabrata genome project [6] the need for an easier and more convenient dsRNA delivery approach for future functional genomic studies in these organisms is overdue. The simplicity of the method described here, paves the way as a first step towards overcoming the challenges faced in using RNAi technology routinely in mollusks.
The PEI soaking method was optimized and evaluated using two snail transcripts (Cathepsin B and peroxiredoxin) that we have previously shown to be expressed, preferentially in the snail's hepatopancreas. Both genes are also early and more significantly induced in the resistant B. glabrata snail stock, BS-90, compared to the susceptible NMRI snail, in response to S. mansoni exposure [20], [21]. With the PEI mediated siRNA and dsRNA gene silencing method developed, we now have the ability to systematically determine the effect of down regulating CathB and Prx as well as other transcripts in the dynamic interplay of the snail- schistosome interaction. Our results showed that the majority of the fluorescent siRNA/PEI complex was detected in the hepatopancreas. It is possible, therefore, that we might have been able to knock- down the two transcripts we chose easily because they are also expressed preferentially in this tissue. In any case, even if limited to this tissue the ability to use a simple and effective whole organism soaking method to conveniently knockdown specific genes in B. glabrata by RNAi technology is an important first step in using this proven gene-silencing tool for snail molecular research.
|
10.1371/journal.pgen.1003694 | A Flexible Approach for the Analysis of Rare Variants Allowing for a Mixture of Effects on Binary or Quantitative Traits | Multiple rare variants either within or across genes have been hypothesised to collectively influence complex human traits. The increasing availability of high throughput sequencing technologies offers the opportunity to study the effect of rare variants on these traits. However, appropriate and computationally efficient analytical methods are required to account for collections of rare variants that display a combination of protective, deleterious and null effects on the trait. We have developed a novel method for the analysis of rare genetic variation in a gene, region or pathway that, by simply aggregating summary statistics at each variant, can: (i) test for the presence of a mixture of effects on a trait; (ii) be applied to both binary and quantitative traits in population-based and family-based data; (iii) adjust for covariates to allow for non-genetic risk factors and; (iv) incorporate imputed genetic variation. In addition, for preliminary identification of promising genes, the method can be applied to association summary statistics, available from meta-analysis of published data, for example, without the need for individual level genotype data. Through simulation, we show that our method is immune to the presence of bi-directional effects, with no apparent loss in power across a range of different mixtures, and can achieve greater power than existing approaches as long as summary statistics at each variant are robust. We apply our method to investigate association of type-1 diabetes with imputed rare variants within genes in the major histocompatibility complex using genotype data from the Wellcome Trust Case Control Consortium.
| Rapid advances in sequencing technology mean that it is now possible to directly assay rare genetic variation. In addition, the availability of almost fully sequenced human genomes by the 1000 Genomes Project allows genotyping at rare variants that are not present on arrays commonly used in genome-wide association studies. Rare variants within a gene or region may act to collectively influence a complex trait. Methods for testing these rare variants should be able to account for a combination of those that serve to either increase, decrease or have no effect on the trait of interest. Here, we introduce a method for the analysis of a collection of rare genetic variants, within a gene or region, which assesses evidence for a mixture of effects. Our method simply aggregates summary statistics at each variant and, as such, can be applied to both population and family-based data, to binary or quantitative traits and to either directly genotyped or imputed data. In addition, it does not require individual level genotype or phenotype data, and can be adjusted for non-genetic risk factors. We illustrate our approach by examining imputed rare variants in the major histocompatibility complex for association with type-1 diabetes using genotype data from the Wellcome Trust case Control Consortium.
| Despite the recent successes of genome-wide association studies (GWAS), which can be well powered under the common disease, common variant hypothesis, the majority of the genetic component of many complex traits remains unexplained. For example, hundreds of common genetic variants, in at least 180 loci, have been associated with height in studies of up to more than 180,000 individuals. However, the individual effects of these variants are modest and their cumulative effect explains just over 10% of the phenotypic variation in height [1], [2], [3], [4]. Rare variants may play an important role in explaining the “missing heritability” of complex traits. Due to recent advances in high-throughput re-sequencing technology, it is becoming financially feasible to assay rare genetic variation in thousands of individuals on the scale of the whole-exome, or even the whole genome. Furthermore, with the availability of whole-genome re-sequencing reference panels, such as those made available through the 1000 Genomes Project [5], imputation allows the possibility to predict genotypes at rare variants not present on, or captured by, GWAS genotyping arrays. Therefore, we now have an exciting opportunity to explore a range of models that may help to explain the missing heritability of complex traits using rare genetic variation. One such model is that where a gene or region affects a complex trait as a consequence of the combined effects of its constituent rare variants. The effects at each rare variant can be either modest or highly penetrant, and can act to either increase or decrease the trait or disease risk.
Recently published methods for the analysis of multiple rare variants illustrate that power can be greatly increased by combining information in a joint analysis in comparison to studying individual variants one at a time [6], [7], [8], [9], [10], [11]. These so called “burden tests” are optimal when all variants have the same direction of effect. However, these variants may act individually to either increase or decrease trait values, or they may be neutral (i.e. no effect on the trait). Ideally, we wish to test for the presence of a mixture of increaser, decreaser and neutral effects at multiple rare variants on a complex binary or quantitative trait. Zelterman and Chen [12] describe tests of homogeneity against such central mixture alternatives for general sampling distributions that are based on the score function. These so called “C-alpha” tests are powerful for detecting the presence of central mixtures [13]. Neale et al. [14] proposed a C-alpha test for the analysis of sequence level data for association with binary (disease) traits based on binomially distributed measures of effect at each site. Their approach has the advantage of allowing for a mixture of risk, protective and neutral effects, but cannot explicitly be applied to quantitative traits, account for non-genetic risk factors as covariates, or allow for imputed variation. More recently, score-based variance component tests SKAT (sequence kernel association test) [15] and an optimized version (SKAT-O) [16] have been proposed for the detection of a mixture of effects which can be applied to both binary and quantitative traits and which can adjust for covariates. These tests have been shown to outperform burden tests and the Binomial C-alpha test in a wide range of scenarios.
Here, we introduce a C-alpha test for the analysis of rare genetic variation for association with both binary and quantitative traits based on normally distributed measures of effect at each site. Measures of effect at each site can be calculated from re-sequencing, array genotyping or imputed data or taken directly from summary measures of effect available, for example, from meta-analysis or published data. Our test assesses the evidence for a mixture of increaser, decreaser and neutral effects in a gene, region or pathway and can be applied to both population and family-based association studies and can adjust for covariates to allow for non-genetic risk factors, such as indicators of population stratification. We refer to our test as the Generalised C-alpha test. We report the results of simulations to investigate the power of our test to detect rare variant association with a quantitative trait, and compare performance with existing approaches.
The HLA class II genes in the major histocompatibility locus (MHC) play a major role in susceptibility to type-1 diabetes (T1D) [17], but common variants mapping to other genes in this region have also been implicated in the disease. Imputation into existing GWAS genotype data up to publicly available reference panels of sequence data can be used to identify novel and refined signals of association with common SNPs (MAF>1%) [18] and is feasible for the evaluation of rare variants [19]. We have used our Generalised C-alpha test to evaluate the evidence for rare variant association with T1D within genes in the MHC using GWAS genotype data from the Wellcome Trust Control Consortium (WTCCC) [20] imputed up to reference panels made available through the 1000 Genomes Project [5].
Consider a gene, region or pathway containing K variants, each with a minor allele frequency (MAF) less than a pre-defined threshold and assayed in a sample of individuals measured for a binary or a quantitative trait. Suppose that at each variant a normally distributed estimate of the effect of the minor allele on the trait of interest can be obtained. For example, in a case-control association study such an estimate may be the log allelic odds ratio obtained as a coefficient in a logistic regression; or in a quantitative trait association study, the estimate may be the per-allele increase in phenotypic value obtained as a coefficient in a linear regression. For each variant alone, there is unlikely to be enough information to make inference about association, unless the sample size is unfeasibly large. However, if the gene is not associated with the trait, then the distribution of estimates across all variants will be Gaussian with mean zero. Conversely, if variants in the gene are associated with the trait, there will be a mixture of Gaussian distributions with different means, manifested as “overdispersion”, which can be detected by a C-alpha test.
More formally, let denote the effect estimate, and it's corresponding estimate of standard deviation, at variant k, k = 1,…,K. We assume that are independent Gaussian distributed random variables with mean and standard deviation . As described, such estimates will typically have been obtained from a logistic (binary trait) or linear (quantitative trait) regression of trait value on genotype. The C-alpha test of homogeneity can be derived for a given sampling model. Here the effects are treated as sampling units from a Gaussian sampling model. Under the null hypothesis of no association with the trait, we assume that all are equal to some fixed, unknown value, denoted . Under the alternative hypothesis, we assume that the take on a mixture of values, centred at . The C-alpha test statistic for a test of homogeneity of against a central mixture of alternative Gaussian hypotheses iswhere is an estimate of under the null hypothesis. In practice, we estimate by the observed standard deviation . Notice that S is simply the sum of the differences between the variance of the observed measures of association and the expected variance under the null hypothesis. To standardise S, we require the estimated normalizing varianceThe standardised C-alpha test statistic is thenwhich is asymptotically standard Gaussian distributed. The null hypothesis of no association is rejected for values of ZNORM significantly larger than that expected using a one-tailed test of size α. The quantities S and c are easily derived using methods detailed in Zelterman and Chen [12] for sampling units from a distribution belonging to the exponential family: in this case, the Gaussian distribution, where is treated as a nuisance parameter. Note that a natural adjustment for the effect of non-genetic risk factors can be achieved by including covariates in the regression model used to estimate . Furthermore, we can consider imputed variation by replacing direct genotypes with dosages under an additive model, or by maximisation of the missing data likelihood of the distribution of genotypes.
For genetic association studies, the expected effect of a minor allele is zero, so that , and the C-alpha statistic reduces to:The assumption that the distribution of ZNORM is Gaussian depends on: (i) the degree of sparseness in the data, as summarised by the relationship between sample size and MAF at each variant; (ii) the number of variants that are considered and (iii) the independence of variants. When the data are too sparse, because the sample size is too small and/or the MAF too low, the maximum likelihood estimates of effect size computed at each site are typically unstable. Furthermore, the discrepancy between the empirical variance of the estimates, and their variance under the reference asymptotic distribution can be large, resulting in inaccurate type I error [21]. It is reasonable to assume that large numbers of individuals will be genotyped because in a practical study design, tests require large numbers of individuals for adequate power, however the minimum MAF must be constrained to ensure stability of estimates in the presence of, for example, private mutations. The second and third requirements ensure convergence of the null distribution of the ZNORM to Gaussian by the central limit theorem. To estimate significance accurately for low MAF, where small numbers of variants are considered or where variants are correlated, standard permutation testing is required. See Text S1 for details of the standard permutation approach utilised here.
We conduct simulations to investigate the performance of the Generalised C-alpha test for the identification of rare variants associated with a binary or quantitative trait. We compare the performance of the Generalised C-alpha test to three existing approaches: (i) the optimized score-based variance component test (SKAT-O, by Lee et al. [15] (ii) the Binomial C-alpha rare variant test by Neale et al. [14], and (iii) GRANVIL, a burden test of association of binary or quantitative traits with accumulations of minor alleles at rare variants in a generalised linear modelling framework by Morris and Zeggini [10]. A short summary of these tests is given here.
Our simulations make use of a simple model of population genetics to generate high-density haplotype data in 30–200 kb genomic regions, designed to represent a gene. Haplotypes are then randomly paired together to form individuals for analysis, and quantitative trait values are generated according to their genotypes at rare causal variants, selected at random according to the underlying trait association model. In the trait association model that we consider here, we assume that the expected phenotypic value of an individual is determined by the net effect of a combination of increaser causal variants, which serve to elevate the mean trait value in the population, and decreaser causal variants, which serve to reduce it. The trait association model is parameterised in terms of: (i) the maximum MAF of each individual causal variant; (ii) the total MAF of all causal variants in the gene; (iii) the relative proportion of increaser and decreaser causal variants; and (iv) the joint contribution of the causal variants in the gene to the trait variance. Full details of the simulation process are described in Text S1.
The Generalised C-alpha test, SKAT-O and GRANVIL are applied directly to the simulated quantitative trait. However, to apply tests designed for binary traits, we dichotomise the quantitative distribution by assigning individuals as “cases” if they belong to the upper 50% of the trait distribution, or “controls” otherwise. The Generalised C-alpha test, as well as the Binomial C-alpha test, is then applied to the dichotomised trait. The significance of the Generalised C-alpha and Binomial C-alpha test statistics are evaluated empirically by standard permutation testing (see Text S1 for details), whilst GRANVIL relies on the asymptotic properties of a linear regression model and SKAT-O uses Davies method [22] for approximating the distribution of the test statistic. For each simulation, we permute 1,000 or 100,000 times to ensure accurate assessment at 0.05 and 1×10−5 significance levels, respectively. Simulations are repeated 10,000 times for each set of parameter values.
We evaluated the evidence for rare variant (MAF<1%) signals of association with T1D in genes on chromosome 6 using the Generalised C-alpha test applied to rare variants using genotype data from the WTCCC [18]. All WTCCC samples are ascertained from the UK. We applied the same quality control (QC) filters employed and described by the WTCCC to exclude samples and SNPs from the analysis. These high-quality samples were imputed up to the Phase 1 1000 Genomes Project reference panel (June 2011 interim release) [5] comprising 1,094 phased individuals from multiple ancestry groups. Adjustment for fine-scale population structure is critical in rare variant analysis because recent founder effects can exert greater impact on association analyses with rare variants than with common variants [23]. To control for population structure we constructed principal components to represent axes of genetic variation within the UK and included these as covariates in association analyses to obtain estimates of effect at each SNP that are adjusted for ancestry. These procedures for imputation and control of fine-scale population structure are the same as those utilised by Magi et al. [24], full details of which are presented in their paper.
For each gene, the Generalised C-alpha test was applied to SNPs in two MAF ranges: 0.1%<MAF<0.5% (very rare) and 0.5%<MAF<1% (rare). Measures of effect at each SNP used in the Generalised C-alpha test were the log odds ratios estimated from single SNP additive tests of association using simple logistic regression. The Generalised C-alpha test was applied to the original data and then, in order to determine a permuted p-value, to repeated permutations of the case/control status and covariate data (see Text S1 for details of the standard permutation approach). We performed two separate analyses with and without adjustment for the lead MHC SNP for T1D, rs9268645. Assuming there are approximately 30,000 genes in the human genome [25], a p-value of less than 0.05/30,000 = 1.7×10−6 is required to ensure genome-wide significance. Hence for each analysis, we performed 600,000 permutations and declared genome-wide significance for a given gene if less than 1 of 600,000 (<1.7×10−6) permutations resulted in a C-alpha test statistic larger than the original.
The assumption that the C-alpha statistic is normally distributed under the null hypothesis depends on the quantity and independence of the variants considered as well as the accuracy of the individual estimates at each variant, which in turn depends on the sample size and the MAF. By considering regions of a fixed size and varying the minimum MAF of alleles considered and the sample size, we were able to effectively vary the number of variants and the allele frequency distribution in order to explore type I error and power.
After QC and imputation, the WTCCC data comprised 2,938 T1D cases and 1,963 controls with directly or imputed genotypes available at 490,888 SNPs with 0<MAF<1%, located in 1,611 distinct genes on chromosome 6; gene boundaries were identified from the UCSC human genome database (build 37). Table 2 shows the genes demonstrating genome-wide significant evidence of rare variant association with type-1 diabetes on chromosome 6. Genome-wide significant (Bonferroni correction for 30,000 genes at a 5% significance level: p<1.7×10−6) evidence of association with T1D were observed with rare variants in 17 genes throughout the 7.5 Mb extended Major Histocompatibility Complex (MHC) region (ranging from the GNL1 gene to the COL11A2 gene). The strongest signal of association was observed at C6orf10 (ZNORM = 89.1, p<1.7×10−6), which contains rare variants previously implicated in susceptibility to T1D [26].
Common SNPs in the MHC have been previously associated with T1D [17], [20]. Exactly which and how many loci in the MHC determine susceptibility remains unclear as a consequence of the high gene density and the strong association between alleles in the region. To take account of established associations in the MHC, we repeated our analyses on the genes with rare variants showing genome-wide significance evidence of association with T1D with adjustment for the lead MHC SNP (rs9268645) [17]. The common SNP explained the rare variant association in 11 of the MHC genes; 6 MHC genes achieved genome-wide significant evidence of rare variant association with T1D after adjustment for the lead MHC SNP.
We have developed the Generalised C-alpha test for the analysis of multiple rare variants that display a mixture of increaser and decreaser effects on a binary or quantitative trait. The Generalised C-alpha test is a score test combining routinely calculated Gaussian distributed measures of effect at multiple variants in order to increase the power to detect an effect at the gene, region or pathway level. The Binomial C-alpha test for binary traits, [14] and, more recently, SKAT-O [15], have been shown to have several advantages over previously proposed tests by Li and Leal [8], Madsen and Browning [9] and Price et al. [11]: most notably increased power in the presence of a mixture of increaser and decreaser effects. Our results confirm that the Generalised C-alpha test is also robust to the presence of bi-directional effects, with no apparent loss in power across a range of different mixtures.
The Generalised C-alpha test performs better than SKAT-O when the data is not too sparse: in our examples we showed the Generalised C-alpha was optimal as long as there were at least 15–25 copies of a minor allele at each rare variant. When data is sparse, so that either the sample size is too small and/or the MAF is too low, estimates of allelic effects at each SNP are not robust, and the asymptotic assumptions on which the Generalised C-alpha test are based are inappropriate. Similarly, for testing rare variant association with a binary trait, we have shown that the Generalised C-alpha test has lower power that the Binomial C-alpha test in the presence of variants with very low minor allele counts: a minimum MAF>∼0.5% is recommended in order to achieve comparable power in these tests.
In any application, the Generalised C-alpha test works on the assumptions that there are (i) a sufficiently large set of variants; (ii) that estimates of effect based on these variants are robust and independent and; (iii) normally distributed. These assumptions are often unrealistic: they are violated for example, in the presence of linkage disequilibrium, small sample size, low MAF or few variants. Hence, it is imperative that permutation testing is employed for accurate estimation of significance. For analysis of the whole genome, 1,000 permutations, for which a simply coded version of the test can be run in a matter of seconds, is recommended as a first approach; regions where the test is significant with a p-value<0.001 can then be rerun with 100,000 or more permutations for an accurate estimate of genome-wide significance.
Unlike the Binomial C-alpha test, the Generalised C-alpha test can naturally adjust for additional covariates and can easily incorporate imputed variation. Unlike SKAT-O, the Generalised C-alpha test can be applied to summary statistics, without requirement of the individual level genotype data. For example, the Generalised C-alpha test can be quickly and easily applied to published data. However, this is recommended only for discovery as permutation testing cannot be implemented in this case and test statistics are likely to be inflated leading to increased type I errors: In this case, any regions identified would require further investigation for any confirmation of association.
Evaluation of rare variants extracted from existing GWAS data via imputation up to re-sequencing reference panels, such as those made available by the 1000 Genomes Project, has been demonstrated to be feasible [18]. We applied the Generalised C-alpha test to rare variants imputed into the WTCCC T1D GWAS across the MHC where genes have been shown to play the single most important role in susceptibility to T1D in both common variant and haplotype analyses. Genome-wide significant association with T1D, independent of the lead common variant GWAS signal in the region, was observed at multiple genes. These included HLA class II genes, DR and DQ, where coding polymorphisms have been shown to account for most of the association with T1D observed at the HLA locus [27], [28], [29]. The identification of rare disease-associated variants within genes in this region highlights the complex genetic architecture of T1D in the MHC, and requires further investigation to disentangle the effects of common and rare variation on immune disease susceptibility.
In summary, the Generalised C-alpha test is a novel, flexible and powerful method for the analysis of rare genetic variation. There is no single alternative test, amongst those we have considered, that is uniformly most powerful over all models and genetic architectures. Our test, however, has the unique advantage that it can be applied to summary statistics from published literature, without the need for individual level genetic data. The fact that the Generalised C-alpha test simply aggregates data from summary statistics allows for great flexibility in general allowing direct application to both binary and quantitative traits, to population (using summary statistics from generalized linear models, as illustrated here) and family based data (using summary statistics from the transmission disequilibrium test, for example), and to imputed genotype data whilst simultaneously allowing for the adjustment of additional covariates. We are already using the method in our analyses and it is currently implemented using the R-PLINK/SEQ library available from: http://atgu.mgh.harvard.edu/plinkseq/. R package is available from http://www.well.ox.ac.uk/~rivas/calphanorm.tar.gz.
|
10.1371/journal.pntd.0003687 | Phlebotomus sergenti in a Cutaneous Leishmaniasis Focus in Azilal Province (High Atlas, Morocco): Molecular Detection and Genotyping of Leishmania tropica, and Feeding Behavior | Phlebotomus (Paraphlebotomus) sergenti is at least one of the confirmed vectors for the transmission of cutaneous leishmaniasis caused by Leishmania tropica and distributed widely in Morocco. This form of leishmaniasis is considered largely as anthroponotic, although dogs were found infected with Leishmania tropica, suggestive of zoonosis in some rural areas.
This survey aimed at (i) studying the presence of Leishmania in field caught Phlebotomus sergenti, (ii) investigating genetic diversity within Leishmania tropica and (iii) identifying the host-blood feeding preferences of Phlebotomus sergenti. A total of 4,407 sand flies were collected in three rural areas of Azilal province, using CDC miniature light traps. Samples collected were found to consist of 13 species: Phlebotomus spp. and 3 Sergentomyia spp. The most abundant species was Phlebotomus sergenti, accounting for 45.75 % of the total. 965 female Phlebotomus sergenti were screened for the presence of Leishmania by ITS1-PCR-RFLP, giving a positive rate of 5.7% (55/965), all being identified as Leishmania tropica. Nucleotide heterogeneity of PCR-amplified ITS1-5.8S rRNA gene-ITS2 was noted. Analyses of 31 sequences obtained segregated them into 16 haplotypes, of which 7 contain superimposed peaks at certain nucleotide positions, suggestive of heterozygosity. Phlebotomus sergenti collected were found to feed on a large variety of vertebrate hosts, as determined by Cytochrome b sequencing of the DNA from the blood meals of 64 engorged females.
Our findings supported the notion that Phlebotomus sergenti is the primary vector of Leishmania tropica in this focus, and that the latter is genetically very heterogeneous. Furthermore, our results might be suggestive of a certain level of heterozygosity in Leishmania tropica population. This finding, as well as the feeding of the vectors on different animals are of interest for further investigation.
| In Morocco three Leishmania species have been reported to cause cutaneous leishmaniasis: Leishmania major, Leishmania tropica and less frequently Leishmania infantum. Amongst these clinically important Leishmania species, Leishmania tropica is considered as a public health problem by the Ministry of Health in Morocco and other endemic countries. Phlebotomus sergenti is the known vector, which is thought to take blood meals mainly from humans, since they appear to be the sole reservoir, considering anthroponosis of the cutaneous leishmaniasis caused by L. tropica in many endemic areas. In the present study, we investigated by molecular tools the presence of Leishmania in field caught Phlebotomus, as well as the heterogeneity of Leishmania tropica in a cutaneous leishmaniasis focus in High Atlas of Morocco. Our results showed a high infection rate of Phlebotomus sergenti, which may be a consequence of high level of the parasite circulating in this region; and underlined the important genetic heterogeneity of Leishmania tropica in Morocco. Analysis of the blood meals of the vectors showed that Phlebotomus sergenti fed on a variety of vertebrates, including wild animals, such as rodent, monkey and bat. Whether these animals play any role in the maintenance of Leishmania tropica in this focus awaits further investigation.
| Leishmaniases are complex diseases of worldwide distribution caused by >20 Leishmania species, which are parasitic protozoa transmitted by the bite of infected female sand flies. The disease affects 98 Mediterranean and other endemic countries putting a population of 350 million people at risk of infection and causing 1.3 million new cases and 20,000 to 30,000 deaths annually [1, 2].
In Morocco, leishmaniasis is a public health problem, with 8,862 cases notified in 2010 alone. It is widely distributed from the mountains of the Riff to the peri-arid foothills of the Anti-Atlas [3, 4]. Three Leishmania species are responsible for cutaneous leishmaniasis (CL) in Morocco: L. major, L. tropica and L. infantum. CL due to L. tropica is endemic in the semi-arid regions and the sand fly species Phlebotomus (Paraphlebotomus) sergenti Parrot is considered to be the vector [5, 6]. This form of CL was described for the first time in Tanant, a rural locality in Azilal province in High Atlas [7]. Thereafter, a large rural focus has been further identified in the centre and the south of Morocco [6]. In the mid 1990s, CL caused by L. tropica was then reported in Taza province, in northern Morocco [8] and early in 2000, outbreaks occurred in emerging foci in the central and northern parts of the country, and L. tropica was also even found in areas previously only known as L. major foci [9, 10]. CL due to L. tropica is basically anthroponotic, although the parasite was also isolated from dogs in some leishmaniasis foci in Morocco, suggesting that L. tropica could also be zoonotic at least in some areas [5, 11, 12].
Blood meal analysis of haematophagous insects has been informative in elucidating their natural host-feeding preference or host-feeding pattern for identifying potential reservoirs [13]. For sand flies, the blood meal sources have been identified initially by serological techniques, like ELISA [14, 15], agarose gel diffusion [16], counter immunoelectrophoresis [17], precipitin test [18, 19] and a more laborious histological technique [20]. Although all these methods have been useful in identifying vertebrate hosts for many haematophagous insects, they lack sensitivity and are time-consuming.
In the last decade, several PCR-based molecular approaches have been developed to enhance the specificity of insect blood meal identification [21, 22]. Analysis of blood meal sources of the engorged vectors has been greatly facilitated by the increasing availability of animal genomic database at the genus and species levels (NCBI, EBI…). Downstream applications of PCR based on primers designed from multiple alignments of the mitochondrial Cytochrome b gene of different vertebrate species have identified avian and mammalian blood sources in various species of sand flies. The sequencing of Cyt b gene, as well as the Cyt b-PCR-RFLP and the Cyt b-PCR combined with reverse line blot have been used to identify the blood meals sources of sand flies [23–25].
The vector incrimination is classically based on the dissection of freshly caught individual sand flies and subsequent culture of parasites present in the gut [26]. This method needs dissecting expertise and is time consuming, since the Leishmania infection rate in sand flies is usually very low [27]. In recent years, several molecular techniques have been developed to identify Leishmania infection in infected phlebotomine sand flies. Because of excellent sensitivity and specificity, these methods have the potential to challenge the time-consuming isoenzyme technique for timely characterization of Leishmania strains. They also play a prominent role in epidemiological prospective surveillance activities which are critical for planning targeted control measures of leishmaniasis [15, 28, 29].
In the present study, we used molecular tools (ITS1-PCR-RFLP) to detect and identify Leishmania species within naturally infected sand flies collected from three rural localities of Azilal province in Morocco. We also applied nested-PCR of Leishmania ITS-rDNA genes for sequencing analyses to characterize L. tropica. The blood meals of engorged flies were also analyzed to determine the putative animal reservoirs of L. tropica, using a PCR technique based on the Cyt b gene of vertebrate mitochondrial DNA (mtDNA).
The sand flies were collected in three neighboring rural areas in the province of Azilal, High Atlas of Morocco. The three regions are: Ait Makhlouf II: 31°01’23”N, 6°58’53”O and Guimi: 32°00’12”N, 6°55’03”O, which are located in Beni Hassan sector, and Agmeroul: 31°58’38”N, 6°51’12”O in Tabia sector. These localities are at different altitude from 679 m to 840 m, separated from one another at a distance of 5 to 11 km (Fig. 1). Vegetation is sparse and mainly dominated by cactus, jujube plant, olive, and almond trees. Agriculture remains the primary source of income and is mainly based on the production of wheat, almonds and olives. Livestock include chicken, sheep, cattle and horses and most homes have at least one dog. Houses are mainly of the traditional type, constructed with adobe. The villages are surrounded by mountains, with some groundwater sources used for irrigation.
Phlebotomine sand flies were caught during three consecutive nights from June to October 2011, using CDC light traps placed inside the houses and in domestic animal shelters, in the three rural sites. The traps placed about 1.5 meter above the ground were set before sunset and collected the next morning. The collected sand flies were then placed in 1.5 mL Eppendorf tubes, transferred to the lab in dry ice and kept at -80°C. All the sand flies were washed and dissected. The head and genitalia were used for morphological identification using Moroccan morphologic key [4]. To distinguish P. perniciosus and P. longicuspis a very closely related species, we used the morphological criteria updated by Benabdennbi et al. (1999) [30] and Guernaoui et al. (2005) [31] based on the morphological features of genitalia and the number of coxite hairs. P. perniciosus is characterized by 2 aedeagi forms: (i) copulatory valve with apex bifid; and (ii) copulatory valve with curved apex, the number of coxite hairs being up to 14. P. longicuspis has copulatory valve ending with a single and long point, and 19 or more coxite hairs.
The remainder of the body of engorged and unengorged females after dissection was stored sterilely in 1.5 mL microtube at -20°C until use.
Genomic DNAs were extracted from unengorged and engorged female P. sergenti sand fly by using the Kit PureLink™ Genomic DNA Mini Kit and the QIAamp® DNA Blood Mini Kit, respectively, according to the manufacturer’s instruction.
For P. sergenti females, the leishmanial ribosomal internal transcribed spacer 1 (ITS1) region was amplified, using the primers LITSR (5′-CTGGATCATTTTCCGATG-3′) and L5.8S (5′-TGATACCACTTATCGCACTT-3′), following the protocol described by Schonian et al. (2003). This PCR was used to amplify a 300 to 350 bp fragment; then the ITS1-PCR products were digested with restriction endonuclease HaeIII enzyme for Leishmania species identification [32]. The restriction profiles were analyzed by electrophoresis on 3% agarose gel containing ethidium bromide, and compared with Leishmania reference strains: L. major (MHOM/SU/73/5ASKH), L. tropica (MHOM/SU/74/K27) and L. infantum (MHOM/TN/80/IPT1). A 100 bp DNA size marker was used. Negative controls (containing water, without DNA) were added to each PCR run.
Further identification of Leishmania parasites was done using nested PCR amplification of ITS1-5.8S rRNA gene-ITS2. The primers IR1 and IR2 were used for the first PCR stage and ITS1F and ITS2R4 for the second stage, as previously described by Parvizi and Ready [33]. Negative controls were included to each PCR run. All PCR products were analyzed by electrophoresis on 1% agarose gel containing ethidium bromide. We used the standard DNA fragment 100 bp ladder as a size marker. The nested-PCR products (~460 bp including primers) were sequenced, using the primers ITS1F and ITS2R4. The sequences obtained were edited using Seq Scape, BioEdit softwares and aligned using ClustalW in MEGA6 software.
A phylogenetic tree was constructed by using the Maximum Likelihood method based on the Tamura-Nei model. Initial tree(s) for the heuristic search were obtained automatically by applying the Maximum Parsimony method. Bootstrap replicates were performed to estimate the node reliability, and values were obtained from 1,000 randomly selected samples of the aligned sequence data. Sequences were compared with entries retrieved from GenBank. Evolutionary analyses were conducted in MEGA6.
Morphologically identified P. sergenti females were confirmed by sequence analyses of PCR-amplified mitochondrial DNA fragment encompassing Cyt b region using PhleF (5’-AAT AAA TTA GGA GTA ATT GC-3’) and PhleR (5’-GCC TCG AWT TCG WTT ATG ATA AAT T-3’) primers [34]. A 500 bp fragment was PCR-amplified using the following reagents: 1X PCR buffer, 2.5 mM MgCl2, 0.4 mM primers (F and R), 0.2 mM dNTPs, 4 μL DNA, made up with distilled water to a final volume of 25 μL for each reaction. The PCR runs were each initiated as follows: 94°C for 12 min, followed by 5 cycles of 94°C for 30 s; 52°C for 30 s, 72°C for 1 min. The PCR reaction was then continued for 30 cycles under the same conditions, except for the annealing step (54°C for 1 min) and a final extension at 72°C for 10 min. The PCR products were sequenced, employing the same primers used for the PCR. Sequences were processed and aligned, using the multiple alignment programs Seq Scape and BioEdit.
A phylogenetic tree was constructed by using the Neighbor-Joining method, in agreement with Kimura 2-parameter model, at uniform rate for transitions and transversions. Bootstrap replicates were performed to estimate the node reliability, and values were obtained from 1,000 randomly selected samples of the aligned sequence data. Sequences were compared with entries retrieved from GenBank. The phylogenetic tree was constructed by using the MEGA6 software.
A region of Cyt b gene from host mtDNA was amplified for blood meal source identification of the blood-fed female specimens. The assay is based on specific amplification and sequencing of the blood meal–derived the region of the host mtDNA Cyt b gene. Modified vertebrate-universal specific primers L14841 (5’-CCATCCAACATCTCAGCATGATGAAA-3’) and H15149 (5’-CCCCTCAGAATGATATTTGTCCTCA-3’) were used to amplify a 359 bp segment of the Cyt b gene [35, 36].
PCR reaction mixture contained 1x PCR Rxn Buffer, 3 mM MgCl2, 300 μM deoxynucleotide triphosphates, 0.4 μM of each primer, 1 unit of Taq DNA polymerase, 3 μL of DNA template solution and distilled water in a final volume of 25 μL. Samples were incubated at 95°C for 5 min; followed by 36 cycles each at 95°C for 30 s, 60°C for 50 s, and 72°C for 40 s; and a final extension at 72°C for 5 min. A negative control was included for each run. Amplified products were analyzed by electrophoresis on 1% agarose gels stained with 2 mg/mL ethidium bromide and visualized under ultraviolet (UV) light. The amplified products were sequenced on both strands, using a Big Dye Terminator v3.1. Sequences were edited and aligned using Seq-Scape and BioEdit softwares.
The sequences obtained were aligned with those deposited in the GenBank database through the BLAST program for the identification of blood meal vertebrate sources. Sequences of a given pairwise alignment with the lowest E-value were selected as the most likely species of blood meal vertebrate source.
In order to characterize the species diversity of sand fly populations, several parameters were used: (i) species richness (S) corresponding to the number of species present in the studied habitat; (ii) relative abundance (pi) reflecting the proportion of individuals belonging to the species i.
These parameters were used to calculate different ecological indexes [37, 38].
Simpson’s diversity index (D) that quantifies the biodiversity of a habitat. It takes into account the number of species present as well as the abundance of each species. D=1∑i=1Spi2
Shannon index (H’) that quantifies species diversity: H′=−∑i=1spilnpi
Pielou index (J’): to calculate the equitability of the distribution of species.
J′ = H′/ln(S) where H’ is the value of Shannon index and S is the species richness.
A total of 4,407 sand flies (2,007 females and 2,400 males) were collected in the three rural areas of Azilal province. Morphological analysis identified 10 Phlebotomus species of three subgenera as follows: P. sergenti, P. kazeruni and P. chabaudi of the subgenus Paraphlebotomus; P. perniciosus, P. longicuspis, P. perfiliewi, P. ariasi, P. langeroni and P. chadlii of the subgenus Larroussius and P. papatasi of the subgenus Phlebotomus. Three Sergentomyia species were also identified: Sergentomyia minuta, S. antennata and S. fallax.
The predominant species was P. sergenti; with a relative abundance of 47.32% of the sand flies belonging to the genus Phlebotomus and 45.75% of the total sample.
P. perniciosus and P. longicuspis are the next most abundant species, representing 26.82% and 14.23% of the sample respectively (Table 1). July had the highest sand fly abundance, whereas the greatest specific richness (S) was found in September, with the occurrence of eleven species against ten, nine and five species recorded in June, July and October, respectively. The three biodiversity indices were maximal in September (Table 2).
All the 39 L. tropica infected female P. sergenti that were identified morphologically, were validated by molecular characteristics based on the sequencing of Cyt b fragment. The alignment of the Cyt b sequences obtained confirmed that all samples corresponded to P. sergenti. All sequences were similar to P. sergenti Cyt b sequences: DQ840350, JN036763, DQ840392 and JN036764. Phylogenetic analyses grouped them in a distinct clade with other P. sergenti retrieved from GenBank (Fig. 4).
One hundred and twelve full or partially blood-fed females (11.6% of the total P. sergenti females collected) were tested for their blood feeding preferences. The Cyt b gene sequencing revealed that 64 P. sergenti fed on a variety of vertebrate hosts, including humans, chickens (Gallus gallus), rodents (Myomyscus brockmani, Mesocritus brandti, Dipodomys ordii: 83–91% of identity), birds, cattle (Bos Taurus), rabbit (Oryctolagus cuniculus), bat (Molossus molossus 89% of identity) and monkey (Macaca nigra 95% of identity) (Table 4).
CL due to L. tropica occurs in many endemic areas, especially North Africa, Middle East, Central and South Asia. L. tropica has not been isolated from P. sergenti frequently [39, 40] as reported here. The transmission cycles of L. tropica vary with different geographic locations and generally have not been known to require a sylvatic reservoir. One noticeable exception is the atypical focus in Northern Israel, where L. tropica is transmitted by P. (Adlerius) arabicus with hyraxes as the reservoir, adjacent to a classical anthroponotic focus [41].
A high biodiversity of sand flies was observed in the region under study, representing 57% (13/23) of all Phlebotomus species hitherto described in Morocco. We have demonstrated that P. sergenti was the predominant vector species (45.75%) present abundantly throughout the collection period.
In agreement with previous reports, our findings showed that the rate of L. tropica infection is higher in fed compared with unfed females, as the latter include largely newly emerged adults, which are expected to contain no L. tropica before taking blood meals [42]. Our overall rate of infection with L. tropica is as high as 5.69%, increasing from 2.95% in June to 10.79–13.09% in September-October. These data are consistent with the previous report in the period of high transmission during fall season [43]. On the other hand, the rate of infection found in this area was much higher than the 1.44–3.66% found in other emerging foci in Morocco [44, 45].
The prevalence of infection seen in the study area may result from the increase in the circulation of L. tropica to a high level. Indeed, the number of CL cases has increased steadily, reaching a hundred per year since the first outbreak of CL in 2006 [46]. Our results confirm that P. sergenti is most likely the primary and the only vector of L. tropica in this region, not only due to the fact that it was the only species found infected with L. tropica but also because of its high abundance and infection rate. In Morocco, L. tropica was first detected in P. sergenti sand flies over 30 years ago [5] and more recently in emerging CL foci in the Centre and the North of Morocco [45, 47].
Species of the sub-genus Larroussius constitute more than 41% of sand flies collected in the studied area with P. perniciosus being the most abundant. Members of sub-genus Larroussius are known to be the vectors of L. infantum. However, PCR screening revealed no Leishmania infection in species belonging to this sand fly sub-genus.
Molecular methods are useful for the detection and characterization of parasites directly in field-collected samples without culturing. Previous evidence indicates that cultivation selects subpopulations of the parasites in the biological samples, including the emergence of nonpathogenic trypanosomatids [48]. In addition, direct analysis of Leishmania DNA in field collected samples can produce unexpected findings, including co-infections by more than one Leishmania species in a single subject [49]. In the present study, PCR-based amplification of the ITS1-5.8S rRNA gene-ITS2 proved to be highly sensitive for identifying Leishmania in sand flies to the species and strain levels [28, 47]. Analysis of ITS1-5.8S rRNA gene-ITS2 sequences from 31 P. sergenti specimens makes it possible to show great heterogeneity of L. tropica, segregating them into 16 haplotypes and revealing their phylogenetic relatedness to Indian strains (FJ948458, FJ948456, FJ948465 and FJ948455) and one Moroccan strain isolated from CL patient (KC145159).
L. tropica is a very heterogeneous species and a high degree of intraspecific polymorphism has been described based on isoenzyme analysis and other molecular methods [50–52]. Morocco is known for being the country with the highest number of L. tropica zymodemes described so far: 8 zymodemes were characterized from human, dogs and the sand flies, e.g. P. sergenti [5, 11]. On the other hand, L. tropica strains from P. sergenti showed the largest range of zymodemes compared to isolates from the vertebrate hosts [12]. Our results underscore the high polymorphism of L. tropica detected in P. sergenti.
Furthermore, sequence analysis revealed seven sequences with two superimposed peaks at a single nucleotide position. There are at least three possible interpretations for this: (i) sequence heterogeneity of different copies in ITS1-5.8S rRNA gene-ITS2 organized as tandem repeats in the Leishmania genome; (ii) sequence heterogeneity of the DNAs in different individual cells, since our samples were amplified from a cell population, which can be heterogeneous in a field collected sample; (iii) these allelic base substitutions are suggestive of heterozygosity of L. tropica, indicating possible sexual reproduction in L. tropica. To confirm the last hypothesis, the two first possibilities must be ruled out by sequencing genes known to be single-copy per haploid genome and by amplifying single-copy genes from a single cell or a population grown up from a single cell. Hybrid marker profiles detected in field isolates have been considered as evidence for sexual recombination in Leishmania [53–55], even if the principal reproductive mechanism in Leishmania is asexual via clonal reproduction. Rogers et al. (2014) used whole genome sequencing to investigate genetic variation in L. infantum parasites isolated from naturally infected sand flies. They showed, for the first time, that variation in these parasites arose following a single cross between two diverse strains and subsequent recombination, despite mainly clonal reproduction in the parasite population. Their results are the most direct evidence of sexual recombination in a natural population of Leishmania [56].
Although cutaneous leishmaniasis caused by L. tropica is usually considered an anthroponotic infection [57], zoonotic transmission was reported in Jordan, the Palestinian Authority and Israel [58, 59]. In the present rural CL focus, zoonotic transmission may occur, as the human cases of CL are sporadic and the major vector species feeds not exclusively on humans. P. sergenti was found in high densities inside houses and animal shelters, which indicates a very close contact between this species and humans and domestic animals. The results of the blood meal analyses showed the important relationship between P. sergenti populations and humans, since 41 P. sergenti (64%) fed on humans. Chicken constituted the second highest source of blood taken by this vector (12.5%), however chickens are not susceptible to Leishmania infection, because of some physiological characteristics, including their body temperature of 41.0°C. Furthermore, infected sand flies may eliminate Leishmania parasite, when they take a second blood meal from chickens or birds [60, 61].
Among 7 P. sergenti females feeding on rodents, one specimen was found to be infected by L. tropica, which make rodents a suspected reservoir of this Leishmania species. Indeed, several rodent species are assumed to have a role in transmission of L. tropica [58, 59, 62, 63]. P. sergenti was found also to feed on wild animals such as monkeys and bats suggestive of their potential role as a reservoir for further investigation.
In the present study, we demonstrated a high infection rate of L. tropica in a single Phlebotomus species, that predominantly, but not exclusively, feeds on humans. This confirms the status of P. sergenti as the primary vector of L. tropica in this focus. Our results highlighted the high diversity of L. tropica in Morocco. A zoonotic transmission of L. tropica in this region is further suggested from the blood meal analysis of the vector. The potential for an animal reservoir host for this species awaits further exploration.
|
10.1371/journal.pntd.0003710 | Using Hospital Discharge Database to Characterize Chagas Disease Evolution in Spain: There Is a Need for a Systematic Approach towards Disease Detection and Control | After the United States, Spain comes second in the list of countries receiving migrants from Latin America, and, therefore, it is the European country with the highest expected number of infected patients of Chagas disease. We have studied the National Health System’s Hospital Discharge Records Database (CMBD) in order to describe the disease evolution from 1997 to 2011 in Spain. We performed a retrospective descriptive study using CMBD information on hospitalizations including Chagas disease. Data was divided in two periods with similar length in time: 1997-2004 and 2005-2011. Hospitalization rates were calculated and clinical characteristics were described. We used multivariable logistic regression to calculate adjusted odds-ratio (aOR) for the association between various conditions and being hospitalized with organ affectation. A total of 1729 hospitalization records were identified. Hospitalization rates for the two periods were 18 and 242.8/100000 population, respectively. The median age was 35 years (range 0-87), 74% were female and the 16-45 age-group was mostly represented (69.8%). Overall, 23.4% hospitalizations included the diagnosis of Chagas disease with organ complications. Being male [aOR: 1.3 (1.00-1.77)], aged 45 and 64 years [aOR: 2.59 (1.42-4.71)], and a median hospitalization cost above 3,065 euro [aOR: 2.03 (3.73-7.86)] were associated with hospitalizations with organ affectation. Since 2005, the number of detected infections increased in Spain. The predominant patients’ profile (asymptomatic women at fertile age) and the conditions associated with organ affectation underlines the need for increased efforts towards the early detection of T cruzi.
| Chagas disease, caused by the protozoa Trypanosoma cruzi, is endemic in most Latin American countries and it is considered a neglected tropical disease (NTD). T. cruzi transmission is feasible in vector-free world regions. The main non-vectorial routes are congenital transmission, blood transfusion, and solid organ transplant. According to WHO, in Spain, which is the second country in the world receiving migrant population from endemic countries, there are up to 40.000 persons at risk of developing symptomatic Chagas disease. Since 2005, serological screening for at-risk blood and solid organ donors is mandatory. Lately, some regulations regarding screening of pregnant women from Latin American endemic countries have also been implemented, but only in some autonomous regions. Up to date, there is no surveillance system in place. Thus, the Chagas disease burden remains unknown. In this paper we use hospitalized Chagas disease discharge data, in order to describe Chagas disease’s geographical and temporal trends before and after the implementation of control measures in Spain from 1997 to 2011. We also assess patients’ main clinical characteristics by organ affectation.
| Chagas disease, also known as American trypanosomiasis, is a potentially life-threatening illness caused by the protozoan parasite, Trypanosoma cruzi (T. cruzi). This disease presents itself in two phases. The initial, acute phase lasts for about two months after infection. In most cases, symptoms are absent or mild, but can include fever, headache, enlarged lymph glands, pallor, muscle pain, difficulty in breathing, swelling and abdominal or chest pain [1,2]. During the chronic phase, the parasites are hidden mainly in the heart and digestive muscle. Up to 30% of patients suffer from cardiac disorders (sudden death, complex arrhythmias, ventricular aneurysms, heart failure, and thromboembolism) and up to 10% suffer from digestive (typically enlargement of the esophagus or colon), neurological or mixed alterations [3]. In later years the infection can lead to sudden death or heart failure caused by progressive destruction of the heart muscle [1]. Chagas disease can be effectively treated with benznidazole and also nifurtimox. Both drugs are almost 100% effective in curing the disease if given soon after infection at the onset of the acute phase [4]. However, the efficacy of both diminishes the longer a person has been infected, thus the importance of an early diagnosis [5]. Additionally, specific treatment for cardiac or digestive manifestations may be required at latest stages of the infection [6].
Worldwide, between seven and eight million people are estimated to be infected with Chagas. The disease is endemic in 21 Latin American countries [7], where it is mostly transmitted to humans through the bite of the triatomine bug, Chagas disease natural reservoir [8]. Nevertheless, T. cruzi transmission is also possible in vector-free world regions [9]. The main non-vectorial routes are congenital transmission, blood transfusion, and solid organ transplant, these routes being characteristics but not exclusive to non-endemic countries [10].
Although most of the cases occurr in Latin America [11], Chagas disease in non-endemic countries, such as North America and the Western Pacific Region (mainly Australia and Japan), has come to light since the beginning of 2000, and only more recently in Europe [12]. In 2010, the World Health Organization (WHO) estimated that 80.000 persons could be infected in Europe, making Chagas disease one of the predominant emerging parasitic infections in the Old World [13].
The emergence of the disease in non-endemic countries is mainly linked to population mobility, notably migration from endemic countries [12]. Beginning with 2010, Spain had the highest immigration rates from Latin America within the EU, (second after the USA worldwide) [14]. A study published by the European Center for Disease Prevention and Control (ECDC) estimated that in 2009, 53% (more than 1.7 million people) of the European migrant population coming from Latin America were living in Spain [15]. The same study indicated Spain as the country with the highest prevalence of Chagas disease in migrant population from endemic countries (2.3–3.8%). Estimates that take into account the prevalence of T. cruzi infection in Latin America suggest that between 40,000 and 65,000 infected people currently reside in Spain [16].
In the last decade some progress has been made regarding the management of Chagas disease in Spain. Since September 2005, in accordance with the Royal Decree RD1088/2005, serological screening of the population considered to be at risk is mandatory in all blood transfusion centers [17]. Regulations regarding screening of pregnant women from Latin American endemic countries and their newborns are in place in 4 out of 17 autonomous regions (Valencia, Catalonia, Galicia and Basque Country) [18–21]. In the rest of the country, the detection of congenital Chagas cases depends mainly on the initiative of the health professionals of the National Health System [15] (Table 1).
Up to date, there is no surveillance system for Chagas disease implemented in Spain. However, hospitalized cases are recorded within the National Health System′s Hospital Discharge Records Database (CMBD in Spanish) belonging to the Spanish Ministry of Health. In this paper, we describe for the first time the Chagas disease related hospitalizations in Spain between 1997 and 2011, in terms of time, geographical distribution, and disease related individual characteristics.
We performed a retrospective descriptive study using CMBD information on Chagas disease related hospitalizations in Spain between January 1st, 1997 and December 31st, 2011. CMBD database receives notification from around 98% of the public hospitals in Spain [22]. Compulsory health insurance covers an estimated 99.5% of the Spanish population, although persons not covered by health insurance can receive treatment in public hospitals. Since 2005, CMBD also has a gradual coverage from private hospitals [23].
All CMBD’s hospital discharges having included the Chagas disease diagnosis were reviewed, in any of 14 possible diagnostic positions. For each entry, we collected socio-demographic (sex, age and autonomous region of residency) and clinical data (type and department of admission, average length of hospitalization, non-invasive procedures and history of surgical intervention during the hospitalization, re-admission, outcome, hospitalization′s cost to the health care system, financing regime and diagnosis related group (DRG). Other concurrent clinical diagnoses, in any of the 14 positions, were also assessed. International Classification of Diseases, Ninth Revision, Clinical Modification (ICD 9 CM) were used for this purpose [24].
We calculated the hospitalization rates by sex and autonomous regions using as denominators the Latin American officially registered population in Spain from the 21 countries endemic for Chagas disease. Latin American population figures were available since 1998 from the Spanish National Institute of Statistics (INE) website [25]. The rates increase among the two study periods (1998–2004 and 2005–2011) were mapped using the Geographical Information System Arcgis version 10.0.
We described the clinical characteristics separately for: a) all hospitalizations, regardless the Chagas diagnosis position; b) hospitalizations with Chagas as first diagnosis and; c) hospitalization with child delivery related stay as first diagnosis. For all hospitalizations, we also examined the differences between the patients notified with Chagas disease with and without an organ (heart or digestive system) complication. We assessed the group distribution by using chi-squared and Fischer exact test, when needed, and evaluated the association level between organ complication and clinical individual risk factors through univariate analysis. We included the risk factors for being hospitalized having Chagas disease and a form of organ affectation with a statistically significant association (p<0.1) in a multivariable logistic regression model (backward stepwise procedure), in order to control for possible confounders. Data analysis was performed using STATA software version 12.
This study involves use of patient medical data from The Spanish National Hospital Database (CMBD). These data, are hosted by the Ministry of Health Social Services and Equality (MSSSI). Researchers working in public and private institutions can request the databases by filling, signing and sending a questionnaire available at the MSSSI website. In this questionnaire a signed Confidentiality Commitment is required. All data is anonymized and de-identified by the MSSII before it is provided to applicants. According to this Confidentiality Commitment signed with the MSSSI, researchers cannot provide the data to other researchers, thus other researchers must request the data directly to the MSSSI [22].
Between 1997 and 2011, 1729 hospitalizations including Chagas disease in any diagnosis position were recorded in Spain, out of which 546 (31.6%) were re-admissions. Up to 90% (1630/1729) of these cases were notified between 2005 and 2011. For the two study periods (1998–2004 and 2005–2011), we obtained hospitalization rates of 18 and 242.8/100000 population, respectively.
While the Latin American population migrating form Chagas disease endemic countries began to increase steadily from 2001, the hospitalization rate showed a stationary tendency in the first time period. In the second time interval (2005–2011), this rate increased while the Latin American population began a slow decrease, starting with 2009 (Fig 1). The increase in hospitalization rate was registered in all but two autonomous regions (Fig 2 and S2 Table). For eight autonomous regions the hospitalization rate increased more for women, in two increased more for men, while in seven autonomous regions there was no difference among sexes (Fig 3).
The median age of the 1729 hospitalizations was 35 years (range 0–87) with the 16–45 age-group being mostly represented with 1207/1729 (69.8%). A total of 74% hospitalized were female, predominating in all age-groups (Fig 4). For 206/1729 (11.9%) records, Chagas disease was registered as the first diagnosis while for 878/1729 (50.7%) appeared as second diagnosis. Other diagnoses in first position are summarized in S1 Table. In the remaining 37.4% records, Chagas disease appeared in any of the last 12 positions. The most frequent main diagnostics associated with Chagas disease were related to pregnancy, giving birth or postpartum complications (36.6%), heart and circulatory conditions (15.3%) and digestive system conditions (9.1%). The hospital departments with higher number of Chagas related admissions were Obstetrics-Gynecology (37.1%), Cardiology (13.1%), Internal Medicine (11.2%) and Digestive Diseases (7.9%) (S3 Table). Overall, 1302/1792 (75.3%) hospitalizations were related to Chagas disease without organ complications, while 388/1792 (23.4%) hospitalizations were recorded as Chagas disease with organ complications, of which 317/388 (81.7%) were due to heart complications and 71/388 (17.9%) were due to another organ complications. Invasive or non-invasive medical procedures were documented for 1463/1729 (84.6%) hospitalizations. We identified 290 different types of medical procedures; 579/1463 (33.5%) procedures were related to obstetrical surgery, 375/1463 (21.7%) were diagnosis or therapy procedures and 190/1463 (11%) were related to heart surgery. Of the total 1729 hospitalizations, 487 (28.2%) had registered a surgical intervention. The admission period was inferior to one week in 71% of Chagas related hospitalizations. The predominant admission type was urgent (76.6%) and 61.5% had a severity level of 2 (out of 4). Around 95% of them were discharged at home, decease occurring in 1.2% of overall Chagas related hospitalizations. The hospitalization median cost was 3,064.9 euros (range 287.3–939,324.1). Costs were covered by social security health care service for 95% of the cases (Table 2).
A total of 55.6% (114/206) hospitalized were female, with the 16–45 and 46–65 age-groups being mostly represented (49.3% and 32.7%, respectively). Around 50% of the hospitalizations were related to Chagas disease with organ complication. Surgical intervention was documented for 25/206 (12.2%) hospitalizations and invasive or non-invasive medical procedures were documented for 147/206 (71.4%) hospitalizations. The predominant admission type was urgent (58.5%) followed by programmed (41.5%). Median time of hospitalization was 7 days, and up to 40% stayed more than one week. Around 90% (184/206) were discharged at home; decease occurred in 2.4% of these hospitalizations. The hospitalization median cost was 4195.6 euros.
From 1997 to 2011, 614 hospitalizations related to Chagas disease with child delivery as first diagnosis occurred. The 16–45 age-group was predominant with 612/614 (99.7%) and heart and other organ complications were present in 23/614 (3.8%) and 4/614 (0.6%) of these hospitalizations, respectively. Surgical intervention was recorded for 43.8% of them, and they were mostly urgent admissions (93.2%). Invasive or non-invasive medical procedures were documented for all the hospitalizations, mainly obstetrical procedures. Up to 95% stayed less than one week in the hospital, being the median average time of 3 days. Around 99% hospitalizations were discharged at home, only 30/614 (4.9%) need to be re-admitted. The hospitalization median cost was 2,207.9 euros.
Through bivariate analysis, we identified several conditions associated with hospitalization due to Chagas disease with organ affectation. Being male, aged over 45 years, Chagas disease as main diagnosis, a history of re-admission in the hospital, and hospitalization’s cost higher than the median (3,065 euros), were significantly associated with Chagas disease with organ affectation. On the other hand, being aged between 16 and 45 years, hospitalization stay up to one week, a history surgery, the urgent admission type and the medium level of disease severity where inversely associated with organ affectation (Table 3).
All these significant variables where included in the final multivariable logistic regression model. Results indicated that the patients being hospitalized with organ affectation Chagas disease were 1.3 times more likely to be male, 2.6 times more likely to be aged between 45 and 64 years, 5.4 times more likely to be admitted with Chagas disease as main diagnosis, and almost 3 times more likely to be admitted for the first time in the hospital, compared to those admitted with Chagas disease without organ affectation. Also, hospitalizations with organ affectation were two times more likely to cost above the median cost of 3065 euro that the hospitalizations of asymptomatic Chagas disease. Not having a personal surgery history and medium severity level appeared to have a negative association with Chagas disease with organ affectation (Table 3).
Although there are few studies performed locally in Spain [16,26,27], to our knowledge, this is the first one addressing Chagas disease epidemiology nationwide. Moreover, the use of hospitalized data for characterizing Chagas disease epidemiology in non-endemic countries also represents a premier. We relied on a database providing information from a net of hospitals that covers more than 98% of the population living in Spain [24], therefore we belief that our results are representative for the Spanish territory.
We aimed to describe the Chagas disease epidemiology between 1998 and 2011 in Spain. As CMBD do not provide enough information that would allow for differentiation between T. cruzi infection and Chagas disease, we have chosen the second one as term for our discussion. We have observed an overall increase in hospitalizations including Chagas disease, beginning with 2005 (Fig 3). This occurs while the figures for Latin American migrant population from countries endemic to Chagas were increasing in the 2000–2004 period, but become stable after 2005. This increase in hospitalizations that includes Chagas disease might be due to the introduction in 2005 of the national program for controlling the Chagas disease horizontal transmission (blood and organ donors screening). Another explanation might surge from the implementation of screening programs for pregnant women. However, it is difficult to quantify the impact of these measures over the hospitalization rate. Community of Valencia in 2007 [18], Catalonia in 2010 [19,28], Galicia in 2012 [20], Basque Country in 2008 [21] and more recently Murcia in 2013 [29] (5 out of 17) have implemented protocols for screening and diagnosis of T. cruzi infection in pregnant women in order to allow for early diagnostic and infection treatment in newborns. The hospitalization rates increased more for women in Basque country, Murcia and Valencian Community, while the increase′s rates were higher in men in Catalonia and similar for both sexes in Galicia. In the Basque Country the recommendation for screening pregnant women from endemic countries includes only those women with clinical symptoms or EKG modifications [21]. Protocols also exist in other autonomous communities, but their implementation is confined only to certain hospitals. This is the case of Madrid [30] and Murcia (before the implementation of the regional screening in 2013) [31]. These protocols are heterogeneous in terms of target population as well as the region coverage and the year of implementation (Table 1).
Otherwise, the number of hospitalizations has also increased in those communities without official screening practices and/or regulations in place for pregnant women. Increased awareness of Chagas disease in both healthcare professionals and patients and relatives might be a possible explanation for this trend. It is also feasible that local initiatives towards early diagnosis of Chagas disease might exist in other autonomous regions, these actions remaining unknown at central level. It remains however unclear why in two communities the overall hospitalization rate decreased in 2005–2011 compared to 1998–2004.
Our study indicates that most of the hospitalizations which included Chagas disease diagnosis at any position were females. Nevertheless, the percentage of female sex decreased from 74% to 55.6% when we analyzed only those hospitalizations with Chagas disease as first diagnosis. Another striking feature is the age-group, as most of the female cases were aged between 16–45 years, while the immigrant population distribution in Spain, reveals however uniformity between sexes [32]. Similar findings were described in another Spanish study on patients with Chagas disease attending the hospital care [33]. This can have several explanations: according to the Spanish Institute of Statistics, the migrant Latin American women are predominant in Spain, the 16–45 age-group corresponding to the most well represented age group, with more than 70% being aged between 15 and 45 years [25]. Additionally, screening programs targeting pregnant women implemented in the above mentioned autonomous communities could represent an alternative explanation for the high percentage of hospitalized women at fertile age. This is concordant with a previous estimation indicating that women at childbearing age represent more than 60% of the overall estimated T. cruzi infections in Spain [10]. This finding supports even more the need of extended screening for pregnant women coming from Latin American countries endemic for Chagas. On the other hand, when we assessed only those hospitalizations with Chagas disease as first diagnosis (n = 206), the 16–45 years old group remain the most frequent (49.3%), but closely followed by the 46–65 years old (32.7%). These figures are similar with the Chagas’ disease prevalence rates provided by WHO Global burden of Chagas’ disease estimations [34].
The most important aspect in Chagas disease evolution is represented by the heart or digestive system affectation, secondary to prolonged T. cruzi infection. In our study, the percentage of the reports with cardiac or digestive affectation was around 22%. Of these hospitalizations, the majority had developed a Chagas related cardiac chronic condition. These figures are concordant with a previous estimate of potentially infected immigrants in Spain [10,35]. On the other hand, the high percentage of diagnosed asymptomatic T.cruzi in our study (77%) could be seen as due to implementation of nationwide and local screening programs. These is however difficult to prove. The analyzed database do not registers information related to the reason of testing for T. cruzi existence. Moreover, the rates of symptomatic Chagas disease increase to 50% when we assess only those hospitalizations with Chagas disease as first diagnosis, the heart being the main affected organ. This supports our hypothesis and also bring to light that Chagas disease could be consider a reason for hospitalization, even if organic complications do not exist. A possible explanation might be that these hospitalizations are related with diagnostic procedures. In fact, 74% of these hospitalizations include invasive or non-invasive medical procedures.
The majority of hospitalizations (1524/1729) didn’t have Chagas disease recorded as the main diagnosis; therefore one could consider that even though Chagas disease was present as pathology, it was not the main reason for hospital admission. Almost half of the hospitalizations with Chagas disease as main principal diagnosis had documented chagasic heart condition and another 17% other organ affectation. On the other hand, for the reports where Chagas is not the main diagnosis, 82% had documented asymptomatic Chagas disease. Moreover, less than 5% of the hospitalizations with child delivery as first diagnosis include Chagas disease with any organ complication. Therefore one could speculate that unless the T. cruzi has evolved into a chronic condition, it represents most of the time a silent disease and can easily go under-diagnosed if not properly searched for. This again stresses the importance of developing efficient strategies for an early T.cruzi detection, even more when considering that the efficacy of Chagas disease drugs diminishes the longer a person has been infected [4].
History of surgical intervention was especially prevalent in child delivery hospitalizations, probably related to the delivery itself, as it occurred with the admission type, most commonly urgent in this kind of hospitalizations. On the other hand, hospitalizations with Chagas disease as first diagnosis stayed longer in the hospital than those where it was placed at any position or where child delivery was listed at first diagnosis. Moreover, we saw that the percentage of deaths was slightly higher in those reports with Chagas disease as first diagnosis. We know that Chagas disease is a major cause of mortality in Latin America due to a wide range of pathogenic processes [36]. Our results might suggest higher severity (and probably more symptomatic) if Chagas is placed as first diagnostic, however this is difficult to prove. The analyzed database do not registers information related to the cause of death. Finally, the median cost was 37% and 90% more in hospitalizations with Chagas disease as first diagnosis compared to records with Chagas diseases in any diagnostic position and delivery as first diagnosis, respectively. This finding suggests that the expansion of the screening program and other prevention activities might be a good public health policy in concept of saving costs.
We have identified the male gender as being associated with Chagas disease involving organ affectation. The association was borderline significant, indicating a possible similarity between genders when it comes to chronic Chagas disease. Reports on sex differences in progressive Chagas’ disease are controversial and previous studies have found either that it is more common in men or that it is unrelated to sex [37]. Eventually, selective preventive measure such as Chagas screening only and exclusively in pregnant women might also result in gender differences.
The association of the 45–64 years age-group with Chagas disease with organ affectation is concordant with previous knowledge related to a 20–30 years of asymptomatic infection between contracting the infection and developing the chronic form [38]. Also reports relating Chagas disease with organ conditions were associated with having Chagas disease as main diagnosis and being hospitalized for the first time. Chagas disease associated to an organ condition was 2 times more likely to have a hospitalization cost above the median value than those without organ complications, this finding being especially relevant from a public health perspective for Spain. The cost could be even higher if one is considering the indirect costs related to work absenteeism, incapacity, etc. This might be also the case in any non-endemic country with important immigration from Latin America. Sicuri et al. indicated in their economic evaluation of Chagas disease screening in Catalonia, that targeting all Latin American women giving birth in Spain and of their infants is the best strategy compared to the non-screening option, showing not only obvious economic advantages but also providing useful information for health policy makers in their decision making process [39].
We calculated the hospitalization rate using as denominators Latin American immigrants officially registered in Spain. Non-registered immigrants could enlarge both the population denominator and the number of cases, thus the results should be interpreted with caution. Another limitation when using the CMBD database is that the country of origin is not recorded, therefore the hospitalization rates can be calculated only as overall figures. This is an important aspect being known that the endemicity levels varies widely among Latin American countries [40]. On the other hand, the estimation of the risk factors for hospitalizations with organ complications might be influenced by this and other factors not present in the CMBD database.
One of the strong points of CMBD is that it provides the total number of hospitalizations records without being subject to the limitations of outpatient surveillance systems, such as under-diagnosis or reporting deficiencies. It remains dependent only by the population’s health seeking behavior and healthcare accessibility. The Spanish Government enacted a Royal Decree in 2012 which basically limits access to free services at the point of delivery for all population (including non-residents migrants), undermining the principle of universal coverage which exist in this country since 1986 [41]. This can have as consequence a decrease in the number of Latin American immigrants attending to the healthcare services. Subsequently, and in the absence of systematic and coordinated actions towards early detection of T. cruzi, a larger number of asymptomatic infections might remain under-diagnosed, increasing the number of patients suffering from Chagas disease chronic conditions in the future.
In conclusion, using hospitalized cases discharge database can be a useful tool when describing the Chagas disease epidemiology. However, our work underlines the need of a nationwide systematic approach towards T. cruzi early detection, especially targeting fertile aged women and relatives from endemic Latin American countries. Health inequalities on immigrants health status might exist across our country if screening programs differ from one region to another and, furthermore, are not nationwide implemented. Moreover, a national surveillance system that would allow a more accurate data collection, analyzing and interpretation, could be of added value in completing a more accurate picture of Chagas disease in Spain, resulting useful both in gaining extended disease knowledge, but especially in evaluating implemented control actions.
|
10.1371/journal.ppat.1000487 | The Glyceraldehyde-3-Phosphate Dehydrogenase and the Small GTPase Rab 2 Are Crucial for Brucella Replication | The intracellular pathogen Brucella abortus survives and replicates inside host cells within an endoplasmic reticulum (ER)-derived replicative organelle named the “Brucella-containing vacuole” (BCV). Here, we developed a subcellular fractionation method to isolate BCVs and characterize for the first time the protein composition of its replicative niche. After identification of BCV membrane proteins by 2 dimensional (2D) gel electrophoresis and mass spectrometry, we focused on two eukaryotic proteins: the glyceraldehyde-3-phosphate dehydrogenase (GAPDH) and the small GTPase Rab 2 recruited to the vacuolar membrane of Brucella. These proteins were previously described to localize on vesicular and tubular clusters (VTC) and to regulate the VTC membrane traffic between the endoplasmic reticulum (ER) and the Golgi. Inhibition of either GAPDH or Rab 2 expression by small interfering RNA strongly inhibited B. abortus replication. Consistent with this result, inhibition of other partners of GAPDH and Rab 2, such as COPI and PKC ι, reduced B. abortus replication. Furthermore, blockage of Rab 2 GTPase in a GDP-locked form also inhibited B. abortus replication. Bacteria did not fuse with the ER and instead remained in lysosomal-associated membrane vacuoles. These results reveal an essential role for GAPDH and the small GTPase Rab 2 in B. abortus virulence within host cells.
| A key determinant for intracellular pathogenic bacteria to ensure their virulence within host cells is their ability to bypass the endocytic pathway and to reach a safe replication niche. Brucella bacteria reach the endoplasmic reticulum (ER) to create their replicating niche called the Brucella-containing vacuole (BCV). The ER is a suitable strategic place for pathogenic Brucella. Bacteria can be hidden from host cell defences to persist within the host, and can take advantage of the membrane reservoir delivered by the ER to replicate. Interactions between BCV and the ER lead to the presence of ER proteins on the BCV membrane. Currently, no other proteins (eukaryotic or prokaryotic) have yet been associated with the BCV membrane. Here we show that non-ER related proteins are also present on the BCV membrane, in particular, the glyceraldehyde-3-phosphate dehydrogenase (GAPDH) and the small GTPase Rab 2 known to be located on secretory vesicles that traffic between the ER and the Golgi apparatus. GAPDH and the small GTPase Rab 2 are involved in Brucella replication at late post-infection. Similarly, integrity of secretory vesicle trafficking is also necessary for Brucella replication. Here, we show that recruitment of the two eukaryotic proteins GAPDH and Rab 2 on BCV membranes is necessary for the establishment of the replicative niche by sustaining interactions between the ER and secretory membrane vesicles.
| Brucella abortus invades both phagocytic and non-phagocytic cells [1]–[6] residing inside a membrane-bound compartment called the Brucella-containing vacuole (BCV). Bacteria ensure their survival and replication within host cells by avoiding fusion with lysosomes and by controlling interactions with the endoplasmic reticulum (ER) [1],[5]. The membrane of the BCV is converted into an ER-derived organelle that is permissive for replication [1]. Interactions between BCV and ER occur at dynamic membrane complexes named ERES for ER exit sites, where membrane fusion and fission events take place. These events are regulated by the small GTPase Sar 1. Sar 1 controls the assembly of COPII complexes on the ER mediating vesiculation and tubulation of the ER membrane towards the Golgi apparatus [7]–[9]. These events were shown to be essential for B. abortus intracellular replication at early stages of infection [10]. The Brucella replicative organelle has been, until now, characterized by the presence of ER chaperones such as calnexin, calreticulin, the translocator sec61β, and the ER resident enzyme protein disulfide-isomerase PDI [1],[5],[11]. Aside from these ER resident proteins, no other eukaryotic or prokaryotic proteins have been associated with the BCV membrane as yet, identification of these proteins is essential to understand how Brucella maintains interactions with the ER and keeps replicating within this compartment. In this work, we investigated by proteomic approaches, the composition of the BCV membrane and characterized 2 eukaryotic proteins that are essential for B. abortus survival. We modified a fractionation method, initially used to analyse latex bead-containing phagosomes [12],[13], to isolate BCVs obtained from cells infected with B. abortus. Mass spectrometry analysis of BCV proteins separated by two-dimensional (2D) gel electrophoresis revealed the presence of the eukaryotic protein GAPDH (glyceraldehyde-3-phosphate dehydrogenase). Further work on GAPDH revealed a role for GAPDH and the small GTPase Rab 2 in the intracellular replication of B. abortus.
To analyse the protein composition of the BCV membrane, a large number of purified BCVs is required. We first tried to determine which cell type was more susceptible for B. abortus infection by monitoring the infection of primary phagocytic cells (bone marrow-derived macrophages: BMDM) and phagocytic and non phagocytic cell lines (Raw 264.7, HeLa, baby hamster kidney: BHK-21). Although no difference was observed in the percentage of infected cells at 48 h post-infection (p.i.) between the different cell types (nearly 35% infected cells), the intracellular replication of B. abortus was 10 times higher within BHK-21 cells than in the other cell types (Figure 1A).
As in macrophages [1] and in dendritic cells [14], B. abortus GFP (in green) in BHK-21 cells were located inside a membrane-bound vacuole labelled with the ER marker calnexin (in red), suggesting that B. abortus replicates in ER-derived compartments within BHK-21 (Figure 1B). This result was confirmed by electron microscopy analysis of infected cells (Figures 1C, 1D and 1E). B. abortus was located inside a membrane-bound compartment resembling the ER with ribosomes lining the vacuolar membrane. However, unlike what has previously been described in BMDM and HeLa cells [1],[5], several bacteria resided inside a unique vacuole (Figures 1C and 1D). This may explain the increase of B. abortus intracellular replication within BHK-21 cells. Taken together, these results show that B. abortus extensively replicates in an ER-derived compartment in BHK-21 cells, validating BHK-21 cells as a good cell model for studying the proteic composition of BCV membranes.
In order to obtain high concentration of membrane proteins, we optimized a fractionation method to isolate and purify BCVs. Within the post-nuclear supernatant (PNS) of BHK-21 cells infected with B. abortus for 48 h, approximately 1.5% of vesicles were GFP positive as detected by flow cytometry and we found that 67% of BCVs remained positive for the ER marker calnexin (Figure S1). After PNS preparation, BCVs were first purified on a 50%–12% sucrose gradient. BCVs were present at the interface, which corresponded to 37% sucrose as indicated by the densitometer measurement (data not shown). Bacteria were only detected in the interface fraction (Figure 2A, lane F) by the presence of the Brucella transmembrane outer membrane protein Omp 25 [15]. On the contrary, ER was detected in each fraction of the sucrose gradient with an anti-calnexin antibody (Figure 2A).
This first purification step allowed the elimination of Rab 7-positive late endosomes and cathepsin D-positive lysosomes from the BCV membrane fraction (Figure 2A, lane F). The interface fraction was then loaded onto a second sucrose step gradient. This gradient allowed the removal of early endosomes (Figure 2A, lane H) and ER structures non-associated with the BCV as detected by immunofluorescence and electron microscopy (data not shown). Although the fraction of BCVs was now devoid of endocytic organelles (Figure 2A, lane H), a few mitochondria detected by the anti-mitochondrial VDAC 1 protein antibody were still present. These could be eliminated by incubating the interface fraction with dynabeads (M-500 subcellular) coated with anti-VDAC 1 antibody, but most of BCVs were lost (data not shown). As a consequence, for proteomic analysis we chose not add the dynabeads step. We analysed BCVs by immunofluorescence and electron microscopy to determine if BCVs were still intact after fractionation. 77% of B. abortus GFP were surrounded by ER-positive vacuoles (Figure 2B). Electron microscopy analysis showed that BCVs were still intact after subcellular fractionation (Figure 2B). Bacteria were surrounded by an ER membrane-bound compartment and residual ribosomes were still located on the ER vacuolar membrane. Taken together, these data indicate that we successfully isolated BCVs and preserved their membrane integrity.
To determine the protein composition of the BCV membrane, vacuolar proteins were solubilized by a mild Triton X-100 treatment, precipitated by a trichloroacetic acid/acetone and analysed by 2D gel electrophoresis. Approximately one hundred spots were detected by silver gel staining (Figure 2C) and further analysed by mass spectrometry. Proteins of the BCV membrane identified by mass spectrometry are listed in Table S1. Spot numbers on 2D gel (Figure 2C) correspond to the numbers of identified proteins listed in Table S1. As expected, 18% of proteins identified were ER proteins (i.e. calreticulin, ERP 57 and PDI) and 13% ribosomal proteins confirming that the replicative niche of B. abortus is derived from the ER. Interestingly, 29% of proteins were not ER-related proteins (19% bacterial and 10% eukaryotic proteins). The remaining 40% proteins were mitochondrial proteins. Among the identified proteins, we focused on one peculiar eukaryotic protein: the Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) annotated by spot number 49 on the 2D gel (Figure 2C) and in Table S1. GAPDH has multiple functions within host cells such as glycolysis and apoptosis [16],[17] More interestingly, GAPDH interacts with the small GTPase Rab 2 to control vesicular retrograde transport between the ER and the Golgi [18]. The presence of GAPDH on the BCV membrane leads us to hypothesise that vesicular retrograde transport may be involved in Brucella replication within the ER. Therefore, we investigated the role of GAPDH and Rab 2 in Brucella survival within host cells.
First, the presence of GAPDH on the BCV membrane was confirmed by immunoblotting (Figure 3A, lane C). Its partner, the small GTPase Rab 2, was also detected in the enriched BCV fraction (Figure 3A, lane C). The small GTPase Rab 1, which is known to be a resident of the ER as well as vesicular and tubular clusters (VTCs) and Golgi apparatus [19],[20] was also detected in BCVs (Figure 3A, lane C). As no commercial antibody was suitable for immunofluorescence detection of GAPDH on BCVs obtained from BHK-21 cells, we analysed the presence of its partner the small GTPase Rab 2. We found that 35% of isolated BCVs were surrounded by Rab 2 staining (Figure 3B). Together these results confirm the presence of GAPDH and Rab 2 on BCVs. We further studied the role of Rab 2 and GAPDH in HeLa cells, a well-established cell culture model for Brucella infection.
We showed that the GTPase Rab 2 is present on the vacuolar membrane of purified BCVs at 48 h p.i. (Figure 3A and 3B). In order to determine the kinetics of Rab 2 acquisition on BCV membranes, we first examined the presence of endogenous Rab 2 on BCVs within infected HeLa cells. Although we could detect Rab 2 on isolated BCVs, labelling of infected cells was extremely weak. As a consequence, we overexpressed Rab 2 within HeLa cells by using a dominant positive form of Rab 2: Rab 2 Q65L, which corresponds to Rab 2 locked in its GTP-bound form. HeLa cells were transfected with Myc Rab 2 dominant positive Q65L for 24 h and then infected with B. abortus GFP. The intracellular replication of Brucella within cells transfected or not with Q65L Rab 2 was similar (data not shown). Figure 4A represents the quantification of BCVs positive for Rab 2 Q65L at different times p.i. At 6 h p.i. few BCVs were surrounded by the active form of Rab 2 (20±4.2%). Then, at 10 h p.i. 68.4±5.7% of BCV had acquired Rab 2 Q65L and remained positive for Rab 2 Q65L until 48 h p.i. Figure 4B illustrates the recruitment of Rab 2 on BCVs at 10 h p.i. as indicated by the arrow. These results indicate that the recruitment of the active form of Rab 2 takes place just before the interaction of BCVs with the ER, known to occur around 12 h p.i [5]. Indeed, most of BCVs surrounded by Rab 2 Q65L were still LAMP-1-positive at 10 h p.i. (data not shown).
Retrograde vesicle formation from the VTC is mediated by the exchange of GDP to GTP on the GTPase Rab 2 [21]. In order to inhibit the retrograde transport between the ER and the Golgi, we used a dominant negative form of Rab 2: Rab 2 I119, which corresponds to Rab 2 locked in its GDP-bound form. HeLa cells infected with B. abortus Ds Red were transfected with either GFP plasmid as a transfection control, myc Rab 2, myc Rab 2 dominant negative I119, GFP Rab 1 or GFP Rab 1 dominant negative S25N. Figures 5A, 5B and S2A illustrate the level of intracellular replication at 48 h p.i. in the different transfected cells. We observed extensive Brucella replication in control cells or cells transfected either with GFP or myc Rab 2 (Figure 5A, 5B and S2A). Interestingly, transfections with either the GFP Rab 1 or its dominant negative did not affect Brucella replication, contrasting with results obtained from cells infected with Legionella pneumophila and transfected with the dominant negative Rab 1 S25N (Figure S2) [22]. On the contrary, in cells transfected with myc Rab 2 dominant negative I119 (Figure 5A and 5B), Brucella replication was strongly decreased (six times) and 89.8±1.62% of the bacteria were located in a lysosomal LAMP-1-positive compartment at 48 h p.i (Figure 5C and S2B), a time point were most of the Brucella (88%) are within an ER-positive, LAMP-1-negative compartments (Figure 5C and 5D). This shows that in cells transfected with the Rab 2 dominant negative form, Brucella was not able to reach the ER. Indeed, only a small percentage of BCVs were positive for cathepsin D at 48 h p.i. (22% of BCVs within cells overexpressing Rab 2 dominant negative) (Figure 5E). In contrast 92% of heat-killed Brucella BCVs were already cathepsin D-positive at 2 h p.i. (data not shown). Therefore, inhibition of retrograde vesicle formation from the VTCs mediated by Rab 2 affects Brucella replication. Taken together these results indicate that the trafficking between ER and Golgi controlled by Rab 2 is important for entry of Brucella in the ER and subsequent intracellular replication.
To investigate the role of GAPDH on Brucella pathogenesis, we down-regulated the expression of GAPDH in HeLa cells infected with B. abortus by using small interfering RNA. HeLa cells transfected with GAPDH siRNA for 72 h efficiently and specifically reduced the expression level of GAPDH (Figures 5F and 5H), whereas, siRNA control did not affect the GAPDH expression (Figures 5F and 5H). Inhibition of GAPDH expression induced a 10 fold reduction in Brucella replication as compared to non-transfected cells or cells transfected with the siRNA control (Figure 5G and 5H). This result indicates that the presence of GAPDH on the BCV membrane is required for Brucella replication. We showed above that in cells transfected with the Rab 2 dominant negative form, Brucella was not able to reach the ER and remained in a LAMP-1 compartment. Similarly, in siRNA GAPDH-treated cells, Brucella was found in a LAMP-1-positive compartment (Figure S3). In addition, inhibition of GAPDH expression prevented Rab2 recruitment on BCVs, as shown after BCV purification (Figure S3). Quantification showed that 77% of BCVs were positive for Rab 2 in control cells whereas only 12% of siRNA GAPDH-treated cells were able to recruit Rab 2. These results show that GAPDH is an important host factor for BCV biogenesis.
GAPDH is known to play several functions within host cells [16]–[18],[23]. To confirm the involvement of the retrograde transport of the early secretory pathway in Brucella pathogenicity, we investigated the role of other key components known to control the vesicular trafficking between ER and Golgi, such as the kinase PKC ι/λ and the coat COPI complex. Inhibition of these components was performed by infecting HeLa cells transfected with small interfering RNAs. We used a PKC ι siRNA to silence the expression of the kinase PKC ι/λ, a COP B siRNA to silence the subunit β of the COPI complex, a Rab 2 A siRNA to silence the GTPase Rab 2 and a α-Enolase siRNA to silence the Enolase, an enzyme involved in glycolysis. Cellular extracts prepared from HeLa cells transfected with the appropriate siRNA for 72 h efficiently and specifically reduced the expression of PKC ι, COPI, Rab 2 and Enolase (Figure 6A), whereas, siRNA control did not (Figure 6A). Figure 6B shows the intracellular replication of B. abortus at 48 h p.i. within infected-HeLa cells transfected with different siRNAs. Brucella replication was reduced 2.2, 2.3 and 5 fold in cells transfected with PKC ι siRNA, Rab 2 A siRNA and COP B siRNA, respectively as compared to cells transfected with the siRNA-A control. This result indicates that each member of the complex GAPDH/COPI/Rab2/PKCι/λ is required for Brucella replication. Surprisingly, we noticed that the intracellular replication of Brucella was reduced 4 fold under the inhibition of Enolase expression. This result suggests that host glycolysis is necessary for Brucella survival within host cells. Taken together, these results demonstrate the role played by the early secretory pathway, in particular the GAPDH/COPI/Rab2/PKCι/λ retrograde vesicles, to ensure replication within host cells at late stages of infection.
Many intracellular bacteria, with the aim of generating a suitable niche of replication, have been shown to alter the phagosomal membrane composition to avoid fusogenic interactions with lysosomes [24]. For example, Salmonella typhimurium secretes multiple effector molecules onto the vacuolar membrane (via its type III secretion systems), which interact with host proteins to modulate vesicle transport and vacuolar membrane dynamics [25]. This enables Samonella to replicate in a vacuole that interacts with late endosomes whilst avoiding fusion with lysosomes. In contrast, Legionella and Brucella replicate in ER-derived compartments [1],[26]. In the case of Legionella, several recent studies have highlighted the role of type IV secreted proteins in recruiting eukaryotic proteins to the vacuolar membrane and by mechanisms that are still unclear sustain intracellular replication [27]. Much less in known for Brucella, particularly regarding the membrane composition of BCVs. Previous work has demonstrated that the small GTPase Sar1 is implicated in Brucella intracellular survival [10]. However, apart from ER-resident proteins no other eukaryotic molecules have been associated with the BCV membrane.
Phagosomal proteomic studies using latex bead-containing phagosomes have significantly helped to decipher phagosome biology [28]. Using a modified procedure for phagosome purification, we analysed in detail the protein composition of the BCV membrane in order to identify eukaryotic proteins recruited to BCVs during intracellular replication. We developed a fractionation method to isolate intact BCVs from infected BHK-21 cells. This method allowed us to establish for the first time the BCV membrane protein map of the replicative niche of Brucella. As expected, a proportion of proteins on the BCV membrane were ER and ribosomal proteins.
Interestingly, one of the proteins identified was GAPDH, a non-ER eukaryotic protein which is normally located on VTCs between ER and the Golgi apparatus. This protein has been extensively studied by Tisdale et al [29]–[35]. GAPDH forms an active complex with the small GTPase Rab 2 and the protein kinase C (PKCι/λ), which is necessary for secretory vesicular transport. First studies showed that an inactive form of Rab 2 had a negative effect on anterograde transport of vesicles from the ER to the Golgi [21]. Recently, the group of Tisdale has shown that Rab 2 modulates protein retrograde transport from the Golgi to the ER by recruiting GAPDH to VTCs which allows the release of retrograde-directed vesicles [31],[35]. This retrograde transport requires a functional GAPDH/COPI/Rab2/PKCι/λ complex. Presence of GAPDH and Rab 2 on the BCV membrane suggests that Brucella is somehow interacting with VTCs or intercepting vesicle trafficking of the retrograde transport. Consistent with this hypothesis we found that inhibition of GAPDH resulted in reduced intracellular replication of Brucella. However, we cannot exclude that its role in the host cell glycolysis also contributes to the intracellular survival of Brucella. Indeed, silencing of enolase, another enzyme involved in glycolysis, also resulted in inhibition of Brucella replication. Further work is necessary to determine if Brucella is directly using the host glycolysis to its advantage, for example as a source of energy.
Nevertheless, the implication of retrograde transport in Brucella virulence is clearly demonstrated by the inhibition of the bacterial replication upon silencing of each member of the complex GAPDH/COPI/Rab2/PKCι/λ.
In addition, we found that Rab 2 is recruited on the BCV membrane before fusion with the ER suggesting that the retrograde transport might have an important role in the establishment of the bacterial replicative niche. Previous work has demonstrated that BCV-ER fusion events occur specifically at ER exit sites and are mediated by the small GTPase Sar1. This work also demonstrated that the anterograde pathway mediated by COPI/ARFI-coated vesicles are not involved in the BCV-ER fusion and that COPII complex formed on ER membranes was found in close apposition to BCV [10]. Our results also implicate the small GTPase Rab 2, in the early events of BCV biogenesis since inactivation of Rab 2 using the dominant negative prior to infection prevented fusion of BCVs with the ER. When inactivation is performed after infection, Brucella is able to start to replicate at 24 h (data not shown) whereas at 48 h p.i. there is a strong replication effect. Overall, these results suggest that Rab 2 is also necessary for Brucella survival after it has established its ER-derived replication niche.
Consistent with this hypothesis, we found that Brucella replication was also affected with prolonged brefeldin A treatment, which causes the Golgi apparatus to redistribute to the ER (data not shown). Regeneration of the Golgi apparatus after brefeldin A treatment allowed Brucella to recover and replicate (data not shown). Therefore, it is possible that trafficking of secretory vesicles from the Golgi apparatus to the ER is beneficial for Brucella replication. Perhaps, extensive bacterial replication requires an additional membrane input that may come from vesicles from the secretory pathway.
Brucella seems to manipulate host cells differently than Legionella pneumophila, another bacterial pathogen which replicates in the ER. Indeed, Legionella was shown to manipulate host cells by secreting specific type IV secretion system bacterial effectors, for example DrrA (or LidA) [36],[37]. DrrA is able to mimic the guanosine exchange factor (GEF) of the small GTPase Rab 1 to catalyse the exchange of GDP to GTP which lead to the activation and the recruitment of Rab 1 on the Legionella-containing vacuole [36],[37]. Interestingly, another secreted effector LepB acts like a GTPase-activating protein (GAP) which inactivates Rab 1 [38]. Inhibition of Rab1 impairs Legionella replication [22]. In this study, we demonstrate that Rab 1 is not involved in intracellular replication of Brucella.
Rab 1 is implicated in the anterograde pathway whereas Rab 2 is also essential for the retrograde pathway. After 24 h of replication within its ER-derived vacuole, Legionella lyses its vacuole to reach the cytosol and to infect the neighbouring host cells [39]. This suggests that Legionella does not need an influx of membrane since the bacteria lyse the host cell relatively quickly compared to Brucella, which continues to replicate within vacuoles for a long period of time (at least another 48 h) [1]. Therefore it is possible that Brucella might require an extensive influx of membrane for its own survival. Targeting of the retrograde pathway could ensure continued fusion with incoming vesicles and an influx of membrane.
At this stage, it is unclear how Brucella is controlling biogenesis of ER-derived BCVs. It is possible that some secreted effector proteins not yet identified are directly recruiting GAPDH and Rab 2 to the BCV membrane. The type IV secretion system VirB, essential for sustained interaction and fusion with ER membranes [1], could secrete these effectors to recruit Rab 2 onto the BCV membrane. We are currently undertaking further studies to identify the bacterial effector(s) which target(s) Rab 2 and GAPDH.
The bacterial strains used in this study were the smooth virulent B. abortus strain 2308 GFP [1] or 2308 Ds Red kindly provided by Jean-Jacques Letesson (URBM, Immunology Laboratory, FUNDP, Namur, Belgium). Bacteria were inoculated in tryptic soy broth (TSB. Sigma-Aldrich) with kanamycin and grown at 37°C overnight (16 h) as previously described for infections [1].
GFP vector used was pEGFP-C1 (Clontech). Plasmids pGEM Rab 2, pGEM Rab 2 I119 and pGEM Rab 2 Q65L were kindly given by Craig Roy (Section of Microbial Pathogenesis, Yale University School of Medicine, New Haven, USA). The rab2 sequence from the vectors pGEM Rab 2, pGEM Rab 2 I119 and pGEM Rab 2 Q65L was amplified by PCR with two specific primers: O-403 (5′GGGGACAAGTTTGTACAAAAAAGCAGGCTTCGCGTACGCCTATCTCTTCAAGTACAT3′) and O-404 (5′GGGGACCACTTTGTACAAGAAAGCTGGGTCCTAACAGCAGCCGCCCCCAGCCTGCTG3′). The myc tag (pCMV-myc) was added by Gateway procedure according to the manufacturer's instructions (Invitrogen) to obtain myc Rab 2, myc Rab 2 I119 and myc Rab 2 Q65L. Plasmids GFP Rab 1 and GFP Rab 1 S25N were also kindly given by Craig Roy. The silencing of endogenous GAPDH was performed by small interfering RNA with the commercial siRNA GAPDH (Applied Biosystems Ambion). A negative control siRNA was also used (Applied Biosystems Ambion) as a transfection control. Silencing of endogenous PKC ι, COP B, Rab 2 A and α-Enolase were performed by small interfering RNA with the corresponding commercial siRNA (Santa Cruz Biotechnology). A negative control siRNA-A was also used (Santa Cruz Biotechnology) as a transfection control.
The primary antibodies used were: a mouse monoclonal anti-actin given by Lena Alexopoulou (CIML, Marseille, France); a rabbit polyclonal anti-calnexin (Stressgen); a rabbit polyclonal anti-cathepsin D [40]; a goat polyclonal T-14 COP B (Santa Cruz Biotechnology); a mouse monoclonal anti-GAPDH (Sigma Aldrich); a rabbit polyclonal anti-giantin (Covance); a mouse monoclonal anti-GM130 (Transduction Laboratories); a rabbit polyclonal anti-LAMP-1 (Abcam); a mouse monoclonal 4A1 anti-LAMP-1 kindly given by Jean Gruenberg (University of Geneva, Switzerland); a mouse monoclonal 9E10 anti-c-myc tag (Santa Cruz Biotechnology); a mouse monoclonal anti-Omp 25 (A59/05F01/C09) kindly given by Michel Zygmund (INRA, Tours, France); a mouse monoclonal H-12 PKC ι (Santa Cruz Biotechnology); a rabbit polyclonal anti-Rab 1 (Santa Cruz Biotechnology); a polyclonal rabbit anti-Rab 2 and a mouse monoclonal anti-Rab 5 kindly given by Marino Zerial (Max Planck Institute of Molecular Cell Biology and Genetics, Dresden, Germany); a rabbit polyclonal anti-Rab 7 [40] and a rabbit polyclonal anti-VDAC 1 (Abcam). The secondary antibodies used were: goat anti-mouse HRP (Sigma Aldrich); goat anti-rabbit HRP (Sigma Aldrich) for western blotting; and donkey anti-rabbit Texas red (Jackson ImmunoResearch); donkey anti-mouse Texas red (Jackson ImmunoResearch); donkey anti-mouse Cy3 (Jackson ImmunoResearch); donkey anti-mouse FITC (Jackson ImmunoResearch); donkey anti-mouse Cy5 (Jackson ImmunoResearch); phalloidin-tetramethylrhodamine isothiocyanate (TRITC. Sigma Aldrich) for immunofluorescence and a IgG anti-rabbit coupled to phycoerytrin (PE. Serotec) for flow cytometry.
Baby hamster kidney (BHK-21) cells were cultured at 37°C with 5% CO2 atmosphere in GMEM (Glasgow's modified Eagle's medium. Gibco) supplemented with 10% Tryptose Phosphate Broth (Sigma Aldrich), 5% FCS (Perbio) and 1% L-glutamine (Gibco) and seeded 24 h before infection on 55 cm2 culture dishes (1.6×106 cells per dish) for fractionation or at a surface ratio of 1/10 in 24-well plates containing 12-mm glass coverslips for immunofluorescence and CFUs. HeLa cells and Raw 264.7 macrophages were cultured at 37°C in a 5% CO2 atmosphere in DMEM supplemented with 10% FCS, 1% non-essential amino acids and 1% L-glutamine and seeded 24 h before infection at a surface ratio of 1/10 in 24-well plates containing 12-mm glass coverslips. Bone marrow-derived macrophages (BMDM) were isolated from femurs of 6 to 10-week-old C57Bl/6 female mice and differentiated into macrophages as previously described [41].
Infections were performed at a multiplicity of infection of 200∶1 by centrifuging bacteria onto BHK-21, HeLa cells, BMDM or Raw 264.7 macrophages at 400 g for 10 min at 4°C, and then by incubating the cells for 1 h (for BHK-21 and HeLa cells) or 15 min (for BMDM and Raw 264.7 macrophages) at 37°C under a 5% CO2 atmosphere. Cells were extensively washed with their respective medium to remove extracellular bacteria and were incubated for an additional hour in their respective medium supplemented with 50 µg/ml gentamycin to kill extracellular bacteria. Thereafter, the antibiotic concentration was decreased to 10 µg/ml.
To monitor Brucella intracellular survival, infected cells were washed three times with PBS and lysed with 0.1% (vol/vol) Triton X-100 in PBS. Serial dilutions in PBS of lysates were plated onto TSB agar plates to enumerate CFUs.
The expression of GFP, myc Rab 2, myc Rab 2 I119, myc Rab 2 Q65L, GFP Rab 1, GFP Rab 1 S25N were performed by transfecting HeLa cells using the FuGENE transfection reagent (Roche), according to the manufacturer's instructions. Depending of experiment, transfections were performed either 24 h before or 2 h after Brucella infection and were left to proceed until the time of analysis.
The transfection of small interfering RNA si GAPDH and siRNA control were performed on HeLa cells using the siPORT™ Amine transfection agent (Applied Biosystems Ambion) according to the manufacturer's instructions. The transfection of PKC ι, COP B, Rab 2 A, α-Enolase siRNAs and control siRNA-A were performed on HeLa cells using the siRNA Transfection Reagent (Santa Cruz Biotechnology) according to the manufacturer's instructions. 24 h later, transfected HeLa cells with the specific siRNA or siRNA-A control were infected with B. abortus as described before. To maintain the silencing of these specific proteins, HeLa cells were again transfected 2 h p.i.
Cellular extracts were prepared after lysis of HeLa cells transfected 72 h with either siGAPDH or siRNA control with 0.1% (vol/vol) Triton X-100 in PBS.
At 48 h p.i., 6 infected BHK-21 55 cm2 dishes were washed once with GMEM. 1 ml of GMEM was added and infected cells were recovered by scraping with a rubber policeman. Then, several steps of washes were done with PBS, PBS/1 mM EDTA, pH 7.4, homogenization buffer (3 mM imidazole/250 mM sucrose/0.5 mM EDTA/0.5 mM EGTA, pH 7.4). A centrifugation at 80 g for 5 min at 4°C was performed between each wash. Pellets were resuspended very gently in the homogenization buffer and cells were mechanically broken through 5 passages into a 22G needle. PNS were recovered after a centrifugation at 80 g for 10 min at 4°C. Then, a first step of purification was performed by loading the PNS on the top of a 50%–12% sucrose gradient. After centrifugation at 800 g for 45 min at 4°C, a cloudy layer containing BCVs appeared at the 50%–12% interface. This cloudy layer (called interface fraction) was carefully recovered and loaded at the bottom of a SW 60 centrifugation tube. Three layers of sucrose were sequentially added on top of the interface fraction: 30%, 20% and 5% sucrose respectively. The enriched BCV fraction was obtained after an ultracentrifugation at 35000 rpm for 1 h at 4°C and was localized in the pellet. The pellet was then resuspended in the homogenization buffer. Depending of experiments, a supplementary step was added between the two steps of sucrose gradient in order to eliminate mitochondria from the enriched BCV fraction. This step consisted of incubating the interface fraction with dynabeads (M-500 subcellular, Invitrogen) coated with a rabbit anti-VDAC 1 antibody (according to the manufacturer's instructions) overnight recovering the supernatant from dynabeads retained with a magnet (according to the manufacturer's instructions).
BCV staining within PNS was adapted from what was previously described [40]. PNS was incubated with a rabbit anti-calnexin antibody and then incubated with a rabbit-PE antibody 30 min on ice. The preparation was fixed for 20 min in 3% final PFA and then diluted to 1% final PFA before analysis on a FACScalibur cytometer (Becton Dickinson). Data were analysed using FlowJo software (Tree Star).
250 µg of the enriched fraction of BCVs was treated for 30 min at room temperature with 0.1% Triton X-100. The BCV membranes were separated from bacteria by centrifugation at 10000 rpm for 5 min at 4°C. The corresponding supernatants enriched in BCV membrane proteins were precipitated with trichloroacetic acid (Sigma Aldrich) 10% final for 5 min on ice and then with trichloroacetic acid 5% final for 5 min on ice. Trichloroacetic acid was then removed by three washes with 90% acetone. Each step was followed by a centrifugation at 10000 rpm for 3 min at 4°C. The pellet was finally resuspended in 400 µl of destreack rehydration buffer containing 2% carrier ampholytes pH 3–10 (IPG Amersham Biosciences).
BCV membrane proteins were separated by isoelectric focusing (IEF). The IEF was performed using 18 cm gels with an immobilized linear pH gradient of 3–10 (Immobiline DryStrips, Amersham Biosciences) in IPGphor strip holders (Amersham Biosciences) on a MultiphorII machine (Amersham Biosciences). The IEF protocol was as follows: 300 V for 1 min; 500 V gradient for 30 min; 3500 V gradient for 1.5 h; 3500 V for 6 h. Temperature was set at 20°C. Prior to SDS PAGE, IPG strips were equilibrated during 20 min in an equilibration buffer (6 M urea, Tris, pH 8.8, 50 mM, 2% SDS, 65 mM DTT, 38.5% glycerol). The second dimension was performed using a Protean II xl Multicell separation unit (Biorad) and home-made 10% SDS PAGE gels. Temperature was set at 20°C. Gels, made of Tris-HCl, 0.1% SDS and 10% acrylamide were run at 20°C using the following running buffer (25 mM Tris, 192 mM glycine and 0.1% SDS) for the cathode part and 2×running buffer for the anode. Electrophoresis was conducted at 10 mA per gel overnight and stopped when the bromophenol blue front dye reached the bottom of the gel. Proteins were stained by a PlusOne Silver Staining Kit, Protein (GE Healthcare) according to the manufacturer's instructions (without glutaraldehyde to allow the mass spectrometry analysis).
Protein spots immediately excised from silver-stained gels were destained and subjected to in-gel digestion with trypsin (Sequencing grade modified porcine trypsine; Promega, Madison, WI, USA) according to a modified protocol from Shevchenko et al. [42]. Tryptic peptides were then extracted from the gel by successive treatment with 5% formic acid and 60% acetonitrile/5% formic acid. Extracts were pooled and dried in a Speedvac evaporator. Peptides resuspended in an α cyano-4-hydroxycinnamic acid matrix solution (prepared by diluting 6 times a saturated solution in 50% acetonitrile/0.3% trifluoroacetic acid), were then spotted on the metal target. Mass analyses were performed on a MALDI-TOF Bruker Ultraflex spectrometer (Bruker Daltonics, Wissembourg, France). Mass spectra were internally calibrated using autolytic peptides from trypsin.
The peptide mass lists were used to identify the protein using Mascot software available on site. Criteria used for protein identification are given by Mascot as a Probability Based Mowse Score. Ions score is −10*Log(P), where P is the probability that the observed match is a random event. Protein scores greater than X (X is a number between 60 and 74, from one search to the other) are significant (p<0.05).
The protein concentration was determined with the BCA™ Protein Assay Kit (Pierce). Volumes corresponding to 60 µg of proteins from each main step of the fractionation were resuspended in 1×laemmeli-buffer and loaded on a 12% SDS polyacrylamide gel. Then proteins were transferred onto a PVDF membrane using a semi-dry transfer. The PVDF membrane was then blocked for 1 h with 4% milk/PBS/0.1% Tween 20 and incubated with different antibodies. The PVDF membrane was washed 3 times with PBS/0.1% Tween 20 before incubation with the secondary antibody and the detection. The detection was carried out using the ECL™ western blotting detection kit (Amersham).
To analyse the PNS and the sucrose step gradient enriched in BCVs, 10 µl of sample were put in 24-well plates containing 12-mm glass coverslips pre-treated with poly-L-lysine and incubated for 15 min at 37°C to allow the adherence onto glass coverslips. Then BCVs or infected cells were fixed with 3% paraformaldehyde, pH 7.4 at room temperature for 15 min, and then processed for immunofluorescence staining as described previously [5]. Specimens were observed on a Zeiss LSM 510 laser scanning confocal microscope for image acquisition. Images of 1024×1024 pixels were acquired and assembled using Adobe Photoshop CS2.
B. abortus-infected BHK cells were fixed for 1 h at room temperature with 2.5% glutaraldehyde (Sigma, St Louis, MO, USA) in 0.1 M cacodylate buffer, pH 7.2, containing 0.1 M sucrose, 5 mM CaCl2 and 5 mM MgCl2. After two successive 15 min washes with the same buffer, cells were postfixed for 1 h at room temperature with 1% osmium tetroxide (Electron Microscopy Sciences, Hatfield, PA, USA) in the same buffer devoid of sucrose. The cells were scraped off the culture dishes with a rubber policeman and concentrated in 2% agarose in the same buffer. After 1 h incubation at room temperature with 1% uranyl acetate in veronal buffer, the samples were dehydrated in a graded series of acetone and embedded in Epon resin. Thin sections were stained with uranyl acetate and lead citrate. The PNS and enriched BCVs obtained from the BHK-infected cells were first prefixed for 20 min at room temperature with 5% glutaraldehyde in cacodylate buffer diluted at a 1∶1 volume ratio with the PNS or BCV fractions. These fractions were then processed as described above for BHK-infected cell.
|
10.1371/journal.ppat.1002765 | The Link between Morphotype Transition and Virulence in Cryptococcus neoformans | Cryptococcus neoformans is a ubiquitous human fungal pathogen. This pathogen can undergo morphotype transition between the yeast and the filamentous form and such morphological transition has been implicated in virulence for decades. Morphotype transition is typically observed during mating, which is governed by pheromone signaling. Paradoxically, components specific to the pheromone signaling pathways play no or minimal direct roles in virulence. Thus, the link between morphotype transition and virulence and the underlying molecular mechanism remain elusive. Here, we demonstrate that filamentation can occur independent of pheromone signaling and mating, and both mating-dependent and mating-independent morphotype transition require the transcription factor Znf2. High expression of Znf2 is necessary and sufficient to initiate and maintain sex-independent filamentous growth under host-relevant conditions in vitro and during infection. Importantly, ZNF2 overexpression abolishes fungal virulence in murine models of cryptococcosis. Thus, Znf2 bridges the sex-independent morphotype transition and fungal pathogenicity. The impacts of Znf2 on morphological switch and pathogenicity are at least partly mediated through its effects on cell adhesion property. Cfl1, a Znf2 downstream factor, regulates morphogenesis, cell adhesion, biofilm formation, and virulence. Cfl1 is the first adhesin discovered in the phylum Basidiomycota of the Kingdom Fungi. Together with previous findings in other eukaryotic pathogens, our findings support a convergent evolution of plasticity in morphology and its impact on cell adhesion as a critical adaptive trait for pathogenesis.
| Although morphogenesis and virulence are commonly associated in many eukaryotic pathogens, the nature of such association is often unknown. For example, Cryptococcus neoformans, a fungal pathogen that causes cryptococcal meningitis, typically undergoes morphological transition between the yeast and the filamentous form during mating. However, molecules that are critical for mating do not directly impact fungal virulence. Thus, the nature of the long observed association between morphotype and virulence in this microbe remains elusive despite decades of effort. Here we demonstrate that constitutively activated pheromone signaling is insufficient to drive morphological transition under mating-suppressing conditions, including those relevant to host physiology. Rather, we demonstrate that sex-independent morphological switching is driven by the transcription factor Znf2 and this regulator controls the ability of this fungus to cause disease. Znf2 governs Cryptococcus morphotype and virulence potential at least partly through its effects on cell surface proteins. One novel adhesin Cfl1functions downstream of Znf2 and it orchestrates morphological switch, cell adhesion, biofilm formation, and pathogenicity. Thus, cell adhesion at least partly underlies the link between morphological transition and pathogenicity in C. neoformans. Our findings provide a platform for further elucidation of the impact of morphotype on virulence in this ubiquitous pathogen. The discovery of Cfl1 and other novel adhesins in Cryptococcus could lay a foundation for the development of vaccines or alternative therapies to combat the fatal diseases caused by this fungus.
| Adaptation to the host environment by many eukaryotic pathogens is often companied by transition in cellular morphology [1], [2], [3], [4], [5], [6], [7], [8], [9]. The ubiquitous fungal pathogen Cryptococcus neoformans causes more than half a million deaths each year [10]. It can grow in the yeast form as well as the filamentous form. Earlier pre-genetic studies indicate an inverse relationship between filamentation and virulence [11], [12], [13], [14], [15], [16], [17]. These studies also point to the potential of filament-specific antigens as vaccines against Cryptococcus infections [18], [19], [20].
Because Cryptococcus typically grows in the yeast form and the morphological transition from the yeast form to the filamentous form appears to be coupled with mating, signaling pathways that lead to bisexual mating (a-α mating) and unisexual mating (mostly α-α mating) have been intensively investigated [21], [22], [23], [24]. The roles of these signaling components in fungal pathogenicity are also scrutinized in animal models. However, accumulating evidence indicates that key signaling components that specifically lead to mating, such as those in the pheromone sensing pathway, have no or minimal direct effect on virulence [25], [26], [27], [28]. Furthermore, conditions relevant to host physiology (e.g. aqueous environment, high temperatures, and high levels of CO2) are mating-suppressive, suggesting sex-independent mechanisms in orchestrating morphotype and virulence in Cryptococcus [29]. Therefore, the existence and the nature of the link between morphological transition and virulence in Cryptococcus remain enigmatic.
Although Cryptococcus morphological transition from the yeast form to the filamentous form is historically associated with mating, the observations that filamentation can be achieved in strains in the absence of key pheromone signaling components or meiotic genes [30], [31], [32], [33], lead us to hypothesize that pheromone signaling pathways are not essential or sufficient for filamentation per se, but they are critical in stimulating filamentation in response to mating cues. To test this hypothesis, we decided to examine the effect of constitutive activation of the pheromone signaling circuit on morphogenesis under mating-inducing and mating-suppressing conditions.
It is known that the expression of genes in the pheromone signaling pathway, such as those encoding the pheromone Mf1α, the pheromone receptor Ste3α, the pheromone transporter Ste6, and the key pheromone response regulator Mat2 (Figure 1A), is low under mating-suppressing conditions but is dramatically higher during a-α bisexual mating (Figure 1B and data not shown) [30], [32]. We found that the expression level of these pheromone signaling genes in wildtype α strain H99 alone was low when cells were cultured on either mating-inducing condition (V8 agar) or mating-suppressing conditions (YPD agar and serum) (Figure 1B, C, D, E and F). This is consistent with the well-noted poor ability of the H99 strain to undergo unisexual mating. In fact, filamentation has never been observed when H99 was cultured alone under mating-inducing conditions (Figure 2A) [34]. We chose this strain to study the link between morphogenesis and virulence because H99 is one of the most virulent clinical strains tested in various animal models and it is also widely used as a reference strain in Cryptococcus research.
When we placed the MAT2 gene under the control of the constitutively active promoter of GPD1 (glycerol-3-phosphate dehydrogenase 1) and introduced this construct to H99, the transcript level of MAT2 was dramatically increased under mating-inducing as well as mating-suppressing conditions (Figure 1C). As expected for a key regulator of the pheromone signaling, overexpression of MAT2 led to high expression of MF1α, STE3α, and STE6 under both mating-inducing and mating-suppressing conditions (Figure 1D, E and F). This result indicates that constitutively overexpression of MAT2 is sufficient to induce pheromone signaling circuit independent of mating cues.
We next tested the effect of activation of pheromone signaling on filamentation under different conditions. The PGPD1-MAT2 conferred filamentation to H99 during unisexual mating (α cells alone) and it significantly enhanced filamentation during bisexual mating (a-α coculture) under the mating-inducing condition (Figure 2A). However, under mating-suppressing conditions, overexpression of MAT2 failed to stimulate filamentation in the α alone culture or in the a-α coculture (Figure 2A), and this was not due to insufficient activation of pheromone signaling. Effective activation of pheromone signaling in the PGPD1-MAT2 strain is supported by both the high expression levels of genes involved in pheromone signaling (Figure 1D, 1E, and 1F) and the formation of shmoo-like cells under both mating-inducing and mating-suppressing conditions (Figure 2B). Shmoo-like cells are typically observed when cells respond to mating signals prior to cell fusion. These observations indicate that activation of pheromone signaling alone is not sufficient to initiate filamentation under mating-suppressive conditions, including conditions relevant to host physiology. Thus mating signaling is unable to coordinate the yeast-filament morphological transition and virulence during infections.
We previously showed that the deletion of ZNF2, which encodes a zinc-finger transcription factor, locked cells in the yeast form during mating without impairing pheromone signaling [28]. This suggests that Znf2 is not essential for mating signal relay; rather, it is crucial for filamentation. Although Znf2 functions downstream of Mat2 during mating [28] and its gene expression was significantly induced by MAT2 overexpression under the mating-inducing condition (Figure 2C and 2D), activation of the pheromone signaling pathway was unable to induce ZNF2 expression in the absence of mating stimuli. This was evidenced by the low expression level of ZNF2 in the PGPD1-MAT2 strain under mating-suppressing conditions (Figure 2C and 2D). The ability of Cryptococcus to undergo filamentation correlates with the expression level of ZNF2, but not that of MAT2 (Figure 1C, Figure 2A, 2C and 2D). Thus, Znf2 could be a master regulator that dictates Cryptococcus morphotype irrespective of environmental stimuli or mating type.
To test this hypothesis, we constructed the PGPD1-ZNF2 strains. Indeed, the PGPD1-ZNF2 triggered filamentation in Cryptococcus strains of either mating types a or α in both serotype A and serotype D backgrounds under all tested conditions, including those that are inducing or suppressive to mating (Figure 3A and Figure S1). In contrast to the PGPD1-MAT2 strain, filaments produced by the PGPD1-ZNF2 strain under mating-inducing condition maintained their filamentous morphology after being transferred to mating-suppressive conditions (Figure S2). However, it is notable that the PGPD1-ZNF2 strain produces more robust hyphae under mating-inducing condition, suggesting that other factors induced under mating-inducing condition could further activate Znf2.
The PGPD1-ZNF2 also conferred filamentation to mutants that harbor deletions in the key mating components under various conditions tested (Mfα1-3, Mat2, or Ste12 functioning in a branching pathway in pheromone signaling) (Figure 3B). To confirm that filamentation conferred by Znf2 activation is not due to some cryptic restoration of mating ability, we measured the efficiency of cell fusion of the wildtype, the mat2Δ mutant, and the mat2Δ+PGPD1-ZNF2 strain during bisexual a-α mating. Indeed, overexpression of ZNF2 did not rescue the cell fusion defects of the mat2Δ mutant (Figure 3C). Consistently, gene ontology analyses of our previous transcription data indicate that Znf2, unlike Mat2, does not regulate genes involved in the cell fusion event critical for mating (Figure S3) [28]. Taken together, the results indicate that filamentation can be independent of mating and Znf2 is one key determinant of this sex-independent morphogenesis.
To verify the correlation of ZNF2 expression and Cryptococcus morphology, we constructed the ZNF2 gene driven by two inducible promoters: the galactose-inducible GAL10 promoter (data not shown) [35] or the copper transporter CTR4 promoter (Figure 3D) (copper deprivation–on; copper repletion–off) [36]. Transformation of the PGAL10-ZNF2 or the PCTR4-2-ZNF2 construct into wildtype either the serotype D reference strain JEC21 or the serotype A reference strain H99 conferred filamentous growth under promoter-inducing conditions. These strains grew as yeasts under promoter-repressive conditions (Figure 3D and data not shown). Increasing the concentration of the copper chelator BCS (inducer) increased the frequency of filamentation in the PCTR4-2-ZNF2 strain (Figure 3D), indicating that the expression level of ZNF2 dictates Cryptococcus cellular morphology.
To examine the effect of ZNF2 on the dynamic morphological transition, we incubated the PCTR4-2 -ZNF2 strain in H99 background in liquid YPD medium containing 200 µM BCS (inducer) and examined cell morphology over time. Morphological transition from the yeast form to the filamentous form completed by 60 hours (Figure 4). At this time, the hyphae were transferred to YPD medium containing copper sulfate (inhibitor). Cryptococcus cells then switched from the filamentous form to the yeast form over time (Figure 4). The control of bi-directional morphological transition by Znf2 is also observed when cells were cultured in serum (data not shown), indicating that this control is independent of environmental cues. These results demonstrate that (i) the expression level of ZNF2 determines Cryptococcus cell morphology: high expression level of ZNF2 drives the cells to the filamentous form and low expression level of ZNF2 renders cells unicellular yeast; (ii) Znf2 is necessary and sufficient to initiate morphological transition; (iii) High Znf2 activity is required to maintain cells in the filamentous morphotype.
The relationship between morphotype and pathogenicity is typically defined through studying morphological mutants that are otherwise isogenic to the wildtype strains and are able to maintain given morphotype under host relevant conditions, even though mutants with such extreme phenotypes are unlikely to be encountered clinically due to natural selections in the host [1], [2], [3], [4]. For Cryptococcus, host physiological environment (e.g. high body temperature, aqueous environment, and high levels of CO2) is extremely inhibitory to mating. Consistently, constitutively activated mating signaling induced filamentation under mating-inducing conditions, such filaments could not be maintained when transferred to in vitro conditions that mimicked host physiological environment (Figure S2). In contrast, the PGPD1-ZNF2 strain can readily initiate and maintain filamentous growth under such host-relevant conditions (Figure S2). Thus ZNF2 overexpression strains could serve as a model to investigate the relationship between morphotype and pathogenicity.
We tested the virulence of the wildtype H99 and the PGPD1-ZNF2 strain in the murine inhalation model of cryptococcosis. The PGPD1-ZNF2 strain exhibited heterogeneity in cell morphology and a mixture of cell types is always present in this strain. To obtain accurate inoculation and to avoid potential problems caused by differences in cell types at initial infection, only cells in the yeast form were used for animal inoculation. Remarkably, the PGPD1-ZNF2 strain was completely avirulent (Figure 5A). By day 60 post infection (DPI 60) when the study was terminated, the PGPD1-ZNF2 cells were either completely cleared from animal lungs or existed in very low numbers (1000 fold lower than the original inocula). We further examined the fungal burden in the lungs and the brain of animals infected with H99, the znf2Δ mutant, and the PGPD1-ZNF2 strain at DPI 10 before any animal succumbed to cryptococcosis. Consistent with the animal survival rates, the lung fungal burden in animals infected with the znf2Δ mutant and the PGPD1-ZNF2 strain was 236% and 0.6% respectively compared to those infected with the wildtype (Figure 5B). The brain fungal burden showed a similar trend with larger variations due to individual differences in the timing of dissemination in this inhalation model (Figure S4), and no fungal cells were recovered from the brains of animals infected by the PGPD1-ZNF2 strain. To examine the effects of Znf2 on fungal morphology in vivo, we infected animals intranasally with H99, the znf2Δ mutant, and the PGPD1-ZNF2 strain and performed histological examination of lung tissues at DPI 1, 7, and 12. Remarkably, even though only yeast cells from the PGPD1-ZNF2 strain were used in the original inoculation into animals, lungs infected by the PGPD1-ZNF2 strain contained Cryptococcus cells of mixed morphology: yeast, pseudohyphae, and hyphae in all the time points examined (Figure 5C and Figure S5). This is consistent with the morphological heterogeneity of the PGPD1-ZNF2 strain in vitro (Figure S2). In comparison, only yeast cells were observed in the wildtype H99 or the znf2Δ mutant infected animals (Figure 5C and Figure S5). This histological examination indicates that activation of Znf2 can drive filamentation in vivo.
Tolerance of host temperatures is a pre-requisite of fungal virulence. In some fungal pathogens, morphological changes are often a response to temperature and some morphological defective mutants lose the ability to cause diseases in mammalian hosts due to growth inhibition by high temperatures in vivo. To determine if alteration of virulence potential in the znf2 mutants are simply due to altered sensitivity to high temperature, we compared the growth of the wildtype H99, the znf2Δ mutant, and the PCTR4-2-ZNF2 strain at 30°C and 37°C on a variety of media via the spot assay. No apparent growth defects were observed in the znf2Δ mutant or the ZNF2 overexpression strain when compared to the wildtype under the conditions tested (Figure S6). Furthermore, the observation that the ZNF2 overexpression strain was capable of amplification during early stages of infection based on the fungal burden time course experiment (Figure S7) also suggests that factors other than growth inhibition by high temperature are mainly responsible for the effects of Znf2 on virulence.
As morphological changes reflect changes in cell surface properties, we predict that Znf2 controls cell surface constitutes. One property likely regulated by Znf2 is cell adhesion, as supported by the following observations. First, increasing the ZNF2 expression led to increasingly wrinkled colony morphology and flocculation (Figure 3D, and Figure 6A, B and C). Both phenotypes are likely caused by increased expression of flocculins (adhesins or adhesion proteins), as previously shown in bacteria and in yeasts [37], [38]. Second, aerial hyphae of the ZNF2 overexpression strains formed on solid media also tended to attach to each other, forming bundles (Figure 6D), as observed in flocculated strains of the filamentous fungus Ashbya gossypii [39]. Third, deletion of ZNF2 impairs agar invasion whereas overexpression of ZNF2 remarkably promoted invasive growth (Figure 6E), and invasive growth reflects cell-substrate adhesion. The results suggest that Znf2 plays a pivotal role in morphogenesis-associated cell flocculation in Cryptococcus.
Given that Cryptococcus strains with increased flocculation are reduced in virulence [40], [41], this transcription factor likely impacts pathogenicity at least partly through its effects on cell adhesion. Ontology analysis of our previous transcriptional profiling data [42] revealed that of those genes that are differentially expressed in the znf2Δ mutants, 23% encode secretory proteins based on WoLF PSORT prediction (http://wolfpsort.org/) (Figure 7A). We selected 9 such genes and examined their transcript level in a ZNF2 overexpression strain incubated in serum at 37°C in 5% CO2 by quantitative realtime PCR. All genes tested were also differentially expressed in the ZNF2 overexpression strain (Figure 7B).
We overexpressed these 9 genes using the constitutively active GPD1 promoter and examined if their overexpression could recapitulate some of the phenotypes caused by the ZNF2 overexpression (Figure 7C). Interestingly, strains with overexpression of CNAG_00795 (designated as CFL1: Cell FLocculin 1) formed extremely wrinkled colonies, like ZNF2 overexpression strains (Figure 6A). Interestingly, the expression of CFL1 was also most dramatically induced by the ZNF2 overexpression (Figure 7B). Because acapsular Cryptococcus mutants also form wrinkled colony, we examined capsule production in the CFL1 overexpression strain and cfl1Δ mutants. No apparent defect in capsule production was detected based on microscopic examination (data not shown).
To confirm that cell adhesion is indeed caused by increased CFL1 expression, we then constructed PCTR4-2-CFL1 strains. These strains grew as yeast cells in liquid cultures. A sharp increase in cell aggregation was observed when PCTR4-2-CFL1 cells were cultured under promoter-inducing conditions, a reminiscence of some of the phenotypes of the PCTR4-2-ZNF2 strains (Figure 7D).
To further confirm that CFL1 is regulated by Znf2, we engineered a reporter strain where ZNF2 expression is inducible by galactose and the fluorescent Cfl1 is driven by its native promoter. We grew the reporter strain under mating-suppressing conditions to avoid complication due to potential activation of mating signaling. Under such conditions, the colony formed by the reporter strain became fluorescent and wrinkled when the ZNF2 expression was induced in the presence of galactose (Figure 8A and B), while the colony was non-fluorescent and smooth when the ZNF2 expression was inhibited in the presence of glucose (Figure 8A and B). Thus the expression of the fluorescent Cfl1 is driven by Znf2. Taken together, Znf2 triggers morphological switch as well as flocculation (cell adhesion), and its downstream factor Cfl1 regulates cell adhesion.
We examined the sub-localization of Cfl1 using a strain harboring the mCherry fused Cfl1 protein driven by its native promoter. Because Cfl1 is induced during mating and controlled by key components of mating signaling (Figure S8A and B), we examined microscopically the expression of CFL1-m-cherry during mating. Cfl1 was rarely detected in yeast cells (Figure 8C), but it was highly expressed in hyphae during both unisexual mating and bisexual mating (Figure 8D). The fluorescent Cfl1 delineated the periphery of hyphal cells, consistent with the function of adhesins on the cell surface and the prediction that Cfl1 is a secretory protein based on the presence of an N-terminal signal peptide for secretion.
Secretion is required for Cfl1's function as an adhesin. This is supported by the observation that overexpression of the fluorescent Cfl1 that lacks the N-terminal signal peptide [Cfl1(sigPΔ)-mCherry] failed to confer wrinkled colony morphology or cell aggregation to Cryptococcus (Figure S9 and Figure 8E). This is not due to a failure of producing the mutant allele protein, as abundant Cfl1(sigPΔ)-mCherry protein was produced by the cells (Figure 8E). However, no fluorescence was detected from the culture supernatant (Figure 8F), indicating defects in secretion. A few other fungal adhesins are also reported to be associated with cell surface as well as being released into surrounding environment [43], [44]. Such property may facilitate their roles in mediating both cell-cell adhesion and cell-substrate adhesion, and it may also help circumvent the blockage by other extracellular components. Consistent with its role as an adhesin, Cfl1 regulates a broad spectrum of cell adhesion-related biological processes, including complex colony morphology [45], [46] and formation of different biofilms (Figure S10).
Remarkably, deletion of CFL1 dramatically reduced hyphal production during either bisexual or unisexual mating while overexpression of CFL1 enhanced the hyphal formation (Figure 9A and B). Thus, both the expression pattern of CFL1 and the observed effects of CFL1 deletion or overexpression on hyphal development indicate the importance of this adhesin in hyphal morphogenesis. Like the ZNF2 overexpression strain, hyphae formed by the PGPD1-CFL1 strain on YPD medium (mating-suppressive) tended to attach to each other, forming bundles (Figure 6C and Figure 9A).
Previous studies implicate an inverse association between flocculation and virulence in Cryptococcus [40], [41]. Consistently, we found that overexpression of CFL1 resulted in attenuation in virulence, indicating that Cfl1-mediated cell adhesion negatively modulates virulence (Figure 9C). Consistently, organ fungal burdens were maintained at low level in the PGPD1-CFL1 and PGPD1-ZNF2 infected animals at DPI 7, whereas the wildtype H99 strain proliferated significantly (Figure 9D). Unlike the PGPD1-ZNF2 strain, the PGPD1-CFL1 strain was not completely avirulent and the PGPD1-CFL1 strain proliferated significantly when examined at DPI 12 (Figure S11). This is surprising but not unexpected as the impact of ZNF2 overexpression is likely the combinational effect of additional adhesion proteins and morphogenesis factors. As noted for znf2 mutations, deletion or overexpression of CFL1 did not cause any apparent change in growth compared to wildtype when cultured at 37°C with 5% CO2 (Figure S12A and B). Cells aggregated when CFL1 was overexpressed at both 30°C and 37°C as expected.
C. neoformans is the major fungal pathogen from the phylum Basidiomycota in the Kingdom Fungi. Its morphological differentiation is typically heterogeneous and stochastic, and has been historically associated with mating. Pheromone signaling is the master regulation system in fungal mating, and it is required for early mating events such as cell recognition, mating projection formation, and initiation of cell contact and cell fusion [47], [48], [49]. However, increasing evidence implies that filamentation in Cryptococcus is a plastic process that is not limited to mating or the production of recombinant progeny: Filamentation is occasionally observed under mating-suppressing conditions, even in some attenuated strains isolated from infected host tissues [50], [51], [52], [53]; Filamentation can occur in the absence of some key components of pheromone signaling or meiosis machinery [30], [32], [33], [54]. Thus, filamentation could be used in behaviors unrelated with mating, such as foraging nutrients or defending predation. Such sex-independent cellular differentiation likely involves signaling pathways in response to cues other than the mating signal.
Here we show that sex-independent morphogenesis is linked with virulence in this fungus. We further demonstrate that the transcription factor Znf2 plays a pivotal role in cryptococcal morphological transition, and it is necessary and sufficient to drive filamentation irrespective of environmental cues, mating types, or pheromone signaling. Znf2 not only controls morphogenesis in vivo, but also the ability of this fungus to cause diseases. Thus Znf2 provides the key link between morphogenesis and virulence in Cryptococcus.
The exact mechanism by which Znf2 controls morphogenesis and links Cryptococcus pathogenicity is of great interest. Previous and this current in vitro studies indicate that Znf2 does not affect typical Cryptococcus virulence traits (e.g. melanization, capsule production, growth at high temperatures, growth in minimal media, and resistance to salt or H2O2 [42], [55]). Although the PGPD1-ZNF2 strain is avirulent, this strain was capable of propagation during the first two weeks of infection (Figure S7). This is in contrast with other avirulent strains such as cna1 or capsule mutants, which are less fit under various stress conditions and are rapidly cleared by the host [56], [57]. These lines of evidence point to new traits regulated by Znf2 that influence pathogenicity.
Our observation that genes encoding secretory proteins are enriched within the regulon of Znf2 emphasizes the importance of changes in cell surface during morphogenesis. Given that Cryptococcus strains with increased flocculation have been noted to be reduced in virulence [40], [41], Znf2 likely impacts pathogenicity at least partly through its effects on cell adhesion (flocculation). Cell adhesion mediated by microbial pathogens usually involves a repertoire of extracellular adhesion proteins. One of Znf2's downstream factors, Cfl1, is a prominent adhesion protein which orchestrates filamentation, cell adhesion, and virulence. To our knowledge, Cfl1 is the first Cryptococcus adhesin discovered. Interestingly, Cfl1 does not resemble any known adhesins characterized in ascomycetous fungi in terms of primary sequences and functional domains based on Pfam prediction (http://pfam.sanger.ac.uk/). There are four other homologues of CFL1 in the genome of Cryptococcus and in some other species in the phylum of Basidiomycota (Figure S13), in which no adhesin has been identified so far. This suggests that Cfl1 and its homologues represent a novel adhesion family specific to Basidiomycota.
Unlike Znf2, overexpression of CFL1 attenuates but does not abolish Cryptococcus virulence in the murine model of cryptococcosis. This is not unexpected as studies show that microbes are typically endowed with multiple adhesins. The master regulator Znf2 likely controls additional adhesins and other morphogenesis factors, and it is the orchestrated effects of its downstream targets that give rise to its overall impact on morphogenesis and virulence. Further characterization of Cfl1, other adhesins, and morphogens downstream of Znf2 can help parse out the effects of cell morphotype and other cell properties (e.g. changes in cell surface proteins like adhesins) on Cryptococcus virulence. Such investigation may lay a foundation for future endeavors to develop vaccines or alternative therapies against cryptococcosis.
This study was performed according to the guidelines of NIH and Texas A&M University Institutional Animal Care and Use Committee (IACUC). The animal models and procedures used have been approved by the Institutional Animal Care and Use Committee (IACUC) at Texas A&M University (protocol number: 2011-22).
Strains used in this study are listed in Table S1. For mating assays, parental strains (a and α) with equal number of cells were cocultured together on V8 medium in the dark at 22°C, and mating was examined microscopically for formation of mating hyphae and spores [58]. For cell fusion assays, the coculture of marked parental strains were removed after 48 hours of incubation on V8 medium, washed, and plated on selective media to select fusion products at 37°C as described previously [28], [33], [59]. For self-filamentation assays, cells were patched on V8 medium alone and hypha formation was examined microscopically. Phenotypical assays in vitro were performed as previously described [59]. The serotype A strain H99 is highly virulent and has been widely used in pathogenesis studies. Thus strains generated in this genetic background were used in the animal experiments and many of the in vitro characterization experiments. However, because wildtype H99 has not been observed to undergo unisexual mating and its bisexual mating is rather weak compared to the well-characterized but less virulent serotype D strains such as JEC21 and XL280, strains generated in these genetic backgrounds were used in some of the morphogenesis and mating assays.
Plasmids and primers used in this study are listed in Table S2 and S3. For gene deletion, overlap PCR products with an appropriate selection marker connected with the 5′ and 3′ flanking regions of gene of interests were introduced into Cryptococcus strains by biolistic transformation and transformants with homologous replacement were selected as described previously [60]. For overexpression, genes were amplified by PCR and the amplified fragments were digested and inserted into pXL1 after the GPD1 promoter [61]. The PGPD1 of the resulting plasmids was replaced with either the PCTR4-2 or the PGAL10 to generate the copper or the galactose inducible system. The PCTR4-2 and the PGAL10 were amplified from the plasmid pNAT/CTR4-2 and H99 genomic DNA respectively [35], [62].
Because Cfl1 contains a predicted secretory signal peptide at its N-terminus, the mCherry [63] was fused to the C-terminus. The fragment including CFL1 coding region and 1 kb upstream sequences (NCfl1) was pieced together with the mCherry by an overlap PCR. The resulting products were introduced into plasmid pXL1 to generate pXL1-NCfl1-mCherryA (for the serotype A H99 allele) and pXL1-NCfl1-mCherryD (for the serotype D JEC21 allele). The CFL1-mCherry without the CFL1 promoter was amplified and introduced into pXL1 to produce the plasmid pXL1-Cfl1-mCherry. The PGPD1 in pXL1-Cfl1-mCherry was replaced with the PCTR4-2 to generate plasmid pXC-Cfl1-mCherry. To construct overexpression of the fluorescent Cfl1 that lacks the N-terminal signal peptide [Cfl1(sigPΔ)-mCherry], primers primers Linlab948 and Linlab864 were used to generate CFL1(sigPΔ)-mCherry allele and pXC-Cfl1-mCherry was used as the template. The resulting PCR product was introduced into pXC to produce pXC- Cfl1(sigPΔ)-mCherry. Plasmids were linearized before introduced into relevant Cryptococcus strains. To examine the sub-cellular localization of Cfl1::mCherry, strains were grown on V8 agar medium at 22°C for 72 hrs before examined with a BX50 (Olympus) microscope.
Total RNA was purified using the purelink RNA purification kit (Invitrogen) and was used as the template for the first strand cDNA synthesis using the Superscript III cDNA synthesis kit (Invitrogen). Relative expression level of selected genes was measured by real time PCR using power SYBR qPCR premix reagents (Invitrogen) in a Realplex system (Eppendorf). Primer efficiency was determined by serially diluting the cDNA and monitoring DNA amplification by real-time PCR. Primers for qPCR used in this study are listed in Table S3. Gene-expression levels were normalized using the endogenous control gene TEF1. The relative transcript levels were determined using the comparative CT method as described previously [64].
RNA was separated on agarose gels blotted to nylon membrane. Redi-Prime II kit (Amersham) was used to generate probes. The C. neoformans actin gene transcript served as a control. mRNA purification was performed using the PolyATtract mRNA Isolation System III (Fisher) according to the manufacture's instruction.
The cells were cultured in 96-well microtiter plates under a variety of growth conditions. The air-liquid interface biofilm was only observed in CFL1 overexpression strains. The strains were grown in YPD liquid medium for 8 days. Crystal violet method was used for the quantitative assessment of the ability of Cryptococcus strains to form biofilm as previously described [65].
Animals were infected essentially as previously described [59], [66]. Groups of 6- to 8-week-old female A/J mice (Jackson Labs) were infected intranasally with 1×105 Cryptococcus cells in 50 µl PBS. For the PGPD1-ZNF2 strain, the culture of cells with mixed morphotype was centrifuged briefly at a low speed to allow the enrichment of yeast cells on the top. The top culture was then centrifuged again and only yeast cells were collected for infection. Ten mice per group were used for survival studies, and four or five were used for organ fungal burden studies and histological examinations. For organ fungal burden studies, fungal CFUs from lungs, kidneys, spleen, and the brains of sacrificed mice at each time point were measured as described previously [59], [67]. Dunnett's two-tailed t test was used to test statistical differences (P≤0.05). For histological examinations, organs from the sacrificed animals were fixed in 10% formalin, embedded in paraffin, sectioned at 5 µm in thickness, and stained with hematoxylin and eosin (H&E) and Gomori methenamine silver (GMS) as previously described [56], [68]. For mortality studies, the infected animals were monitored until all mice were sacrificed due to sickness or up to DPI 60 when the experiment was terminated. If the experiment was terminated, surviving animals were examined for the presence of Cryptococcus cells. Statistical significance (P≤0.05) of the survival data between different groups was assessed by the Mantel-Cox log-rank test [69].
C. neoformans var. grubii (H99): ZNF2 (CNAG_03366); MAT2 (CNAG_06203); STE3α (CNAG_06808); STE6 (CNAG_03600); MF1α (CNAG_07407), CFL1 (CNAG_00795) and other secretory protein encoding genes controlled by Znf2 (CNAG_00596, CNAG_00925, CNAG_01211, CNAG_05778, CNAG_07422, CNAG_06239, CNAG_06411, CNAG_05729), KEL1 (CNAG_01149), CDC10 (CNAG_01373), CDC12 (CNAG_01740), cnCDC11 (CNAG_02196), cnMUC1 (CNAG_03234), cnCDC24 (CNAG_04243), cnCDC3 (CNAG_05925).
C. neoformans var. neoformans (JEC21): ZNF2 (CNG02160), MAT2 (CNM02020), STE12α (CND05810), CFL1 (CNA07720). (Gene ID numbers were obtained from either NCBI Entrez or the Cryptococcus genome website at the Broad Institute http://www.broadinstitute.org/annotation/genome/cryptococcus_neoformans/MultiHome.html)
|
10.1371/journal.pbio.0060081 | Electrical Neuroimaging Reveals Timing of Attentional Control Activity in Human Brain | Voluntarily shifting attention to a location of the visual field improves the perception of events that occur there. Regions of frontal cortex are thought to provide the top-down control signal that initiates a shift of attention, but because of the temporal limitations of functional brain imaging, the timing and sequence of attentional-control operations remain unknown. We used a new analytical technique (beamformer spatial filtering) to reconstruct the anatomical sources of low-frequency brain waves in humans associated with attentional control across time. Following a signal to shift attention, control activity was seen in parietal cortex 100–200 ms before activity was seen in frontal cortex. Parietal cortex was then reactivated prior to anticipatory biasing of activity in occipital cortex. The magnitudes of early parietal activations were strongly predictive of the degree of attentional improvement in perceptual performance. These results show that parietal cortex, not frontal cortex, provides the initial signals to shift attention and indicate that top-down attentional control is not purely top down.
| To extract important details about objects in the environment, people must focus their attention on a specific location in space at any given moment. Research using functional magnetic resonance imaging (fMRI) has suggested that regions of the frontal and parietal lobes work together to control our ability to direct attention to a specific location in space in preparation for an expected visual object. However, the sluggishness of the hemodynamic response has made it difficult to obtain information from fMRI about the timing of activity. Electroencephalography (EEG) has provided information about the timing of neural activity, but the limitations of traditional source estimation techniques have made it difficult to obtain information about the precise location in the brain that the EEG signals are coming from. Thus, the sequence of activities within this frontal-parietal network remains unclear. We used a recently developed electrical neuroimaging technique—called beamforming—to localize the neural generators of low-frequency electroencephalographic (EEG) signals, which enabled us to determine both the location and temporal sequence of activations in the brain during shifts of visuospatial attention. Our results indicate that low-frequency signals in parietal cortex provide the initial signal to shift attention.
| Shifting attention to the expected location of an impending visual stimulus will improve the perception of that stimulus once it occurs there [1]. This perceptual improvement is considered to be a consequence of attentional-control operations that are performed by frontal and parietal regions of the human brain [2,3]. According to the widely accepted top-down model of voluntary attentional control, neural activities in frontal and parietal regions control the deployment of attention in space and eventually modulate the excitability of neurons in sensory-specific areas, which are responsible for processing of the upcoming stimulus. Traditionally, it has been assumed that the frontal lobes initiate top-down attentional control, because regions in frontal cortex are involved in the executive control of other cognitive and motor operations [3]. This assumption has been built into neural models of attentional control, in which one-way pathways from frontal cortex to parietal cortex to low-level visual areas subserve the voluntary control of spatial attention (Figure 1A) [2]. However, there is still much debate about the precise sequence of activity in the fronto-parietal network. Some evidence has suggested that frontal cortex becomes active before parietal cortex [4], while other evidence has suggested the opposite sequence [5,6]. This issue needs to be resolved in order to pin down the attentional control operations performed by the various regions in the network. For example, the latter sequence would suggest that parietal lobe is involved in the initiation of attentional control rather than the deployment or maintenance of attention in space, and thus necessitate a revision of current models of attentional control.
A number of functional magnetic resonance imaging (fMRI) studies have confirmed the involvement of frontal and parietal lobes in the control of visual spatial attention [7–15], but the changes in blood flow that give rise to the fMRI signal are too sluggish to investigate the time courses of activities within these brain areas (however, attempts have been made to identify temporal order of activities using analytical techniques; see [16–18]). Advances in event-related fMRI have enabled researchers to separate attentional-control activity from subsequent attention effects on the neural responses to visual stimuli [11]. However, the hemodynamic response lasts for 10–20 s, whereas the neuro-cognitive operations involved in the control and deployment of attention in space each take only a fraction of a second [19]. Thus, the sequence of neural activations within the frontal-parietal network for attentional control cannot be elucidated with hemodynamic neuroimaging methods. By comparison, the scalp-recorded electroencephalogram (EEG) and event-related potentials (ERPs) triggered by sensory or cognitive processes reveal precisely the timing of brain activity associated with specific mental operations but traditionally have failed to provide precise information about the locations of active neurons.
In both ERP and fMRI studies, the neural correlates of attentional control are often investigated by examining the neural activity elicited by a symbolic cue (e.g., an arrow) that indicates which location to attend to in preparation for an upcoming target [20]. Typically, the neural responses between leftward-directing and rightward-directing cues are compared to one another to identify brain regions that are spatially selective for shifts of attention to particular locations [5,6,21–26]. Although this type of comparison has been useful for examining pre-target biasing in sensory areas, it has two important limitations with regards to identifying attentional control activity. First, not all of the spatially specific activities observed in the cue-target interval are related to attentional control. Some of these activities have been linked to low-level sensory responses elicited by the cue [14,24], motor preparation [21], saccadic suppression [23], and other nonattentional processes. Second, this method cannot detect any activity that is associated with shifts of attention to both left and right locations, because such spatially nonspecific activity is subtracted away. If, for example, activity in the right parietal lobe controls shifts of attention to both left and right visual fields [27], then that activity would go undetected.
To better isolate activity related to the control of attention shifts, researchers have begun to compare activity associated with the presentation of attend cues to activity associated with the presentation of neutral cues that either provide no information about the location of the impending target (i.e., noninformative cues) [6] or signify that the target will not occur (i.e., interpret cues) [4,14,28]. This method controls for the presentation of the sensory cue stimuli and also permits the detection of both spatially specific and spatially nonspecific neural responses. ERP and fMRI studies using this isolation method have provided converging evidence for bilateral activity in frontal and parietal regions of cortex [4,6,14]; but unfortunately, the sequence of attentional control activities in these regions has remained unclear.
One recent study that isolated attentional control with an interpret cue reported findings consistent with the top-down model of attentional control illustrated in Figure 1A using an fMRI-constrained dipole-modeling approach [4]. Neural sources of the grand-averaged attend-minus-interpret ERP difference waveforms were modeled with four dipoles placed at the coordinates of the bilateral frontal and parietal activations observed in a similar fMRI task [14]. The orientations of the dipoles were varied until the dipole model accounted for as much of the scalp-recorded ERP data as possible in the 400–1,900-ms time interval. The resulting fMRI-constrained model suggested that the left parietal source was active 200–300 ms after cue onset. Subsequent bilateral frontal source activity began 400 ms after cue onset and was sustained until target onset. Sustained bilateral activity was also seen in the parietal source waveforms beginning at 650 ms. Follow-up analyses suggested that the early left parietal source activity was not statistically significant; thus it was concluded that frontal cortex initiated attentional control about 400 ms after cue onset. However, the early parietal activity may have been obscured in three ways. First, the ERPs elicited by leftward and rightward directing cues were averaged together, thereby minimizing any spatially specific effects that might have occurred early in parietal cortex. Second, the analyses were not ideally designed to pick up small, transient ERP effects that may have occurred early in the cue-target interval. For example, differences between attend-cue ERPs and interpret-cue ERPs were analyzed statistically by measuring mean ERP amplitudes within consecutive 100-ms intervals that were not centered on any peaks in the attend-interpret difference waveforms. Moreover, the fMRI-constrained dipoles were not fit to the difference waveforms in the early (0–300 ms) portion of the cue-target interval. Third, the fMRI-constrained dipoles may have been at suboptimal locations to pick up any early activity in the parietal lobes.
Another study that isolated attentional control with a spatially noninformative cue reported findings that were inconsistent with the top-down model of attentional control illustrated in Figure 1A [6]. Bilateral activity was observed over frontal and parietal scalp sites, primarily at electrodes on the same side (ipsilateral) as the to-be-attended location, in the 300–450 ms time interval. This fronto-parietal activation was preceded by activity over the right parietal scalp at 250 ms, which suggests that right parietal cortex might initiate the sequence of attentional control. However, dipole source modeling of the isolated attentional control activity revealed sources in temporal, rather than parietal, cortices and was rejected as being physiologically implausible. Consequently, the neural sources of the early ERP activity seen over the parietal scalp remain unknown. In addition, some of the activities seen in the attend-neutral ERP difference waveforms may have reflected differences in overall arousal or motivation, because the attend cues and neutral cues were presented in separate tasks.
Given the results of the two studies that isolated attentional control with neutral cues, it is possible that parietal, rather than frontal, cortex initiates attentional control in the spatial cueing paradigm. To date, however, the methodological and analytical procedures used to investigate the sequence of attentional control in the fronto-parietal network have been insufficient to verify this hypothesis. Here we capitalized on recent advances in EEG source reconstruction to clarify the timing and sequence of activity related to attentional control. We examined event-related changes in EEGs recorded from 11 participants during an attention-cueing task [20], in which a cue presented at fixation indicated the likely location of an impending target (Figure 2A). This task enabled us to separate the neural activities associated with the cue-induced orienting of attention from the subsequent effects of attention on target processing. In addition, we included a subset of trials on which the cue provided no information about the location of the upcoming target. By comparing activity elicited by these noninformative (no-shift) cues with the activity elicited by the informative (shift) cues, we were able to isolate further the neural activities associated with attentional control from those associated with the sensory processing of the cue itself.
We reconstructed the neural sources of EEG attentional control activity using a beamformer spatial filtering method [29,30]. The beamformer approach has several advantages over the dipole modeling approach. First, the beamformer method does not require a priori determination of the number of neural sources that may be giving rise to the scalp-recorded electrical fields. Second, the beamformer method outputs a volumetric image of neural activity throughout the brain, thereby facilitating the comparison of our results with those obtained from previous fMRI studies. Third, the beamformer method can be used to reconstruct neural sources of EEG in specific frequency bands. This enabled us to focus on oscillatory activity that we hypothesized would be important for visualizing attentional control activity across the entire cortex.
Prior studies have linked alpha band (8–14 Hz) and gamma band (>30 Hz) oscillations to attention and perception [31], but scalp-recorded oscillations in these frequency bands are primarily associated with the consequences of attention on activity in visual sensory areas [32–34] rather than the preceding attentional control operations in frontal and parietal cortices. To specifically examine attentional control activity, we opted to focus our beamformer analysis on the low-frequency theta band (4–7 Hz) oscillations. Although there is little or no existing evidence linking theta band activity to attention, we hypothesized that focusing on theta band oscillations would enable us to visualize attentional control activity across the cortex, because theta band oscillations have the following properties: (1) they reflect long-range communications between distant brain areas [35]; (2) they are carrier frequencies for high-frequency oscillations that reflect communications between nearby neurons (e.g., within a region) [36]; and (3) they have been previously linked to the working memory system [37], which is known to overlap with the spatial attention system [38]. To maximize our ability to home in on the attentional control areas that were identified in previous fMRI studies, we included both the evoked (phase-locked) and induced (non–phase-locked) activities in the analysis, because both would contribute to the hemodynamic response measured with fMRI.
We imaged neural sources of theta activity in each of 18 consecutive 50-ms intervals between cue and target. The reconstructed EEG source activities were then subjected to nonparametric statistical analyses [39] to determine which brain areas showed significant increases in activity associated with shifting attention. Based on previous electrophysiological studies, we made two predictions about the sequence of theta band activity during the voluntary control of visual attention. If voluntary attentional control is initiated in a completely top-down manner [4], activity would be seen first in frontal cortex, then in parietal cortex. Alternatively, if attentional control is initiated in parietal regions [5,6], activity should be seen first in parietal cortex and then in frontal cortex. Our results supported this latter hypothesis. Following a signal to shift attention, control activity was seen in parietal cortex 100–200 ms prior to activity in frontal cortex. Parietal cortex was then reactivated prior to anticipatory biasing of activity in occipital cortex.
Participants were most accurate to respond to targets that were validly cued (79%) and least accurate to respond to targets that were invalidly cued (69%), with intermediate accuracy for noninformatively cued targets (75%), F = 67.3, p < 0.0001. The location of the target neither influenced accuracy, F = 2.43, p = 0.12, nor interacted with cue validity, F = 1.14, p = 0.35. Follow-up comparisons revealed that accuracy for validly cued targets was significantly higher than for invalidly cued targets, t = 10.29, p < 0.00001, and for noninformatively cued targets, t = 6.86, p < 0.0001. Accuracy for noninformatively cued targets was also significantly higher than that for invalidly cued targets, t = 5.95, p = 0.0001. These behavioral results indicate that participants shifted their attention to the location indicated by the cue on shift trials and that target discrimination was improved when the cue accurately predicted the location of the upcoming target. A similar pattern of effects was observed for response times, with the shortest response times to validly cued targets, intermediate response times to noninformatively cued targets, and the longest response times to invalidly cued targets (674 ms, 714 ms, and 755 ms, respectively, F = 12.07, p = 0.002).
Figure 2B displays surface-rendered maps of significant theta band activity for shift-up cues (relative to noninformative cues) in six representative time intervals. Activity associated with attentional control was observed in posterior brain areas during the first 300 ms following the appearance of the attention-directing cue. Initially, the activity was confined primarily to extrastriate regions of the occipital lobe, but by 200 ms, both the superior and inferior parietal lobes became active, and by 300 ms, the frontal lobes became active. Between 400 and 600 ms, the activity was confined to the inferior, middle, and superior frontal gyri. Following the activity in the frontal lobes, posterior parietal cortex became active for a second time (600–700 ms post-cue). During this second activation, activity was seen in the inferior, but not the superior, parietal lobule. This parietal activity was then followed by a second phase of activity in extrastriate visual cortex that extended along the middle and inferior occipital gyri into the inferior temporal lobes.
To better characterize the spatio-temporal sequences of neural activities involved in attentional control, we plotted the normalized power changes in theta band activity for the shift-up cue relative to the noninformative cue across the entire cue-target interval in occipital, parietal, and frontal regions of interest (ROIs) (Figure 3A). Activity in the inferior occipital gyrus (IOG) occurred in two phases, with an early peak at approximately 150–200 ms after the cue and a late phase that began approximately 600 ms after the cue and continued until the onset of the target stimulus. Activity in the inferior parietal lobule (IPL) showed a similar biphasic pattern. Notably, however, the first phase peaked later than in IOG, and the second phase peaked earlier. Activity in the superior parietal lobule (SPL) peaked early, around the same time as the initial peak activation in IPL, whereas activity in the middle frontal gyrus (MFG) peaked in the middle of the cue-target interval (300–600 ms post-cue). The sequence of peak activations across these ROIs suggests that an initial feed-forward sweep of activity sends information to executive control areas in frontal cortex, which then sends information back to lower areas.
Similar patterns of attentional control activity were observed following shift-left and shift-right cues. In the case of shift-left and shift-right cues, however, some of the attention-related activity was lateralized (i.e., spatially specific). As shown in Figure 4, initial occipital activity following these cues was observed predominantly in the hemisphere contralateral to the to-be-attended location (i.e., the right hemisphere for shift-left cues and the left hemisphere for shift-right cues). The early activity in SPL was bilateral, whereas the early activity in IPL was greater in the hemisphere ipsilateral to the to-be-attended location than in the hemisphere contralateral to the to-be-attended location. Subsequent activations in MFG and occipital cortex were also larger in the ipsilateral hemisphere, whereas the late activity in IPL was bilateral.
The early occipital and parietal activations are inconsistent with current models of top-down attentional control, according to which the signal to shift attention originates in frontal cortex [2]. Because our informative cues differed from the noninformative cue in one important respect—they contained a specific color that was known in advance to be predictive of target location—it is possible that the early activity was associated with attentive processing of the cues rather than control of attention shifts to the cued locations. To evaluate this possibility, we performed a follow-up experiment in which informative and noninformative cues did not differ on the basis of a simple feature. Letters were used to cue attention to the left, upper-middle, and right locations (L, U, and R, respectively) as well as for the non-informative cue (X). The results were almost identical to those obtained in the first experiment with the exception that no early occipital activity was observed (Figure 3B). This shows that the early occipital activity seen in the main experiment reflected attentional processing of the cue but that the early parietal activity reflected control of attentional shifts to the cued location.
To determine whether the activations in occipital, parietal, and frontal regions led to modulation of perceptual processing of the subsequent target, we examined correlations between the activation magnitudes and the attention effects on target discrimination accuracy (Figure 5). All peak activations in the ROI time-courses correlated significantly with performance (rs > 0.78), except the initial activation in occipital cortex (Table 1). The lack of significant correlation with early occipital activity bolsters the conclusion that the early occipital activity reflected attentional processing of the informative cue itself. The significant correlations only at the peaks of activity in the ROIs provide compelling evidence that the early parietal activations as well as the later frontal, parietal, and occipital activations reflect attentional control operations that enhance processing of the impending visual target. Taken together, these peak activations accounted for 93% of the variability in attention effects on target discrimination accuracy (R = 0.97; R2 = 0.93; p < 0.006). That is, the net activity within the attention-control areas identified here strongly predicts the level of attentional improvement in visual processing across participants.
The present study used a recently developed technique for localizing the neural sources of scalp-recorded EEG to investigate the time course of brain activity associated with voluntary control of visuospatial shifts of attention. Although converging lines of evidence have pointed to the involvement of the frontal and parietal lobes in attentional control, the sequence of activity within the fronto-parietal control network has remained unclear due to the poor temporal resolution of fMRI and the limitations of ERP dipole source modeling. A number of alternatives have been proposed, including an entirely top-down system wherein shifts of attention are initiated by executive control regions of the frontal cortex [4] and a system wherein shifts of attention are initiated by activity in posterior brain regions that precedes frontal lobe activity [6]. Recent findings have provided support for the top-down model proposing that the frontal lobes initiate the sequence of attentional operations involved in the voluntary control of visuospatial attention shifts (Figure 1A). Our results, however, did not support this model. Instead, attentional-control activities in the parietal lobes were found to precede activity in the frontal lobes, which demonstrates that voluntary attentional control is not initiated solely by frontal cortex.
Given that IPL was active twice and SPL was active only early on, the two regions appear to mediate different attentional-control operations. The combined early activity in parietal cortex likely reflects a signal to switch attention to a specific location that is sent to executive control structures in frontal cortex. Recent neuroimaging studies indicate that activity in SPL is associated with shifting attention in spatial [40] and nonspatial [41] visual tasks, as well as in auditory and audiovisual tasks [42,43]. On this basis, we believe that SPL supplies the initial signal to switch attention, whereas IPL supplies spatial information about the to-be-attended location. The spatially nonspecific (bilateral) activation of SPL coupled with the spatially specific (predominantly ipsilateral) activation of IPL early in the cue-target interval following shift-left and shift-right cues supports this interpretation. The late activity in IPL may reflect operations involved in the marking of the to-be-attended location [9] or the actual deployment of attention to that location [11]. The late IPL activity was not sustained until target onset; thus, it is unlikely to reflect operations involved in maintenance of attention at the cued location.
The late activity in occipital and inferior temporal cortices began after the second phase of activity in IPL and was sustained until target onset. These areas are part of a ventral visual pathway that is involved in object processing and recognition [44]. Thus, the late occipito-temporal activity likely reflects anticipatory modulation of neuronal excitability in brain areas that would be responsible for processing sensory features of the upcoming target [45,46].
Following cues to shift attention to the nonlateralized location above fixation, attentional control activities in frontal and parietal areas as well as subsequent pre-target biasing in occipital cortex (relative to the noninformative cue) were largely bilateral. In contrast, attentional control activities in occipital, inferior parietal, and frontal cortices were lateralized following cues to shift attention to the left or right side of fixation. The spatially specific nature of the lateralized attentional control activity and subsequent pre-target biasing is in line with the lateralized organization of the primary visual pathways and is consistent with the observation of lateralized activity in ERP and fMRI studies examining activity following leftward and rightward-directing cues [4–6,11,14]. Increases in theta band activity were seen predominantly in cortical regions on the same side as the cued location, which suggests that this activity may be more closely associated with the anticipatory suppression of the to-be-ignored locations than the anticipatory enhancement of the to-be-attended location. The suppression of uncued locations has previously been linked to alpha band activity in this type of spatial cueing task [34]. The current results suggest that theta band activity also plays a role in the suppression of irrelevant information in order to maximize the attentional benefits for perception.
Our main finding—that voluntary attentional control is initiated in parietal cortex—is inconsistent with data from a recent combined ERP-fMRI study that reported initial activity in frontal cortex [4]. This discrepancy may be due to differences in the methods used to model brain activity. The electrical neuroimaging approach employed here used a spatial filtering technique unconstrained by any previous results or a priori hypotheses about the number of activated brain regions or the locations of the activated regions, whereas the conventional ERP-fMRI approach models electrical activity with a few discrete (dipolar) sources constrained to be at locations of fMRI activations. In addition, the beamformer technique enabled us to reconstruct the distributed neural sources of all oscillatory activity in the theta band, rather than just the evoked activity that is observed in the ERP. By comparison, the combined ERP-fMRI method faces the potential problem that induced changes in post-synaptic neural potentials are not seen in ERP waveforms (because they are not precisely phase-locked to events) but are likely associated with changes in hemodynamic responses. Such differences between the physiological contributions to ERP and fMRI signals may lead to errors in estimating the locations of ERP sources, which would, in turn, lead to errors in estimating the timing of ERP source activities. Unfortunately, the combined ERP-fMRI approach also eliminates the opportunity to use the fMRI data to evaluate the validity of the ERP source model, because data from the two methods are integrated.
We hypothesized that event-related changes in low-frequency theta band EEG oscillations would enable us to examine the spatial and temporal characteristics of activity in the voluntary attentional control network without any bias from previous fMRI results. To facilitate comparison of the present results with the results of recent fMRI studies of voluntary attentional control, we summarized the cortical sites of theta band activity across the entire cue-target interval in one image along with loci of fMRI activations [7–15]. This image, shown in Figure 6, reveals clusters of activations in occipital, parietal, and frontal regions of cortex. Although our use of standard head models, MRIs, and electrode positions likely limit our accuracy in identifying precisely the regions where attentional control activity took place, the loci of the frontal, parietal, and occipital theta band sources dovetail nicely with the foci observed in previous fMRI studies. In light of this converging evidence, it is clear that these frontal, parietal, and occipital regions play important roles in the control of spatial attention. In addition, these results provide evidence for a link between low-frequency theta band oscillations and attentional processes that heretofore has not been explored in the literature.
Our results show a clear link between low-frequency theta oscillations and attention. Prior studies have linked event-related changes in alpha and gamma band oscillations to attention and perception [32–34], but to date, theta band activity has been most closely associated with learning and memory [35–37]. Our focus on theta band activity was motivated by the hypothesis that theta band oscillations are critical for long-range communications between distant brain regions [35]. From this view, any cognitive operation that requires communication between distant brain regions should involve changes in theta-band activity. However, these low-frequency oscillations overlap in space and time with oscillations in many other frequency bands and are even coupled with high-frequency oscillations (e.g., high gamma [36]). Thus, it is unlikely that activity in any particular frequency band—such as theta, alpha, or gamma—is fully responsible for the many different attentional control operations performed by the fronto-parietal network. Other frequency bands may show different sequences of activities (i.e., frontal activity preceding parietal), and it remains to be seen how sequences of activity in different frequency bands relate to different attentional control processes. It is possible that event-related changes in specific frequency bands relate to specific attentional control operations performed by a given brain region, but it is also possible that the dynamics of attentional control activity across the cortex are more closely linked to coupling between different frequency bands (e.g., between theta and high gamma).
The electrical neuroimaging data provided here show that attentional control operations that follow the appearance of a symbolic spatial cue involve not just top-down signaling from frontal cortex but also an initial signaling from parietal cortex to indicate the need for an attention shift (Figure 1B). Moreover, the magnitude of the early parietal activity accurately predicted behavior on the subsequent perceptual task, indicating the importance of this early activity for accurate target identification. While it is possible that the attention system may be flexible and display different sequences of parietal and frontal activations with varying task requirements, it is clear that models of top-down control that posit a one-way passage of information from frontal to parietal cortex are insufficient to explain the complexities of voluntary attentional control.
|
10.1371/journal.pntd.0003366 | Quantitative PCR in Epidemiology for Early Detection of Visceral Leishmaniasis Cases in India | Studies employing serological, DTH or conventional PCR techniques suggest a vast proportion of Leishmania infected individuals living in regions endemic for Visceral Leishmaniasis (VL) remain asymptomatic. This study was designed to assess whether quantitative PCR (qPCR) can be used for detection of asymptomatic or early Leishmania donovani infection and as a predictor of progression to symptomatic disease.
The study included 1469 healthy individuals living in endemic region (EHC) including both serology-positive and -negative subjects. TaqMan based qPCR assay was done on peripheral blood of each subject using kDNA specific primers and probes.
A large proportion of EHC 511/1469 (34.78%) showed qPCR positivity and 56 (3.81% of 1469 subjects) had more than 1 calculated parasite genome/ml of blood. However, the number of individuals with parasite load above 5 genomes/ml was only 20 (1.36% of 1469). There was poor agreement between serological testing and qPCR (k = 0.1303), and 42.89% and 31.83% EHC were qPCR positive in seropositive and seronegative groups, respectively. Ten subjects had developed to symptomatic VL after 12 month of their follow up examination, of which eight were initially positive according to qPCR and among these, five had high parasite load.
Thus, qPCR can help us to detect significant early parasitaemia, thereby assisting us in recognition of potential progressors to clinical disease. This test could facilitate early intervention, decreased morbidity and mortality, and possibly interruption of disease transmission.
| Anthroponotic VL caused by Leishmania donovani in the Indian subcontinent accounts for 70% of the world burden of VL. Among the estimated 100,000 cases of VL acquired annually in India, 90% occur in the state of Bihar. Leishmania infection can result in either symptomatic or asymptomatic infection. L. donovani infection can also manifest as post-kala azar dermal leishmaniasis, a chronic cutaneous form thought to provide the reservoir for anthroponotic transmission of VL in regions endemic for this parasite species. We hypothesized that, in areas endemic for L. donovani, asymptomatic infections might also play a crucial role in disease transmission. This study describes use of quantitative PCR (qPCR) to determine the infection status in individuals living in an endemic region of India. We hypothesized that parasite load estimation by qPCR of peripheral blood cells among healthy individuals living in the endemic region might reveal the true frequency of infections through direct evidence of parasitemia. We reasoned this test would detect both asymptomatic non-progressors as well as asymptomatic individuals who will progress to fully symptomatic VL. Serologic testing by ELISA or DAT showed poor agreement with molecular detection of parasite DNA by qPCR, suggesting the tests differentiate between infection and immune response. Amongst ten healthy individuals who progressed to VL, only six were serologically positive whereas eight were initially qPCR positive, among whom five had high parasite loads in their blood. Thus, deployment of qPCR technique to estimate the presence and level of parasitemia in healthy individuals from Leishmania endemic regions may contribute to early case detection, thereby reducing morbidity and mortality. Consistent with the goals of the VL control and elimination program, this early intervention approach could help interrupt disease transmission.
| The Leishmania spp. parasites of humans are endemic in 98 countries, and more than 350 million people are at risk of infection [1]. Leishmaniasis is a neglected tropical disease, and the most severe form visceral leishmaniasis (VL, also known as kala-azar) is fatal if untreated. VL is primarily an anthroponotic infection caused by Leishmania donovani in India, transmitted by the sand fly vector Phelobotomus argentipes [2], [3].The state of Bihar in India accounts for 90% of cases in the country [4]. A majority of infected individuals do not develop clinical illness [5], [6], [7]. According to a serology-based epidemiological survey, the prevalence of asymptomatic Leishmania donovani infection in Bihar is 110 per 1,000 persons, and the rate of progression to symptomatic VL is 17.85 per 1,000 persons [8]. The kinetics of parasite amplification during the progression from infection to disease is as yet uncharacterized. We have recently shown that a highly quantitative qPCR test of blood can track the decrease in parasite load during successful treatment of infection [9]. The current study was based on the hypothesis that the number or the kinetics of circulating parasites in asymptomatically infected individuals, as measured by qPCR, might provide the most sensitive early indicator of infected subjects apt to progress to full blown disease.
Alternate techniques to detect parasites in persons with VL include direct histological examination and/or culture of bone marrow and splenic aspirates. However these methods are not feasible for screening methods or epidemiological research due to their invasive nature. Serological methods are simple, non-invasive means of detecting specific antibodies, but it is already shown that there is a lack of correlation between serology and nucleic acid methods for parasite detection [10], [11], [12]. This could reflect the inability of serology to distinguish past from ongoing infection, and therefore might result in overestimation of the number of infected asymptomatic individuals.
A large proportion of infected individuals are reportedly asymptomatic according to both serology and PCR surveys in India and nearby endemic countries [13]. Recent epidemiological reports from Brazil, Spain, and France have shown that detectable parasite DNA is present in the blood of asymptomatic infected individuals [14], [15], [16]. qPCR based epidemiological studies in the Mediterranean region have described a threshold and reference value for asymptomatic infection [17]. A similar study from our population in Bihar suggested the equivalent of 5 L. donovani parasite genomes detected/ml of blood is the threshold for clinical symptoms of VL to occur [18]. Data from our prior work in India suggest that serologic status is not a good predictor of conversion to symptomatic VL. Indeed, only 3.48% of seropositive individuals converted to active VL, whereas the conversion rate was 2.57% among seronegative individuals from the same endemic region [19]. We therefore investigated the potential for molecular quantification of parasite genome equivalents in blood as a more sensitive measure of asymptomatic infection likely to progress to disease.
Early case detection and treatment are the most important control measures for Leishmaniasis. Thus, the inability to identify individuals with asymptomatic infection, and among these to discern the individuals that are likely to progress to disease, presents a problem for clinical management. In response to this need, the current study constitutes a comparison of qPCR, serological testing with direct agglutination test (DAT), and the rK39 ELISA as predictors of progression from asymptomatic infection to fully symptomatic VL. We performed this study in a population of individuals living in the highly endemic Muzaffarpur region of the state of Bihar, India.
The work was carried out in the Department of Medicine, Banaras Hindu University, Varanasi and at its field site Kala-Azar Medical Research Centre, Muzaffurpur, Bihar and villages of Muzaffarpur district. The study was approved by the Ethics Committee of the Institute of Medical Sciences, Banaras Hindu University, the University of Iowa and the National Institutes of Health. The IRB at Banaras Hindu University is registered with the US NIH. Written informed consent was obtained from each participating individual.
The study was carried out in villages of Muzaffarpur district, which is endemic for VL. To identify individuals who had recently seroconverted, an epidemiological sero-survey was performed for two consecutive years (2009 to 2012). Villages from which large numbers of VL cases originated were identified from hospital records at Kala Azar Medical Research Centre. The research team enrolled all consenting adults age 18 and above in these villages. In the first survey, serology was done using DAT and rK39 ELISA from figure prick blood collected on filter paper. All individuals who were seronegative on the first survey were selected for testing for seroconversion by DAT and rK39 during the second serosurvey conducted 12 months later. To extract blood leukocyte DNA, two ml of blood were collected in the citrate-containing tubes from 401 recent seroconverters (seropositive) as well as 1068 randomly selected seronegative individuals within 15 days of serologic test. Buffy coat cells were isolated, and were transported from Muzaffarpur on ice to the central laboratory in Varanasi and stored at −20°C until use. 36 nonendemic healthy person's blood were also taken for qPCR assay.
Sera were eluted from filter papers containing finger prick blood and used to perform serology by DAT and rK39 ELISA as described previously [20], [21]. Individuals who are either DAT or rK39 ELISA positive were considered seropositive.
DNA was extracted using the QIAamp DNA mini kit (Qiagen, Hilden Germany) as per the manufacturer's instructions Only those DNA samples that had an optical density (OD) 260/280 ratio of 1.8–2.0 and an OD 260/230 ratio >1.5 by spectrophotometer measurements (ND-2000 spectrophotometer; Thermo Scientific, Waltham, MA, USA) were taken for qPCR experiments. The TaqMan based qPCR assay was performed in a final volume of 10 µL containing 5 µl TaqMan master mixture (2×) [Applied Biosystems (ABI), Carlsbad, CA, USA], 4 µl of DNA template and 0.25 µl (5 µM) of forward and reverse primer and 0.375 µl of probe (Integrated DNA Technologies, Coralville, IA, USA) (Table 1.). Primer-probe sequences are listed in Table 1. Amplification was conducted in a 7500 Real-Time PCR system [Applied Biosystems (ABI), Carlsbad, CA, USA]. The standard curve method for absolute quantification of parasite numbers was used as described previously [9]. All assays included no-DNA template controls, as well DNA from a negative control unexposed healthy subject. Cutoff values to consider a test positive were Cyclic threshold (Ct) value of 39. According to the standard curve, 0.001 parasite genome equivalents in the well corresponded to a CT value of 39.
Data analysis was done by non parametric Mann-Whitney test using SPSS 16 (IBM, Somers, NY, USA) and Prism (Graph pad software).
A flow chart of the progression of results is shown in Fig. 1. Serological results were interpreted in light of our original description of the first serosurvey. Cutoff values for positive serology were chosen considering results from study of negative control unexposed Indian subjects, positive control subjects with acute or successfully treated VL, and recommendations from the serological test manufacturers [20]. Considering conversion of either the DAT or the rK39 ELISA as a seroloconversion, 401 subjects converted from seronegative to seropositive between the first and the second serosurvey, whereas the remaining 1068 individuals remained seronegative.
Quantitative PCR of DNA extracted from circulating blood cells was used to assess the proportions of individuals from the endemic region with evidence of asymptomatic parasitemia, and the correlation with serological conversion. The kDNA4 probe set and taqman assay was chosen from our previously published diagnostic criteria, because of the efficient amplification of L. donovani sequences and the lack of primer-dimers complicating quantification of low numbers of parasites [22]. Data were carefully controlled, and results of individual qPCR runs were only accepted when there was a lack of kDNA amplification in no-DNA and negative controls. A standard curve was run with each assay, using the same stock of promastigote DNA extracted from an Indian isolate, to ensure consistency between assays. Data were expressed as “genome equivalents” compared to this uniform DNA standard. Notably, more than four “genome” per ml were present in individuals with symptomatic VL according to our prior publication [9].
Among a total of 1469 healthy individuals living in the endemic villages, 511 (34.78%) were positive by qPCR for amplification of any parasite DNA (CT less than 39). 171/401 (42.8%) were from seropositive group and the remaining 340/1068 (31.6%) were seronegative (Fig. 2, Table 2). The median value of parasite genomes/ml of blood was less than one and found to be 0.11 and 0.15 in seropositive and seronegative group respectively. Ten individuals who were initially belong to both seropositive and seronegative category progressed to symptomatic VL by the time of the follow up. Six (60%) were both qPCR and serology positive, two (20%) were qPCR positive but seronegative, whereas two (20%) were both qPCR and serologically negative (Table 3). Among these qPCR positive progressor five of the progessors had parasitemia levels equal to or more than the threshold value for ocuurence of symptomatic VL due to L.donovani.
The non-randomness of qPCR results is illustrated by the fact that all noendemic healthy were negative for qPCR.
In this study Leishmania DNA was detected in large proportions of both seropositive and seronegative endemic healthy groups (Fig. 1). In contrast, to nonendemic healthy who were negative for the test. Similar findings were reported in a study of L. infantum infection, a cause of VL in the Mediterranean and in Latin America [17]. In one of our earlier study we showed that the parasite load in individuals with acute symptomatic VL due to L. donovani was at least 20, and at day 30 of treatment was >1.12 genome equivalents/ml [9]. In other study we found 5 parasite genome/ml of blood as the threshold value to differentiate asymptomatic from symptomatic [18]. Mary et al. cited a persistent level of more than 1 parasite/ml as a risk for relapse of L. infantum disease [17]. Our ability to quantify the parasite load in asymptomatic individuals led us to examine a potential threshold for progression to active infection in previously uninfected individuals.
A positive serologic test for L. donovani in individuals living in endemic regions who have no symptoms of VL could indicate prior exposure without substantial active infection, ongoing asymptomatic infection which will not lead to disease, or early infection that will progress. In this situation it would be extremely valuable to perform additional diagnostic testing that could be used as a marker of infection, and also be capable of differentiating those likely to progress from those at low risk for progression to disease. Given our results suggesting the magnitude of parasitemia is related to the risk of disease, a quantitative test such as the qPCR reported herein represents a candidate test for this distinction.
Both our study and the reported study of L. infantum parasitemia cite very low numbers of parasite genome equivalents in the blood as indicative of infection. It is important to appreciate the distinction between the calculated number of parasite genomes is not equivalent to the actual number of parasites in a ml of drawn blood. We previously reported that the number of kDNA copies varies between amastigotes and promastigotes, and that copy number is highly variable between strains of the same parasite species [22]. Given this variability as well as the fact that there is an anticipated loss of DNA in the extraction process itself, one can assume that the numbers of genomes quantified on a standard curve will be relatively quantitative compared to comparison samples treated in the same manner. However one cannot draw conclusions about the absolute numbers of parasites present in the subject based on these relative numbers. It is nonetheless important to use a standard DNA so as to obtain as equivalent quantitative measures between assays as possible.
A study of asymptomatic L. donovani infection in Nepal cited poor agreement between serological and molecular tests, i.e. DAT and routine PCR [5]. Our study similarly showed a lack of agreement between serology and qPCR (k = 0.1303). Herein 42.8% or 31.6% of healthy subjects from the endemic neighborhood who were seropositive or seronegative, respectively, contained Leishmania specific DNA in their blood (Fig. 2.). Potential reasons that seropositive individuals might become qPCR negative could include degradation and clearance of Leishmania DNA after infection, corresponding with development of protective immunity. A positive qPCR test in seronegative individuals could occur if the individual was bitten by a Leishmania infected sand fly, but either immunity has not yet developed or antibody levels are too low to be detectable by the methods employed. Analogous to infection with hepatitis B, it is possible that parasite DNA, detected by PCR of peripheral blood, could be the first marker of the infection prior to antibody seroconversion. Consistent with this hypothesis, during canine VL, kDNA-PCR is significantly more sensitive than the other parasitological and serological methods, allowing the identification of infected dogs even before the appearance of antibodies [23].
Quantification on the standard curve revealed that among qPCR positives, 56 subjects (10.95% of total qPCR positive) had more than one parasite genome/ml of blood, and among them 20 (3.91%) had five or more parasites (Table 2.). Although progression to disease occurred both in seropositive and seronegative groups, 8/10 (80%) of those converting to clinical VL were qPCR positive and 5/10 (50%) had relatively high parasite loads. This suggests that asymptomatic individuals who have high parasite load may be more likely to progress to disease than individuals whose parasite loads are low (Table 3.). Other reports of asymptomatic infection suggest that parasite DNA does not often persist for more than one year, but that rarely detectable asymptomatic infection may last for decades [24]. Further their follow up is necessary to know their conversion into symptomatic cases or they remain asymptomatic.
Our recent serological study from same population area shows there is an increased risk of progressing to disease among individuals with high titers of DAT or rk39 serology [21]. Although our study suggested that DAT/ELISA titers are less sensitive and specific than qPCR with high parasite load for detection of progressors, neither approach was perfect. It may be that a combination of qPCR to detect the presence and quantity of parasite nucleic acid, coupled with serology to identify individuals with very high titers, may be a practical and sensitive means of detecting infection, for use in early case detection. The qPCR measure serves as well as an effective tool to monitor clinical management. Early case detection and treatment are the most important control measures for leishmaniasis. In anthroponotic leishmaniasis in which humans are the only reservoir, early detection by qPCR should also be explored as a means of identifying individuals who might also pose a reservoir for disease transmission.
Limitations of qPCR include high initial investment, relatively higher cost per test compared to serology. Requirement of skilled personnel can be another limiting factor, however, if completely equipped and manned central laboratories are established at strategic locations to cater to one or several districts, a reliable diagnosis can be provided to population living in endemic regions for VL which will give more possibility of identification of symptomatic condition of VL disease in infected persons.
|
10.1371/journal.ppat.1000865 | Novel Riboswitch Ligand Analogs as Selective Inhibitors of Guanine-Related Metabolic Pathways | Riboswitches are regulatory elements modulating gene expression in response to specific metabolite binding. It has been recently reported that riboswitch agonists may exhibit antimicrobial properties by binding to the riboswitch domain. Guanine riboswitches are involved in the regulation of transport and biosynthesis of purine metabolites, which are critical for the nucleotides cellular pool. Upon guanine binding, these riboswitches stabilize a 5′-untranslated mRNA structure that causes transcription attenuation of the downstream open reading frame. In principle, any agonistic compound targeting a guanine riboswitch could cause gene repression even when the cell is starved for guanine. Antibiotics binding to riboswitches provide novel antimicrobial compounds that can be rationally designed from riboswitch crystal structures. Using this, we have identified a pyrimidine compound (PC1) binding guanine riboswitches that shows bactericidal activity against a subgroup of bacterial species including well-known nosocomial pathogens. This selective bacterial killing is only achieved when guaA, a gene coding for a GMP synthetase, is under the control of the riboswitch. Among the bacterial strains tested, several clinical strains exhibiting multiple drug resistance were inhibited suggesting that PC1 targets a different metabolic pathway. As a proof of principle, we have used a mouse model to show a direct correlation between the administration of PC1 and the reduction of Staphylococcus aureus infection in mammary glands. This work establishes the possibility of using existing structural knowledge to design novel guanine riboswitch-targeting antibiotics as powerful and selective antimicrobial compounds. Particularly, the finding of this new guanine riboswitch target is crucial as community-acquired bacterial infections have recently started to emerge.
| During the last 30 years, bacterial resistance to antibiotics has become a major problem. This situation is partly because today's antibiotics are mainly based on a limited selection of chemical scaffolds, which makes it easier for bacterial pathogens to quickly develop resistance against new drug derivatives. This recurrent problem of multiple drug resistance implies a constant need to search for novel microbial targets and to modulate their activity using artificial molecules. Riboswitches are newly discovered gene regulatory elements that represent attractive targets for antimicrobial drugs. Riboswitches are RNA structures located in untranslated regions of messenger RNAs that regulate the expression of genes involved in the transport and metabolism of small metabolites. We have identified a new antibiotic specifically targeting riboswitches found in a subgroup of bacteria including Staphylococcus aureus and Clostridium difficile, which are nosocomial pathogens responsible for a significant mortality rate in hospitals, and increased health care costs. The riboswitch controls the expression of guaA that appears essential for virulence in the mammalian host. A murine model was used as a proof of principle to show that such an antibiotic could inhibit the growth of S. aureus in a mammal. Our work provides new insights into the discovery and design of novel antimicrobial agents against bacterial pathogens.
| Multiple drug resistance (MDR) has been a growing problem during the last decade, partly due to excessive use of antibiotics in human medicine and food animal production. MDR also stems from the fact that drug design has been largely based on limited chemical scaffolds leaving an opportunity for pathogens to circumvent antibiotic action mechanisms [1]. Staphylococcus aureus and Clostridium difficile are nosocomial pathogens responsible for a significant mortality rate in hospitals and increased health care costs [2]. Recently, community-acquired methicillin-resistant S. aureus (MRSA) infections have emerged and are commonly responsible for skin and soft-tissue infections that may rapidly evolve in severe and life-threatening infections [3], [4]. Moreover, some emerging clones were shown to be resistant to vancomycin, which is considered as the last chance antibiotic [5]. The pathogen C. difficile has also dramatically increased the hospital-associated deaths in recent years due to the MDR emergence and spreading of the hypervirulent and high toxin-producing strain BI/NAP1/027 [4], [5], [6]. This particular strain is spreading in North America and Europe with currently little therapeutic solutions besides the use of metronidazole and vancomycin, which are increasingly associated with relapses and poor treatment outcome [7].
Previous attempts to discover alternative antibacterial drugs targeting RNA were mainly based on a fortuitous interaction between an exogenous ligand and its RNA target [1], [8], [9], [10]. Metabolite-responsive riboswitches represent a novel solution to MDR since they could be considered as antimicrobial targets when agonistic ligands are employed as demonstrated for lysine, thiamine pyrophosphate (TPP), flavin mononucleotide (FMN) and guanine responsive riboswitches [1], [11], [12], [13], [14], [15]. In the case of lysine and TPP riboswitches, previously described ligand analogs were reported to have a multitude of cellular effects in addition to inhibition of gene expression via riboswitch binding [16], [17], [18], [19]. Pleiotropic effects were also observed for compounds targeting the guanine riboswitch and at least one analog was reported to be possibly incorporated in DNA during replication [15], [20]. Thus, while it is of interest to select antibiotics that are chemically distinct from natural ligands to avoid cellular efflux or chemical modification, it is important to consider that these chemical differences will potentially help avoid patient toxicity due to off-target binding. It is also important that the antibiotic provokes a bacteriostatic or bactericidal effect either by targeting a single gene, or a collection of genes, that is necessary for growth, or essential for bacterial survival or virulence. Thus, because modified pyrimidines can specifically bind the purine riboswitch with affinities in the low nanomolar range [21], they make excellent candidates to target purine riboswitches which are likely potent drug targets given their role in regulating purine metabolic pathways (Figure S1). For instance, the inactivation of the E. coli GMP synthetase guaA leads to guanine auxotrophy [22] whereas the inactivation of the B. subtilis IMP dehydrogenase guaB is lethal [23]. Here we show that the guanine riboswitch in S. aureus and C. difficile controls the expression of guaA and that this gene appears essential for virulence in a murine model.
Guanine-sensing riboswitches are members of the purine riboswitch class, which also comprises adenine and 2′-deoxyguanosine [24]. The guanine riboswitch negatively regulates transcription elongation at high guanine concentration in Bacillus subtilis [25] (Figure 1A). The guanine aptamer is organized around a three-way junction connecting three helices, in which a critically important nucleotide is involved in a Watson-Crick base pair interaction with the bound ligand [25] (Figure 1B). The ligand binding site contains a cavity in which the metabolite is completely surrounded by RNA contacts suggesting that most atomic positions are important for the formation of the native ligand-RNA complex [26], [27]. By using appropriate aminopyrimidines, it is also possible to recreate the correct network of hydrogen bonds required to ensure proper complex formation as previously shown for the adenine riboswitch [21]. Thus, by taking advantage of the fact that purine riboswitches efficiently bind pyrimidines, it may be possible to design novel antibiotics that bind to guanine riboswitches and therefore inhibit bacterial growth.
Pyrimidine-based molecules that could fit into the guanine riboswitch aptamer binding site were selected based on molecular modeling of crystal structures [26], [27] (Figure 2A). Using this approach, we identified two pyrimidine compounds 2,5,6-triaminopyrimidin-4-one (PC1) and 2,6-diaminopyrimidin-4-one (PC2) that satisfied defined criteria such as geometrical constraint, hydrogen bonding pattern and molecule planarity (Figure 2B). As opposed to guanine, PC1 and PC2 lack one aromatic ring, making them electronically distinct from guanine despite their similarity to guanine in terms of H-bond donating and accepting potential. Next, using the established in-line probing assay [25], [28], we monitored PC1/PC2-induced riboswitch conformational changes (Figure 2C). In absence of ligand, several cleavage products that map to previously reported single-stranded regions were observed [25], [28]. However, a cleavage reduction consistent with a reorganization of the structure upon ligand binding was observed in the core domain in presence of guanine (Figure 2C). In-line probing assay with PC1 and PC2 instead of guanine showed an identical cleavage pattern for both pyrimidine compounds and guanine, suggesting that the core is reorganized similarly in presence of these compounds, consistent with the recently reported pyrimidine-bound riboswitch crystal structure [21].
To determine whether PC1 and PC2 repress gene expression, we transformed B. subtilis with transcriptional fusions in which a guanine riboswitch was fused to a lacZ reporter gene (Figure 2D). When cells were grown in minimal medium with increasing concentrations of PC1 and PC2, beta-galactosidase activity was clearly repressed in a dose-dependent manner suggesting a modulation of the guanine riboswitch gene regulation by both molecules. We also performed growth inhibition experiments using various concentrations of both PC1 and PC2 (Figure S2). While growth inhibition was observed in minimal medium, no such inhibition was observed using a richer medium such as cation-adjusted Muller-Hinton broth (CAMHB). This selective growth inhibition can be explained by PC1/PC2 inhibiting the biosynthesis or transport of essential metabolites, which are present in CAMHB but not in minimal medium. For instance, it was recently shown that guanine-related compounds can only inhibit B. subtilis growth in a minimal medium but not in Luria broth; the growth inhibition was partly attributed to the riboswitch-mediated repression of de novo purine synthesis [15].
The S. aureus ATCC 29213 genome contains a unique guanine riboswitch located immediately upstream of the xpt gene (Figure S3). Very interestingly, RT-PCR experiments identified that the riboswitch controls a four-gene operon consisting of xpt, pbuX, guaB and guaA, thus placing guaA and guaB under the control of a riboswitch in S. aureus (Figure S3). To determine if PC1 and PC2 have antibiotic activities by targeting the guanine riboswitch in S. aureus, we performed antibiograms with PC1 and PC2 as well as with three additional molecules having similar structures (compounds 3, 4 and 5). While compounds 3 and 5 are structurally very close to PC1 and PC2, compound 4 is a guanine analog (Figure 3A). Surprisingly, among the five compounds tested, only PC1 inhibited bacterial growth in Muller-Hinton agar, which is consistent with its ability to modulate riboswitch gene expression in B. subtilis (Figure 2D). The absence of PC2 antibiotic activity is consistent with the ∼5-fold lower PC2-mediated gene expression modulation in B. subtilis, which may result from the lower number of riboswitch-ligand interactions (Figure 2B). The binding affinity of PC1 suggests that the guanine riboswitch can tolerate modifications on the ligand pyrimidine ring that are not strongly deleterious for complex formation (∼100 nM vs ∼5 nM for PC1 and guanine, respectively). The binding affinity of PC1 is very similar to that of hypoxanthine, which is a naturally occurring guanine analog [25].
To explore the antibacterial activity spectrum of PC1, we used several Gram-positive bacterial species which are potential human pathogens containing guanine riboswitches. Of the 15 species tested, 9 showed marked cellular growth inhibition, including MDR strains and the C. difficile CD6 isolate representing the hypervirulent NAP1/027 strain (Figure 3B). Interestingly, when analyzing guanine riboswitch-regulated genes, we observed that all PC1-responsive strains had guaA under riboswitch control whereas the PC1-unresponsive ones did not employ riboswitch regulation to control guaA. The best example of this correlation is that while 16S rDNA sequence analysis indicates that B. subtilis and Bacillus halodurans are very closely related species [29], B. halodurans has a guaA-controlled riboswitch and is sensitive to PC1 whereas B. subtilis lacks a guaA-controlled riboswitch and is resistant to PC1. Antibiogram results also showed that strains exhibiting pronounced MDR phenotypes are sensitive to PC1 suggesting that the antimicrobial activity does not involve action mechanisms common to other known antibiotics.
Because our data suggest that PC1 acts by repressing the GMP synthetase guaA, we reasoned that the PC1 inhibitory activity should be reduced by GMP supplementation. S. aureus cells were thus grown with or without supplemented GMP, and colony forming units (CFU) were determined following serial microdilutions (Figure 3C). As predicted, bacterial growth inhibition was relieved when GMP was provided to cells grown in presence of PC1 for 2 h or 4 h, supporting the hypothesis that bacterial growth inhibition is caused by the riboswitch-mediated guaA gene repression that results in GMP cellular depletion.
The PC1 specificity was also confirmed using the Gram-negative bacterium Escherichia coli ATCC 35695, a strain that does not contain guanine riboswitches. As expected, E. coli showed no growth inhibition in presence of PC1 even when using strains deficient for the AcrAB efflux system or having increased membrane permeability (Figure S4). These results suggest that the inability of PC1 to inhibit E. coli most probably results from guaA not being under the control of a guanine riboswitch in E. coli.
To further characterize the riboswitch inhibitory action mechanism of PC1, S. aureus cells were grown in CAMHB in presence of various ligand concentrations. We obtained a PC1 dose-dependent growth inhibition response characterized by a MIC of 0.625 mg/mL (Figure 4A). PC2 was also used and its antibiotic activity was found to be less efficient than PC1, as observed in B. subtilis (Figure S2). When compared to known antibiotics, PC1 was found to have an extremely rapid bactericidal activity similar to ciprofloxacin, one of the most bactericidal antibiotics (Figure 4B). For instance, a 4 h treatment with PC1 led to 6.67±0.58 and 5.42±1.02 log reductions in CFU/mL compared to the untreated control for cultures of S. aureus ATCC 29213 and C. difficile CD6, respectively. When the same experiment was repeated by adding either GMP or AMP to the culture for 8 hours, bacterial growth was restored by a factor of 103 only in presence of GMP (Figure 4C), suggesting that PC1 growth inhibition activity is specific to guanine metabolism.
To analyze the PC1-mediated riboswitch inhibition on S. aureus gene expression, we performed a transcriptomic microarray analysis containing a selection of S. aureus genes involved in different cellular processes such as virulence, secretion, general stress responses, sensory/regulatory systems, antibiotic resistance, iron transport and general biosynthesis [30] (Figure 4D). Among the 468 genes analyzed, 72% were repressed by at least two folds when S. aureus was treated with PC1 where the 16S rRNA gene was the most repressed (Table S1). This result is consistent with a riboswitch-mediated guaA gene expression inhibition leading to GMP cellular depletion and RNA synthesis inhibition. This is supported by the low expression of the guanine riboswitch operon (xpt, pbuX, guaB and guaA) as well as the two DNA gyrase subunits (gyrA and gyrB), which were used as housekeeping gene controls (Figure 4D). Of all the monitored genes in the microarray analysis, only ahpF and ahpC, two genes involved in stress response mechanisms, were activated by the PC1 treatment. However, when S. aureus was treated with PC1 and GMP, the microarray data showed an expression profile in which only 21% of the genes surveyed were repressed. Whereas the housekeeping gyrase genes were no longer repressed, the expression of the guanine riboswitch operon was still reduced, consistent with PC1 binding the riboswitch operon and inhibiting gene expression. The other repressed genes mainly comprised those involved in virulence and cell wall synthesis suggesting that the GMP supplemented cells were still under stress [31], which is in agreement with the partial growth rescue observed in Figure 4C. GMP is able to rescue PC1-treated cells in a dose-dependent manner (data not shown) but its low solubility prevents full recovery at higher doses. It is also probable that GMP-related feedback inhibitory mechanisms were responsible for some of the gene repression observed (as in the case of the guanylate kinase gmk). Taken together, these results are consistent with PC1 mainly acting through a riboswitch inhibition mechanism that ultimately results in GMP cellular deprivation and S. aureus growth inhibition.
Because our data showed that the growth repression activity of PC1 is influenced by the presence of GMP, we decided to assess the bactericidal activity of PC1 in a murine mastitis model of S. aureus infection, which adequately represents the clinical context. Indeed, in addition to morbid nosocomial infections caused by S. aureus, this bacterium is one of the major pathogen leading to bovine mastitis, which is the most frequent and costly disease for dairy producers with current antibiotic therapies usually failing to eliminate infections from dairy herds [32]. The antimicrobial activity of PC1 was therefore first tested on several S. aureus isolates from mastitic cows, some of which having persisting chronic infections (Figure 5A). A bactericidal effect of at least 4 orders of magnitude was observed after a 4 h treatment with PC1. Next, to ascertain that guaA was expressed in vivo and that this gene may be important during infection, we monitored the expression level of guaA by real-time PCR. When strain 1290 was grown either in broth culture in vitro or when it was directly isolated from the mastitic milk of infected cows (M. Allard and F. Malouin, in preparation), very similar expression levels were found for guaA and the essential gene gyrB in both environments. This suggests that PC1 could have an impact on guaA expression in vivo and thus be used to treat S. aureus infections.
The proof of concept for the therapeutic efficacy of PC1 was established in our murine model of S. aureus-induced mastitis [33]. At 4 h post-infection, different concentrations of PC1 were administered to infected mice that were sacrificed 6 h later (Figure 5B). When compared to mice that were not treated with PC1, viable bacterial counts in the mammary gland were drastically reduced in a dose-dependent manner. This strong therapeutic effect was highly comparable to what we observed with known antibiotics. For example, amoxicillin decreased the bacterial load in the mammary gland to a log10 median value of 3.97 CFU/g of gland at a dose of 50 µg/gland. Noteworthy, a dose of 50 µg of amoxicillin would represent 100xMIC/g of gland, whereas a similar dose would only represent a twelfth of the MIC/g of gland for PC1. This result is consistent with the idea that PC1 is most efficient in the mammary gland environment suggesting that the microaerobic condition (i.e., low oxidative environment) of the mastitic milk[34] helps PC1 therapeutic efficacy. Consistent with these results, we found that the potency of PC1 was significantly increased by preventing its oxidation using a reductive agent such as DTT in susceptibility tests in vitro (Figure S5).
Despite previous large scale screen data suggesting that guaA is not essential for S. aureus growth [23], [35] in relatively rich media, we show here that blocking guaA expression can lead to bactericidal activity in various bacterial species. In support of guaA for cell viability, it has been recently reported that mutations occurring in guaA prevent Streptococcus suis [36] and Salmonella thyphimurium [37] from properly infecting porcine and murine models, respectively, suggesting that GMP bioavailability may be reduced during host infection and that guaA is likely to be crucial for bacterial infection in mammals. Thus, together with studies showing the importance of guaA for bacterial growth in urine or blood [38], [39], our data suggest that mammalian infection sites may significantly differ in their nutrient compositions from those used in large scale screens [23], [35], and that care should be taken when assessing the “essentiality” of a gene. Furthermore, when assessing whether S. aureus could develop resistance to PC1, no resistant bacteria were obtained after more than 30 passages suggesting that maintaining a functional guaA-regulated riboswitch is a vital process (Figure S6). Taken together, the demonstration that guaA expression is normally maintained in S. aureus grown in vivo and the strong therapeutic effect resulting from PC1 treatment indicate that guaA is an important contributor to the survival of S. aureus during infection and that it can be used as an antibiotic target.
The major limitation to validate an antibiotic that targets riboswitches is to evaluate the antibiotic specificity of action. In this particular case, PC1 is not a broad-spectrum antibacterial drug given that it does not target all bacteria containing guanine riboswitches, but only those in which guaA is under the control of a riboswitch. It is not excluded that other riboswitch-controlled genes may participate in the PC1-dependent bacterial growth inhibition (e.g., guaB), or that PC1 may bind other cellular targets, which alone or in combination with the riboswitch-controlled gene repression, would repress bacterial growth. Nevertheless, the restricted nature of growth inhibition likely indicates that PC1 inhibits bacterial growth through riboswitch binding and not via an alternative mechanism such as DNA incorporation (Figure 6). For instance, when performing antibiograms using the guanine analog 6-thioguanine, a general growth inhibition was observed in E. coli and S. aureus (Figure S7), consistent with its incorporation into DNA that perturbs the epigenetic pathway of gene regulation [40]. The selective antibacterial activity of PC1 toward S. aureus was also supported by the lack of apparent toxicity for mice treated with the experimental compound at concentrations as high as 100 µg/gland with no sign of discomfort including vocalizations, curved back, piloerection and hypothermia. There was also no apparent cytotoxicity upon histological observations of mammary tissues in PC1-treated mice compared to PBS-treated glands (Figure S8).
Recently, the Breaker group used purine derivatives modified in position 2 or 4 to target guanine riboswitches and found three molecules that inhibited bacterial growth [15]. Among these molecules, only one was able to repress the expression of a guanine riboswitch-controlled reporter gene suggesting that at least two molecules inhibited cell growth through a different mechanism of action. It is possible that the mode of action of these two molecules involves nucleic acids incorporation following ribosylation at position 9 of the purine analog, as demonstrated for 6-thioguanine. It is also interesting to mention that the antibiotic activity was only observed using B. subtilis strains cultivated in minimal medium whereas no growth inhibition was detected in rich media. In our study, we selected guanine riboswitch ligands that cannot be ribosylated to prevent alternative modes of action, and favored a pyrimidine compound (PC1) that retained most of the key functional groups. By testing various bacterial species, we observed that the bactericidal activity of PC1 was only seen against bacteria using a guanine riboswitch to control guaA expression, and this even when bacteria were grown in a rich medium. This demonstrates that analog binding to riboswitch aptamers is not the only determinant to achieve selective and efficient bactericidal effects.
This study shows for the first time that antibiotics targeting riboswitches may be efficient to kill bacterial pathogens in vitro as well as in mammalian infection models. We found here that the selective bacterial killing of PC1 is only achieved when guaA, a GMP synthetase, is under the control of the riboswitch and when the antimicrobial agent cannot be ribosylated. The narrow spectrum of activity we demonstrated here for PC1 is very interesting since two of the target bacteria, S. aureus and C. difficile, are among the most problematic nosocomial pathogens. The spread of MDR in those bacterial species also stresses the importance to develop new antibiotics that avoid current mechanisms of resistance. The use of narrow spectrum drugs should be encouraged whenever possible to reduce any selective pressure for resistance in non-targeted bacteria. Here, we also showed that the development of resistance toward PC1 is likely to be infrequent.
The institutional ethics committee on animal experimentation of the Faculté des sciences of the Université de Sherbrooke (QC, Canada) approved these experiments and the guidelines of the Canadian Council on Animal Care were respected during all the procedures.
4-hydroxy-2,5,6-triaminopyrimidine (PC1), 2,4-diamino-6-hydroxypirimidine (PC2) and 9-methylguanine were purchased from Fluka. 2-amino-5-bromo-6-methyl-4-pyrimidol and 2-amino-4-hydroxy-6-methylpyrimidine were purchased from Aldrich.
Our ligand selection took into account guanine binding requirements [25] and crystal structure interactions [21], [26], [27]. Planar molecules were selected to preserve stacking interactions with adenines 21 and 52 in the guanine aptamer binding site. One of our important selection criteria was to avoid the presence of functional groups that could serve as a ribosylation site in purine or pyrimidine analogs that would allow subsequent nucleic acid incorporation and non-specific antibiotic effect. Successfully identified molecules were drawn using Chem3D Pro (CambridgeSoft) and docked onto the guanine aptamer crystal structure (PDB 1U8D). The in silico procedure was important to validate aptamer-ligand interactions and to avoid sterical obstructions that would perturb ligand binding.
For the production of guanine riboswitch aptamers, DNA templates were prepared from partial duplexes and transcribed using T7 RNA polymerase as previously described [28]. The aptamer sequences used in this study are based on the genomic sequence to which a GCG sequence is added to the 5′ side to allow high transcription yield and to minimize the 5′ heterogeneity [28].
[5′-32P] RNA molecules were incubated for 96 h at 25°C in 50 mM Tris-HCl buffer, pH 8.5, 20 mM MgCl2 and 100 mM KCl in absence or in presence of indicated ligand concentrations. The reactions were stopped with a 97% formamide solution containing 10 mM EDTA and samples were purified by electrophoresis in 10% polyacrylamide gels (acrylamide:bisacrylamide; 19:1) containing 8 M urea. Gel were dried and exposed to Phosphor Imager screens.
Regulation of the beta-galactosidase reporter gene expression in presence of PC1 or PC2 was determined using an xpt-lacZ transcriptional fusion construct integrated in the genome of B. subtilis by recombination. The beta-galactosidase activity was measured after 4 h of growth at 37°C in minimal medium in absence or presence of the indicated ligand concentrations [41].
Bacteria were inoculated at 105 CFU/mL in melted Muller-Hinton agar. After agar medium was solidified, six wells of 4 mm in diameter were made and filled with 10 µL of the tested molecules (5 mg/mL). Plates were incubated for 16 h at 37°C.
The minimal inhibitory concentration (MIC) of PC1 and PC2 against S. aureus strain ATCC 29213 was determined using a microdilution method in 96-well plates [30]. Bacteria were inoculated at 105 CFU/mL and incubated at 37°C for 24 h in cation-adjusted Muller-Hinton broth (CAMHB). Bacterial growth was detected by measuring the OD at 595 nm on a microplate reader.
Time-kill experiments were performed for the determination of the bactericidal effect of test antibiotics. Bacteria were inoculated at 105 CFU/mL in CAMHB in absence or presence of the antibiotic at its MIC with or without 100 µM GMP. Bacterial permeability to GMP was increased by adding 0.002% Triton X-100. At several time points, bacteria were sampled and serially diluted before spreading on tryptic soya agar (TSA) plates for CFU determinations. Plates were incubated for 24 h at 37°C.
Bacteria were inoculated at 108 CFU/mL in CAMHB in absence or presence of 600 µg/mL PC1 or 600 µg/mL PC1 supplemented with 100 µM GMP. After 30 min of growth, RNA was extracted and 2.5 µg of RNA were submitted to reverse transcription to generate fluorescent probes through an aminoallyl cDNA labeling procedure before being hybridized on the microarray [30].
Experimental conditions used here for the mastitis model were previously optimized for S. aureus Newbould and antibiotic treatment [42]. CD-1 lactating mice (Charles River, St. Constant, Canada) were used 12 to 14 days after offspring birth and typically weighed 35 to 40 g. Pups were removed 1 h before bacterial inoculation of mammary glands and a mixture of ketamine/xylazine at 87 and 13 mg/kg of weight, respectively, was used for anesthesia of lactating mice. A 100 µl syringe with a 33-gauge blunt needle was used to inoculate both L4 (on the left) and R4 (on the right) abdominal mammary glands. These large glands constitute the fourth pair found from head to tail. Each udder canal was exposed by a small cut at the near end of the teat under a binocular and 100 µL of bacterial suspension (1 CFU/µL) was injected through the orifice. Mice mammary glands were treated 4 h after infection with PBS or PBS with 10, 50 and 100 µg/gland of PC1 and mice were sacrified 6 h later for mammary gland sampling and homogenization. The tissues used for CFU counts were homogenized in 2 mL of PBS and the bacterial content was evaluated by serial logarithmic dilutions on agar. The detection limit was 100 CFU/g of gland.
|
10.1371/journal.pntd.0002750 | Sm10.3, a Member of the Micro-Exon Gene 4 (MEG-4) Family, Induces Erythrocyte Agglutination In Vitro and Partially Protects Vaccinated Mice against Schistosoma mansoni Infection | The parasitic flatworm Schistosoma mansoni is a blood fluke that causes schistosomiasis. Current schistosomiasis control strategies are mainly based on chemotherapy, but many researchers believe that the best long-term strategy to control disease is a combination of drug treatment and immunization with an anti-schistosome vaccine. Numerous antigens that are expressed at the interface between the parasite and the mammalian host have been assessed. Among the most promising molecules are the proteins present in the tegument and digestive tract of the parasite.
In this study, we evaluated the potential of Sm10.3, a member of the micro-exon gene 4 (MEG-4) family, for use as part of a recombinant vaccine. We confirmed by real-time PCR that Sm10.3 was expressed at all stages of the parasite life cycle. The localization of Sm10.3 on the surface and lumen of the esophageal and intestinal tract in adult worms and lung-stage schistosomula was confirmed by confocal microscopy. We also show preliminary evidence that rSm10.3 induces erythrocyte agglutination in vitro. Immunization of mice with rSm10.3 induced a mixed Th1/Th2-type response, as IFN-γ, TNF-α, and low levels of IL-5 were detected in the supernatant of cultured splenocytes. The protective effect conferred by vaccination with rSm10.3 was demonstrated by 25.5–32% reduction in the worm burden, 32.9–43.6% reduction in the number of eggs per gram of hepatic tissue, a 23.8% reduction in the number of granulomas, an 11.8% reduction in the area of the granulomas and a 39.8% reduction in granuloma fibrosis.
Our data suggest that Sm10.3 is a potential candidate for use in developing a multi-antigen vaccine to control schistosomiasis and provide the first evidence for a possible role for Sm10.3 in the blood feeding process.
| Schistosomiasis mainly occurs in developing countries and is the most important human helminth infection in terms of global mortality. This parasitic disease affects more than 200 million people worldwide and causes more than 250,000 deaths per year. Current schistosomiasis control strategies are mainly based on chemotherapy, but many researchers believe that the best long-term strategy for controlling schistosomiasis is a combination of drug treatment and immunization with an anti-schistosome vaccine. Consequently, significant effort has been dedicated to developing and characterizing an anti-schistosome vaccine. Over the last five years, considerable data have been generated regarding the genomics, transcriptomics and proteomics of Schistosoma mansoni. In the present study, we characterize the Sm10.3 protein and evaluate its potential to protect against S. mansoni infection in a murine model. We demonstrate that Sm10.3 is primarily expressed during the stages of the parasite life cycle that involve infection and disease development in the human host. Sm10.3 is located on the surface of the digestive epithelia of adult female worms, an important host/parasite interface. Moreover, the vaccination of mice with rSm10.3 confers partial protection against S. mansoni. Taken together, our data suggest that Sm10.3 may be a useful component of a multi-antigen vaccine against schistosomiasis.
| Schistosomiasis occurs primarily in developing countries and is the most important human helminth infection in terms of global mortality. This parasitic disease affects more than 200 million people worldwide, causing more than 250,000 deaths per year [1]. Furthermore, schistosomiasis is responsible for the loss of up to 4.5 million DALYs (disability adjusted life years) annually [2]. Current schistosomiasis control strategies are mainly based on chemotherapy but, despite decades of mass treatment, the number of infected people has not decreased considerably in endemic areas [3]–[5]. The extent of endemic areas, constant reinfection of individuals and poor sanitary conditions in developing countries make drug treatment alone inefficient [6]. It is thought that the best long-term strategy for controlling schistosomiasis is through immunization with an anti-schistosome vaccine combined with drug treatment [7]. A vaccine that induces even a partial reduction in worm burdens could considerably reduce pathology and limit parasite transmission [8].
Currently, the most promising schistosome vaccine candidates are proteins located on the surface of the worms [9], such as the tegument proteins TSP-2 [10] and Sm29 [11]. The tegument is a dynamic host-interactive surface involved in nutrition, immune evasion/modulation, excretion, osmoregulation, sensory reception, and signal transduction [12], [13]. Other surface-exposed proteins with high potential as vaccine targets are located in the digestive tract of lung-stage schistosomula and adult worms [14]–[18].
In this study, we evaluated the potential of Sm10.3, a member of the micro-exon gene 4 (MEG-4) family, to serve as a component of a recombinant subunit vaccine. The Sm10.3 antigen was first described in 1988 [19], but the Sm10.3 and MEG family genes were not completely characterized until the recent publication of the S. mansoni genome [20]. There are multiple copies of some MEGs in the S. mansoni genome, arranged as tandem, symmetrically organized exons with lengths that are a multiples of three bases (from 6 and 36 base pairs) [20], [15]. It is thought that this arrangement may lead to protein variation through alternative splicing. Moreover, most of the MEGs are up-regulated during the stages in the parasite life cycle that involve establishment in the mammalian host [15].
In this study, we determined that Sm10.3 is located on the surface of the digestive tract of S. mansoni adult female worms and lung-stage schistosomula. We detected higher levels of Sm10.3 mRNA in the schistosomula stage of the parasite life cycle. We also show preliminary evidence that Sm10.3 plays a role in erythrocyte agglutination. Furthermore, we report that vaccination with rSm10.3 induces a mixed Th1/Th2-type immune response in mice, which correlates with a reduction in the worm burden and liver pathology.
All animal experiments were conducted in accordance with the Brazilian Federal Law number 11.794, which regulates the scientific use of animals, and IACUC guidelines. All protocols were approved by the Committee for Ethics in Animal Experimentation (CETEA) at Universidade Federal de Minas Gerais UFMG under permit 179/2010.
Female C57BL/6 mice aged 6–8 weeks were purchased from the Federal University of Minas Gerais (UFMG) animal facility. S. mansoni (LE strain) cercariae were maintained in Biomphalaria glabrata snails at CPqRR (Centro de Pesquisa René-Rachou-Fiocruz) and prepared by exposing infected snails to light for 1 h to induce shedding. Cercarial numbers and viability were determined prior to infection using a light microscope.
The plasmid pJ414 containing the sequence for rSm10.3 (pJ414::Sm10.3) was manufactured by DNA 2.0 (https://www.dna20.com) using DNA2.0 optimization algorithms for expression in Escherichia coli. This plasmid was transformed into E. coli Rosetta-gami (Merck KGaA, Darmstadt, Germany) competent cells. Transformants harboring the designed plasmid were screened on LB agar plates containing ampicillin (50 µg/ml) and cloranphenicol (34 µg/ml) and the selected transformant was designated as rSm10.3-Rosetta. One liter of rSm10.3-Rosetta was cultured in a three-liter erlenmeyer on a rotary shaker at 200 rpm at 37°C to an optical density at 600 nm of approximately 0.5–0.8 and gene expression was induced by using 1 mM isopropylthiogalactoside (IPTG). After 5 h of induction, the bacterial cells were harvested by centrifugation at 4,000× g for 20 min. Using gently vortexing or pipetting, the pellet was resuspended in 50 ml of 10 mM Na2HPO4, 10 mM NaH2PO4, 0.5 M NaCl and 20 mM imidazole. Subsequently, the cells were submitted to three cycles of sonication lasting 30 s each and centrifuged at 5400× g for 20 min. The rSm10.3 was recovered solubilized in the supernatant and purified by affinity chromatography on a Ni-Sepharose column (Hitrap chelating 5 mL) using an AKTA explorer chromatography system (GE Healthcare, São Paulo, Brazil). After protein binding to the Ni-Sepharose column, washes with 50 mM imidazole were performed and the protein was eluted with 500 mM imidazole. Fractions containing the protein were determined through Bradford's method (Coomassie Protein Assay Kit, Pierce) and also SDS/PAGE-12% and dialyzed against PBS pH 7.0. The dialysis was carried out at 4°C using a Spectra/Por2 membrane (MWCO 6 to 8 kDa; Spectrum Medical Industries, Inc., Laguna Hills, CA). The recombinant protein was quantified using the Bradford's method and used as antigen for vaccination and immunological experiments. All reagents were purchased from Sigma-Aldrich, CO (St. Louis, MO, USA) unless otherwise specified.
Total RNA was isolated from adult parasites, eggs, miracidia, cercariae or schistosomula using published procedures [21]. Total RNA was extracted with Trizol (Invitrogen, Carlsbad, CA, USA) according to the manufacturer's instructions. RNA samples were treated with DNAse, purified and concentrated using the RNeasy Micro Kit (QIAGEN, Valencia, CA, USA) according to the manufacturer's instructions. A 1.5 µg portion of each sample was reverse transcribed using the SuperScriptH III First-Strand Synthesis SuperMix (Invitrogen, Carlsbad, CA, USA). Specific primer pairs (5′-CTT AAT CAA TAA GCC AAA GG-3′ and 5′-TAT TGA TTT GTC GTA ATA GT-3′) were designed using the Primer Express program (Applied Biosystems, Foster City, CA, USA) and default parameters and arbitrarily named primers 1 and 2, respectively. Real-time RT-PCR reactions were conducted in triplicate in a 20 µL volume containing 10 µL of Sybr Green PCR Master Mix (Applied Biosystems, Foster City, CA, USA), 160 nmol of each primer (primers 1 and 2) and 0.30 µL of the cDNA generated by reverse transcription. Real-time RT-PCR was performed using the 7300 Real-Time PCR System (Applied Biosystems, Foster City, CA, USA) and the following cycling parameters: 60°C for 10 min, 95°C for 10 min and 40 cycles of 95°C for 15 sec and 60°C for 1 min. A dissociation curve was generated using the following conditions: 95°C for 15 sec, 60°C for 1 min, 95°C for 15 sec and 60°C for 15 sec. Real-time data were normalized to the expression level of NADH dehydrogenase. p-values were determined by Student's t-test, using one-tailed distribution and heteroscedastic variance.
Purified rSm10.3 was analyzed on 15% polyacrilamide SDS-PAGE gels prepared and run as previously described [22]. The gel was then transferred to a nitrocellulose membrane [23]. The membrane was blocked with TBS-T (0.5 M NaCl−0.02 M Tris (pH 7.5), 0.05% Tween 20) containing 5% dry milk for 16 h at 4°C. The membrane was then incubated with a mouse monoclonal antibody to the 6×His-tag (GE Healthcare, Pittsburgh, PA, USA) diluted 1∶2,000 or a mouse polyclonal antibody to rSm10.3 diluted 1∶1,000 for 1 h at room temperature. After three washes with TBS-T, the membrane was incubated in 1∶2,000 goat anti-mouse IgG conjugated to alkaline phosphatase (AP), then treated with a developing buffer containing nitroblue tetrazolium (NBT) and 5-bromo-4-chloro-3-indolyl-1-phosphate (BCIP). After the membrane was imaged, it was washed with distilled water and dried by sandwiching between two sheets of filter paper for storage. All reagents were purchased from Sigma-Aldrich, CO (St. Louis, MO, USA) unless otherwise specified.
Six to eight week-old female C57BL/6 mice were divided into two groups of ten mice each. Mice were injected subcutaneously at the nape of the neck with 25 µg of rSm10.3 protein or PBS (for the control group) on days 0, 15 and 30. The protein mixture was formulated using Complete Freund's Adjuvant (CFA) for the first immunization and Incomplete Freund's Adjuvant (IFA) for the last two immunizations (Sigma-Aldrich, CO, St. Louis, MO, USA).
For the microscopy studies, adult worms were recovered from perfused mice, and lung-stage schistosomula were prepared in vitro as described by Harrop & Wilson [24]. Parasites were fixed in Omnifix II (Ancon Genetics, St Petersburg, FL, USA) for sectioning. For the sectioning assays, 7 µm slices of Paraffin-embedded adult male or female parasites were deparaffinized using xylol and hydrated with an ethanol series, [25]. For experiments using in vitro cultured lung-stage schistosomula, a whole-mount protocol was chosen, lung stage schistosomula were treated with permeabilizing solution (0.1% Triton X-100, 1% BSA and 0.1% sodium azide in PBS pH 7.2) overnight at 4°C [25]. Following, permeabilized schistosomula and parasite sections were blocked with 1% BSA (bovine serum albumin) in PBST (phosphate buffered saline, pH 7.2 with 0.05% Tween-20) for 1 h and incubated with anti-rSm10.3 serum diluted 1∶20 in blocking buffer. Serum from non-immunized mice was used as a negative control. The samples were washed three times with PBST and incubated with an anti-mouse IgG antibody conjugated to FITC (Molecular Probes, Carlsbad, CA, USA) diluted 1∶100 in blocking buffer containing rhodamine phalloidin (Molecular Probes, Carlsbad, CA, USA) to stain the actin microfilaments. The samples were washed four times and mounted with ProLong Gold anti-fading mounting medium containing DAPI (Molecular Probes, Carlsbad, CA, USA). Schistosomula were imaged on a Nikon A1R confocal microscope (60× NA1.4-CFI-Plan-Apo oil objective), and adult worms were imaged on a Nikon Eclipse Ti microscope from the Microscopy Center of the Biological Sciences Institute (CEMEL) at the Federal University of Minas Gerais (UFMG). All reagents were purchased from Sigma-Aldrich, CO (St. Louis, MO, USA) unless otherwise specified.
Mice blood for the hemagglutination assays were collected and processed as described previously [26]. Blood was withdrawn from mice with a syringe, added to tubes containing EDTA (at a final concentration of 12 mM) and centrifuged for 15 min at 3000× g at room temperature. After centrifugation, the plasma was transferred to a fresh tube. In a second tube, erythrocytes were washed three times and suspended in PBS (pH 7.2) to a hematocrit of 20 or 40%. The erythrocyte suspensions were used for the agglutination assays on glass slides, in the hemagglutination endpoint dilution assays or in the hemagglutination kinetics assays.
Agglutination assays on glass slides was performed as previously described [26], the erythrocyte suspensions were prepared as above to a final hematocrit of 20% diluted in PBS. Next, 9 µl of each erythrocyte suspension was combined with 1 µl of rSm10.3 containing 5 µg of purified protein, 1 µl of PBS or 1 µl of PBS containing 5 µg of rSm29 (an unrelated S. mansoni antigen) as negative controls. Recombinant protein rSm29 [11] was expressed with a 6×-histidine tag and purified in a similar way to rSm10.3, as described in previous sections. The 10 µl mixture was loaded onto glass slides, covered with coverslips and immediately visualized using 10× and 40× objective lenses on a microscope equipped with a JVC TK-1270/RBG micro camera.
Hemagglutination endpoint dilution assays were performed as previously described [27], [28]. A serial two-fold dilution of 50 µl of rSm10.3, rSm29 and Concanavalin A solutions (protein concentrations ranged from 500 µg/mL to 0.97 µg/mL) in microtiter U-plates was mixed with 50 µl of a 2% suspension of mice erythrocytes in PBS at 37°C, resulting in a final erythrocytes suspension volume of 100 µl and a hematocrit of 1%. In the blank wells 50 µl of a 2% suspension of mice erythrocytes were mixed with 50 µl of PBS. Concanavalin A was used as positive control of the hemagglutination process [29]–[31] and rSm29 (an unrelated S. mansoni antigen) as negative control. The results were read after approximately 1 h when the blank had fully sedimented. The endpoint was defined as the highest dilution showing complete hemagglutination. The hemagglutination titer, defined as the reciprocal of the highest dilution exhibiting hemagglutination, was defined as one hemagglutination unit. Specific activity is the number of hemagglutination units per mg of protein per milliliter [32].
Mice polyclonal antibodies raised against rSm10.3 were used as inhibitor in hemagglutination inhibition assays according to a protocol previously described [33], with modifications. Twenty five microliters of rSm10.3 solutions with three different protein concentrations (1000 µg/mL, 500 µg/mL and 250 µg/mL) were mixed with a serial two-fold dilution of 25 µl of anti- rSm10.3 mice serum, ranging from 1∶1 to 1∶32 dilution and incubated at 37°C for 30 min. Following, 50 µl of a 2% suspension of mice erythrocytes in PBS was added to the wells and incubated for 1 h at 37°C.
The kinetics of the hemagglutination process were monitored by analyzing variations in turbidity [26]. Briefly, 100 µl of the erythrocyte suspension (in PBS) at a hematocrit of 5% was added to 96-well plates, and the agglutination was triggered by adding 0.5 to 5 µg of rSm10.3 diluted in 100 µl of PBS, 100 µl of PBS containing 5 µg of Concanavalin A, as positive control of the hemagglutination process and 100 µl of PBS or 100 µl of PBS containing 5 µg of rSm29 as negative controls, resulting in a final volume of 200 µl and a hematocrit of 2.5%. Plates were incubated at 37°C and spectrophotometric readings were taken at 655 nm every 13 s, with 8 s of shaking between each reading in a VersaMax Tunable Microplate reader (Molecular Devices, Sunnyvale, CA, USA). All reagents were purchased from Sigma-Aldrich, CO (St. Louis, MO, USA).
Fifteen days after the final immunization, the mice were challenged with 100 cercariae (LE strain) by percutaneous exposure of the abdominal skin for 1 h. Forty-five days after the challenge, the adult worms were perfused from the portal veins, as described previously [34]. Two independent experiments were performed to determine protection levels. The degree of protection was calculated by comparing the number of worms recovered from each vaccinated group to the respective control group, using the following formula:where PL indicates the protection level, WRCG indicates the number of worms recovered from the control group, and WREG indicates the number of worms recovered from the experimental group.
Following immunization, sera were collected from ten mice in each experimental group at two week intervals. The levels of specific anti-Sm10.3 antibodies were measured by indirect ELISA. Maxisorp 96-well microtiter plates (Nunc, Roskilde, Denmark) were coated with 5 µg/mL rSm10.3 in carbonate-bicarbonate buffer (pH 9.6) for 16 hat 4°C, then blocked for 2 h at room temperature with 200 µL/well PBST (phosphate buffer saline, pH 7.2 with 0.05% Tween-20) plus 10% FBS (fetal bovine serum). One hundred microliters of serum diluted 1∶100 in PBST was added to each well, and the plates were then incubated for 1 h at room temperature. Plate-bound antibodies were detected using peroxidase-conjugated anti-mouse IgG, IgG1 and IgG2a (Southern Biotechnology, CA, USA) diluted 1∶10000, 1∶5000 and 1∶2000 in PBST, respectively. The color reaction was induced by adding 100 µL of 200 pmol OPD (o-phenylenediamine) in citrate buffer (pH 5.0) plus 0.04% H2O2 to each well for 10 min. The color reaction was stopped by adding 50 µL of 5% sulfuric acid to each well. The plates were read at 495 nm in an ELISA plate reader (BioRad, Hercules, CA, USA). All reagents were purchased from Sigma-Aldrich, CO (St. Louis, MO, USA) unless otherwise specified.
The cytokine experiments were performed using cultured splenocytes from individual mice immunized with rSm10.3 or PBS (n = 4 for each group). The splenocytes were isolated from the macerated spleens of individual mice one week after the third immunization and washed twice with sterile PBS. After washing, the cells counts were adjusted to 1×106 cells per well in RPMI 1640 medium (Gibco) supplemented with 10% FBS, 100 U/mL of penicillin G sodium and 100 µg/mL of streptomycin sulfate. The splenocytes were maintained in culture with medium alone or stimulated with rSm10.3 protein (15 µg/mL), concanavalin A (ConA) (5 µg/mL), or LPS (1 µg/mL), as previously described [34]. The 96-well plates (Nunc, Roskilde, Denmark) were maintained in a 37°C incubator with a 5% CO2 atmosphere. The culture supernatants were collected after 24 h to measure IL-5 levels, after 48 h to measure TNF-α levels and after 72 h to measure IFN-γ and IL-10 levels. The cytokine measurement assays were performed using the DuoSet ELISA kit (R&D Systems, Minneapolis, MN) according to the manufacturer's instructions. All reagents were purchased from Sigma-Aldrich, CO (St. Louis, MO, USA) unless otherwise specified.
Following perfusion to recover the schistosomes, liver samples were collected from 8 animals each from the control and experimental groups to evaluate the effect of immunization on granuloma formation. The liver samples, which were taken from the central part of the left lateral lobe, were fixed with 10% buffered formaldehyde in PBS. Histological sections were performed using microtome at 6 µm and stained on a slide with picrosirius-haematoxylin-eosin (PSHE). The count of granulomas was performed at a microscope with 10× objective lens. Each liver section was scanned for calculating its whole area (mm2) using the ImageJ software (http://rsbweb.nih.gov/ij/index.html). For measurement of the total area of granulomas, a microscope with 10× objective lens was used; images were obtained through a JVC TK-1270/RBG microcamera attached to the microscope. Twenty granulomas with a single-well-defined egg were randomly selected, in each liver section and the granuloma area was measured using the ImageJ software. Granuloma fibrosis was analyzed using the software analysis getIT (Olympus Soft Imaging getIT) and images to illustrate the fibrosis area were edited using Adobe Photoshop software. All reagents were purchased from Sigma-Aldrich, CO (St. Louis, MO, USA).
The results from the two experimental groups were compared by Student's t-test using the software package GraphPad Prism (La Jolla, CA). Bonferroni adjustments were included for multiple comparisons. The p-values obtained by this method were considered significant if they were <0.05.
Sm10.3 (M22346.1), Sm29 (AF029222.1), Tsp2 (AF521091.1).
The expression of the Sm10.3 gene was detected by real-time PCR at different stages in the S. mansoni life cycle. The only stage during which Sm10.3 mRNA was not detected was the miracidium stage. The highest level of Sm10.3 mRNA expression was observed in lung-stage schistosomula. Sm10.3 expression was also detected in eggs, adult worms and cercariae, but at lower levels than in the schistosomula (Fig. 1).
Cloning and heterologous expression of the Sm10.3 gene was performed as described in the material and methods section. Recombinant Sm10.3 (rSm10.3) was purified from bacteria, and the first seven fractions that were eluted from an affinity chromatography column were combined and dialyzed in PBS pH 7.2 and then analyzed by SDS-PAGE followed by Coomassie blue staining (Fig. 2A). The strong band visible at approximately 26 kDa, the expected mass of the purified rSm10.3, indicates the success of the purification protocol (Fig. 2A). To further evaluate the specificity of the purification procedure, the purified rSm10.3 was analyzed by western blot using a mouse monoclonal anti-His tag antibody (Fig. 2B). The western blot analysis confirmed that the protein around 26 kDa had a histidine tag and this is the molecular weight expected for rSm10.3 (Fig. 2A, B).
The localization of Sm10.3 was determined in S. mansoni lung-stage schistosomula (Fig. 3C and D), female adult parasites (Fig. 3G and H) and male adult parasites (Fig. 3K and L) using specific mouse polyclonal antibodies to rSm10.3 and fluorescence microscopy. Rhodamine phalloidin (red) was used as an actin marker to label the cytoskeletal tegument components and muscle layers (Fig. 3B, D, F, H, J and L). The cell nuclei were stained with DAPI (blue) (Fig. 3). The native Sm10.3 protein (green) was located exclusively in the internal tissues of lung-stage schistosomula (Fig. 3C and D), as well as on the surface of the esophageal and intestinal epithelia of adult parasites (Fig. 3G and H). No Sm10.3-specific signal was detected in sera from naïve mice (pre-serum) in either the adult parasites (Fig. 3E, F, I and J) or the lung-stage schistosomula (Fig. 3A and B).
To evaluate the role of Sm10.3 in the digestive epithelia of adult worms and its possible contribution to the erythrocyte feeding process, we analyzed the effect of rSm10.3 on mouse erythrocytes suspended in PBS. The kinetics of the hemagglutination process were monitored analyzing variations in turbidity by an adapted protocol that was previously described [26]. This methodology allows the measurement of hemagglutination due to the formation of erythrocyte clumps and a subsequent reduction in absorbance. Different amounts of rSm10.3 were added to erythrocyte suspensions at a haematocrit of 5% and followed over time by reading the absorbance at 655 nm with a spectrophotometer. The recombinant protein rSm29 [11] was used as a negative control and the lectin concanavalin A as a positive control of the erythrocytes hemagglutination process, as previously demonstrated [29]–[31]. The addition of a similar amount of rSm29 did not induce any changes in the absorbance. However, the addition of 5 µg/mL or 10 µg/mL of rSm10.3 resulted in a 5% reduction in absorbance, and higher amounts of rSm10.3 (50 µg/mL) reduced the absorbance by 10–15% as compared to rSm29, while 50 µg/mL of concanavalin A decreased the absorbance by 28–32% after 400 s compared to rSm29 (Fig. 4I). The hemagglutination process induced by rSm10.3 can also be observed microscopically, as shown by the formation of erythrocytes clumps upon the addition of recombinant Sm10.3 (Fig. 4E and F).
Later, the hemagglutinating effect of rSm10.3 in mice erythrocytes was clearly evidenced by hemagglutination endpoint dilution assays. Protein concentrations of rSm29 up to 250 µg/mL were not able to induce hemagglutination (Fig. 4G), while rSm10.3 caused erythrocytes agglutination with the lowest protein concentration of 31.2 µg/mL (Fig. 4G). The minimal protein concentration of concanavalin A required to induce hemagglutination was 3.9 µg/mL, approximately eight times lower than the amount observed for rSm10.3 (Fig. 4G). These results confer to concanavalin A a total hemagglutinating activity of 1.56×102 units (U) and a specific hemagglutinating activity of 40000 U/mg/mL (Table S1), while these values for rSm10.3 were 13 times lower and 100 times lower, respectively, 0.12×102 units and 400 U/mg/mL (Table S1).
By means of hemagglutination inhibition assays using anti-rSm10.3 it was demonstrated the specific and protein concentration dependent pattern of the rSm10.3 hemagglutinating activity (Fig. 4H). There was no hemagglutination inhibition with 250 µg/mL of rSm10.3, even at the highest anti-rSm10.3 concentration (1∶1 dilution). However, with lower rSm10.3 concentrations its hemagglutinating activity was inhibited by anti-rSm10.3 (Fig. 4H). When using 125 µg/mL of rSm10.3 there was inhibition at anti-rSm10.3 dilutions ranging from 1∶1 to 1∶8, while with 62.5 µg/mL of rSm10.3 the inhibition was evident from 1∶1 to 1∶16 dilutions (Fig. 4H).
Sera from ten animals from each vaccination group were tested by ELISA to evaluate the levels of specific IgG, IgG1 and IgG2a antibodies to rSm10.3. Significant titers of anti-rSm10.3 IgG antibodies were detected at all time points tested after the first immunization (Fig. 5A). The levels of specific IgG1 antibodies increased at 30 days after the first immunization (Fig. S1A), while the levels of specific IgG2a antibodies continued to increase up to day 75 (Fig. S1B). Furthermore, the IgG1/IgG2a ratio was reduced at days 45, 60, 75 and 90 (Fig. 5B), which correlates with the elevation in anti-rSm10.3 IgG2a production (Fig. S1B).
To test the potential usefulness of rSm10.3 as part of an anti-schistosome vaccine, we asked whether this recombinant antigen could induce protection in a murine model of S. mansoni infection. Two independent vaccination trials were conducted and C57BL/6 mice were immunized three times with rSm10.3 formulated with Freund's adjuvant and then challenged with 100 S. mansoni cercariae. The control group received adjuvant only in phosphate-buffered saline. Mice vaccinated with rSm10.3 showed a 32.0% reduction in the adult worm burden in the first trial (Fig. 5C) and 25.5% reduction in the second trial (Fig. 5D). Regarding the number of eggs in mice livers, it was observed 43.6% and 32.9% reduction in the number of eggs per gram of liver tissue in the first and second trials, respectively (Fig. 5E).As shown in Figure 6A–D, histopathological analysis of the hepatic tissue, from animals of the first vaccination trial, demonstrated that rSm10.3 immunization reduced the extent of fibrosis compared to control animals. These analysis showed a 23.8% reduction in liver granuloma counts (Fig. 6E), an 11.8% reduction in granuloma area (Fig. 6F), and a 39.8% reduction in granuloma fibrosis (Fig. 6G), as compared to control mice.
To evaluate the cytokine profile of mice immunized with rSm10.3, splenocytes were isolated from spleens of vaccinated and control animals after the third immunization. Statistically significant levels of IFN-γ, the signature cytokine of a Th1-type immune response, and TNF-α, a pro-inflammatory cytokine, were detected in the supernatant of cultured splenocytes from immunized animals compared to the control group (Fig. 7A and B, respectively). IL-5, a characteristic Th2-type cytokine, was also detected in the supernatant of cultured splenocytes from rSm10.3-immunized mice at statistically significant levels compared to the PBS control (Fig. 7-D). Furthermore, high levels of the modulatory cytokine IL-10 were also observed in vaccinated animals (Fig. 7C). Concanavalin A (ConA) and LPS were used as positive controls to confirm that the splenocytes were responsive to stimuli. As shown in Figure 7, ConA induced the production of IFN-γ, IL-5 and IL-10, while LPS induced the production of TNF-α.
Schistosomiasis is one of the most important neglected tropical diseases. Effective control is unlikely in the absence of improved sanitation and the development of a vaccine. Proteins located at the host/parasite interface, particularly molecules that are secreted or surface-exposed on the tegument and the intestinal epithelia, are the most promising targets for developing an anti-schistosomiasis vaccine [18]. We therefore evaluated the potential of the Sm10.3 antigen as a vaccine candidate, as it was previously reported to be localized on the esophageal gland in schistosomula and adult worms [15], as well as on the gut primordium of the nonfeeding cercaria [15]. The deduced amino acid sequence of Sm10.3 contains a signal peptide, and the protein was predicted to be secreted or localized to the exterior surface of the cell. The gene products of several other MEG family members contain signal peptides for secretion and are secreted from different schistosomal glands and epithelia [20], [15].
We confirmed previous reports [20], [15] that Sm10.3 is mainly expressed in the schistosomulum stage, as well as in other stages that involve contact with the mammalian host, such as eggs, cercariae and adult worms. MEG genes are difficult to clone, primarily due to extensive alternative splicing that generates variant transcripts of different sizes through exon skipping and the arbitrary combination of exons [19], [20], [15]. This variation in MEG gene products may represent a strategy used by members of the Schistosoma genus to confuse the host immune system, similar to the mechanisms of surface protein variation in Trypanosoma brucei and Plasmodium falciparum [20], [15], [18]. In this study, we produced the recombinant Sm10.3 protein from a synthetic gene that allowed us to express a protein corresponding to the largest transcript from the Sm10.3 gene to optimize codon usage and avoid errors in the amino acid sequence.
Our fluorescence microscopy data confirm the in silico prediction that Sm10.3 is secreted or located on the cell surface, demonstrating that the native Sm10.3 protein localizes to the epithelia and lumen of the intestinal tract in adult parasites, as well as to the internal tissues of lung-stage schistosomula. A previous microarray study demonstrated an increase in Sm10.3 expression in lung-stage schistosomula [15]. In the same study, Sm10.3 localization was examined using an antibody against a synthetic peptide (immunocytochemistry) and RNA hybridization (WISH). The authors found Sm10.3 proteins in the esophageal gland and the esophageal lumen of adult worms, but were unable to define the localization of Sm10.3 in the larval stage [15]. A more recent study from this group demonstrated that the Sm10.3 antigen is highly O-glycosylated, which means that native Sm10.3 a very sticky macromolecule that is highly likely to adhere to surfaces [35]. The authors suggest that this adherence could be responsible for the immunodetection of Sm10.3 not only in the esophageal gland but also throughout the entire esophagus (distal to secretion sites) [35], [15]. We did not detect any Sm10.3 accumulation in the esophageal glands, but we did find Sm10.3 on the esophageal epithelia, in the esophageal lumen and in the gut epithelia. It was observed immunoreactivity in the gut epithelia and sometimes in the gut lumen in all samples that were imaged. However, it was not possible to obtain image sections showing the entire gut length, which prevent us to state that Sm10.3 is present in the whole gut. These apparent differences in Sm10.3 localization could be due to the different antibodies that were used in the studies. Alternatively, variant Sm10.3 transcripts could be differentially regulated depending on the experimental conditions. It has been previously shown that environmental stimuli can affect gene expression in S. mansoni, as the presence or absence of erythrocytes altered the transcription levels of genes expressed in the tegument and related to feeding [14].
To further investigate the role of Sm10.3 protein in the esophagus and gut of adult worms, we analyzed the effect of rSm10.3 protein on erythrocytes and found that rSm10.3 induced hemagglutination in vitro. This effect could be related to erythrocyte digestion and nutrition in adult worms, and may represent a role for Sm10.3 in adult worms. However, it is necessary to evaluate in vivo the impact of this Sm10.3-induced hemagglutination on the blood feeding process. In addition to the possible role of Sm10.3 protein in blood feeding, it is also likely that Sm10.3 contributes to protein variation, a role that has been previously proposed for the MEG family members [20], [15], [18]. The digestive tracts of schistosomes in the blood feeding stages are accessible to macromolecules such as albumin and immunoglobulins [36], [37], which implies that there may be direct contact between the digestive epithelia and the host immune system.
We also assessed the interaction of Sm10.3 with the host immune system and its potential as a vaccine candidate. rSm10.3 induced high levels of anti-rSm10.3 IgG production in the sera of immunized mice after the second immunization. A decrease in the ratio between the IgG subtypes (IgG1/IgG2a) was observed 45 days after the first immunization. Furthermore, cytokine analysis of the supernatants of cultured splenocytes stimulated with rSm10.3 suggests a mixed Th1/Th2-type immune response. Studies using the irradiated cercariae model, which induces high levels of protection in mice, suggest that effective protection can be based on a mixed Th1/Th2 response, a polarized Th1 response or even a polarized Th2 response [38]. In previous studies performed by our group, the majority of S. mansoni antigens tested as recombinant protein vaccines that conferred partial protection against cercariae challenges induced a Th1-type immune response [11], [34], [39], [40] or a mixed Th1/Th2 response [41], [42], [43]. IFN-γ is involved in protective immunity against schistosomiasis, as specific anti-IFN-γ antibodies completely abolish the protection conferred by vaccination with irradiated cercariae [44]. Similar results were obtained in a study using IFN-γ knockout mice [45]. The partial protection conferred by vaccination with rSm10.3 resulted in 25.5% to 32% reduction in worm burden, and the overall pathology was reduced, as shown by 32.9% to 43.6% reduction in the number of eggs per gram of hepatic tissue, a 23.8% reduction in the number of granulomas, an 11.8% reduction in the area of the granulomas and a 39.8% reduction in granuloma fibrosis. It is possible that the reduced liver pathology is related to the elevated levels of IL-10 detected in immunized mice, which may regulate Th2 responses and/or prevent the development of a polarized Th1 response, consequently reducing inflammation and liver injury [46], [47].
In conclusion, our results confirm that Sm10.3 is mainly expressed during the stages of the Schistosoma mansoni life cycle that involve contact with the mammalian host. We show that the Sm10.3 protein is located in the intestinal tract of adult worms, providing the first evidence for a possible role for Sm10.3 in the blood feeding process. Finally, our data suggest that Sm10.3 is a potential candidate for use in developing a multi-antigen vaccine to control schistosomiasis.
|
10.1371/journal.pntd.0004565 | A Simple and Novel Strategy for the Production of a Pan-specific Antiserum against Elapid Snakes of Asia | Snakebite envenomation is a serious medical problem in many tropical developing countries and was considered by WHO as a neglected tropical disease. Antivenom (AV), the rational and most effective treatment modality, is either unaffordable and/or unavailable in many affected countries. Moreover, each AV is specific to only one (monospecific) or a few (polyspecific) snake venoms. This demands that each country to prepare AV against its local snake venoms, which is often not feasible. Preparation of a ‘pan-specific’ AV against many snakes over a wide geographical area in some countries/regions has not been possible. If a ‘pan-specific’ AV effective against a variety of snakes from many countries could be prepared, it could be produced economically in large volume for use in many countries and save many lives. The aim of this study was to produce a pan-specific antiserum effective against major medically important elapids in Asia. The strategy was to use toxin fractions (TFs) of the venoms in place of crude venoms in order to reduce the number of antigens the horses were exposed to. This enabled inclusion of a greater variety of elapid venoms in the immunogen mix, thus exposing the horse immune system to a diverse repertoire of toxin epitopes, and gave rise to antiserum with wide paraspecificity against elapid venoms. Twelve venom samples from six medically important elapid snakes (4 Naja spp. and 2 Bungarus spp.) were collected from 12 regions/countries in Asia. Nine of these 12 venoms were ultra-filtered to remove high molecular weight, non-toxic and highly immunogenic proteins. The remaining 3 venoms were not ultra-filtered due to limited amounts available. The 9 toxin fractions (TFs) together with the 3 crude venoms were emulsified in complete Freund’s adjuvant and used to immunize 3 horses using a low dose, low volume, multisite immunization protocol. The horse antisera were assayed by ELISA and by in vivo lethality neutralization in mice. The findings were: a) The 9 TFs were shown to contain all of the venom toxins but were devoid of high MW proteins. When these TFs, together with the 3 crude venoms, were used as the immunogen, satisfactory ELISA antibody titers against homologous/heterologous venoms were obtained. b) The horse antiserum immunologically reacted with and neutralized the lethal effects of both the homologous and the 16 heterologous Asian/African elapid venoms tested. Thus, the use of TFs in place of crude venoms and the inclusion of a variety of elapid venoms in the immunogen mix resulted in antiserum with wide paraspecificity against elapid venoms from distant geographic areas. The antivenom prepared from this antiserum would be expected to be pan-specific and effective in treating envenomations by most elapids in many Asian countries. Due to economies of scale, the antivenom could be produced inexpensively and save many lives. This simple strategy and procedure could be readily adapted for the production of pan-specific antisera against elapids of other continents.
| Antivenom is the most effective treatment modality for snake envenoming. However, they are specific and effective against only one or a few snake venoms. Production of antivenom against many snake species covering a wide geographic area of some countries or regions e.g., Asia and Africa, is not yet possible. This study aimed to use a simple procedure to produce horse antiserum which could neutralize many or all medically important elapid (neurotoxic) snakes (cobras, kraits) of Asia. The venoms of 6 elapid species were obtained from 12 different regions/countries of Asia. Nine of these venoms were ultra-filtered to remove high molecular weight, non-toxic proteins to obtain the toxin fractions (TFs) for use as immunogen and thus enable inclusion of greater variety of elapid venoms. The 9 ‘toxin fractions’ together with the remaining 3 crude venoms were used to immunize 3 horses. The antisera of the horses obtained were shown to neutralize, in mice, the lethal effects of the venoms used in the immunization and 16 other Asian/African elapid venoms not used in the immunization. Thus, the simple strategy could broaden the neutralizing capacity of the resulting antiserum. The antiserum could be processed into antivenom with wide paraspecificity and effective against many elapid snakes of Asia.
| Snake envenoming is an important medical problem in various developing countries of Asia and Africa [1, 2]. It has been estimated that at least 1.2 million people are affected annually with about 20,000 deaths [3]; however these figures likely represent merely the tip of the iceberg as a result of poor epidemiological records [4–6]. This serious envenoming problem has led WHO to recognize it as one of the neglected tropical diseases [7], in addition to the long recognized status of snakebite as an occupational hazard and a disease of poverty [2]. Despite this, the provision of global funding and technologies aimed at solving the global snakebite envenomation problem has been limited. Moreover, efforts to solve the problem have largely been taken up by regional toxicologists through research initiatives designed to gain a better understanding of the compositional variation of venoms and facilitate the production of effective, affordable, broad-spectrum antivenoms [8, 9]. Indeed, the rational and most effective treatment for snake envenomation is the administration of specific antivenom which remains unavailable in many parts of the world. Fortunately, attempts are being made on various research fronts to produce these antivenoms [9, 10].
Antivenoms (AVs) are usually produced in horses, although other large animals like camels, donkeys and sheep can also be used [11–14]. AV can be monospecific or polyspecific. Monospecific AV is specific and effective against the snake venom used as the immunogen and some related cross-reacting species. Therefore, identification of the culprit snake is necessary. Polyspecific AVs are effective against the multiple venoms that are used in the immunization and some other cross-reacting venoms. The cross-reactivity or paraspecificity of polyspecific antivenom is an important property useful to pan-specific antivenom production. However, one problem with the production of polyspecific AVs is that only about 3–6 snake venoms can be used in the immunization [15, 16]. Although the maximum number of venoms has not been established and reported, it is believed that immunization of more than 5–6 venoms resulted in lower potency of the polyspecific AVs. Thus, in order to increase coverage, 2 (or more) polyspecific AVs were often combined to yield a mixed-polyspecific AV that is effective against a broader spectrum of venoms [16]. Although the mixed-polyspecific AVs are very useful against a wide range of venoms and make identification of the culprit snake prior to AV administration unnecessary, it is probably more expensive to produce since 2 or more groups of horses are needed.
In order to prepare a truly polyspecific AV with wider venom coverage, various technologies have been employed. For example, antivenomics studies of a number of polyspecific AV have identified which of the heterologous venom’s lethal toxins interact, or fail to interact, with the AV antibodies [17–19]. From such studies, those venom toxin antigens that interact weakly or fail to interact with the AV antibody can be isolated and added to the immunogen mix to improve the coverage of the AV. Antivenomics studies could also provide valuable information about common venom antigens and aid in the selection of appropriate venom antigens to be included in the immunogen mix.
Another interesting approach is the use of DNA immunogens termed ‘epitope strings’ [20]. DNA immunization of mice with the ‘epitope string’ resulted in antibodies that could neutralize the toxins of several species of African viper [20]. This technology could be a useful possibility in the future.
These very interesting approaches are promising and could eventually result in the production of effective pan-specific antivenoms. Meanwhile, some simple, readily applicable protocols that would result in effective antivenoms would be desirable.
Snake venoms contain mixtures of more than 100 proteins with different molecular weights and biological activities [21]. Bites by elapids, such as cobras, kraits and mambas, are considered ‘neurotoxic’ as they cause neuromuscular paralysis mediated by low molecular weight toxins of the 3 finger toxin family (3FTs) [22]. The most important toxins among the 3FTs are the postsynaptic neurotoxins (PSNT). PSNTs bind specifically, and quasi-irreversibly, to nicotinic acetylcholine receptors (nAchR) at the neuromuscular junction [23, 24]. Toxin binding results in inhibition of neuromuscular transmission, muscle paralysis and death by respiratory failure [23, 25]. Another important subtype of the 3FTs is the cytotoxins (cardiotoxins), which have cytolytic activity and are mainly involved in local tissue necrosis [26, 27].
In addition to 3FTs, the krait venoms (genus Bungarus) also contain lethal basic phospholipase A2 presynaptic neurotoxins. These toxins have MWs of about 21–30 kDa [28]. They can cause damage to motor neuron terminals and cause release of the neurotransmitter acetylcholine, from the nerve endings [29]. Depletion of acetylcholine in the nerve terminals results in neuromuscular transmission blockage with death occurring as a result of respiratory failure.
These postsynaptic and presynaptic neurotoxins, as well as acidic phospholipases, are therefore the most important cause of death resulting from elapid envenoming. AVs must be able to neutralize these toxins in order to be effective and life-saving. The other high MW and thus highly immunogenic elapid proteins, mostly hydrolytic enzymes are usually not lethal [30]; antibodies to neutralize these venom proteins are not essential in saving the lives of victims.
From the above information, it should be possible to prepare AVs effective against elapid venoms using only the 3FTs and presynaptic neurotoxins as immunogens. This approach would have the advantage of reducing the total number of venom antigens the horse is exposed to. Consequently, it should be possible to significantly increase the number of different snake venoms used in the immunogen mix. This will expose the horse immune system to a wide variety of lethal toxins from numerous venoms with diverse repertoires of toxin epitopes. This, in turn, should broaden the paraspecificity of the AV antibodies and increase the cross-neutralization of the AV against venoms of species not included in the immunogen mix, hence extending its use to more countries, particularly those which are socioeconomically disadvantaged.
The rationale described above suggests that it should be possible to prepare a ‘pan-specific’ AV effective against most of the elapids of Asia. In the present study, a mixture of ‘toxin fractions’ (TFs) and venoms from 6 medically important elapid venoms (WHO category 1) obtained from 12 regions/countries, was used at low doses to immunize horses. It was shown that the horse antiserum could neutralize all the 27 homologous/heterologous elapid venoms (including 4 African Najas) tested. Details of the preparation and characterization of the immunogens and the in vitro and in vivo potency of the antiserum are described.
The horses were of mixed breed and were 3–6 years old and weighed about 420–480 Kg. They were under the care of veterinarians with expertise in equine health at the Animal Hospital of the Animal and Remount Department, The Royal Thai Army, and the veterinarians of the Faculty of Veterinary Science, Mahidol University. There were dewormed to remove gut helminthes and were free of external parasites. The horses were vaccinated against rabies, tetanus, and equine encephalytis. Their hematologic, hepatic and kidney status were tested prior to and monitored during the experiment. They were kept in clean well-ventilated brick-made stables, and were allowed to stay on the pasture for several hours every day. Albino mice (ICR strain, 20-25g) were supplied by the Animal Experimental Unit, Faculty of Medicine, University of Malaya.
Experiments involving the care, bleeding and immunization of horses with various venoms were reviewed and approved by the Animal Care and Use Committee of the Faculty of Veterinary Science, Mahidol University, Protocol and clearance no.MUVS-2012-69 in accordance with the Guidelines of the National Research Council of Thailand.
The protocol of animal study on mice was based on the guidelines given by the Council for International Organizations of Medical Sciences (CIOMS) and was reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Malaya (Ethical clearance No. 2014-09-11/PHAR/R/TCH).
The TFs of elapid venoms were prepared by separately dissolving the venom in 100 mM ammonium acetate, pH 5.0 to make 1.0 mg/ml final protein concentration. High molecular weight venom proteins of the Naja spp. were removed by filtration of the venom solution through 30 kDa molecular weight cut off (MWCO) ultra-filtration membrane (Amicon) at 14,000 x g for 10 min at 4°C. Venoms of the Bungarus spp. were filtered through 50 kDa MWCO ultra-filtration membrane. The volume of the filtrates and the retentates were recorded and their protein contents were assayed. The filtrates were called ‘toxin fraction’ (TF) which were further characterized by one-dimensional SDS-PAGE, RP-HPLC and protein determination. The crude venoms and the TFs were kept frozen at -20 °C until used.
The antibody titer in each serum sample was determined by an indirect ELISA as described by Rungsiwongse and Ratanabanangkoon [38]. A polyvinyl microtiter plate (Costar) was coated with 50 μl/well of 5 μg/ml (protein content) of each TF in 0.05 M sodium carbonate-bicarbonate buffer pH 9.6 for 18 hr at 4°C. The plate was washed 4 times with 0.05% Tween-20 in normal saline (NSST). The serum from each horse was 4-fold diluted, starting from 1: 100 to 1: 2.6 x 107 in diluting buffer (0.15 M PBS, pH 7.4 containing 0.05% Tween-20 and 0.5% BSA). Then, 50 μl of the sera at various dilutions were added into each well and incubated for 1 hr at room temperature. The wells were washed 4 times with NSST before 50 μl/well of 1: 40,000 diluted sheep anti-horse IgG-horseradish peroxidase conjugate (Sigma) in diluting buffer was added and incubated at room temperature for 1 hr. After 4 washes, 100 μl/well of freshly prepared substrate solution (0.01% 3,3′,5,5′-Tetramethylbenzidine (TMB) and 0.03% hydrogen peroxide in 0.075 M citrate-phosphate buffer, pH 5.0) was added into each well and incubated in the dark for 15 min at room temperature. The reaction was terminated by adding 25 μl of 4 N sulfuric acid. The plates were read at 450 nm against blank using an ELISA reader (Labsystem Multiskan, Ascent). The highest dilution giving an absorbance reading of 0.4 was regarded as the end point titer using GraphPad Prism 6 program. A positive reference serum (pooled monospecific anti-N. kaouthia horse serum at dilution 1:6400) and the principal postsynaptic neurotoxin (NTX3) of N. kaouthia used as a standard reference antigen, were added in every plate to correct for day-to-day and plate-to-plate variations.
Median lethal doses (LD50) of the venoms and median effective doses (ED50) of pAS are expressed with 95% confidence interval (CI). LD50, ED50 and the 95% CI were calculated using the probit analysis method of Finney [39] with the Biostat 2009 Analysis software (AnalySoft Corp., Bracknell, UK).
Protein concentration was determined as described by Lowry et al. [41] using bovine serum albumin as standard. The mass of venoms and other proteins referred to in this study was in protein content as assayed by Lowry et al [41].
When the individual venoms of the Naja spp. were ultra-filtered through the 30 kD MWCO (molecular weight cut-off) membrane and the Bungarus spp. venoms through the 50 kD MWCO filters, the filtrates containing the toxin fractions (TFs) of the venoms were obtained. Fig 1 shows the SDS-PAGE of the crude N. kaouthia (Thailand) venoms and the corresponding TF. Both the gels and the scanning profiles showed that the high molecular weight proteins (>30 kDa, including protein bands B1, B2, B3 and B4, total 5.2% of protein content by protein assay) were effectively removed while the low MW toxic proteins (proteins B5 to B10) were all present (as C1 to C6, total 94.8% of protein). Similar findings were obtained for the other venoms and their corresponding TFs. The SDS-PAGE and the amount of high molecular weight proteins removed by ultra-filtration are shown in Table 3 and S1 Fig. The protein content of high MW venom proteins removed by the ultrafiltration process range from 3.13% (B. multicinctus, Taiwan) to 17.52% (N. kaouthia, Vietnam).
We also examined the proteome of the TF of N. kaouthia (Thailand) and compared with the crude venom proteome. TF was subjected to RP-HPLC and the elution profiles are shown in Fig 2. Comparison of the RP-HPLC patterns with that of the crude venom [34] indicated that for N. kaouthia venom, the ultrafiltration successfully removed the bulk of the high molecular weight venom proteins while retained the lethal, low molecular weight toxins of the venom. All the six groups of venom toxins (short postsynaptic neurotoxin (SNTX), long postsynaptic neurotoxin (LNTX), muscarinic toxin-like proteins (MTLP), weak postsynaptic neurotoxin (WTX), cytotoxin (CTX) and phospholipase A2 (PLA2) were almost fully retained in the TF whereas the high molecular weight snake venom metalloproteinase (SVMP) and L- amino acid oxidase (LAAO) were totally removed. Table 4 shows the lethality of TFs and venoms from four representative venoms: N. kaouthia (Thailand and Malaysia), B. candidus (Northeast Thailand) and B. multicinctus (China) in a mouse model. The results show that the crude venoms and the corresponding TFs exhibited statistically comparable values of LD50s. Since the amount of high MW proteins removed by ultrafiltration was quite small and the LD50 determination was subjected to high biological variations, the expected lower LD50 values of the TFs were not observed (except in the case of N. kaouthia Vietnam). However, as mentioned above, the RP-HPLC and venomics data together indicated that all the major lethal toxins of each venom were indeed recovered in their corresponding TFs.
The lethal effects measured in median lethal doses (LD50s) of all the 27 homologous and heterologous venoms including 4 African elapids are shown in Table 6. The LD50s of these snake venoms varied widely. The most lethal venom was that of B. sindanus (LD50 of 0.018 μg/g) while the least toxic (LD50 of 1.80 μg/g) was N. naja (India) venom. In general, the Bungarus venoms were more lethal than those of the Naja. It should be mentioned that the amount of B.candidus (South Thailand) venom which was used as immunogen was not enough for the in vivo neutralization assay.
The effective doses at 50% survival rate (ED50s) of the pAS against the homologous and heterologous venoms are shown in Tables 7 and 8, respectively. Overall, the pAS could neutralize all the venoms tested including 4 African najas with different degree of effectiveness. For the homologous venoms, the potency value (P) which is theoretically independent of the challenging dose [40] ranged from 0.712 mg/ml against N. kaouthia (Malaysia) venom to 0.101 mg/ml against N. kaouthia (Thailand). Among the heterologous venoms, the potency value (P) of pAS ranged from 0.672 mg/ml against B. caeruleus (India) to 0.0297 mg/ml against N.haje (Egypt) venom. A higher potency value (P) implies a better capability of venom neutralization by the pAS.
There were 4 heterologous venoms (B. fasciatus, N. nigricollis (Cameroon), N. naja (India) and N. naja (Sri Lanka)) that were neutralized by pAS only when the venom challenge dose were low at 1.5LD50. At this venom dose in the absence of pAS, all the mice died.
In the present study, a novel yet simple approach was used to produce a pan-specific snake antiserum with wide paraspecificity. Nine toxin fractions (TFs) derived from various elapid venoms together with other three crude venoms were used at very low dose as immunogen. The use of TFs from 9 elapid venoms as immunogen reduced the amount of high MW, highly immunogenic and largely non-toxic venom proteins in the immunogen. The low doses of TFs/venoms not only reduced the possible toxicity on the horse but could induce high affinity neutralizing antibody [43]. The horse antiserum obtained showed wide paraspecificity and neutralized the lethality of 27 homologous and heterologous elapid venoms from various countries of Asia including 4 from Africa.
There are a few methods whereby high MW venom proteins can be selectively removed from the venom. Since the elapid 3FTs (neurotoxins and cytotoxins) and phospholipase A2 are relatively heat-stable at acidic pHs, heating of elapid venoms at 100°C at pH 5.0 could quantitatively recover the α-neurotoxin, cytotoxins and phospholipase A2 of N. kaouthia venom [44]. However, this method may not be applicable to the venoms of Bungarus spp. because the presynaptic neurotoxins (toxic phospholipases A2) may not be able to withstand high temperature. Alternatively, separation of the high MW venom proteins from the toxic components could also be achieved by size-exclusion chromatography. In this work, we chose to use ultra-filtration method as the method is simpler, faster, economical as well as effective. This method works by filtering proteins below a certain molecular weight through a non-denaturing polyethersulfone membrane with specific pore size selected prior to ultrafiltration. TFs of lower MW hence could be optimally recovered from the filtrate and was devoid of high MW proteins as shown by SDS-PAGE (Fig 1 and S1 Fig). The amount of the high MW venom proteins removed by ultra-filtration ranged from 4.7% for B. multicinctus (China) venom to 17.52% for N. kaouthia (Vietnam) venom. Though relatively small in quantity, these high MW proteins are highly immunogenic and have been shown by Western Blot to induce high titer of antibody [45]. Although ‘antigenic competition’ [46] in antivenom production has not yet been established, the presence of these proteins in the immunogen mix could possibly interfere with the production of antibodies specific against the lethal toxins of the venom as have been observed with other venoms [47–49].
In this study, the venomic method of comparing the proteomes of TF and crude venom of N. kaouthia provided evidence that all the important toxins of the venom were recovered in the TF. Also, LD50’s of the crude venoms and the TFs of the 4 selected venoms were found to be comparable, further confirming that essentially all the principal toxic components of the venom were present in the respective TF. The results support the use of TFs as immunogen in the preparation of antiserum, as the TFs contain essentially all the lethal components that should be neutralized for effective treatment.
With the ‘low dose, low volume multi-site immunization protocol’ employed [36, 37], the sera antibody ELISA titers of all 3 horses rose rapidly and reached plateau in only about 4–6 weeks. The fast kinetics of antibody production, observed previously with elapid and viperid venoms [36, 37, 42], could significantly reduce the cost of antivenom production. However, with the many and heterogeneous venom proteins used as immunogens and as ELISA antigens of this study, the maximum ELISA titers varied widely. It seemed that, for the homologous venoms, the titers of the pAS against the Naja venoms were higher than those against the Bungarus venoms. When compared with the in vivo neutralization results, the ELISA results did not seem to correlate well with the in vivo neutralization results. Thus, the ELISA titers could only be used as a rough guide on the kinetics of antibody production but not on the neutralization capacity of the sera antibody. The reason behind this observation is not apparent at the moment. The relative content of the lethal toxins in the TF proteins is likely to be involved in this correlation. For example, while all the members of the 3FTs (postsynaptic neurotoxins, cytotoxins, weak neurotoxins) could act as antigens and contribute to the ELISA titers, but only the postsynaptic neurotoxins, short and long, play major parts in the lethality of the venom and in the neutralization of the pAS.
The TFs used as immunogens were derived from venoms of WHO Category 1 medically important snakes of several countries in Asia, so were the three venoms used. Moreover, some species were taken from different regions or countries. The reason for this was that intra-specific variations in venom compositions and clinical manifestations have been widely observed [50–54] and this variation could increase the diversity of the venom toxins and enhance the snake coverage of the resulting antiserum. For example, N. kaouthia venoms were from Thailand, Vietnam and Malaysia; these venoms had been shown to exhibit intra-specific variations in toxin profiles and lethality [34]: N. kaouthia venom from Thailand is high in long neurotoxin, the venom from Malaysia is low in neurotoxin but very high in cytotoxin, whereas venom from Thailand is high in weak neurotoxin; and both venoms from Malaysia and Vietnam contain more short neurotoxins. N. philippinensis and N. atra venoms are both known to contain very high content of short neurotoxin [55, 56]. N. sputatrix venom also contained mainly short neurotoxin, large amount of cytotoxin and lethal PLA2 [57]. Thus, together, the Naja venoms selected contain a good balance of 3FTs, including long, short and weak neurotoxin as well as cytotoxins.
On the other hand, antigens of krait venoms were contributed by B. candidus and B. multicinctus venoms to cover the lethal β-neurotoxins that induce presynaptic neurotoxic effect, distinct from the post-synaptic neurotoxicity caused by the cobra venoms. In the case of B. candidus, the venom was from Bandung (Indonesia), Thailand (northeast and south). We did not have the B. candidus venom from Malaysia at the time of immunization, B. candidus venom from southern Thailand was therefore used as a substitute.
Overall, the purpose of this approach (use of very low doses of many TFs/venoms from different geographical areas as immunogen) was to increase the diversity and repertoire of the epitopes of the lethal toxins exposed to the horse, and thereby increasing the paraspecificity of the resulting antiserum. Moreover, the use of very low doses of immunogens was to induce high affinity specific antibodies [43]. Three crude venoms were included as immunogen because these venoms were available in very limited amounts at the time, and were not enough for the entire immunization program if they were prepared as TFs. Whether the small amount of high MW proteins of these 3 crude venoms had any inhibitory effect [47–49] or adjuvant effect [58] on the antibody production observed here is not known.
It is encouraging that the pAs produced using the mixture as immunogen could neutralize all the 27 homologous and heterologous elapid venoms tested albeit with varying potencies. The amount of B. candidus (South Thailand) venom which was used as immunogen was not available for the in vivo neutralization assay. It can be seen that the neutralization potencies (P) which theoretically is independent of the challenge venom dose [40], varied widely. For example, among the homologous venoms, the P value against N. kaouthia (Thailand) was only 0.101 mg/ml while the P value against N. kaouthia (Malaysia) was 0.712 mg/ml. The LD50s of the Thai and the Malaysian venoms were 0.18 and 0.90 μg/g mouse, respectively. Moreover, the heterologous N. oxiana (Pakistan) venom with the LD50 of 0.90 was neutralized by pAS with P of 0.525 mg/ml. Thus, it seems that the difference in neutralization potency of pAs against these venoms correlated with the venom lethality which in turn related to the difference in the contents of the lethal postsynaptic neurotoxin of these venoms [34]. However, this might be a simplistic interpretation and the situation may be more complex. The number of epitopes on the lethal toxins and the binding affinity of the pAS antibodies against these epitopes, are likely to play important roles in determining the neutralization potency of the antiserum. Thus, among the heterologous venoms, the LD50s of B. fasciatus and B. sindanus were 1.50 and 0.018 μg/g mouse, respectively. The former krait venom was about 83 folds less toxic than the latter, yet the P values of the pAS against these two venoms were quite comparable (Table 8). More detailed information of the biochemistry, immunochemistry and pharmacology of the toxins of these venoms are needed to gain a thorough understanding of the relative neutralization capacity of the pAS against them.
It should be emphasized that the relatively low neutralization potencies against some venoms reported here are those of the crude horse antiserum, and could not be strictly compared with those of the fractionated and concentrated therapeutic IgG or F(ab’)2 antivenom. Depending on the manufacturer and the starting potency of the horse sera, the antisera are usually processed and concentrated several folds so that the final antibody product passes the minimal potency requirement. Thus, the pAS prepared in the present study could be successfully processed into an effective pan-specific antivenom against most of the medically important cobra and kraits of many parts of Asia.
It is interesting to note that the pAS could also neutralize the venoms of 4 African cobra: N. melanoleuca (North Cameroon), N. nubiae (Egypt), N. nigricollis (Cameroon) and N. haje (Egypt), though with the neutralization potencies not quite as high as those against the homologous Asiatic cobras. Inclusion of these venoms (as TFs) in the immunogen mix could possibly provide better neutralization against these and possibly some other African elapid venoms as well.
As discussed above, at this stage we could not compare the potencies of our antiserum directly to the potencies of commercial antivenoms available in the region. However, the commercial antivenoms available in this region are usually country-specific and are raised against only limited species important to the country; antivenoms produced as such generally have limited therapeutic value for use against other species from another region/country, due to vast variations affected by geographical and inter-species factors. These variations were believed to be taken care of by the unique method innovated in this study.
The current study thus demonstrated the feasibility of producing a pan-specific antivenom against many cobras and kraits in Asia through a novel yet simple immunization strategy. The procedure involved simple and inexpensive ultrafiltration method and small quantities (about 5 mg/venom) of the homologous venoms for the entire immunization program. It could readily be applied to the production of any pan-specific AS against elapid (Naja and Bungarus) snakes. The findings should provide useful insights into the optimization of immunogen preparation with the aim to broaden the paraspecificity of antivenom for clinical use, an effort in line with the Global Snakebite Initiative [59]. With the economy of scale, the pan-specific antivenom could be produced economically and offered as a more sustainable and affordable supply to many countries; hence saving the lives of many victims succumbed to snakebite envenomation.
It is also hoped that the simple ‘low dose, diverse toxin repertoires’ strategy employed in this study could be adapted in the preparation of pan-specific antisera against medically important elapids of other continents.
|
10.1371/journal.pgen.1006495 | Meta-GWAS Accuracy and Power (MetaGAP) Calculator Shows that Hiding Heritability Is Partially Due to Imperfect Genetic Correlations across Studies | Large-scale genome-wide association results are typically obtained from a fixed-effects meta-analysis of GWAS summary statistics from multiple studies spanning different regions and/or time periods. This approach averages the estimated effects of genetic variants across studies. In case genetic effects are heterogeneous across studies, the statistical power of a GWAS and the predictive accuracy of polygenic scores are attenuated, contributing to the so-called ‘missing heritability’. Here, we describe the online Meta-GWAS Accuracy and Power (MetaGAP) calculator (available at www.devlaming.eu) which quantifies this attenuation based on a novel multi-study framework. By means of simulation studies, we show that under a wide range of genetic architectures, the statistical power and predictive accuracy provided by this calculator are accurate. We compare the predictions from the MetaGAP calculator with actual results obtained in the GWAS literature. Specifically, we use genomic-relatedness-matrix restricted maximum likelihood to estimate the SNP heritability and cross-study genetic correlation of height, BMI, years of education, and self-rated health in three large samples. These estimates are used as input parameters for the MetaGAP calculator. Results from the calculator suggest that cross-study heterogeneity has led to attenuation of statistical power and predictive accuracy in recent large-scale GWAS efforts on these traits (e.g., for years of education, we estimate a relative loss of 51–62% in the number of genome-wide significant loci and a relative loss in polygenic score R2 of 36–38%). Hence, cross-study heterogeneity contributes to the missing heritability.
| Large-scale genome-wide association studies are uncovering the genetic architecture of traits which are affected by many genetic variants. In such efforts, one typically meta-analyzes association results from multiple studies spanning different regions and/or time periods. Results from such efforts do not yet capture a large share of the heritability. The origins of this so-called ‘missing heritability’ have been strongly debated. One factor exacerbating the missing heritability is heterogeneity in the effects of genetic variants across studies. The effect of this type of heterogeneity on statistical power to detect associated genetic variants and the accuracy of polygenic predictions is poorly understood. In the current study, we derive the precise effects of heterogeneity in genetic effects across studies on both the statistical power to detect associated genetic variants as well as the accuracy of polygenic predictions. We present an online calculator, available at www.devlaming.eu, which accounts for these effects. By means of this calculator, we show that imperfect genetic correlations between studies substantially decrease statistical power and predictive accuracy and, thereby, contribute to the missing heritability. The MetaGAP calculator helps researchers to gauge how sensitive their results will be to heterogeneity in genetic effects across studies. If strong heterogeneity is expected, random-effects meta-analysis methods should be used instead of fixed-effects methods.
| Large-scale GWAS efforts are rapidly elucidating the genetic architecture of polygenic traits, including anthropometrics [1, 2] and diseases [3–5], as well as behavioral and psychological outcomes [6–8]. These efforts have led to new biological insights, therapeutic targets, and polygenic scores (PGS), and help to understand the complex interplay between genes and environments in shaping individual outcomes [7, 9, 10]. However, GWAS results do not yet account for a large part of the estimated heritability [1, 2, 7, 8]. This dissonance, which is referred to as the ‘missing heritability’, has received broad attention [11–17].
Differences across strata (e.g., studies and populations), in genetic effects, phenotype measurement, and phenotype accuracy, lead to loss of signal [18–20]. Hence, such forms of heterogeneity attenuate the statistical power of a GWAS [17, 18, 21, 22] and the predictive accuracy of a PGS in a hold-out sample [23], and, thereby, contribute to the missing heritability. Since large-scale GWAS results are typically obtained from a meta-analysis of GWAS results from many different studies, we focus on the attenuation resulting from heterogeneity at the level of studies included in such a meta-analysis. Given the importance of discovering trait-affecting variants and obtaining accurate polygenic predictions, it is vital to understand to which extent cross-study heterogeneity attenuates the statistical power and predictive accuracy of GWAS efforts. By considering cross-study differences in genetic effects and heritability, we can quantify this attenuation.
Despite empirical evidence of transethnic genetic heterogeneity in diseases [24] and the fact that cross-study heterogeneity has been found to decrease the chances of a study to yield meaningful results [22, 25], a theoretical multi-study framework that quantifies the effect of cross-study heterogeneity on statistical power and predictive accuracy is still absent. We bridge this gap by developing a Meta-GWAS Accuracy and Power (MetaGAP) calculator (available at www.devlaming.eu) that accounts for the cross-study genetic correlation (CGR). This calculator infers the statistical power to detect associated SNPs and the predictive accuracy of the PGS in a meta-analysis of GWAS results from genetically and phenotypically heterogeneous studies, and quantifies the loss in power and predictive accuracy incurred by this cross-study heterogeneity. Using simulations, we show that the MetaGAP calculator is accurate under a wide range of genetic architectures, even when the assumptions of the calculator are violated.
Although meta-analysis methods accounting for heterogeneity exist [26–31], large-scale GWAS results are typically still obtained from fixed-effects meta-analysis methods [32, 33] such as implemented in METAL [34]. Therefore, the MetaGAP calculator assumes the use of a fixed-effects meta-analysis method. Thus, the calculator will help researchers to assess the merits of an intended fixed-effects meta-analysis of GWAS results and to gauge whether it is more appropriate to apply a meta-analysis method that accounts for heterogeneity.
In an empirical application, we use genomic-relatedness-matrix restricted maximum likelihood (GREML) to estimate the SNP-based heritability (h SNP 2) and CGR of several polygenic traits across three distinct studies: the Rotterdam Study (RS), the Swedish Twin Registry (STR), and the Health and Retirement Study (HRS). For self-rated health, years of education, BMI, and height, we obtain point-estimates of CGR between 0.47 and 0.97. Based on these estimates of h SNP 2 and CGR, we use the MetaGAP calculator to quantify the expected number of hits and predictive accuracy of the PGS in recent GWAS efforts for these traits. Our theoretical predictions align with empirical observations.
For height, under an estimated CGR of 0.97, the expected relative loss in the number of genome-wide significant hits is 8–9%, whereas, for years of education, under an estimated CGR of 0.78, we expect a relative loss of 51–62% in the number of hits. Moreover, we find that the relative loss in PGS R2 is expected to be 6–7% for height and 36–38% for years of education. Hence, our findings show that cross-study heterogeneity attenuates the statistical power and PGS accuracy considerably, thus, contributing substantially to the missing heritability, and, more specifically, to the ‘hiding heritability’ [15–17]—defined as the difference between the SNP-based heritability estimate [35] and the proportion of phenotypic variation explained by genetic variants that reach genome-wide significance in a GWAS.
The MetaGAP calculator is based on theoretical expressions for statistical power and PGS accuracy, derived in S1 Derivations and S2 Derivations. In these expressions, within-study estimates of SNP heritability (e.g., inferred using GCTA [36]) are required input parameters. Estimates of CGR (e.g., inferred as genetic correlations across studies using pairwise bivariate methods as implemented in GCTA [37] and LD-score regression [38, 39], or as genetic-impact correlation from summary statistics [24]) also play a central role in those expressions. As we show in S1 Note, such estimates of CGR are affected by the cross-study overlap in trait-affecting loci as well as the cross-study correlation in the effects of these overlapping loci. In our derivations of statistical power and predictive accuracy, we assume, however, that the set of trait-affecting loci is the same across all studies and that CGRs are, consequently, shaped solely by cross-study correlations in the effects. Using simulation studies, discussed in S1 Simulations, we assess how violations of this assumption affect our results.
In addition, genetic correlations as inferred using GCTA [37] or LD-score regression [39] effectively estimate the cross-trait and/or cross-study correlation in the effects of standardized SNPs. This correlation has been referred to as the genetic-impact correlation [24]. The scale of rare variants is inflated most by standardization (i.e., genotypes are scaled by 1 / 2 f ( 1 - f ), where f denotes the allele frequency of the SNP of interest). Therefore, the scale of the effects of these variants is decreased most by standardization of SNPs (i.e., when standardizing a SNP, the effect is scaled by 2 f ( 1 - f )). Hence, the genetic-impact correlation emphasizes the contribution of common variants [24]. If rare alleles tend to have larger effects than common alleles, as assumed in GCTA [36] and LD-score regression [38], these two opposing forces may cancel each other out; the effects of rare alleles are then bigger, but also scaled downwards more strongly by considering standardized SNPs. Alternatively, one can also consider the correlation in the effect of non-standardized SNPs, referred to as the genetic-effect correlation [24]. This genetic-effect correlation gives rare and common variants equal weight in theory. However, in case rare alleles have larger effects than common alleles, this genetic-effect correlation, in practice, gives a disproportional weight to rare variants.
A clear definition of genetic correlation can be further complicated by the presence of allele frequency differences across samples. Whereas GCTA assumes fixed allele frequencies across the samples included in the analysis [36], there also exist methods which allow for differences in allele frequencies. Ideally, estimates of cross-study genetic-impact correlation accounting for allele frequency differences [24] should be used in the MetaGAP calculator as input for CGR. However, provided the genetic drift is small, whether to account for allele frequency differences across samples or not, will—in all likelihood—hardly affect the CGR estimates. Therefore, under little genetic drift, estimates of CGR obtained by methods ignoring cross-study differences in allele frequencies (e.g., bivariate GREML [37]), suffice as input for the MetaGAP calculator.
In line with other work, we define the effective number of SNPs, S, as the number of haplotype blocks (i.e., independent chromosome segments) [40], where variation in each block is tagged by precisely one genotyped SNP. By genotyped SNPs we also mean imputed SNPs. Hence, in our framework, there are S SNPs contributing to the polygenic score. Due to linkage disequilibrium (LD) this number is likely to be substantially lower than the total number of SNPs in the genome [41], and is inferred to lie between as little as 60,000 [15] and as much as 5 million [41].
In terms of trait-affecting variants, we consider a subset of M SNPs from the set of S SNPs. Each SNP in this subset tags variation in a segment that bears a causal influence on the phenotype. We refer to M as the associated number of SNPs. We assume that the M associated SNPs jointly capture the full SNP-based heritability for the trait of interest and, moreover, that each associated SNP has the same theoretical R2 with respect to the phenotype. In the simulation studies, we also assess the impact of violations of this ‘equal-R2’ assumption.
By considering only independent genotyped SNPs that are assumed to fully tag the causal variants, we can ignore LD among genotyped variants and between the causal variant and the genotyped variants. Thereby, we can greatly reduce the theoretical and numerical complexity of the MetaGAP calculator. However, a genotyped tag SNP does not necessarily capture the full variation of the causal variant present in that independent segment. Nevertheless, the inputs for SNP heritability used in the MetaGAP calculator are within-study GREML estimates of heritability, based on the available SNPs. Therefore, if these genotyped SNPs are in imperfect LD with the causal variants, this will lead to a downward bias in the SNP-based heritability estimates [42]. Hence, the imperfect tagging of the causal variants is likely to be absorbed by a downward bias in the SNP-based heritability estimates.
The theoretical distribution of the Z statistic, resulting from a meta-analysis of GWAS results under imperfect CGRs, can be found in S1 Derivations. These expressions allow for differences in sample size, h SNP 2, and CGR across (pairs of) studies. For intuition, we here present the specific case of a meta-analysis of results from two studies with CGR ρG, with equal SNP-based heritability h SNP 2, and equal sample sizes (i.e., N in Study 1 and N in Study 2). Under this scenario, we find that under high polygenicity, the Z statistic of an associated SNP k is normally distributed with mean zero and the following variance:
Var Z k = E Z k 2 ≈ 1 + h SNP 2 M N 1 + ρ G . (1)
We incorporate cross-study genetic heterogeneity by assuming that the data-generating process follows a random-effects model, where cross-study correlations in SNP effects shape the inferred CGRs. When one has random effects, under the null hypothesis a SNP effect follows a degenerate distribution with all probability mass at zero, whereas under the alternative hypothesis a SNP effect follows a distribution with mean zero and a finite non-zero variance. Bearing in mind that we can write a meta-analysis Z statistic as a weighted average of true effects across studies and noise terms, the null hypothesis leads to a Z statistic with a mean equal to zero and a variance equal to one, whereas the alternative hypothesis does not lead to a non-zero mean in the Z statistic, but rather to excess variation (i.e., a variance larger than one).
The larger the variance in the Z statistic, the higher the probability of rejecting the null. The ratio of h SNP 2 and M can be regarded as the theoretical R2 of each associated SNP with respect to the phenotype. Eq 1 reveals that (i) when sample size increases, power increases, (ii) when h SNP 2 increases, the R2 per associated SNP increases and therefore power increases, (iii) when the number of associated SNPs increases, the R2 per associated SNP decreases and therefore power decreases, (iv) when the CGR is zero the power of the meta-analysis is identical to the power obtained in each of the two studies when analyzed separately, yielding no strict advantage to meta-analyzing, and (v) when the CGR is positive one, the additional variance in the Z statistic—compared to the variance under the null—is twice the additional variance one would have when analyzing the studies separately, yielding a strong advantage to meta-analyzing.
Notably, our expression for E [ Z k 2 ] bears a great resemblance to expressions for the expected value of the squared Z statistic when accounting for LD, population stratification, and polygenicity [38, 43, 44]. Consider the scenario where the CGR between two samples of equal size is positive one. Based of Eq 1, we then have that E [ Z k 2 ] ≈ 1 + h SNP 2 M N T for a trait-affecting haplotype block, where NT = 2N denotes the total sample size. This expression is equivalent to the expected squared Z statistic from the linear regression analysis for a trait-affecting variant reported in Section 4.2 of the Supplementary Note to [44] as well as the first equation in [38] when assuming that confounding biases and LD are absent.
In order to compute statistical power in a multi-study setting, we first use the generic expression for the variance of the GWAS Z statistic derived in S1 Derivations to characterize the distribution of the Z statistic under the alternative hypothesis. Given a genome-wide significance threshold (denoted by α; usually α = 5 ⋅ 10−8), we use the normal cumulative distribution function under the alternative hypothesis to quantify the probability of attaining genome-wide significance for an associated SNP. This probability we refer to as the ‘power per associated SNP’ (denoted here by β). Given that we use SNPs tagging independent haplotype blocks, we can calculate the probability of rejecting the null for at least one SNP and the expected number of hits, true positives, false positives, false negatives, and positive negatives, as functions of α, β, the number of truly associated SNPs (denoted by M), and the number of non-associated SNPs (denoted by S − M). Letting ‘#’ denote the number of elements in a set, we have that
ℙ [ # true positives ≥ 1 ]=1−(1−β)M,ℙ [ # hits ≥ 1 ]=1−[ (1−β)M(1−α)S−M ],E [ # hits ]=βM+α(S−M),E [ # true positives ]=βM,E [ # false positives ]=α(S−M),E [ # false negatives ]=(1−β)M, andE [ # true negatives ]=(1−α)(S−M).
In S2 Derivations we derive a generic expression for the theoretical R2 of a PGS in a hold-out sample, with SNP weights based on a meta-analysis of GWAS results under imperfect CGRs. We consider a PGS that includes all the SNPs that tag independent haplotype blocks (i.e., there is no SNP selection).
For intuition, we here present an approximation for prediction in a hold-out sample, with SNP weights based on a GWAS in a single discovery study with sample size N, where both studies have SNP heritability h SNP 2, and with CGR ρG, between the studies. Under high polygenicity, the R2 of the PGS in the hold-out sample is then given by the following expression:
R 2 ≈ h SNP 2 ρ G 2 h SNP 2 S N + h SNP 2 . (2)
In case the CGR is one, and we consider the R2 between the PGS and the genetic value (i.e., the genetic component of the phenotype) instead of the phenotype itself, the first two terms in Eq 2 disappear, yielding an expression equivalent to the first equation in [40]. Assuming a CGR of one and that all SNPs are associated, Eq 2 is equivalent to the expression in [23] for the R2 between the PGS and the phenotype in the hold-out sample.
From Eq 2, we deduce that (i) as the effective number of SNPs S increases, the R2 of the PGS deteriorates (since every SNP-effect estimate contains noise, owing to imperfect inferences in finite samples), (ii) given the effective number of SNPs, under a polygenic architecture, the precise fraction of effective SNPs that is associated does not affect the R2, (iii) R2 is quadratically proportional to ρG, implying a strong sensitivity to CGR, and (iv) as the sample size of the discovery study grows, the upper limit of the R2 is given by h SNP 2 ρ G 2, implying that the full SNP heritability in the hold-out sample cannot be entirely captured as long as CGR is imperfect.
An online version of the MetaGAP calculator can be found at www.devlaming.eu. This calculator computes the theoretical power per trait-affecting haplotype block, the power to detect at least one of these blocks, and the expected number of (a) independent hits, (b) true positives, (c) false positives, (d) false negatives, and (e) true negatives, for a meta-analysis of GWAS results from C studies. In addition, it provides the expected R2 of a PGS for a hold-out sample, including all GWAS SNPs, with SNP weights based on the meta-analysis of the GWAS results from C studies. Calculations are based on the generic expressions for GWAS power derived in S1 Derivations and PGS R2 derived in S2 Derivations.
The calculator assumes a quantitative trait. Users need to specify either the average sample size per study or the sample size of each study separately. In addition, users need to specify either the average within-study SNP heritability or the SNP heritability per study. The SNP heritability in the hold-out sample also needs to be provided. Users are required to enter the effective number of causal SNPs and the effective number of SNPs in total. The calculator assumes a fixed CGR between all pairs of studies included in the meta-analysis and a fixed CGR between the hold-out sample and each study in the meta-analysis. Hence, one needs to specify two CGR values: one for the CGR within the set of meta-analysis studies and one to specify the genetic overlap between the hold-out sample and the meta-analysis studies.
Finally, a more general version of the MetaGAP calculator is provided in the form of MATLAB code (www.mathworks.com), also available at www.devlaming.eu. This code can be used in case one desires to specify a more versatile genetic-correlation matrix, where the CGR can differ between all pairs of studies. Therefore, this implementation requires the user to specify a full (C+1)-by-(C+1) correlation matrix. Calculations in this code are also fully in line with the generic expressions in S1 Derivations and S2 Derivations.
We simulate data for a wide range of genetic architectures in order to assess the validity of our theoretical framework. As we show in S1 Simulations, the theoretical expressions we derive for power and R2 are accurate, even for data generating processes substantially different from the process we assume in our derivations. Our strongest assumptions are that all truly associated SNPs have equal R2 with respect to the phenotype, regardless of allele frequency, and that genome-wide CGRs are shaped solely by the cross-study correlations in the effects of causal SNPs. When we simulate data where the former assumption fails and where—in addition—allele frequencies are non-uniformly distributed and different across studies, the root-mean-square prediction error of statistical power lies below 3% and that of PGS R2 below 2%. Moreover, when we simulate data where the CGR is shaped by both non-overlapping causal loci across studies and the correlation of the effects of the overlapping loci, the RMSE is less than 2% for both statistical power and PGS R2.
Using 1000-Genomes imputed data from the RS, STR, and HRS, we estimate SNP-based heritability and CGR respectively by means of univariate and bivariate GREML [36, 37] as implemented in GCTA [36]. In our analyses we consider the subset of HapMap3 SNPs available in the 1000-Genomes imputed data. In S1 Data we report details on the genotype and phenotype data, as well as our quality control (QC) procedure. After QC we have a dataset, consisting of ≈ 1 million SNPs and ≈ 20,000 individuals, from which we infer h SNP 2 and CGR. In S1 Estimation we provide details on the specifications of the models used for GREML estimation.
Written informed consent was provided by all participants and the research project was approved by the Ethics Committee of Erasmus Medical Center (MEC 02.1015), the Ethics Committee of Stockholm (2007-644-31, 2011-463-32, 2012/270-31/2), the ERIM Institutional Review Board (2014-04), and dbGaP (#3544, #5752, #5082, #5285).
Using the MetaGAP calculator, we assessed the theoretical power of a meta-analysis of GWAS results from genetically heterogeneous studies and the theoretical R2 of the resulting PGS in a hold-out sample, for various numbers of studies and sample sizes, and different values of CGR and h SNP 2.
In Table 1 we report univariate GREML estimates of SNP heritability and bivariate GREML estimates of genetic correlation for traits that attained a pooled sample size of at least 18,000 individuals, which gave us at least 50% power to detect a genetic correlation near one for a trait that has a SNP heritability of 10% or more [45]. The smallest total sample size is NT = 19,184 for self-rated health. Details per phenotype on sample size, univariate estimates of SNP heritability, and bivariate estimates of genetic correlation, stratified across studies, and cross-study averages, are provided in S1 Table. Results stratified across sexes are listed in S2 Table.
The univariate estimates of SNP heritability based on the pooled data assume perfect CGRs. Therefore, such estimates of SNP heritability are downwards biased when based on data from multiple studies with imperfect CGRs. To circumvent this bias, we estimated SNP heritability in each study separately, and focused on the sample-size-weighted cross-study average estimate of SNP heritability.
For both height and BMI, we observed genetic correlations close to one across pairs of studies and between females and males. For years of schooling (EduYears) we found a CGR around 0.8 when averaged across pairs of studies. Similarly, the genetic correlation for EduYears in females and males lies around 0.8. The CGR of self-rated health is substantially below one across the pairs of studies, whilst the genetic correlation between females and males seems to lie around one. The reason for this difference in the genetic correlation of self-rated health between pairs of studies and between females and males may be due to the difference in the questionnaire across studies, discussed in S1 Data. The questionnaire differences can yield a low CGR, while not precluding the remaining genetic overlap for this measure across the three studies, to be highly similar for females and males. For CurrCigt and CurrDrinkFreq, the estimates of CGR and of genetic correlation between females and males are non-informative. For these two traits the standard errors of the genetic correlations estimates are large, mostly greater than 0.5. In addition, for CurrDrinkFreq there is strong volatility in the CGR estimate across pairs of studies.
Considering only the traits for which we obtained accurate estimates of CGR and SNP heritability (i.e., with low standard errors), we used the MetaGAP calculator to predict the number of hits in a set of discovery samples and the PGS R2 in a hold-out sample, in prominent GWAS efforts for these traits. Details and notes on the results from existing studies, used as input for the MetaGAP calculations, can be found S3 Table. Importantly, as reported in S4 Table, for the traits under consideration here, large-scale GWAS results to date have been obtained using fixed-effects meta-analyses.
Since we only had accurate estimates for height, BMI, EduYears, and self-rated health, we focused on these four phenotypes. For these traits, we computed sample-size-weighted average CGR estimates across the pairs of studies. Table 2 shows the number of hits and PGS R2 reported in the most comprehensive GWAS efforts to date for the traits of interest, together with predictions from the MetaGAP calculator. We tried several values for the number of independent haplotype blocks (i.e., 100k, 150k, 200k, 250k) and for the number of trait-associated blocks (i.e., 10k, 15k, 20k, 25k). Overall, 250k blocks of which 20k trait-affecting yielded theoretical predictions in best agreement with the empirical observations; we acknowledge the potential for some overfitting (i.e., two free parameters set on the basis of 17 data points; 10 data points for the reported number of hits and 7 for PGS R2).
For height—the trait with the lowest standard error in the estimates of h SNP 2 and CGR—the predictions of the number of hits and PGS R2 for the two largest GWAS efforts are much in line with theoretical predictions. For the smaller GWAS of 13,665 individuals [47], our estimates seem slightly conservative; 0 hits expected versus the 7 reported. However, in our framework, we assumed that each causal SNP has the same R2. Provided there are some differences in R2 between causal SNPs, the first SNPs that are likely to reach genome-wide significance in relatively small samples, are the ones with a comparatively large R2. This view is supported by the fact that a PGS based on merely 20 SNPs already explains 2.9% of the variation in height. Hence, for relatively small samples our theoretical predictions of power and R2 may be somewhat conservative. In addition, the 10k SNPs with the lowest meta-analysis p-values can explain about 60% of the SNP heritability [1]. If the SNPs tagging the remaining 40% each have similar predictive power as the SNPs tagging the first 60%, then the number of SNPs needed to capture the full h SNP 2 would lie around 10k/0.6 = 17k, which is somewhat lower than the 20k which yields the most accurate theoretical predictions. However, as indicated before, the SNPs which appear most prominent in a GWAS are likely to be the ones with a greater than average predictive power. Therefore, the remaining 40% of h SNP 2 is likely to be stemming for SNPs with somewhat lower predictive power. Hence, 20k associated independent SNPs is not an unreasonable number for height.
The notion of a GWAS first picking up the SNPs with a relatively high R2 is also supported by the predicted and observed number of hits for the reported self-rated-health GWAS [51]; given a SNP heritability estimate between 10% [51] and 16% (Table 2), according to our theoretical predictions, a GWAS in a sample of around 110k individuals is unlikely to yield even a single genome-wide significant hit. Nevertheless, this GWAS has yielded 13 independent hits. This finding supports the idea that for various traits, some SNPs with a relatively high R2 are present. However, there is uncertainty in the number of truly associated loci. More accurate estimates of this number may improve the accuracy of our theoretical predictions.
For BMI our predictions of PGS R2 were quite in line with empirical results. However, for the number of hits, our predictions for the largest efforts seemed overly optimistic. We therefore suspect that the number of independent SNPs associated with BMI is higher than 20k; a higher number of associated SNPs would reduce the GWAS power, while preserving PGS R2, yielding good agreement with empirical observation. Nevertheless, given the limited number of data points, this strategy of setting the number of causal SNPs would increase the chance of overfitting.
For EduYears we observed that the reported number of hits is in between the expected number of hits when the CGR is set to the averaged GREML estimate of 0.783 and when the CGR is set to one. Given the standard errors in the CGR estimates for EduYears, the CGR might very well be somewhat greater than 0.783, which would yield a good fit with the reported number of hits. However, as with the number of truly associated SNPs for BMI, in light of the risk of overfitting, we can make no strong claims about a slightly higher CGR of EduYears.
Overall, our theoretical predictions of the number of hits and PGS R2 are in moderate agreement with empirical observations, especially when bearing in mind that we are looking at a limited number of data points, making chance perturbations from expectation likely. In addition, regarding the number of hits, the listed studies are not identical in terms of the procedure to obtain the independent hits. Therefore, the numbers could have been slightly different, had the same pruning procedure been used across all reported studies.
Regarding attenuation, we observed a substantial spread in the predicted number of hits and PGS R2 when assuming either a CGR equal to one, or a CGR in accordance with empirical estimates, with traits with lower CGR suffering from stronger attenuation in power and predictive accuracy. In line with theory, R2 falls approximately quadratically with CGR. For instance, for self-rated health, the estimated CGR of about 0.5, would yield a PGS that retains approximately 0.52 = 25% of the R2 it would have had under a CGR of one. Hence the approximated attenuation is 75%. This approximation is corroborated by the theoretical relative attenuation of 78%.
Given our CGR estimates, the theoretical relative loss in PGS R2 is 6% for height, 14% for BMI, 36% for EduYears, and 78% for self-rated health, when compared to the R2 of PGSs under perfect CGRs (Table 2). These losses in R2 are unlikely to be reduced by larger sample sizes and denser genotyping.
Somewhat contrary to expectation, the number of hits seems to respond even more strongly to CGR than PGS R2. However, since in each study under consideration the average power per associated SNP is quite small, a small decrease in power per SNP in absolute terms can constitute a substantial decrease in relative terms. For instance, when one has 2% power per truly associated SNP, an absolute decrease of 1%—leaving 1% power—constitutes a relative decrease of 50% of power per causal SNP, and thereby a 50% decrease in the expected number of hits. This strong response shows, for example, in the case of EduYears, where the expected number of hits drop by about 37% when going from a CGR of one down to a CGR of 0.783.
We have shown that imperfect cross-study genetic correlations (CGRs) are likely to contribute to the gap between the phenotypic variation accounted for by all SNPs jointly and by the leading GWAS efforts to date. We arrived at this conclusion in five steps. First, we developed a Meta-GWAS Accuracy and Power (MetaGAP) calculator that accounts for the CGR. This online calculator relates the statistical power to detect associated SNPs and the R2 of the polygenic score (PGS) in a hold-out sample to the number of studies, sample size and SNP heritability per study, and the CGR. The underlying theory shows that there is a quadratic response of the PGS R2 to CGR. Moreover, we showed that the power per associated SNP is also affected by CGR.
Second, we used simulations to demonstrate that our theory is robust to several violations of the assumptions about the underlying data-generating process, regarding the relation between allele frequency and effect size, the distribution of allele frequencies, and the factors contributing to CGR. Further research needs to assess whether our theoretical predictions are also accurate under an even broader set of scenarios (e.g., when studying a binary trait).
Third, we used a sample of unrelated individuals from the Rotterdam Study, the Swedish Twin Registry, and the Health and Retirement Study, to estimate SNP-based heritability as well as the CGR for traits such as height and BMI. Although our CGR estimates have considerable standard errors, the estimates make it likely that for many polygenic traits the CGR is positive, albeit smaller than one.
Fourth, based on these empirical estimates of SNP heritability and CGR for height, BMI, years of education, and self-rated health, we used the MetaGAP calculator to predict the number of expected hits and the expected PGS R2 for the most prominent studies to date for these traits. We found that our predictions are in moderate agreement with empirical observations. Our theory seems slightly conservative for smaller GWAS samples. For large-scale GWAS efforts our predictions were in line with the outcomes of these efforts. More accurate estimates of the number of truly associated loci may further improve the accuracy of our theoretical predictions.
Fifth, we used our theoretical model to assess statistical power and predictive accuracy for these GWAS efforts, had the CGR been equal to one for the traits under consideration. Our estimates of power and predictive accuracy in this scenario indicated a strong decrease in the PGS R2 and the expected number of hits, due to imperfect CGRs. Though these observations are in line with expectation for predictive accuracy, for statistical power the effect was larger than we anticipated. This finding can be explained, however, by the fact that though the absolute decrease in power per SNP is small, the relative decrease is large, since the statistical power per associated SNP is often low to begin with.
Overall, our study affirms that although PGS accuracy improves substantially with further increasing sample sizes, in the end PGS R2 will continue to fall short of the full SNP-based heritability. Hence, this study contributes to the understanding of the hiding heritability reported in the GWAS literature.
Regarding the etiology of imperfect CGRs, the likely reasons are heterogeneous phenotype measures across studies, gene–environment interactions with underlying environmental factors differing across studies, and gene–gene interactions where the average effects differ across studies due to differences in allele frequencies. Our study is not able to disentangle these different causes; by estimating the CGR for different traits we merely quantify the joint effect these three candidates have on the respective traits.
However, in certain situations it may be possible to disentangle the etiology of imperfect CGRs to some extent. For instance, in case one considers a specific phenotype that is usually studied by means of a commonly available but relatively heterogeneous and/or noisy measure, while there also exists a less readily available but more accurate and homogeneous measure. If one has access to both these measures in several studies, one can compare the CGR estimates for the more accurate measure and the CGR estimates for the less accurate but more commonly available measure. Such a comparison would help to disentangle the contribution of phenotypic heterogeneity and genetic heterogeneity to the CGR of the more commonly available measure.
In considering how to properly address imperfect CGRs, it is important to note that having a small set of large studies, rather than a large set of small studies, does not necessarily abate the problem of imperfect genetic correlations. Despite the fact that having fewer studies can help to reduce the effects of heterogeneous phenotype measures, larger studies are more likely to sample individuals from different environments. If gene–environment interactions do play a role, strong differences in environment between subsets of individuals in a study can lead to imperfect genetic correlations within that study. The attenuation in power and accuracy resulting from such within-study heterogeneity may be harder to address than cross-study heterogeneity.
Our findings stress the importance of considering the use of more sophisticated meta-analysis methods that account for cross-study heterogeneity [26–31]. We believe that the online MetaGAP calculator will prove to be an important tool for assessing whether an intended fixed-effects meta-analysis of GWAS results from different studies is likely to yield meaningful outcomes.
|
10.1371/journal.ppat.1004893 | Circumventing Y. pestis Virulence by Early Recruitment of Neutrophils to the Lungs during Pneumonic Plague | Pneumonic plague is a fatal disease caused by Yersinia pestis that is associated with a delayed immune response in the lungs. Because neutrophils are the first immune cells recruited to sites of infection, we investigated the mechanisms responsible for their delayed homing to the lung. During the first 24 hr after pulmonary infection with a fully virulent Y. pestis strain, no significant changes were observed in the lungs in the levels of neutrophils infiltrate, expression of adhesion molecules, or the expression of the major neutrophil chemoattractants keratinocyte cell-derived chemokine (KC), macrophage inflammatory protein 2 (MIP-2) and granulocyte colony stimulating factor (G-CSF). In contrast, early induction of chemokines, rapid neutrophil infiltration and a reduced bacterial burden were observed in the lungs of mice infected with an avirulent Y. pestis strain. In vitro infection of lung-derived cell-lines with a YopJ mutant revealed the involvement of YopJ in the inhibition of chemoattractants expression. However, the recruitment of neutrophils to the lungs of mice infected with the mutant was still delayed and associated with rapid bacterial propagation and mortality. Interestingly, whereas KC, MIP-2 and G-CSF mRNA levels in the lungs were up-regulated early after infection with the mutant, their protein levels remained constant, suggesting that Y. pestis may employ additional mechanisms to suppress early chemoattractants induction in the lung. It therefore seems that prevention of the early influx of neutrophils to the lungs is of major importance for Y. pestis virulence. Indeed, pulmonary instillation of KC and MIP-2 to G-CSF-treated mice infected with Y. pestis led to rapid homing of neutrophils to the lung followed by a reduction in bacterial counts at 24 hr post-infection and improved survival rates. These observations shed new light on the virulence mechanisms of Y. pestis during pneumonic plague, and have implications for the development of novel therapies against this pathogen.
| The pathogen Yersinia pestis is the causative agent of pneumonic plague, as well as a potential bioweapon. The nature of this disease involves an initial non-inflammatory phase where the influx of neutrophils to the lungs is suppressed, allowing bacterial propagation in this organ. Using the mouse model of pneumonic plague, we demonstrate that the early expression of neutrophil chemoattractants and adhesion molecules in the lungs is delayed concomitant with a delayed recruitment of neutrophils to the lung. We also show that the Y. pestis virulence factor YopJ is involved in the early suppression of chemoattractants mRNA expression in the lung early after infection, but it seems that additional Y. pestis factors interfere with the protein synthesis of these chemoattractants. Indeed, administration of recombinant KC and MIP-2 to the infected lung of G-CSF treated mice restored the early neutrophil influx to the lungs, leading to a significant reduction in bacterial burden. The treatment has also proved efficacious in reducing mortality. This study highlights the complex virulence mechanisms employed by Y. pestis to diminish the early homing of neutrophils to the lungs thereby allowing bacterial propagation and disease progression.
| The recruitment of neutrophils is a fundamental component of the initial phase of the innate immune response to bacterial lung infections, as demonstrated by the selective depletion of neutrophils and the consequences on pathogen clearance from the lungs [1]. In response to infection, neutrophils are mobilized from the bone marrow (BM), resulting in a rise in circulating neutrophils in the blood within a few hours after infection [2, 3]. The robust expression of G-CSF modulates the production of neutrophils to meet the increased need of the host during infection [4]. Circulating neutrophils migrate to the infection site along a chemotactic gradient of potent chemoattractants, such as KC (CXCL1 or IL-8 in humans) and MIP-2 (CXCL2), produced at the infection site [5, 6]. To allow circulating neutrophils to cross the vascular wall and arrive at the site of infection, multiple adhesion molecules are induced on endothelial cells adjacent to the inflamed tissue. E- and P-selectins are known to be involved in the initial attachment of neutrophils to the endothelium as well as their rolling behavior. Intracellular adhesion molecule 1 (ICAM-1) and vascular cell adhesion molecule 1 (VCAM-1) mediate the subsequent step of tight adhesion to the endothelium, allowing neutrophils to transmigrate to the site of inflammation [7]. After migration, neutrophils phagocytose and digest the invading pathogen and produce pro-inflammatory cytokines [8], thereby serving a beneficial role for the host. However, their excessive and uncontrolled activity may also cause severe damage to the host [9, 10]. The important role of neutrophils in protecting the host against infection with respiratory pathogens has been investigated primarily with regard to pathogens such as Pseudomonas aeruginosa [11], Legionella pneumophila [12], Klebsiella pneumonia [13] and Yersinia pestis [14, 15].
Y. pestis gained notoriety as the causative agent of plague [16]. Inhalation of Y. pestis droplets or aerosols leads to the development of primary pneumonic plague, which is a rapidly progressing fatal disease with the capability of spreading from person to person [17, 18]. These characteristics also led to the recognition of Y. pestis as a potential biological threat agent [19].
Recent in vivo studies in animal models of pneumonic plague have revealed the biphasic nature of the progression of this disease [20–22]. The observed initial delay in the recruitment of immune cells and neutrophils in particular to the lungs of Y. pestis-infected mice is correlated with the limited up-regulation of multiple inflammatory cytokines and chemokines. Additionally, pulmonary infection with Y. pestis creates a permissive environment for the proliferation of other avirulent bacterial species [23].
Bacterial pathogens have developed a variety of mechanisms to inhibit immune cell functions as a means to disarm the host defense. For example, Y. pestis utilizes the pCD1-encoded type III secretion system (TTSS) composed of a secretory apparatus, chaperones and several translocated effectors (Yops) to disable the early innate immune response [24–26] and the activity of neutrophils in particular [27, 28]. Recently, neutrophils were found to be an important cellular target of Y. pestis Yop secretion during the early stage of pneumonic plague [29].
While the lack of an adequate early immune response in the lungs during pneumonic plague is well described, the cascade of neutrophil recruitment from the circulation into the lungs during pneumonic plague and the identification of Y. pestis virulence factors involved in suppressing this process have yet to be fully elucidated.
We previously reported that an early immune response is initiated by bone-marrow (BM) cells after airway infection of mice with a fully virulent Y. pestis strain, causing rapid mobilization of neutrophils from the BM to the blood circulation [30]. In the present study, we analyzed the interference of Y. pestis with the recruitment of neutrophils from the circulation to the lungs of infected mice. Our observations indicate that in the lungs of infected mice, the induction of the major neutrophil chemoattractants KC, MIP-2 and G-CSF as well as the leukocytes adhesion molecules E-selectin, P-selectin, ICAM-1 and VCAM-1 is delayed. In addition, we describe the role of YopJ in preventing the induction of the chemoattractants mRNA at the early stage of disease progression. Finally, we demonstrate that early attraction of neutrophils to the infected lung by intranasal installation of exogenous chemoattractants improves bacterial clearance as well as survival rate.
Studies in animal models of pneumonic plague have revealed the biphasic nature of the progression of this disease. The early phase of the disease, which takes place during the first 24–36 hours post infection (hpi), involves a limited pro-inflammatory response in the lung, whereas the later phase of disease progression (48–72 hpi) is associated with an excessive pro-inflammatory response [20–22].
To further characterize the early innate immune response during pneumonic plague, C57BL/6 mice were exposed i.n. to a lethal dose of 1x105 cfu (100 LD50) of the highly virulent Y. pestis strain Kim53, which typically kills mice within 72 hpi (S1 Fig). As neutrophils are one of the first innate immune cells recruited to the site of infection, we measured the levels of neutrophils infiltrating into the lungs at the early time point of 24 hpi. No significant change was observed in the absolute number or percentage of neutrophils at this time point in comparison to naïve mice (Fig 1A and 1B). Because the metalloproteinases MMP8 and MMP9 are produced and released by neutrophils while combating invading pathogens [31], we measured the levels of MMPs in lung extracts at the early time point of 24 hpi. Consistent with the limited infiltration of neutrophils into the lungs, the expression of MMP8 (Fig 1C and 1D) and MMP9 (Fig 1E and 1F) in the lungs of infected mice did not change in comparison to naïve mice. These findings indicate that the early pulmonary innate response, associated with neutrophils homing to the lung, is impaired after infection with the virulent Y. pestis strain. Moreover, the increased number of live Y. pestis bacilli detected in the lungs at 24–48 hpi (Fig 1G) suggests that the delayed influx of neutrophils into the lungs allowed the pathogen to rapidly proliferate and overwhelm the innate immune system. As the disease progressed into the excessive pro-inflammatory state at 48 hpi, massive infiltration of neutrophils into the lungs was observed (Fig 1A and 1B) together with a dramatic increase in the expression of both MMP8 and MMP9 (Fig 1C–1F). Evidently, this late intensive pulmonary immune response was unable to prevent the pathogen from propagating to high levels in the lung tissue (Fig 1G).
This phenomenon was in contrast to the kinetics of neutrophil infiltration after infection with an equivalent infective dose of the avirulent Y. pestis strain Kim53Δ70Δ10, which lacks the pCD1 and pPCP1 plasmids that carry essential virulence factors including the TTSS and Pla protease (Fig 1A and 1B). The rapid elevation in neutrophil counts in the lungs of mice infected with the avirulent Y. pestis strain was accompanied by a significant up-regulation of the levels of MMP8 and MMP9 expression (Fig 1C–1F) and by a significant decrease in bacterial loads in the lungs at 24 hpi (Fig 1G). Notably, this early recruitment of neutrophils to the lungs was transient, as depicted by the return of neutrophil numbers in the lungs to their basal level by 48 hpi (Fig 1A and 1B). As a result, the levels of MMP8 and MMP9 expression were also reduced (Fig 1C–1F).
We previously demonstrated that i.n. infection of mice with the virulent Y. pestis strain is sensed by the BM compartment early after infection, resulting in the subsequent release of neutrophils to the blood by 12–24 hpi [30]. The observed delay in the homing of neutrophils from the circulation to the lungs motivated us to study the impairment of this pathway during the progression of pneumonic plague.
Because the chemoattractants KC, MIP-2 and G-CSF are of central importance for the recruitment of neutrophils to infected organs [32], their levels in the blood of mice infected with the virulent Y. pestis strain were measured by ELISA. As shown, the levels of KC, MIP-2 and G-CSF at the early stage of 24 hpi were comparable to those observed in naïve mice, and these levels increased significantly only during the late stage of infection at 48 hpi (Fig 2A). This result suggests that although neutrophils are released from the BM to the blood early after airway infection, their ability to navigate towards the infected lungs is impaired. To better understand the limited ability of infected lungs to induce neutrophil chemotaxis and infiltration, we performed a Transwell-migration assay of naïve BM-derived neutrophils towards lung supernatants obtained from mice at several time points after infection with Y. pestis Kim53. Only lung supernatants obtained from mice at late stages of disease progression e.g., 48 hpi, demonstrated the potential to induce in vitro Transwell-migration of naïve neutrophils (Fig 2B), implying that at this time point, the lungs are enriched with chemotactic factors that facilitate neutrophil migration.
Next, we measured the levels of mRNA and protein expression of KC, MIP-2 and G-CSF in lung extracts and BALF at 12, 24 and 48 hpi with Kim53. The levels of mRNA (red line) and protein (blue line) of KC, MIP-2 and G-CSF were significantly increased in the lungs only at 48 hpi (Fig 2C–2E). These observations are in line with previous reports describing the delayed pulmonary pro-inflammatory response to Y. pestis infection using various animal models of pneumonic plague [20–22, 33].
In contrast, infection with the avirulent Y. pestis strain Kim53Δ70Δ10 was characterized by early and moderate induction of the mRNA levels of KC, MIP-2 and G-CSF in the lungs at 12 hpi, accompanied by elevated protein levels in the BALF (Fig 2F and 2H).
The recruitment of neutrophils to infected tissue is a complex process dependent on orchestrated and tightly regulated communication between neutrophils and endothelial cells, resulting in a multistep adhesion cascade. This process includes the initial attachment of the neutrophils to the endothelium, rolling along the endothelial surface and arrest at the final destination to allow complete transmigration [34]. This sequence of coordinated and transient interactions relies on the synchronized expression of several adhesion molecules by endothelial cells adjacent to the site of inflammation. Hence, we decided to examine the expression levels of four molecules that participate in two different stages of neutrophil transmigration: E-selectin, P-selectin, ICAM-1 and VCAM-1 [7]. Quantitative PCR analysis of lung mRNA obtained from mice infected with the virulent Kim53 strain revealed a delay in the up-regulation of the expression of all four adhesion molecules during the first 24 hpi (Fig 3A). Again, this delayed expression was in contrast to the early induction of adhesion molecule mRNA in the lungs of mice infected with the avirulent strain Kim53Δ70Δ10 (Fig 3B). Together, these data suggest that the rapid propagation of Y. pestis in the lungs during the early stages of pneumonic plague results from an impaired innate immune response associated with delayed neutrophil infiltration to the lung, presumably due to a combined delay in the expression of chemotactic signals and adhesion molecules by lung resident cells.
The virulence characteristics of Y. pestis are mostly attributed to the pCD1 plasmid that encodes the TTSS and its six effectors proteins, termed Yops. During interactions with host target cells, these proteins are transported into the cytosol of the host cell via a needle-like apparatus. Together, the translocated Yops target the phagocytic machinery and deregulate signaling pathways, resulting in a reduced immune response by the host [35–37]. As described, infection of mice with the avirulent strain Kim53Δ70Δ10 that does not express the entire TTSS and the Pla protease, was characterized by the early induction of neutrophil chemoattractant mRNA and protein in the lungs (Fig 2F and 2H), and by the early recruitment of neutrophils to the lungs (Fig 1A and 1B). We suspected that one of the Yop effectors may be involved in the early suppression of the up-regulation of chemoattractants expression in the lungs after exposure to the virulent Y. pestis strain. To decipher which Yop was responsible for the early inhibition of KC, MIP-2 and G-CSF up-regulation, we used two different Yop-null derivatives of the fully virulent Y. pestis strain Kim53, and we performed a series of in vitro infection experiments using alveolar-derived macrophages (MH-S) and lung epithelial (TC-1) cell lines. We used the avirulent Y. pestis strain Kim53Δ70Δ10 and the wild-type Y. pestis strains as controls in these experiments. As shown, infection with 50 MOI of the Kim53ΔYopJ strain, but not with the Kim53 derivative lacking YopH, resulted in increased levels of KC and MIP-2 mRNA and protein in MH-S cells (Fig 4A and 4B) and of MIP-2 and G-CSF mRNA and protein in TC-1 cells (Fig 4C and 4D). The induction of these chemokines production by the cell lines following infection with the YopJ mutant was similar to infection with the avirulent strain Kim53Δ70Δ10. Additionally, infection of both cell lines with a YopJ mutant of the Y. pestis EV76 vaccine strain and with an EV76 derivative lacking the pCD1 plasmid led to induction of chemoattractants mRNA, whereas infection with EV76 and other Yop-null mutants including EV76ΔYopE, EV76ΔYopK and EV76ΔYopH did not (S2 Fig). These results point to the involvement of YopJ in the regulation of KC, MIP-2 and G-CSF expression by lung-derived cells in vitro.
Because YopJ activity was associated with suppression of KC, MIP-2 and G-CSF expression in Y. pestis-infected alveolar macrophages and epithelial cells, we infected C57BL/6 mice i.n. with a dose of 1x105 cfu of Kim53ΔYopJ and monitored the disease progression in the lung. All infected mice succumbed within 4 days of infection (S1 Fig). Bacterial counts in the lungs were elevated to 1x109 cfu by 48 hpi (Fig 5A), and no significant change was observed in neutrophil numbers or percentage in the lungs after the first 24 hpi (Fig 5B and 5C). Massive infiltration of neutrophils to the lungs was apparent only at 48 hpi (Fig 5B and 5C), and the delayed kinetics of neutrophil influx to the lungs following i.n. infection with Kim53ΔYopJ and bacterial propagation resembled the kinetic responses observed following infection with the wild-type Kim53 strain (Fig 1). Taking into account the observed involvement of YopJ in preventing KC, MIP-2 and G-CSF up-regulation by lung resident cells in vitro (Fig 4), we further evaluated the changes in mRNA and protein levels of these chemokines in lung extracts and BALF of Kim53ΔYopJ-infected mice. Surprisingly, we observed that while the mRNA levels of these chemokines were significantly elevated at 12–24 hpi, their protein levels remained constant and relatively low during this time frame. A significant increase in the protein levels of G-CSF, KC and MIP-2 was detected only at 48 hpi with Kim53ΔYopJ (Fig 5D and 5F). Notably, significantly higher levels of G-CSF, KC and MIP-2 mRNA were measured at 48 hpi in the lungs of mice infected with the YopJ mutant compared to those of mice infected with the wild-type Y. pestis strain (Fig 5G). This difference may result from earlier induction of chemoattractants mRNA following infection with the YopJ mutant, leading to the accumulation of higher levels of chemoattractants mRNA at later stages of disease progression. The data indicate that YopJ mediates the delayed up-regulation of chemoattractant mRNA expression in the lungs of Y. pestis-infected mice. However, unlike in the in vitro infection system we utilized, in the absence of YopJ, the early elevation of chemoattractant mRNA levels did not yield a subsequent increase in the protein levels. These observations suggests that additional virulence mechanisms involving host-pathogen interactions play a role in modulating the early expression of chemoattractants in a complex multicellular organ such as the lungs, thereby preventing the early recruitment of neutrophils to the infected lung.
The early recruitment of neutrophils to the lungs appears to be of central importance for the defense against pneumonic plague. Due to the fact that some key players in this process (e.g. KC, MIP-2 and G-CSF) are targeted by the bacterium early after the infection, we tested the potential of chemokine therapy for early neutrophil recruitment to the lung. The treatment included subcutaneous administration of G-CSF to synchronize and overload the circulation with newly formed neutrophils, combined with i.n. instillation of KC and MIP-2 to guide the neutrophils and stimulate their recruitment and homing to the infected lungs (Fig 6A).
We first examined the potential of this treatment regimen to promote neutrophil recruitment to the lungs of naïve mice. As depicted in Fig 6B, treatment with G-CSF alone for 5 consecutive days significantly increased the numbers of neutrophils in the blood but not in the lung, whereas the combined treatment (GKM), which included an additional intranasal administration of KC and MIP-2 (1 μg/mouse each), led to a significant accumulation of neutrophils in the lungs (Fig 6B). This treatment was not associated with deleterious effects on animal morbidity or mortality (S1 Fig). Next, we assessed the ability of GKM treatment to induce the early recruitment of neutrophils to the lungs of mice infected i.n. with the virulent Y. pestis strain.
Similar to naïve mice, the percentage of neutrophils measured at 24 hpi in the blood of Y. pestis-infected mice treated for 5 consecutive days with G-CSF alone (starting 3 days before the infection) was elevated in comparison to the percentage in control-treated mice, whereas the percentage of lung neutrophils did not change (Fig 6C, G-CSF). In contrast, a rapid increase in the percentage and total number of neutrophils was detected at 24 hpi in the lungs of GKM-treated mice that received KC and MIP-2 at 6 hpi (Fig 6C and 6D, GKM). Injection of the specific anti-neutrophil antibody anti-Ly-6G into GKM-treated mice resulted in a reduction in the influx of neutrophils to the lungs by nearly 30%, verifying the specificity of the response with regard to the involvement of neutrophils (Fig 6C and 6D, GKM+αLy-6G). We further assessed whether the early influx of neutrophils to the lungs of GKM-treated mice led to induction of the MMP8 and MMP9 metalloproteinases. Indeed, their expression was significantly higher in lung extracts obtained from GKM-treated mice as compared to control-treated mice (Fig 6E). Again, injection of anti-Ly-6G to GKM-treated mice lowered the expression of MMPs (Fig 6E), indicating that these early recruited neutrophils are in an active state once they reach the lung.
To determine whether early recruited neutrophils to the lungs are able to clear Y. pestis, we measured the bacterial burden in the lungs of GKM-treated mice at 24 hr following infection with 1x105 cfu (100LD50) of the virulent Kim53 strain. A substantial reduction of almost a thousand fold was observed for the load of Y. pestis in the lung in comparison to untreated mice (Fig 7A). This beneficial effect of GKM treatment was mediated by neutrophils, as the injection of anti-Ly-6G neutralizing antibody to GKM-treated mice decreased bacterial clearance, reflecting the importance of neutrophils for lung defense against Y. pestis (Fig 7A). Furthermore, analysis of the relationships between the numbers of neutrophils and bacterial loads in the lungs of GKM-treated versus sham-treated mice at the early time point of 24 hpi, indicated that under this treatment levels of neutrophils greater than 1–2×106 were associated with effective bacterial clearance (Fig 7B).
Following the demonstration of early migration of neutrophils to the lung by the treatment with the recombinant proteins and the pronounced effect on Y. pestis propagation, it was interesting to evaluate this treatment in animals exposed to a lethal challenge. Relatively high protection level of 60% was observed in GKM-treated mice that were exposed i.n. to a dose of 2x103 cfu of the virulent Kim53 strain, whereas only 10% of the sham-treated mice survived this infection (Fig 7C).
The ability of bacterial pathogens to prevent the early recruitment of neutrophils to infected organs provides an obvious advantage during infection because these cells, with their various antimicrobial capabilities, would otherwise kill the pathogen. Y. pestis, the causative agent of plague, exploits a variety of mechanisms for evading and coping with the host immune response during the early stages of infection. Accumulating evidence based on studies in various animal models of pneumonic plague indicates that following airway infection with Y. pestis, the early induction of a pro-inflammatory immune response in the lungs as well as the recruitment of neutrophils to the lungs are delayed [20–22].
We previously showed that the immune response is initiated by BM cells early after i.n. infection of mice with a fully virulent Y. pestis strain, causing rapid modulation of the BM CXCR4-SDF-1 axis and prompt mobilization of neutrophils into the circulation within 12–24 hpi [30]. These observations raised intriguing questions, namely, at what stage of neutrophil recruitment to the lungs does the pathogen interfere and which Y. pestis virulence factor is involved in this process. In this study, we further analyzed the mechanisms involved in the late homing of neutrophils to the lungs following i.n. infection of mice with the fully virulent Y. pestis strain Kim53.
Our results clearly indicate that the influx of neutrophils to the lungs in Y. pestis-infected mice is delayed. In addition, the delayed recruitment of neutrophils to the lungs is associated with a significant increase in bacterial burden. Cytokines and chemokines act in a coordinated manner to mobilize and recruit neutrophils to the site of inflammation. Because production of these factors represents the first step in the neutrophil recruitment process, we monitored the expression of several chemokines critical for the chemoattraction of neutrophils, in the lungs and plasma of infected mice during disease progression. Up-regulation of G-CSF, KC and MIP-2 in the plasma of infected mice was delayed until the late stages of disease progression (i.e., 48 hpi), consistent with the absence of a pro-inflammatory response in the lungs at the early stage of disease progression [20–22]. In addition, the levels of CXCR2 (KC and MIP-2 receptor), did not change at the first 24 hpi on circulating neutrophils in the blood (S3 Fig).
Using a Transwell-migration assay, we found that lung extracts from the early stage of disease progression were incapable of inducing neutrophil migration, in contrast to lung extracts obtained from mice at later stages of the disease. This result suggests that the intrinsic induction of chemoattractant production in lung resident cells early after Y. pestis infection is inhibited. Moreover, the delayed induction of KC, MIP-2 and G-CSF mRNA and protein in the lungs of Kim53-infected mice corroborates this observation.
The expression of adhesion molecules is up-regulated on endothelial cells located at the site of inflammation [38]. Circulating neutrophils that egress from the BM undergo E- and P-selectin-mediated rolling along the endothelial surface, followed by firm attachment via ICAM-1 and VCAM-1 [39]. In addition to the delay in chemokine up-regulation, the ability of Kim53-infected lung cells to support neutrophil transmigration into the lungs also appears to be impaired at the early stage of infection. This is due to the delayed induction of the adhesion molecules E- and P-selectin as well as ICAM-1 and VCAM-1. Because chemokines are involved in the expression of adhesion molecules on capillary endothelia [1], the delayed induction of adhesion molecules in Y. pestis-infected lungs might result from delayed expression of the KC and MIP-2 chemokines. Alternatively, direct interaction of Y. pestis with endothelial cells might affect the expression of these molecules, as demonstrated for ICAM-1 during in vitro infection of human umbilical vein endothelial cells (HUVECs) with the related enteropathogen Y. enterocolitica [40].
The virulence of pathogenic Yersinia strains is mostly attributed to the TTSS and its effector proteins which are used by the pathogen to subvert early innate immune responses [26, 41]. In striking contrast to the impaired innate immune response in the lungs of mice infected i.n. with the virulent Y. pestis strain Kim53, rapid and moderate induction of the expression of chemokines and adhesion molecules followed by an influx of neutrophils to the lungs was observed early after pulmonary infection of mice with the avirulent Y. pestis strain (Kim53Δ70Δ10) that lacks the TTSS and Pla protease virulence factors. Consequently, this prompt response was associated with effective bacterial clearance from the lungs.
TTSS Yop effectors of pathogenic Yersinia species are known for their ability to suppress the induction of innate immune responses in various types of mammalian cells through disruption of the target cell signaling network. For example, YopE has been reported to inhibit the production of IL-8 (the human homologue of KC) in Y. pseudotuberculosis-infected HeLa cells [42, 43], and YopJ-dependent suppression of TNFα secretion has been reported in Yersinia-infected macrophages [44–46]. Human bronchial epithelial cells co-transfected with cDNAs encoding Y. pseudotuberculosis YopJ also exhibited reduced transcription of IL8, RANTES and ICAM-1 in a promoter activity assay [47]. Moreover, the YopH effector was shown to inhibit the expression of monocyte chemoattractant protein 1 (MCP-1) in macrophages infected with Y. enterocolitica [48] and to suppress early pro-inflammatory cytokines in the lungs during pneumonic plague [49].
We investigated the involvement of YopJ and YopH in the modulation of KC, MIP-2 and G-CSF expression by alveolar macrophages (MH-S) and lung epithelial cell lines (TC-1) following infection of the cells with Y. pestis Kim53 YopJ and YopH deficient mutants. The in vitro results indicated that inhibition of KC, MIP-2 and G-CSF mRNA and protein expression was mediated by the YopJ effector in a similar manner that was shown after infection with the avirulent Y. pestis strain—Kim53Δ70Δ10, which lacks the entire TTSS.
YopJ (YopP in Y. enterocolitica) belongs to a family of proteases related to the ubiquitin-like protein proteases [50], and YopJ was shown to be a deubiquitinating cysteine protease capable of removing ubiquitin moieties from IκBα, thereby inhibiting its proteasomal degradation and leading to the down-regulation of NF-κB function [51]. In addition, YopJ was shown to acetylate Ser/Thr residues in the activation loop of MAPK kinases (MKKs) and IκB kinases (IKKs), thereby preventing their activation by phosphorylation [52, 53]. This acetyltransferase activity of YopJ may well account for its ability to inhibit MAPK pathways and NF-κB activation. Interestingly, expression of the KC and MIP-2 genes is known to be tightly regulated by the NF-κB transcription factor [54, 55], and NF-κB activation in lung epithelial cells was shown to be important for the migration of neutrophils to the lungs [56]. Therefore, YopJ-mediated inhibition of NF-κB activity may be involved in the suppression of KC and MIP-2 expression in lung-derived cell lines infected with Y. pestis.
Studies on the involvement of YopJ in Y. pestis virulence have indicated that this effector is not essential for virulence in various rodent models of plague [57–59]. Similar results were obtained in our study using C57BL/6 mice infected i.n. with the Kim53ΔYopJ mutant. Close examination of the expression of neutrophil chemoattractants in the lungs of Kim53ΔYopJ-infected mice revealed that while the mRNA levels of KC, MIP-2 and G-CSF were induced in the lungs early after infection, the proteins levels were up-regulated only at the late stage of the disease, i.e., 48 hpi. In line, the influx and homing of neutrophils to the lungs of Kim53ΔYopJ-infected mice was also delayed to late stage of the disease as observed following exposure to the wild-type Y. pestis strain. These intriguing data may indicate that Y. pestis affects the early recruitment of neutrophils to the lungs by regulating the expression of neutrophil chemoattractants at both the transcriptional level via the YopJ effector and at the posttranscriptional level by another, yet unknown virulence factor. One possible candidate is the YopH tyrosine phosphatase that was shown to suppress early pro-inflammatory cytokine up-regulation in the lungs of Y. pestis-infected mice; this factor was also shown to be essential for Y. pestis virulence in the mouse model of pneumonic plague [49]. Another candidate is the Pla protease that was shown to be important for Y. pestis proliferation in the lungs [60] and for degradation of Fas ligand to manipulate host cell death and inflammation [61]. The apparent discrepancy between the in vitro and in vivo infection systems suggests a different mode of chemokines expression in the absence of YopJ. While YopJ depletion resulted in augmented mRNA and protein expression levels of KC, MIP-2 and G-CSF by infected macrophages and epithelial cell lines in vitro, lungs obtained from infected mice exhibited changes only in the mRNA and not in the protein levels of these chemokines early after infection. We assume that these differences emphasize the major impact of the alveolar niche and its various resident cells on the complex virulent mechanisms generated by Y. pestis. Furthermore, the lack of host defense components (such as other white blood cells, immunoglobulins, complement, cytokines, defensins and more) in in vitro models, may have a great influence on the virulence quality of the pathogen.
Because we previously showed that neutrophils egressed from the BM at 12–24 hr after i.n. infection with Y. pestis [30], our current findings point to the delayed induction of chemokines as the major reason for the inability of circulating neutrophils to rapidly infiltrate the site of infection in the lungs. To test this hypothesis, we administered recombinant KC and MIP-2 by i.n. instillation into the lungs of Y. pestis-infected mice that were pre-treated systemically with G-CSF to synchronize and overload the circulation with primed neutrophils [62, 63].
Whereas systemic pre-treatment of mice with G-CSF alone prior to pulmonary infection with Y. pestis did not lead to early recruitment of neutrophils to the lungs, additional i.n. administration of KC and MIP-2 several hours after the infection resulted in rapid mobilization of neutrophils to the lungs. Furthermore, the recruitment of neutrophils to the lungs of treated mice was accompanied by an increase in neutrophil-associated MMP8 and MMP9 expression and induction of the expression of the adhesion molecules E- and P-selectin in the lungs (S4 Fig).
The early influx of neutrophils to the lungs of Y. pestis-infected mice led to a rapid and significant reduction of nearly 1,000-fold in the average bacterial cfu in the lungs of mice infected with a dose of 100 LD50 of the virulent Y. pestis Kim53 strain. Moreover, this treatment was also successful in improving the survival rates of mice following i.n. exposure to a lethal dose of 2 LD50 of the fully virulent Y. pestis strain. Injection of treated mice with neutralizing anti-Ly-6G antibodies reduced the percentages of neutrophils by 30% and diminished the beneficial antibacterial outcome of treatment as well as the expression of MMPs, highlighting the contribution of recruited neutrophils to lung defense against Y. pestis.
Collectively, our results indicate that Y. pestis-mediated interference with the early induction of G-CSF, KC and MIP-2 expression by lung resident cells and the recruitment of neutrophils to the lungs are important for the manifestation of Y. pestis virulence during pneumonic plague. These observations are consistent with previous reports demonstrating (a) the delayed expression of pro-inflammatory cytokines and chemokines during pneumonic plague [20–22] and (b) the importance of neutrophils for defense against Y. pestis pulmonary infection [15]. Notably, Goldman and his colleagues have recently reported that the number of neutrophils in the lungs of Y. pestis-infected mice increases rapidly within 24 hr post intranasal infection [29]. Differences in the experimental systems may account for this discrepancy, as this group infected the mice with an inoculation dose of 106 cfu of the CO92 strain pre-grown at 37°C, whereas in our experiments mice were infected with a lower dose of 105 cfu of the Kim53 strain pre-grown at 28°C.
Directed guidance of immune cells to the lungs following i.n. administration of recombinant proteins has become an important tool for understanding the lung defense mechanisms against bacterial infections. Intranasal administration of exogenous KC ameliorated B. pertussis-mediated inhibition of neutrophil recruitment to the lungs in infected mice [64]. In addition, the recruitment of neutrophils to the lungs of S. pneumonia-infected mice was observed following i.n. administration of IL-12 and was associated with increased levels of KC, decreased bacterial burden and improved survival [65]. The early arrival of neutrophils to sites of infection may also influence the outcome of disease progression by regulating other types of immune cells. Bi Y et al. recently reported that production of IL-17A by neutrophils coordinates the antimicrobial activity of neutrophils and macrophages against Y. pestis infection during pneumonic plague [66]. In addition, the secretion of neutrophil-derived granule proteins and antibacterial peptides was associated with the migration of inflammatory monocytes to the site of infection [67, 68]. The therapeutic potential of the early recruitment of neutrophils to the lungs may also be attributed to the involvement of these cells in disease resolution and anti-inflammatory processes. Neutrophils have recently been shown to play an important role in the repair of damaged tissue through the expression of MMP9, which degrades intracellular matrix (ICM) components, thus promoting the removal of damage-associated molecular pattern (DAMP)-containing ICM proteins released from damaged cells (Reviewed in [69]).
Taken together, this study highlights (a) the complex virulence mechanisms employed by Y. pestis to minimize its early encounter with neutrophils in the lungs following airway infection and (b) the beneficial effect of modulating neutrophil chemotaxis into the lungs at the early stage of Y. pestis infection by treatment with chemoattractants. This therapeutic approach could be useful for improving treatments against plague as well as other pathogens that suppress the recruitment of neutrophils to sites of infection. The existence of antibiotic-resistant Y. pestis strains [70] further emphasizes the importance of modulating the host defense for the treatment of plague infections.
This study was carried out in strict accordance with the recommendations for the Care and Use of Laboratory Animals of the National Institute of Health. All animal experiments were performed in accordance with Israeli law and were approved by the Ethics Committee for animal experiments at the Israel Institute for Biological Research (Permit Numbers: IACUC-IIBR M-07-2012, IACUC-IIBR M-28-2013). During the experiments, the mice were monitored daily. Humane endpoints were used in our survival studies. Mice exhibiting loss of the righting reflex were euthanized by cervical dislocation. Analgesics were not used, as they may have affected the experimental outcomes of the studies.
The following Y. pestis strains were used in this study: the Y. pestis virulent strain Kimberley53 (Kim53) [71], the avirulent Kim53Δ70Δ10 strain that is spontaneously cured for pPCP1 and pCD1 [72], Kim53 deleted for YopJ (Kim53ΔYopJ) [73] and the Kim53 deleted for YopH (Kim53ΔYopH) (Constructed by Zauberman A, replacing the yopH gene with a kanamycin resistance cassette as described in [73]). The Y. pestis vaccine strain EV76 [71], EV76 spontaneously cured for pCD1 (EV76Δp70) [74], EV76ΔYopJ [46], and the EV76 deleted mutants EV76ΔYopK, EV76ΔYopE, EV76ΔYopH (Constructed by Zauberman A., replacing the yopK, yopE and yopH genes with a kanamycin resistance cassette as described in [73]). The strains were routinely grown on brain heart infusion agar (BHIA, BD, MD USA) for 48 hr at 28°C. The Y. pestis Yop-deleted strains were grown on BHIA supplemented with 100 μg/ml kanamycin (Sigma-Aldrich, Israel).
Deletion mutagenesis of the Y. pestis Kim53 and EV76 strains was performed by replacing the central region of the genes with a kanamycin resistance cassette (Pharmacia) by homologous recombination. The protocol used was based on previously established methodologies [75, 76]. The linear PCR fragment in which kanamycin sequences were flanked by yop sequences was electroporated into Y. pestis bacteria expressing the λ phage red system from pKOBEG::sacB (generous gift from Dr. E. Carniel). Electroporation was performed in 10% glycerol and 10% PEG-8000 (Sigma-Aldrich, Israel), and the bacteria were incubated in HIB for 2 h at 28°C. Transformants were selected on BHIA containing 50 μg/ml kanamycin, and then the pKOBEG::sacB plasmid was removed from the bacteria by growing the bacteria on BHIA supplemented with 10% sucrose. The expected knockout phenotype was verified by PCR and Western blot analyses.
Female C57BL/6 mice (6–10 weeks old) were purchased from Harlan Laboratories (Rehovot, Israel) and maintained under defined flora conditions at the animal facilities of the Israel Institute for Biological Research. The i.n. infections were performed as described previously [77]. Briefly, bacterial colonies were harvested and diluted in heart infusion broth (HIB) (BD, USA) supplemented with 0.2% xylose and 2.5 mM CaCl2 (Sigma-Aldrich, Israel) to an OD660 of 0.01 and grown for 22 h at 28°C in a shaker (100 rpm). At the end of the incubation period, the cultures were washed, diluted in PBS solution to the required infectious dose and quantified by counting colony forming units (cfu) after plating and incubating on BHIA plates (48 hr at 28°C). Prior to infection, mice were anesthetized with a mixture of 0.5% ketamine HCl and 0.1% xylazine and then infected i.n. with 35 μl/mouse of the bacterial suspension, whereas naïve mice were instilled i.n. with PBS only. The intranasal LD50 of the Kim53 strain under these conditions is 1,100 cfu. LD50 values were calculated according to the method described by Reed and Muench [78].
Three days prior to infection, mice received daily subcutaneous injections of recombinant G-CSF (rhG-CSF 300 μg/kg/Nupogen 48 MU/0.5 ml, Roche Applied Science) for 5 consecutive days. Six hours after i.n. infection with Y. pestis Kim53, mice were euthanized, and 1 μg of each recombinant KC and MIP-2 (recombinant MCXCL1/KC, recombinant MCXCL2/MIP-2, R&D Systems), diluted in 25 μl of PBS, or 25 μl of PBS alone (sham) were instilled i.n. Mice were either sacrificed and analyzed 24 hpi or followed to observe the rates of morbidity and mortality. For the depletion of neutrophils, 100 μg purified anti-Ly-6G antibody clone 1A8 (Biolegend, USA) diluted in 300 μl PBS was administered intraperitoneally twice at 24 hr prior to infection and 3 hpi.
The murine alveolar macrophage cell line MH-S was obtained from ATCC. TC-1 is a tumor cell line derived from primary lung epithelial cells of C57BL/6 mice[79]. This cell line was a kind gift from the laboratory of Prof. T.C. Wu (Johns Hopkins University). Both cell-lines were grown in RPMI 1640 medium supplemented with 10 mM HEPES, 2 mM L-glutamine, 1 mM sodium pyruvate, 0.1 mM non-essential amino-acids and 10% fetal bovine serum. Cell cultures were maintained at 37°C with 5% CO2. Cell infection studies were performed as previously described [73]. Briefly, bacteria were grown by shaking (150 rpm) for 22 h at 28°C in HIB. The resulting cultures were diluted in HIB medium to OD660 0.05 and allowed to grow for 3 h at 37°C (100 rpm). Bacteria were harvested, washed once and re-suspended in complete RPMI supplemented with 10% fetal calf serum and added to the cells at a multiplicity of infection (MOI) of 50. Bacteria were adhered onto the cells by centrifugation at 130 g for 5 min followed by incubation for an additional 1 h at 37°C and 5% CO2. Gentamicin was then added to the cultures to a final concentration of 50 mg/ml, and the cultures were incubated for an additional 4 h before using the cells for RNA extraction and RT-PCR analysis and the media for ELISA.
To prepare lung cell suspensions, mice were euthanized, and blood was withdrawn from the heart using a heparinized syringe. Lungs were then removed and placed on a 70-μm nylon cell strainer (BD Falcon,USA) dipped in 2 ml PBS containing 1% protease inhibitor cocktail (Sigma-Aldrich, Israel). Cell suspensions were pelleted at 260 g for 10 min at 4°C, fixed in 4% paraformaldehyde in PBS for 1 h at room temperature and washed twice in flow cytometry buffer. Neutrophils (CD11b+/Gr-1high) were stained with PerCP-Cy5.5-conjugated anti-mouse CD11b antibody (clone M1/70) (eBioscience, USA) and APC-conjugated anti-mouse Ly-6G (Gr-1) antibody (clone RB6-8C5) (eBioscience, USA). CXCR2 staining was performed with PE-conjugated anti-CXCR2 antibodies (clone 242216) (R&D, USA). Cells were stained using standard protocols with appropriate matched isotype control antibodies. The analysis was performed using a FACSCalibur flow cytometer with CellQuest Pro software (BD Bioscience, USA).
Lung cell suspensions were prepared as previously described. Total RNA was extracted using Tri-reagent (Sigma-Aldrich, Israel) according to the manufacturer’s instructions. Two micrograms of total RNA were reverse-transcribed using Moloney murine leukemia virus reverse transcriptase and oligo-dT primers (Promega, USA). Quantitative PCR analysis was performed using an ABI 7500 instrument (Applied Biosystems, USA) with SYBR green PCR master mix (Applied Biosystems, USA). The fold change in the quantity of gene transcripts was measured and compared to the hypoxanthine phosphoribosyl transferase (HPRT) gene using the comparative (-2ΔΔCt) method. Forty cycles of PCR amplification were performed in duplicate for each primer set. Primer sequences used are listed in Table 1.
Blood was collected and centrifuged at 260 g for 10 min, and the plasma was collected, filtered and stored at -70°C. Bronchoalveolar lavage fluid (BALF) was collected by exposing the trachea and injecting, and then removing twice, a total of 1 ml PBS containing 1% protease inhibitor cocktail (Sigma-Aldrich, Israel). BALF was then filtered and stored at -70°C. Before analysis, samples were centrifuged again at 13,000 g for 5 min. The levels of KC, MIP-2 and G-CSF in the plasma and BALF and the levels of MMP9 in the BALF, were measured by enzyme-linked immunosorbent assay (ELISA) according to the manufacturer’s protocol (R&D Systems, MN, USA).
Gelatin zymography for MMP9 activity was performed as previously described [80]. Briefly, whole lung supernatants were mixed with a non-reducing sample buffer, and an equal amount was loaded onto 10% SDS-polyacrylamide gels co-polymerized with 1 mg/ml gelatin derived from porcine skin (Sigma-aldrich, Israel). After electrophoresis, the gels were washed for 30 min in Triton X-100, followed by 3 washes with H2O and incubation at 37°C for 16 hr in developing buffer. The gels were then stained with SeeBand Forte (Gene Bio-Application Ltd) until clear bands appeared, indicating the presence of MMP9 activity. Conditioned media of HT-1080 cells secreting MMP9 served as a control.
Transwell-migration assays were performed as previously described [81]. Briefly, total BM cells were extracted from the femur and tibias of naïve mice and suspended in complete RPMI media. Prior to the migration assay, the total BM cells were labeled with the neutrophil markers Gr-1 and CD11b to evaluate the levels of neutrophils. The cells were counted, and 250,000 cells/100 μl were allowed to migrate towards a total of 600 μl media containing 150 μl of lung supernatant through 24-well filters with a pore size of 5 μm (Corning, NY, USA) at 37°C for 3 hr. The cells were then collected and counted using flow cytometry. In parallel, a portion of the migrated cells was labeled with Gr-1 and CD11b to evaluate the number of migrated cells. The percentage of migrated cells was then calculated by comparing the number of neutrophils before and after migration.
Bronchoalveolar lavage fluid (BALF) was collected as previously described, and equal amount of BALF samples were subjected to 10% SDS-PAGE followed by immunoblot with anti MMP8 polyclonal antibody (Proteintech cat:17874-1-AP).
Statistical significance was determined using the nonparametric Mann-Whitney test. A Kaplan-Meier survival estimate of treated and non-treated mice (of at least 10 mice per group) was also performed. Calculations were made using GraphPad Prism software.
|
10.1371/journal.pcbi.1006793 | Available energy fluxes drive a transition in the diversity, stability, and functional structure of microbial communities | A fundamental goal of microbial ecology is to understand what determines the diversity, stability, and structure of microbial ecosystems. The microbial context poses special conceptual challenges because of the strong mutual influences between the microbes and their chemical environment through the consumption and production of metabolites. By analyzing a generalized consumer resource model that explicitly includes cross-feeding, stochastic colonization, and thermodynamics, we show that complex microbial communities generically exhibit a transition as a function of available energy fluxes from a “resource-limited” regime where community structure and stability is shaped by energetic and metabolic considerations to a diverse regime where the dominant force shaping microbial communities is the overlap between species’ consumption preferences. These two regimes have distinct species abundance patterns, different functional profiles, and respond differently to environmental perturbations. Our model reproduces large-scale ecological patterns observed across multiple experimental settings such as nestedness and differential beta diversity patterns along energy gradients. We discuss the experimental implications of our results and possible connections with disorder-induced phase transitions in statistical physics.
| The diversity, stability and functional structure of microbial communities have dramatic effects on the health of humans and of ecosystems. The complexity of these communities has so far precluded the development of a general predictive model that would capture the dependence of these features on environmental conditions. We confronted this challenge by constructing a flexible simulation framework, and randomly sampling parameters under a variety of modeling assumptions to identify generic patterns. We found two qualitatively distinct regimes of community structure, which reproduce observed patterns of biodiversity, and make new predictions about stability and function.
| Microbial communities inhabit every corner of our planet, from our own nutrient-rich guts to the remote depths of the ocean floor. Different environments harbor very different levels of microbial diversity: in some samples of non-saline water at mild temperature and pH, nearly 3,000 coexisting types of bacteria can be detected, whereas at ambient temperatures warmer than 40° C, most cataloged samples contain fewer than 100 distinct variants [1]. The functional structure of these communities is also highly variable, with functional traits often reflecting the environment in which the communities are found [1, 2]. A central goal of microbial community ecology is to understand how these variations in diversity, stability and functional structure [3] arise from an interplay of environmental factors such as energy and resource availability [4, 5] and ecological processes such as competition [6–9] and stochastic colonization [10–13].
This endeavor is complicated by the fact that microbes dramatically modify their abiotic environments through consumption and secretion of organic and inorganic compounds. This happens on a global scale, as in the Great Oxidation Event two billion years ago [14, 15], and also on smaller scales relevant to agriculture, industry and medicine. In this sense, every microbe is an “ecosystem engineer” [16]. Metabolic modeling and experiments suggests that metabolically mediated syntrophic interactions should be a generic feature of microbial ecology [17–19] and that complex microbial communities can self-organize even in constant environments with no spatial structure or predation [17, 20]. For these reasons, there has been significant interest in developing new models for community assembly suited to the microbial setting [21–25].
Here, we present a statistical physics-inspired consumer resource model for microbial community assembly that builds upon the simple model introduced in [17] and explicitly includes energetic fluxes, stochastic colonization, syntrophy, and resource competition. We focus on modeling complex communities with many species and metabolites. By necessity, any mathematical model of such a large, diverse ecosystem will contain thousands of parameters that are hard to measure. To circumvent this problem, we take a statistical physics approach where all consumer preferences and metabolic parameters are drawn from random distributions.
This approach to modeling complex systems has its root in the pioneering work of Wigner on the spectrum of heavy nuclei [26] and was adapted by May to ecological settings [27]. Recently, there has been a renewed interest in using these ideas to understand complex systems in both many-body physics (reviewed in [28]) and community assembly [12, 17, 25, 29–35]. The key insight underlying this approach is that generic and reproducible large-scale patterns observed across multiple settings likely reflect typical properties, rather than fine tuned features of any particular realization or community. Consistent with this idea, it was recently shown that a generalized consumer resource model with random parameters can reproduce many of the patterns observed in experiments where natural communities were grown in synthetic minimal environments [17].
In this paper, we ask how varying the energy flux into an ecosystem and the amount of cross-feeding affects microbial community assembly. We find that the resulting communities generically fall into two distinct regimes, characterized by qualitative differences in their community-level metabolic networks, functional structures, responses to environmental perturbations, and large-scale biodiversity patterns. We show our model predictions are consistent with data from the Tara Oceans database [36] and the Earth Microbiome Project [1], and propose feasible experimental tests using synthetic communities.
The starting point for our analysis is a new model that adapts MacArthur’s Consumer Resource Model [7] to the microbial context by including energetics, stochastic colonization, and the exchange and consumption of metabolites. We consider the population dynamics of S species of consumers (e.g., microbes) competing for M types of substitutable resources. We are interested in large, diverse ecosystems where S, M ≫ 1. A schematic summarizing our model is shown in Fig 1.
A natural setting for considering substitutable resources is when all essential biomass components are supplied in excess, and the limiting factor for growth is the supply of usable energy. In this context, one only needs to keep track of resources from which energy can be harvested. All other nutrients are included implicitly, under the assumption that some of the energy budget is used to import whatever materials are required for growth and reproduction. Terminal waste products from which no more energy can be extracted are likewise treated implicitly, and are not included among the M resource types.
In our model, the rate at which an individual of species i harvests energy from resource α depends on the resource concentration Rα as well as on the consumer’s vector of resource preferences ciα through the expression:
J i α in = w α σ ( c i α R α ) , (1)
where σ(x) encodes the functional response and has units of mass/time, while wα is the energy density of resource α with units of energy/mass. In the microbial context the consumer preferences ciα can be interpreted as expression levels of transporters for each of the resources. In the main text, we focus on Type-I responses where σ(x) = x, and we set wα = 1 for all α, but most of our results still hold when σ(x) is a Monod function or the wα are randomly sampled, as shown in Section 3 of S1 Text.
We model leakage and secretion by letting a fraction lα of this imported energy return to the environment, so that the power available to the cell for sustaining growth is equal to
J i grow = ∑ α ( 1 - l α ) J i α in . (2)
This parameterization guarantees that the community does not spontaneously generate usable energy in violation of the Second Law of Thermodynamics. We assume that a fixed quantity mi of power per cell is required for maintenance of a steady population of species i, and that the per-capita growth rate is proportional to the remaining energy flux, with proportionality constant gi. In typical experimental conditions, cell death is negligible, and mi is the energy harvest required for the replication rate to keep up with the dilution rate. Under these assumptions, the time-evolution of the population size Ni of species i can be modeled using the equation
d N i d t = g i N i [ J i grow - m i ] . (3)
The leaked energy flux J i out = ∑ αl α J i α in from each cell of species i is partitioned among the M possible resource types via the biochemical pathways operating within the cell. We assume that all species share a similar core metabolism, encoded in a matrix Dβα. Each element of Dβα specifies the fraction of leaked energy from resource α that is released in the form of resource β (note that by definition, ∑β Dβα = 1). Thus, in our model the resources that are excreted into the environment are intimately coupled to the resources a cell is consuming. The outgoing energy flux contained in metabolite β is given by
J i β out = w β ν i β out = ∑ α D β α l α J i α in . (4)
The dynamics of the resource concentrations depend on the incoming and outgoing mass fluxes ν i α in = σ ( c i α R α ) and ν i α out, which are related to the energy fluxes via the energy densities wα. In terms of these quantities, we have
d R α d t= h α + ∑ j N j ( ν j α out - ν j α in ) , (5)
with hα encoding the dynamics of externally supplied resources. In this manuscript, we focus on the case where the microbial communities are grown in a chemostat with a single externally supplied resource α = 0 (Fig 1). In this case, the resource dynamics can be described by choosing h α = κ α - τ R - 1 R α, with all the κα set to zero except for κ0. These equations for Ni and Rα, along with the expressions for J i α in and J i α out, completely specify the ecological dynamics of the model.
This model has been implemented in a freely available open-source Python package “Community Simulator.” The package can be downloaded from https://github.com/Emergent-Behaviors-in-Biology/community-simulator.
To assess the typical community structure and resource pool stability for ecosystems obeying Eqs (1)–(5), we randomly generated an M × M metabolic matrix Dαβ, and a binary S × M consumer preference matrix ciα with S = 200 species and M = 100 resources. We chose ciα so that each species had 10 preferred resource types on average, with ciα = 1, while the rest of the resources were consumed at a baseline level of ciα = 0.01. The metabolic matrix Dαβ was sampled from a Dirichlet distribution, which guarantees that all the columns sum to 1 as required by the definition of this parameter. In Section 3 of S1 Text, we show that the qualitative patterns we observe are unchanged if ciα is drawn from a Gaussian or Gamma distribution, or if the Dαβ matrix is made less sparse. The full sampling procedure is detailed in S1 Text Section 1.
We chose our units of energy flux such that the mean maintenance cost mi over all species in the regional pool is equal to 1. To break ties between species with similar consumption profiles, we added a Gaussian random offset to the mi of each species with standard deviation 0.1. In S12 Fig., we show that these intrinsic fitness differences do not dominate the ecological dynamics, and that many species with relatively high maintenance costs are able to reach large population sizes in the steady-state communities. Finally, we set all the wα equal to 1, and made all the leakage fractions identical, with lα = l for all α.
To assess the amount of variability in the results, we initialized 10 different communities by seeding each one with a random subset of 100 species from the full 200-species pool. This simulates the stochastic colonization frequently observed in microbial ecosystems, where the community composition can randomly vary depending on the set of microbes this particular local environment happened to be exposed to [37]. Fig 1 shows typical dynamical trajectories in the presence of high (l = 0.8) and extremely low leakage (l = 0.001).
Our numerical simulations display a transition between two qualitatively different community structures as we vary the externally supplied energy flux w0κ0 and the leakage/syntrophy l. In the “thermodynamic limit” of M, S → ∞, the communities exhibit signatures of a phase transition analogous to those found in disordered systems in physics (see Discussion and S1 Text Section 5). Fig 2 shows the effect of this transition on community diversity at our chosen finite values of S and M. At low levels of energy flux or syntrophy, the diversity is severely limited by resource availability. In the limit of high supplied energy flux and high leakage, a maximally diverse regime is obtained, where the number of surviving species is limited only by the similarity between consumption profiles within the regional species pool, in accordance with classical niche-packing theory [7] as we will discuss below.
The difference between the two regimes is most apparent from the perspective of the energy flux networks. Because our model explicitly accounts for the flow of energy from one resource type into another, we can compute all the steady-state fluxes and represent them graphically, as shown in Fig 3 for some representative examples. Each node in this network is a resource type, and each directed edge represents the steady-state flux Jβα of energy conversion from resource α to resource β, mediated by one or more syntrophic consumers:
J β α= ∑ i N i J i β α out = D β α l α ∑ i N i w α c i α R α . (6)
The resource-limited regime produces a unidirectional flow of energy, which is converted from the externally supplied resource type into an orderly succession of secreted resources. For the sparse metabolic matrix shown in the top row of Fig 3, most resource types also have extremely small incoming flux vectors in this regime, with magnitudes less than 1% the size of the largest flux in the network. The diverse regime displays a qualitatively different structure, where all resources have significant incoming fluxes (regardless of the choice of Dαβ), and the large number of loops in the network makes it impossible to put the resource types into any definite order. In S10 Fig., we plot the fraction of samples from Fig 2 whose (pruned) flux networks are free of cycles, and confirm that this observation is generic. The dramatic contrast between the community-level metabolism of the two regimes affects many other global features of the ecosystem, which we will explore in the following sections.
To better understand the behavior of consumers in the two regimes, we examined the functional traits of members of typical communities in each one. In the resource-limited regime, many surviving species derive most of their energy directly from the externally supplied resource (Fig 4A). In the diverse regime, by contrast, only a minority of the steady-state community members can consume this resource at all, and even these species receive most of their energy from a diverse array of metabolic byproducts (Fig 4B). We quantified this observation using the Simpson Diversity M i eff of the incoming resource flux vectors J i α in, which measures the effective number of resources consumed by each species, and is closely related to the inverse participation ratio in statistical physics. The Simpson Diversity is defined by
Mieff=[ ∑α(JiαinJiin)2 ]−1, (7)
where J i in = ∑ αJ i α in is the total incoming energy flux for each cell of this species. M i eff approaches 1 for species that obtain the bulk of their energy from a single resource type and approaches M when all resource types are consumed equally. In the resource-limited regime, the distribution of these values is sharply peaked around 2. In the diverse regime, the peak is located around 10, which is the average number of resources with high transporter expression in our binary sampling scheme for ciα. This shows that most community members in the diverse regime utilize multiple energy sources, with the incoming flux spread evenly over all resource types they are capable of consuming.
Another important property of microbial ecosystems is how they respond to environmental perturbations. Previous theoretical studies have shown that sufficiently diverse communities can “pin” the resource concentrations in their local environment to fixed values, which are independent of the magnitude of externally supplied fluxes [21, 38, 39]. In these studies, resource pinning occurs only when the community saturates the diversity bound imposed by the principal of competitive exclusion, i.e. when the number of coexisting species is at least as large as the number of resource types. Such maximally diverse communities typically require fine-tuning of the resource utilization profiles or imposition of universal efficiency tradeoffs in cellular metabolism.
In our stochastically assembled communities, the diversity is always much lower than the number of resource types, so we hypothesized that the resource concentrations should not be pinned. To test this idea, we measured the response of the steady-state concentrations R ¯ α to changes in external supply rates κα, in terms of the “resource susceptibilities” ∂ R ¯ α / ∂ κ α plotted in Fig 4D [34]. Our hypothesis was valid in the resource-limited regime, where many resource susceptibilities are comparable to the susceptibility in the empty chemostat ∂ R ¯ α / ∂ κ α = τ R = 1. But in the diverse regime, we were surprised to find that the susceptibilities are 100 times smaller than this maximum value. This suggests that resource pinning may be a generic phenomenon, observable in real ecosystems when the energy supply is sufficiently large.
In the diverse regime, the number of coexisting species (“richness”) is not limited by energy availability or by access to secreted metabolites, but is still much less than the maximal value of M = 100 set by the competitive exclusion principle [8], even though almost all M resource types are present at non-negligible levels (as shown in S11 Fig.). We hypothesized that the diversity in this regime is limited by the degree of similarity between consumption preferences of members of the regional species pool. This can be quantified in terms of the niche overlap [9, 40], whose average value in a large regional pool is given by:
〈 ρ i j 〉 ≡ 〈 ∑ αc i α c j α ∑ αc i α 2 ∑ αc j α 2 〉 = 〈 c i α 〉 2 〈 c i α 2 〉 . (8)
Fig 5 shows how the richness varies as a function of 〈ρij〉. In the diverse regime the mean richness decreases approximately linearly with increasing overlap. The richness of the resource-limited regime, on the other hand, has only a very weak dependence on the niche overlap. These results suggest that the distribution of consumption preferences in the regional pool is the primary driver of community assembly in the diverse regime. Importantly, non-zero niche overlap limits the number of coexisting species well below the upper bound imposed by the competitive exclusion principle.
Our aim in developing this model is to identify and understand generic patterns in community structure, that are independent of particular biological details. In large-scale surveys of natural communities, subject to many sources of noise and environmental heterogeneity, one expects that only sufficiently generic patterns will be detectable. The simplest observable to examine in such survey data is the list of species that are present or absent in each sample. We obtained these presence/absence vectors from the simulations of Fig 2, and found that when we sorted species by prevalence (rows in Fig 6A) and samples by richness (columns in Fig 6A), the community composition generically exhibited a nested structure—less diverse communities tended to be subsets of more diverse communities [41, 42]. We quantified this result using an established nestedness metric, as described in S1 Text and S7 Fig., and found that the actual nestedness exceeds the mean value for a randomized null model by more than 100 standard deviations. This suggests that nested structures may generically emerge in community assembly through the interplay of stochastic colonization, competition, and environmental filtering.
Next, we asked if we large-scale beta-diversity patterns could be used to distinguish the resource-limited and diverse regimes. We initialized 200 new communities with 100 randomly chosen members from the full regional species pool and simulated these communities to steady state in both the resource-limited and diverse regimes (see S1 Text Section 2B for details). This sub-sampling of the full regional species pools mimics the effect of stochastic colonization, where a different random subset of species seeds each community. To better understand beta-diversity signatures in the two regimes, we performed a Principal Component Analysis (PCA) on community composition and projected the data onto the first two principal components, as shown in Fig 6B–6D. In the resource-limited regime, the communities form distinct clusters that are distinguished by different highly abundant species. This suggests that harsh environments only allow a few species from the regional pool to rise to dominance, and that these dominant species induce clustering of communities. Such “enterotype”-like behavior is a common feature observed in many microbial settings [43]. In contrast, the diverse regime exhibited neither well-defined clusters nor dominant, highly abundant species.
The preceding results suggest that the resource-limited and diverse regimes can be distinguished using beta-diversity patterns. Rigorous testing of this prediction is beyond the scope of the present work. But as an illustration of the kind of data we hope to explain, we examined the natural gradient of solar energy supply in the Tara Oceans survey, which collected microbial community samples from a range of depths across the world’s oceans [36]. Explicitly including light as an energy source would require some modification to the structure of the model equations, but we expect that the large-scale features of sufficiently diverse ecosystems should not be sensitive to changes involving just one resource. We analyzed the 16S OTU composition of tropical ocean communities for all 30 sea-surface samples, where solar energy is plentiful, and all 13 samples from the deep-sea mesopelagic zone where energy fluxes are limited. We projected these composition vectors onto their first two principal components as in Fig 6 above, and plot the results in Fig 7. The sea surface data superficially resembles our diverse regime, with a relatively uniform distribution of possible community compositions. In contrast, the Mesopelagic Zone is similar to our resource-limited regime: the dominance of the most abundant species is much more pronounced, and the compositions appear to cluster into four discrete types. While these results are consistent with our model predictions, the number of samples at each depth is still too small to draw any definitive conclusions about clustering.
As mentioned above, our model also gives a natural explanation for the nestedness in the Earth Microbiome Project community composition data [1], suggesting that it may be a natural byproduct of complex microbial communities shaped by competition, environmental heterogeneity, and stochastic colonization. To test how generic this feature is, we plotted presence/absence community compositions of all samples from the Tara Oceans dataset, sorting samples by richness and OTU’s (“species”) by prevalence. Each sample contains thousands of low-abundance OTU’s, which can obscure ecological patterns through their susceptibility to sequencing noise and transient immigration. We therefore imposed a 0.5% relative abundance threshold for an OTU to count as “present.” The resulting pattern in Fig 7 is qualitatively similar to our simulations (Fig 6D), and to the phylum-level data of the Earth Microbiome Project [1], with the region below the diagonal significantly less populated than the region above the diagonal (although the signal is much weaker). In S1 Text Section 4 and S7 Fig., we quantify the nestedness using the same metric employed in the Earth Microbiome Project analysis [1, 44], and show that the score is significantly higher than the mean scores from two standard null models.
Advances in sequencing technology have opened new horizons for the study of microbial ecology, generating massive amounts of data on the composition of both natural and synthetic communities. But the complexity of these systems make it difficult to extract robust general principles suitable for guiding medical and industrial applications. Numerical simulations provide a powerful tool for addressing this problem. By rapidly iterating numerical experiments under a variety of modeling choices with random parameters, one can identify robust patterns and use the resulting insights to guide targeted experiments.
Following this strategy, we developed a thermodynamic consumer resource model that explicitly includes energetic fluxes and metabolically mediated cross-feeding and competition. Using this model, we identified two qualitatively distinct regimes as we varied the amount of energy supplied to ecosystem and the fraction of energy leaked back into the environment: a low diversity “resource-limited” regime and a “diverse” regime. The structure of the resource-limited regime is strongly constrained by species- and community-level environmental filtering. Each community is dominated by a handful of species, making the community properties sensitive to the idiosyncratic characteristics of these species and susceptible to environmental fluctuations. In the diverse regime, communities exhibit more universal features because they substantially engineer their environments. In particular, the concentrations of resources at steady state are more narrowly distributed and insensitive to perturbations in the external supply rates. Moreover, each species draws its energy roughly equally from all resources, rather than subsisting on the externally supplied resource as in the resource-limited regime.
The emergence of environmental engineering from this community-scale model makes it a valuable tool for testing and refining existing conceptual frameworks employed by empirical biologists [45]. A major limitation of the dominant paradigms for evolution and ecology from the last century is the implicit assumption of a constant environment [46]. The generalized Lotka-Volterra model, for example, remains a standard lens for reasoning about ecological dynamics, both quantitatively and qualitatively [47–49]. It assumes that the dynamics emerge from the sum of pairwise interactions among species, and that the sign and strength of these interactions are intrinsic properties of the species. This can be a good assumption in some circumstances [47, 48], but fails to accurately describe the behavior of simple models that explicitly account for the state of the environment [50]. Our work provides a starting point for determining the conditions under which pairwise models will generically succeed or fail in describing the behavior of large ecosystems.
Our model complements other recent efforts at understanding microbial community ecology. Taillefumier et al. proposed a similar model of metabolite exchange, and focused on the case where the number of resource types M is equal to 3 [21]. In this case, repeated invasion attempts from a large regional species pool produced optimal combinations of metabolic strategies. Goyal et al. examined the opposite limit, with M = 5, 000, but allowed each species to consume only one type of resource [22]. This generated communities with a tree-like metabolic structure, where each species depends directly on another species to generate its unique food source. In our model, the large number of resource types (M = 100 in the current study) makes spontaneous strategy optimization extremely unlikely. And our generic protocol for sampling the metabolic matrix Dαβ allows a variety of community-level energy flux topologies to emerge, as illustrated in Fig 3, which can sometimes be quite different from the tree networks of Goyal et al. The absence of highly specialized metabolic structure in our model makes it especially well-suited for interpreting patterns in large-scale sequence-based datasets such as the Earth Microbiome Project [1].
Our model predictions can also be directly tested using experiments with natural communities in synthetic laboratory environments [17, 51]. Our model predicts that beta-diversity patterns and community-level metabolic networks can be significantly altered by increasing the ecosystem’s energy supply, inducing a transition from the resource-limited to the diverse regime. In the experimental set-up of [51], this can be done by directly adding chitinase enzymes to the sludge reactor to increase the degradation of chitin-based organic particles on which the ocean-derived microbial communities subsist. One could then look for shifts in the resulting diversity patterns, and observe any changes in the topology of the metabolic flux network using isotope labeling.
In this work we have largely confined ourselves to studying steady-state properties of well-mixed microbial communities. Microbial communities often exhibit complex temporal dynamics with well-defined successions [51–53]. It will be interesting to explore these dynamical phenomena using our model. It is also well established that spatial structure can give rise to new ecological phenomena [54, 55] and an important area of future work will be to better explore the role of space in microbial community assembly.
Finally, we have obtained strong numerical evidence that the two regimes are separated by a phase transition, which is likely closely related to disorder-induced phase transitions in statistical physics [32]. In Supporting Text Section 5, we examine the steady-state richness in the three examples of Fig 2 under increasing values of M from M = 40 to M = 560. We find that the richness is proportional to M in the diverse regime, but scales sub-linearly with M in both examples from the resource-limited regime. In the M → ∞ limit, therefore, we expect to find a sharp line between the regimes, with the ratio of the richness to M vanishing on the resource-limited side. But we do not yet know the exact location of this boundary, or the critical exponents describing the behavior of the system near the transition.
|
10.1371/journal.pbio.1001497 | Indirect Evolution of Hybrid Lethality Due to Linkage with Selected Locus in Mimulus guttatus | Most species are superbly and intricately adapted to the environments in which they live. Adaptive evolution by natural selection is the primary force shaping biological diversity. Differences between closely related species in ecologically selected characters such as habitat preference, reproductive timing, courtship behavior, or pollinator attraction may prevent interbreeding in nature, causing reproductive isolation. But does ecological adaptation cause reproductive incompatibilities such as hybrid sterility or lethality? Although several genes causing hybrid incompatibilities have been identified, there is intense debate over whether the genes that contribute to ecological adaptations also cause hybrid incompatibilities. Thirty years ago, a genetic study of local adaptation to copper mine soils in the wildflower Mimulus guttatus identified a locus that appeared to cause copper tolerance and hybrid lethality in crosses to other populations. But do copper tolerance and hybrid lethality have the same molecular genetic basis? Here we show, using high-resolution genome mapping, that copper tolerance and hybrid lethality are not caused by the same gene but are in fact separately controlled by two tightly linked loci. We further show that selection on the copper tolerance locus indirectly caused the hybrid incompatibility allele to go to high frequency in the copper mine population because of hitchhiking. Our results provide a new twist on Darwin's original supposition that hybrid incompatibilities evolve as an incidental by-product of ordinary adaptation to the environment.
| Adaptive evolution by natural selection is the primary force generating biological diversity. A critical question is whether the evolution of hybrid incompatibility, which is essential for the maintenance of species diversity, is caused by adaptive evolution. In this article, we investigate one of the most widely cited examples of ecological divergence driving the evolution of reproductive incompatibility, the strong association between hybrid lethality and copper tolerance in a copper mine population of the wildflower Mimulus guttatus. Hybrid lethality and tolerance of high levels of copper co-segregate as a single Mendelian locus. While copper tolerance and hybrid lethality are nearly universal in the mine population at Copperopolis, California, they are absent from adjacent off-mine populations, suggesting that reproductive isolation evolved rapidly as a pleiotropic by-product of recent adaptation to the mine environment. We find that copper tolerance and hybrid lethality are controlled by distinct loci, in tight genetic linkage. We also demonstrate that this genomic region has experienced strong recent selection and conclude that ecological selection for copper tolerance indirectly caused the neighboring hybrid lethality allele to hitchhike to high frequency. To our knowledge, this is the first case to demonstrate that reproductive isolation factors can evolve as an incidental by-product of adaptation to novel environments through genetic hitchhiking.
| Adaptation to local environmental conditions by natural selection is the primary cause of evolutionary change in natural populations. Ecological adaptation can cause reproductive isolation when selection acts on traits that influence the likelihood of intermating in nature, such as habitat preference, reproductive timing, courtship behavior, or pollinator attraction [1],[2]. However, it is controversial whether adaptation to local environmental conditions promotes the development of reproductive incompatibilities such as hybrid sterility or lethality. As Darwin [3] first discussed, reproductive incompatibility should not evolve directly via natural selection (but see [4]) and would only arise as an incidental by-product of interspecific divergence. Darwin realized that reproductive isolation could evolve if it is not expressed as lineages diverge, perhaps due to populations inhabiting geographically distinct regions, and it would only manifest in cases in which these populations hybridized subsequent to divergence. Classic models of the evolution of hybrid incompatibility, independently developed by Bateson [5], Dobzhansky [6], and Muller [7] (BDM), predict that alleles at different loci may accumulate within distinct lineages, and although they may be neutral or adaptive in an ancestral population, they will produce deleterious interactions when brought together by hybridization. This model is now thoroughly supported by genetic mapping studies of hybrid incompatibility loci. The role of natural selection in driving the evolution of incompatibility alleles continues to be an area of rich investigation [8]–[10].
Selection may drive the evolution of BDM incompatibility alleles in a variety of ways. With the molecular genetic dissection of multiple hybrid incompatibility systems, researchers have determined that many of the underlying loci have experienced rapid evolutionary divergence consistent with natural selection [9],[10], but there is little evidence of what may be causing this rapid divergence. Evidence for the traditional view that hybrid incompatibility evolves as a by-product of adaptation to divergent environments is supported by only a few lab-based experimental evolution studies [11]–[13]. Of the hybrid incompatibility genes identified in natural populations, it seems unlikely they function in classical ecological adaptation. Instead researchers have speculated that these incompatibilities evolved due to intragenomic conflict [14],[15]. For example, the incompatibility Overdrive locus causes both segregation distortion and hybrid male sterility in Drosophila psuedoobscura [16]. Ultimately, it is considerably difficult to determine the evolutionary forces acting upon hybrid incompatibility alleles during the process of divergence because these studies focus on species reproductively isolated for thousands of generations.
In this study, we investigate whether recent adaptation to a copper mine habitat by the wild flower Mimulus guttatus caused the evolution of hybrid incompatibility. Extreme soil environments, either naturally occurring or created anthropogenically, characterized by low nutrient and high heavy-metal concentrations, impose strong selection on plant communities [17],[18]. M. guttatus has colonized multiple mine sites in the Sierra Nevada foothills region of central California, the largest of which is located at the Keystone Union mine complex in the town of Copperopolis [19]. A survey of tolerance to high copper concentration in plants from Copperopolis and adjacent off-mine populations suggests this phenotype is under strong selection in the mine habitat. Copper tolerance is nearly fixed in the mine population (99.77%, N = 2,796) and segregating at very low frequency in a sample of 15 off-mine populations located with 40 km of Copperopolis (8.7%, N = 1,440) [20]. An investigation into the genetic basis of this trait initiated over 30 years ago identified a locus that appeared to cause both copper tolerance and hybrid lethality in crosses to other off-mine populations [21],[22]. Copper tolerance is controlled by a dominant allele at a single Mendelian locus, when measured as a threshold character in lab-based root growth assays. Crosses between tolerant lines from a recombinant backcross population (Figure S1) and the off-mine population, Cerig-y-drudion, Wales, United Kingdom (hereafter referred to as Cerig), yielded various levels of F1 hybrid lethality, whereas crosses between nontolerant lines from the same recombinant backcross population always produced viable offspring [22]. Additional crosses demonstrated there is a single incompatibility locus in the Copperopolis population and that variation in F1 hybrid lethality is caused by multiple loci segregating in the Cerig population [22]. Cerig was used in the initial crosses by M. R. Macnair and continues to be a focal population in this study, because it was the first population identified to produce inviable hybrid offspring in crosses with Copperopolis [22]. Subsequent experiments have demonstrated that the Cerig hybrid lethality factors are geographically widespread and segregate in three California populations. In contrast, the Copperopolis lethality factor is geographically restricted; no incompatible plants were identified in test crosses with plants from 15 other populations (N = 10 plants/population) [23]. The lethality factor is at high frequency in the Copperopolis population; 16 Copperopolis genotypes (five of which are from this study, see details below) have been tested and all of them produce lethal offspring when crossed to Cerig tester lines [22],[23]. Hybrid lethality in this system is consistent with the BDM model and was not observed to be genetically de-coupled from the copper tolerance phenotype [22],[24]. One attractive hypothesis is that they are controlled by the same locus and F1 hybrid lethality evolved as a pleiotropic by-product of the locally adaptive copper tolerance. An alternative hypothesis is that the two phenotypes are not pleiotropic and are instead controlled by two distinct loci in tight genetic linkage.
The physiological manifestation of BDM hybrid incompatibility in this system reveals possible genetic mechanism for inviability, while maintaining support for both the pleiotropy and linkage hypotheses. Hybrid lethality manifests in F1 plants as tissue yellowing and death in the early stages of development and prevents the plants from reaching reproductive maturity [22],[24]. This lethality is similar to the hybrid necrosis phenotype known for many years in the plant agriculture literature [25]. Recently, it has been demonstrated that hybrid necrosis is an autoimmune-like response due to negative interactions between divergent pathogen-resistance proteins, often involving nucleotide-binding leucine-rich repeats (NB-LRRs) and their interacting partners [26],[27]. Although there is no support for NB-LRR's dual function in disease resistance and trafficking of heavy-metal ions, this could be the case for their partner proteins. NB-LRR protein–protein interactions are difficult to examine, and most partners remain unknown [28]. Additionally, there is strong evidence of pleiotropy between disease resistance and heavy metal tolerance from an independent class of pathogen-resistance proteins, defensins, which contribute to enhanced zinc tolerance in Arabidopsis halleri [29]. It remains to be determined whether selection on the copper tolerance allele in M. guttatus resulted in the near fixation of a hybrid incompatibility factor because these phenotypes are the result of pleiotropy at a single locus.
The primary aim of this study is to investigate the genetic basis of the copper tolerance and hybrid necrosis phenotypes to determine whether they are pleiotropic or controlled by distinct genes in tight genetic linkage. We also use population genetic approaches to infer the nature of natural selection acting on this copper tolerance and hybrid necrosis genomic region.
To determine whether copper tolerance and hybrid lethality are controlled by the same locus, we conducted a high-resolution genetic mapping experiment. Our mapping approach focused on first identifying the location of copper tolerance locus because, of the two phenotypes, it is easier to measure, and then screen a subset of plants with recombination events near the tolerance locus for hybrid inviability. In order to genetically map these phenotypes with high resolution, we used a near isogenic line (NIL) created through recurrent backcrossing of a Copperopolis by Cerig35 F1 (this line is compatible with Copperopolis) to lines derived from Stinson Beach, California. For seven generations, only tolerant plants were selected to be backcrossed to Stinson Beach (Figure S1). The Stinson Beach population is nontolerant and lacks any intrinsic incompatibilities with Copperopolis [23]. We used multiple outbred Stinson Beach lines to avoid the development of inbreeding depression in our backcross lines. Each backcross generation segregated tolerant and nontolerant progeny in equal ratio, in accordance with predictions for a single, dominant tolerance locus. We created our mapping population by crossing a single tolerant (BC7T) and a single nontolerant (BC7NT) line to produce F1BC7 progeny (Figure S1). We set about identifying the introgressed tolerance region in the NIL lines (BC7T, BC7NT) by screening 541 genetic markers distributed across the genome (mimulusevolution.org). We identified 42 markers that were homozygous in the BC7NT line and heterozygous in the BC7T line, possibly indicating a region of introgression from the tolerant parent. To establish which heterozygous markers are linked to the copper tolerance locus, we tested for an association between phenotype and genotype in 80 F1BC7 lines. We measured copper tolerance in the F1BC7 plants as a threshold character in accordance with previously published hydroponic methods [20]. We found the genotypes of one marker, MgSTS242, located on linkage group (LG) 9, were very tightly correlated with tolerance phenotype (Figure 1). We henceforth refer to this copper tolerance locus as Tol1.
To fine map the genomic position of Tol1, we created a mapping population of 4,340 F1BC7 plants, phenotyped each plant for copper tolerance, and genotyped each plant with the MgSTS242 marker. Initially, we identified 83 recombinant plants, however we repeatedly phenotyped and genotyped these lines and we were able to confirm 36 true recombinant plants, indicating that this marker is 0.83 cM from the Tol1 (Figure 2). To fine map Tol1, we designed additional genetic markers near MgSTS242 using the currently unpublished M. guttatus genome assembly (www.phytozome.net). This assembly is not contiguous on the scale of chromosomes; the genome is instead composed of many, relatively small, genomic scaffolds that do not assemble into large chromosome-length blocks. The assembly in the Tol1 genomic region is particularly poor because this is a highly repetitive pericentromic region of LG 9 and none of the scaffolds that map to this region in the reference are longer than 1 MB. We attempted to determine the location of 10 target scaffolds relative to MgSTS242 and Tol1 by designing multiple new markers per scaffold and screening them in our recombinant lines. We identified three scaffolds that are in tight linkage to Tol1 (Figure 2). We determined that a marker on scaffold 84 is in tightest association with Tol1; we found 14 plants have a recombination event between Sc84_37kb and Tol1, indicating this marker is located within 0.32 cM of the tolerance locus (Figure 2). We identified three additional scaffolds that map to this region (97c, 157, and 238), but they do not reside within the Sc84_37kb–Tol1 interval (Text S1). We have not included these scaffolds in our map because we have a limited amount of genetic material and we could not test all 36 recombinants. We attempted to identify a flanking marker for Tol1 by designing markers in the 3′ end of scaffold 63b, which is predicted to be adjacent to the 5′ end of scaffold 84 (unpublished data, Uffe Hellsten, Joint Genome Institute, Walnut Creek, CA). However, we found no evidence of linkage between our markers and the tolerant phenotype, likely indicating that scaffold 63b is outside of the NIL introgression region or there are errors in the reference genome assembly (Figure 2 and Text S1). We observed the same pattern of no linkage between Tol1 and markers located in two additional scaffolds (460 and 925) that are predicted to be located in this region in the reference genome assembly.
In order to determine if the tolerance and incompatibility phenotypes are controlled by the same locus, we crossed 18 (nine tolerant and nine nontolerant) recombinant plants to an incompatible genotype, Cerig10 (Figure S2). Crosses were conducted in each direction to determine if there is asymmetry in the incompatibility. We measured hybrid inviability as the percentage of offspring with a majority of tissue being yellow or necrotic in a block of 60 plants (Figure 3A and 3B). We scored hybrid inviability in 2–8 replicate blocks per line. We used Cerig10 as our tester line because this line produced the highest level of hybrid lethality in crosses to Copperopolis lines. In crosses between Cerig10 and five Copperopolis genotypes, we found an average of 83% of the offspring were inviable (Figure S2).
Our F1BC7×Cerig10 crosses produced a bimodal distribution of hybrid inviability. The majority of crosses with nontolerant lines produced 5%–15% inviable progeny, whereas the majority of crosses with tolerant lines produced 45%–62% inviable progeny (Figure 3C and 3D). We identified two plants, 25_E01 and 25_E11, with recombinant phenotypes (Figure 3C). Hybrid inviability of nontolerant line 25_E01 is significantly elevated compared to all nontolerant control lines (Wilcoxon test, z = 2.46, p<0.014; Table S1) except 46_B12, which had only two replicate measurements (Wilcoxon test, z = 1.84, p = 0.065; Table S1). Hybrid inviability of tolerant line 25_E11 is significantly lower than the two tolerant control lines (Wilcoxon test, z = 2.46, p<0.014; Table S1). This establishes that copper tolerance and hybrid inviability are controlled by two distinct, but tightly linked, loci and that hybrid necrosis is not a pleiotropic by-product of copper tolerance. We henceforth refer to the hybrid inviability locus as Nec1.
To map the genomic location of Nec1, we compared the Tol1 genotyping information for these lines to our lethality data and determined that Nec1 maps to scaffold 84. To fine map Nec1, we designed markers in this region and genotyped plants with informative recombination events. We found that the incompatibility locus maps to an intergenic region between 283 kb and 293 kb on scaffold 84 (Figure 4). This region in the M. guttatus genome (www.phytozome.net) contains a single Gypsy3 transposable element (TE) (Figure 4). We are prevented from continuing to narrow this region, despite identifying three lines (26_C07, 25_E01, and 48_B01) with recombination events within this locus (Figure 4) because of the repetitive nature of this element. We attempted to sequence across this genomic interval using long-range PCR, but these attempts were unsuccessful. Our long-range PCR may have failed because repetitive DNA in this TE element or because this genomic interval is much larger in the Copperopolis genome compared to the reference genome.
Nec1 maps to an intergenic region containing a Gypsy3 TE. Unregulated TE replication can cause hybrid sterility in Drosophila; in this system, hybrid sterility manifests in crosses in a single direction because hybrids lack the maternally inherited TE replication-suppression mechanisms [30]. It is unlikely that this is the cause of inviability in our system because significant levels of hybrid lethality occur in both crossing directions, although it is higher when Cerig is the paternal parent (Table S2). Instead, the functional changes at Nec1 likely result from altered gene expression or gene function.
The Copperopolis allele at Nec1 may cause hybrid lethality by altering the cis-regulatory expression of a gene within this region, and the functional gene may flank our mapped interval or reside even further away along the chromosome. The nearest gene, MGV1A004323M, resides 5 kb from Nec1 and encodes a glycosyltransferase metabolic enzyme (www.phytozome.net). This protein has greatest homology to A. thaliana gene AT3G18170, which attaches glyosyl groups to N-linked glycan molecules critical for construction of plant cell membranes [31]. The second flanking gene, MGV1A026627M, is located 41 kb from Nec1 and encodes a Jumonji-C histone demethylation protein. The Jumonji-C proteins are part of the histone demethylase family, whose function reverses the epigenetic silencing of transposons and genes in eukaryotic genomes [32]. Neither of these genes, nor is any gene within 100 kb of Nec1, is a known participant in plant disease-resistance pathway. Alternatively, Nec1 may involve another gene that is absent from the reference genome. This gene may have undergone recent functional or expression changes to cause negative interactions with a member of the plant disease-resistance pathway, producing the hybrid necrosis phenotype. This hypothesis is especially attractive because the NB-LRRs often exhibit large variation in copy number within species [33]. Although additional work is required to distinguish between the possible mechanisms of Nec1 function, our results clearly demonstrate that hybrid inviability did not evolve as a pleiotropic by-product of the evolution of copper tolerance. What evolutionary force caused the lethality allele at Nec1 to go to high frequency in the Copperopolis population?
The hybrid lethality phenotype reached high frequency in the Copperopolis population because of genetic drift or natural selection. Genetic drift may have been the primary evolutionary mechanism causing the lethality allele at Nec1 to rise to high frequency if the initial population of mine colonists experienced a bottleneck event. This scenario would cause a genome-wide reduction in genetic variation in the mine population relative to neighboring off-mine populations [34]. Alternatively, the hybrid lethality allele at Nec1 may have been directly selected because this allele confers some fitness advantage in the mine habitat. A variant of the selection hypothesis is that the lethality allele at Nec1 may have hitchhiked to high frequency because of selection on the tightly linked Tol1 locus. In this case, the tolerance and incompatibility alleles would have resided on the same haplotype in the founders of the mine population. Selection on this genomic region would leave a molecular genetic signature of reduced genetic variation in the mine population and increased genetic differentiation between mine and off-mine populations in this genomic region, compared to the rest of the genome [35]–[38].
Studies that attempt to identify these molecular signatures of selection in natural populations have been used to identify putatively selected loci in many different systems [39]–[44]. The power of population genetic studies to detect signatures of selection can be diminished by multiple interdependent processes: selection on standing genetic variation [45], population structure [46], and time since a selective sweep [34]. These studies can most readily identify a molecular signature of selection in the classic case of a hard sweep, in which strong selection on a new mutation causes a single haplotype to rapidly go to fixation and tightly linked alleles hitchhike to high frequency [34]–[38]. However, if this mutation is segregating in the ancestral population and has recombined onto multiple haplotypes prior to selection, hitchhiking of tightly linked alleles is reduced [45]. This event, termed a “soft sweep,” will not decrease genetic variation or increase population differentiation at linked alleles to the same degree as a hard sweep. The number of generations since a selected allele goes to high frequency affects the molecular signature of selection because each generation there is new opportunity for mutation, migration, and recombination to add genetic variation to the derived population [34]. The young age of the mine habitat (150 y) suggests that if a sweep occurred, the molecular signature of this event would still be apparent. The mine and off-mine populations are in close geographic proximity, and there is opportunity for gene flow, suggesting there will be little effect of population structure. Gene flow will increase the amount of shared genetic variation at selectively neutral sites in the genome, but strong selection in the mine environment could maintain the original haplotype (or haplotypes) containing the tolerance allele in the face of gene flow.
In order to evaluate the selection hypothesis, we compared genetic variation at markers linked to Tol1 and markers randomly distributed throughout the genome. We sampled individual plants, collected as seed from unique maternal families in 2005 and 2007, from the mine population at Copperopolis (N = 108) and from two off-mine populations: O'Byrnes Ferry Road (N = 33), located 2 km from Copperopolis, and Hunt Road (N = 39), located 9 km from Copperopolis (Figure S3). The geographic proximity of these populations and recent divergence of the mine population, a maximum of 150 generations, suggests any signal of selection will not be lost in a haze of population structure. Furthermore, this population sampling is designed to explicitly test for an effect of habitat-mediated selection on a single genomic region with known phenotypic effect, Tol1/Nec1 loci, and not a genome-wide analysis. Lastly, this design will not be confounded by the effect of hybrid inviability because these off-mine populations lack the incompatible allele found in the Cerig population [23].
We measured genetic variation of eight loci in the Tol1 fine mapped region (Tol1Link) and 11 unlinked loci (Tol1UnLink) for the Copperopolis and two off-mine populations. The Tol1UnLink markers serve as a control to estimate the average level of genetic variation and population differentiation. We used co-dominant DNA fragment length polymorphism markers [47] to estimate the within-population genetic variation, which we report as number of alleles (Na) and expected heterozygosity (He) [48]. We report both values, but these are often correlated when using fragment length markers. We calculated genetic differentiation between populations as Fst for two marker classes: Tol1Link and Tol1UnLink using Fdist [36] implemented in the program LOSITAN [49]. To evaluate whether Fst is significantly elevated in the Tol1 fine mapped region, we compared the observed levels of Fst at the Tol1Link markers to a null distribution of Fst generated from coalescence simulations of population divergence using the Tol1UnLink dataset in the program LOSITAN. Our hypothesis is that Fst for the Tol1Link markers will be significantly higher than the Tol1UnLink markers. The coalescence simulations assume two populations and use Wright's symmetrical Island Model of migration [50] to calculate the probability that two randomly chosen chromosomes in a population have a most recent common ancestor within that population without an intervening migration or mutation event [36],[49]. This model also assumes that the two populations are at equilibrium with constant effective population size. Our populations likely violate this assumption because the mine population was only recently established. However, Beaumont and Nichols found that a recent colonization event had no significant effect on their estimates of the Fst distribution [36].
We observe low Fst for all Tol1UnLink markers, and for all but a single marker, there is no significant differentiation for the three pairwise population comparisons (Table S3). Using the Tol1UnLink dataset, we also estimate the relatedness between the three populations using the program STRUCTURE version 2.3 [51]. In this analysis, the two off-mine populations cluster more tightly with each other than with Copperopolis (Figure S4); thus, we combined the two off-mine populations into a single sample for subsequent analyses.
We found mine and off-mine populations have no difference in Na and only slightly reduced He for the Tol1UnLink markers (Table 1). This suggests that the mine population did not undergo a dramatic population bottleneck during colonization, or if it did, this signal has been erased by ongoing gene flow. Conversely, we find that Na and He of the Tol1Link markers is sharply reduced in the Copperopolis population compared to the off-mine population, and this pattern is strongest at the four markers in tightest linkage to Tol1 and Nec1 (Table 1). We found that the four markers that are most tightly linked to Tol1 and Nec1 have significantly elevated Fst when compared to the Tol1UnLink distribution (Figures 5 and S5). Fst values varied between the four markers; the markers with the two highest values are Sc84_364kb (Fst = 0.546) and Sc84_37kb (Fst = 0.317) are in tightest linkage to the Nec1 and Tol1 loci, respectively. These data support the two predictions arising from the selection hypothesis: low genetic variation within Copperopolis and high differentiation between populations at markers linked to the tolerance and hybrid lethality loci.
We demonstrate that copper tolerance is not pleiotropic with hybrid inviability. Instead, both phenotypes are controlled by distinct loci that are in tight genetic linkage. We genetically mapped the copper tolerance locus to a highly repetitive pericentromeric genomic region. The large amount of repetitive DNA and relatively rare euchromatic sequence has inhibited the assembly large genomic scaffolds in this region of M. guttatus reference genome (unpublished data, Uffe Hellsten, Joint Genome Institute, Walnut Creek, CA). Our attempts to fine map the Tol1 locus to a genomic scaffold, or even identify a flanking marker, have been stymied by the complex nature of this genomic region despite our large mapping population. In contrast, we were able to fine map the hybrid lethality locus to a 10 kb intergenic region with a contiguous genomic scaffold. The Copperopolis allele at Nec1 may cause hybrid lethality by altering the cis-regulatory expression of a neighboring gene, or this region in the Copperopolis genome may harbor genes that are missing in the reference M. guttatus genome. We attempted to amplify the 10 kb region from the Copperopolis genome using long-range PCR, but these attempts were unsuccessful. The failed PCR attempts may be caused by repetitive DNA within the gypsy-3 TE or because this region may be expanded or rearranged in the Copperopolis genome compared to our reference.
We observed a large amount of variance between independent replicates of the high-lethality lines (Figures 3C, 3D, and S2). This variance is likely caused by two factors—segregating variation at multiple incompatibility loci in our tester line, Cerig10, and variation in temperature in the greenhouses in which the F1BC7×Cerig10 offspring were grown. There are multiple incompatibility loci segregating within the Cerig population that interact with the Copperopolis allele at the locus we have now identified as Nec1 [22],[23]. Our tester line, Cerig10, was chosen because it gave the highest level of incompatibility, but it may be heterozygous at some incompatibility loci, inflating variation in hybrid lethality. An additional source of variation may be the temperature at which the offspring were grown; research on hybrid necrosis in A. thaliana has demonstrated that this phenotype manifests at 16°C but is absent at 23°C [27]. Our plants were grown under controlled greenhouse conditions, but daily or seasonal temperature fluctuations may have exceeded the capability of the greenhouse to maintain constant temperature.
We find strong support for the hypothesis that adaptation to the mine habitat imposed selection on the Nec1/Tol1 genomic region in the Copperopolis population. The high level of genetic variation in the Tol1UnLink dataset in the Copperopolis population argues against a dramatic bottleneck event during mine colonization. The four markers in tight linkage to Tol1 and Nec1 have reduced Na and He in the mine population, as well as significantly elevated Fst. We interpret these data as evidence of a sweep at Tol1, driven by selection for copper tolerance, causing the Nec1 BDM incompatibility allele to hitchhike to high frequency in the mine population. This model predicts that the marker in tightest linkage to Tol1, Sc84_37kb, would have the highest Fst and sharpest reduction in He. Contrary to this prediction, we found that marker Sc84_364kb, the marker nearest Nec1, has the highest Fst and sharpest reduction in He. This finding suggests that adaptation to the mine environment was more complex than the classic hard selective sweep at Tol1. Although we cannot completely rule out the possibility that the incompatibility allele at Nec1 has also experienced selection in the mine population, the physiological effect of the tolerance allele and the biogeographic distribution of copper tolerant plants, which are nearly fixed in multiple mine populations and are segregating at 8.3% of neighboring off-mine populations, is strong evidence that copper tolerance is under selection in the mine habitat [21]. Because the lethality allele at Nec1 has only been found in a single, very recently derived population [23] and has no discernible phenotypic effect beyond causing hybrid lethality, it is less likely this allele is experiencing selection in the mine environment.
We suggest multiple factors that may explain why the patterns of He and Fst we observe do not conform to the predictions of a classic hard selective sweep. Variation in the mutation rate between markers Sc84_37kb and Sc84_364kb may explain the differences in He and Fst. Additionally, we note that we have been unable to fine map the precise location of Tol1 because this is a highly repetitive pericentromeric region and the Sc84_367kb marker may actually be closer to Tol1 in the Copperopolis genome than its position in the reference genome suggests. Finally, we suggest that the tolerance allele at Tol1 may not have experienced a hard selective sweep. If the tolerance allele was segregating at low frequency in the ancestral population and experienced a soft selective sweep, tightly linked sites would be affected differently depending on the amount of genetic variation at these sites in the ancestral population and the level of linkage disequilibrium (LD) between the selected allele and tightly linked sites when selection first acted upon this population [45]. Tolerant plants are segregating at 8.3% in off-mine populations [23], and although we are unable to determine whether this is caused by segregating ancestral variation or recent migration from mine to off-mine populations, it indicates that the tolerant allele may predate the development of the copper mines and recently rose to near fixation via a soft selective sweep. We investigate whether a soft sweep at Tol1 could have caused the lethality allele at Nec1 to hitchhike to high frequency using a quantitative model of genetic hitchhiking.
To simulate hitchhiking on tightly linked sites following a hard or soft selective sweep, we use the two-locus model of genetic hitchhiking described by Maynard Smith and Haigh [35] and Barton [38]. In the classic model of a hard selective sweep, the beneficial allele resides on a single haplotype and the selected and linked alleles are initially in complete LD. Because copper mining was initiated in the 1860s, we focused on strong selection situations that would cause a beneficial allele to become nearly fixed within 150 generations. We model the effects of hitchhiking on neutral allele segregating at low frequency (5%) in the ancestral population because the hybrid lethality allele has not been found in neighboring off-mine populations [23]. Additional details of our hitchhiking model are provided in Text S1. Consistent with previous findings [38], we find that during a hard selective sweep, tightly linked alleles within 0.5 cM hitchhike to high frequency in the derived population (Figure S6A, S6D, and S6G). We next model the soft sweep scenario, in which the selected allele is segregating at low frequency (5% and 10%) in the ancestral population. We find that a soft selective sweep can cause sites within 0.5 cM of the beneficial allele to hitchhike to high frequency for a wide range of parameter values (Figure S6B, S6C, S6E, S6F, S6H, and S6I). The primary determinant of hitchhiking effects is the initial level of LD between the selected and neutral alleles. In the hard sweep model, LD is 1, however this parameter can vary when the focal beneficial mutation and a tightly linked allele are segregating in the ancestral population. We find that when we set initial levels of LD to high levels (0.8–1.0), a tightly linked allele within 0.5 cM can hitchhike to high frequency in the derived population (Figure S6). Overall, these findings demonstrate that for this situation, in which a M. guttatus population has rapidly adapted to the copper mine environment, a soft or hard selective sweep can produce large hitchhiking effects on tightly linked sites that happen to be in strong LD in the ancestral population. Although this hitchhiking model makes assumptions that are likely violated in natural populations—deterministic selection on an allele, bi-allelic loci, and constant population size—it provides a useful quantitative framework to describe the effects of genetic hitchhiking on loci in tight genetic linkage.
The variation we observe in of He and Fst between Tol1Link markers suggests that adaptation in the mine environment may have occurred in a manner more complex than the classic hard sweep model. A soft selective sweep may not uniformly reduce genetic variation or increase genetic differentiation at tightly linked sites if there were differences in the amount of ancestral genetic variation at these sites or variation in the levels of LD between the selected allele and linked sites in the ancestral population [45]. In a soft sweep, the primary determinant of the strength of hitchhiking effects is the initial level of LD between the selected allele and tightly linked alleles. Unfortunately, the extent of LD in the ancestral population for this pericentromeric region is not known. Although we cannot conclusively state whether there was a hard or soft sweep at the Tol1/Nec1 region, we do find strong evidence of selection on this region in the mine population, and the results of our hitchhiking simulations of a hard or soft sweep are consistent with the hypothesis that selection at Tol1 caused the hybrid lethality allele at Nec1 to hitchhike to high frequency in mine population.
In summary, we have determined that the copper tolerance and hybrid inviability phenotypes are controlled by two distinct, tightly linked loci. Hybrid inviability is not a pleiotropic by-product of adaptation to the mine environment, as originally proposed [22]. We find that Tol1 maps to a highly repetitive, unassembled region of the M. guttatus genome, whereas Nec1 maps to a 10 kb region containing a single TE. Furthermore, we demonstrate that there has been strong positive selection on the Nec1/Tol1 genomic region. We interpret these data as evidence that the Nec1 lethality allele rose to high frequency because of genetic drift induced by selection on the tightly linked Tol1. We demonstrate that natural selection on one locus can cause a tightly linked hybrid lethality allele to hitchhike to high frequency, providing empirical evidence for a new twist on Darwin's hypothesis that reproductive isolation can evolve as an incidental by-product of adaptation to novel environments.
Our mapping population was created with repeated backcrossing and phenotypic selection for copper tolerance [20],[21]. The initial cross between Copperopolis and Cerig was conducted in 1981, but after the discovery of hybrid inviability loci segregating in this population, all subsequent crosses were made to the Stinson Beach population. The Tol1/Nec1 mapping population was grown and maintained at the University of Exeter greenhouse from 2005–2009. Tolerance was scored as a threshold character as previously described [20]. Tolerance was measured repeatedly (2–6 times) for every putative recombinant line, although a few lines did perish in the greenhouse before they could be re-tested. To score hybrid lethality, we grew 60 seedlings from a focal cross and scored the number of plants with majority yellow or necrotic tissue after 3 wk of growth [22]. Hybrid lethality was scored in multiple (2–8) growouts per line from 2008–2009.
Genetic mapping of Tol1 and Nec1 was conducted at Duke University using tissue samples shipped from University of Exeter. Genomic DNA was extracted from plant tissue using a modified hexadecyl trimethyl-ammonium bromide chloroform extraction protocol [52]. Most of the genetic markers used in this analysis are DNA fragment-length polymorphism markers analyzed using capillary electrophoresis on an ABI 3730×l DNA Analyzer [53]. These genetic markers score variation in PCR fragment length created by insertion-deletions. For most markers, primers reside in conserved exon sequence and amplify intron sequence [47]. The size of the amplified fragments was scored automatically by the program GENEMARKER (SoftGenetics, 2005, State College, PA) and was confirmed by eye. For additional fine-mapping markers, we used di-deoxy sequencing of PCR products to identify SNPs distinguishing Copperopolis and Stinson Beach alleles (Markers: Sc84_180kb, 234kb, 252kb, 277kb, 281kb, 283kb, 293kb, 297kb; Sc86_144kb, 400kb; Sc341_99kb). SNPs were scored manually using Sequencher (Gene Codes Corp., Ann Arbor, MI). To be sure of genotypes for markers defining the Nec1 locus, Sc84_283kb, 293kb, and 297kb, we ran 2–3 independent reactions for each genetic line.
We sampled Copperopolis, O'Byrnes Ferry Road, and Hunt Road populations in 2005 and 2007. We collected seed from unique maternal plants and grew a single plant from each maternal line at Duke University greenhouse for genetic testing. We measured genetic variation using fragment-length polymorphism markers [47]. We calculated the number of alleles and heterozygosity using the program Arlequin [48]. We calculated observed levels of Fst for Tol1Link and Tol1UnLink markers and generated null distribution of Fst using Tol1UnLink markers using coalescence simulations implemented in the program LOSITAN [49] with the following settings: two populations in the Island Model, 1,000,000 simulations, neutral mean Fst, and infinite allele mutation model as suggested for microsatellite markers [34]. To conduct numerical simulations of the two-locus hitchhiking model based on Maynard Smith and Haigh [35] and Barton [38], we wrote a program in C. Constant variables in each simulation were the population size, N, set at 1,000, and the allele frequency of tightly linked neutral allele, U0, set at 0.05. We ran simulations varying the strength of selection, initial frequency of selected allele, and LD between the selected allele and tightly linked sites (Figure S6).
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.