doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.pntd.0003141 | Strongyloides stercoralis: Systematic Review of Barriers to Controlling Strongyloidiasis for Australian Indigenous Communities | Strongyloides stercoralis infects human hosts mainly through skin contact with contaminated soil. The result is strongyloidiasis, a parasitic disease, with a unique cycle of auto-infection causing a variety of symptoms and signs, with possible fatality from hyper-infection. Australian Indigenous community members, often living in rural and remote settings, are exposed to and infected with S. stercoralis. The aim of this review is to determine barriers to control of strongyloidiasis. The purpose is to contribute to the development of initiatives for prevention, early detection and effective treatment of strongyloidiasis.
Systematic search reviewing research published 2012 and earlier was conducted. Research articles discussing aspects of strongyloidiasis, context of infection and overall health in Indigenous Australians were reviewed. Based on the PRISMA statement, the systematic search of health databases, Academic Search Premier, Informit, Medline, PubMed, AMED, CINAHL, Health Source Nursing and Academic was conducted. Key search terms included strongyloidiasis, Indigenous, Australia, health, and community. 340 articles were retrieved with 16 original research articles published between 1969 and 2006 meeting criteria. Review found barriers to control defined across three key themes, (1) health status, (2) socioeconomic status, and (3) health care literacy and procedures.
This study identifies five points of intervention: (1) develop reporting protocols between health care system and communities; (2) test all Indigenous Australian patients, immunocompromised patients and those exposed to areas with S. stercoralis; (3) health professionals require detailed information on strongyloidiasis and potential for exposure to Indigenous Australian people; (4) to establish testing and treatment initiatives within communities; and (5) to measure and report prevalence rates specific to communities and to act with initiatives based on these results. By defining barriers to control of strongyloidiasis in Australian Indigenous people, improved outcomes of prevention, treatment of strongyloidiasis and increased health overall are attainable.
| Strongyloides stercoralis, a nematode parasite, has a well-documented history of infecting human hosts in tropic and subtropic regions mainly through skin contact with inhabited soil. The result is strongyloidiasis, a human parasitic disease, with a unique cycle of auto-infection contributing to a variety of symptoms, of which, hyper-infection causing fatality may occur. In Australia, Indigenous community members often located in rural and remote settings, are exposed to and infected with strongyloides. Previous researchers report strongyloidiasis as a recurrent health issue for Indigenous Australians. This is a systematic review to determine the barriers to control for this pernicious pathogen. Barriers to control can be defined across three key themes: (1) health status, (2) socioeconomic status, and (3) health care literacy and procedure. By conceptualizing these barriers and addressing steps to control as outlined in this study, there is potential for improvement in prevention and treatment outcomes of strongyloidiasis and subsequently, overall health for Australian Indigenous people. This study contributes to furthering prevention and treatment of strongyloidiasis, increasing exposure to the issue of strongyloidiasis in Australian Indigenous people. It is the intent of this paper to express the need to have continued research and further health policy directed specifically to eradicate strongyloidiasis in Australian Indigenous communities.
| Strongyloidies stercoralis, a nematode parasite, is well documented as a potentially fatal soil transmitted helminth, described as a unique and complex human parasite in Speare [1]. S. stercoralis is a cosmopolitan parasite, but is more prevalent in tropical regions of the world, including tropical Australia. Rural and remote regions of Australia, in particular, Queensland, Northern Territory, Western Australia, north of South Australia and northern areas of New South Wales, endemic rates [1]-[5]. Australia's Indigenous communities have high prevalence of strongyloidiasis (disease resulting from S. stercoralis) as do immigrants from other endemic countries, travellers to these countries and military personnel who have spent time in endemic regions [6], [7]. Soulsby, Hewagama and Brady [8] report four cases of strongyloidiasis in non-Indigenous people resulting from work-related exposure presenting at Alice Springs Hospital and by implication acquired indirectly from Indigenous populations. Those infected included a teacher at an Indigenous school, a child care worker, an ex-nurse and a paediatrician. Very high prevalence rates are reported for Australian Indigenous communities [3], [4], [6], [7], [9], [10]. Johnston, Morris, Speare, et al. [7] describe strongyloidiasis as a clinically important condition in Australia. Kline, McCarthy, Pearson, et al. [11] discuss major neglected tropical diseases in Oceania and emphasize strongyloidiasis as an important infection despite the lack of data on overall prevalence rates and clinical impact.
Strongyloidiasis in a community is evidence that individual(s) in that community has been exposed to S. stercoralis from soil contaminated by human faeces [6]. Infected individuals pass first stage larvae in the faeces; these develop on the soil to infective larvae which penetrate the skin of the next host. After a blood-lung migration, parasitic adult females (there is no parasitic male) molt and develop into adult female worms in tunnels in the small intestinal mucosa [12]. Eggs are then laid in the tunnels, hatch, and produce first stage larvae in the intestinal lumen. Most of these pass out in the feces. A small number, however, change to infective larvae in the gut. These autoinfective larvae penetrate the wall of the large intestine and re-enter the body. Hence, S. stercoralis is a very unusual nematode, producing infective larvae not only externally in the soil, but also internally [12].
The occurrence of the autoinfective larvae is the main reason strongyloidiasis is such a serious disease [12], [13]. Infection is life-long since adult worms are replaced by young worms and the infection does not end when the original crop of adults die. Worm numbers can rise incrementally to produce severe disease, known as the hyperinfection syndrome. Autoinfective larvae, migrating from the lumen of the large intestine, can carry enteric bacteria into the body, resulting in sepsis in any organ. Of patients with the hyperinfection syndrome, 50% present with a septic event (pneumonia, septicaemia, meningitis, peritonitis) usually caused by an enteric bacteria or polymicrobial suite of enteric bacterial [14]. Complicating this is that S. stercoralis has an immunosuppressive effect [15], [16]. Hyperinfection occurs mainly, but not exclusively, in the people who are immunocompromised or immunodeficient with a high case fatality rate of hyperinfection, at least 60% [6], [7], [9], [10], [13], [17], [18].
Strongyloidiasis is usually symptomatic [14] but most signs and symptoms are non-specific. The exception is with larva currens, a rapidly moving urticarial linear rash that marks the passage of an autoinfective larvae through the skin [14], [19]. This is pathognomonic of strongyloidiasis. The other non-specific signs and symptoms can include gastrointestinal (e.g., abdominal pain, nausea, diarrhea, weight loss), respiratory (e.g., cough (productive and non-productive), haemoptysis, cutaneous (e.g., urticara) and general malaise [7], [10], [14], [20]. Hyperinfective strongyloidiasis, in addition to the spectrum of acute-infection symptoms, can also clinically present as paralytic ileus, pulmonary haemorrhage, pneumonia, meningitis, septicaemia or other bacterial infections [6], [10], [14], [16], [18], [20]–[22].
Diagnostic testing includes serology and faecal examination. Once diagnosed, strongyloidiasis can be eradicated with specific anthelmintics, ivermectin being the drug of choice [6], [7], [12], [17]. The recommended treatment for strongyloidiasis has changed with the development of more effective anthelmintic drugs. Thiabendazole was the first moderately effective anthelmintic introduced in the mid-1970s [23], [24]. Albendazole, a benzimidazole like thiabendazole, was recommended as the treatment of choice for strongyloidiasis about the mid-1990s [25]. It was replaced by ivermectin as first line recommended anthelmintic in the early 2000s [10].
In Australia, ivermectin is not licensed for children <5 years or for use in pregnancy [26], [27], although there is no evidence of harm in these groups [10]. Albendazole is used for > 6 months and <10 kg to adults, not licensed for use during pregnancy [26]–[28]. Fatality from strongyloidiasis most often results from missed or late diagnosis, inadequate treatment and/or the use of immunosuppressant drug therapy in high risk groups [6], [10], [17]. Co-infection of strongyloidiasis with HTLV-1 is associated with more serious strongyloidiasis and potential resistance to treatment [10], [15]. In addition, HTLV-1 carriers are more likely to develop T-cell leukaemia when infected with S. stercoralis [29]–[32].
There are questions about the limited information available about the prevalence, clinical picture, diagnosis and public health approaches to manage strongyloidiasis in rural and remote Indigenous communities in tropical regions of Australia [5], [33]. Programs based on the treatment of stool positive individuals have also been associated with decreases in prevalence [7]. Researchers suggest that little published evidence of public health approaches to control strongyloidiasis exists [7], [34] and there is a need to consider mass drug administration in Indigenous Australian communities with high prevalence of strongyloidiasis [10], [11].
This systematic review attempts to answer the questions, what is the epidemiology of strongyloidiasis in Australian Indigenous people, and, what, if any, are the mentioned barriers to control? The aim of this review is to identify research focused on strongyloidiasis in this specific population and to collect and analyse available data specific to symptoms, diagnosis and treatment to determine barriers to control of strongyloidiasis. For the purpose of this paper, we respectively use the term Indigenous to represent Australian Aboriginal people and Torres Strait Islanders.
The outline and focus of this paper is framed on the concept of a translational research framework described by Thomson [35] within the Australian Indigenous HealthInfoNet. This systematic review was designed as a narrative review of the evidence as a way to summarise, explain and interpret evidence with thematic analysis [36].
This systematic review was based on the PRISMA statement, a tool to summarize accurate, reliable, quality evidence by way of transparent reporting (Checklist S1) [37], [38]. A systematic search of health databases, Academic Search Premier, Informit, Medline, PubMed, AMED, CINAHL, Health Source Nursing and Academic was performed to search for all articles published 2012 and prior were included in the search. Articles were searched through the online academic search site, Google Scholar and internet searches for websites containing information about strongyloidiasis. Key search terms included strongyloidiasis, Indigenous, Australia, health, and community with search strategy developed to access the broadest range of articles about strongyloidiasis are presented in Table 1. Reference lists of original articles, review articles, grey literature and websites were searched for potential articles to review for inclusion. Language restrictions were not imposed.
To meet inclusion criteria, original qualitative or quantitative research articles contained content addressing one or more of the following: symptoms, diagnosis, treatment, and barriers to control of strongyloidiasis. The location of the studies had to be Australia and include Australian Indigenous people. Exclusion criteria included, review articles and non-peer reviewed literature, original research articles with animal only studies, pharmaceutical therapy only studies and studies not differentiating S. stercoralis or strongyloidiasis from amongst other parasites or parasitic infections.
Based on these selection criteria, articles were reviewed in two stages. First stage, article titles and abstracts were screened to meet the requirements of strongyloidiasis as topic, Australian location and inclusion of Indigenous Australians. Second stage, articles were read as full text. Articles meeting final criteria were included in the study. Figure 1 represents the overall article search outcome.
From the original research questions, (1) what is the epidemiology of strongyloidiasis in Australian Indigenous people? and (2) what, if any, are the mentioned barriers to control? Description of studies was collected and a thematic analysis conducted [36]. Key data extracted were: purpose of study, study design, participant description, symptoms, diagnosis, treatment, barriers to control, and author's conclusions. Articles were presented in a database with publisher details and summarized key data. The categories of symptoms, diagnosis, treatment and barriers to control were further assessed and coded using thematic analysis to determine recurring items in each. Symptoms were defined as manifestations of strongyloidiasis and included symptoms and signs due to strongyloidiasis and other existing concurrent conditions. Diagnosis was defined medical diagnoses including health status, tests performed and results.
Assessment of treatment of strongyloidiasis was based on the recommended therapy at the time of publication and defined as details on therapy provided and the comments on outcomes. Barriers to control were defined as a medical context, symptom and/or condition, or social determinant (derived from categories of symptoms, diagnosis, treatment and each authors' summary and conclusions) that inhibited overall health and/or recovery from strongyloidiasis of the individual(s). Once the barriers to control items were documented, they were then coded into barrier themes and health level. Detailing each barrier and the associating theme and level supports the translational knowledge concept by assisting to identify the relevant stakeholders [39].
Figure 1 provides an overview of the literature search results. 340 articles were retrieved with a total of 16 articles, published between 1969 and 2006, eligible for the systematic review and are summarized in Table 2. Eleven eligible articles were from electronic library databases. Google Scholar revealed two additional eligible articles. The reference lists reviewed from published articles, grey literature and internet websites reporting on strongyloidiasis infections of Indigenous people of Australia revealed three eligible articles. Study design included case studies, retrospective and prospective comparison and non-comparison studies. Participant numbers ranged from 1 to 683. Indigenous Australian children were reported in 12/16 studies, of those 8/12 reported children only. Indigenous Australian adults were reported in 7/16 studies, of which 4/7 reported adult only. Thirteen studies were conducted in hospital and four in Indigenous communities. Eleven studies examined strongyloidiasis only with the remaining discussing the parasitic infection in the context of other infections [40], [41] or while examining gastrointestinal issues [42]–[44].The 16 papers included 2537 Indigenous participants and 272 non-Indigenous participants.
Eleven papers described manifestations of strongyloidiasis, including symptoms and signs due to strongyloidiasis as well as other concurrent conditions (Table 3). Studies noted strongyloidiasis symptoms such as diarrhoea, malnutrition and anorexia, abdominal pain, abdominal distension, anemia, septicaemia, and fever. Other concurrent conditions including Type 2 Diabetes, Lupus, Chronic Liver Disease and Chronic Lung Disease, Alcoholism, Pneumonia, Bronchitis, COPD, Acute Rheumatic Fever, Acute Renal Failure and/or general gastrointestinal, cardiac and respiratory problems were reported. Gunzburg, Gracey, Burke, et al. [43] reported only diarrheal symptoms as this was the scope of the study. Page, Dempsey, and McCarthy [28] and Prociv & Luke [5], although studying strongyloidiasis specifically, did not focus on symptomology. Four studies [4], [15], [40], [42] did not discuss symptomology due to the aim of the study.
All sixteen studies provided data on diagnosis of strongyloidiasis determined by one or more tests (Table 4). Nine studies performed purposeful testing [4], [5], [21], [28], [40]–[43]. Five studies reported strongyloidiasis had been diagnosed when not suspected [15], [22], [42], [45], [46].
Articles were reviewed for the adequacy of treatment noting that recommended therapy has changed with time (Table 5). Eight articles discussed the use of one or a combination of albendazole, thiabendazole and ivermectin. Three articles described a subgroup of patients receiving no therapy [28], [42], [45] and one article mentioned the use of pyrantel only for strongyloidiasis [5]. Pyrantel is ineffective against S. stercoralis [47]. In two articles, prednisolone or prednisone, a treatment which suppresses the immune system and as a result can increase the severity of strongyloidiasis, was administered to patients. Walker-Smith [42] discussed diagnoses of giardiasis and strongyloidiasis in children and provided no data on treatment. Einsiedel & Fernandes [15] detailed treatment therapies across four case studies, of which, only one case received correct strongyloidiasis treatment with ivermectin. Overall, adequate treatment was documented in publications in only 5.2% of cases.
Barriers to control of strongyloidiasis were summarized in terms of item, theme and health access level (Table 6). Three barriers themes emerged as items contributing to adequate management of strongyloidiasis: (1) health status; (2) socioeconomic status; (3) health care literacy and procedures. Theme 1, health status was defined patients' health prior to and at the time of diagnosis of strongyloidiasis. This included concurrent infections (e.g., meningitis, pneumonia), concurrent chronic health conditions (e.g., Lupus, Chronic Liver Disease, Chronic Lung Disease, Acute Rheumatic Fever, HTLV-1, Hepatitis B, alcoholism, immunocompromised, immunosuppressed) and the phenomenon of strongyloidiasis (e.g., re-infection, hyperinfection, at times asymptomatic, chronic diarrhoea, septicaemia). Theme 2, socioeconomic status included living conditions, racial disparities, communication (e.g., interaction between community, patients, health professionals/institutions).Theme 3, health care literacy and procedures involved barriers that influence the diagnosis and treatment outcomes (e.g., delayed diagnosis, difficult to detect, failure to recognize symptoms, inadequate knowledge/treatment/treatment dose, serology test cut off, lack of communication, lack of screening, lack of follow-up, treatment non-compliance).
Einsiedel & Fernandes [15] had the largest number of symptoms and signs and other conditions associated with barriers to control of strongyloidiasis. The top four barriers listed most often (determined by the most barriers per article, total of 4) were delayed diagnosis, inadequate treatment, living conditions and malnutrition. Barriers to control are located across all four health access levels: (1) Individual; (2) Public/Community; (3) Organization; and (4) Healthcare system.
This study reviewed original articles on strongyloidiasis in Indigenous Australian people. Articles were analyzed for symptoms, diagnosis and treatment and barriers to control of Strongyloidiasis. Overall outcomes are presented as symptomology, diagnosis and treatment protocols, community research and action and addressing barriers to control.
The broad spectrum of symptoms, as represented in manifestations of strongyloidiasis in Table 3, illustrates the complex nature of Strongyloidiasis that is so often misdiagnosed. Many of these manifestations, such as diarrhoea, stomach pain, malnutrition, dehydration and vomiting are common to many illnesses and diseases. As described by researchers [6], [15], [16], [20], [43], [45], [46], strongyloidiasis can present many varying symptoms or be asymptomatic [43], [46]. It is important to recognize that strongyloidiasis can potentially exist for years presenting often with non-specific symptoms and signs (e.g., diarrhoea) as well as at times with periods without symptoms.
Delayed diagnosis, inadequate knowledge/treatment/treatment dose, lack of communication and lack of follow up by health professionals were described as particular issues in the majority of studies [5], [15], [16], [22], [29], [40], [44], [45], [50], [51]. Infection should be suspected in every person with unexplained abdominal pain, diarrhoea, cutaneous symptoms or eosinophilia and the laboratory alerted of a provisional diagnosis [45]. Testing for strongyloidiasis is particularly important for patients from populations in S. stercoralis endemic areas. Rural and remote Indigenous communities (more specifically northern Australia) and including immunocompromised patients are at particular risk for hyperinfecion before administering immunosuppressive medication [22]. Protocol including clinical screening index, stool microscopy and culture, full blood count, immunoglobulin levels, and serological testing is recommended [22].
Majority of studies reported Indigenous Australian children with strongyloidiasis suggesting a diagnosis of strongyloidiasis should be considered when Indigenous children presenting with even non-suspecting general gastro-intestinal symptoms. Mucosal damage in Indigenous Australian children is possibly a result of damage produced by repeated episodes of gastroenteritis and/or parasitic infection, including strongyloidiasis [42]. Reduction in the frequency of gastroenteritis and parasitic infection in Indigenous children should greatly reduce incidence of small intestinal mucosal damage [42]. Working to eradicate or reduce strongyloidiasis infection in children with early detection and immediate treatment could decrease strongyloidiasis and mucosal damage. Given the challenges of diagnosing infection, standardizing treatment in communities for an extended period could potentially decrease infections rates [5].
Parasitic diseases have significant health risk and morbidity for Australian Indigenous people [11], [20]. Rural and remote communities are the most affected [3], [18]; mainly in children; and those immunocompromised with a number of cases of fatality reported [15], [22], [40], [41]. Studies in 2002 and 2005 report there are limited published examples of community interventions in Australia to control strongyloidiasis [7], [52]. Johnston, Morris, Speare, et al. [7] found no evidence of studies examining roles of environmental interventions and expressed the need to do so. The need for initiatives for housing and sanitation are imperative [15]. Issues of environmental health must be addressed concurrently with health service initiatives to develop long term and sustainable improvements in control of infectious parasitic and non-parasitic diseases in rural and remote Indigenous communities in Australia [10], [11], [20]. There may be increased risks associated with a casual approach to management and may be significantly higher for Indigenous Australian people living in HTLV-1 endemic Central Australia [10], [40]. Einsiedel and Woodman [40] further state the risk of strongyloidiasis in Indigenous communities and HTLV-1 infection may further predispose people to complicated strongyloidiasis.
Steps to address the barriers to control should include: (1) development of S. stercoralis and strongyloidiasis reporting protocols across health care system and communities (e.g., consistent case study reporting methods, documentation of current infection sites) [6], [40]; (2) testing all Indigenous Australian patients, immunocompromised patients and those exposed to or living in areas of strongyloidiasis (e.g., rural/remote communities) presenting with gastrointestinal or respiratory symptoms (take particular notice of individuals from these groups with repeated visits to hospital) [7], [15], [16], [48]; (3) requirement of health professionals to have detailed information and education regarding strongyloidiasis and the potential for exposure in Indigenous Australian communities (e.g., understanding of the expanse of symptoms and potential for asymptomology, difficulty in diagnosis, need for variety of tests and retesting, accurate follow-up to confirm patient cleared of infection) [5], [15], [21], [42]; (4) establishment of testing and treatment initiatives in the community (e.g., over extended periods and periodically and treat symptomatic and asymptomatic strongyloidiasis carriers) [6], [10], [12], [15], [45]; (5) measure and report prevalence specific to Indigenous Australian communities and to act with initiatives based on these results [6], [12], [40].
|
10.1371/journal.ppat.1004738 | Attenuation of Tick-Borne Encephalitis Virus Using Large-Scale Random Codon Re-encoding | Large-scale codon re-encoding (i.e. introduction of a large number of synonymous mutations) is a novel method of generating attenuated viruses. Here, it was applied to the pathogenic flavivirus, tick-borne encephalitis virus (TBEV) which causes febrile illness and encephalitis in humans in forested regions of Europe and Asia. Using an infectious clone of the Oshima 5–10 strain ("wild-type virus"), a cassette of 1.4kb located in the NS5 coding region, was modified by randomly introducing 273 synonymous mutations ("re-encoded virus"). Whilst the in cellulo replicative fitness of the re-encoded virus was only slightly reduced, the re-encoded virus displayed an attenuated phenotype in a laboratory mouse model of non-lethal encephalitis. Following intra-peritoneal inoculation of either 2.105 or 2.106 TCID50 of virus, the frequency of viraemia, neurovirulence (measured using weight loss and appearance of symptoms) and neuroinvasiveness (detection of virus in the brain) were significantly decreased when compared with the wild-type virus. Mice infected by wild-type or re-encoded viruses produced comparable amounts of neutralising antibodies and results of challenge experiments demonstrated that mice previously infected with the re-encoded virus were protected against subsequent infection by the wild-type virus. This constitutes evidence that a mammalian species can be protected against infection by a virulent wild-type positive-stranded RNA virus following immunisation with a derived randomly re-encoded strain. Our results demonstrate that random codon re-encoding is potentially a simple and effective method of generating live-attenuated vaccine candidates against pathogenic flaviviruses.
| The arbovirus Tick-borne encephalitis virus (TBEV; genus Flavivirus) is transmitted by ticks of the Ixodes genus. TBEV causes febrile illness and encephalitis in humans in forested regions of Europe and Asia. The incidence of TBE is increasing across Central and Eastern European countries despite the availability of several licensed inactivated vaccines and appropriate vaccination programmes. Large-scale codon re-encoding, a recently developed attenuation method that modifies viral RNA nucleotide composition of large coding regions without alteration of the encoded proteins, has been successfully applied to a variety of RNA viruses. In contrast with previous empirical methods of generating live attenuated vaccines, large-scale codon re-encoding facilitates rapid generation of vaccine candidates using reverse genetics methods, by direct control of the attenuation phenotype. Additional benefits include reduced costs and induction of long-term immunity. Here, we have applied the large-scale codon re-encoding method to the TBEV to demonstrate the principle of developing a live attenuated virus vaccine which protects mice against subsequent infection with the wild type virulent virus. This study therefore illustrates that codon re-encoding is potentially an easily derived and effective method of producing live attenuated vaccine candidates against positive-stranded RNA viruses.
| The genus Flavivirus (family Flaviviridae) includes important human pathogens such as yellow fever virus (YFV), dengue virus (DENV), Japanese encephalitis virus (JEV), West Nile virus and tick-borne encephalitis virus (TBEV). Flaviviruses are enveloped, single-stranded positive-sense RNA viruses with virions, close to 50nm in diameter, and a viral genome of ca. 11 kb which includes a unique open reading frame (ORF) encoding structural (C-prM-E) and non-structural proteins (NS1–2A-2B-3–4A-4B-5) [1,2]. Some mosquito-borne flaviviruses also harbour sequences that induce a proportion of translating ribosomes to shift-1 nt and continue translating in the new reading frame to produce a ‘transframe’ fusion protein [3,4]. Most flaviviruses are arboviruses and are therefore maintained in nature by circulating between haematophagous arthropod vectors and vertebrate hosts. Arthropod-borne flaviviruses are sub-divided into two major groups: the tick-borne and mosquito-borne flaviviruses (TBFVs and MBFVs respectively) [1,2,5]. TBFVs include a heterogeneous group called seabird tick-borne flavivirus group (S-TBFV) [6] and the mammalian tick-borne flavivirus group (M-TBFV), with all known pathogenic TBFVs causing febrile illness, encephalitis and/or haemorrhagic fever in humans. In the latter group, TBEVs are recognised in 25 European and 7 Asian countries and transmitted by Ixodes species ticks [7]. The TBEVs are subdivided into three sub-types, namely Siberian, Western European and Far Eastern viruses [8,9], the latter being responsible for the most severe forms of central nervous system (CNS) disorders associated with high fatality rates (5–20%) [10]. Despite the availability of several licensed inactivated vaccines and vaccination programmes [1,11], the incidence of TBEV infections is increasing across much of Central and Eastern European countries, currently with an estimated 9,000 cases per year [12,13,14].
Live attenuated vaccines provide effective and affordable protection against major flaviviral infections. One dose of the widely used 17D YF vaccine used since 1937, provides long-lasting immunity [15]. The live attenuated JE vaccine (strain SA-14–14–2) has been successfully used in China with over 100 million children vaccinated [16]. Attenuated strains of virus have been obtained in the past using empirical methods such as serial passage of wild-type (WT) viruses in cell cultures and/or chicken embryos. Whilst several hundred million yellow fever vaccine doses have been administered, and proven to be safe and highly efficacious [16,17], the attenuation mechanism is associated with a number of non-synonymous mutations (31 in the case of the 17D-204 YF vaccine strain when compared with the WT Asibi virus). These modifications can occasionally generate new biological properties, e.g., a neurovirulent phenotype for YF 17D in contrast to the viscerotropic phenotype of WT yellow fever viruses. New approaches to resolve such problems are required.
Large-scale codon re-encoding is a recently developed method with which to attenuate virus by introducing a large number of slightly deleterious synonymous mutations into the protein coding region of the viral genomic RNA without alteration of the encoded proteins. Genomic re-encoding has been successfully applied to a variety of RNA viruses: poliovirus, influenza A virus, chikungunya virus, human respiratory syncytial virus, human immunodeficiency virus, Japanese encephalitis virus and porcine reproductive and respiratory syndrome virus [18,19,20,21,22,23,24,25,26,27]. In each of these examples, re-encoding modulated virus fitness thus generating potential vaccine candidates with antigenically indistinguishable proteins, therefore alleviating the generation of novel and therefore undesirable biological properties. To date, most published studies have utilised specific re-encoding approaches, including codon deoptimisation, codon-pair deoptimisation or increase of CpG/UpA dinucleotide frequency. The choice of these methods is based on the assumption that global modification of the viral genome induces attenuation because synonymous sites have been shaped by genome-wide mutational processes during virus evolution [28]. However, a random re-encoding approach, previously applied in cellulo to the chikungunya virus, also produced an attenuated phenotype without modifying the global properties of the viral genome, thus underlining the role of local constraints that also shape synonymous sites such as secondary structures or interactions between viral RNA and capsid proteins [27,29,30,31]. Therefore, whilst the efficacy of re-encoding methods for attenuating RNA viruses has been widely demonstrated, the precise—and presumably multiple- mechanisms and their respective contributions to attenuation remain to be analysed in more detail. In cellulo experiments showed, for poliovirus, chikungunya virus and bacterial virus T7, that the phenotype of the re-encoded viruses was stable, and that the evolutionary response to re-encoding was compensatory in nature, with very few reversion mutations [22,27,32].
In the present study, we have applied the random large-scale codon re-encoding method to the TBEV Oshima 5–10 strain, isolated in Japan in 1995 [33]. This strain belongs to the Far Eastern TBEV subtype which is characteristically highly neurovirulent for mice [34,35]. Our studies demonstrate that this re-encoded TBEV Oshima 5–10 strain exhibited an attenuated phenotype in vivo and induced robust protective immunity in mice subsequently infected with the WT virus.
An infectious clone derived from the wild-type Oshima 5–10 TBEV strain was constructed using reverse genetics methods (see the Materials and Methods section for more details). This infectious clone was recovered from cell cultures and designated “WT_IC virus”. In addition, a re-encoded “NS5_Reenc_IC virus” was derived from WT_IC virus by substituting a cassette of approximately 1.4kb, in the corresponding NS5 coding region, with the re-encoded counterpart (Fig. 1). We have chosen to introduce mutations into the NS5 coding region because the first re-encoded cassette introduced into the Chikungunya virus, as described in our previous work [27], was located in the nsP4 coding region which also encoded the viral RNA dependent RNA polymerase. This enabled comparisons, in cellulo and in vivo, of the biological properties of the WT and re-encoded viruses.
A total of 273 synonymous mutations, located between positions 8,619 and 10,019 (with reference to the complete genome sequence, GenBank accession number KF623542), was introduced in the specified NS5 coding region using a random distribution algorithm with restrictions [27] (Table 1). The codon usage (measured using the effective number of codons; eNC) and G+C% of the NS5_Reenc_IC virus was slightly modified compared with that of the WT_IC virus: 53.96 vs 55.46 and 54.3% vs 53.8%, respectively (Table 1). When compared with 85 TBEV and 56 other TBFV complete ORF sequences retrieved from GenBank, the eNC value of the NS5_Reenc_IC virus was higher than the maximum eNC value of TBEV sequences but fell within the extreme values of all available TBFV sequences. On the other hand, the G+C% value of NS5_Reenc_IC virus fell within the extreme values of TBEV sequences.
WT_IC and NS5_Reenc_IC viruses were recovered following transfection of the corresponding infectious clones in BHK21 cells. These two viruses were then passaged once on BHK21 cells and their replicative fitness was subsequently studied.
The laboratory mouse model has been the primary global choice with which to study the CNS pathology induced by TBEVs. Pathologic changes in mouse brains as well as clinical signs are similar of those observed in humans [35,36,37]. In agreement with previous findings, virus recovered from infectious clones displayed lower mortality than the original cell culture derived Oshima 5–10 strain which induces high mortality in mice following intra-peritoneal inoculation [34]. Therefore, mortality was not used as a comparative criterion: we only observed late and low mortality rates when the mice were inoculated with the lower dose of TBEV (Fig. 3). It should also be noted that TBEVS are known to infect small rodents chronically/persistently as suggested by field studies [38,39].
Five-week-old C57Bl/6J female mice were inoculated intra-peritoneally with 200μL containing 2.105 TCID50 [low dose] or 2.106 TCID50 [high dose] of virus. Clinical monitoring included (i) the clinical manifestation of the disease (shivering, humpback, dirty eyes, weak paws, hemiplegia or tetraplegia) and (ii) the body weight curve (a weight loss of more than 6% of the initial weight was chosen as a disease recognition criterion, as described in S1 Fig in S1 Text). Periodically, groups of mice were sacrificed to conduct virological investigations. A close relationship between viral load values determined using either quantitative RT-PCR (qRT-PCR) or the TCID50 method was observed. Consequently, the virological follow-up of sera and brains was performed using qRT-PCR (S2A and S2B Fig in S1 Text). Sera were also used for viral serological analysis by ELISA and virus neutralisation assays.
We have evaluated in cellulo and in vivo the effect of genomic large-scale random re-encoding on TBEV), a pathogenic TBFV that causes febrile illness and encephalitis in humans. Encephalitic flavivirus infections provoke CNS pathology which can be correlated with the observed morbidity and mortality: viruses replicating in the CNS induce direct neuronal damage causing severe CNS dysfunction often involving long-term neurological sequelae in non-fatal cases [36,40,41]. In addition, it has been demonstrated that host immune response is a critical determinant of clinical outcome [36,42,43,44,45].
Here, an infectious clone of a neurovirulent strain of TBEV (Far Eastern subtype) was used to perform in vivo studies in a mouse model that faithfully mirrors many aspects of the infection in humans. Our experiments showed that decreased replicative fitness (as determined by in cellulo competition assays and in vivo viraemia measurement) was associated with reduced neuroinvasiveness as previously described [46]. In vivo competition experiments shed further light on the mechanisms of pathogenesis: when a 50/50 (WT/Reenc) TCID50 initial ratio was used, as expected, the wild-type virus rapidly out-paced the re-encoded virus in blood and was the only virus detected in mouse brains. However, when a 10/90 (WT/Reenc) TCID50 initial ratio was used, the re-encoded virus was the only one detected in brain, suggesting an early and selective neuroinvasion process: only the majority virus present in blood during the first hours of the viraemia invades the central nervous system. Of course, more experiments are needed to confirm these findings and the potential role of viral interferences have to be assessed [47,48].
In this model, TBEV infection is frequently associated with asymptomatic persistence of the virus in mouse brain as previously described [38,39] and confirmed by our experiments (we detected viral RNA of the re-encoded virus but failed to isolate the virus). Previous reports of virus reactivation, years after the initial infection, suggest that this specific phenomenon might also occur in humans [49].
The main objectives of this study were (i) to analyse the effect of genome random re-encoding on the fitness and clinical phenotype of a virulent TBEV strain and (ii) to perform a complete set of in vivo experiments including immunisation with a re-encoded virus and follow-up challenge experiments with an infectious wild-type virus. It has previously been demonstrated that large-scale re-encoding generates attenuated viruses and the studies support the proposal that relative degree of attenuation or replicative fitness can be regulated by modulating the number of introduced synonymous mutations [18,19,20,21,22,23,24,25,26,27]. Indeed, a re-encoded strain of influenza A virus that displayed limited fitness in cellulo was highly attenuated when tested in a mouse model and showed potential for use as a vaccine candidate [25]. Similarly, re-encoding a 1.4 kb region of the CHIKV genome by introducing 300 random synonymous mutations was associated with only limited in cellulo attenuation [27]. Therefore, we hypothesised that appropriate re-encoding of a TBEV strain might result in limited fitness reduction in cellulo and thus provide a relevant candidate to study the relative degree of in vivo attenuation. Accordingly, we introduced 273 random synonymous mutations in the NS5 gene of a neurovirulent strain of TBEV (Oshima, Far Eastern subtype), with limited modification of G+C content or codon bias of the genome. This re-encoding protocol produced a virus variant that displayed no fitness difference when growth kinetics were compared with WT TBEV in mammalian cell cultures. However, more sensitive in cellulo competition experiments revealed that the replicative fitness of the wild-type virus was indeed higher than that of the re-encoded virus. Moreover, in vivo experiments in immunocompetent mice fully validated our starting hypothesis: the re-encoded virus could reproduce efficiently in mice but in competition experiments the wild-type virus had a significantly higher fitness. In vivo experiments also revealed the attenuated characteristics of the re-encoded virus, namely reduced neurovirulence in terms of weight loss and appearance of neurological symptoms and reduced neuroinvasiveness (i.e. lower proportion of mice with virus in the brain). It can therefore be concluded that the re-encoding process, although restricted to the NS5 gene, decreased the pathogenicity and led to the production of an attenuated phenotype of the normally highly virulent strain of TBEV.
Concerning the mouse protection experiments, the results showed that, at 40 days post-infection, neutralising antibodies were produced by 100% of mice, i.e. infected by either the wild-type or the re-encoded virus. Moreover, no significant differences in antibody titres were observed between mice infected with either virus. Thus, mice “immunised” by the re-encoded virus were likely to have been protected against subsequent TBEV infection. This was verified at 40 days post-infection by challenging the “immunised” mice using a high dose of intraperitoneally administered wild-type virus. Protection was effective in terms of viraemia, neurovirulence and neuroinvasion. Likewise, inoculation of CD155 tg mice with re-encoded strains of poliovirus induced the production of neutralising antibodies and protects the mice against a subsequent challenge with a lethal dose of virus [24].
The results provide a robust proof of concept: large-scale random re-encoding can be used to produce in vivo attenuated strains of TBEV and infection of mice by re-encoded viruses can induce neutralising and protective immune responses against challenge with virulent homologous viruses. This represents evidence that a positive-stranded randomly re-encoded RNA virus could be developed and trialled as a potential vaccine to protect humans and/or animals against viral pathogens.
Many current virus vaccines were derived empirically and carry an inherent risk of vaccine-associated complications. Our findings open up new perspectives for the development of new-generation custom-designed re-encoded live-attenuated vaccines which are potentially safe, induce high levels of protective immunity and are relatively easy to produce.
The use of a highly neurovirulent strain of TBEV in our experimental model enabled us to identify the in vivo modification of the clinical picture provided by genome re-encoding. The results strongly suggest that re-encoding could be used in the future for attenuation of highly pathogenic viral species. However, it would be wiser and more practical in the specific case of TBEV, to develop a live attenuated re-encoded vaccine using a strain known to have a naturally lower association with neurovirulence and by inserting additional synonymous mutations in coding regions with a view to reducing the encephalitic potential of the virus and thus produce a potentially safer vaccine candidate. Whilst effective inactivated vaccines are available to prevent TBEV infections, the use of a live attenuated vaccine may have specific advantages, e.g. long-term protection and reduced costs [50].
Baby hamster kidney BHK21 (BHK21) cells (ATCC, number CCL10) and mouse (L929) cells (ATCC, number CCL1) were grown at 37°C with 5% CO2 in Minimum Essential Medium with 7% fetal calf serum (Life Technologies) and 1% Penicillin/Streptomycin (5000U/mL and 5000μg/mL; Life Technologies). Five-week-old C57Bl/6J mice females were provided by Charles River laboratories.
Animal protocols were reviewed and approved by the ethics committee “Comité d’éthique en expérimentation animale de Marseille—C2EA—14” (protocol number 2504). All animal experiments were performed in compliance with French national guidelines and in accordance with the European legislation covering the use of animals for scientific purposes (Directive 210/63/EU).
A cassette of 1,412 bp located in the NS5 coding region was randomly re-encoded as described previously for chikungunya virus [27]. Briefly, a computer programme was used to randomly attribute nucleotide codons based on their corresponding amino acid sequence: for example, the amino acid proline was randomly replaced by CCT, CCC, CCA or CCG. The number and the position of rare codons in primate genomes [51] (i.e. CGU, CGC, CGA, CGG, UCG, CCG, GCG, ACG), and unique restriction sites were conserved (S1 Note in S1 Text).
We modified a previously described IC of the Oshima 5–10 strain [34] (GenBank accession number of the parent virus: AB062063) by adding 9 synonymous mutations along the genome to increase the number of unique restriction sites, by replacing the SP6 promoter by a promoter CMV (pCMV) in 5′, by adding in 3′ of the complete viral genome the sequence of the hepatitis delta ribozyme followed by the simian virus 40 polyadenylation signal (HDR/SV40pA). The origin of replication was replaced by a modified pBR322. This IC was designated Cloning vector pTBEV-32.11 ic (GenBank accession number KF623542) and was considered as WT (Fig. 1). The re-encoded cassette (see above) was synthesized de novo by GenScript and inserted into the WT IC by digestion (SacII/SalI; New England Biolabs) (Fig. 1), gel purification of digestion products (Qiagen), ligation (T4 DNA ligase; Life Technologies) and transformation into electrocompetent STBL4 cells (Life Technologies). Before their transfection, both ICs were purified (0.22μm filtration) and their genome integrity was verified using a restriction map and complete sequencing.
Complete open reading frames of TBEV (n = 85) and other TBFV (n = 56) were manually extracted from GenBank (S2 and S3 Tables in S1 Text). G+C% and effective number of codons (eNC) were calculated using Codon W v1.3 software [52,53].
ICs were transfected into a 12.5cm2 culture flask containing sub-confluent BHK21 cells (FuGENE 6 transfection reagent; Roche). After incubation for 6 hours, cells were washed twice with Hank’s Balanced Salt Solution (HBSS, Life Technologies) and incubated until appearance of complete cytopathic effect (CPE). Cell supernatant medium was harvested, clarified by centrifugation and stored at -80°C. Each virus was then passaged in BHK21 cells at a calculated moi of 0.5 in a 175cm2 culture flask: after adsorption of the virus for 2 hours, the cells were washed twice (HBSS) and 50mL of medium was added and the flasks were incubated at 37°C for 72 hours. Cell supernatant media were harvested, clarified by centrifugation, aliquoted, stored at -80°C and used to perform in cellulo experiments. A similar experimental procedure was also carried out using L929 cells and the resulting cell supernatant medium was aliquoted, stored and used to perform the in vivo experiments. The integrity of the genome of all the viruses produced to perform in cellulo and in vivo experiments was verified using sequencing methods (Sanger methods).
A calculated moi of 200 or 0.5 was used to infect a 25cm2 culture flask of confluent BHK21 cells. Cells were washed twice (HBSS) 30 minutes after the infection and 7mL of medium was added. 800μL of cell supernatants were sampled just before the washes and at 2, 5, 10, 15, 23, 31 and 48 hours post-infection. They were clarified by centrifugation, aliquoted and stored at −80°C. They were then analyzed using a TCID50 assay (see below).
As described previously for chikungunya virus WT_IC virus was competed with NS5_Reenc_IC virus [27]: five initial TCID50 ratios (WT_IC/NS5_Reenc_IC virus: 1/99, 20/80, 50/50, 80/20, 99/1) were used to infect a 25cm2 culture flask of confluent BHK21 cells at a calculated moi of 0.5. Cells were washed twice with HBSS and then incubated for 48h after addition of 7mL of medium. Recovered infectious cell supernatant was then sequentially passaged 10 times in the same manner with the clarified cell supernatant medium from the previous passage. At each passage, a calculated moi of 1 was used. Aliquots of cell supernatant from each passage were clarified by centrifugation and stored at -80°C. Viral RNA was extracted from clarified culture supernatant medium using the EZ1 Virus Mini Kit v2 on the EZ1 Biorobot (both from Qiagen). Using two specific quantitative real time RT-PCR assays targeting the re-encoded NS5 coding region (see the quantitative real time RT_PCR assays section for more details), the amount of viral RNA was assessed for each virus (WT_IC and NS5_Reenc_IC) and the ratio of the two values (WT_IC/NS5_Reenc_IC) was calculated.
Five-weeks-old C57Bl/6J female mice were intra-peritoneally inoculated with 200μL containing 2.105 TCID50 or 2.106 TCID50 of virus. In some experiments (see details in the results section), a control group of mice was used (they were intra-peritoneally inoculated with 200μL of PBS). The clinical course of the viral infection was monitored by following (i) the clinical manifestation of the disease (shivering, humpback, dirty eyes, hemi- or tetra-paresia, hemiplegia or tetraplegia) and (ii) the weight of the mice. Weights were normalized with the average weight of mice of control group; the normalized weight was expressed as percentage of initial weight and calculated as follows: (% of initial weight: weight/weight at the day of the inoculation or challenge)–(mean of the % of the initial weight for control mice) +100. Brains and blood were sampled from sacrificed mice. Blood was collected by intracardiac puncture. After centrifugation, serum was aliquoted and stored at -80°C. Nucleic acid extraction using 50μL of serum previously inactivated with 50μL of AVL buffer (Qiagen) and spiked with 10μL of MS2 bacteriophage (internal control) was performed using the EZ1 Virus Mini Kit v2 on the EZ1 Biorobot (both from Qiagen). Brains were collected in 1mL of PBS with a tungsten bead and ground using a TissueLyser (Qiagen) for 3min at 30cycles/s. The brain suspensions were homogenized with NucleoSpin filters (Macherey-Nagel). The collected filtrate was then aliquoted and stored at -80°C. Virus TCID50 assays were performed using these filtrates. Nucleic acid extraction using 30μL of filtrate, 270μL of RLT buffer (Qiagen) and 10μL of MS2 bacteriophage (internal control) was performed using the EZ1 RNA Tissue Mini Kit on the EZ1 Biorobot (both from Qiagen).
100μL of the filtrates collected from brain suspensions (see above) were used to inoculate a 12.5cm2 culture flask of confluent BHK21 cells containing 400μ of medium. After incubation for 2 hours, 2.5mL of medium was added and cells were incubated 5 days. A blind passage was then realised using a 25cm2 culture flask of confluent BHK21 cells. 2mL of clarified cell supernatant diluted 1:3 was used to inoculate the cells. After incubation for 2 hours, cells were washed once with HBSS and incubated 5 days. Virus replication was demonstrated using detection of viral genomes in cell supernatant using qRT-PCR assay and detection of cytopathic effect (CPE).
All quantitative real-time PCR (qRT-PCR) assays were performed with SuperScript III Platinium One-Step qRT-PCR kit. The mix content for a final volume of 25μL per sample, was as follows: a standard quantity of 2x of PCR Mastermix and Enzymes, both primers (final concentration: 0.4μM), probe (final concentration: 0.1μM) and 4μL of extracted nucleic acids. qRT-PCR were performed on CFX96 Real-Time System/C1000 Touch Thermal Cycler (Biorad) with the following conditions: 15min at 50°C, 2min at 95°C, then 45 times 15sec at 95°C and 40sec at 60°C, data collection occurring during this last step. Primers and probe sequences are detailed in S1 Table in S1 Text. All sera and brain samples from mice were spiked with MS2 bacteriophage (internal control) prior nucleic acid extraction and a MS2-specific qRT-PCR was performed to monitor the extraction, reverse transcription, and amplification steps as previously described [54]. A universal qRT-PCR assay was used to detect the genomic RNA of all TBEVs (nucleotide position 10,236 to 10,337). The amount of viral RNA was calculated using synthetic RNA transcript for this universal assay. Results from mouse brains were normalized using amplification (qRT-PCR) of the housekeeping gene HMBS as described previously [55]. Two specific qRT-PCR assays were also used to specifically detect either WT_IC viruses or NS5_Reenc_IC viruses (nucleotide position 8,819 to 8,933). Nucleic acids from cell supernatant media of cultured WT_IC virus or NS5_Reenc_IC virus were used as standard for these specific assays. Values for the quantity of viral RNA of each standard used for both specific assays were calculated using the universal assay.
For each determination, a 96-well plate culture of confluent BHK21 cells was inoculated with 150μL/well of serial 10-fold dilutions of clarified (centrifugation) cell supernatant medium, mouse sera or mouse brain filtrates: each dilution was repeated 6 times. The plates were incubated for 7 days and read for absence or presence of CPE in each well. Determination of the TCID50/mL was performed using the method of Reed and Muench [56].
Sera were incubated for 30min at 56°C prior to viral serology. TBEV-specific immunoglobulin G (IgG) antibodies were detected using the Anti-TBE Virus ELISA (IgG) kit (Euroimmun). Sera were diluted 1:64 and then 1:101 prior to the first incubation using the Sample Buffer of the kit. Goat anti-mouse IgG antibodies (Invitrogen) diluted 1:2000 in BSA 0.7% (KPL) as secondary antibodies were used. Plates were read using the Sunrise reader (Tecan) at a wavelength of 450nm.
Sera were incubated for 30min at 56°C prior to viral serology. For each serum, a 96-well plate culture of confluent BHK21 cells was inoculated with 50μL/well of WT_IC virus (final calculated moi: 0.001) and 50μL/well of a serial 2-fold dilution (first dilution at 1:40) of serum. Each row included 5 wells of serum dilution, a positive control (virus only) and a negative control with neither virus nor serum. The plates were incubated for 7 days and read for the absence or presence of a CPE in each well. The 50% plaque reduction neutralization titre (PRNT50/mL) was determined using the method of Reed and Muench [56].
Kaplan-Meier survival analysis with Mandel-Cox’s Logrank tests, Student’s t tests and Fisher’s exact tests were performed using SPSS software package (IBM). p values below 0.05 were considered significant.
|
10.1371/journal.ppat.1000831 | Rapid Evolution of Pandemic Noroviruses of the GII.4 Lineage | Over the last fifteen years there have been five pandemics of norovirus (NoV) associated gastroenteritis, and the period of stasis between each pandemic has been progressively shortening. NoV is classified into five genogroups, which can be further classified into 25 or more different human NoV genotypes; however, only one, genogroup II genotype 4 (GII.4), is associated with pandemics. Hence, GII.4 viruses have both a higher frequency in the host population and greater epidemiological fitness. The aim of this study was to investigate if the accuracy and rate of replication are contributing to the increased epidemiological fitness of the GII.4 strains. The replication and mutation rates were determined using in vitro RNA dependent RNA polymerase (RdRp) assays, and rates of evolution were determined by bioinformatics. GII.4 strains were compared to the second most reported genotype, recombinant GII.b/GII.3, the rarely detected GII.3 and GII.7 and as a control, hepatitis C virus (HCV). The predominant GII.4 strains had a higher mutation rate and rate of evolution compared to the less frequently detected GII.b, GII.3 and GII.7 strains. Furthermore, the GII.4 lineage had on average a 1.7-fold higher rate of evolution within the capsid sequence and a greater number of non-synonymous changes compared to other NoVs, supporting the theory that it is undergoing antigenic drift at a faster rate. Interestingly, the non-synonymous mutations for all three NoV genotypes were localised to common structural residues in the capsid, indicating that these sites are likely to be under immune selection. This study supports the hypothesis that the ability of the virus to generate genetic diversity is vital for viral fitness.
| Since 1995, norovirus has caused five pandemics of acute gastroenteritis. These pandemics spread across the globe within a few months, causing great economic burden on society due to medical and social expenses. Norovirus, like influenza virus, has over 40 genotypes circulating within the population at the same time. However, it is only a single genotype, known as genogroup II genotype 4 (GII.4), that causes mass outbreaks and pandemics. Very little research has been conducted to determine why GII.4 viruses can cause pandemics. Consequently, we compared the evolution properties of several pandemic GII.4 strains to non-pandemic strains and found that the GII.4 viruses were undergoing evolution at a much higher rate than the non-pandemic norovirus strains. This phenomenon is similar to influenza virus, where an increase in antigenic drift has been associated with increased outbreaks. This discovery has important implications in understanding norovirus incidence and also the development of a vaccine and treatment for norovirus.
| Norovirus (NoV), a member of the Caliciviridae family, is now considered the most common cause of viral gastroenteritis outbreaks in adults worldwide [1]. In the US, NoV has been identified as the cause of over 73% of outbreaks of gastroenteritis [1]. Furthermore, outbreak NoV strains spread rapidly causing great economic burden on society due to medical and social expenses. Consequently, a vaccine or treatment for NoV would be useful in reducing its transmission and alleviating disease symptoms. Our current knowledge of NoV replication and evolution has made it difficult to predict the efficacy of a treatment or longevity of a vaccine, as evidence is emerging that NoV, like many other RNA viruses, exists as a dynamic, rapidly evolving and genetically diverse population [2],[3],[4]. The high level of genetic diversity in RNA viruses is recognised as the basis for their ubiquity and adaptability [5]. Therefore, in order to develop a successful treatment or control program it is first necessary to understand the mechanisms behind NoV replication and evolution.
NoV is a small round virion of 27–38 nm in diameter and possesses a single-stranded, positive-sense, polyadenylated, RNA genome of 7400–7700 nucleotides [6]. The human NoV genome is divided into three open reading frames (ORFs). ORF1 encodes for the non-structural proteins, including an NTPase, 3C-like protease and RNA-dependent RNA polymerase (RdRp) [7]. The two structural proteins VP1, the major capsid protein, and VP2, the minor capsid protein are encoded by ORF2 and ORF3, respectively [8],[9].
NoV is a highly diverse genus with up to 61% VP1 amino acid diversity between its five genogroups (GI to GV) [10]. Up to 44% amino acid diversity over VP1 is also observed within the genogroups and has resulted in the further subgrouping of GI, GII and GIII into 8, 17 and 2 genotypes, respectively [10]. VP1 exhibits the highest degree of sequence variability in the genome [11],[12]. It consists of three domains, namely the shell (S) domain connected by a flexible hinge (P1 domain) to a protruding domain (P2) [13]. The highly conserved S domain forms the backbone of the capsid structure [13], while the moderately conserved P1 domain encodes the flexible hinge that connects the S and P2 domains. The protruding P2 domain possesses motifs that are involved in binding to the host cell, and hence, the P2 domain is responsible for the antigenicity of the virus [14],[15].
The most clinically significant of the five genogroups is GII, as it is the most prevalent human NoV genogroup detected and more frequently associated with epidemics compared with other genogroups. Of particular interest is GII genotype 4, (GII.4), because this lineage accounts for 62% of all NoV outbreaks globally [14],[15] and has also caused all five major NoV pandemics in the last decade (1995/1996, U5-95_US strain; 2002, Farmington Hills; 2004, Hunter; 2006, 2006a virus; and 2007, 2006b virus) [16],[17],[18],[19]. The basis for the increased epidemiological fitness [20] of the GII.4 strains, as determined by its high incidence and ability to cause pandemics, is currently unknown. Investigations with influenza indicate a link between increased viral evolution and increased viral incidence [21],[22]. However, because of the non-culturable nature of human NoV, variations in rates of evolution have not been calculated for different NoVs and consequently this has not been investigated as a factor in determining viral incidence and epidemiological fitness.
Replication efficiency and genetic diversity are both important parameters in viral fitness [23]. The aim of this study was to determine if these two parameters are contributing to the increased epidemiological fitness of the GII.4 strains. Replication efficiency and genetic diversity are primarily determined by the viral RdRp, as it controls the rate new sequence is introduced into the genome. Therefore using in vitro RdRp assays together with bioinformatics, the replication efficiency, mutation rate and rate of evolution of GII.4 viruses was compared with other NoV GII genotypes. The results of this study suggest that, like influenza A, the increased incidence of the pandemic GII.4 lineage may be a result of the combined influence of a high mutation, replication and evolution rate which, together culminate in an increased epidemiological fitness for the GII.4 strains.
Stool samples containing NoV were obtained from the Department of Microbiology, Prince of Wales Hospital, Sydney, Australia, with the exception of the stool specimen that contained NoV/Mc17/01/Th (GenBank accession numbers AY237413). This stool specimen was obtained from McCormic Hospital, Chiang Mai, Thailand [16].
The six genetically diverse NoV strains used in this study included: three GII.4 pandemic strains; NoV/Sydney 348/97/AU (of the NoV/US95_96 GII.4 pandemic lineage) [16], NoV/NZ327/06/NZ (NoV/2006a GII.4 lineage) [17] and NoV/NSW696T/06/AU (NoV/2006b GII.4 lineage) [17]. Two recombinant strains; NoV/Sydney C14/02/AU (GII.b ORF1 and GII.3 ORF2/3 [commonly referred to as GII.b/GII.3]) [16] and NoV/Sydney4264/01/AU (GII.4 ORF1 and GII.10 ORF2/3, [GII.4/GII.10]) [16], and a GII.7 NoV, NoV/Mc17/01/Th associated with rare sporadic cases of gastroenteritis [24]. In this study, the RdRp enzymes are referred to by their genotype, except in the case of the GII.4 strains, which are referred to by their pandemic name, eg. GII.4 2006b-RdRp (see Table 1). RdRps from recombinant strains are indicated by an ‘r’ in front of the nomenclature.
Viral RNA was extracted from 140 µl of 20% faecal suspension using the QIAmp Viral RNA kit according to manufacturers' instructions (Qiagen, Victoria, Australia). RNA was resuspended in 50 µl of Baxter Steri-pour H2O and stored at −80°C. cDNA synthesis was performed as described previously [16].
The full length capsid gene, P2 domain and RdRp regions were amplified with specific primers (Table 2) using reverse transcriptase - polymerase chain reaction (RT-PCR) methods described in [17]. The amplified RdRp genes were cloned into pGEM-T Easy vector (Promega, Wisconsin, United States).
Plasmids and PCR products were purified by PEG precipitation and washed with 70% ethanol. Products were sequenced directly on an ABI 3730 DNA Analyzer (Applied Biosystems, Foster City, CA, US) using dye-terminator chemistry.
pGEM-T Easy vectors containing 1736 bp from the 3′ end of ORF1 were purified using the Quantum prep® plasmid miniprep kit (BioRad, California, United States) and used as template DNA for the construction of expression vectors. Strain specific primers incorporating restriction enzyme sites, were designed to amplify the precise RdRp region of each strain (Table 2). PCR was performed as described previously [17]. PCR products were digested with their corresponding restriction enzymes and cloned into the expression vector pTrcHis2A (Invitrogen, Mount Waverley, Australia). Constructs containing the hepatitis C virus (HCV) genotype 3a RdRp (pVRL69) and HCV genotype 1b RdRp (pVRL75), were used as controls and have been described previously [25].
Site directed mutagenesis of residue 291 in the GII.4 US95_96-RdRp and the GII.4 2006a-RdRp was carried out with the Stratagene Quickchange II mutagenesis kit, according to manufacturer's instructions (Stratagene, La Jolla, United States). The primers used to introduce the mutation into the plasmid are listed in Table 2.
The NoV RdRps and control HCV RdRps were expressed in Escherichia coli, as described previously [25], except expression of the NoV RdRps was performed for 4 hr at 30°C. Purity was checked by SDS-PAGE and the identity of the RdRp was confirmed by western blot with an anti-six histidine antibody and peptide sequencing performed by the Bioanalytical Mass Spectrometry Facility (University of New South Wales, Australia). Recombinant RdRp was quantified with a Nanodrop ND-1000 Spectrophotometer (Nanodrop, Wilmington, United States).
Kinetic RdRp assays were performed in a final volume of 15 µl and contained 20 mM Tris-HCl (pH 7.4), 2.5 mM MnCl2, 5 mM DTT, 1 mM EDTA, 500 ng of homopolymeric C RNA template, 2 U RNasin (Promega), 4 mM sodium glutamate and increasing concentrations of [3H]-GTP (Amersham Biosciences, Little Chalfont, UK) ranging from 2 µM to 60 µM. Reactions were initiated with the addition of 50 nM of RdRp and incubated for 9 mins at 25°C. The reactions were terminated by adding EDTA to a final concentration of 60 mM, 10 µg herring sperm DNA and 170 µl of 20% (w/v) trichloroacetic acid. The incorporated radionucleotides were precipitated on ice for 30 min and then filtered through a 96 well GF/C unifilter microplate (Falcon, Franklin Lakes, United States) by a Filtermate harvester (Packard BioSciences, Melbourne, Australia). Using the harvester, the filters were washed thoroughly with water and left to dry. The filter wells were each filled with 25 µl of Microscint scintillation fluid (Packard Biosciences) and radioactivity measured using a Packard liquid scintillation counter (TopCount NXT; Packard Biosciences). Background measurements for each assay consisted of reactions without RdRp and were subtracted from the count per minute (CPM) values obtained for the individual enzyme assays. Results were plotted and statistical analysis performed with the Mann Whitney Test (one-tailed, 95% confidence interval) in GraphPad Prism version 4.02 (GraphPad Software, San Diego, CA).
An in vitro fidelity assay was developed to measure mutation rates and was adapted from Ward et al. [26]. The RdRp assay was performed using conditions described above with a homopolymeric C RNA template, except 82.1 pmoles of [3H]UTP (2 µCi) or [3H]ATP (4 µCi) (Amersham Biosciences) were added (as the non-complementary nucleotides) with an equimolar amount of GTP (82.1 pmoles) (Promega) added as the complementary nucleotide. The total amount of ribonucleotide incorporated was calculated in a parallel experiment with the addition of 1 µCi (164.2 pmoles) [3H]GTP (Amersham Biosciences) as the correct nucleotide. The assay was incubated for 50 min at 25°C. Error frequency of the RdRp was determined by calculating the total number (pmoles) of non-complementary ribonucleotides incorporated and dividing by the total number (pmoles) of [3H]GTP ribonucleotides incorporated.
In order to determine the rate of evolution of the rGII.3, GII.3, GII.4 and GII.7 capsids, the nucleotide sequences of ORF2 were analysed. RNA capsid sequences used for the analysis included eight from this study and 76 sequences from GenBank, with the oldest strains available dating back to 1987. The strains used and their GenBank accession numbers are listed in Text S1. The rate of evolution (substitutions/nucleotide site/year) for GII.3, GII.b/GII.3 GII.4 and GII.7 NoVs was determined by calculating the number of nucleotide substitutions in ORF2 compared to an ancestral strain and this was plotted against time [27]. The rate of evolution was determined by linear regression with the program GraphPad PRISM® version 4 and was equivalent to the gradient of the line. Pairwise alignments of RNA sequences and evolutionary distances between sequences were carried out using the Maximum Composite Likelihood model in Mega 4.0 [28]. Bootstrapped trees (1000 data sets) were constructed using the Neighbour-joining method, also with the program Mega 4.0.
In order to determine the amount of selection each genotype is under, the average Ka/Ks ratio was calculated for each genotype's capsid gene (GII.4, GII.b/GII.3 and GII.7). The Ka/Ks ratio is a measure of nonsynonymous amino acid changes compared to synonymous (silent) changes. Ka/Ks>1 indicates that positive selection is occurring. Ka/Ks = 1 is interpreted as neutral evolution and Ka/Ks<1 is indicative of negative or purifying selection. The program Sliding Windows Alignment Analysis Program (SWAAP) version 1.0.2 [29] was utilised. The Nei-Gojobori model was used to calculate Ka and Ks values [30]. The window size was set at 15 bp (5 aa) and the step size was 3 bp (1 aa).
Predicted secondary structure analysis of the RdRps and capsid protein VP1 were performed by generating a Protein Data Bank (PDB) file from the amino acid sequence in FastA format using software on the CPHmodels 2.0 Server [31]. Three dimensional structures were then generated from the PDB files with PyMol [32].
The GenBank accession numbers for the RdRp and capsid genes described in this paper are listed in Text S1.
Over the last decade five NoV pandemics have occurred approximately every two years and all pandemics have been associated with a single NoV genotype, GII.4 [16],[17],[19],[34]. The reason for the predominance of the GII.4 strains has been the subject of much speculation but is currently unknown primarily due to a limited understanding of NoV population dynamics and evolution [4],[15],[35].
Studies with other RNA viruses indicate that viral fitness is dependent on many factors, such as, viral mutation, replication efficiency, population size and host factors (reviewed in [2]). To date progress has been made in understanding the role host factors have on NoV prevalence with several studies indicating that variations in viral docking to the blood group antigens may affect infectivity of individuals within a population (reviewed in [36]). In particular, GII.4 viruses bind to all blood group antigens, whereas, GII.1 and GII.3 viruses bind fewer blood group antigens and this could account for higher prevalence of GII.4 viruses [36]. This paradigm however remains controversial, especially for GII NoV, as not all studies show an association between blood group antigens and clinical infection [37],[38],[39].
Apart from the host/viral interaction, no other factors have been affiliated with NoV fitness. Recent studies performed with poliovirus have shown that an increase in fidelity leads to less genetic diversity and subsequently a reduction in viral fitness and pathogenesis because of a reduced adaptive capacity of the virus [40],[41]. It has been hypothesised that viruses are fitter if they are able to produce a more robust (diverse) population (reviewed in [42],[43],[44]). In the current study we examined whether there was a link between epidemiological fitness, as defined by their incidence, and the rate and accuracy of viral replication.
In the present study error rates were assessed directly by examining the mutation rate of the viral RdRp and by analysing the rate of evolution for selected GII lineages. Our results are consistent with mutation rates for the poliovirus RdRp [26] and retrovirus reverse transcriptases [45], which range between 10−3 to 10−5 (Table 1). The more prevalent GII.4 strains had a 5 to 36-fold higher mutation rate compared to the less frequently detected GII.b/GII.3 and GII.7 strains, as determined by in vitro enzyme assays. Consistent with this, the rate of evolution of the capsid was on average 1.7-fold higher in GII.4 viruses compared to GII.3, GII.b/GII.3 and GII.7 viruses. The GII.4 capsids also had a larger Ka/Ks ratio than the GII.b/GII.3 and GII.7 strains suggesting that the increased incidence/epidemiological fitness of the GII.4 strains maybe through greater antigenic drift, a consequence of the higher mutation rate of the GII.4 RdRp.
The mutation rates for the control HCV RdRps (average of 1.6×10−3 substitutions per nucleotide site, Table 1) were 2-fold higher compared to the GII.4 RdRps. Evaluation of previously published rates of evolution for the HCV hypervariable region 1 (HVR1) within the envelope 2 glycoprotein (E2) were also higher (6–fold) than the NoV GII.4 rates of evolution calculated in this study [46] (Table 1). HVR1 was chosen for comparison because, like the NoV capsid gene, it is the most variable region in the genome and under the greatest immune selection. Mutation rate and rate of evolution cannot be directly compared as they are indirectly related due to the increased complexity of evolution in vivo [20]. However, in this study we did find a common trend between the two different measurements of diversity with HCV displaying the highest diversity rate for both measurements compared to NoV.
Interestingly, the majority of non-synonymous mutations in the P2 domain for all three NoV genotypes were localised to six common structural sites. These six hypervariable regions within the P2 domain were consistent with hypervariable sites for GII.4 capsids already identified in other studies [4],[19]. We demonstrated that GII.7 and GII.3 viruses shared two and four common hypervariable sites, respectively, with GII.4 viruses (Fig. 5). Substitutions at one of these sites (residue 395) have been shown to alter GII.4 strains antigenic profiles [4]. Localization of the hypervariable sites to common regions on the surface of the P2 domain suggests that these regions are likely to be under immune pressure possibly from a neutralizing antibody response [39]. The lower number of amino acid changes at these sites for viruses with a GII.3 capsid may explain why GII.b/GII.3 is predominantly associated with gastroenteritis cases in children [47]. This suggests that GII.b/GII.3 viruses are not as efficient at escaping herd immunity compared to GII.4 strains and therefore only hosts immunologically naïve to GII.3 infection are susceptible. Similarly, we propose that the low prevalence of the GII.7 strain is also a consequence of a low mutation rate in the RdRp resulting in limited antigenic drift and an inability to escape herd immunity.
Apart from mutation rate, replication rate is considered to be another major determinant in viral fitness [48]. Replication rates are important because an increased replication rate would produce a larger heterogenous population than a slower replicating virus in the same unit of time, given the same mutation rate. Interestingly, the RdRps from the recent 2006 GII.4 pandemic strains had a higher nucleotide incorporation rate than the recombinant GII.4 RdRp and the US95/96-like pandemic GII.4 RdRp, which could be associated with a point mutation in the RdRp (Thr291Lys). Residue 291 is located in the finger domain, which is comprised of five β sheets that run parallel and strongly interact with each other. The innermost of these five β sheets contains motif F which interacts directly with incoming nucleotides [49]. Therefore, it is plausible that substitutions at residue 291 affects the orientation of motif F due to the strong interaction between the five β sheets and subsequently alters the binding affinity to the incoming nucleotide. Fixation of the Thr291Lys point mutation in the GII.4 lineage after 2001 has been paralleled with a reduction in the period of stasis between the emergence of new antigenic variants [4]. Alterations in residue 291 after 2001 could have led to an increase in the rate of evolution of GII.4 strains by increasing the replication rate, however this did not seem to have an effect on mutation rate (Table 1). High replication rates did not always correlate with epidemiological fitness as the NoV strain, GII.7, had the highest incorporation rate but is considered to be the least fit due to it having the lowest incidence. Therefore, this study suggests mutation rate in combination with a high replication rate are key determinates in epidemiological fitness.
Influenza research also indicates a relationship between rate of evolution and epidemiological fitness (reviewed in [21]). New antigenic influenza A variants arise every one to two years and cause more annual epidemics than influenza B, as well as the more devastating pandemics [21]. Once a population has accumulated mass herd immunity to a virus the virus is forced to alter its antigenic determinants, a possibility for viruses with poor fidelity and fast replication rates, or face extinction [50], whereas, viruses such as influenza B, which have higher fidelity and slower antigenic change, are more often associated with sporadic cases [21]. In this study a parallel can be seen in the epidemiology between NoV and influenza, in particular between GII.b/GII.3 viruses and influenza B and GII.4 viruses and influenza A.
In summary, this study supports the hypothesis that epidemiological fitness is a consequence of the ability of the virus to generate genetic diversity, as the NoV pandemic GII.4 strains were associated with an increased replication and mutation rate. Therefore, it would seem that GII.4 viruses, as opposed to GII.b/GII.3 and GII.7 viruses, have reached a balance in their replication rate and mutation rate that is better suited to viral adaptation. In contrast, it would seem that the GII.7 lineage, despite having a high replication rate, has a low mutation rate that limits its adaptation and therefore its incidence. It is important to improve our understanding of the mechanisms underlying NoV epidemiological fitness as future pandemics are expected.
|
10.1371/journal.pmed.1002750 | Patterns of joint involvement in juvenile idiopathic arthritis and prediction of disease course: A prospective study with multilayer non-negative matrix factorization | Joint inflammation is the common feature underlying juvenile idiopathic arthritis (JIA). Clinicians recognize patterns of joint involvement currently not part of the International League of Associations for Rheumatology (ILAR) classification. Using unsupervised machine learning, we sought to uncover data-driven joint patterns that predict clinical phenotype and disease trajectories.
We analyzed prospectively collected clinical data, including joint involvement using a standard 71-joint homunculus, for 640 discovery patients with newly diagnosed JIA enrolled in a Canada-wide study who were followed serially for five years, treatment-naïve except for nonsteroidal anti-inflammatory drugs (NSAIDs) and diagnosed within one year of symptom onset. Twenty-one patients had systemic arthritis, 300 oligoarthritis, 125 rheumatoid factor (RF)-negative polyarthritis, 16 RF-positive polyarthritis, 37 psoriatic arthritis, 78 enthesitis-related arthritis (ERA), and 63 undifferentiated arthritis. At diagnosis, we observed global hierarchical groups of co-involved joints.
To characterize these patterns, we developed sparse multilayer non-negative matrix factorization (NMF). Model selection by internal bi-cross-validation identified seven joint patterns at presentation, to which all 640 discovery patients were assigned: pelvic girdle (57 patients), fingers (25), wrists (114), toes (48), ankles (106), knees (283), and indistinct (7). Patterns were distinct from clinical subtypes (P < 0.001 by χ2 test) and reproducible through external data set validation on a 119-patient, prospectively collected independent validation cohort (reconstruction accuracy Q2 = 0.55 for patterns; 0.35 for groups).
Some patients matched multiple patterns. To determine whether their disease outcomes differed, we further subdivided the 640 discovery patients into three subgroups by degree of localization—the percentage of their active joints aligning with their assigned pattern: localized (≥90%; 359 patients), partially localized (60%–90%; 124), or extended (<60%; 157). Localized patients more often maintained their baseline patterns (P < 0.05 for five groups by permutation test) than nonlocalized patients (P < 0.05 for three groups by permutation test) over a five-year follow-up period.
We modelled time to zero joints in the discovery cohort using a multivariate Cox proportional hazards model considering joint pattern, degree of localization, and ILAR subtype. Despite receiving more intense treatment, 50% of nonlocalized patients had zero joints at one year compared to six months for localized patients. Overall, localized patients required less time to reach zero joints (partial: P = 0.0018 versus localized by log-rank test; extended: P = 0.0057).
Potential limitations include the requirement for patients to be treatment naïve (except NSAIDs), which may skew the patient cohorts towards milder disease, and the validation cohort size precluded multivariate analyses of disease trajectories.
Multilayer NMF identified patterns of joint involvement that predicted disease trajectory in children with arthritis. Our hierarchical unsupervised approach identified a new clinical feature, degree of localization, which predicted outcomes in both cohorts. Detailed assessment of every joint is already part of every musculoskeletal exam for children with arthritis. Our study supports both the continued collection of detailed joint involvement and the inclusion of patterns and degrees of localization to stratify patients and inform treatment decisions. This will advance pediatric rheumatology from counting joints to realizing the potential of using data available from uncovering patterns of joint involvement.
| Juvenile idiopathic arthritis (JIA) is a heterogeneous childhood disease for which joint inflammation is the common denominator.
The current classification system, the International League of Associations for Rheumatology (ILAR) subtypes, categorize patients according to the number of active joints (four or fewer versus five or more affected joints in the first six months of disease) with little evidence to support this cut-off.
Grouping frequently co-involved joints into joint patterns may help to better classify patients and predict disease course.
We analyzed baseline joint involvement data from 640 treatment-naïve patients from 16 Canadian centers participating in the Research in Arthritis in Canadian Children, Emphasizing Outcomes (ReACCh-Out) study between 2005 and 2010.
We identified seven joint patterns that grouped frequently co-involved joints using multilayer non-negative matrix factorization (NMF), an unsupervised pattern recognition technique. We named these patterns as follows: pelvic girdle, fingers, wrists, toes, knees, ankles, and indistinct.
These seven patterns described more homogenous groupings of joints than the ILAR subtypes and further stratified them.
Throughout disease course, we found that patients with joint involvements largely overlapping their baseline patterns, i.e., those having localized involvement, continued to have the same active joints, whereas patients with nonlocalized involvement did not.
Patients with localized joint involvement reached zero joint involvement faster than those with nonlocalized involvement.
Patterns of joint involvement represent prognostic features that should be incorporated into a comprehensive JIA disease classification.
Patients with nonlocalized joint involvement are at high risk of poor outcome.
Grouping patients by joint pattern and degree of localization may help clinicians tailor treatments based on predicted disease trajectories.
| Juvenile idiopathic arthritis (JIA) is the most common chronic inflammatory rheumatic disease in childhood, characterized by joint inflammation of unknown etiology lasting at least six weeks starting at <16 years of age. The International League of Associations for Rheumatology (ILAR) criteria classify JIA into seven subtypes [1].
Clinicians acknowledge that individual joints or groups of joints influence outcome. Distinct joint involvement patterns such as dactylitis, sacroiliac joint involvement, or tarsitis are clearly recognized in patients with JIA [1]. However, despite efforts to characterize distinct disease entities, these patterns remain heterogeneous in clinical presentation and consequently disease course and response to treatment. In addition, individual joints such as the hip, cervical spine, ankle, or wrist may be indicators of poor outcome [2–5] and have therefore been considered in the 2011 American College of Rheumatology (ACR) treatment recommendations as indicator joints for poor prognosis [6–8]. However, existing classification subtypes and treatment recommendations do not take affected joint patterns into account [1,6]. Moreover, these findings only apply to a subset of JIA patients and are based on a small number of joints [9,10]. A significant gap remains in our understanding of the relationship between patterns of joint involvement and disease outcome.
Systematically collected data from the Research in Arthritis in Canadian Children, Emphasizing Outcomes (ReACCh-Out) study have enabled us to address this knowledge gap. ReACCh-Out is a prospective inception cohort study of children with newly diagnosed JIA [11] that has rigorously collected detailed longitudinal clinical data, including information on joint inflammation. A standard 71-joint homunculus precludes the use of traditional statistical—or supervised learning—techniques to associate individual joints, let alone combinations of joints, to descriptors of outcome in an unbiased manner due to multiple hypothesis testing in the context of small patient numbers, a common challenge in rare diseases. To address this challenge, unsupervised learning can help identify a small number of patterns of arthritis in a principled, logical, and expressive manner, whose outcomes we can then explore. We can draw from previous successes applying these approaches in JIA analyzing clinical and biomarker data [12,13]. These patterns not only provide possible predictive features of outcome but also help to organize the patients into homologous groups that may underlie further disease research.
In this study, we pursued a data-driven strategy in a novel data domain, joint involvements, to identify two easily measurable clinical variables—joint pattern and degree of localization—in children with newly diagnosed JIA. We then determined the relationship of these measures to longitudinal outcomes, including pattern stability and response to treatment.
Our analysis, outlined as a flow chart in S1 Fig, can be divided into the following phases: data collection and filtering, dimensionality reduction, clustering, and postclustering analysis. As indicated in our transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) checklist (S1 Checklist), our postclustering analysis plan was developed after we completed the clustering and observed that some patients defied identification with a single joint grouping. This observation prompted us to include a variable describing this in our multivariate prediction model.
For the discovery cohort, patients enrolled in ReACCh-Out were included if they satisfied the classification criteria for any of the seven ILAR JIA subtypes [1]. Enrolled patients were included within one year of JIA diagnosis and were treatment naïve for medications except for nonsteroidal anti-inflammatory drugs (NSAIDs). ReACCh-Out is a prospective inception cohort of newly diagnosed patients with JIA recruited between 2005 and 2010 from 16 tertiary Canadian centers. These 16 centers included 14 academic and two community centers: Royal Jubilee Hospital, Victoria, British Columbia; BC Children’s Hospital, Vancouver, British Columbia; Penticton Regional Hospital, Penticton, British Columbia; University of Calgary, Calgary, Alberta; Alberta Health Services, Edmonton, Alberta; Royal University Hospital, Saskatoon, Saskatchewan; Health Sciences Centre, Winnipeg, Manitoba; The Hospital for Sick Children (SickKids), Toronto, Canada; Children’s Hospital of Eastern Ontario, Ottawa, Ontario; Montréal Children’s Hospital, Montréal, Québec; Université de Montréal, Montréal, Québec; Université de Sherbrooke, Sherbrooke, Québec; le Centre Hospitalier Universitaire de Québec, Québec, Québec; Santé et Services Sociaux, Québec, Québec; IWK Health Centre, Halifax, Nova Scotia; and Janeway Children’s Hospital and Rehabilitation Centre, St. John’s, Newfoundland [11].
The nonoverlapping independent validation cohort comprised prospectively collected patients with newly diagnosed chronic childhood arthritis and detailed information about joint involvement recruited between 1980 and 2007 from two tertiary Canadian sites: Royal University Hospital of Saskatoon, Saskatchewan and Health Sciences Centre of Winnipeg, Manitoba [14]. These sites were chosen because they had standardized collection of prospective data by the same personnel over the study timeframe. For this study, all patients were treatment naïve except for NSAIDs. Prior to the introduction of the ILAR classification criteria, patients in the validation cohort were identified as satisfying ACR criteria for juvenile rheumatoid arthritis (JRA).
Research ethics boards at each participating center approved the study protocols. Informed written consent for participation was obtained from parents, and informed consent or assent was obtained from patients as appropriate.
For the discovery cohort, detailed demographic, clinical, and laboratory data were collected prospectively at study entry using standardized clinical reporting forms. Collected information included key features of the ILAR classification [1], the ACR pediatric core set measures of disease activity [15], standard laboratory markers, and physician assessments of global disease activity. All forms were subsequently validated. Musculoskeletal information for 71 joints as well as treatment regimens were serially collected at six-month intervals for the first 18 months and yearly thereafter for five years after study enrollment.
In the validation cohort, the features of the ILAR classification were collected as well as joint involvements at all subsequent clinical visits.
All analyses were conducted using R version 3.3.0 (Vienna, Austria) and Python version 3.6 and higher (Wilmington, Delaware, US) on Apple computers running Mac OS X 10.10 and higher, as well as at supercomputing facilities at the High Performance Facility at SickKids (Toronto, Ontario, Canada) and the Terrence Donnelly Centre for Cellular and Biomolecular Research at the University of Toronto (Toronto, Ontario, Canada).
Initially, joint co-involvements—which describe pairs of joints that are active together—were analyzed by calculating joint co-involvement frequencies and conditional joint co-involvement frequencies. A joint co-involvement frequency, P(x,y), for a reference joint x and a co-involved joint y is the proportion of patients having both joints involved. The conditional joint co-involvement frequencies, P(y|x)=P(y,x)P(x), are the fraction of all patients with x involved who also have y involved.
To investigate whether joint co-involvement frequencies were asymmetric, i.e., higher for joints on the same side of the body, we developed a measure of same-side versus opposite-side skew in the co-involvement frequencies computed across the discovery cohort. A “joint type” is the pair of corresponding joints on opposite sides of the body, e.g., the knee joint type consists of the left and right knee joints. To compute skew for a reference joint type x and a co-involved joint type y, we counted the number of patients for whom a reference joint was paired with a co-involved joint on the same side (nsame) and on the opposite side (nopposite) and used these to compute a z-score to measure skew for the joint type pair by setting z=m−0.5σ, such that m=nsame+1nsame+1+nopposite+1 and σ=m(1−m)n. If the joint type pair co-involvement was symmetric, z would be close to zero, whereas if co-involved joints tended to occur on the same side of the body, z would be large and positive. To assess the significance of the skew, χ2 tests were conducted using nsame and nopposite under the null hypothesis that both quantities were equal. P values were translated to false discovery rates (FDRs) to account for multiple hypothesis testing. FDR < 0.1 was considered significant.
We selected an analysis strategy suitable for the positive joint counts that would be able to identify sparse patterns of joint involvement if they were present in the data. Sparse patterns better support interpretation and application of these patterns by clinicians. Also, we sought a mixed membership model so that patients could belong to multiple groups. Finally, our preliminary analysis suggested a hierarchical structure to the joint co-involvement patterns, with some tightly coupled joint groups (e.g., toe joints) that often co-occur along with one or more broader patterns of co-activation of these tight groups.
As such, we adapted a form of multilayer non-negative matrix factorization (NMF) [16] to learn hierarchical, sparse representations (S1C Fig). Multilayer NMF progressively applies NMF to the joint patterns. NMF is a dimensionality reduction method that fits summary indicators or “factors” that group frequently co-involved joints together [17]. Broadly, NMF decomposes joint involvements X into a basis/loading matrix W, describing the contributions of joints to factors, and a coefficient/score matrix H that scores patients on factors. Multiplying these two matrices approximates the input data (X ≈ WH). NMF produces an intuitive parts-based representation in which reconstructing patient data involves only adding groups of joints in factors, similar to adding parts of faces (different eyes or noses) to reconstruct facial representations [17]. Unlike other dimensionality reduction techniques like principal component analysis (PCA), NMF constrains the elements of W and H to be positive. This constraint often causes many elements of the two matrices to be zero, a trend that we supplement by setting small elements to zero (a technique called “sparsifying”), as described in S1 Text. NMF can be used to cluster patients by interpreting the nonzero elements of H for each patient as assignments to clusters represented by the factors in W [18].
Multilayer NMF was conducted as described in S1 Text. Briefly, NMF was applied to the joint involvement data, and W was then sparsified to produce low-level factors that correspond to tight joint groupings. NMF was then applied to the H matrix to identify tight joint groupings that often co-occur; the term “high-level factors” refers to these frequently co-occurring tight joint groupings. Through matrix multiplication, joint involvements can be recovered from the high-level factors. “Key joints” for each high-level factor were those appearing with nonzero contributions. For both the high- and low-level factors, patients were assigned into patient groups (“[x]”) corresponding to their highest-scoring factor [19].
The degree of overlap was assessed between patient groups at both levels of the analysis. Patient factor scores were normalized patient-wise to the highest factor score and then z-score–transformed factor-wise. One-sided z-tests determined which patient groups had higher scores than expected on individual factors. FDRs were calculated from P values to account for multiple hypothesis testing. Relationships between patient groups and factors were significant if their FDR was <0.1.
Relationships were visualized between patient groups and ILAR subtypes through a circular figure built using Circos 0.63 [20]. To identify enriched relationships between patient groups and ILAR subtypes, a χ2 test was conducted. Relationships were enriched if their standardized residual was ≥1.96 (i.e., P < 0.05).
To describe how closely patients aligned with their associated high-level factors or patient groups, patients were further stratified by the “degree of localization” of their active joints. Patients with “localized” involvement had ≥90% of active joints being key joints in the high-level factor underlying their patient group. For patients with “partially localized” involvement, this range was ≥60% and <90%. All other patients had “extended” involvement. S2 Text describes how these boundaries were determined.
To determine which patient groups skewed towards any localization, χ2 tests were conducted to compare the distribution of localizations within a single patient group against the global distribution. P values were Bonferroni-adjusted for multiple hypothesis testing.
For each medication at six-month and one-year visits, multivariable logistic regression was conducted to predict medication status as an outcome from both patient groups and degrees of localization. Model significance was assessed using likelihood-ratio tests. A model for a medication and visit was significant if P < 0.05 after a Bonferroni adjustment for multiple hypothesis testing.
To track how joint involvement changed over subsequent visits, high-level patient factor scores and patient group assignments were calculated where possible at six-month, one-year, 18-month, two-year, three-year, four-year, and five-year visits. From baseline patient group assignments and localizations, frequencies of transitioning between two patient groups at any time between six-month and five-year visits were calculated. Significantly overrepresented transitions were determined by a 2,000× permutation test.
A multivariate Cox proportional hazards analysis, as implemented in the survival R package, version 2.41 [21], was conducted to identify which patient groups, ILAR subtypes, and degrees of localization experienced zero joint involvement more quickly. The assumption of proportional hazards was assessed visually and tested using the “cox.zph” function in survival. Hazard ratios (HRs) and 95% confidence intervals (CIs) were computed and patient groups, ILAR subtypes, and degrees of localization were compared using log-rank tests.
To determine whether the identified patterns of joint involvement could generalize beyond the discovery cohort, validation cohort joint involvement data were projected onto high-level factors and patient groups. At each level, the same scaling parameters were applied to joint involvement data or low-level patient factor scores. Patients were assigned to patient groups based on their highest-scoring high-level factors as above. Patterns of joint involvement were also identified independently of discovery joint patterns using the multilayer NMF framework described above.
We included 640 patients with newly diagnosed JIA in the discovery cohort and 119 in the validation cohort. Table 1 outlines demographic data for these cohorts respectively. Discovery cohort patients were diagnosed at a median age of 7.7 years, with a range of 0.57 to 16.6 years, whereas validation cohort patients were diagnosed at a younger age (median: 5.7 years; range: 0.5–18 years). The most highly represented subtype in both cohorts was oligoarthritis (discovery: 47%; validation: 56%), and most patients were female (discovery: 65%; validation: 71%).
To investigate patterns of joint involvement and co-involvement, we computed overall frequencies of individual joint involvement and pair co-involvement in the discovery cohort. Fig 1 depicts overall joint involvement frequencies. Knees, ankles, and wrists had the highest rate of involvement. When we considered conditional co-involvement frequencies partitioned by side of body (Fig 2), 3,271 of 5,041 (65%) of these probabilities were significant (P < 0.05 after Bonferroni adjustment). Examining the heat maps for same-side joints in the top-left and bottom-right quadrants, we observed hierarchical patterns of joint co-involvement along the vertical axis. Joints closer on the vertical body axis were generally more often co-involved joints. For example, index finger joints and middle finger joints were likely to be co-involved. However, there also appeared to be broad, nonlocal patterns of co-involvement, e.g., finger and toe joints were frequently co-involved.
In contrast, the heat maps for opposite-side involvement in the top-right and bottom-left quadrants appeared nearly identical (Frobenius norm = 3.7; P < 0.001 by permutation test), indicating little proximity preference along the horizontal axis. For example, left index finger joints were as likely to be co-involved with right middle finger joints as those on the left middle finger. This observation was consistent with widespread symmetric joint involvement in JIA.
Asymmetric joint involvement is associated with more severe forms of JIA [22]. To study the prevalence of asymmetric co-involvements, we determined which joint types were more likely to be co-involved with joints on the same side of the body. Consistent with our initial impression of Fig 2, and different than what we expected based on previous reports in JIA, we found few statistically significant asymmetric joints: only ankle, midfoot, and subtalar joints (with ankle: χ2 = 14, PFDR = 0.0097; with subtalar joints: χ2 = 11, PFDR = 0.048) had statistically significant same-side co-involvement (S2 Fig).
To characterize hierarchical patterns of joint co-involvements, we applied multilayer NMF on discovery cohort joint involvements to identify groups of frequently co-involved joints. Conventional NMF identified 19 low-level groupings (or factors)—<1–19> (S6A Fig)—of vertically proximal joints, which S3 Text details. Consistent with the pairwise analysis (S2 Fig), most factors grouped joints of the same type (S6A Fig), with exceptions being pairs of groups containing only ankles and subtalar joints, only a single knee, and one group containing joints from the index (second) and middle (third) fingers on the right side. We named these 19 factors (“<x>”) as follows: <1 TMJs>, <2 shoulders>, <3 sternoclavicular joints>, <4 elbows>, <5 thumbs>, <6 sacroiliac joints>, <7 MCPs>, <8 second–third fingers>, <9 hips>, <10 subtalar and midfoot>, <11 finger PIPs>, <12 knee>, <13 knee>, <14 finger DIPs>, <15 subtalar and midfoot>, <16 grand toes>, <17 ankles>, <18 toe IPs>, and <19 MTPs>. Multiple low-level joint groups were significantly co-involved (S7A Fig), suggesting a hierarchical structure not captured by low-level factors.
To identify this hierarchical structure to the co-involvements, we conducted a second round of conventional NMF on low-level patient scores, which identified seven groupings of groups we deem “high-level factors”, <A–G>. S3 Text describes the process of deriving these high-level factors. As expected from our preliminary investigations (Fig 2), each high-level factor combined low-level factors into broader, symmetric groupings covering partially overlapping areas of the vertical body axis (S6 Fig). We named these high-level factors <A pelvic girdle>, <B fingers>, <C wrists>, <D toes>, <E ankles>, <F knees>, and <G sternoclavicular joints>. Finger distal interphalangeal (DIP) joints distinguished <B fingers> from <C wrists>, as <B fingers> skewed towards finger DIPs and <C wrists> towards shoulders and elbows. <A pelvic girdle> included sacroiliac joints and/or hips, and <E ankles> included ankles, subtalar, and midfoot joints.
We classified each patient into one of seven groups (“[x]”) corresponding to the high-level factors based on their highest-scoring high-level factor. Fig 3 shows joint involvement frequencies in each patient group, and S8 Fig depicts individual joint involvements for patients. Key joints—outlined in Fig 3—were those with nonzero weights in the corresponding high-level factors (S6 Fig). Non-key joints were rarely involved except in [G indistinct] and for some finger/wrist involvement in [D toes], suggesting that most patients had what we deem localized involvement; in other words, most or all of their joints overlapped with key joints. Overall, these patient groups corresponded to logical patterns of joint involvement reported at the bedside [23].
Six of seven patient groups associated with at least one ILAR subtype, despite patient groups comprising different stratifications of patients from the ILAR subtypes (χ2 = 313; P < 0.001). Conversely, six of seven ILAR subtypes associated with patient groups. More opaque ribbons in the Circos figure (Fig 4), linking patients common to patient groups and ILAR subtypes, represent these enriched associations encompassing more patients than expected through standardized residuals from the above χ2 test (S2 Table). These standardized residuals quantify how far the number of patients shared by a patient group and ILAR subtype deviated from expectation. Children in [A pelvic girdle] associated with enthesitis-related arthritis (ERA), [B fingers] with rheumatoid factor (RF)-negative polyarthritis; [C wrists] with systemic arthritis and both RF-positive and negative polyarthritis; [D toes] with RF-negative polyarthritis, psoriatic arthritis, and ERA; and [F knees] with oligoarthritis.
Although the high-level groupings completely encapsulated joint involvements for most patients (56%), a small group of patients (25%) had more non-key joints involved than key joints (S9 Fig). We deemed these patients as having extended involvement. Patients with 90% of their joints as key were localized, and those with 60% to 90% were partially localized. S4 Text describes how we determined these thresholds. To determine the clinical significance of these subcategories, we compared the clinical attributes of patients therein.
Patient groups with significantly skewed distributions of localizations included [C wrists], [D toes], [F knees], and [G indistinct] (χ2 ≥ 21.5; P < 0.001; S9C Fig and S3 Table). Patients in [G indistinct] skewed towards extended involvement, [C wrists] towards partially localized involvement, and [D toes] towards both partially localized and extended involvement. Children in [F knees] skewed towards localized involvement.
To determine whether treatment decisions were associated with patient groups and degrees of localization, we conducted multivariable logistic regression. S10A Fig depicts the number of patients in each group and localization with medication data at six-month and one-year visits. S10B Fig depicts the proportion of patients in each patient group, divided by localization, who received biologics, disease-modifying antirheumatic drugs (DMARDs), joint injections, and systemic corticosteroids prior to these visits. Patient groups differed by DMARD usage, with higher than expected utilization in children in [C wrists] and [E ankles] with localized involvement and [F knees] with partially localized and extended involvement, and lower than expected utilization in [F knees] with localized involvement (S4 Table). In terms of systemic corticosteroid usage, patients in [F knees] with localized involvements were less likely to receive such treatment at six-month visits. Therefore, patient joint group and localization influenced treatment.
To observe how the data-driven joint patterns evolved throughout disease course, we traced patient group assignments longitudinally. Fig 5 depicts transition probabilities from baseline patient groups, divided by degree of localization, to groups at any visit up to five years. While patients often transitioned to zero joint involvement, several transitions were enriched (P < 0.05; Holm-Bonferroni–adjusted), with little movement between patient groups given the lack of filled circles outside the diagonals. Among patients with localized involvements (Fig 5A), patients in [A pelvic girdle], [B fingers], [D toes], [E ankles], and [F knees] remained in their own patient group. Even among patients with partially localized involvements (Fig 5B), patients in [A pelvic girdle] often transitioned to zero joint involvement, whereas patients in [B fingers] and [C wrists] transitioned to or remained [C wrists]. Patients in [E ankles] often remained in the same patient group. Among patients with extended involvement (Fig 5C), patients in [A pelvic girdle], [E ankles], and [G indistinct] often remained in their respective patient groups.
Having observed a strong tendency for some groups towards zero joint involvement (disease control/inactive disease), we asked whether the groups differed by the rate in which they achieved this outcome. We constructed a Cox proportional hazards model predicting time to zero joint involvement from patient groups, localizations, and ILAR subtypes. The resulting model did not deviate from the proportional hazards assumption (χ2 = 22.9, P = 0.062) and identified localizations and ILAR subtypes that reached zero joint involvement at different rates than others (R2 = 0.093, P < 0.001). At least half of the patients followed in [F knees] reached this endpoint by six months, and [A pelvic girdle], [C wrists], [D toes], and [E ankles] reached it by one year (S11A Fig). In contrast, patients in [B fingers] reached this endpoint after 18 months, and those in [G indistinct] reached it after three years. Localization was especially important because patients with localized involvement achieved this outcome faster than patients with partial involvement (HR = 0.70, Z = −3.1, P = 0.0018) and extended involvement (HR = 0.63, Z = −2.8, P = 0.0057) (Fig 5D). Among diagnoses, patients with RF-positive polyarthritis (HR = 0.42, Z = 2.4, P = 0.016) and undifferentiated arthritis (HR = 0.64, Z = −2.8, P = 0.0059) reached this outcome more slowly. S5 Table contains additional statistics for the Cox proportional hazards model.
To determine whether factors and patient groups were generalizable beyond the discovery cohort, we projected an independent validation cohort of JIA patients to the joint patterns. These patients’ joint involvements were reconstructed by low-level factors with Q2 = 0.81, high-level factors with Q2 = 0.55, and patient groups with Q2 = 0.48 (S1 Table), comparing favorably against ILAR subtypes, with Q2 = 0.43. Projected patient groups presented with similar joint involvement frequencies (S12 Fig). We then calculated de novo factors and patient groups, which we detail in S4 Text. Validation joint involvements were reconstructed by low-level factors with Q2 = 0.84, high-level factors with Q2 = 0.55, and patient groups with Q2 = 0.35. De novo factors described similar joint groupings as discovery factors (S13C, S13F and S13G Fig).
We then asked whether the projected groups could also predict time to zero joint involvement. Considering individual univariate models due to limited power, patient groups (χ2 = 17, P = 0.0072) and localizations (χ2 = 8.9, P = 0.012; S14 Fig) themselves predicted this outcome. Patients with extended involvement took significantly longer to reach zero joint involvement than patients with localized involvement (HR = 0.53, Z = −2.5, P = 0.011).
We explored patterns of articular involvement in JIA using unsupervised data-driven pattern recognition techniques. We initially observed joints co-occurring in logical and localized groupings without same-side skewing (Figs 1 and 2 and S2 Fig). To better understand these signals in a clinically applicable manner, we conducted a modified version of multilayer NMF, identifying seven high-level groupings of joints (S6 Fig and Fig 3) describing arthritis foci anchored by distinct subsets of joints. The resulting seven patient groups subdivided the ILAR subtypes into distinct subgroups based on patterns of arthritis (Fig 4). Patients with localized involvement often remained in the same patient group after baseline visit (Fig 5A, 5B and 5C) and reached zero joint involvement faster than patients with nonlocalized involvement (Fig 5D). These patterns generalized to an independent validation cohort, supporting their applicability beyond the discovery cohort (S12 Fig and S13 Fig).
Our study is the first to provide a detailed, data-driven description of heterogeneously co-involved joints in JIA. Previous studies have focused on individual joints [2–8] in specific individual ILAR subtypes [9,10], whereas we identified joint groupings in a data-driven fashion independently of these ILAR subtypes. For example, our approach identified differences in presentation among patients with polyarticular involvement, separating these patients based on joint patterns. For example, the small joints of the fingers in [B fingers] and [C wrists] were clearly distinct from the small joints of the toes in [D toes]. The composition of patient groups in RF-negative polyarthritis further supported this distinction, with this ILAR subtype associating with [B fingers], [C wrists], and [D toes]. [B fingers] and [C wrists] may represent a spectrum of disease phenotypes (S7 Fig) that may transition between each other during disease evolution (Fig 5). Wrist involvement has been associated with poor prognosis and decisions to treat more aggressively [6]. Our results reflected this trend as patients in [C wrists] had less defined disease trajectories and longer times to zero joint involvement (S11 Fig) despite more commonly receiving DMARDs and systemic glucocorticoids. Similar findings were observed in patients in [B fingers] which tended to transition into [C wrists], with higher use of (biologic) DMARDs and systemic glucocorticoids at diagnosis. These findings were notable as patients in [B fingers] tended to transition to [C wrists] when they had partially localized involvement.
The stability of the patient groups for up to five years after diagnosis supports how meaningful these patterns are. As patients lose active joints over the course of treatment, we expected their joint patterns to shift as patient factor scores represent weighted sums of individual joints. With fewer active joints, we expected patterns to become more sensitive to the specific joints involved. However, patients remaining in their same groups in at least one subsequent visit suggested that patients with residual joint involvement had it in their group’s key joints. Therefore, joint patterns may represent robust core groupings of joints much like the indicator joints of poor outcome [2–7].
Our study is also among the first to identify the degree of active joint localization with outcomes through an easily measurable clinical variable, the degree of localization. Patients had worse disease outcomes if their active joints did not align strictly with a single pattern. This has been instinctively recognized at the bedside by clinicians, as this aspect has clearly influenced treatment decisions (S4 Table and S10 Fig), but is not included in any clinical guidelines. Furthermore, classifying patients by the degree of localization predicted disease course and time to zero joint involvement. For patients with recognizable patterns of joint involvement, treatment decisions appear effective. However, patients with partially localized or extended joint involvement had the poorest outcomes, taking a longer time to reach zero joint involvement despite receiving stronger medications (e.g., more DMARDs). Patients with nonlocalized joint involvement may therefore represent a high-risk group who require early aggressive therapy.
Our pattern recognition approach is well suited for analyzing joint involvements. Factors supported signals that were apparent based on overall co-involvements. Low-level factors grouped knees, subtalar joints, and midfoot joints on separate sides of the body into separate factors, demonstrating that our approach identifies both asymmetrical and symmetrical patterns of arthritis. As our approach identified patterns with little overlap outside the fingers, patient groups described clinically homogeneous entities. Extending NMF into a multilayer approach bears some similarity to other hierarchical modelling techniques such as deep autoencoders [24], in which each successive layer identifies increasingly broad representations of the data, or Gaussian process latent variable models [25]. However, multilayer NMF provides directly interpretable latent representations. This representation differs from PCA, which produces patterns that are orthogonal to each other, a feature with implications with respect to joint involvements. Furthermore, PCA patterns are less intuitive as joints contribute both positively and negatively to them. To reconstruct a patient’s joint involvements, we would have to add and subtract groups of joints, whereas with NMF, we would only add groups of joints.
Our study has a number of limitations. First, both discovery and validation cohorts were based on sample sizes of convenience. Because discovering joint patterns involved an unsupervised analysis, a priori power analyses were not done. However, the proximity of joints within patterns along the vertical axis and our ability to identify useful clinical measures demonstrated the potential of conducting such an analysis retrospectively. Secondly, the small size of the validation cohort limited our ability to test our findings in a multivariate model in a validation cohort, although we successfully validated our predictors in univariate tests. Furthermore, our validation cohort strongly supported these clinical measures, demonstrating that this concern is one of statistical power rather than approach. Lastly, we required patients to be treatment naïve except NSAIDs, which may have potentially skewed our patient cohorts towards individuals with milder forms of disease.
The identified joint patterns appear to have important prognostic implications. They are conceptually simple to apply at the bedside as they represent an easily computed weighted sum of active joints. Further classifying patients by the degree of localization may help clinicians further tailor treatment decisions as patients with less strongly defined phenotypes may require early aggressive therapy. Patterns of joint involvement may be among key components of a new disease classification for JIA in addition to other data domains, including antinuclear antibody (ANA) status [26], biological measures [13], and other musculoskeletal features such as enthesitis. Beyond JIA, our approach may be a transferrable template for application in rheumatoid arthritis (RA) and spondyloarthropathies (SpAs). Previous efforts in RA have attempted to define a representative pattern of “core joints” as indicators instead of using the full complement of joints [27,28]. This reductionist approach still counts joints. Alternatively, utilizing the rich data for pattern recognition may identify predictors of outcome.
Using multilayer NMF, we identified patterns of joint involvement predictive of disease trajectory in children with arthritis. Our results are consistent with previous observations pointing to key individual joints as predictors of poor outcome. Our hierarchical unsupervised learning approach allowed us to identify a new clinical variable, localization of joint involvement, as a key feature associated with poor outcomes in both our discovery and validation cohorts. Detailed bedside assessment of every joint is already part of every musculoskeletal exam for children with arthritis. Our study supports not only the continued collection of detailed information about joint involvement but also the inclusion of these patterns together with localization data (i.e., how closely affected children fit these patterns) to stratify patients and inform treatment decisions. Our findings will move the field of pediatric rheumatology out of infancy, from joint counts to realizing the potential of using data available from patterns of joints involvement.
|
10.1371/journal.pbio.1001831 | Targeting Of Somatic Hypermutation By immunoglobulin Enhancer And Enhancer-Like Sequences | Somatic hypermutation (SH) generates point mutations within rearranged immunoglobulin (Ig) genes of activated B cells, providing genetic diversity for the affinity maturation of antibodies. SH requires the activation-induced cytidine deaminase (AID) protein and transcription of the mutation target sequence, but how the Ig gene specificity of mutations is achieved has remained elusive. We show here using a sensitive and carefully controlled assay that the Ig enhancers strongly activate SH in neighboring genes even though their stimulation of transcription is negligible. Mutations in certain E-box, NFκB, MEF2, or Ets family binding sites—known to be important for the transcriptional role of Ig enhancers—impair or abolish the activity. Full activation of SH typically requires a combination of multiple Ig enhancer and enhancer-like elements. The mechanism is evolutionarily conserved, as mammalian Ig lambda and Ig heavy chain intron enhancers efficiently stimulate hypermutation in chicken cells. Our results demonstrate a novel regulatory function for Ig enhancers, indicating that they either recruit AID or alter the accessibility of the nearby transcription units.
| During the B cell immune response, immunoglobulin (Ig) genes are subject to a unique mutation process known as somatic hypermutation that allows the immune system to generate high-affinity antibodies. Somatic hypermutation preferentially affects Ig genes, relative to other genes, and this is important in preventing catastrophic levels of general genomic mutations that could lead to B cell cancers. We hypothesized that this preferential targeting of somatic hypermutation is assisted by specific DNA sequences in or near Ig genes that focus the action of the mutation machinery on those genes. In this study, we show that Ig genes across species—from human, mouse, and chicken—do indeed contain such mutation targeting sequences and that they coincide with transcriptional regulatory regions known as enhancers. We show that combinations of Ig enhancers cooperate to achieve strong mutation targeting and that this action depends on well-known transcription factor binding sites in these enhancer elements. Our findings establish an evolutionarily conserved function for enhancers in somatic hypermutation targeting, which operates by a mechanism distinct from the conventional enhancer function of increasing levels of transcription. We propose that combinations of Ig enhancers target somatic mutation to Ig genes by recruiting the mutation machinery and/or by making the Ig genes better substrates for mutation.
| The appearance of point mutations within the rearranged immunoglobulin (Ig) genes of B cells, which leads eventually to the selection and production of high-affinity antibodies, is called somatic hypermutation (SH) [1],[2]. SH requires transcription of the Ig genes [3] and expression of the activation-induced cytidine deaminase (AID) protein encoded by the AICDA gene [4],[5]. AID is believed to initiate all three types of B cell–specific Ig gene diversification—SH, Ig gene conversion (GCV), and Ig class switch recombination—by deaminating cytidines within the Ig loci [6]–[8].
While many non-Ig genes accrue mutations in AID-expressing B cells as a result of SH, Ig genes mutate at levels that are typically several orders of magnitude greater than those of non-Ig genes [9]–[12]. The question of how SH is preferentially targeted to Ig loci has been studied and debated for over 20 years. Pioneering experiments using chimeric gene constructs in transgenic mice indicated that sequences overlapping with the Ig light chain and Ig heavy chain enhancers distinguish the Ig genes as mutation targets [13]–[15]. Other early transgene studies indicated that Ig V region sequences themselves are not required for SH [16] and that active heterologous promoters can support SH [13],[17]. However, further insight into the nature of the putative cis-acting regulatory elements was hampered by the laborious transgene experimental system, the relatively low mutation rates of the chimeric genes, and the fluctuation of mutation rates among transgenic lines, perhaps due to integration site effects and copy number variations. A further problem arose from the fact that the putative hypermutation-stimulating sequences included the known enhancers, making it difficult to differentiate between the effects of these sequences on transgene hypermutation versus transgene transcription (reviewed in [18]).
The hypothesis that SH is targeted preferentially to Ig genes by the Ig enhancers was subsequently called into question when germline deletions of individual murine Ig enhancers—the same sequences previously implicated in the hypermutation of chimeric transgenes—did not abolish SH within the respective loci [19]–[21]. It also became apparent that expression of either AID or the related cytidine deaminases APOBEC-3A or APOBEC-3B increased mutation frequencies in the genomes of fibroblasts [22], Escherichia coli [23], yeast [24], and human breast cancer cells [25]. These findings and others (reviewed in [9],[18]) raised widespread doubts about the relevance of specific cis-acting SH targeting elements in Ig loci. In particular, Ig enhancers were no longer regarded as likely SH targeting elements, and it was increasingly felt that they increased SH solely by increasing Ig gene transcription. Attention has recently focused on RNA polymerase II (Pol II)–associated factors that interact with AID and play roles in transcriptional stalling [26] and RNA processing [27], processes that are likely to be critical for generating the single strand DNA substrate required by AID (reviewed in [9],[28]). However, these broadly acting factors do not provide a ready explanation for the strong preference that SH exhibits for Ig genes over non-Ig genes. Consequently, this has remained a central unresolved issue in the field.
The chicken B cell line DT40, whose genome is easily modified by targeted gene integration [29], is a powerful model to investigate AID-mediated gene diversification [30]. DT40 variegates its rearranged Ig light chain (cIgλ) gene primarily by GCV [31], but diversification occurs by SH if either upstream GCV donor sequences or uracil DNA glycosylase (UNG) are missing [7],[32]. Evidence for the stimulation of cIgλ GCV by cis-acting sequences in DT40 has been detected by the analysis of endogenous cIgλ gene diversification [33], transgene GCV [34], and transgene hypermutation [35]. Reminiscent of the early experiments in transgenic mice, SH of a green fluorescent protein (GFP) knock-in transgene in DT40 cells depended on the nearby presence of a 10-kb fragment of the cIgλ locus, which was named diversification activator (DIVAC) [35]. Deletion analysis of DIVAC led to the identification of two core regions downstream of the cIgλ C-region that cooperate with each other and with other parts of the 10-kb sequence to stimulate SH of the adjacent GFP transcription unit [36]. However, a clearer definition of the DIVAC code proved challenging using the original GFP assay because of functional redundancy within the 10-kb sequence and difficulty in measuring the DIVAC activity of elements shorter than 500 bp [35]–[37]. Furthermore, murine Ig lambda (Igλ) and Ig kappa (Igκ) enhancer sequences displayed disappointingly low DIVAC activity in DT40 cells [36],[38]. Hence, the identity of key SH targeting sequences and the extent to which these sequences have been conserved during vertebrate evolution have remained undetermined.
We have now developed a highly sensitive assay that allows analysis of the SH targeting activity of small DNA elements, largely overcoming the shortcomings of previous experimental strategies. Using this new assay, we demonstrate that chicken, mouse, and human Ig locus enhancers and enhancer-like elements are core DIVAC sequences that work together to target SH. Regardless of which species they derive from, these elements rely for function on a common set of well-characterized transcription factor binding motifs, highlighting the evolutionary conservation of the SH targeting mechanism. These findings are likely to have implications for the mistargeting of SH to non-Ig genes and the origins of B cell lymphoma.
We previously developed an assay for DIVAC function that made use of a reporter cassette, termed GFP2, consisting of a strong viral promoter driving expression of GFP and a drug resistance gene (Figure 1A) [35]. In this assay, GFP2, with or without a flanking test sequence, was inserted by homologous recombination into the DT40 genome, and GFP expression was monitored in subclones by flow cytometry. Loss of GFP expression was entirely dependent on AID, was due to point mutations in GFP, and could be stimulated more than 100-fold by the presence of a strong DIVAC element adjacent to the GFP2 cassette [35]. Importantly, three previous studies demonstrated that DIVAC-dependent stimulation of GFP mutation was not accompanied by substantial changes in GFP transcription as measured by several methods, demonstrating that DIVAC stimulates SH by a mechanism independent of an increase in transcription [35]–[37].
To increase the sensitivity of the DIVAC assay, we modified the GFP2 reporter by the insertion of a 5′ untranslated sequence upstream of the methionine start codon and a hypermutation target sequence between the start codon and the GFP open reading frame, yielding the new reporter GFP4 (Figure 1A). The 249-bp hypermutation target sequence consists of repetitions of TGG, CAA, and CAG codons frequently positioned in the context of SH hotspot motifs WRCY/RGYW (W = A or T; R = A or G; Y = C or T). Transition mutations at the second or third position of the TGG codons or at the first position of the CAA and CAG codons will introduce nonsense mutations, precluding the translation of the GFP open reading frame (Figure S1).
To further increase the frequency at which mutations and stop codons are generated, the GFP4 assay is performed in UNG-deficient cells, which accumulate exclusively C-to-T and G-to-A transition mutations and display a 7-fold increased rate of SH [32], most likely because AID-induced uracils cannot be excised and repaired before replication. To assay DIVAC-GFP4 combinations at a defined chromosomal position, we generated a recipient cell line, UNG−/−AIDR/puro, in which (i) both endogenous UNG genes were disrupted and the coding sequences of both endogenous AICDA genes were deleted, (ii) AID expression was reconstituted by inserting an AICDA cDNA expression cassette under the influence of the β-actin promoter into one AICDA locus, and (iii) the position of the second AICDA locus was marked by a puromycin resistance gene. When this cell line is transfected by AICDA locus–targeting constructs containing DIVAC-GFP4, targeted integrants into the marked AICDA locus are easily identified by the loss of puromycin resistance.
Alignment of the cIgλ locus with the corresponding sequence of turkey, zebra finch, and ground finch revealed seven evolutionarily conserved sequence contigs downstream of the C-region (Figures 1B and S2). Two of these corresponded closely to regions we had previously demonstrated to be important for DIVAC function in the context of larger DNA elements [36]: the cIgλ enhancer (cIgλE) [39] and the 3′Core. The conserved sequence regions were cloned into the upstream DIVAC insertion site of GFP4 (the default site used in all experiments except where indicated) and transfected into UNG−/−AIDR/puro cells. Primary transfectants with targeted integration of a construct were subcloned, and 24 subclones were analyzed for GFP loss by flow cytometry 12 d after subcloning (Figure 1C and 1D). Transfectants containing cIgλE or 3′Core, in either orientation (reverse orientation indicated by “R”), showed median GFP loss levels of 20%–30%, whereas levels of GFP loss in transfectants of the other conserved sequences (Con1–Con5) were close to the 1.7% median value observed in the no DIVAC control transfectant, UNG−/−AIDR. Interestingly, the Con2 sequence, which displayed activity close to background on its own, substantially increased GFP loss when combined with cIgλE in Con2+cIgλE cells (44.6%). The highest levels of GFP loss were seen when cIgλE and the 3′Core were combined (63.7%) or when they were tested together with their intervening sequence (cIgλE↔3′Core; 70.5%). Importantly, GFP loss in UNG−/−AID−/−cIgλE↔3′Core cells (lacking the AICDA expression cassette) was almost 3,000-fold lower than in cIgλE↔3′Core cells and about 60-fold lower than in UNG−/−AIDR cells.
These results illustrate several points. First, the DIVAC-GFP4 assay is capable of detecting robust stimulation of SH by short DNA fragments, which heretofore has not been possible. Second, these results directly confirm the role of cIgλE and 3′Core as core DIVAC elements [36]. Third, in the absence of DIVAC, GFP loss from GFP4 in UNG−/− cells is 15- to 20-fold higher than we detect with GFP2 in wild-type cells (see below, and [35],[36]), likely reflecting both the increased sensitivity of GFP4 and an increase of DIVAC-independent mutations in the UNG-deficient background. Finally, in the absence of AID, UNG deficiency does not lead to substantial GFP loss, even in the presence of a strong DIVAC element. Hence, despite the repair-deficient context, both DIVAC-dependent and DIVAC-independent GFP loss in the GFP4 assay require AID.
Sequencing of the hypermutation target region amplified from cIgλE↔3′Core cells 6 wk after subcloning revealed frequent transition mutations at G/C bases with a hotspot preference as expected for SH in UNG-deficient DT40 cells (Figure S1). Many of these mutations yielded stop codons, explaining the efficient GFP loss seen in cIgλE↔3′Core cells.
cIgλE includes an E-box as well as NFκB (nuclear factor kappa B), MEF2 (myocyte-specific enhancer factor 2), and PU.1-IRF4 (interferon regulatory family-4) binding motifs, all of which are remarkably conserved among avian species (Figure S2B). Deletions starting either from the 5′ or the 3′ end of cIgλE progressively decreased GFP loss in the DIVAC assay (Figure 2A and 2B). Once the 5′ deletions reached the NFκB motif (5′Δ37), GFP loss fell to background levels. Similarly, 3′ end deletions including the IRF4 motif in 3′Δ49 cells strongly reduced GFP loss.
The role of specific binding site motifs was further investigated by mutation of consensus residues in these sites (Figure 2A and 2C). Whereas mutations in the NFκB, MEF2, PU.1, or IRF4 motifs strongly decreased GFP loss, mutations in the E-box caused a more modest reduction, and a mutation in the spacer between the PU.1 and IRF4 motifs was well tolerated (Figure 2C). These results indicate that cIgλE requires the integrity of multiple transcription factor binding sites in its 5′ and 3′ halves for full activity.
Little was known about 3′Core, the second autonomous DIVAC sequence of the chicken Igλ locus. Deletion of the first 42 and the last 99 bp did not affect GFP loss (5′Δ42_3′Δ99), whereas many deletions in the central part of the fragment reduced GFP loss (Figure S3A and S3B). Search algorithms for transcription factor binding motifs predicted, among others, six evolutionarily conserved binding motifs in the parts of 3′Core where deletions compromised activity: three E-boxes and three other putative sites, referred to as pCBF (core binding factor), pC/EBP (CCAAT enhancer binding protein), and pPU.1 (Figure S2C) (where “p” designates a putative binding site for which experimental evidence linking it to the factor is lacking). Deletion or mutation of any one of these motifs, with the exception of pPU.1, reduced GFP loss substantially, with the strongest effects seen for E-box2, pCBF, and pC/EBP, which lie close together in the central part of the fragment (Figure S3C and S3D). Thus, evolutionarily conserved transcription factor binding motifs are also critical for the DIVAC function of 3′Core. We note that many more sites were predicted in silico than were tested, and the factors that might bind to these and the tested sites, particularly pCBF and pC/EBP, remain unknown.
Alignment of human, murine, and chicken Igλ enhancer sequences revealed striking conservation of the E-box and NFκB, MEF2, PU.1, and IRF4 binding motifs [40],[41], while the mammalian sequences possess an additional E-box about 50 bp downstream of the PU.1 site (Figure S4A). Since the conserved transcription factor binding motifs were important for the DIVAC function of cIgλE, we reasoned that the mammalian enhancers might also be active DIVAC elements despite low sequence conservation of the intervening sequences.
We began by testing the human Igλ enhancer (hIgλE) in either the upstream or downstream insertion site of GFP4 (Figure 1A), which yielded a remarkable 46% GFP loss (Figure 3B), almost twice the activity of cIgλE (27.2%). Removal of the upstream E-box in 5′Δ56 did not decrease DIVAC activity, whereas larger 5′ deletions reduced activity (Figure 3A and 3B). However, even after removal of the upstream E-box, NFκB, and MEF2 sites, the 5′Δ84 fragment was still capable of supporting 23.6% GFP loss, almost as high as the activity of full-length cIgλE and much higher than the activity of the comparable deletion fragment (5′Δ59) of cIgλE (Figure 2B). These results suggest that the 3′ portion of hIgλE contains important elements and that the downstream E-box might compensate for loss of the upstream E-box-NFκB-MEF2 sites. Consistent with this, a 3′ deletion including the downstream E-box (3′Δ46) reduced GFP loss to 20%—roughly the activity of full-length cIgλE—and a larger 3′ deletion removing the composite PU.1-IRF4 site (3′Δ108) strongly reduced GFP loss to 6% (Figure 3B), similar to the low activity of the comparable cIgλE 3′Δ68 fragment (Figure 2B). In the strongly active hIgλE, point mutations in individual motifs reduced activity, although typically less than 2-fold, and only mutation of both components of the composite PU.1-IRF4 site had a strong effect on activity (Figure 3A and 3B). Therefore, hIgλE is both more active and apparently more robust than cIgλE, being less sensitive to mutation of individual motifs. The major difference between the human and chicken enhancers appears to lie in sequences in their 3′ portions.
These results demonstrate, to our knowledge for the first time, a substantial conservation of DIVAC function from human to chicken sequences. They also reveal parallels between the enhancement of SH and the enhancement of transcription by the Igλ enhancer because the transcription factor binding sites long known to be important for the regulation of Igλ transcription [41]–[43] are also critical for DIVAC function.
Sequence homologues of mammalian Ig heavy chain intron enhancers (IgHEi) could not be identified in birds, and an enhancer in the intron between the duck Jμ and Cμ segments showed no obvious conservation with mammalian counterparts apart from the presence of multiple E-boxes [44]. Human (hIgHEi) and murine (mIgHEi) enhancer fragments contain conserved YY1 (yin yang 1) (μE1), E-box (μE2 and μE4), Ets1 (μA), PU.1 (μB), IRF, and Octamer transcription factor binding sites, and less well conserved regions μE5 and μE3 [45],[46] (Figure 4A and S4B). Since these sites overlap substantially with those important for DIVAC function in cIgλE, 3′Core, and hIgλE, we reasoned that the mammalian IgHEi elements might also have SH targeting activity. Strikingly, hIgHEi and mIgHEi yielded high levels of GFP loss (62.1% and 47.3%, respectively; Figure 4B), well above that of cIgλE and 3′Core, and similar to that observed with hIgλE.
To investigate the role of the well-known binding sites, hIgHEi was subject to deletion and mutation analysis. Whereas a 5′ deletion of hIgHEi including the μE1, μE2, μA, and μB sites only moderately decreased GFP loss in 5′Δ109 and 5′Δ136 cells, 3′ deletions including the Octamer, μE4, and IRF sites strongly decreased GFP loss in 3′Δ67 and 3′Δ136 cells. Consistent with the importance of the 3′ part of hIgHEi, mutations of either the μE4 or IRF site strongly decreased GFP loss, whereas an Octamer site mutation had little effect. Thus, the binding sites in the 5′ portion, although able to boost activity of the 3′ portion, are unable to compensate for loss of the IRF or μE4 sites in the 3′ portion. We conclude that mammalian IgHEi sequences are potent DIVAC elements in chicken cells.
Homologues of the mammalian Ig kappa chain (Igκ) enhancers are also not present in avian species, which contain only a single Igλ light chain locus. The three Igκ enhancers, intron (IgκEi), 3′ (IgκE3′), and Ed (IgκEd) [47], of mice and humans (Figures 5A and S5) induced low or modest levels of GFP loss when assayed on their own (Figure 5B), consistent with previous analyses [36],[38]. However, when two Igκ enhancers were combined (IgκEi+IgκE3′ or IgκE3′+IgκEd), GFP loss markedly increased, and when the three human Igκ enhancers were combined, GFP loss reached 50.9% (Figure 5B). This shows that the known synergy of the Igκ enhancers with respect to the activation of Igκ transcription ([47] and references therein) also holds true for their DIVAC function, even in an avian B cell line lacking an endogenous Igκ locus.
To confirm our results in a repair-proficient cellular context (UNG-proficient DT40 cells) and in a different genomic integration site (the deleted rearranged Igλ locus), we tested various cIgλ DIVAC elements using the GFP2 assay. The full cIgλ DIVAC region (the 9.8-kb W fragment that includes the rearranged VJλ region and all downstream cIgλ sequences) yielded about 10% GFP loss using GFP2 (Figure S6B and S6C), consistent with our previous study [35]. In general, the rank order of activities of DIVAC elements was similar between the GFP2 and GFP4 assays (compare Figures S6C and 1D). Comparison of median GFP loss levels indicated that the GFP2 assay is approximately 20–50 times less sensitive than the GFP4 assay (e.g., for cIgλE and 3′Core, respectively: 27.2% and 33.2% median GFP loss with GFP4, and 0.54% and 0.75% median GFP loss with GFP2). However, with the cIgλE↔3′Core fragment, GFP loss in the GFP2 assay (6.7%) was only about 10-fold lower than in the GFP4 assay (70.5%), probably because of saturation of the GFP4 assay in the presence of this highly active DIVAC element (see Protocol S1). We also used the GFP2 assay to confirm that Con2 (which lacks activity on its own) was able to substantially boost the activity of cIgλE (Figure S6C). A limited deletion and mutation analysis of Con2 (Figure S6A) using the GFP2 assay (Figure S6C) and the GFP4 assay (Figure S6D) demonstrated that functional cooperation between Con2 and cIgλE required only the 3′ portion of Con2 and was dependent on one of the two putative IRF binding motifs (pIRF-down) in this region. We conclude that there is good congruence between the results of the GFP4 and GFP2 assays and that the less sensitive GFP2 assay is preferable for analysis of highly active DIVAC elements.
The murine Igλ locus contains two enhancers, mIgλE3-1 and mIgλE2-4, due to a duplication of a pair of J-C regions and their downstream enhancer (Figure 6A) [48]. These enhancers are relatively weak DIVAC elements on their own (0.4%–0.5% GFP loss in the GFP2 assay; Figure 6B), consistent with our previous analysis [36]. This suggested the need for other sequences in the locus to cooperate with mIgλE3-1 and mIgλE2-4 to support efficient SH of murine Igλ (note that cooperation between mIgλE3-1 and mIgλE2-4 is not possible in some rearranged Igλ loci because rearrangement of upstream V2 or V3 gene segments to the JC3 or JC1 clusters deletes mIgλE2-4). However, the identity of such putative cooperating elements was unclear because other murine Igλ enhancers were not known.
Intriguingly, BLAST searches revealed the presence of IgλE homologues 20–25 kb downstream of mIgλE3-1 and mIgλE2-4 (Figure 6A), which we refer to as mIgλE3-1s and mIgλE2-4s because of their resemblance to shadow enhancers [49]. The newly identified elements are 95% identical to one another and about 70% identical to the canonical enhancers, with the conservation including many of the transcription factor binding motifs shown to be important for DIVAC function of the chicken and human Igλ enhancers (Figure S4A). When tested for DIVAC function, mIgλE3-1s and mIgλE2-4s were substantially more active than the canonical enhancers in both the GFP2 (Figure 6B) and GFP4 assays (data not shown). Strikingly, the combination of a shadow enhancer with its neighboring canonical enhancer induced GFP loss strongly and synergistically (Figure 6B), in the case of mIgλE2-4 plus mIgλE2-4s to levels almost as high as that seen for the entire cIgλ W fragment. These results reveal that strong SH targeting elements can be constructed from combinations of enhancers and enhancer-like elements in the murine Igλ locus, as is true also for chicken Igλ. Furthermore, they demonstrate our ability to identify strong DIVAC elements in the murine Igλ locus on the assumption that Igλ enhancer-like sequences activate SH.
We extended this by investigating the activity of other combinations of elements, continuing to use the GFP2 assay. Consistent with the GFP4 data, hIgλE, hIgHEi, and the combined murine Igκ enhancers supported levels of GFP loss that were more than 20-fold above the background of AIDR cells (0.1%), whereas the 5′Δ84 deletion mutant of hIgλE was less active (Figure 6C). Duplication of the truncated 5′Δ84 or the full-length hIgλE increased levels of GFP loss from about 0.6% and 2.4% to about 2.0% and 6%, respectively, showing that even the interaction between identical sequences can lead to a synergistic increase of DIVAC function, similar to the well-known effects of multimerization of enhancer sequences on transcriptional activity [50].
Consistent with previous studies of the GFP2 reporter [35],[36] or modifications thereof [37], mRNA levels from GFP4 were either not significantly or only marginally (up to 2-fold) increased by the presence of chicken or mammalian DIVAC fragments compared to the no DIVAC control (Figure 7A–7C). Therefore, as with the GFP2 assay, DIVAC elements stimulate mutation in the GFP4 assay by a mechanism that is independent of an increase in GFP transcription.
Given the relatively strong DIVAC function associated with the mIgλ shadow enhancers, we wondered whether they also possessed transcriptional enhancer activity. To test this, sequences were cloned downstream of a minimal promoter–luciferase reporter and transfected into the UNG−/−AIDR recipient cell line used for the GFP4 studies. Both mIgλE3-1s and mIgλE2-4s were able to stimulate luciferase expression above that of the empty vector (no DIVAC) control, but both exhibited significantly less enhancer activity than their canonical mIgλ enhancer counterparts (Figure 7D), despite being stronger DIVAC elements. This discordance between transcriptional enhancer activity and DIVAC function further supports the conclusion that DIVAC operates by a mechanism distinct from that of stimulating transcription. A very recent study, published while our manuscript was under revision, identified the two mIgλ shadow enhancers based on epigenetic criteria and demonstrated that they possess B lineage–specific enhancer activity [51].
Using a highly sensitive, well-controlled assay we provide conclusive evidence that SH is targeted by Ig enhancer and Ig enhancer-like sequences. The phenomenon is strikingly conserved during vertebrate evolution, as even short mammalian Igλ and IgH enhancer fragments raised mutation rates more than 20-fold in chicken cells. SH activating sequences, or DIVAC, not only physically overlap the Ig enhancers but also closely resemble transcriptional enhancers in their mode of action by (i) requiring multiple transcription factor binding sites, (ii) functioning independent of orientation and when positioned either upstream or downstream of the transcription unit, and (iii) increasing activity through the collaboration of multiple enhancer-like regions, each of which depends on transcription factor binding motifs.
The recognition of Ig enhancers as SH targeting sequences yields a conceptual framework within which to reevaluate earlier studies. Most notably, the new results vindicate the early transgenic experiments that showed overlap of SH stimulating sequences with the Igλ, IgH intron, and Igκ enhancers [13],[14] and synergistic effects between the Igκ intron and Igκ 3′ enhancer sequences [13],[15]. The failure of either Igκ intron or 3′ enhancer knockouts in mice to abrogate hypermutation [19],[20] is consistent with the contributions of multiple, partially redundant Igκ enhancers to DIVAC function. Similarly, the failure of a previous study to identify SH targeting function associated with the Igκ intron and 3′ enhancers in DT40 cells [38] was likely due to use of a less sensitive assay and the absence of the Igκ distal enhancer. In addition, evidence that E-box [37],[52],[53], NFκB [34], MEF2 [34], and PU.1-IRF4 [54],[55] binding sites play a role in the targeting of SH or GCV can be explained by the importance of these sites within the context of Ig enhancers and enhancer-like sequences.
The results presented here provide the foundation for models of the cis-acting regulatory regions that target SH to a variety of Ig loci. The chicken Igλ locus is best understood and offers several lessons that might be generally applicable. In cIgλ, the enhancer cooperates with an evolutionarily conserved downstream element (3′Core) that itself possesses low levels of transcriptional enhancer activity (Figure 7D) but contains functionally important transcription factor binding motifs well known from Ig enhancers (Figures S2 and S3). However, it is clear that these two elements depend on additional sequences (e.g., Con2 and the region between cIgλE and 3′Core) for full DIVAC function (Figures 1 and S6) [35],[36]. The mouse Igλ and human and mouse Igκ loci offer parallels, with DIVAC function involving the combined action of two or more well-separated enhancer or enhancer-like elements. By analogy with cIgλ, it is tempting to think that other surrounding sequences further contribute to the full SH targeting activity of mammalian Ig loci. The human Igλ enhancer, the human and mouse IgH intron enhancers, and a combination of the known Igκ enhancers increase SH 20- to 30-fold in our assays, well below the 100-fold stimulation achieved by the full cIgλ DIVAC (Figure S6C). Indeed, previous analyses showing that deletion of mIgHEi or hIgHEi from the endogenous loci did not abolish SH [21],[56] are consistent with the existence of other compensatory targeting elements, a strong candidate for which is the large 3′ regulatory region more than 200 kb downstream of IgHEi [57],[58].
The identities of the trans-acting factors that bind Ig enhancers to stimulate SH are not known, although some candidates have been identified in previous studies and others can be inferred from the binding motifs whose integrity we show is important for DIVAC function. Substantial data support a role for E-box binding factors, including the E2a-encoded proteins E12 and E47 [53]. Disruption of E2a in DT40 cells reduced the frequency of SH/GCV [59],[60] as did overexpression of the E protein inhibitors Id1 and Id3 [61]. E12 and E47 prefer to bind the CASSTG (S = C or G) subtype of E-box [62], and while mutation of this subtype reduces DIVAC function, mutation of E-boxes predicted to be bound poorly by E12/E47 does also [37]. Existing data leave unresolved the identity of the E-box binding factor(s) that contribute to DIVAC function. Studies in DT40 have also implicated NFκB, PU.1, and IRF4 as trans factors relevant for the targeting of SH/GCV [34],[55]. Despite the fact that transcription and hypermutation enhancers make use of overlapping binding motifs and likely an overlapping set of trans factors, our data provide a compelling argument that the two processes operate by distinct mechanisms and, in particular, that DIVAC does not operate by increasing transcription.
It may not be a coincidence that enhancers, able to exquisitely regulate cell type– and gene-specific expression, have assumed the vital role of targeting SH to the Ig loci. The complex structure of DIVACs—distinct configurations of a common set of transcription factor binding motifs, with robust activity relying on multiple, and to some extent redundant, sequences—may reflect the formidable task of fine tuning and restricting SH. It might also reflect piecemeal evolution of DIVAC, with each Ig locus cobbling together an idiosyncratic collection of SH targeting elements. Chromosomal translocations near DIVACs likely increase the mutation rate in the neighborhood of the translocation breakpoint, as confirmed for the case of IgH to c-Myc locus translocations [63]. It is also possible that non-Ig genes like BCL6 that mutate at substantial rates in AID-expressing B cells [10]–[12] do so because of DIVAC-like sequences in their neighborhoods. In support of this, a recent computational analysis found that promoter-proximal E-box, C/EBPβ, and YY1 binding motifs (all of which are found in some of the DIVAC elements identified here) were predictive of off-target SH of non-Ig genes [64].
Little is known about how gene-specific enhancers and particularly Ig enhancers distinguish themselves from other enhancers that may contain the same or similar transcription factor binding sites. Despite this limitation in our understanding of enhancer function, plausible models for how SH is targeted to Ig genes can be formulated based on what is known about the interaction of enhancers with the transcription initiation complex (Figure S7). One possibility is that a DIVAC-bound factor or a combination of factors actively recruit AID. A not mutually exclusive alternative is that DIVACs induce changes in the Pol II transcription initiation or elongation complex, making the transcribed DNA more accessible to AID. This hypothesis might explain why the accumulation of SH events rises rapidly downstream of the transcription start site and then falls off exponentially [3],[65], and might establish a connection between DIVACs and stalled transcription [9],[36],[66],[67] or RNA exosome complexes [27]. Interestingly, members of the APOBEC family can induce showers of clustered mutations in breast cancer and yeast cells that are believed to be related to single stranded DNA in the neighborhood of DNA double strand breaks [24],[25], setting a precedent for how a change in DNA conformation can target deaminases to particular regions of the genome.
The GFP4 cassette (Figure 1A)—which resembles GFP2 [35] but contains a 5′ untranslated sequence, the hypermutation target sequence (Figure S1), and, for increased GFP brightness, the GFPnovo2 open reading frame [68]—was custom synthesized (Blue Heron Biotechnology) and cloned into the BamHI site of an AICDA locus–targeting construct [69], yielding the GFP4-containing, AICDA locus–targeting construct pAICDA_GFP4. A variant of pAICDA_GFP4, named pAICDA_GFP4D, was made in which the SpeI/NheI sites upstream of GFP4 were deleted and a unique NheI site was introduced downstream of GFP4. The cloning of DIVAC sequences into the GFP4 or GFP2 targeting vectors is described in Protocol S1.
An UNG-deficient DT40 clone with both endogenous AICDA alleles deleted [32] was reconstituted with AID by the targeted integration of a bicistronic AICDA/gpt expression cassette into one of the AICDA loci [35]. The second AICDA locus was subsequently marked by the targeted integration of a puromycin resistance gene driven by the chicken β-actin promoter, yielding the recipient UNG−/−AIDR/puro cell clone for transfections of GFP4 targeting constructs. The ΨV−IgL− clone in which the rearranged Igλ locus was replaced by a puromycin resistance cassette [35] was used for transfections of GFP2 targeting constructs.
DT40 cell culture, transfection, drug selection, and the identification of transfectants with targeted integration of GFP2 constructs were performed as described previously [35]. Transfectants with targeted integration of GFP4 constructs were also detected by the appearance of puromycin sensitivity. The AID-negative UNG−/−AID−/−cIgλE↔3′Core clone was derived from the cIgλE↔3′Core transfectant by cre recombinase–mediated removal of the LoxP-flanked AICDA/gpt expression cassette [70].
GFP expression from GFP2 transfectants was assessed by flow cytometry at day 14 after subcloning, as described previously [35],[36], whereas GFP4 transfectants were assessed at day 12 after subcloning. Details of the flow cytometry analysis are provided in Protocol S1.
Genomic DNA was isolated from a subclone of cIgλE↔3′Core after 6 wk of culture and used for the amplification of GFP4 sequences by PCR using Phusion polymerase (New England Biolabs). The PCR fragments were cloned using the In-Fusion Cloning Kit (Clontech) into the linearized pUC19 provided with the kit and sequenced. Thirty-four sequences covering the first 500 transcribed bases of GFP4 were aligned to the GFP4 sequence to detect sequence variation (Figure S1).
Orthologues of the Igλ locus were identified in the turkey, zebra finch, and ground finch genomes using the W fragment of cIgλ in low stringency blastn BLAST of the reference genome database and Blat genome searches of the respective genome sequences. BLAST and Blat searches were also used to identify the murine Igλ shadow enhancers and map them within the murine Igλ locus. The bird Igλ orthologues were aligned using the ClustalW2 web interface (http://www.ebi.ac.uk/Tools/msa/clustalw2/) to detect sequence contigs conserved during avian evolution. ClustalW2 was also used to create the other sequence alignments shown in Figures S2, S4, and S5. Searches for conserved transcription factor binding sites were performed using the TESS (Transcription Element Search Software) program [71].
Two-tailed unpaired t-tests were used to compare relative GFP transcript and luciferase levels in Figures S2D and S2E.
Reverse transcription quantitative PCR analysis was carried out on transfectants containing various DIVAC-GFP4 constructs after the cells were treated with 4-OH tamoxifen and subcloned to delete the AID expression cassette. This avoided potential effects on transcript levels due to nonsense-mediated mRNA decay. The resulting AID-negative cells used for analysis were stably GFP-positive (data not shown). RNA was extracted from 5×106 cells using the RNeasy Mini kit (Qiagen), and the cDNA was prepared from 1 µg of RNA using the iScript cDNA synthesis kit (Bio-Rad). Quantitative PCR was performed using the DyNAmo HS SYBR Green qPCR kit (Thermo Scientific). GFP transcript levels were normalized to 18S rRNA levels. Samples were denatured for 15 min at 95°C, followed by 40 cycles of 30 s at 94°C, 30 s at 60°C, and 30 s at 72°C. The primers used were as follows: GFPup-F 5′-ggaatatactttgccaagaagcgtt-3′, GFP5up-R 5′-accatcgttgccagaaccatt-3′, GFPcds-F 5′-gagcaaagaccccaacgaga-3′, GFPcds-R 5′-gtccatgccgagagtgatcc-3′, 18S-F 5′-taaaggaattgacggaaggg-3′, and 18S-R 5′-tgtcaatcctgtccgtgtc-3′.
RNA for Northern blot analysis was prepared from GFP2 or GFP4 cell lines with the RNeasy kit (Qiagen) or TRIzol reagent (Invitogen). 10 µg of total RNA was run on a gel, transferred to a membrane, and hybridized with a GFP probe. The blot was then stripped and reprobed with a GAPDH probe as a loading control. Using Image Lab software (Bio-Rad), bands were quantitated and normalized to the corresponding GAPDH signal, and values were presented relative to the GFP4 no DIVAC control. The probes were PCR-amplified DNA products made with the corresponding primers: GFPp-F 5′-accatggtgagcaagggcga-3′, GFPp-R 5′-ctaggacttgtacagctcgtccatgc-3′; GAPDHp-F 5′-accagggctgccgtcctctc-3′, GAPDHp-R 5′-ttctccatggtggtgaagac-3′.
Test sequences were cloned between SalI and BamHI sites downstream of the firefly Luc2 gene of the minimal promoter containing pGL4.23 vector (Promega). 20 µg of the plasmid was co-transfected into UNG−/−AIDR/puro cells with 2.5–5.0 µg of pGL4.75 Renilla luciferase control vector (Promega) using the Amaxa Nucleofector kit V (Nucleofector program B-023) (Lonza). The relative activity of firefly luciferase to Renilla luciferase was determined using the Dual-Glo Luciferase Assay System (Promega) according to the manufacturer's protocol.
|
10.1371/journal.pntd.0003917 | Sequence- and Structure-Based Immunoreactive Epitope Discovery for Burkholderia pseudomallei Flagellin | Burkholderia pseudomallei is a Gram-negative bacterium responsible for melioidosis, a serious and often fatal infectious disease that is poorly controlled by existing treatments. Due to its inherent resistance to the major antibiotic classes and its facultative intracellular pathogenicity, an effective vaccine would be extremely desirable, along with appropriate prevention and therapeutic management. One of the main subunit vaccine candidates is flagellin of Burkholderia pseudomallei (FliCBp). Here, we present the high resolution crystal structure of FliCBp and report the synthesis and characterization of three peptides predicted to be both B and T cell FliCBp epitopes, by both structure-based in silico methods, and sequence-based epitope prediction tools. All three epitopes were shown to be immunoreactive against human IgG antibodies and to elicit cytokine production from human peripheral blood mononuclear cells. Furthermore, two of the peptides (F51-69 and F270-288) were found to be dominant immunoreactive epitopes, and their antibodies enhanced the bactericidal activities of purified human neutrophils. The epitopes derived from this study may represent potential melioidosis vaccine components.
| Melioidosis is an infectious disease caused by Burkolderia pseudomallei that poses a major public health problem in Southeast Asia and northern Australia. This bacterium is difficult to treat due to its intrinsic resistance to antibiotics, poor diagnosis, and the lack of a licensed vaccine. Vaccine safety is a prime concern, therefore recombinant protein subunit and/or peptide vaccine components, may represent safer alternatives. In this context, we targeted one of the main subunit vaccine candidates tested to date, flagellin from B. pseudomallei (FliCBp) that comprises the flagellar filament that mediates bacterial motility. Based on the knowledge that activation of both cell-mediated and antibody-mediated responses must be addressed in a melioidosis vaccine, we identified B and T cell immunoreactive peptides from FliCBp, using both sequence-based and structure-based computational prediction programs, for further in vitro immunological testing. Our data confirm the accuracy of sequence-based epitope prediction tools, and two structure-based methods applied to the FliCBp crystal structure (here-described), in predicting both T- and B-cell epitopes. Moreover, we identified two epitope peptides with significant joint T-cell and B-cell activities for further development as melioidosis vaccine components.
| Burkholderia pseudomallei, a pathogenic Gram-negative bacterium present in soil and water, is responsible for melioidosis, an often fatal infectious disease that is most frequently reported in tropical regions of the world, especially in Thailand and northern Australia, where the disease is endemic [1]. Diagnosis and treatment of melioidosis are far from adequate, as symptoms lack a specific signature necessary for rapid diagnosis, and the bacterium is inherently resistant to many commercially available classes of antibiotics. In addition, due to the intrinsically polymorphic nature of the pathogen, infections can occur in acute and chronic forms, with a plethora of different clinical manifestations [2]. Over the last few years, the study of melioidosis has become increasingly relevant, not only as a public health concern, but also due to bioterrorism implications, since B. pseudomallei is classified as a category B infective agent.
To date, no melioidosis vaccine is available, and no vaccine candidates are close to licensing [3,4]. Furthermore, the use of attenuated forms of B. pseudomallei presents significant safety issues due to its capacity to remain dormant and cause infections years later, therefore it is clear that alternative approaches must be sought.
A subunit vaccine approach, whereby only the microbial components that produce an appropriate immune response are administered, offers a means to significantly improve vaccine safety [5]. One safer alternative is a peptide-based vaccine approach, since peptides are easily and inexpensively produced [6]. Although subunit vaccines present many desirable qualities, their ability to stimulate a potent immune responses is much weaker than traditional whole-cell preparations. This may explain why no effective vaccine candidates against B. pseudomallei infection have emerged, despite several attempts using different approaches [7].
According to previous studies, protection against B. pseudomallei infection requires optimal responses from both cellular and humoral immune systems [7,8]. IFN-γ secreted by NK cells and T cells plays an important role in the control of infection in mice [9]. To date, one of the main candidate antigens is bacterial flagellin (FliC), which assembles to form the flagellar filament that supports bacterial motility. FliC induces IFN-γ responses from human T cells, and is recognized by antibodies from seropositive individuals living in endemic areas [10,11]. In a mouse model of infection, CpG-modified plasmid DNA encoding flagellin, induced responses from Th1 cells and B cells, and was shown to protect against bacterial re-infection and to reduce mortality and morbidity rate [11]. Furthermore, in passive immunization trials, antibodies raised against FliC from B. pseudomallei strain 319a were shown to protect diabetic rats challenged with a heterologous B. pseudomallei strain [12,13].
Interestingly, FliC is recognized as a pathogen associated molecular pattern (PAMP) by Toll-like receptor 5 (TLR5) and by nucleotide-binding oligomerization domain (NOD)-like receptor C4 (NLRC4), activating both innate and adaptive immunity [14,15]. Indeed, both are pattern recognition receptors (PRR), and play a key role in innate immunity, offering a first line of defence against invading pathogens. PAMPs are prevalent in bacterial but not in vertebrate genomic DNA; the immune system appears to exploit these molecules as signaling beacons to reveal the presence of infection and activate appropriate defense pathways [16]. It has been proven that TLR5 interacts with the conserved D1 domain of FliC, responsible for flagellin assembly [17]. Therefore, in addition to the direct stimulation of a protective response, FliC could function as a molecular adjuvant in combination with other specific antigen(s) [18].
With regards to flagellin from B. pseudomallei (FliCBp), epitope mapping, performed in our laboratory at Khon Kaen University, Thailand by synthesis of 38 multiple overlapping peptides extending along the whole protein sequence was carried out and identified several peptides that bound to HLA class II alleles found in Thailand [19]. This approach, however, is costly and time consuming. Sequence-based and 3D structure-based computational predictions for consensus dominant epitopes represent alternative, inexpensive and rapid approaches for epitope discovery and vaccine development [20].
In this report, we focused on FliCBp as a target antigen for sequence and structure-based epitope discovery. On this basis, we designed and synthesized the predicted epitopes as free peptides for consideration as potential vaccine components. To this aim, we used three B-cell epitope prediction servers [21,22] and one T-cell epitope prediction server [23], to identify putative ab initio immunoreactive linear epitopes. Analysis of the sequence outputs from all servers resulted in the identification of three linear epitopes with predicted B-cell and T-cell activities. The three selected sequences were synthesized as conjugated peptides, and tested for their effective T-cell/B-cell activities in immunological studies.
In parallel, we solved the high-resolution crystal structure of FliCBp as a starting point for future structure-based antigen and epitope design and optimization that permitted the identification of additional potentially antigenic regions. To this aim, we used a combination of 3D structure-based epitope predictions [24,25] and an experimental epitope mapping technique to broaden our consensus toward the identification of novel epitopes. Upon comparison of sequence-based and structural T-cell predictions, we identified and synthesized an additional T-cell epitope that was also tested in immunological studies.
FliCBp constitutes a major target for vaccine discovery initiatives to elicit protective immunity against B. pseudomallei infections. The predicted epitope sequences here-identified, coupled to the structural information about their native conformations, hold great potential for a process of rational design toward the optimization of their antigenic properties, and display the characteristics required for presentation as vaccine components.
Ethical permission was obtained from the KKU Ethics Committee for Human Research no. HE470506 and HE561234. All subjects were adults and had received written information before signing the consent form. Heparinized whole blood samples were collected from healthy donors, seropositive (IHA titer > 40) and seronegative (IHA titer ≤ 40) individuals [26,27], at the Blood Bank Center, Khon Kaen University, and from recovered melioidosis patients at Srinakarind Hospital, Khon Kaen University, Khon Kaen, Thailand.
The full-length FliCBp protein sequence was submitted to B cell linear epitope predictors, BepiPred (http://www.cbs.dtu.dk/services/BepiPred/) [21]; threshold = 0.35, BCPred (http://ailab.ist.psu.edu/bcpred/predict.html) [22]; classifier specificity = 75%, and AAP (http://ailab.ist.psu.edu/bcpred/predict.html) [22]; classifier specificity = 75%). Any sequences that were positively selected by more than two prediction methods and that contained more than 15 amino acids were selected as linear B cell epitopes. Predicted linear B cell epitopes were then re-submitted for T cell epitope predictions using the MHC-II Binding Predictor (http://tools.immuneepitope.org/mhcii/) in the Immune Epitope Database (IEDB) [23], with HLA DRB1 alleles common to Thailand; HLA DRB1*0301, 0405, 07, 09, 1202, 1501, 1502 and 1602 [28]. Positive T cell epitopes were identified by a HLA binding prediction score of less than 30.
The BPSL3319; fliC gene encoding for protein residues 25–378 (devoid of the N-terminal signal peptide (1–24) and C-terminal residues 379–388), was amplified from B. pseudomallei strain K96423 (Kindly provided by Prof. R. Titball, University of Exeter, UK), cloned into pGEX4T1 (Life Technologies) and expressed as a GST-fusion protein in BL21 Star (DE3) Escherichia coli cells (Life Technologies) and purified by affinity chromatography on a 5mL GSTrap FF column (GE Healthcare) pre-equilibrated in 1X PBS. On-column cleavage of the GST tag was carried out upon addition of 100 U thrombin (Sigma-Aldrich) in 1X PBS, incubation overnight at RT. Cleaved FliCBp was eluted with 1X PBS and thrombin was removed using a 1ml HiTrap Benzamidine FF column (GE Healthcare), according to the manufacturer’s instructions. FliCBp was exchanged into 10 mM Tris-HCl pH 8.0 and concentrated to 10.5 mg/ml for crystallization trials.
FliCBp crystals were grown by sitting drop at 20°C, in a 300 nl drop containing 50% protein solution (10.5 mg/ml) and reservoir solution (0.1 M HEPES pH 7.5, 25% PEG 6000 and 0.1 M Lithium Chloride). Crystals were cryo-protected in reservoir solution containing 30% glycerol. One crystal was used to collect X-ray diffraction data at the ID23-2 beamline at the European Synchrotron Radiation Facility (Grenoble, France). Data were processed using IMOSFLM and scaled using POINTLESS and SCALA using the CCP4 suite [29,30]. The structure was solved via molecular replacement using Phaser [31] and the structure of domain D1 (residues 62–168 and 281–326) of Sphingomonas sp. A1 flagellin (PDB entry 2ZBI) as a search model (the structure of FliCPa was not available at that time). Residues of the search model not conserved in FliCBp were substituted with alanines. The experimental phases were improved by applying Arp/Warp [32], and the amino acid sequence of the model was then modified to match the correct FliCBp sequence. The FliCBp structure was completed by manually fitting in the electron density residues of the D2 domain, which inserts between residues 167 and 287 of domain D1. Several rounds of manual model building with COOT [33], and refinement with the program REFMAC5 [34], were carried out until refinement reached convergence and the quality of the model was checked with PROCHECK [35]. The final refinement statistics and geometry quality parameters are shown in S1 Table. Atomic coordinates and structure factors were deposited in the Protein Data Bank with PDB code 4CFI [36].
The crystal structure of FliCBp was used as a starting point for three replicas of all-atom Molecular Dynamics (MD) simulations in explicit water at 300K. Each replica was 50 ns long and was carried out in NPT conditions. The simulations and the analysis of the trajectories were performed using the GROMACS 4.54 software package [37],[38], the SPC water model [39] and the GROMOS53A6 force field [40]. The procedure employed is fully described in Gourlay et al., [41].
Epitope predictions were carried out using MLCE on the representative structure of the most populated structural cluster of each MD trajectory [24]. The clustering procedure was performed using the method developed by Daura et al [40]. The MLCE method is based on the calculation of the matrix of inter-residue, non-bonded interaction energies using a MM-GBSA (Molecular Mechanics Generalized Born Surface Area) implicit solvent approximation. The principal components of the interaction matrix are selected by eigenvalue decomposition [42,43,44,45,46] and filtered by the protein contact map to select the substructures presenting low energy couplings with the rest of the protein 3D fold. These substructures are characterized by dynamic properties allowing them to visit multiple conformations, a subset of which can be recognized by the antibody. MLCE is available as the free web tool BEPPE (http://bioinf.uab.es/BEPPE).
The EDP method calculates the free energy penalty for desolvation placing a neutral probe at various protein surface positions. Surface regions with a small free energy penalty for water removal may correspond to preferred interaction sites. Evidence suggests that it is easier for an antibody to bind to an epitope when properties required for high affinity binding like low desolvation penalty are met [25].
Peptide mixtures were obtained by trypsin digestion of recombinant FliCBp in 50 mM ammonium bicarbonate buffer (pH 7.8) at a ratio of 10:1, at 37°C for 3 h. To capture the epitope-containing peptide, a 25 μl suspension of Dynabeads Pan Mouse IgG (uniform, super-paramagnetic polystyrene beads of 4.5 μm diameter coated with monoclonal human anti-mouse IgG antibodies) was used. The beads were washed twice with PBS using a magnet and re-suspended in the initial volume. 50 μl of the murine serum were added and incubated for 30 min at room temperature (RT), after which the beads were washed five times with PBS to remove serum debris. 0.5 μl of Protease Inhibitor Mix (GE healthcare) were added before the peptide mixture to avoid potential antibody degradation. The sample was then incubated for 2 h at RT with gentle tilting and rotation. After incubation, beads were washed three times with 1 ml PBS, and the bound peptides were eluted in 50 μl of 0.2% TFA. The elution fraction was concentrated and washed with C18 ZipTips (Millipore) and eluted in 2 μl of 50% ACN and 0.1% TFA. Subsequent MALDI-MS analysis of the eluted fractions was carried out. For more details, see Gourlay et al., [41].
All peptides were manually assembled by stepwise Fmoc-SPPS onto a 2-Chlorotrityl chloride resin (2-CTC), using HBTU/DIEA for in situ activation of entering amino acids [47]. Piperidine 20% in DMF was used for Fmoc removal steps. Cysteine was added to the N- of peptides to enable specific conjugation to carrier proteins, using poorly immunogenic PEG units as spacers. Upon completion of peptide assembly, peptides were simultaneously cleaved from the resin and side chains were deprotected by treatment with a mixture of 2.5% water, 2.5% thioanisole, 2.5% ethanedithiol, 2.5% triisopropylsilane and 90% trifluoroacetic acid for 2 hours at RT. Crude peptides were precipitated in cold diethyl ether, collected by centrifugation and washed with further cold ether Peptides were subsequently dissolved in 50% aqueous ACN/0.1% TFA and purified by C18-RP-HPLC. Peptide purity and identity was assessed by analytical C18-RP-HPLC and separate ESI-MS analysis.
Peptide N-terminal cysteine residue, preceded by the PEG spacer, allowed selective conjugation to carrier proteins (human serum albumin, (HSA), Hemocyanin from Concholepas (KLH) and Rabbit Serum Albumin (RSA)) using sulfosuccinimidyl 4-(N-maleimidomethyl) cyclohexane-1-carboxylate (Sulfo-SMCC) bifunctional linker. All peptides were prepared in free- and conjugated forms. Polyclonal antibodies were raised against peptide epitopes in rabbits (Primm srl, Milano Italy). Antisera against the FliCBp peptides were immunopurified against the peptides chemically linked to Cyanogen Bromide Activated Sepharose (Sigma-Aldrich).
Rabbit or human antibody recognition was detected as previously described [41]. In brief, 96-well polystyrene plates (Nunc Maxisorp) were uncoated or coated with 50 μl/well of 1 μg/ml of B. pseudomallei K96243 crude extract (crude Bps), 3 μg/ml of FliCBp protein or HSA-conjugated predicted peptides or HSA in 0.1 M carbonate-bicarbonate buffer (pH 9.6), and incubated at 37°C for 3 hr. After washing, 50 μl/well of 1:300 diluted human plasma or 1:3,000 rabbit antisera samples were probed in duplicate. Immunoreactivity was detected and represented as absorbance index of individual samples = (O.D.test−O.D.uncoated) / O.D.uncoated.
PBMCs from each donor were isolated by density gradient centrifugation on Ficoll-Hypaque and stored at -80°C with 10% dimethyl sulfoxide (DMSO) in fetal bovine serum (FBS). Immediately prior to the experiment, frozen PBMCs were thawed and resuspended in RPMI 1640, supplemented with 10% FBS, 200 U/ml penicillin, and 200 mg/ml streptomycin. 5 x 105 PBMCs/well were plated in 96-well culture plates and cultured with fixed Bps (PBMCs:organism = 1:30), 3 μg/ml of phytohaemagglutinin (PHA), 10 μg/ml of recombinant FliCBp, and 50 g/ml peptides for 48 h. Gamma interferon (IFN-γ) and interleukin 10 (IL-10) levels in the supernatant were quantified using a cytokine detection ELISA kit (BD biosciences) according to the manufacturer’s instructions. The sample with a detectable cytokine level higher than the lower limit of detection (>45 pg/ml) was described as responder.
Human PMN cells were purified by 3.0% dextran T-500 sedimentation and Ficoll-PaquePLUS centrifugation (Amersham Biosciences), as previously described [41]. Generally, PMN cell purity was >95%, as determined by Giemsa staining and microscopy; viable cells were counted by trypan blue exclusion (viability >99%). 1 x 109 CFU/ml of intact B. pseudomallei K96243 cells were labeled with 1 mg/ml of FITC in the dark at room temperature for 1 h. Intensity FITC on labeled bacteria was measured by flow cytometry. 7.5 x 107 CFU/ml FITC-labeled bacteria were treated with medium alone or antibody at concentrations of 40 μg/ml, for 1 h at 37°C, before being cultured with purified human PMN cells for 15 min at 37°C. 1:50 rabbit anti-FliCBp antisera was used as positive control for phagocytosis assay, while, 800 ng/ml of PMA (Sigma-Aldrich) was used as a positive control for the oxidative burst assay. Subsequently, 25 μl of 2,800 ng/ml of hydroethidine (HE) (Sigma-Aldrich) was added and incubated for 5 min at 37°C, washed twice and fixed with 2% paraformadehyde. Phagocytosis of FITC labeled Bps+ and oxidative burst activities of PMN cells that turn hydroethidine (HE) to ethidium bromine (EB+) were analyzed by flow cytometry (FACSCalibur; BD Biosciences). Results are represented as % phagocytosis (total % FITC+) and % oxidative burst (% EB+ FITC+) (details shown in S1 Fig). The differences between antibody groups were tested by the paired t test.
Live B. pseudomallei K96243 was opsonized, with or without 1 μg/ml of anti-FliCBp peptide antibodies, as described above. Rabbit pre-breed serum at 1:50 was used as negative control while rabbit anti-FliCBp antiserum at 1:50 was used as positive control.
Purified human PMN cells at 5 x 105 cells were incubated with opsonized bacteria (multiply of infection; MOI = 10) for 30 min at 37°C. Extracellular bacteria were subsequently killed upon addition of 250 μg/ml of kanamycin for 30 min, at this time point; PMN cells were lysed and plated on LB agar for bacterial counts (T0). In another condition, 20 μg/ml of kanamycin was added to maintain extracellular bacteria and incubated for 3 h at 37°C, and bacteria numbers were counted (T3).
Epitope predictions were carried out on the full-length (residues 1 to 388) amino acid sequence of FliCBp (UNIPROT code H7C7G3), using the web-accessible prediction servers BepiPred, BCPred and AAP [21,22]. Six linear B cell epitopes were identified (Fig 1) and subsequently used for T cell epitope predictions using the IEDB server [23,48] (www.iedb.org), selecting for HLA DRB1 alleles common in Thailand; HLA DRB1*0301, 0405, 07, 09, 1202, 1501, 1502 and 1602 [28]. A detailed overview of the consensus between linear epitopes predicted by sequence-based web servers is shown in Fig 1. The top three epitopes that were predicted to be both B and T cell epitopes are F51-69, F96-111 and F270-288. Interestingly, F51-69 overlaps with Peptide 6 (residues 51–70); F96-111 overlaps with Peptide 10 (residues 91–110), and F270-288 overlaps with Peptide 28 (residues 271–290) from the peptide panel synthesized by Musson et al. The following peptides were synthesized as described in the Materials and Methods section: F51-69; 51-TRMQTQINGLNQGVSNAND-69, F96-111; 96-VQASNGPLSASDASAL-111, F270-288; 270-NATAMVAQINAVNKPQTVS-288.
Predicted B and T cell epitopes were initially evaluated for immunorecognition by probing them with rabbit anti-FliCBp anti-sera. All three peptides were recognized (Fig 2). Subsequently, crude B. pseudomallei extract (Bps), recombinant FliCBp, and peptides F51-69, F96-111 and F270-288 were tested against human plasma samples for immunorecognition (Fig 3). In order to investigate a possible role of the peptides in protection, samples were tested from diverse sub-populations, including healthy seronegative, healthy seropositive and recovered melioidosis patient groups. Results showed that F51-69 and F270-288 were recognized by antibodies in human plasma, in contrast to F96-111. In addition, antisera from recovered individuals were shown to recognize Bps extract, FliCBp, and all peptides, except for F96-111, to a greater extent than those from healthy seropositive and seronegative individuals (Fig 3).
Moreover, upon comparison of antibody reactivity between healthy sample groups, we found that the levels of reactivity to crude Bps, F51-69 and F270-288 in the seropositive group were significantly higher than those in the seronegative group, p < 0.001, p < 0.001 and p < 0.05, respectively (Fig 3B). In case of FliCBp, there was a background of antibodies in seronegative group that might be due to B. thailandensis flagellin as it shares 91% similarity to B. pseudomallei flagellin compared by Basic Local Alignment Search Tool (BLAST). However, the levels of antibody against FliCBp in melioidosis recovered group were significantly higher than both seronegative and seropositive groups. Taken together, this suggests that people who are exposed and/or infected with B. pseudomallei develop antibodies against FliCBp, and more specifically against F51-69 and F270-288 epitopes (Fig 3). In contrast, peptide F96-111 was not recognized in human samples, despite high reactivity against rabbit anti-sera. This is likely due to the fact that this peptide is accessible in the recombinant protein, but not when FliCBp is assembled in the flagella, as supported by the observation that antibodies raised against F96-111 do not recognize crude B. pseudomallei (S2B Fig).
An additional confirmation of antibody specificity was illustrated by coating ELISA plates with crude Bps extract and each synthesized peptide. Each plate was probed with human plasma samples, previously neutralized with the same extract/peptide at different concentrations (S3 Fig). According to this test, peptides F51-69 and F270-288 inhibited the activity of antibodies in the recognition of their corresponding sample coated on the plate in a dose-dependent manner (S3 Fig).
Following our analysis of the FliCBp crystal structure (see below) and comparison with FliC from S. typhimurium, it became apparent that F96-111 is located in a region of the protein that mediates monomer-monomer interactions during assembly of FliC into its 11-subunit ring protofilament arrangement [49]. Therefore, when FliCBp is assembled in its native form in the flagella, this region would not be solvent accessible, in agreement with its lack of recognition by human antibody.
To further our studies, we then assessed whether the peptides could be active as human T cell epitopes, as implied by the bioinformatics predictions. Peptides were cultured with healthy PBMCs for 48 h, and the culture supernatants were measured for IFN-γ and IL-10 production using ELISA, as described in the Materials and Methods. We found that almost all samples (19/20; 95%) responded to intact killed Bps via IFN-γ production, some responded to FliCBp (13/20; 65%), F51-69 (15/20; 75%), F96-111 (9/20; 45%) and F270-288 (7/20; 35%), as shown in Fig 4. According to our results, FliCBp protein and predicted epitope peptides, in particular F51-69, elicited IFN-γ production from human PBMCs. With regards to an IL-10 response, all samples responded to fixed Bps, although with variable intensity. Recombinant FliCBp and peptide F51-69 responded in all samples (100%), however only 13/20 samples responded to F96-111 (65%) and 16/20 samples to F270-288 (80%). By combining the results of immunoreactivity against patient sera and the stimulation of cytokine production, we may confirm that epitopes F270-288 and F51-69, in particular, are promising B cell and T cell epitopes.
Our results show that antibodies against F51-69 and F270-288 are present in plasma of healthy donors who have been exposed to B. pseudomallei. Therefore we aimed to investigate whether these antibodies may stimulate phagocytosis by neutrophils and induce bacterial killing. Rabbit polyclonal antibodies were raised against F51-69 and F270-288 peptides (see Materials and Methods). Prior to experiments, the activity of antibodies against intact B. pseudomallei, recombinant FliCBp and peptides were checked by indirect ELISA. Results showed that all rabbit antisera recognized intact B. pseudomallei, recombinant FliCBp and immunized peptides (S3 Fig). Antibodies were subsequently used in B. pseudomallei opsonization tests and analyzed for bacterial phagocytosis and oxidative burst production in human neutrophils. Results revealed that antibodies raised against recombinant FliCBp and predicted epitope peptides enhance both phagocytosis and oxidative burst activity in human neutrophils (Fig 5A and 5B). Interestingly, we also found that these antibodies also enhance bacterial uptake and intracellular bacterial killing in human neutrophil infection studies (Fig 5C). These results suggest that antibodies against F51-69 and F270-288 may enhance host resistance to B. pseudomallei infection by stimulating neutrophil phagocytosis and bacterial killing.
FliCBp (residues 25 to 378) was crystallized using the sitting drop vapor diffusion method, as described in the Materials and Methods section. The crystal structure of FliCBp was solved at a resolution of 1.3 Å by molecular replacement, using the structure of domain D1 (residues 62–168 and 281–326) of Sphingomonas sp. A1 flagellin (PDB entry 2ZBI) as an initial search model, and refined to satisfactory Rgen and Rfree values of 13.6% and 16.4%, respectively (statistics for the data collection and model refinement are shown in S1 Table). Electron density was visible for residues 69–326; the first 44 N-terminal residues and the 52 C-terminal residues were absent, however experimental epitope mapping revealed two immunocaptured peptides containing N-terminal residue 37 and residue 361, therefore the residues that are not visible in the electron density are likely to be present in disordered regions of the protein.
Flagellins vary in dimension (28–65 kDa) and contain at least two essential highly conserved domains, D0 and D1 that mediate assembly of the flagellar filament, and may contain a second (D2) or third (D3) variable domain. Typically, removal of the D0 domain is necessary for successful crystallization. This was not the case for FliCBp where, however, the first 44 N-terminal residues (corresponding to the D0 domain) are not visible in the electron density map because of structural disorder. Overall, FliCBp adopts the canonical flagellin fold and presents two domains, a conserved D1 domain (residues 69–167, 287–326), and a variable D2 domain (168 and 286) (Fig 6). D1 forms a helical rod (α1, α2 and α5) involving residues contributed by both the N-terminus (residues 69–170) and C-terminus (residues 285–326), a β-hairpin domain (β1-β2) and a 310 α-helix (η1). Domain D2 (residues 172–287) contains a three-stranded antiparallel β-sheet, two α−helices (α3 and α4) and two 310 α-helices (η2 and η3). Domain D2 is connected to D1 via two anti-parallel regions that house two short 310 α-helices (η1 and η3) that interact with each other, thus stabilizing the relative positions of the two domains (Fig 6).
The closest structural homolog to FliCBp available in the Protein Data Bank is flagellin from Pseudomonas aeruginosa (FliCPa; PDB entry 4NX9) [50], with whom it shares 36.5% sequence identity and 96.2% secondary structure similarity and a rms deviation of 1.34 Å over 168 aligned Cα pairs (S4 Fig). The structural similarity is limited to the D1 domains, while domain D2 is different in FliCBp and FliCPa. The main difference arises due to the 7 β-strands present in the D2 domain of FliCPa in comparison with three present in FliCBp (S4 Fig). Furthermore, due to the diverse tertiary positioning of the D2 domains, when the D1 domains are aligned, the D2 domains do not superimpose (S4 Fig).
Both D1 and D2 domains have been shown to be important for innate immune recognition [51,52]. With regards to the location of the predicted epitope peptides on the 3D structure, F51-69 could not be mapped onto the structure as electron density was absent for this region, indicating that it is likely to be non-structured. F96-111 is housed in domain D1 and is composed of two short α-helical segments at the C- and N- termini of α-helices 1 and 2, connected by a short loop.
The third predicted peptide F270-288, is located in domain D2 and is formed by α-helix 4, the 310 α-helices η3, and their connecting loop region (Fig 6).
In the context of a Structural Vaccinology project, structure-based computational epitope predictions have proved successful for the identification and design of B. pseudomallei epitopes with improved immunological properties in comparison with their cognate recombinant proteins, commencing from the crystal structures of the target antigens [41,53].
In order to expand the investigation on FliCBp beyond the reach of the bioinformatics sequence-based predictions, and possibly complementing it, we employed two structure-based computational methods for epitope identification, and compared the results with the previous consensus. The newly solved 3D structure of the antigen, allows taking a different perspective on epitope discovery. Here, two complementary computational methods for epitope prediction were combined and applied to representative structures obtained from molecular dynamics (MD) simulations run on the FliCBp crystal structure (see Materials and Methods) [54]. MLCE selects antigenic B- and T-cell epitopes while the EDP identifies general protein-protein interaction interfaces. Predictions produced individually by MLCE and EDP are aligned with the sequence-based consensus and presented in S5 Fig.
Epitope mapping experiments were carried out using recombinant FliCBp and cognate polyclonal sera. To this aim, we adopted and extended an immunocapturing approach successfully used previously with monoclonal antibodies [55]. The approach involves proteolytic digestion (using diverse proteases) of the target antigen prior to immunocapturing, and subsequent analysis of antibody-bound peptides (containing epitopes or fragments thereof) by mass spectrometry. Using trypsin for the partial digestion of FliCBp, 5 different peptides were captured by the polyclonal IgGs and subjected to MS/MS analysis: a 1529.807 Da peptide corresponding to the N-terminal segment 37-INSAADDAAGLAIATR-52 (1529.6 Da); a peptide of 2231.122 Da corresponding to 300-QAMVSIDNALATVNNLQATLGAAQNR-325 (2685 Da); two 2771.356 Da and 3310.671 peptides Da represented the same region of the protein with segment 98-ASNGPLSASDASALQQEVAQQISEVNR-124 (2770.9 Da) and 93-QLAVQASNGPLSASDASALQQEVAQQISEVNR-124 (3310.5 Da), and a 3782.885 Da peptide corresponding to 326-FTAIATTQQAGSNNLAQAQSQIQSADFAQETANLSR-361 (3783 Da) (S5 Fig).
Overall, the addition of experimental mapping, as well as EDP and MLCE analysis confirm the accuracy of the previous results, being among the areas of the alignment displaying the best consensus (S5 Fig). The other putative active epitopes, consisting of the overlap of two or more methods, include fragments V148-T155; D188-T203; G236-F250 and L309-F326. According to the new consensus, these sequences represent good candidates for further tests of immune recognition, and are currently being investigated.
Among the activities presented in this communication, we selected an additional T cell epitope to be tested, based on the consensus between IEDB (sequence based) (Table 1) and MLCE (structure-based) (S5 Fig), both having specificity for T-cell epitopes. The linear stretch K213-V231 is not selected by any other B cell specific prediction method, and may be an interesting candidate to be assessed for T cell activation. In the 3D FliCBp structure, this peptide mainly is formed by α-helix 3 and β-strand 7 that pertain to Domain 2 (Fig 6). We synthesized the new epitope F213-231 and repeated the experiment of PBMC stimulation for cytokine production (Fig 4). This peptide epitope was shown to stimulate IFN-γ production in 8/20 samples (40%) and IL-10 production in all 20 samples tested (100%). These results correlate with our previous results that show that this peptide segment can elicit IFN-γ [19]. In parallel, we assessed immune sera recognition of epitope F213-231 and observed that it is not seroreactive, suggesting that it is a potential T-cell epitope but not a B cell epitope (Fig 3C).
In the melioidosis vaccine development field, several protein antigens have been tested for their ability to trigger a protective immune response in animal models of disease, however, protection has proven to be limited in all cases [7]. In addition to the adaptive immune response, innate immunity also plays an important role in resistance to B. pseudomallei infection [3,56]. Neutrophils are one of the key players in cellular immunity that stimulate bacterial killing, both directly or indirectly via cytokine production [57,58]. Therefore, it is likely that a future melioidosis vaccine should comprise antigens/epitopes that trigger both arms of the immune system.
Antigen design guided by computational biology is at the forefront of vaccine development, focusing on the design of only immunogenic portions of the antigen that may be domains or even linear peptides. In fact, epitopes may display elevated antigenic activities in comparison with their parental antigen, as previously observed for two acute phase antigens from B. pseudomallei [41,53].
In this context, we focused on flagellin from B. pseudomallei as a target for epitope discovery and design, using both sequence-based and 3D structure based approaches. Flagellins from diverse bacteria have been studied as subunit vaccine candidates; among these, S. typhimurium FliC induces responses from Th2 in primary immunization tests, followed by Th1-dependent responses that occur later during subsequent infection, which leads to bacterial clearance [59]. Furthermore, protective antibody production against B. pseudomallei infection has also been reported for mice immunized through a modified plasmid encoding the fliC gene, combined with CpG oliogodeoxynucleotide [11]. Several immunogenic epitopes have been identified from FliC homologs, e.g. FliC from Borrelia burgdoferi [60] and FliC from Salmonella typhimurium, from which four epitopes were identified as CD4+ specific T cells epitopes [61,62].
Using sequence-based computational predictions, we identified and synthesized three peptides (F51-69, F96-111 and F270-288) with predicted T-cell and B-cell stimulatory activities. Two peptides (F51-69 and F270-288) were in fact found to be B-cell epitopes and were reactive against human antibodies (IgG) isolated from melioidosis-infected subjects and healthy seropositive subjects. The third peptide (F96-111) was not reactive with sera and, based on 3D-structural considerations, and in light of the crystal structure here presented, it is unlikely to be accessible to antibodies when FliC is assembled in the flagella. In fact, structural comparisons of FliCBp and S. typhimurium flagellin (FliCSt), for which several structural and flagellin assembly studies have been carried out, indicate that F96-111 is housed in a key site that in FliCSt mediates axial interactions and protofilament assembly, being also responsible for TLR5 recognition of the monomeric protein [63,64]. As only monomeric FliC activates TLR5, it has been suggested that cellular recognition of FliC requires flagella degradation, for example via phagocytosis or other depolymerization mechanisms [63]. Such an interface location for F96-111 in the assembled protofilament structure was confirmed by the inability of anti-F96-111 antibodies to recognize crude B. pseudomallei extract, and accounts for the fact that the epitope was not recognized by human antibodies. Thus, while it remains a possible T-cell epitope, F96-111 is unlikely to be a B-cell epitope in FliCBp oligomeric native state.
Based on sequence alignment with FliCSt for which residues involved in intermolecular flagellin contacts are known [63,64], F51-69 should also be part of the assembly interface, and should be buried in the flagella, thus inaccessible to antibodies. However, residues 51–69 are absent in the FliCBp structure, suggesting that they are structurally disordered, in contrast to equivalent residues from FliCSt that instead pertain to the N- terminal α-helix, suggesting that the assembly interfaces may differ between the species. In fact, the F51-69 peptide is recognized by human antibodies, demonstrating that this region is accessible in the flagella and not buried as in FliCSt.
With regards to the cell-mediated response, our results show that all predicted T/B cell FliCBp epitopes (in particular F51-69) induce cytokine production (IFN-γ and IL-10) from human PBMCs, and that their cognate antibodies also enhance human PMN phagocytosis and bacterial killing by increasing oxidative burst activity.
We also report the 1.3Å resolution FliCBp crystal structure as the basis for in silico epitope predictions using two diverse computational methods. Our approach, besides allowing us to cross-validate sequence-based and 3D-structure-based epitope prediction methods (resulting in the identification of consensus epitope regions), identified a fourth epitope (F213-231) that, when synthesized as a peptide, was found to be a T-cell epitope but not a B-cell epitope. Significantly, the in silico (structure and sequence-based) predictions were complemented by in vitro epitope mapping that identified residues predicted by both sequence-based and structure-based methods.
As previously mentioned, it is likely that several protective antigens will be required to formulate an effective protective vaccine [7,65]. Previous experimentation on mice immunization showed that the addition of adjuvants is often necessary in combination with subunit vaccine candidates to elicit protection [65]. TLR agonists are widely used as adjuvants, including flagellin, which is a TLR5 agonist and a potent activator of both the innate and adaptive immune responses [18,50]. This implies that the potential of the T-cell/B-cell FliCBp peptides identified in this study could represent adjuvants to be administered with other antigen candidates to develop a protective melioidosis vaccine. In fact, a truncated form of FliCSt was shown to induce a significantly stronger cellular immune response when administered in a chimeric subunit vaccine together with Eimeria tenella immune mapped protein-1 (EtIMP1), in comparison to Freund’s Complete Adjuvant [66]. An additional successful application of FliC as a carrier protein was reported for the glycoconjugate subunit vaccine that improved protection against fetal infection disease [67].
In conclusion, our results shows that both sequence-based and 3D-structure-based epitope discovery methods, and the derived consensus sequence segments, proved successful in the identification of both T-cell and B-cell epitopes from FliCBp. Validatory in vitro epitope mapping and immunological tests fully confirmed the predictive analyses. Taken together, our data suggest that the predicted peptide epitopes with confirmed T-cell/B-cell activities should proceed to further in vivo immunological tests as potential vaccine components, possibly in conjunction with other known B. pseudomallei antigens. To this aim, further analyses of additional peptides suggested by the 3D-structure-based predictions will be undertaken to enrich the repertoire of potential protective epitopes.
|
10.1371/journal.pntd.0004044 | Analysis of Dengue Virus Genetic Diversity during Human and Mosquito Infection Reveals Genetic Constraints | Dengue viruses (DENV) cause debilitating and potentially life-threatening acute disease throughout the tropical world. While drug development efforts are underway, there are concerns that resistant strains will emerge rapidly. Indeed, antiviral drugs that target even conserved regions in other RNA viruses lose efficacy over time as the virus mutates. Here, we sought to determine if there are regions in the DENV genome that are not only evolutionarily conserved but genetically constrained in their ability to mutate and could hence serve as better antiviral targets. High-throughput sequencing of DENV-1 genome directly from twelve, paired dengue patients’ sera and then passaging these sera into the two primary mosquito vectors showed consistent and distinct sequence changes during infection. In particular, two residues in the NS5 protein coding sequence appear to be specifically acquired during infection in Ae. aegypti but not Ae. albopictus. Importantly, we identified a region within the NS3 protein coding sequence that is refractory to mutation during human and mosquito infection. Collectively, these findings provide fresh insights into antiviral targets and could serve as an approach to defining evolutionarily constrained regions for therapeutic targeting in other RNA viruses.
| Dengue viruses cause debilitating and potentially life-threatening acute disease throughout the tropical world. While drug development efforts are underway, there are concerns that drug-resistant strains will emerge rapidly. Indeed, many antiviral drugs for other RNA viruses lose efficacy over time as the virus mutates. Here, we sought to determine if there are regions in the dengue virus genome that are constrained in their ability to mutate and could therefore serve as better targets for antiviral drugs. Deep sequencing of the dengue virus 1 genome directly from the blood of twelve dengue patients and from mosquitoes given this blood showed consistent and distinct mutation patterns during infection. Importantly, we identified regions within the viral genome that are resistant to mutation during human and mosquito infection. Collectively, these findings provide fresh insights into potential antiviral targets and could serve as an approach to defining better regions for therapeutic targeting in other RNA viruses.
| Dengue, caused by one of four dengue viruses (DENV), is the most important arboviral disease in the world. Recent estimates indicate DENV infects 400 million people annually and over half of the world’s population lives in regions endemic for this debilitating and potentially life-threatening disease [1]. DENVs are single-stranded, positive sense RNA viruses within the Flaviviridae family of viruses and are transmitted from human-to-human in most parts of the world by Aedes aegypti and Aedes albopictus [2]. The DENV genome is approximately 11kb long and encodes a single viral polyprotein that is then post-translationally cleaved into three structural proteins—the capsid (C), pre-membrane (prM) and envelope (E)—and seven non-structural (NS) proteins, NS1, NS2a, NS2b, NS3, NS4a, NS4b and NS5. The known and suspected functions of these proteins have been reviewed elsewhere [3,4]. The viral coding region is flanked by a short 5’ untranslated region (UTR) and a longer 3’ UTR, both of which have been shown to associate with host factors and form secondary and tertiary structures that are required for viability of the virus [3]. Dengue prevention relies solely on vector control, which in most places has not resulted in sustainable reduction in disease incidence. While vaccine development has made important strides recently, the efficacy against all four DENV serotypes is variable and protection against infection is incomplete [5,6]. An antiviral drug that specifically combats DENV remains a much-needed tool against this global scourge.
Antiviral drug development has mostly focused on compounds targeting conserved regions of the viral genome. Despite such an approach, drug resistance has developed rapidly, particularly for RNA viruses. RNA viruses are indeed notorious for their ability to adapt quickly to selective pressure from the host immune system and/or antivirals [7,8]. This adaptability can largely be attributed to their existence as a population and the error-prone characteristics of their RNA-dependent RNA polymerase (RdRp) [9–11]. These features combine to make RNA viruses able to quickly adapt to selective pressure from the host or antiviral treatment by exploring available sequence space [12–22]. Combination therapy is thus required to prevent rapid emergence of drug resistant strains and this strategy have been successful for human immunodeficiency virus (HIV) and hepatitis C virus (HCV) [23,24]. However, such a therapeutic approach may not be suitable for viruses such as dengue or chikungunya. The cost of treatment would increase with each additional drug and the tropical world, where these viruses are prevalent and cause significant economic burden, may not be able to afford the treatment needed. Identification of regions within the DENV genome that are not only evolutionarily conserved but also genetically constrained could thus pinpoint potent and resilient targets for monotherapy that minimizes risk of resistance emergence.
To this end, we analyzed intra-host genetic diversity of DENV1 at day 1–3 and again at 4–7 following onset of fever in 12 dengue patients. The sera from these patients were then intra-thoracically inoculated into both Ae. aegypti and Ae. albopictus and analyzed after 10 days of infection (Fig 1a). This method of viral delivery to the mosquito bypasses the bottlenecking event the virus encounters in the midgut barrier [25] and was necessary due to the limited amount of patient sera available. This method does, however, allow us to explore the full mutational space available to the virus when not confronted by this bottleneck thereby allowing a more complete picture of which areas in the genome tolerate a degree of variability without sampling hundreds of natural infections. Conversely, we were also able to identify those regions where variability was significantly reduced. These areas of reduced variation, hereby referred to as constrained, likely represent residues lethal to the virus if mutated. Using the resolution enabled by next generation sequencing (NGS) technologies [12], we show that there is an abundant accumulation of intra-host viral population diversity in both humans and mosquitoes. Unexpectedly, we observed specific variations in the DENV genome in Ae. aegypti not present in Ae. albopictus, suggesting that amid the stochastic variations, there are distinct changes critical for DENV to thrive in each mosquito host. Importantly, we also show that there are regions of constraint within the viral genome that are refractory to variation in both human and mosquito.
The intra-host genetic diversity of DENV1 was analyzed in 12 individuals that were enrolled in the early dengue infection and outcome (EDEN) study (S1 File) [26,27]. Consensus sequences of DENV1 isolated from these 12 individuals and grown in C6/36 cells have been reported previously [28]. DENV1 genomic material from paired serum samples was also taken at fever day 1–3 (early) and 4–7 (late) from each patient. The patients were a mixture of primary and secondary infection and the final diagnosis for each of them are shown in S1 File [28]. We tested for differences between primary and secondary infections using the non-parametric Wilcoxon-Mann-Whitney test but none were statistically significant. DENV1 genomic material from these samples was PCR amplified and deep sequenced. These same serum samples were also inoculated intrathoracically into 4-day old female Ae aegypti and Ae albopictus. After 10 days of incubation, the ten mosquitoes for each serum sample were pooled to minimize sampling bias from individual mosquitoes. DENV1 was then PCR amplified from the total RNA and deep sequenced on an Illumina or Solid sequencing platform (Fig 1a and S2 File). In order to check whether different sequencing technologies (Solid and Illumina) had an effect on the type of SNPs detected, we performed Fisher's exact test on the number of Transition and Transversion SNPs of serum samples sequenced by Solid and Illumina. The results of this analysis suggest that there are no significant differences in any gene between Solid and Illumina sequencing (S3 File).
Overall, our deep sequencing data shows positional variance throughout the DENV1 genome (S4, S5 and S6 Files). The 17 positions where consensus discordance was observed in at least three of the twelve viruses are shown in Fig 1b and the complete list of all consensus changes observed in our data set are described in S7 File. These consensus changes fall within the coding sequence and the 3’UTR. Two of these consensus changes, one in prM and the other in NS5 also resulted in changes to the protein coding sequence.
Besides the limited number of consensus changes, there are a large number of positions throughout the viral genome that display a degree of intra-host viral diversity. We refer to these types of positions as having ‘variance’. To distinguish variants from the average sequencing error rate, we used the program Lofreq, which identifies single nucleotide polymorphisms by incorporating base-call quality scores as error probabilities into its model and assigns a p-value to each variant [29]. These analyses identified seven positions within the DENV genome that possessed this level of reproducible plasticity: two in the E gene, one in the NS1 gene, one in the NS3 gene, one in the 2k peptide at the C terminus of the NS4a gene and two in the NS5 gene (S4 File).
Since our samples were extracted from the same patients at two time points during their infection and then directly inoculated into the two mosquito vector species, we were able to track these changes in the genetic diversity across the viral genome over time. More specifically, we compared the proportion of base calls at each position in the DENV genome in early and late serum samples as well as between human and Ae. aegypti or Ae. albopictus in both early and late stages of acute dengue. Our results indicate that during the course of the human infection, changes in the intra-host genetic diversity were more prevalent in the NS1, NS2A and E genes (NS2A vs NS2b Bonferroni corrected p-value [Bcp] = 0.008; NS2A vs NS3 Bcp<0.001; NS2A vs NS4B Bcp = 0.02; NS2A vs NS5 Bcp<0.001; E vs NS3 Bcp = 0.006; also NS1 vs NS3 Bcp = 0.002). The average number of changes occurring over the course of four days of human infection is 86 or ~0.0020 changes/position/day of human infection. In Ae. albopictus, changes were observed in E, NS1, NS4A (2k peptide) and NS5 genes. In Ae. aegypti, changes were observed in prM, E, NS1, NS3, NS4A (2k peptide) and NS5 genes (Fig 2 and S5 File). Two of the most commonly observed changes at 2719 (NS1) and 6782 (2k peptide) were observed in both species of mosquito and suggests that selection pressure on these residues is likely to be a common mechanism shared between the species. Interestingly, there were changes that were unique to Ae. aegypti infection, notably at 9986 and 9998 in the NS5 gene. These changes suggest that differential selection pressures may be applied on selected nucleotide residues in the DENV genome by Ae. aegypti but not by Ae. albopictus.
For each detectable change in the DENV genome over the course of the human infection, we defined whether the proportion of base calls at each position moved towards or away from the consensus base after 10 days of incubation in the vector. Our results indicate that the variance acquired between early and late serum samples undergo a reversion back towards the sequence in the early serum sample after 10 days of incubation in either vector (Fig 3). We then asked whether this reversion was happening in specific regions or was a more general mechanism. Our data suggest that reversion is largely a general phenomenon that occurs across the majority of the viral genome (S8 File). The majority of the reversion events are small oscillations in the overall composition at each position; however, larger consensus-level changes were also observed (S9 File).
Selection pressure and genetic drift were measured by calculating the dN/dS ratio. Overall it can be seen from the mean dN/dS ratio for each group that there is likely a purifying selection pressure against non-synonymous mutations (S10 File). The Mann-Whitney test was used to compare single protein coding sequences against the rest of the polyprotein and results suggest that there are no significant differences within the human samples (Early and Late), whereas significant differences were identified within the Ae. aegypti and Ae. albopictus samples (S11 File). The transition/transversion analysis and the Shannon diversity index and Shannon equitability measurements suggest that there is a decrease in mutation frequency from early to late samples in Ae. aegypti and human whereas an increase in the mutation rate was observed in the Ae. albopictus samples (S12 and S13 Files respectively). We also questioned whether the two time points for each species were likely to come from the same population and results suggest that the dN/dS ratio is significantly different between the Ae. aegypti early and late samples (S14 File). From the Ts/Tv ratio comparison, the ratio is significantly different (<0.01) between the Early and Late samples in Ae. aegypti, Ae. albopictus and human. This trend is consistent with the results from the Shannon diversity index and Shannon equitability measurements (S13 File).
The heterogeneity observed at positions 9986 and 9998 (NS5) in the DENV genome fall within the RdRp domain of NS5 and correspond to amino acids 541 (Thr → Ala) and 545 (Leu → Leu) at junction of the “palm” and the α14 alpha helix “finger” of the RdRp domain of the protein, respectively (Fig 4) [30]. That these observations were unique to Ae. aegypti suggests that these are not random events but are responses to species-specific selection pressure. To test this possibility experimentally, we constructed an infectious clone of DENV1 isolated from the same outbreak in Singapore in 2005 [26] but from a patient not among the 12 studied here. This infectious clone was constructed with the exact nucleotide sequence of the virus (GenBank: EU081230) that was isolated in the C6/36 Ae. albopictus derived cell line [26]. In vitro transcribed RNA was electroporated into BHK cells and harvested supernatant was inoculated intrathoracically into both Ae. aegypti and Ae. albopictus. The initial starting material and time points of 5, 10 and 21 days post intrathoracic inoculation were then sequenced and the data analyzed in the same manner as described above. Although the number of replicates in this experiment is limited, our results nevertheless indicate that the changes observed in NS1 and the 2k peptide (Fig 2) are recapitulated in both Ae. aegypti and Ae. albopictus (Fig 4). The two Ae. aegypti specific residues in the NS5 gene, 9986 and 9998 (amino acid positions 541 and 545), arose 21 days after the infectious clone-derived virus was incubated in Ae. aegypti but not in Ae. albopictus (Fig 4). Collectively, these findings demonstrate that species-specific selective pressures act to select for variance in specific positions on the DENV genome.
The “palm” domain of the DENV RdRp is the most structurally conserved domain among all known polymerases [30]. Although there is no specific catalytic activity associated with residue at position 541, the Thr→Ala substitution may alter the angle of the “finger” relative to the “palm” and by doing so; alter the enzymatic properties of the RdRp. The nucleotide change at position 9998 does not translate to an amino acid shift at position 545 and its functional significance in regards to RdRp activity is not clear. However, examination of the predicted RNA secondary structure in this region suggests that nucleotides 9986 and 9998 interact with each other in a previously uncharacterized stem-loop structure (Fig 4C and 4D) [31]. The observed A→G and/or C→U changes at bases 9986 and 9998 respectively are predicted to strengthen this interaction (see reduced dG in MFE structure). Other structures and sequences within the virus have been shown to be essential for the virus in a species-specific manner and this may be another mechanism the Ae. aegypti vector uses to control DENV replication [32].
To identify regions of constraint within the DENV1 genome, we aggregated all predicted single nucleotide variants (SNVs) to detect regions with (i) a local enrichment in intra-host SNV calls (mutational hotspots) and (ii) a significant depletion in variants (mutational cold-spots) (Fig 5). This type of analysis complements classical approaches of finding evolutionarily conserved regions through multiple sequence alignments and can reveal functionally important, though otherwise not easily detectable regions [29]. No hotspots predicted in more than one sample of either the mosquito or the human isolates could be detected. However, at least four samples of the infectious clone experiment were observed to have a hotspot in the envelope protein (bases 1789–1814). As the virus used in the infectious clone experiment was in vitro transcribed and electroporated into BHK cells for packaging we suggest that this particular hotspot can be attributed to the markedly different selection pressures in cell culture conditions than encountered by the virus in vivo [33]. Overall the mosquito samples had 12 coldspots covering 1064 total positions whereas the human samples show only 2 coldspots covering 220 total positions, which is consistent with previous reports [29]. The absence of coldspots in NS1 and NS2A has been observed before [29], but its significance is unclear. Mosquito, human and infectious clone samples largely display an absence of shared mutational coldspots (i.e. regions that show intra-host constraint) with the notable exception of coldspots within the multifunctional NS3 gene (Fig 5). NS3 is comprised of a protease domain and an ATP-driven helicase with two subdomains. The large coldspot discovered in the mosquito samples covers all three of these domains. Structurally, this region clusters around the ATP-binding domain and around the interaction site with NS2B (Fig 5b) [34]. The coldspot in NS5 is in the fingers domain of the RdRp and forms part of the zinc-binding pocket (Fig 5a and 5c) [30]. Although the function of this zinc-binding pocket is unknown, it is a feature shared with the West Nile virus RdRp and is likely to be functionally important [34].
In order to test the hypothesis that coldspot regions would make good antiviral targets, siRNA’s were designed to target the hotspot in the E gene (starting at position 1848), the coldspot in NS3 (starting at position 4794) and a region near the Ae. aegypti specific mutations in NS5 (starting at position 10014). A non-targeting (NT control) was used as a control for the experiment and DENV1 genome copies relative to GAPDH were measured by RTPCR at 24 and 48 hours post infection. All DENV1 siRNA’s were significantly different than the NT control (p<0.001) at both 24 and 48 hours post infection (Fig 5d). The siRNA’s targeting the E gene and NS3 were indistinguishable from each other in their effect at both 24 and 48 hours post infection. Interestingly, although the siRNA targeting NS5 that contains the Ae. aegypti specific mutations was not statistically significant from the E gene and NS3 siRNA’s at 24 hours post infection, it was statistically less effective by the 48 hour timepoint (p = 0.0008 and p = 0.0003 respectively) (Fig 5d).
In this study, we have used high-throughput parallel sequencing to analyze intra-host genetic diversity in the DENV genome directly from serum samples obtained from dengue patients and intra-thoracically infected mosquitoes. Our data show that numerous positions within the DENV genome exhibit a high degree of plasticity over the course of infection within the human and mosquito hosts and that several of these changes are in functionally significant domains of the viral coding sequence.
Maintenance of genome plasticity within viral populations is a poorly understood process however; it is possible that it may be critical for the overall fecundity of the virus [9]. Interestingly, members of the mosquito-borne clade of flavivirus have been observed to be more genetically stable over time than other RNA viruses [10,33,35–39]. Previous studies have found that repeated passage in a single host is likely to result in a consensus genome that is highly divergent from its original source [40,41]. In studies measuring viral variance by clonal analysis of viral amplicons, significant intra-host variation of the virus has been observed in laboratory passaged DENV and other flaviviruses such as West Nile virus [33,36,37,42,43]. Serial in vitro or in vivo passages in mosquitoes show significant amounts of consensus changes throughout the genome [15,33,37,44]. The long-term stability of these viruses may therefore be due to differential selection pressures exerted by the human and mosquito hosts that result in a net conservation of the viral genome [33,37]. Our results suggest that there is an abundant accumulation of intra-host viral population diversity in humans and mosquitoes. Consistent with the prevailing theory however, our study indicates that the changes that accrued during infection in the human host predominantly revert back to the ‘original state’ as the virus transits through the mosquito. Intriguingly, this reversion is occurring despite bypassing the bottlenecking event of midgut barrier escape. This suggests that even when these humanized variants are given the opportunity to replicate in the mosquito body, they are outcompeted by the original population observed early in the human infection. Given the broad distribution across the viral genome and their stochastic appearance in our data set, these larger changes likely represent changes to locations tolerated in the human host but not in the mosquito vectors. The proportion of these variants that are then able to disseminate through the salivary gland infection/escape barrier and thus infect the next human host is also of interest and will be the subject of future study [45]. Taken together, these data provide evidence that cycling between the two hosts restricts the overall diversity in DENV genome, making it phylogenetically more stable than other RNA viruses that propagate within a single species.
Our study also cautions on performing phylogenetic analyses on consensus sequences alone; especially those derived from serially passaged virus. We have also identified species-specific variations that have not been previously reported. Some of these variations are recurrent among the samples we have tested suggesting that there are positions with a high degree of plasticity in the DENV genome. The locations of these positions appear to depend on the host, which provides new insights into how the different vectors may influence DENV evolution. These differences also suggest that viruses transmitted predominantly in an Ae. aegypti-human cycle may produce viruses genetically distinct from those transmitted predominantly in an Ae. albopictus-human cycle. Indeed, it may be a molecular basis for which epidemic emergence is more often associated with Ae. aegypti than Ae. albopictus [46]. Further studies are necessary to determine whether these same residues arise after passing through the midgut barrier and are ultimately present in the saliva of the infected mosquitoes [47,48].
Interestingly, our samples did not display the extensive mutations in the 5’ or 3’UTR that have been identified in recent studies [49–51]. We observed only a single position in the 3’UTR to change consensus and only for two of the isolates (S7 File). This consensus change falls within the unstructured region between DB1 and DB2 and is not predicted to have a substantial impact on the overall 3’UTR structures (S15 File). The reasons for the observed stability in this region are unclear. The aforementioned studies were primarily conducted in cell lines with DENV2 and, to a lesser extent, DENV3 [49–51]. Whether the DENV1 serotype is fundamentally different in this regard or whether this can simply be attributed to our limited sample size is an interesting question and deserving of additional studies.
The finding of areas within the DENV genome that are constrained in nucleotide variation in the human and mosquito hosts are interesting. These cold-spots are consistent even in two disparate host species and thus suggest that these positions may encode protein-protein interactions that are functionally vital to the DENV lifecycle. The cold-spot at the interface between NS2b and NS3 is particularly interesting. NS3 requires a direct interaction with NS2b as a cofactor for its proteolytic activity. Our findings suggest that if not outright lethal, mutations within this interaction site are likely to cripple the virus. Given that sequence in this region of the genome is highly constrained, a potentially attractive antiviral strategy may be RNA interference (RNAi) due to its potential for high specificity to the viral genome [52,53]. In this study, we tested three different siRNA’s targeting the E gene, NS3 and NS5. Although all were able to significantly reduce viral copy number, the siRNA targeting NS5 was not as effective as the other two after the first 24 hours of infection. Given the plasticity observed in this region when these isolates were passaged into Ae. aegypti, this may represent a ‘flexible’ part of the genome and a less than ideal target for this type of antiviral strategy. This may indeed also explain the lack of difference between siRNA that targeted the E gene compared to NS3 as, while there is a degree of variance in the E gene, it is not a ‘flexible’ part of the genome that alters depending on the host species. Indeed, it is possible that sub-therapeutic doses of siRNA against the E gene would be more likely to generate resistant mutants over repeated passages compared to that against the cold spot in NS3. This could be a useful focus in future investigations.
While we have used siRNA in this study, small molecule therapeutics against the cold spot on NS3 would be another option. Binding to either interaction surface should interfere with the catalytic function of the viral protease although our data suggests that the NS3 is a more constant target. A small molecule inhibitor would also have the potential advantage of inexpensive mass production; a distinct advantage for the treatment of affected populations unable to afford more expensive therapies.
The changes observed in the PrM and NS5 genes are remarkable as they are highly conserved across most flaviviruses [30,54–56]. The Leu → Phe mutation in the prM gene at position 138 occurs in the C-terminal transmembrane domain of the “M” residue that is embedded in the lipid bilayer of the mature virion [54]. The functional significance of this particular amino acid change is not immediately clear. The membrane composition of mosquito cells are substantially different from mammalian membranes, particularly in their cholesterol content [57]. It is conceivable that an alteration at this position may be in response to these differences and plays a role in the infectivity of the virus. The Val → Ala mutation at position 324 in the NS5 gene occurs at the N-terminus of the RdRp domain in a region involved in the binding of β-importin and NS3 [30]. Alteration in the ability of NS5 to interact with these proteins could directly impact the ability of NS5 to shuttle into the nucleus and the ability of the virus to replicate its genome respectively [30,58]. The lack of extensive cold-spots in the NS5 gene was surprising to us although this might be due to the fact that we pooled ten mosquitoes inoculated with the same serum together. While this methodology has the advantage of being more rigorous when trying to identify common variants, it may obscure authentic coldspots due the averaging effect of combining individuals with stochastic mutations in these regions. One cold-spot was identified in the fingers subdomain of NS5 though not in the thumb or palm domain, which contains the catalytic active site. The latter has been the focus of attention in anti-dengue drug development [59,60]. Furthermore, that we also found variance in two nucleotides in the RdRP domain, one of which changes the protein coding sequence, in DENV that replicated in Ae. aegypti but not Ae. albopictus also raises additional concerns on antiviral drug development efforts that target the RdRP. Most laboratories culture DENV in C6/36 cell line, which is derived from Ae. albopictus. Compounds that show attractive efficacy to DENV cultured in such cells may thus not achieve anticipated efficacy in humans who acquire infection from Ae. aegypti, which is the epidemiologically more important vector.
Finally, the analysis we have employed in this study can readily be adapted for other pathogen-host studies affecting the developing world such as influenza, chikungunya and ebola viruses. We suggest that our approach could serve not only to identify areas of constraint in viral genomes but also to monitor the emergence of escape mutants following vaccination or initiation of antiviral therapies [61].
The samples used in this study were collected under the Early Dengue infection and outcome study (EDEN). This prospective study was approved by the National Healthcare Group Domain Specific Review Board (DSRB B/05/013) and the Institutional Review Boards of the National University of Singapore and DSO National Laboratories. Enrollment of participants into the study was conditional upon written informed consent administered by a designated research nurse. All biological specimens collected for this study were de-identified following collection of demographic and clinical data.
Both Ae. aegypti and Ae. albopictus mosquitoes were obtained from a colony at the Duke-NUS Graduate Medical School. The colony was established in 2010 with specimens collected in Ang Mo Kio, Singapore, and infused monthly with field-collected mosquitoes to maintain genetic diversity. Female mosquitoes, three to five days old, were intrathoracically inoculated with 0.017 μl of serum from the Early Human sample (fever day 1–3) and the Late Human sample (taken four days after the initial sample) for 11 out of 12 patients as previously described [62]. Insufficient sera remained from one patient for the mosquito inoculations. Mosquitoes inoculated with DENV-1 clinical serum were incubated for 10 days, while mosquitoes inoculated with DENV-1 derived from pOEEic infectious clone were incubated for 5, 10 and 21 days respectively at 28°C and 80% humidity, with access to 10% sucrose and water. Surviving mosquitoes were killed by freezing and examined for the presence of viral antigens in head tissue by direct immunofluorescence assay (IFA). Infected mosquitoes were stored at -80°C until assayed. For each viral sample, including each time point of the infectious clone experiment, 10 infected mosquitoes were pooled and triturated with a pellet pestle (Sigma Aldrich, St. Louis, MO, USA) in 250 μl of 1x PBS. DENV RNA was extracted from the sample using TRIzol RNA isolation reagent (Invitrogen, Carlsbad, CA, USA) according to the manufacturer's protocol and stored at −80°C until use.
A pool of 10 inoculated mosquitoes was triturated in 250ul L-15 (Gibco, Life Technologies, Carlsbad, CA, USA) maintenance medium. Total RNA was extracted using TRIzol® (Life Technologies) according to the manufacturer’s protocol and stored at -80°C until use. Two separate reverse transcription reactions were carried out using the SuperScrip III First-Strand Synthesis System (Life Technologies) according to the manufacturer’s protocol (1) cDNA was synthesized with random hexamers for downstream amplification of fragments 1, 2, 3 and 4. (2) cDNA was synthesized with 10μM of reverse primer 10693R for downstream amplification of fragment 5. cDNA was amplified in 5 fragments by PCR (S16 File). Using the respective primers for each fragment, PCR was carried out using Phusion High-Fidelity PCR Master Mix with HF Buffer (Thermo Fisher Scientific, Waltham, MA, USA). 2μL of cDNA was mixed with 1μL of each primer (10μM), 25μL of 2X master mix and 21μL of water. The PCR conditions were: 30sec at 98C followed by 40 cycles of PCR at 98C for 10sec, 55C for fragment 1 or 57C for fragments 2,3,4 and 5 for 20sec, 72C for 2min and final extension at 72C for 10min. The PCR products were run on a 1.5% agarose gel. Bands of the correct size were excised and gel purified using Qiagen QiaQuick gel extraction kit (Qiagen, Valencia, CA, USA) according to the manufacturer’s protocol. NGS libraries were constructed according to the methods described in Aw et al. [63].
pOEEic was digested with SacII at 37°C for 2 h. Linearized DNA was purified using ultrapure Phenol: Chloroform: Isoamyl Alcohol (Invitrogen) according to the manufacturer’s protocol. In vitro transcription of the purified DNA was performed to generate full-length genomic DENV RNA using MEGAscript T7 kit (Ambion, Life Technologies, Carlsbad, CA, USA) according to the manufacturer’s protocol. The reaction was spiked with additional rATP after incubation at 37°C for 30 minutes, and further incubated at 37°C for 2 h. 5 μg of RNA was electroporated into approximately 5.0×106 BHK cells in a 4 mm cuvette using the Bio-Rad Gene Pulser II with the settings adjusted to 850 V, 25 μF. Each cuvette was subjected to 2 pulses at an interval of 3s. The cells were allowed to recover for 10 mins at 37°C and transferred into 15 ml of pre-warmed culture medium in a T75 flask. Cell culture supernatant was collected 5 days after infection and tested for the presence of infectious DENV-1 using plaque assay. The BHK cells were also scraped off and analyzed for the presence of DENV-1 envelope antigen by indirect fluorescent assay (IFA) using the anti-DENV-1 envelope primary mAb HB-47.
The data analysis pipeline used in this study is built upon open-source tools, which are freely available. The bulk of the sequencing analysis was done using the viral analysis pipeline ViPR (https://github.com/CSB5/vipr), which mainly handles mapping of reads and calling of SNVs with LoFreq [29]. For mapping of Illumina paired-end reads to the Sanger sequenced reference genomes we used BWA version 0.6.2-r126 for all Illumina and version 0.5.9 for all SOLiD sequencing datasets [64]. LoFreq (version 0.6.1) was used for SNV calling using default options and regions overlapping primer positions were ignored. Cold-spot analysis was performed as described in [29]. In brief, SNVs predictions from groups of samples are pooled and then scanned for SNV free regions that are larger than expected (binomial test; Bonferroni corrected p-value < 0.05). Reference genomes used for mapping and annotation were specific for each sample and come from the Sanger consensus sequences reported for viruses in Schreiber et al. 2009 [28].
An in-house R script was developed for calculations and visualization [65]. First, the numbers of reads across the entire sequence were extracted from pileup data files and used as default data values. Additional reads for were extracted from.snp format files for all positions where at least one non-consensus nucleotide was present, over-writing the pileup data for that position. Regions where primers bound to amplify the DENV1 genome, corresponding to positions 1–70, 2065–2084, 4221–4241, 6442–6461, 8519–8540, 10645–10735, were excluded from all analysis. As the typical number of reads is ~100,000, any difference between the proportions of non-consensus bases at different time points or in different hosts that is biologically significant is also statistically significant.
Selection pressure and genetic drift were measured by calculating the dN/dS ratio, which is the ratio of non-synonymous mutation changes per non-synonymous sites (dN) to the synonymous mutation changes per synonymous sites (dS) (S10 File). The Mann-Whitney test was then performed on the dN/dS of the samples between a particular protein coding sequence and compared to the rest of the protein coding sequences within an experimental condition (e.g. the C gene is compared to PrM, E, NS1, NS2a, NS2b, NS3, NS4a, NS4b, 2K and NS5 from the early human samples) (S11 File). The frequency of mutation type was measured by calculating the ratio of transitions (A ↔ G, C ↔ T) to transversions (A↔ C, G ↔ T, G ↔ C, A ↔ T) for each protein coding sequence (S12 File). Diversity and evenness across the viral polyprotein was calculated using the Shannon diversity index and Shannon equitability measurements respectively and the overall number of mutations across the polyprotein (either transitions or transversions) was calculated using a 100bp window (S13 File). We used the Mann-Whitney-Wilcoxon test to assess whether the two samples come from the same population (S14 File).
Plasticity at each position-individual-time point was determined by the proportion of the reads that did not agree with the consensus. Three thresholds (0%, 1% and 5%) of plasticity were recorded. For each time point, the number of individuals (out of 12 in human, 11 in mosquitoes) with plasticity above that threshold were counted to assess localized reproducibility across samples. If more than a third of the samples for that position-individual-time point exhibited plasticity (of >0%), this was declared significant: with 99.6% of positions agreeing with the consensus, this threshold leads to a p-value of < 10−7. Although this proportion of samples was arbitrarily defined, it nonetheless provides a conservative approach to differentiate stochastic from biologically important variances.
Differences between (logically comparable) pairs of time points were assessed, for each position-individual-pair, by counting differences, differences of at least 0%, 1%, or 5%, between the proportion of reads of the consensus (vs. all others) nucleotide. If more than a third of the samples for that position-individual exhibited a difference between the two time points (of >0%), this was declared significant: with 99.2% of positions not exhibiting differences within the two serum samples, this threshold leads to a p-value of < 10−5. Prevalence of differences in different genes were tested via χ2 tests between all gene pairs, with Bonferoni’s correction for multiple testing.
For triples of time points (e.g. human early, human late, mosquito late), we quantified pairs of differences for each position-individual combination. In this analysis, we looked for differences of at least 0.1% between the proportions of reads (of the consensus nucleotide at the first time point) for each pair of time points, and coded the sign of the differences. If both differences were positive, this position-individual was coded as changing towards consensus; if negative, as changing away from consensus; and if alternating (positive then negative or vice versa), as a reversion (if either difference were less than 0.1%, the position for that individual was ignored). The number of individuals with each change was recorded as a measure of reproducibility. Reversions with differences of >5% were stored and plotted individually. Consistent reversions across individuals (in more than 25% of samples) were declared significant: with 99.6% of positions in Ae. aegypti, and 99.5% in Ae. albopictus, not undergoing reversions, this leads to p<10−4. At this evidence threshold, if throughout the entire sequence reversions occurred by chance alone, we would expect to see 0.1 to 0.2 locations declared significant, and there is ~15% chance of seeing 1 false positive, and ~1% chance of seeing 2 or more.
In addition, we identified positions in which the consensus nucleotide itself changed between time points. To do this, we scanned over positions and noted any in which the dominant nucleotide for any two time points, for any individual, differed. Positions in the list for with several such switches were plotted.
|
10.1371/journal.pntd.0003876 | Development of a Novel Rabies Simulation Model for Application in a Non-endemic Environment | Domestic dog rabies is an endemic disease in large parts of the developing world and also epidemic in previously free regions. For example, it continues to spread in eastern Indonesia and currently threatens adjacent rabies-free regions with high densities of free-roaming dogs, including remote northern Australia. Mathematical and simulation disease models are useful tools to provide insights on the most effective control strategies and to inform policy decisions. Existing rabies models typically focus on long-term control programs in endemic countries. However, simulation models describing the dog rabies incursion scenario in regions where rabies is still exotic are lacking. We here describe such a stochastic, spatially explicit rabies simulation model that is based on individual dog information collected in two remote regions in northern Australia. Illustrative simulations produced plausible results with epidemic characteristics expected for rabies outbreaks in disease free regions (mean R0 1.7, epidemic peak 97 days post-incursion, vaccination as the most effective response strategy). Systematic sensitivity analysis identified that model outcomes were most sensitive to seven of the 30 model parameters tested. This model is suitable for exploring rabies spread and control before an incursion in populations of largely free-roaming dogs that live close together with their owners. It can be used for ad-hoc contingency or response planning prior to and shortly after incursion of dog rabies in previously free regions. One challenge that remains is model parameterisation, particularly how dogs’ roaming and contacts and biting behaviours change following a rabies incursion in a previously rabies free population.
| Rabies in domestic dog populations still causes >50,000 human deaths worldwide each year. While its eradication by vaccination of the reservoir population (dogs and wildlife) was successful in many parts of the world, it is still present in the developing world and continues to spread to new regions. Theoretical rabies models supporting control plans do exist for rabies endemic regions; however these models usually provide information for long-term programs. Here, we describe a novel rabies simulation model for application in rabies-free regions experiencing an incursion. The model simulates a rabies outbreak in the free-ranging dog population in remote indigenous communities in northern Australia. Vaccination, dog density reduction and dog confinement are implemented as control strategies. Model outputs suggest that the outbreak lasts for an average of 7 months and typically spreads through all communities of the region. Dog vaccination was found to be the most effective response strategy. The model produces plausible results and can be used to provide information for ad-hoc response planning before and shortly after rabies incursion.
| Rabies is among the most lethal infectious diseases, present on all populated continents except Australia [1]. The domestic dog remains the most important vector worldwide, causing >95% of all human cases [2–4]. Despite availability of an effective vaccine for more than a century and repeated demonstration that vaccinating the domestic dog population is the most effective way to eliminate the disease [5–8], rabies remains endemic in large areas in Africa and Asia. Moreover, the disease has (re)emerged in areas previously free (such as Bhutan [9,10], Indonesia [11,12], and the Central African Republic [13]).
Rabies continues to spread through the Indonesian archipelago via human mediated domestic dog movements [11,12,14], most recently through the previously rabies-free province of Maluku in eastern Indonesia [11,15]. The risk of incursion into rabies-free areas − Timor, Irian Jaya, Papua New Guinea (PNG) and northern Australia − is therefore high. Possible incursion scenarios into Australia include yachts or fishing boats hosting latently rabies infected dogs traveling from Indonesian islands to remote areas in northern Australia [16]. Also, close cultural ties between PNG and Torres Strait Island communities exist, increasing the risk of movements of dogs incubation rabies from PNG to Australia, if an incursion in PNG occurs [16]. In these regions large, free-roaming domestic dog populations [17,18] increase the risk of rabies establishment, which would subsequently impact human and wildlife populations. Because there are no historical precedents, the spread and final impact of such rabies incursions is difficult to estimate. However, such knowledge is critical to informing preparedness and response plans prior to an incursion, and to design the most effective strategies.
Descriptions and applications of several rabies models in wildlife [19–24], domestic dogs [5,7,25–28] or a combination of these [8,29] have been published. All have been based on empirical field data in rabies endemic regions and typically aim to inform policy on reducing rabies prevalence and thus impacts. However, for a region in which rabies is exotic, predictions of the effectiveness of different interventions following the initial detection of rabies are more relevant. An issue is how rabies behaves when introduced to a previously free population, particular the effect of rabies on contact rates and biting rates. Evidence on these behaviour changes from previous rabies incursion may serve as an approximation but is typically vague and therefore equivocal. To our knowledge, epidemic models simulating rabies invasion in regions never exposed to rabies do not exist, a barrier to rabies preparedness planning.
Here, we describe the development of a novel simulation model of rabies epidemics in domestic, free-roaming dog populations in remote Indigenous communities in northern Australia–as an example of the potential scenario in many regions of the world where rabies is absent but where the risk of a rabies incursion is present and its spread is likely due to large populations of free-roaming domestic dogs. Results of a systematic sensitivity analysis are also presented and model application options are discussed.
Data collection required to estimate model parameters has been approved by the Human Ethical Committee of the University of Sydney, grant no. 2013/757 and the Animal Ethical Committee of the University of Sydney, grant no. N00/7-2013/2/6015.
The rabies simulation model development was based on data from two distinct regions in northern Australia, the Northern Peninsula Area (NPA) of Cape York, Queensland and Elcho Island, the Top End of the Northern Territory. Characteristics of the two study sites are presented elsewhere [18]. Briefly, in the NPA five Aboriginal and Torres Strait Islander communities are located in close proximity (2−4 km) to each other. On Elcho Island one larger Aboriginal community is present. The dogs are typically owned but unrestrained and build a large population in all communities (human:dog ratio 2.7−8.8 per community, Table 1) [18]. The dog population in the NPA − which informs the simulation model − is based on the most recent dog census conducted by the NPA Regional Council in 2009. As such information was not available from Elcho Island, the number of dogs are calculated based on the average human:dog ratio of the NPA communities and official human census data from Elcho Island in 2011 (http://www.censusdata.abs.gov.au/census_services/getproduct/census/2011/quickstat/SSC30094) and similar household sizes as in the NPA are assumed.
The model developed is stochastic, spatially explicit, based on individual dog data and assumes a daily simulation time unit. It starts with the introduction of a latently (non-clinical) infected dog and ends when no infected dog remains. The exact location of each dog’s home is known and a closed dog population within the region is assumed, but dog movements between regional communities are simulated. In the model, 429 and 410 dogs in 175 and 163 households are included in the NPA and on Elcho Island, respectively (137−451 and 68−142 households per km2, respectively, Table 1). The two regions are simulated separately. The average number of dogs per dog holding household ranges from 1.9–3.2 per community.
Parameter value definition is a critical component in modelling studies, driving the outcome of any simulation or mathematical model. While some parameter values can be taken from the literature (e.g. disease or vaccine related parameters), other parameter depend on the settings in the specific environment in which the model was developed or applied. Seven out of 37 (19%) parameter values of the model presented here are sourced from the literature (those of rabies virus and vaccine related parameters) and 23 (62%) are based on assumptions (S1 Table). The latter can further be classified as experimental parameters (parameters defining control strategy implementation as e.g. delay in starting control strategies or vaccination coverage, 16/37 [43%]) and parameters for which value information are currently lacking (e.g. bite probability given a contact, owner compliance to cease dog movements, 7/37 [19%]).
The remaining seven (19%) parameter values were estimated based on our field collected data, including contact data within and between communities and mean distance between households used for dog confinement strategy (S1 Table). Data used to calculate the distance kernel function applied for contact rates between dogs of different households was derived from a large scale GPS study on 69 domestic dogs in all of the six communities (S5 and S6 Tables) [18]. The number of contacts between each pair of dogs within the same community was extracted using the definition of contact being within 20 meters during the same minute. As the model runs on a daily basis, the contact information was converted into a binary variable with two dogs having at least one (1) or no (0) contacts within 24 hours. This binary outcome was analysed by a logistic regression model with the known distance between the two dogs’ homes as the explanatory variable. The outcome variables estimated by the logistic regression (intercept α, coefficient β and standard error of the coefficient βse) were further used to build the distance kernel (Eqs 1 and 2). Daily contact probability of two dogs living in the same household was estimated based on the same dataset plus similar data collected during the post-wet season (monsoon) in the same communities. One of 31 (3%) pairs of dog living in the same household was not observed to have at least one contact per 24 hours. This within-household contact probability was implemented as a uniform distributed parameter with 97% as the mean.
Four parameters defining the dog movements between communities − both short term and permanent − were estimated from questionnaire data collected in the NPA (S2 Table, approved by the Human Ethical Committee of the University of Sydney, # 2013/757) together with observations of short term movements by GPS and of permanent movements of dog owners from one NPA community to another during a year. Twenty-nine dog owners were interviewed in September 2013 and one in September 2014, including questions on frequency of dog movements to other NPA communities due to pig hunting or other trips (e.g. visits or work). A daily movement probability of 0.058 per dog was calculated from these reported data, while from the study we observed that only 8 of 81 (10%) dogs were moved during an observation period of 6.2 days, resulting in daily movement probability of 0.016 per dog. Combining these reported and observed data and giving twice the weight to observed data, 0.03 was defined as the beta-pert distribution mode for daily short term movement probability per dog; 2- and 0.5-fold values were used for the minimum and maximum limits of the distribution, respectively. The duration of the short term movements were derived from the questionnaire in which all hunters reported trips of one to two days and observations from the GPS study in which all 8 dogs stayed less than one day in the community visited. The frequency of permanent movements was estimated from questionnaire data and observations of permanent movements during September 2013 and September 2014. Dog owners reported that 18% (6/33) of the NPA dogs originated from a different community within the NPA, which resulted in an estimated probability of permanent movements of 1.64*10−4 per dog per day assuming an average dog life of three years. In addition, owners of 6% (3/49) of the dogs were observed to have permanently moved between NPA communities during the year, resulting in a probability of daily movements of 1.64*10−4 per dog. The sum (3.3*10−4) was selected as the mode of the beta-pert distributed daily probability of permanent movements per dog, with 0.5 and 2-fold values for minimum and maximum. Finally, the destination community for both permanent and short term movements was observed to be more frequently a neighbouring community than any other; consequently a neighbouring community as the destination was assumed to be twice as likely as for any other community in the NPA.
The median distance between each household and its closest neighbour–used in the model to truncate the distance kernel for the dog confinement control strategy–were estimated for each community separately using the coordinates of all households per community. Household coordinates were derived from Google Earth (http://earth.google.com/) where placemarks were set on all private dwellings, identified with the help of community maps. The distances between each household and its closest neighbour were calculated and the mean per community implemented in the model as a fixed value parameter.
In the first step of the sensitivity analysis (SA), all model parameters used for six different control strategies were tested using the strategy’s default values: a) vaccination with 70% coverage either pre-emptively (PV) or reactive (RV); b) culling of dogs contacted by a rabid dog (CC) or reactive culling (RC) with culling levels of 80 and 50%, respectively; c) dog confinement plus movement bans between communities (MB) with 80% and 90% compliance, respectively; and d) a non-intervention strategy (NI). For all of these strategies, including NI, culling of dogs detected rabid (DC) was applied. The number of index dogs was set to 1 and randomly selected in the region. For each of the 12 combinations of the two regions (NPA and Elcho Island) and six control strategies, 1000 model repetitions were simulated.
For stochastic parameters, the mode or mean (for beta-pert distributed and uniform distributed parameters, respectively) was allowed to range between ±25% around the default value while the difference between the minimum and maximum values was kept fixed (no variation of the distributions’ shape; S3 Table, S1 Fig). Deterministic parameters were allowed to vary ±25% around the default value. Variation of 25% has been chosen to allow enough variation for parameters with wide distributions (e.g. the infectious period) and avoid too large distinction between the lower and upper limit of narrow distributed parameters (e.g. the rabies transmission probability given a bite). The distance kernel was tested using three different shapes representing a minimal kernel and increased probabilities for short and long distance contacts, respectively (S2 Fig). The influence of the index community within the NPA region was investigated by defining one of the five communities hosting the index dog. To explore the sensitivity of the model on the weighting matrix to choose the destination community for between-community movements, an alternative matrix was tested beside the default with equal chance for all communities to be selected as the destination.
The values of all parameters tested were randomly selected from the described ranges so that for each simulation, an individual set of parameter value combination was chosen. For each of the 12 region-strategy combinations, linear multivariable regression analysis was modelled with the outbreak duration and–where applicable–outbreak size as the response variable and the parameter values as explanatory variables. For the stochastic parameters the mode (beta-pert distribution) and mean (uniform distribution) values were modelled as explanatory variable values. Correlations between the parameter values were explored using Kendall’s tau correlation, Chi-Square and Wilcoxon Rank Sum test for two continuous, two categorical and a continuous and categorical parameter, respectively. Because the assumption of a normally distributed response variable was not always met (S3 Fig), logistic regression following the same principle was modelled defining an outbreak with a duration or size above the median as 1 and as 0 otherwise. Based on both the linear and logistic regression analyses, parameters were defined to be highly (statistically significant p-values < 0.05 in ≥ ¾ of all tested regression models), low (statistically significant p-values < 0.05 in < ¼ of all tested regression models) or moderate (otherwise) sensitive to the outbreak duration and size. Additionally, scatter plots of the outbreak duration and size over the range of each parameter value were visually analysed and correlations were calculated between outbreak duration and size and parameter values with continuous scale. A correlation of >|0.1| was considered as a threshold to distinguish between sensitive and non-sensitive parameters.
Parameters found to be highly sensitive in either of the regression analyses during the first step of the SA were further explored in a second step to identify the influence of their mode or mean (default, large, small) and shape (default, narrow, wide) on the model’s outcome. For the beta-pert and uniform distributed parameters nine combinations per parameter with the three values of mode or mean (default and ±10%), and three values of difference to the minimum and maximum (default and ±10%) of the distribution were defined (S4A–S4C Fig). For the vaccine efficacy parameter, the variation of 10% had to be reduced to 4%, which was the highest variability still ensuring a maximal value < 1, a requirement for probabilities (S4D Fig). For each parameter, 1000 repetitions were simulated for the same 12 region-strategy combinations described in step 1 of the SA (in case of the vaccine efficacy only the scenarios for RV and PV), where one of the nine options was randomly chosen while all other parameters in the model were kept at their default value. The distance kernel, defined by the three variables α, β and βse, was explored by varying the three variables around a default value ±50%, resulting in 27 combinations (S5 Fig). Six thousand repetitions were simulated for all 12 region-strategy combinations with a randomly selected distance kernel out of the 27 options while all other model parameters were kept at their default values. The outcome was analysed visually comparing boxplots of the outbreak duration and number of rabid dogs.
A critical question for stochastic models is always, how many repetitions are required to sufficiently reflect the variability of the model? The coefficient of variation (CV = standard deviation/mean) of model outputs’ mean has been proposed as a measurement to determine the critical number of repetitions required [30,31]. The CV of the estimated mean of outcome of interest (e.g. outbreak size or duration) over n model simulations is expected to approach 0 for infinite sample sizes n and a threshold of the CV of 15% has proposed to predict outputs with acceptable precision [30]. We used the same approach, but reduced the CV threshold value to 3%.
To demonstrate the model’s functionality for the different control strategies, 1000 outbreaks were simulated for the default strategies: 1. reactive vaccination with 70% coverage at the household level (RV); 2. reactive culling with 50% of the dogs culled in affected communities (RC); and 3. dog confinement between and within communities with 90 and 80% compliance, respectively (MB). Culling of dogs detected rabid (DC) was applied for all of these strategies. The model was simulated in both regions, NPA and Elcho Island, separately. As outcome measures, the epidemic duration and size, i.e. the number of rabid dog and the number of dead dogs (including rabid and culled dogs) were calculated and visually compared between the different scenarios.
Two measures of outbreak size are the cumulative number of rabid dogs and cumulative number of dead dogs (due to rabies plus culled). The outbreak duration is defined as the number of days from the introduction of the latently infected index dog until the death of the last infectious dog. The simulation model resulted in plausible results comparing outputs for the three different default control strategies (Fig 4). Rabies spreads through the communities in a wave pattern and, depending on the control strategy, can kill the entire dog population (S6 Fig illustrates an example epidemic curve). The reactive culling (RC) strategy reduces the number of rabid dogs; however the number of dead dogs is only slightly less than the total dog population. For the reactive vaccination (RV) strategy the number of rabid dogs (equal to the number of dead dogs) showed higher variability among the 1000 model simulations compared to RC, but in both regions, the median of RV was lower than for RC. Obviously, vaccination saves the dogs from death in contrast to culling strategies. The movement ban (MB) strategy showed a slight decrease of the outbreak size in the region of Elcho Island whereas no effect was observed in the NPA. However, it was found to be the strategy with the longest durations of outbreaks, demonstrating that movement bans (if not 100% compliant) only slow the speed of spread rather than reducing its size. The reduction in the number of movements between communities for MB was obvious, decreasing from a median of 52 (RV) and 46 (RC) to 19 (S7 Fig). Overall, outbreak duration ranged from 1−20 months (median 6.7 months) and was more homogenous between interventions than the outbreak size (Fig 4). Outbreaks lasting for one month did not spread beyond the index dog. For the RV strategy, the vaccination coverage was set at 70% of the households, producing dog level vaccination coverage of 59−75% and 56−76% for the NPA and Elcho Island region, respectively.
The control strategy with the largest number of simulations required (n = 490) to capture the variability in the model’s output with the defined CV threshold of 3% was found to be the MB strategy in the NPA (S8 Fig).
The number of secondary cases was reported for every rabid dog over the duration of the outbreak. From these records, the basic reproductive ratio R0 was calculated and defined as the mean number of secondary cases for dogs becoming infectious within the first phase–i.e. up to its peak–of an epidemic. The peak of the epidemic is defined as the day with the highest number of newly infectious animals over the entire outbreak. R0 ranged from 0−6.1 (median 1.8) for RV, 0−6.1 (median 1.7) for RC and 0−5.7 (median 1.7) for MB (S9 Fig) with an overall median of 1.7 (Fig 5A). The epidemic peak was reached on average after 93 days (Fig 5B) with a mean of 17 newly infected dogs (Fig 5C). The number of secondary cases derived from each index dog was highly variable and ranged from 0 to 79 (median of 25) for NPA and 4 to 106 (40) for Elcho Island. Over the duration of the outbreak, the effective reproductive ratio Rt and the number of dogs in the susceptible population decreased in a wave pattern (S10 Fig). The value of 1 for mean Rt is reached during the second or third wave. This reflects that Rt depends on the dogs remaining in the population and finally the outbreak dies out because there are no susceptible dogs left that are close enough to the infectious dogs.
The simulation model outputs were highly sensitive to seven parameters: incubation period (G1 in S4 Table), transmission probability given a bite (G8), distance kernel (G5), bite probability given a contact between dogs of different households (G7), vaccine efficacy (V5), index community (G12) and delay in starting the control strategy of movement restrictions between communities (B2). The same sensitive parameters, in addition to the detection delay of the first clinical case, were also identified via correlation tests, with the exception of B2. These outcomes were also confirmed by scatterplots, which express particularly dependencies between the outbreak duration and the incubation period, distance kernel and index community (S11 Fig). Significant correlations between parameters included in the regression analyses were only observed between categorical and continuous parameters where 2.1−12.1% (mean 5.4%) of all parameter combinations resulted in Wilcoxon Rank Sum test p-values <0.05.
The influence of this subset of parameters was further explored in step 2 of the SA, with the exception of the index community and the delay in commencing movement restrictions because these two parameters directly relate to incursion and intervention scenarios. For all parameters, except the distance kernel, it was found that both the mode and mean (for beta-pert and uniform distributions, respectively) and the shapes influence outbreak duration and number of rabid dogs (S12 Fig). The mean and mode were found to have a greater impact, particularly for the incubation period and rabies transmission probability. For the distance kernel, the regression coefficient β was most influential on both the number of rabid dogs and the outbreak duration, followed by βse (standard error of β), particularly for the Elcho Island region (S13 Fig). Outputs were less sensitive to the intercept α.
The model described herein provides insights into short-term rabies epidemics occurring within a small spatial extent in previously rabies-free regions. This is of crucial value for contingency planning in areas where rabies is exotic and the model fills a gap in the published literature on rabies models. The example of Bali, Indonesia demonstrates the impact of a rabies incursion on an under-prepared region [12]. Late detection of the disease, lack of surveillance strategies and an unsuccessful initial response (focused on dog culling) resulted in island-wide disease spread and high impacts on both dog and human populations [7,12,32]. Another example in Indonesia–comparable to communities in northern Australia with a high density of free-roaming dogs and limited veterinary health services − is the island of Flores in East Nusa Tengarra province [14,33,34]. There, one or more latently infected dogs were introduced, developed clinical rabies and transmitted the disease to local dogs. The disease consequently spread throughout island with a considerable impact on dogs and humans. This is another example of a rabies invasion in a new area with very severe impact.
Another novel aspect of the current model is the inclusion of individual susceptible and rabid dogs modelled within a continuous spatial dimension, an approach previously used to simulate highly infectious diseases of livestock (e.g. foot-and-mouth-disease [35,36]) but not rabies. To date, published dog rabies models have been based on mathematical differential equations [5,24,25,27] or spatially explicit models simulating the spread of rabies within grids [7]. Our approach has several advantages, including stochasticity to capture epidemic variability, incorporation of detailed population structure to better represent real target regions of interest, and detailed spatio-temporal model outputs.
By simulating three default control measures–vaccination, culling and dog confinement–our model produced plausible results, suggesting it adequately captures how an epidemic of an infectious disease with a relatively long incubation period would develop in a previously uninfected population. The disease spread temporally in wave-like patterns, peaking on average about three months after an incursion. The average R0 that was estimated from the model outputs was 1.7, consistent with previous estimates of rabies spread in endemic countries [16,37] and estimated from the Bali outbreak [7].
For the calculation of R0 we considered rabid dogs up to the peak of the epidemic, while peak has been defined as the day during the epidemic with the highest number of newly rabies infectious animals. It has been demonstrated that both, clustering within the susceptible population and high number of repetitive contacts among individuals in the population–thus a non-random mixing situation–affects the dynamics of disease spread via depletion of susceptible population within a cluster and therefore also affects R0 [38]. A re-evaluation of R0 of a H1N1 infection revealed that it was overestimated during the early stage of the outbreak when only cases within a cluster has been considered rather than the population-wide epidemic [39]. In our model, random mixing of infectious and susceptible animals has to be rejected as rabies transmission depends on the distance between infectious and susceptible animals. However, neither fixed repetitive contacts (as in network based models) nor significant clusters within communities are included in the model. The only functionally relevant cluster structure is present in the region of the NPA with the five communities building distinct clusters. We respected the cluster structure in the R0 calculation by not only considering cases to the peak within the first cluster, i.e. the community of the index dog, but all cases occurring before to the epidemic peak within the total population at risk in the respective region.
As illustrative examples we simulated the most often discussed and applied control strategies for dog rabies, namely vaccination and culling of dogs as a reactive action, as well as dog movement restrictions. According to the model, the only beneficial measure (based on outbreak size) is vaccination. This is consistent with a range of studies in regions where a rabies incursion has been observed: culling as a single control measure was unsuccessful [7,12,14,33], whereas vaccination was demonstrated to be a successful strategy to control recent rabies invasions [7,10,32,40]. However, success of vaccination campaigns obviously depends on the vaccination coverage, as the example of a unsuccessful rabies control via low level vaccination coverage demonstrated in Flores [34]. Also, culling can lead to an eradication of timely detected outbreaks, as for example in region in Bhutan [9], however might be impractical because of non-acceptance, depending on culture and religion of the community. In Australian Indigenous communities, culling of dogs is unlikely to be culturally acceptable. According to our model, movement bans as a single strategy does not seem to be sufficient to reduce rabies spread from one community to another nor within a community–at least for the simulated dog owner compliance (80–90%) that was simulated here. Movement bans would culturally also be difficult to implement as travel with dogs between communities and regions is common. These results are similar for both regions, the NPA and Elcho Island. Within the NPA, rabies epidemics were able to sustain after incursion in each the five communities, as it does for the one community on Elcho Island, identifying the here considered regions and communities equally susceptible. Targeting surveillance should therefore be based on information revealed by risk assessment pathways exploring high risk regions for a rabies incursion. Further detailed simulations exploring combinations of response measures and their threshold values for effectiveness–including the effect of surveillance intensity [40]–is warranted as future research.
The non-intervention strategy was implemented in the model and applied during the sensitivity analysis as a “baseline” to quantify the benefit of other control strategies, although it will most certainly never be observed in the field. Keeping in mind the assumption of a closed dog population with no influx of new dogs (neither birth nor immigration), the high densities of dogs that roam freely in the modelled community and the assumption of an infectious period of up to 12 days combined with a fully naïve population, it is expected that the epidemic will eventually kill the modelled dog population, in the case of no intervention applied.
Model outputs were sensitive to the assumed distance kernel for rabies transmission between dogs from different households. This highlights that the distance kernels should be empirically developed for the particular regions in which the model is to be applied. In our case, the kernel was estimated from roaming dog data collected within the actual study regions [18]. Model outputs (size and duration of the epidemics) were also sensitive to the assumed rabies incubation period and probability of rabies transmission given a bite. These parameters were derived from previously published field observations in Africa (S1 Table), and are likely applicable to many situations, as is the vaccine efficacy parameter. For the probability of bite given a contact between dogs, no specific published data could be found. We assumed this parameter to be >50% due to the aggressiveness that rabies can cause [41], but we also included a large range of uncertainty (60–80%).
Critical parameter values (as for contact rates and the population structure) in this model are informed by data collected in the field. This guarantees the fit of the model to the intended environment of application, although limitations still do occur. Data informing the distance kernel were based on the roaming behaviour of healthy dogs [18]. A rabid dog might change its normal roaming behaviour, as for example reported for rabid racoons in New Jersey (USA) which moved over significantly larger distances than healthy racoons [42]. Apparently the effect of rabies on the roaming behaviour of domestic dogs has not been reported, but considering the observed changes in racoon behaviour, our model might underestimate the spread of rabies. When comparing contact distances observed in our study (healthy dogs, mean distance = 103 meters) that inform this model with the published spatial infection kernel of contacts between rabid dogs and resulting infection (mean = 880 meters, [37]), rabid dogs might roam up to 8.5 times further than healthy dogs. This again indicates that in our model, although long distance movements have been included otherwise, transmission events might be underestimated. Studies on the nature of roaming, contact rates and biting rates of dogs in rabies-endemic studies–comparing these between infected and apparently uninfected communities in the same environment–are needed to address this gap in our knowledge and to more realistically parameterise models of rabies incursion.
In the model, we have focused on the domestic dog population and ignored possible spread to wildlife (in this region, wild dogs and dingoes) because we were interested in the initial epidemic behaviour of a rabies incursion − and hence its impact on domestic dog health and by implication, human health. We assumed a closed dog population without introductions, births and natural deaths, again because our focus was on exploring initial disease response actions rather than the design of long-term rabies control programs.
Although a critical issue for simulation model development, we have not yet validated the model using real outbreak data. Since Australia is historically rabies-free, a direct validation of the model in the region where it is intended for use is impossible, until an actual rabies incursion occurs. However, outbreak data from regions with similar dog population characteristics and recent rabies epidemics might be used, for example the island of Bali [12] or parts of Bhutan [9,10]. Validation of the model using data from a rabies outbreak in domestic dogs in Bhutan in 2008 is currently being planned.
Global control of rabies is a declared goal of the World Health Organization and the Global Alliance for Rabies Control [43]. Within endemic, mostly developing countries, knowledge of effective control strategies is advanced and the challenge is to transfer this knowledge into successful actions [44]. Effectiveness of control strategies to prevent rabies from establishing in previously free regions–as is currently occurring in several areas with large free-roaming dog populations and limited public health services–has not received the same attention. The model described here is a tool to generate such information for remote northern Australia; and it is flexible enough to be adapted to other regions.
|
10.1371/journal.pgen.1006039 | The Splicing Efficiency of Activating HRAS Mutations Can Determine Costello Syndrome Phenotype and Frequency in Cancer | Costello syndrome (CS) may be caused by activating mutations in codon 12/13 of the HRAS proto-oncogene. HRAS p.Gly12Val mutations have the highest transforming activity, are very frequent in cancers, but very rare in CS, where they are reported to cause a severe, early lethal, phenotype. We identified an unusual, new germline p.Gly12Val mutation, c.35_36GC>TG, in a 12-year-old boy with attenuated CS. Analysis of his HRAS cDNA showed high levels of exon 2 skipping. Using wild type and mutant HRAS minigenes, we confirmed that c.35_36GC>TG results in exon 2 skipping by simultaneously disrupting the function of a critical Exonic Splicing Enhancer (ESE) and creation of an Exonic Splicing Silencer (ESS). We show that this vulnerability of HRAS exon 2 is caused by a weak 3’ splice site, which makes exon 2 inclusion dependent on binding of splicing stimulatory proteins, like SRSF2, to the critical ESE. Because the majority of cancer- and CS- causing mutations are located here, they affect splicing differently. Therefore, our results also demonstrate that the phenotype in CS and somatic cancers is not only determined by the different transforming potentials of mutant HRAS proteins, but also by the efficiency of exon 2 inclusion resulting from the different HRAS mutations. Finally, we show that a splice switching oligonucleotide (SSO) that blocks access to the critical ESE causes exon 2 skipping and halts proliferation of cancer cells. This unravels a potential for development of new anti-cancer therapies based on SSO-mediated HRAS exon 2 skipping.
| HRAS was the first human proto-oncogene reported and p.Gly12Val in codon 12 the first oncogenic mutation described. Somatic mutations in HRAS are important drivers in cancer, and germline mutations cause Costello syndrome. Until now, it has been believed that the severity of mutations located in the HRAS codon 12/13 mutational hot spot sequence is exclusively determined by the activity of the encoded proteins. Here we show that mutations in exon 2 of the HRAS proto-oncogene can have a previously unrecognized effect on splicing efficiency and thereby determine HRAS activity. We report a patient with Costello syndrome (CS), who despite having the classical, severe, oncogenic p.Gly12Val mutation in HRAS has a mild clinical phenotype. We show that this is due to his specific sequence change, c.35_36GC>TG, encoding p.Gly12Val, which disrupts an ESE fundamental for efficient splicing of exon 2. We have explored this in detail and were able to show that HRAS exon 2 is a weak exon, which is dependent on the balance between positive and negative splicing regulatory factors, like SRSF2 and hnRNPF/H in order to be properly included. We show that different mutations in the mutational hot spot in HRAS codon 12 and 13 (c.34-39) affect splicing of HRAS differently, suggesting that this mechanism may also influence their occurrence in CS and their oncogenic potential in somatic cancers. Finally, we show that blocking access to the fundamental ESE, located in HRAS c.34-39, using a splice switching oligonucleotide (SSOs), causes exon 2 skipping, abolishes production of functional HRAS protein and halts proliferation of cancer cells. This shows a previously unknown weakness of the HRAS proto-oncogene, namely that exon 2 is weakly defined and therefore is a suitable target for SSO-based therapy, thereby pointing towards a potential new direction for therapeutic targeting of RAS in anti-cancer treatment.
| The Harvey rat sarcoma viral proto-oncogene homolog (HRAS) was the first human proto-oncogene to be identified [1]. The HRAS protein is a GTPase, which mediates signal transduction from growth factor receptors important for cellular proliferation, growth and survival. Somatic mutations in HRAS are present in many cancers and the vast majority of mutations affect codons 12 and 13 (c.34-39) leading to a constitutively active protein. The c.35G>T mutation (p.Gly12Val) in HRAS was the first mutation in a proto-oncogene that was implicated in cancer [2,3] and it is the second most frequently reported HRAS mutation in human cancers (Cosmic database: http://cancer.sanger.ac.uk/cosmic/gene/overview?ln=HRAS). The p.Gly12Val mutant has the lowest GTPase activity [4] and highest transformation potential among HRAS mutants [5,6].
Germline mutations in HRAS cause Costello syndrome (CS) (MIM: 218040), which is a congenital disease characterized by postnatal growth retardation, short stature, tumor predisposition, developmental delay, and abnormalities of the heart (cardiomyopathy), skin and skeletal muscles [7].
The vast majority (75%) of CS is caused by heterozygous c.34G>A (p.Gly12Ser) activating mutations in HRAS [8]. Heterozygous c.35G>C (p.Gly12Ala) and c.37G>T (p.Gly13Cys) HRAS mutations are also frequent in CS patients, making up 10% and 7% of alleles, respectively [8]. Whereas these mutations result in a relatively homogenous phenotype, a particularly severe, early lethal form of CS has been observed in a few patients with the less frequently observed p.Gly12Val (encoded by c.35G>T, c.35_36GC>TT or c.35_36GC>TA), p.Gly12Asp (c.35G>A) or p.Gly12Cys (c.34G>T) mutations [9–12]. These severe CS phenotypes are consistent with the higher transforming potential of the p.Gly12Val, p.Gly12Asp and p.Gly12Cys mutant proteins [5] and are also reflected in the higher frequencies of these mutations in cancer (Cosmic).
So far it has therefore been believed that the relative frequency of the different HRAS mutations in cancer and CS simply reflects differences in the oncogenic potential of the encoded proteins and differences in the rate they occur by spontaneous mutations. Recently, it was, however, demonstrated that a phenomenon termed “selfish selection” may be an important factor determining the mutation spectrum observed in CS. Selfish selection both offers an explanation for the puzzling fact that CS has an extreme paternal bias in origin and occurs with a frequency two—three orders of magnitude higher than expected from the background mutation rate [8]. This is explainable by the fact that HRAS mutations offer different selection advantages in male germ cells depending on their oncogenic potential, and that their proportion increases with paternal age dependent on the selective advantage.
Exonic mutations may, however, also have other effects than the expected changes in the protein, which can be predicted based on the resulting change from one codon to another according to the genetic code. A so called “splicing code” [13], which predicts that exonic mutations may have effects by impacting splicing regulatory elements, is beginning to emerge as an important cause of human disease, including cancer. This was elegantly demonstrated in a study by Supek and co-workers, who showed that synonymous mutations frequently act as driver mutations in cancer by altering exonic splicing regulatory motifs [14]. In particular, it was demonstrated that cancer-associated synonymous mutations frequently create exonic splicing enhancers (ESEs) and destroy exonic splicing silencers (ESSs) which regulate oncogene splicing in tumors. Additionally, expression and activity of splicing regulatory factors, like the SR-proteins and the hnRNP proteins are often dysregulated in cancer leading to aberrant splicing of oncogenes and tumor suppressor genes. A prominent example is the SRSF2 splicing regulatory factor, which is frequently mutated to a dominant negative form in myelogenic diseases [15]. Other important splicing regulatory factors such as SRSF1, SRSF3 and SRSF6 have been described as potent oncogenes [16–18], underscoring the importance of dysregulated splicing in cancer.
Here we demonstrate for the first time that HRAS exon 2 is a vulnerable exon, which is dependent on binding of splicing regulatory proteins to ESEs in order to be correctly included in the HRAS mRNA. Importantly, we show that a mutation in the codon for Glycine 12 (c.34-39) can abrogate formation of constitutively active p.Gly12Val HRAS in a CS patient by mediating pronounced exon 2 skipping. Moreover, we show that this vulnerability of HRAS exon 2 splicing can be exploited by employing splice switching oligonucleotides (SSOs) to induce HRAS exon 2 skipping. This provides proof of principle for a new mechanism for knocking out oncogenic HRAS, which may be used therapeutically to treat cancer.
We investigated a 12-year-old boy with an attenuated CS phenotype (see online methods). Sequence analysis of his DNA from multiple tissues showed a heterozygous c.35_36GC>TG germline mutation that was predicted to result in p.Gly12Val (Fig 1) There was no evidence of mosaicism (S1 Fig). Presently only a few individuals with CS with a p.Gly12Val mutation (c.35G>T, c.35_36GC>TT, c.35_36GC>TA) have been reported and all had a very severe clinical presentation, with death typically within the first months of life [9,10].
Because mosaicism did not explain the mild clinical presentation of the severe p.Gly12Val mutation in this boy, RNA from lymphocytes was examined and showed extensive exon 2 skipping, and showed that the c.35_36GC>TG mutation was nearly absent in correctly spliced HRAS mRNA (Fig 1). Low levels of exon 2 skipping were also observed in cells from controls and individuals with CS heterozygous for other HRAS mutations, indicating that exon 2 may be difficult to splice efficiently. Consistent with this, different human tissues also show low levels of HRAS exon 2 skipping and the amount varies between tissues, possibly reflecting tissue specific differences in the splicing regulatory factors which mediate exon 2 inclusion (S2 Fig). An mRNA without exon 2 will not produce a functional protein as the switch domains, switch I (amino acids 32–38) and switch II (amino acids 59–67), are fundamental for RAS-GTP/GDP binding, and thus for biological function. Exon 2 (amino acids 1–37) encodes a major part of the switch I domain. In particular threonine 35, which binds the terminal phosphate (γ-phosphate) of GTP in the active site, is encoded by exon 2. Therefore, if a protein were to be produced from an mRNA lacking exon 2 it would not be functional. Additionally, the normal ATG start codon is located in exon 2. A potential alternative in-frame ATG start codon is located in exon 3, but it’s use would produce a protein with a deletion of 66 amino acids from the amino terminal end, and thereby exclude both switch I and switch II. Consequently, loss of HRAS activity due to exon 2 skipping from the c.35_36GC>TG mutation can explain the attenuated phenotype in the individual with CS.
Since individuals with CS with other HRAS mutations, c.35G>T, c.35_36GC>TT and c.35_36GC>TA encoding the p.Gly12Val mutant protein have suffered from a very severe CS phenotype, we investigated the effect of these mutations on HRAS exon 2 splicing using a HRAS minigene. Transfection of the wild type and mutant minigenes into HepG2 cells confirmed that the c.35_36GC>TG mutation by itself causes high levels of exon 2 skipping, whereas the c.35_36GC>TT and c.35_36GC>TA p.Gly12Val mutations cause a low or modest increase in exon 2 skipping, respectively (Fig 2). The severe p.Gly12Val, c.35G>T mutation did not increase exon 2 skipping. Taken together, these data show that different nucleotide changes in codon 12 (c.34-36) of HRAS exon 2 can regulate HRAS activity by affecting the efficiency of HRAS exon 2 inclusion into mRNA during pre-mRNA splicing. In their examination of the spontaneous mutation rate and selective advantage of HRAS mutations in the paternal germline, Giannoulatou and co-workers [8] did not observe the c.35_36GC>TG mutation, whereas the c.35G>T, c.35_36GC>TT and c.35_36GC>TA p.Gly12Val mutations were observed. This is consistent with their hypothesis of selective advantage contributing to the observed abnormally high mutation rates in sperm, since the c.35_36GC>TG mutation would be selected against due to its deleterious effect on exon 2 inclusion.
Pre-mRNA splicing of constitutive exons with weakly defined splice sites is dependent on a delicate balance between exonic splicing enhancers (ESE) and exonic splicing silencers (ESS) [13,19]. ESEs bind positive splicing factors, typified by the serine/arginine-rich (SR) proteins [20]. In contrast, ESSs bind proteins from the heterogeneous nuclear ribonucleoprotein (hnRNP) family [21], which inhibits splicing. Consequently, a mutation which either disrupts/weakens a binding motif in an ESE or creates/strengthens an ESS can result in exon skipping if the exon is weakly defined. In silico analysis showed that HRAS exon 2 is weakly defined due to a weak 3’ splice site with a non-consensus G nucleotide at position -5 in intron 2 and a GGG triplet at positions -14 to -16 disrupting the polypyrimidine tract (Fig 3). A short polypyrimidine tract is difficult to recognize for the U2AF65 splicing factor, and furthermore GGG triplets can bind hnRNPF/H family proteins, which could compete with U2AF65 binding and thereby decrease 3’-splice site splicing efficiency [22]. We hypothesized that this weak 3’ splice site makes inclusion of HRAS exon 2 dependent on the binding of splicing regulatory proteins to ESEs and that this is disrupted by the c.35_36GC>TG mutation.
In order to make exon 2 recognition independent of the ESE/ESS balance we strengthened the weak 3’ splice site by replacement of the non-consensus G and the GGG triplet in the polypyrimidine tract with consensus T nucleotides (Fig 3). This improved splicing from the wild type HRAS minigene and abolished exon 2 skipping from the c.35_36GC>TG mutant. Interestingly, c.35_36GC>TG creates a known splicing inhibitory motif (GTGGGTG) (S3 Fig), which binds proteins from the hnRNPF/H family. There are several examples from other genes where single nucleotide substitutions have created or abolished this motif in a weak constitutive exon causing aberrant splicing and disease [22–27] (S3 Fig). This suggests that the c.35_36GC>TG mutation creates an hnRNPF/H binding ESS, which inhibits inclusion of exon 2. Consistent with this hypothesis, replacement in the HRAS minigene of the wild type sequence c.34-39 with known hnRNPF/H binding ESS motifs results in exon 2 skipping (S4 Fig). These motifs have previously been demonstrated to cause exon skipping in other genes [26,28]. Moreover, introduction of a single c.36C>G mutation, creating the hnRNPF/H (DGGGD) binding motif [29] also causes complete exon 2 skipping (S4 and S5 Figs) underscoring the importance of c.36G in disruption of splicing. However, both introduction of CC at position c.37-38 (S4 Fig) and deletion of nucleotides c.32-37 (Fig 3) in the wild type cause exon 2 skipping, indicating that a fundamental ESE is also present in this region of wild type HRAS exon 2.
We used two different splicing reporter minigenes [30,31] to demonstrate that this part of HRAS exon 2 harbors ESEs, which can drive splicing in other genomic contexts, and that exon inclusion is abolished by the c.35_36GC>TG mutation (S5 Fig). Interestingly, testing of the c.35G>T mutant sequence indicated that it results in more efficient splicing than the wild type sequence. This is consistent with the results from the HRAS minigene and indicates that this mutation may improve splicing. ESE finder analysis [32] also suggests that the region around c.35 harbors potential binding sites for the SRSF1 and SRSF2 splicing stimulatory proteins, which usually bind ESEs to stimulate splicing. The c.35_36GC>TG mutation directly abolishes SRSF1 motifs, but does not directly affect the SRSF2 motif, whereas a deletion of nucleotides c.32-37 disrupts both the SRSF1 and SRSF2 motifs (Fig 3). RNA affinity purification employing wild type and c.35_36GC>TG mutant RNA oligonucleotides combined with ITRAQ labeling followed by MS/MS analysis indicated that the c.35_36GC>TG mutation increases binding of hnRNP F/H proteins and decreases binding of SRSF2, but not SRSF1 and this could be demonstrated by western blot analysis (Fig 4). Binding of hnRNPF/H proteins may also disrupt binding of SRSF2 and other splicing stimulatory proteins to an overlapping ESE.
Since hnRNPF/H binding to GGG triplets in a pre-mRNA is cooperative and synergistic [29], mutations creating new GGG triplets in HRAS exon 2 are likely to inhibit splicing by acting in synergy with pre-existing GGG triplets, such as the flanking GGG triplets and the GGG triplet in the weak 3’-splice site.
Interestingly, expression of hnRNPF/H proteins is low in cardiomyocytes [33], suggesting that inclusion of c.35_36GC>TG mutant exon 2 could be high in the heart from our patient.
Taken together our data suggest that c.35_36GC>TG simultaneously disrupts an ESE and creates a strong hnRNPF/H binding ESS (S6 Fig). Consistent with this, siRNA mediated knock down of SRSF2 caused exon 2 skipping both from the wild type HRAS minigene and from endogenous HRAS in T24 and HepG2 cells (Figs 4 and S7), whereas SRSF1 knockdown had no effect on HRAS exon 2 inclusion (S8 Fig). This does of course not exclude that other splicing regulatory factors may also bind to the HRAS ESE and stimulate exon 2 inclusion.
To further substantiate that HRAS exon 2 skipping leads to inactivation of HRAS and that an ESE fundamental for HRAS exon 2 inclusion is present in the region harboring c.35G, we designed a splice switching oligonucleotide (SSO) that would block binding of splicing regulatory proteins to the ESE. Consistent with this proposed effect, the SSO caused exon 2 skipping from the wild type HRAS minigene. Interestingly, the effect of the SSO was alleviated when the 3’-splice site was strengthened in the minigene (Fig 5). This substantiates that vulnerability of exon 2 is determined by the weak 3’-splice site.
Next we demonstrated that the SSO causes exon 2 skipping from the endogenous HRAS gene in both T24 and HepG2 cells. This was reflected in reduced levels of HRAS protein and by decreased growth and proliferation (Fig 5). This indicates that HRAS mRNA with exon 2 skipped is either not translated due to the lack of the normal ATG start codon, or if a protein is produced from an alternative start codon the resulting protein is unstable. These data show that skipping of HRAS exon 2 leads to decreased growth and proliferation consistent with reduced HRAS activity. Moreover, they confirm that an important ESE is located around position c.35 and that SSO-mediated blocking of access to this ESE reduces exon 2 inclusion.
Because recognition of ESEs by splicing stimulatory proteins is highly sequence specific, it is likely that other sequence variants in codon 12 and 13 may influence exon 2 inclusion and play a role in determining their phenotypic consequences. Consequently, we employed minigene transfection of T24 and HepG2 cells to test codon 12 and 13 mutations, which are known to cause CS [8] or cancer (Cosmic) for their effect on exon 2 inclusion (Fig 6). This showed that mutations like c.35G>C and c.37G>T, which are frequent in CS, but infrequent in cancer, have a relatively high level of exon 2 skipping, potentially attenuating their deleterious effect, whereas mutations like c.35G>A and c.35G>T, which are very frequent in cancers and rare in CS, have a high level of exon 2 inclusion (Fig 6).
In line with this, it is quite obvious that the c.35_36GC>TG mutation despite encoding the most severe mutant protein, p.Gly12Val, would most likely never be observed in cancer (c.35_36GC>TG is not present in the Cosmic database) due to the high level of exon 2 skipping and conversely, c.35G>T, which encodes an identical protein, is the second most frequent mutation in cancer and very rare in Costello syndrome due to the very efficient inclusion of exon 2.
The most frequent CS mutation, c.34G>A, has a very modest, nearly neutral, negative effect on exon 2 inclusion and its high occurrence in both CS and cancer is thus probably mainly due to a high mutation rate (due to CpG hypermutability) and a modest transforming potential of the encoded p.G12S protein.
We show that a particular mutation, c.35_36GC>TG, which encodes the prototypical oncogenic, constitutively active p.G12V HRAS protein, causes exon 2 skipping in an individual with CS with an attenuated clinical phenotype. This key finding demonstrates, in vivo, in a patient that exon 2 skipping leads limited production of the constitutively active oncogenic HRAS, thereby attenuating clinical symptoms. Simultaneously, this points to a previously unrecognized “Achilles heel” of the HRAS gene, namely that exon 2 is weakly defined due to a suboptimal 3’splice site. It’s inclusion in the mRNA is therefore dependent on binding of splicing stimulatory proteins, like SRSF2 to ESEs, and that binding of splicing inhibitory proteins, like hnRNPF/H to ESSs is avoided. Thus HRAS exon 2 inclusion can be affected by mutations altering the balance between ESEs and ESSs and this could in turn also result in cell type specific differences in splicing efficiency dependent on the relative levels/activities of splicing regulatory proteins, like SRSF2 or hnRNPF/H in the relevant tissue. Consistent with this, we show that mutations in codon 12 and 13 (c.34-39) impact exon 2 inclusion differently, and that HRAS exon 2 splicing efficiency is different (e.g. T24 and HepG2 cells). Thus, our results illustrate that the oncogenic effect of different mutations in HRAS may be determined also by their effect on exon 2 splicing efficiency. This adds an additional layer to the complex interpretation of the molecular consequences of mutations in HRAS exon 2. We posit that a delicate balance exists between the mutability of the different nucleotides, the resulting efficiency of exon 2 inclusion, and the oncogenic effect of the encoded mutant protein. We postulate that the observed frequencies of the various mutations in codon 12 and 13 in CS and cancers are a result of this balance.
This has clear implications for our understanding of the correlation between genotype and phenotype in diseases caused by HRAS mutations and highlights the general importance of the “splicing code” [13], by providing a striking example on how exonic mutations, like c.35_36GC>TG, can affect splicing and have dramatically different effects than those predicted based solely on the genetic code.
Finally, we show that this previously unknown weakness of the HRAS gene points to a new mechanism for knocking out oncogenic HRAS by employing SSOs that block binding of the required splicing regulatory factors, resulting in exon 2 skipping and decreased growth of cancer cells. SSOs targeting HRAS exon 2 splicing may represent a new therapeutic approach either when used alone or in combination with other therapies. In contrast to traditional drugs, SSOs are highly specific for a single gene and there are currently promising clinical trials employing SSOs for treating human disease. SSOs that are able to inhibit tumorigenesis in vivo by altering splicing of genes, like Bcl-X, STAT3 and MDM4 have been reported [35–37]. A particularly appealing characteristic of SSOs is that delivery of several different SSOs targeting different cancer genes, like those mentioned above, could be performed simultaneously using the same delivery method. In this way multiple oncogenic mechanisms could be targeted in a single approach. Our data suggest that SSO-based therapy targeting HRAS could also be included in such a future strategy.
Written informed consent was obtained from all participants and the study was approved by the Institutional Review Board at the University of Utah (IRB#00013747).
This 12-year-old boy had a history of hypertrophic cardiomyopathy status post septal myomectomy at 11 months of age. An echocardiogram at 12 years showed only mild septal hypertrophy with trace aortic insufficiency. A Nissen fundoplication and gastrostomy tube (GT) placement were performed at 2 months of age due to swallowing dysfunction and aspiration. The GT was used intermittently for 10 years but subsequently removed. He had one generalized seizure at 1 year without recurrence. A brain MRI at 17 months showed mild enlargement of the lateral and third ventricles with a mild Chiari I malformation. He received growth hormone injections starting at 9 years of age due to growth hormone deficiency. At 12 years, growth parameters were as follows: height = 142 cm (10th centile), weight = 32 kg (5th centile), and head circumference = 55.5 cm (75th centile). Other clinical features included ptosis, telecanthus, posteriorly rotated ears, deviated nasal septum, dental crowding, retrognathia, slightly large appearing hands without significant skin redundancy/deep creases or ulnar deviation, pes planus, an asymmetric anterior chest wall deformity, hyperflexibility, and mild kyphosis. He had one large nevus on his leg but otherwise did not have any additional dermatologic abnormalities and his hair appeared normal. There was no history of malignancies. He had mild developmental delay with good verbal skills and a full scale IQ of 76. He required resource classes for approximately 40% of his classes but was in the mainstream educational system for the remaining classes with some modification.
Genomic DNA was used for PCR amplification of a fragment of the human HRAS gene (NC_000011.9) encompassing exons 1–4 using Platinium Pfx DNA Polymerase supplemented with enhancer solution (Invitrogen) and primers HRAS1sNheI: 5’-GGCCCCGCTAGCAGTCGCGCCTGTGAA-3’ and HRAS1asXhoI: 5’-GTGAAGGACTCGAGTGACGTGCCCAT-3’. The amplified fragment was digested with NheI and XhoI and cloned into the polylinker of pcDNA.3.1+ (Invitrogen). Mutations were introduced by site-directed mutagenesis using standard methods either by the authors or by GeneScript Inc. (GenScript, Piscataway, NJ, USA). All plasmids were sequenced by GATC Biotech AG (Germany) in order to exclude any PCR derived errors.
HRAS exon 2 and variant double stranded DNA oligonucleotides corresponding to c.13_47 of HRAS exon 2 were inserted into the alternatively spliced second exon in the RHCglo splicing reporter minigene [31]. To generate pSXN constructs [30] we used sense and antisense oligonucleotides with desired sequences. The integrity of all constructs was confirmed by sequencing.
T24 human urinary bladder cancer cells were obtained from Coriell Institute (https://catalog.coriell.org/). HepG2 human hepatocellular carcinoma cells were obtained from American Type Culture Collection (ATCC) (http://www.lgcstandards-atcc.org/en.aspx). HepG2 or T24 cells were grown under standard conditions using 10% RPMI (Lonza RPMI 1640 added 10% FCS, glutamine (100x) and pen/strep (1000 U/ml)) or 5% RPMI (Lonza RPMI 1640 added 5%FCS, glutamine (100x) and pen/strep (1000 U/ml)), respectively. Twenty-four hours before transfection the cells were seeded 9.6 cm2 6-well plates (Nunc) at a density of 1.7×105 (HepG2) or 1.2×105 (T24) in 2 ml 5% or 10% RPMI (25% confluence) and grown O.N. to a density of 50% confluence on the day of transfection. Cells were transfected with a total DNA amount of 800 ng per well using X-tremeGene 9 DNA Transfection Reagent (Roche). Cells were transfected with 600 ng of plasmid DNA of interest and co-transfected with 200 ng MCAD 362T plasmid [19] as positive control. Forty-eight hours after transfection cells were harvest for RNA using Isol-RNa Lysis Reagent (AH Diagnostic) and RNA isolated using phenol-chloroform extraction. cDNA synthesis was performed using Superscript VILO cDNA Synthesis Kit (Invitrogen). Splicing analysis was carried out by PCR amplification and agarose gel electrophoresis. For HRAS constructs we used a specific primer T7-EXT: 5’- ATTAATACGACTCACTATAGGG-3’ and a primer spanning the exon 3-exon 4 junction of the HRAS gene (RasEx4Ex3: 5’-CGTTTGATCTGCTCCTGTAC-3’). For the RHCglo constructs we used primers RSV5U: 5′-CATTCACCACATTGGTGTGC-3′ and TNIE4: 5′-AGGTGCTGCCGCCGGGCGGTGGCTG-3′. For the pSXN construct we used primers pSXN12s2: 5'-AAGGTGAACGTGGATGAAGTTGGTGGTG-3' and pSXN13as: 5'-CCCACGTGCAGCCTTTGACCTAGTA-3'. All transfections were performed in triplicates.
Affinity purification of RNA binding proteins was performed with 3’-biotin coupled RNA oligonucleotides (DNA Technology, Denmark) as previously described [19]. The sequences of the RNA oligonucleotides were: HRAS-wt (5′-GUCGACUGGUGGGCGCCGGCGGUGUGGGCAAGAGUG-3′) and HRAS-mut (5′-GUCGACUGGUGGGCGCCGUGGGUGUGGGCAAGAGUG-3′) corresponding to position c.17_52 of HRAS mRNA. For each purification 100 pmol of RNA oligonucleotide was coupled to 100 μl of streptavidin-coupled magnetic beads (Invitrogen) and incubated with HeLa nuclear extract (Cilbiotech S.A., Belgium). Eluted proteins were analyzed by western blotting using a monoclonal anti-SRSF2 antibody (sc-041550 from Millipore) or a polyclonal anti-hnRNPF/H (sc-15387 from Santa Cruz).
SSOs were phosphorothioate oligonucleotides with 2'-O-methyl modification on each sugar moiety (DNA-technology, Denmark). HRAS-SSO-A: 5’-CGCACUCUUGCCCACACCGCCGGCG-3’ (Nucleotides corresponding to pos. c.51_30) and control SSO: 5’-GCUCAAUAUGCUACUGCCAUGCUUG-3'.
Approximately 3×105 HepG2 or T24 cells were reverse transfected with 50 pmol (20 nM), 75 pmol (30 nM), 100 pmol (40 nM) or 250 pmol (100 nM) of SSOs using Lipofectamine RNAiMAX transfection reagent (Invitrogen). Forty-eight hours after transfection cells were harvest for RNA using Isol-RNa Lysis Reagent (AH Diagnostics, Denmark) and RNA isolated using phenol-chloroform extraction or analyzed by the WST-1 viability assay. cDNA synthesis was performed using Superscript VILO cDNA Synthesis Kit (Invitrogen). For exogenous splicing analysis cells were transfected with minigene constructs 24 hours after SSO treatment. Splicing analysis of endogenous HRAS transcripts were performed by PCR with primers located in exon 1 (HRAS1sNheIS: 5’-GGCCCCGCTAGCAGTCGCGCCTGTGAA-3’) and a primer spanning the exon 3-exon 4 junction of the HRAS gene (RasEx4Ex3AS: 5’-CGTTTGATCTGCTCCTGTAC-3’). For exogenous splicing analysis we used primers T7-EXT: 5’- ATTAATACGACTCACTATAGGG-3’ and RasEx4Ex3AS: 5’-CGTTTGATCTGCTCCTGTAC-3’. All experiments were performed at least in triplicate. To check the effect of SSO treatment on HRAS protein levels, protein was extracted for SDS-PAGE and western blotting using a mouse polyclonal anti-HRAS antibody (SAB1405964 from Sigma-Aldrich) or an rabbit anti-beta actin antibody (ab8229 from AbCam).
Cell viability was determined by the WST-1 viability assay in 96 well plates following the manufacturer’s instructions (Roche). Approximately 3×105 T24 cells/well were reverse transfected with 30nM of SSOs using Lipofectamine RNAiMAX transfection reagent (Invitrogen) and incubated for 72h. Absorbance was measured on a VERSAmax tunable microplate reader (Molecular devices) at 3, 4 and 5 hours after addition of the WST-1 reagent. Non-treated cells, cells treated only with Lipofectamine RNAiMAX transfection reagent (Invitrogen) and a non-targeting scrambled SSO served as controls. All WST-1 viability assays were performed at least in triplicate.
Knockdown of SRSF1 or SRSF2 was performed using 100 pmol siRNA SMARTpools (Thermo Scientific) and scrambled control. For endogenous knock down T24 and HepG2 cells were grown to a density of 60–80% confluence when treated with siRNA using Lipofectamine RNAiMAX transfection reagent according to manufacturer’s instructions. Forty-eight hours after transfection cells were harvest for RNA using Isol-RNa Lysis Reagent (AH Diagnostics) and RNA isolated using phenol-chloroform extraction or protein was extracted for SDS-PAGE and Western blotting. For exogenous knock down cells were transfected with minigene constructs 24 hours after siRNA treatment.
Approximately 9000 T24 cells/well were reverse transfected with 20 nM or 30 nM of SSO-A or control SSO using Lipofectamine RNAiMAX transfection reagent (Invitrogen) and cell index was continuously monitored for 140 h. Real-time proliferation analysis was conducted using E plates (Roche, Basel, Switzerland) and xCELLigence kinetic Systems (ACEA Biosciences, San Diego, CA). The xCELLigence software (RTCA 1.2) was used to collect impedance measurements (reported as Cell Index) every 10 min for up to 72 hours.
Real-time (RT) qPCR and analysis were performed using a LightCycler with software version 1.5.1.62 (Roche). The RT qPCR master mix was prepared using the FastStart Essential DNA Green Master (Roche). For quantification of total HRAS we used primers HRASEX1S: 5’-CAGTCGCGCCTGTGAACGGTGG-3’ and HRASEX3-2AS: 5’-CCTGCTTCCGGTAGGAATCCTCTATAGTGGG-3’. For exon 2 skipping analysis we used primers HRASEX1-3S: 5’-CGCGCCTGTGAACGGATTCC-3’ and HRASEX4-Ex3-QPCR2AS: 5’-CACCCGTTTGATCTGCTCCTGTACT-3’.
|
10.1371/journal.ppat.1006219 | Zika Virus infection of rhesus macaques leads to viral persistence in multiple tissues | Zika virus (ZIKV), an emerging flavivirus, has recently spread explosively through the Western hemisphere. In addition to symptoms including fever, rash, arthralgia, and conjunctivitis, ZIKV infection of pregnant women can cause microcephaly and other developmental abnormalities in the fetus. We report herein the results of ZIKV infection of adult rhesus macaques. Following subcutaneous infection, animals developed transient plasma viremia and viruria from 1–7 days post infection (dpi) that was accompanied by the development of a rash, fever and conjunctivitis. Animals produced a robust adaptive immune response to ZIKV, although systemic cytokine response was minimal. At 7 dpi, virus was detected in peripheral nervous tissue, multiple lymphoid tissues, joints, and the uterus of the necropsied animals. Notably, viral RNA persisted in neuronal, lymphoid and joint/muscle tissues and the male and female reproductive tissues through 28 to 35 dpi. The tropism and persistence of ZIKV in the peripheral nerves and reproductive tract may provide a mechanism of subsequent neuropathogenesis and sexual transmission.
| Although it was first identified almost 70 years ago, Zika virus had rarely been associated with pathology in humans until the 21st century. Recent outbreaks in the South Pacific and the Americas have been characterized by numerous confirmed cases, some involving neurologic sequelae and, of most concern, birth defects following infection of pregnant women. Here, we present the results of experimental infection of adult rhesus macaques with a strain of Zika virus isolated during the recent epidemic in the Western hemisphere. Following infection, Zika virus was detected in the sera and urine of all infected animals. Further, we detected virus in multiple tissues of infected animals as late as 35 days post infection, indicating viral persistence. The apparent tropism of the virus for tissues of the peripheral nervous system as well as the reproductive tracts of males and females has implications for the further characterization of the mechanism(s) of Zika virus pathogenesis. Additionally, this model provides a platform for development and testing of preventative or therapeutic interventions to combat the emergence of this virus.
| Zika virus (ZIKV), once a little-studied member of the family Flaviviridae, forcefully emerged across the Western Hemisphere in 2015–16. As of October, 2016, the Centers for Disease Control and Prevention (CDC) lists 60 countries worldwide, including the continental U.S., that have reported autochthonous transmission of the virus, primarily through an Aedes spp mosquito vector (https://www.cdc.gov/zika/geo/active-countries.html). The virus is also believed to be endemic in multiple countries in Africa and Southeast Asia. The World Health Organization (WHO) has estimated that 3–4 million individuals will be infected with ZIKV in the next year [1]. While an estimated 80% of all infections are asymptomatic or subclinical, the remaining 20% of ZIKV infections often resembles infection by co-circulating dengue (DENV) or chikungunya (CHIKV) viruses, with typical symptoms including fever, rash, headache and arthralgia, although ZIKV does appear to have a fairly distinctive association with conjunctivitis [2,3]. ZIKV was first isolated in 1947, from serum taken from a febrile sentinel rhesus macaque (RM) in the Zika Forest region of Uganda [4,5]. Although reported instances of human ZIKV-associated disease during the 20th century had been sporadic with generally mild disease, large outbreaks were reported in Yap State, Micronesia in 2007 and in French Polynesia in 2013, resulting in ~900 and 30,000 symptomatic cases, respectively [6,7]. Illness during these outbreaks was initially characterized as self-limiting and did not require hospitalization. However, 74 patients in French Polynesia who experienced confirmed or probable ZIKV infection later presented with neurological complications. Over half of these were characterized as Guillain-Barré syndrome (GBS); the remainder included various encephalitides, paraesthesia, facial paralysis and myelitis. Similarly, increases in GBS have been reported in 12 countries worldwide, together with laboratory confirmation of ZIKV infection associated with these cases [8]. Of particular concern is the presumed causal relationship between ZIKV and microcephaly in developing fetuses. In several cases, ZIKV infection of the mother during pregnancy, as well as the presence of ZIKV in the amniotic fluid or tissue of fetuses showing evidence of microcephaly was reported [9–12]. However, the specific mechanisms by which ZIKV causes fetal neurological defects in humans remain unknown. Mouse models, both with and without intact innate immune signaling, of ZIKV infection during pregnancy have proven susceptible to infection of placental and fetal tissue, resulting in intrauterine growth restriction and fetal death [13,14]. However, differences in placental architecture and fetal development in the mouse vis-à-vis humans suggest that certain aspects of ZIKV pathogenesis during pregnancy may not be reflected in the murine infection model [15].
Non-human primates, by virtue of their relatedness to humans, are valuable models for the study of human disease. For example, experimental infection of RM with yellow fever virus (YFV) results in viscerotropic disease that closely parallels the course observed in humans [16,17]. Additionally, we have recently developed a model of CHIKV infection in the RM that recapitulates several aspects of human disease, including joint tropism and inflammation, viremia, and robust innate and adaptive immune responses [18,19]. Both DENV [20] and WNV [21] infection of RMs results in detectable viremia and immune response, although infection is not associated with overt pathology. Prior to the current epidemic, the outcome of ZIKV infection in RM model had not been well characterized. However, the fact that the virus was originally isolated from a febrile RM suggests that viral replication, immune response, and aspects of pathogenesis may be modeled in RMs. Here, we report outcomes of infection of adult RMs with a ZIKV strain currently circulating in the Western hemisphere (a 2015 isolate from Puerto Rico). Sub-cutaneous inoculation of animals produced transient, detectable viremia and viruria, as well as clinical symptoms described in human ZIKV infections (e.g. fever, rash, conjunctivitis). These data are comparable to those additional recently published studies of ZIKV infection in RM [22–24]. Herein we also extend these data by examination of tissue tropism during infection. Cohorts of animals necropsied at day 7, 28 or 35 pi indicated that viral RNA was present within secondary lymphoid tissues, joints, peripheral nervous tissue and organs of the female reproductive tract. Further, a robust immune response, including production of neutralizing antibodies, was observed in all infected animals. Our data suggest that RM may provide a useful model for the study of ZIKV pathogenesis, as well as a platform for the testing of vaccines or anti-viral therapeutics.
All Zika virus infection experiments utilizing animals were performed in compliance with guidelines established by the Animal Welfare Act for housing and care of laboratory animals and conducted in accordance with Oregon National Primate Research Center (ONPRC) Institutional Animal Care and Use Committee approved protocol (IACUC #0993). RM studies were performed in ABSL-3 or ABSL-2 containment facilities at the Oregon National Primate Research Center (ONPRC), which are accredited by the Assessment and Accreditation of Laboratory Animal Care (AAALAC) International. Appropriate procedures were utilized in order to reduce potential distress, pain and discomfort. Ketamine (10 mg/kg) was used to sedate the animals during all procedures including routine blood draws performed by trained veterinary staff. Rhesus monkeys were fed standard monkey chow twice daily and the amount was matched to each animal according to body weight, age and sex and intake as monitored. Animals also received daily food supplements and other enrichment devices. The infected animals were caged with partners or caged separately but within visual and auditory contact of other animals in order to promote social behavior. At the designated time points, the animals were euthanized according to the recommendations of the American Veterinary Medical Association 2013 panel on Euthanasia.
Zika virus train PRVABC59 was isolated by the Centers for Disease Control (CDC) from an individual in Puerto Rico in December 2015 [25]. PRVABC59 was obtained from the CDC, and passaged twice in C6/36 cells (American Type Culture Collection, ATCC). To prepare virus stock, infected C6/36 tissue culture supernatant was concentrated through a 20% sorbitol cushion and titered in Vero cells (ATCC) using a focus-formation assay. The virus stock was sequenced and found to conform to the previously described sequence (Genbank accession #KU501215.1) with the following four single base pair substitutions: G-1964-T (Envelope protein V to L; frequency 82.7%); T-3147-C (NS2A protein M to T; frequency 14.5%); C-5676-T (NS3 protein S to F; frequency 36.8%); and C-7915-T (NS5 protein silent mutation; frequency 13.5%). All cells were cultured in Dulbecco’s Modified Eagle Medium (DMEM; Corning) containing penicillin-streptomycin-glutamine (PSG; Corning) and 5–10% fetal calf serum (FCS; HyClone). Vero cells were grown at 37°C and C6/36 cells were grown at 28°C.
Serial dilutions of virus were plated in 96-well plates seeded with Vero cells, allowed to adsorb for 1 h, followed by overlay with 0.5% carboxymethyl-cellulose (CMC; Sigma). At 30 h pi, cells were fixed with 4% paraformaldehyde, washed twice with PBS and blocked/ permeabilized for 1 h in PBS supplemented with 2% normal goat serum (NGS; Sigma) and 0.4% triton X-100. Cells were then washed twice with PBS followed by incubation with 0.3 μg/ml anti-flavivirus monoclonal antibody 4G2 [26] in PBS supplemented with 2% NGS for 1 h, washed twice more with PBS, incubated with anti-mouse IgG-horseradish peroxidase (Santa Cruz Biotech) for 1 h, and washed twice with PBS. Foci were visualized by incubation with the Vector VIP peroxidase substrate kit (Vector Labs) according to manufacturer’s specifications and counted using an ELIspot reader (AID).
Seven Indian-origin RMs (3 females and 4 males) were divided into three cohorts (Fig 1 contains a description of the animals within each cohort). Cohort 1 was infected subcutaneously with a total of 1x104, 1x105, or 1x106 focus forming units (ffu) of ZIKV diluted in 1ml of PBS and delivered by ten 100μl injections into the hands and arms bilaterally. These doses are comparable to the 1x105 pfu median dose of the flavivirus West Nile virus (WNV) previously determined to be delivered by the bite of infected Culex spp mosquitoes [27]. Cohorts 2 and 3 were similarly infected with 1x105 ffu. Peripheral blood and urine (collected in the cage pan) samples were obtained at 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 14, 21, 28 and 35 dpi. Peripheral blood mononuclear cells (PBMCs) and plasma samples were separated by centrifugation over lymphocyte separation medium. PBMCs were analyzed for immune cell phenotype and frequency by flow cytometry. Plasma was assessed for viral loads by qRT-PCR and the levels of cytokines by Luminex multiplex-bead based assay, as described below. Urine was assessed for viral RNA by qRT-PCR and for infectious virus by co-culture on C6/36 cells and focus-forming assays using Vero cells. Cohort 1 was euthanized at 28 dpi, Cohort 2 animals at 7 dpi, and Cohort 3 at 35 dpi. Samples of tissues (joints, muscles, organs, brain, spinal cord, peripheral nerves, glands, and lymph nodes) and biological fluids (cerebral spinal fluid, blood, and urine) were collected and stored in RNAlater, Trizol (RNA isolation), medium (virus isolation) as well as fixed and embedded in paraffin.
Plasma anti-ZIKV antibody concentrations were measured by end point dilution ELISA. For this assay, high-binding polystyrene 96-well plates (Corning) were coated with PBS containing a 1:1000 dilution of 4x108 ffu/ml stock of purified ZIKV particle preparations The plates were incubated overnight at 4°C and then blocked with PBS containing 2% milk and 0.05% Tween (ELISA-Block) for 1 hr at room temperature. Plates were washed with 0.05% Tween-PBS (ELISA-Wash) and incubated with two-fold dilutions of RM plasma in ELISA-Block starting at a dilution of 1:50. The plate was incubated at room temperature for 2 hrs. Plates were washed several times with ELISA-Wash and then incubated with secondary anti-monkey IgM or IgG (Rockland, Inc.) conjugated with horseradish peroxidase for 30 mins. Plates were washed with ELISA-Wash and bound secondary antibody was detected using the OPD substrate (Life Technologies) followed by HCl stop the assay. The plates were read within 10 minutes using a Synergy HTX Microplate Reader (BioTek) at 490nm. Endpoint titers of ZIKV binding antibodies were determined using a Log/Log transformation method and the results were analyzed and graphed using GraphPad Prism v6 software.
Neutralization assays were used to measure the concentration of serum that can neutralize 50% of a fixed number of ZIKV (50% Plaque Reduction Neutralization Test (PRNT50)). Sera from infected RM were serially diluted 4-fold from starting dilutions of 1:10 and following dilution were mixed with an equal volume of ZIKV (30–50 PFU) for final serum dilutions ranging from 1:20 to 1:5120. Sera and virus were incubated for 1hr at 37°C. The mixtures were added to individual wells of 24-well plates seeded with Vero cells at 90% confluence for 1hr at 37°C on a rocker and then overlaid with 1% methylcellulose in OPTI-MEM (Gibco). Plates were incubated for 3 days at which time the cells were fixed and counterstained with methyl blue. Plaques were visualized and counted using a light box. Raw counts were entered into GraphPad Prism v. 6.0, converted to a percent of mock neutralized input virus and PRNT50 values were calculated utilizing the sigmoid dose-response curve fitting function with upper and lower limits of 100 and 0, respectively.
RNA from tissue samples, blood, urine, and cerebrospinal fluid (CSF) was isolated using TRIzol (Invitrogen) according to the manufacturer’s protocol. ZIKV RNA levels were measured by a one-step quantitative real time reverse transcription polymerase chain reaction assay (qRT-PCR) using TaqMan One-Step RT-PCR Master Mix (Applied Biosystems). 250ng total RNA from tissue samples or 1/10th volume of RNA isolated from 100 μl of liquid samples was used in each reaction. Primers and probes were as follows: Forward: 5’-TGCTCCCACCACTTCAACAA (ZIKV PRVABC59 genome sequence nucleotides 9797–9816); Reverse: 5’-GGCAGGGAACCACAATGG; (complement of nucleotides 9840–9857); and TaqMan probe: 5’ Fam-TCCATCTCAAGGACGG-MGB (nucleotides 9819–9834). Forward and reverse primers were used at 250 nM in the reaction, and the probe at 200 nM. Validation of the qRT-PCR assay is shown in S1 Fig. For RNA standards, RNA was isolated from purified, titered stock of ZIKV (PRVABC59). RNA yield was quantified by spectrometry and the data was used to calculate genomes/μl. Focus-forming units (ffu)/ μl was calculated based on titer of stock. ZIKV RNA was serially diluted 1:10 into Vero cell RNA (25 ng/μl) and amplified in triplicate using primers and conditions described above.
Tissues were homogenized in 1ml of DMEM cell culture medium containing 5% FBS and PSG plus approximately 250μl of SiLiBeads using a bead beater (Precellys 24 homogenizer, Bertin Technologies), and cellular debris were pelleted by centrifugation (5,000 × g for 2 min). A 50μl or 500μl sample of the clarified lysate was applied to one well of a 6-well plate of C6/36 cells for seven days. Supernatant titers from these cultures were transferred to Vero cells, incubated at 37°C for an additional 5 d and assayed for the presence of infectious virus by indirect immunofluorescent staining using mAb 4G2 and an anti-mouse IgG Alexa-488 conjugated secondary antibody. This method proved must sensitive in isolation of infectious virus, as compared to initially culturing tissue samples with Vero cells.
Flow cytometry was used to quantify the immune cell phenotype as well as the level of cellular proliferation and activation for peripheral blood mononuclear cells (PBMCs) isolated at the time points defined above. The panel of antibodies used for the analysis of innate immune cells consisted of HLA-DR, CD14, CD11c, CD123, CD20, CD3, CD8, CD16, and CD169. To differentiate between monocyte/macrophages, DCs, and NK cells the following gating strategy was utilized: monocyte/macrophages (CD3-CD20-CD14+HLA-DR+), myeloid DCs (CD3-CD20-CD14-HLA-DR+CD11c+), plasmacytoid DCs (CD3-CD20-CD14-HLA-DR+CD123+), other DCs (CD3-CD20-CD14-HLA-DR+CD123-CD11c-), and natural killer (NK) cells (CD3-CD20-CD8+CD16+). The percentage of activated cells (CD169+) within each subset was calculated as a representation of the cellular activation profile [28]. T cells were analyzed with the following panel of antibodies directed against CD4, CD8β, CD95, CD28, CD127 and for intracellular levels of Ki67 (proliferation marker). The T cell subset was identified as CD4+ or CD8+ and within the CD4+ and CD8+ T cell subsets, the naïve (CD28+CD95-), central memory (CD28+CD95+), and effector memory (CD28-CD95+) subsets are displayed. B cells were analyzed using the following antibodies: CD3, CD20, CD27, and IgD to delineate naïve (CD3-CD20+CD27-IgD+), memory (CD3-CD20+CD27-IgD-) and marginal-zone like B cells (CD3-CD20+CD27+IgD+) as well as Ki67 to identify proliferating cells. The percentage of proliferating (Ki67+) B and T cells within each subset was calculated as well as for granzyme B (activation marker). The gating strategies and definition of the different cellular subsets were performed as previously described [18]. Phenotyping was performed using an LSRII instrument (BD bioscience) and the data was analyzed with FlowJo Software (TreeStar).
Monkey Cytokine Magnetic 29-plex Panel (Luminex Platform Kit from Invitrogen) was used to quantify cytokine and chemokine expression in blood plasma and CSF samples. According to the manufacturer’s instructions, antibody-conjugated polystyrene magnetic beads were plated onto a 96-well plate and washed with buffer. Beads were incubated with a 7-point standard curve along with 25μl of rhesus monkey plasma or CSF plus 25μl of blocking buffer for 2h. Beads were washed with wash buffer and labeled with the biotinylated detector antibody for 1hr. Beads were washed and then incubated with Streptavidin conjugated to R-Phycoerythrin for 30 minutes and washed. After final wash, cytokines were identified and quantified using a Luminex 200 Detection system (Luminex). Statistical analysis was performed using Sidak’s multiple comparison tests and data was graphed using GraphPad Prism v6 software.
Whole blood chemistry analysis was performed on a VetScan VS2 system (Abaxis, Union City CA) using the 14-analyte Mammalian Comprehensive Diagnostic Profile Panel (#500–0038) according to the manufacturer instructions.
Complete necropsies were performed and tissues were collected for microscopic examination. Tissues were fixed in 10% buffered formalin, embedded in paraffin, sectioned at 5μm and stained with hematoxylin and eosin.
In situ hybridization studies were performed on formalin fixed paraffin-embedded (FFPE) tissue sections of 5μm using two different Zika-specific commercial RNAscope Target Probes (Advanced Cell Diagnostics, Hayward, CA; catalog #464531 and #463781) complementary to sequences 866–1763 and 1550–2456, respectively. Pretreatment, hybridization and detection techniques were performed according to manufacturer’s instructions. In the absence of control specimens of Zika virus infected cells/tissues, FFPE brain tissue from mice infected with West Nile Virus [29] were used as positive controls. The target probe #463781 detected WNV infected brain tissue while the probe #464531 did not. Both probes detected ZIKV in the test specimens. Tissue sections were counterstained with hematoxylin. As a negative control, a probe specific for Influenza A virus (ACD catalog #313241 was used to probe contiguous sections of Zika positive tissues.
Spleen and lymph node tissues were dissected at necropsy to produce single cell suspensions by first pushing the tissues through a wire mesh filter followed by extensive washing, and lysis of red blood cells. These cell preparations were put through a 70μm filter and counted prior to being frozen in RPMI containing 50% fetal calf serum plus 10% DMSO in liquid nitrogen. For MACS the cells were thawed, pelleted by low speed centrifugation (1,500 rpm for 10 minutes) and resuspended in MACS buffer at a concentration of approximately 2x107 cells/mL. Prior to magnetic separation, cells were passed through a 70μM filter, to remove cell clumps. The cells were then divided into two aliquots of 1x107 cells/mL per tissue type. One aliquot was used to obtain pure populations of CD14+ macrophages and CD3+ T cells and the second to obtain pure populations of CD20+ B cells and CD1c+ dendritic cells. All cell types were isolated using a two-step magnetic isolation method with RM-specific reagents (MACS, Miltenyi Biotec, Germany).
For isolation of T cells and macrophages, approximately 1x107 cells were incubated with magnetic beads coated with anti-CD14 (Miltenyi Biotech) for 15 min at 4°C. The cells were then washed and resuspended in 500μL MACS buffer before being loaded onto a LD magnetic separation column. After sample loading the column was washed with 2mL of MACS buffer. The magnetically labeled CD14+ macrophages retained on the column were eluted by removing the column from the magnetic field and flushing with MACS buffer using the provided plunger. A portion of the eluted cells were saved for purity analysis via flow cytometry, the remaining eluted cells were pelleted and resuspended in Trizol reagent. The CD14-depleted cells in the flow-through were pelleted, resuspended in MACS buffer and incubated with magnetic beads coated with anti-CD3-biotin antibody (Miltenyi Biotech) for 10 min at 4°C, followed by subsequent incubation with anti-biotin MicroBeads (Miltenyi Biotech) for 15 min. The cells were then washed and resuspended in 500μL MACS buffer before being loaded onto a MS magnetic separation column for positive selection. Labeled CD3+ T cells, retained on the column, were flushed out of the column with 1mL MACS buffer with the provided plunger in the absence of a magnetic field. Eluted cells were analyzed for purity and vRNA as above.
For isolation of CD20+ B cells and CD1c+ DCs, the total cell mixtures were first incubated with a FcR blocking reagent (MACS) and magnetic beads coated with anti-CD1c-PE (MACS) for 5 mins at 4°C followed by a 15 mins incubation with magnetic bead coated with anti-CD20 (MACS). Cells were then washed, resuspended in 500μL MACS buffer, and loaded onto a LD magnetic separation column. The CD20+ B cells retained on the column were eluted with MACS buffer as described above, and a portion was analyzed for purity via flow cytometry. The remaining eluted cells were pelleted and resuspended in Trizol reagent. To positively select for CD1c+ DCs the B cell-depleted flow-through fraction was pelleted, resuspended in MACS buffer, and then incubated with anti-PE Microbeads (MACS) for 15 mins at 4°C. The cells were then washed and resuspended in 500μL MACS buffer before being loaded onto a MS magnetic separation column. The CD1c+ DCs retained on the column were eluted and analyzed as above.
Luciferase assays were performed as previously described [30]. Briefly, the firefly luciferase (LUC) open reading frame downstream of an NF-κB promoter element was transduced via lentivector (Qiagen) into RM fibroblasts whose functional lifespan was extended through the stable introduction of human telomerase. Cells were grown, infected, and treated in 96 well plates as indicated. After adding Steady Glo lysis and luciferin reagent (Promega) luminescence was read on a BioTek Synergy plate reader.
Three cohorts of RMs were used for this study (Fig 1). The first cohort consisted of two adult females and one adult male. Animals were infected with 1x104, 1x105, or 1x106 focus forming units (ffu) of ZIKV (PRVABC59). The infectious dose was divided over 10 subcutaneous injections over bilateral hands and arms. Blood and urine were sampled daily through 10 dpi, as well as on 14, 21, and 28 dpi. Euthanasia was performed at 28 dpi and tissues collected at necropsy for analysis of viral loads. A second cohort, consisting of two adult RM (one male, one female) was infected with 1x105 ffu, followed by daily sampling of blood and urine through 7 dpi, at which time animals were euthanized as above. The third cohort, consisting of two adult male RM, was infected with 1x105 ffu followed by daily sampling of blood and urine through 35 dpi. All animals developed a transient fever, rash on the arms and upper torso, as well as lymphadenopathy of their axillary lymph nodes. Additionally, 3 of 7 animals developed conjunctivitis lasting 3–5 days. None of the infected animals experienced weight loss or signs of clinical disease other than those described above. Analysis of blood chemistry revealed no significant changes following ZIKV infection (S2 Fig).
All infected animals developed plasma viremia as detected by RT-qPCR of viral genomes that typically peaked at 2 dpi and was detectable out to 5–7 dpi (Fig 2A and 2B). We were unable to titer virus directly from plasma samples. However, infectious virus, detected by co-culture of plasma with C6/36 cells, was observed in indicated cases between 2 to 4 dpi (Fig 2A, stars). Viral RNA in the urine was detected from 3–10 dpi with peak levels at 5 dpi (Fig 2A and 2B). We detected viral RNA positive urine samples outside of the initial 3–10 dpi window, which is consistent with other reports of ZIKV infections of NHP [22]. This finding indicates that ZIKV infection in NHP is dynamic and remains persistent.
Following euthanasia and necropsy, RNA was isolated from individual tissues and the viral genomes were quantified by qRT-PCR (viral loads of positive tissues in Fig 3A and 3B; complete list of tissues samples in S1 Table). At 7 dpi (cohort 2, 1x105 ffu), viral RNA was detected in multiple tissues: lymphoid tissue—including lymph nodes distributed throughout the body as well as the spleen; joints—most prominently joints near the site of inoculation but in some instances more distal joint tissue as well; peripheral nervous tissue, specifically the sciatic nerve, brachial plexus and trigeminal ganglion. Additionally, viral RNA was found associated with the spinal cord (cervical, lumbar and thoracic), but not in CSF or in the brain, at this time point. This may indicate neurological tropism, but an inability to effect retrograde transport of infectious virus into the CNS or that infection of the CNS requires additional time. Viral RNA was detected in the kidney and bladder of the male animal (27679), although not in the testes or prostate. Viral RNA was found in the uterus of the female (24504).
We were able to co-culture infectious ZIKV in C6/36 insect cells from homogenates of the axillary and inguinal lymph nodes, finger joints, kidney and bladder derived from the male monkey (27679) (Fig 3A, black arrows). Together these data indicate that ZIKV quickly disseminates to many tissues throughout the body including lymph nodes/spleen, peripheral nerves, and skin as well as the genital/urinary tract.
At 28 dpi, (cohort 1, inoculated with 1x104 1x105 and 1x106 ffu) viral RNA was still detected in both lymphoid and joint tissues in all animals (Fig 3B). In general, vRNA tissue distribution was greatest for the 1x105 and 1x106 ffu infected animals. The axillary (draining) lymph nodes and spleen showed the highest level of viral RNA in all three animals, while other lymph nodes were positive for at least 2 of 3 animals. Joint tissues close to the site of inoculation (wrist and finger) were also positive in all three animals 28 dpi. Additional joints, muscles of the arms and legs, and heart were positive for ZIKV RNA in subsets of animals. In animal 25421 (female, 1x106 ffu) viral RNA was detected in the reproductive tissues (uterus and vagina) suggesting that the virus can infect these tissues and persist there for at least 4 weeks post infection. This finding may have important implications for viral transmission and fetal infections during pregnancy. Viral RNA was also detected in the sciatic nerve and eyes from this subject. Interestingly, ZIKV RNA was detected in the cerebellum of animal 24561 (female, 1x104 ffu), indicating penetration to the CNS. Co-culture of homogenates from tissues collected at 28 dpi with C6/36 cells did not amplify infectious virus.
At 35 dpi, (cohort 3, inoculated with 1x105 ffu) positive viral RNA detection occurred in neuronal tissues, lymph nodes, and joint/muscle tissues (Fig 3C). Animal 26023 displayed extensive neuronal tissue involvement with viral RNA detected in the occipital and parietal lobes of the brain, lumbar region of the spinal cord, dorsal root ganglia, brachial plexus, and eye. In Animal 26021, ZIKV RNA was not detected in the brain but was present in the trigeminal ganglia, as well as cervical, lumbar and thoracic regions of the spinal cord and peripheral nerves (brachial plexus and sciatic nerve). Interestingly, in situ hybridization on cross sections of sciatic nerve using ZIKV-specific chromogenic probes detected robust virus RNA levels in the perineurial adventitial space from animal 26023 (Fig 3D). Virus was not detected in the nerve fibers. Both animals had viral RNA in their axillary lymph nodes. These results combined with the viral detection data from the day 28 animals confirm the long-term persistence of ZIKV RNA in neuronal, lymph node and joint/muscle tissues.
Histologic examination of sections taken from tissues of infected RM found few specific abnormalities, although several areas of inflammation were observed (S3 Fig). An uncharacteristic prostatitis characterized by interstitial neutrophilic and lymphoplasmacytic cellular infiltrates and glandular microabscesses were noted 7dpi in animal 27679 infected with 1x105 ffu (S3A Fig). Minimal perivascular lymphocytic or lymphoplasmacytic inflammatory cell infiltrates were present in sections of skin from the upper torso affected with a rash for both animals examined 7dpi (S3B Fig). Viral RNA was also detected in this area of skin in this animal. Similarly, variable perivascular inflammatory infiltrates composed of lymphocytes, eosinophils and plasma cells were observed in the joints and muscles of animal 24504 (S3C & S3D Fig). Focal lymphohistiocytic inflammation was associated with a meningeal vessel in the cerebrum of the high dose (1x106 ffu) animal 25421 at 28 dpi suggestive of an ongoing infection of the brain vasculature (S3E Fig). Animal 25147 (28 d pi, 1x105 ffu) had focal lymphocytic infiltration of the dorsal root ganglion of the cervical spinal cord (S3F Fig). The lack of correlation of detection of viral RNA with sites of inflammation in the prostate, brain, and DRG may indicate highly focal areas of infection, or clearance of virus prior to resolution of inflammation.
In order to determine which cell types within lymphoid tissues were positive for viral RNA, we sorted macrophage, dendritic cell, B-cells and T-cells from the splenocytes and axillary lymphocytes by positive magnetic bead selection (S4 Fig). RNA was isolated from each cell population and ZIKV RNA quantified by qRT-PCR. As shown, at 28 dpi RNA is primarily found in the macrophage and B cell subsets with reduced levels in DC subsets but rarely present in the T cell fractions (Fig 4A). In situ hybridization with ZIKV- and Influenza-specific probes detected ZIKV but not Flu RNA in multiple axillary lymph node follicles from animal #25421 (Fig 4B), confirming the presence of the ZIKV RNA in the macrophage, B cell and DC rich regions of the germinal center. Overall, this data indicates that ZIKV spreads to multiple tissue types and the infection of many of these tissues persists in macrophages, as well as other cell types for at least 4–5 weeks post infection.
We performed a detailed phenotypic analysis of immune cell subsets by flow cytometry to characterize activation of innate immune cells (monocyte/macrophage/DC/NK cells) as well as adaptive immune cell proliferative responses (T and B cells). We also characterized cytokines and antibodies present in the sera of infected RM. Within 1–2 days pi, all of the animals showed innate immune cell activation, as demonstrated by the presence of CD169+ staining (Fig 5). RM 24961 (1x104 ffu) displayed a more protracted innate immune response, compared to 25421 (1x106 ffu), 25147, 26021 and 26023 (1x105 ffu). While all animals showed an increase in CD169+ monocytes and DCs at 2–4 dpi, the number of activated cells waned between day 8–10 pi in animals infected with 1x105 or 1x106 ffu, while the number of activated cells in the animal infected with 1x104 ffu did not return to baseline levels until 14–21 dpi.
Cytokine expression in the plasma largely did not change following infection. However, expression of 4 cytokines (IL-1RA, MCP-1-CCL2, IP-10-CXCL10, and I-TAC-CXCL11) was induced over background levels in the plasma (Fig 6A–6D). Expression of these cytokines was elevated within the first several days post infection but returned to baseline levels by 10 dpi. Low levels of cytokine activation in vivo may be an indirect effect of routine ketamine treatment [31] or a result of direct inhibition by ZIKV of innate immune pathways that direct synthesis and secretion of pro-inflammatory cytokines. To examine the latter possibility we employed a reporter assay for which the readout is luciferase expression that responds to NF-κB or JAK/STAT pathway (type I IFN) activation [32,33]. As shown in Fig 6E and 6F, rhesus fibroblasts infected with ZIKV for 48h at 5 FFU/cell showed significantly diminished LUC signal relative to uninfected cells following treatment with either poly(I:C) or human IL-1β. These represent distinct NF-κB-terminal signaling pathways with poly(I:C) induction resulting from activation of the TLR3 pattern recognition receptor and TRIF adaptor protein as well as IL-1β triggering the IL1 receptor and associated MyD88 adaptor protein. Similarly, activation of JAK/STAT signaling by IFNβ1 treatment was also repressed by ZIKV infection (Fig 6G). These results agree with previous observations that ZIKV infection promotes degradation of STAT2 and subsequent inhibition of type I IFN signaling [34]. As such we hypothesize that ZIKV exhibits an inhibitory phenotype that operates downstream of the convergence of these pathways, likely targeting activation of NF-κB itself but delimiting the mechanisms associated with this point will require further experimentation.
Proliferating CD4+ and CD8+ T-cells were present in all infected animals by 6–8 dpi. CD8+ T cell proliferative responses (Ki67+ cells) were evident at 6 dpi, maximal at 8–9 dpi and returned to background levels by 14 dpi (Fig 7B and 7D). Both central memory and effector memory CD4+ T cell proliferative responses were maximal at 7 dpi (Fig 7A and 7C) but took longer to return to baseline levels compared to the CD8+ T cells. Consistent with these findings, Granzyme B expression in naïve and central memory CD4+ and CD8+ T cells peaked between 7 and 10 days post infection (S5 Fig).
B cell proliferative burst responses were maximal at 14 dpi (Fig 7E–7G). Interestingly, when comparing the cohort 1 animals, the B-cell proliferative responses in RM 24961 (1x104 ffu) were observed slightly earlier than in RMs 25147 (1x105 ffu) and 25421 (1x106 ffu) and represented a greater percentage of cells within each subset. Proliferating T-cells also appeared for a longer time post infection in this animal (Fig 7A–7D).
ZIKV virion-reactive IgM and IgG antibodies in sera were quantified by ELISA. Levels of anti-ZIKV IgM became detectable between 7–10 dpi and were maintained through 28 dpi in 2 of 3 animals, while one animal (#25421) showed reduced titers after 10 dpi. (Fig 8A). IgG levels increased beginning between d8 to d14 pi and plateaued around 21 dpi (animals #25147 and #25421) or continued to increase in the low dose animal (#24961). Western blotting of ZIKV infected cell lysates revealed that antibody responses targeted at least two proteins that were 38 and 55 kDa, respectively, consistent with viral proteins NS1 and E (S6 Fig), which elicit antibody responses during ZIKV infection of humans and are also major antibody targets during other flavivirus infections [35,36]. The neutralizing capacity of the ZIKV-directed antibodies was quantitated, and robust neutralizing antibody responses were detected at 28 or 35 dpi (Fig 8C and 8D) in all animals, regardless of the infectious dose.
In 1947, the Zika virus was originally isolated in Uganda from a febrile RM used as a sentinel in a study of yellow fever virus transmission. Given this initial finding, we hypothesized that ZIKV infection of RM could be used as a model of viral replication, pathogenesis, and immune response. Our results demonstrate that the 2015 Puerto Rico ZIKV strain productively infects the RM, characterized by clinical symptoms comparable to that described in human infection, such as fever, rash, and conjunctivitis. The RM further develops viremia, viruria, widespread tissue infection and a robust adaptive immune response. These data are generally in agreement with recently published studies of RM infected with a 2013 isolate from French Polynesia [22].
Because ZIKV has shown a breadth of tissue tropism not seen in other human flavivirus infections, we sought to characterize distribution of ZIKV in RMs as broadly as possible. As such, our study significantly advances what is known about the tissue tropism of ZIKV on 7, 28 and 35 dpi, a key feature of infection not examined by previous RM studies at these time points. Analysis of viral genome load in tissues revealed a tropism for lymphoid, joint and peripheral nerve tissue. The apparent persistence of viral RNA in various tissues after resolution of primary viremia is not unique to ZIKV. Persistence of flaviviruses in the infected host after cessation of viremia and recovery from clinical symptoms (if any) has been observed in several instances. In humans who have been infected with WNV, prolonged viruria (in several cases >6 years) is sometimes observed [37]. This viral persistence may correlate with long-term neurological and renal sequelae [38,39]. Models of WNV persistence in RM, mice, and hamsters indicate that virus can persist for months in tissues of the CNS, kidney, and lymphoid organs, and the presence of virus also correlates with long term neurological sequelae and motor neuron loss in hamsters [21,40–43]. Herein, we demonstrate that ZIKV RNA was detected in peripheral nerves (5 out of 7 RM) and spinal cord (4 out of 7 RM) and viral RNA can persist in these tissues for up to 5 weeks post infection, which could lead to long-term neuropathology. Indeed, post-recovery complications from ZIKV infection, primarily GBS or other neurologic manifestations, have been documented in the French Polynesian and subsequent South American outbreaks. GBS is characterized by the presence of certain autoreactive antibodies and immune cells and is often associated with previous infection, including viral infection. It is unknown how either acute or chronic ZIKV infection might contribute to GBS or other symptoms, and warrants further study.
Additionally, viral RNA was detected in tissues of the male and female reproductive tracts, which may have implications for the association of ZIKV infection of pregnant women with aberrant fetal development and sexual transmission. These findings also correspond to a similar observation of ZIKV RNA in the genital tract of a human female [44], and may also suggest a mechanism of a recently described instance of female to male sexual transmission [45]. In male RMs, we were unable to detect viral RNA in the testes, which is, perhaps, surprising given the reports of male to female ZIKV sexual transmission. However, we were able to detect viral RNA in the prostate and seminal vesicles, which may represent a potential reservoir and mode of sexual transmission. In addition, the presence of virus in the bladder and urine suggests virus seeding into the semen in the urethra may also be a possible route of transmission. Further intensive study with regard to sexual transmission is clearly warranted. We also note that the challenge route during sexual transmission may affect the biological outcomes to ZIKV infection including pathogenesis. While our studies were designed to mimic mosquito transmission, further studies to elucidate the effect of route of infection on disease are possible in RM. Viral RNA detected in the vagina and uterus of infected females may also be relevant to the association between ZIKV infection during pregnancy and microcephaly/fetal abnormalities. Interestingly, early results from one such study observed prolonged viremia in females infected during the first trimester of pregnancy, leading the authors to speculate that the fetus may be the source of this virus, or that the placenta may serve as a primary reservoir of ongoing ZIKV infection [22]. Our results suggest that other tissues within the infected dam, such as the female reproductive tract, lymphoid or joint tissue may also be considered as a potential source of persistent virus, given that pregnancy is associated with partial immune suppression, the pregnant animal maybe unable to completely clear viral infection. Again, much more intensive analysis of how ZIKV affects pregnancy in RM will be informative.
Several recent studies have examined ZIKV infection in vivo using various mouse models. These include ZIKV challenge of mice deficient in type I or type I and II interferon (IFN) receptors, in which infection is lethal. Additionally, some strains of immunocompetent mice also appear susceptible to ZIKV infection with regard to displaying detectable viremia and susceptibility of fetuses to developmental defects, although the mechanisms underlying the outcomes in this model are unclear. In type I IFN receptor deficient mice (IFNAR-/- or A129) viral tropism appears broader than we observe in the infected RM. Notably, high viral loads were observed in the brain and testes of these mice, tissues that in the RM, had undetectable or minimal levels of viral RNA. These results suggest that restriction of viral tropism in vivo may be due to the action innate immune factors. Therefore, robust and highly relevant animal models of disease are critical to advancing our understanding of ZIKV disease pathogenesis, immune responses, and potential vaccination and antiviral strategies and these will include both mouse and nonhuman primate models. Although it is a flavivirus, ZIKV clinical and virologic features are distinct and strikingly different from other flavivirus infections, challenging pre-existing assumptions about how this virus behaves in an intact host. Results presented here establish the outstanding potential of the RM ZIKV model for dissecting many of these unique features—specifically effective infectious dose, tissue tropism, fluid compartment infection, and chronicity of infection—in a host that both naturally develops disease and has a fully intact immune system. The magnitude and ongoing expansion of the current ZIKV outbreak calls for rapid initiation of more comprehensive studies to further validate and expand this RM model as well as begin to explore specific in vivo questions that are critical to controlling the ZIKV epidemic.
|
10.1371/journal.pcbi.1002491 | Tension and Robustness in Multitasking Cellular Networks | Cellular networks multitask by exhibiting distinct, context-dependent dynamics. However, network states (parameters) that generate a particular dynamic are often sub-optimal for others, defining a source of “tension” between them. Though multitasking is pervasive, it is not clear where tension arises, what consequences it has, and how it is resolved. We developed a generic computational framework to examine the source and consequences of tension between pairs of dynamics exhibited by the well-studied RB-E2F switch regulating cell cycle entry. We found that tension arose from task-dependent shifts in parameters associated with network modules. Although parameter sets common to distinct dynamics did exist, tension reduced both their accessibility and resilience to perturbation, indicating a trade-off between “one-size-fits-all” solutions and robustness. With high tension, robustness can be preserved by dynamic shifting of modules, enabling the network to toggle between tasks, and by increasing network complexity, in this case by gene duplication. We propose that tension is a general constraint on the architecture and operation of multitasking biological networks. To this end, our work provides a framework to quantify the extent of tension between any network dynamics and how it affects network robustness. Such analysis would suggest new ways to interfere with network elements to elucidate the design principles of cellular networks.
| Multitasking pervades our daily lives. For example, the technological devices that we increasingly rely upon are now engineered with such multifunctionality or “integration” in mind. Similarly, cellular networks also multitask in that they generate multiple, distinct dynamics according to their operating context. Here we show that differences in parameter spaces that underlie different dynamics thus cause a “tension”, which ultimately constrains network operation. In particular, our analysis reveals that tension negatively impacts robustness by reducing accessibility of parameters able to accomplish two tasks and reduces their ability to withstand perturbations. The presence of tension and its negative impact on network robustness represents a fundamental, generic constraint on the operation of different multitasking networks.
| Decades of experimental studies have established detailed “wiring diagrams” of diverse cellular networks. A striking property of many networks is multitasking – the ability to generate different dynamics according to their operating context (Figure 1A). For example, the mitogen-activated protein kinase (MAPK) pathway involving RAF-MEK-ERK responds to epidermal growth factor (EGF) by triggering transient ERK activation in a graded fashion whereas nerve growth factor (NGF) induced sustained ERK in bistable manner [1]. These tasks directly underlie contrasting biological outcomes: EGF induces proliferation whereas NFG induces differentiation into neurons. Another example concerns the p53 stress response network that mediates arrest, death, and DNA repair functions [2]. In response to ionizing radiation, the network generates multiple pulses of p53 with constant amplitude (i.e. digital) [3] whereas UV-radiation generates a single, broad pulse whose amplitude follows a graded dose-response (i.e. analog) [4]. Insight into how distinct p53 tasks translate into biological outcomes is just beginning to emerge [5].
Multitasking networks are speculated to have arisen through successive elaboration on pre-existing “core” processes, representing an evolutionarily feasible route to generate novel biological attributes [6]. Intuitively, reusing a common set of components to multitask can be an economical way to accomplish multiple biological goals. Yet, such a strategy can pose an operational challenge: A dynamic may require network states (each being defined by a set of parameter values) that are ill suited for other dynamics. This concept is related to applications of multi-objective optimization (MOO) algorithms in engineering [7], where two or more, possibly conflicting design aspects are considered. Recently, these approaches have been adopted for biology in problems involving classification, system optimization, and gene regulatory network inference [8]. Here, we use “tension” to describe the difference in parameter spaces for distinct dynamics. Intuitively, tension increases with the number of tasks that a network is charged with as each task invariably requires a different subset of parameter values. In the extreme, tension can constrain a network to the point that few additional changes to the network can be tolerated.
A full understanding of network design principles requires an appreciation of where such tensions can arise within networks, their consequences on the robustness of each dynamic, and the strategies used to overcome them. Thus far, however, such concepts have been neglected in quantitative analysis of natural and synthetic pathways. To this end, we have developed a generic computational framework to allow streamlined examination of these questions. We illustrate the use of this framework by analyzing a well-studied RB-E2F network, which plays a pivotal role in regulating cell cycle entry.
The RB-E2F network has been examined in detail under both normal [9] and pathological [10] circumstances (Figure 1B and supporting text (Text S1)). In quiescent cells, E2F is silenced by RB [9] whereas E2F expression and activity is modulated by growth stimulation through four “modules” – interconnected subsets of the network with a distinct regulatory effect. A sensor module links extracellular growth stimulation and E2F activity: Upon growth stimulation, the MYC level increases and facilitates E2F expression [11] directly and via D-type Cyclins (CYCD), which potentiate kinases (CDK4/6) to inactivate RB. A positive feedback module (PFB) reinforces E2F expression by two routes: E2F can bind to its own promoter and maintain an activated state [12] and E2F increases expression of Cyclin E (CYCE) [13], which activates another RB-kinase (CDK2). A negative feedback module (NFB) consists of E2F-regulated genes that include Cyclin A [14] and SKP2 [15] that inactivate E2F binding and induce proteolysis, respectively. Finally, a repression module (R) consists of MYC-regulated genes that down-regulate E2F expression, which may include microRNAs within the miR-17-92 cluster [16] and the ARF tumor suppressor [17].
Three distinct E2F dynamics underlie the response to growth stimuli, depending on the operating context of the network. First, E2F is bistable with respect to serum. Once activated, E2F remains ON even if serum is reduced below the threshold required to activate E2F [18]. In particular, the serum response of E2F exhibits hysteresis, whereby activation of E2F from the OFF state (by increasing serum) and shutting-OFF from the ON state (by decreasing serum) follow different trajectories (Figure 1B). This property provides a mechanism for cells to enforce two distinct states, quiescence and proliferation [19]: Cells will commit to the cell cycle when a growth stimulus exceeds an activation threshold and to quiescence when signals drop below a maintenance threshold.
Second, E2F exhibits biphasic response to direct MYC stimulation: E2F expression increases with the MYC level when the latter is low, but is repressed when the MYC level is too high [20]. This response restricts the range of MYC levels that can activate E2F. It may represent a safeguard mechanism that allows cells to distinguish physiological levels of MYC induced by serum from transient, potentially oncogenic levels resulting from gene mutation or stochastic gene expression.
Third, in normal cells strongly stimulated by serum, E2F expression exhibits temporal adaptation: It increases to a high level leading up to the end of G1 before being down-regulated as cells enter the S-phase [21]. As E2F controls expression of many genes involved in DNA synthesis [22], adaptive E2F can both promote coherent induction of DNA replication activities and restrict them to a brief period in S-phase. Indeed, precocious or prolonged E2F activity has been shown to cause replicative stress resulting from deregulated DNA synthesis followed by a DNA damage checkpoint [23], [24].
The starkly different dynamics generated by the same network led us to hypothesize the existence of conflicts that constrain its operation. To examine this issue, we probed several questions by modeling: How (dis)similar are the solution set of parameters that underlie different dynamics? What is the relative difficulty in identifying such parameter sets and what properties do they demonstrate in terms of network performance? In short, for a specific set of dynamics, what is the relationship between tension and robustness?
Here, we have developed a generic computational approach to examine these questions (Figure 1C). Candidate parameter sets were used to simulate from the model and assigned a score based upon an objective function (Figure S1A). In a single iteration of the algorithm, randomly initialized parameter sets were subjected to successive rounds of ‘mutation’ followed by scoring. If a solution was identified, the iteration was terminated or it was terminated without a solution after a defined number of consecutive mutations (in this case 100) without an improvement in the objective score.
This analysis allowed us to enumerate parameter sets that satisfy each particular task (i.e. single) or biologically relevant pairings (i.e. dual). For two tasks (e.g., A and B), tension is calculated as the weighted sum of the log-ratio of median parameter values (Figure 1D). In the case that each parameter receives equal weighting (i.e. 1/n, where n is the number of free parameters), tension is the average extent each parameter shifts between single tasks. We evaluate robustness according to the “accessibility” of dual solutions and “resilience” of single-task or dual solutions to parameter perturbation. Accessibility is defined as the fraction of single-task solutions identified as dual. A decrease in accessibility indicates increasing difficulty in locating dual solutions. Resilience is defined as the ability of a solution to maintain some minimal performance after a perturbation (in this case at least 10% of the objective score). This framework can be applied to any kinetic model of cellular networks where objective functions can be quantified.
We first compared the bistable response to serum and the biphasic response to MYC. From 10,000 iterations we identified a large fraction of solutions for each single task (Figure 2A). However, only 146 dual solutions were present amongst 4,541 for hysteresis and 14 dual solutions were present in the 4,878 for biphasic. This result corresponds to a dual-solution accessibility of AHB = 0.017, that is, dual solutions represent 1.7% of the total. The rate of solutions identified per iteration and dual accessibility was similar even when only 500 iterations were performed (Figure S1B), indicating that the result from 10,000 iterations is representative.
Reduced accessibility may reflect tension in the network that arises because single dynamics may adopt disparate states. To examine the correlation between shifts in dynamics and corresponding changes in parameters, we determined the median value of each parameter from all the solutions. By using the values for hysteresis as a reference, we isolated changes specifically associated with biphasic response. According to their influence on each module (i.e., synthesis rates are proportional whereas degradation constants are inversely proportional to module strength), parameters were grouped into four modules (sensor, NFB, PFB, and R). This analysis identified biases in the solution parameters associated with sensor, NFB, and R modules (Figure 2B), whereas changes to the PFB parameters were divergent (see supporting Text S1, Discussion). Note that the overall distribution of NFB values were quite similar between different dynamics despite a change in median (Figure S2A). The changes across all median parameter values resulted in a tension of 0.34 (i.e. average shift in parameter value) between hysteretic and biphasic tasks.
Parameter distributions may be highly irregular, raising the issue of how the median may perform as a summary of each solution set. An alternative approach to compare distributions is to calculate the Kullback–Leibler (KL) divergence (Text S1). Consistent with the results obtained using median values, the largest KL divergence involved parameters of R (Figure S2B). More subtle distances in NFB, Sensor, and NFB were also present. In this case, the tension value (0.08) was calculated as the average KL divergence. All subsequent analyses were done by using the median values.
The difference between the two dynamics can be largely accounted for by the strength of sensor and R modules - the product of free parameters constituting each module (Figure 2C). Given this observation, an effective strategy to reconcile the tension is to dynamically configure these modules: Increasing their strengths would favor biphasic response, while decreasing them would favor hysteresis. In contrast, changes in other modules would be less critical. We term this dynamical ‘network reconfiguration’.
The overlap between R and sensor (Figure 2C), however, also suggests the possibility to accommodate the two dynamics by using common parameter sets, which, by definition, represent dual solutions. We performed 10,000 search iterations using a composite objective function that represents the product of hysteretic and biphasic objectives (see Text S1, Materials and Methods) which allowed us to identify an additional 1,290 dual solutions (Figure 2A). Most dual solutions were concentrated in the overlap between single-solution sets, consistent with the notion that they represent a hybrid of parameters from single dynamics (Figure S2C). To validate the distribution of these dual solutions, we also attempted a search using a dual objective function composed of the sum of individual objectives. In addition, we performed a search with single hysteretic and biphasic solutions as a starting point, mimicking the successive elaboration of network tasks. In each case, the distribution of solutions parameters was virtually indistinguishable (Figure S2 A and C). This supports the notion that the distribution of dual solutions is representative.
Simulations show that a typical dual solution could indeed generate both dynamics (Figure 2D). Consistent with Figure 2C, weakening the R module (by substituting it with the median value from hysteretic solutions) diminished the repression of E2F at high MYC, thus diminishing the biphasic response. In contrast, strengthening the R module (by substituting it with the median value from biphasic solutions) maintained the biphasic response to MYC but eliminated hysteresis by weakening overall E2F response. Weakening the sensor shifted the hysteretic response to higher serum inputs but diminished the E2F levels achieved in response to MYC (Figure S2D); strengthening the sensor eliminated hysteresis and broadened the biphasic response by stimulating an increase in E2F at relatively low doses of input.
A caveat of such dual solutions is their reduced accessibility (Figure 2A). In addition, it is interesting to examine if tension could also impact their resiliency to perturbation. To examine this, we selected fifty representative solutions from each category in the vicinity of their respective medians, subjected each one to 10,000 parameter perturbations, and determined the fraction of perturbations that retained at least 10% of the initial objective score. This analysis revealed that biphasic response was a more resilient property than hysteresis overall (Figure 2E and Figure S2E).
Although the median resiliency of dual solutions was slightly lower than single solutions, this change was not significant, suggesting this tension had a minimal impact on the performance of dual solutions. As such, properly configured sensor and R modules can accommodate both dynamics. This could be achieved by engaging the R module only when MYC is sufficiently high, yet simultaneously enhancing the sensitivity of E2F to MYC stimulation. This notion is consistent with the distinct modes of MYC regulation in physiological and pathological contexts. Physiological stimulation, e.g., by serum, of arrested cells leads to a pulse of MYC that drops to a low level throughout the cell cycle [11], which is unable to trigger the R module. Still, a strong sensor module would enable robust generation of E2F switching behavior despite relatively low MYC levels (second column of Figure 2D). In contrast, more elevated and persistent levels of MYC, due to overexpression or stochastic gene expression, would trigger the R module and result in biphasic response.
Using the same approach, we found that the accessibility of dual solutions involving hysteretic and adaptive dynamics was 7-fold lower compared to biphasic behavior (AHA = 0.0024 compared to AHB = 0.0170) (Figure 3A). This decrease was accompanied by an elevated tension between hysteresis and adaptation (THA = 0.48 compared to THB = 0.34). Compared to hysteresis, adaptation is associated with parameters defining moderately enhanced sensor and R modules, and a drastically stronger NFB module (Figure 3B and Figure S3A). Changes in parameters associated with PFB were without coherent bias (Text S1, Discussion). Consistent with these results, the dominant shift in KL divergence involved NFB parameters (Figure S3B). Furthermore, the tension (average KL divergence) between hysteretic and biphasic dynamics (0.08) is lower than that between hysteretic and adaptive dynamics (0.11). These observations suggest that an effective strategy to reconcile the drastic tension is to dynamically configure these modules, particularly for the NFB: Increasing its strength favors adaptation, while decreasing it favors hysteresis (Figure 3C). Reflecting their ‘hybrid’ nature, dual solutions were concentrated in the overlap between individual dynamics when plotted as a function of sensor and NFB strengths (Figure S3 A and C).
To examine the specific contribution of NFB and sensor in modulating these dynamics, we varied its strength in a typical dual solution. Simulations confirmed its ability to generate both dynamics (Figure 3D). Weakening the NFB module (by substituting it with the median value from hysteresis solutions) eliminated the adaptive response and strengthening it (by substituting it with the median value from adaptive solutions) increased the precision of adaptation, consistent with its requirement for this behavior [25]. Weakening the NFB module also enhanced hysteresis to the point that E2F expression became irreversible (i.e. the solid and dotted curves do not meet at low serum). Yet, strengthening it diminished hysteresis by interfering with maintenance of the E2F ON state upon reduction in serum. The sensor strength had a more general impact (Figure S3D). The sensor strength for the dual solution seemed to be near optimal for hysteresis; either weakening or strengthening it led to almost elimination of hysteresis.
The strong tension between dynamics corresponds to a greatly reduced dual accessibility and suggests that they may be operational over a much restricted parameter space (Figure 3C). Indeed, here tension penalized the performance of individual dual solutions: Dual solutions were significantly less resilient to perturbations than single solutions in maintaining both hysteretic and adaptive dynamics (Figure 3E and Figure S3E).
The drastically reduced accessibility and robustness of dual solutions suggests that they would be ineffective in accommodating both dynamics. Instead, dynamic network reconfiguration is likely critical, which is consistent with the operation of the network: the negative feedback on E2F has a significant time-delay in its operation. In the G1 phase, the Anaphase-Promoting Complex/Cdh1 (APCCdh1) keeps negative feedback from both CYCA [26] and SKP2 [27] low by targeting them for proteasomal-mediated degradation. Upon progression to G1/S, E2F activity increases and induces CYCE - which is resistant to APCCdh1 - engaging sole positive feedback. SKP2 and CYCA levels are eventually allowed to increase through E2F-mediated induction of Emi1 [28] which targets APCCdh1 for destruction. This is reinforced through positive feedback as CYCA itself can also target APCCdh1 for destruction [29]. Delay is also achieved at the transcriptional level through ordered release of Cyclin E and Cyclin A from RB-mediated repression [30]. This temporal coordination has been speculated to enforce a brief time window between DNA replication origin licensing mediated by CYCE and origin deactivation and initiation of DNA synthesis mediated by CYCA [31], [32]. Sequential triggering of positive and negative feedback appears to be a generic, systems-level organizational principle of networks underlying cell cycle control conserved throughout evolution [33], [34]. Our analysis suggests an additional role for the temporal coordination: it represents dynamic network reconfiguration that accommodates robust hysteretic and adaptive E2F responses.
In addition, the tension between dynamics can potentially be alleviated by increasing network complexity. For example, eight E2F members of the E2F family have been identified in mammals; some members can functionally substitute for one another [35]. E2F1 and E2F3 are part of the “activator” subgroup required for cell cycle entry of fibroblasts from quiescence [36]. We wondered if such apparent redundancy could reduce tension. To test this notion, we extended our model to include an additional E2F member (E2F′) expressed in parallel with E2F (Figure 4A and Text S1, Mathematical Model). In particular, the model includes distinct parameters governing production and degradation of each E2F copy. On the other hand, we assumed that the biochemical activity of each E2F copy was indistinguishable and could contribute in an additive manner to overall E2F output (i.e. hysteresis and adaptation) as well as to downstream gene expression (i.e. Cyclin E and Cyclin A) via shared parameters.
The added complexity indeed led to a 3.1-fold increase in dual solution accessibility (A2xE2FHA = 0.0052 compared to AHA = 0.0024) (Figure 4B). This was accompanied by a reduction in network tension with dual E2F (T2xE2FHA = 0.39 versus THA = 0.48) (Figure 4C). This is reflected in the more modest extent to which the NFB and R modules shifted between hysteretic and adaptive dynamics (Figure 4C and Figure S4A) and the greater extent of overlap in their distributions (Figure S4B). Importantly, inclusion of an additional E2F copy was sufficient to increase the resilience of dual solution adaptation such that the median was not significantly different from single task solutions (Figure 4D). In contrast, this additional complexity did not have a significant impact on the resilience of hysteresis associated with dual solutions. Why this fragility of hysteretic dynamics persists in such dual solutions is not clear. Nevertheless, these results are consistent with the notion that increasing network complexity reduces tension and the corresponding penalty on some aspects of robustness.
Quantitative modeling has been widely adopted to examine design principles of biological networks. Many studies have provided important insight into the ways networks generate particular dynamic responses [25], [37]. To date, however, how a multitasking cellular network accommodates different dynamics is poorly understood, despite the recognition of their wide presence and importance. Here we have developed a general approach to quantify tension between different dynamics, which we have applied to a well-established network underlying cell cycle progression. In general, our analysis is consistent with an inverse relationship between tension and the overall robustness of network operation (Figure 5). In the face of moderate tension, common or ‘one-size-fits-all’ parameter sets could be attractive as they avoid the need for additional, possibly complex, mechanisms to coordinate system parameters. However, dynamic network reconfiguration may be critical to resolve strong tension. Though dual solutions exist, there is a pronounced penalty on the accessibility and resilience of these solutions. Our approach is general in that it can be applied to any other network with behaviors that are distinct and quantifiable. In the case where a network demonstrates numerous tasks (including the RB-E2F network), accessibility, tension, and resiliency can be reported by an “adjacency matrix”, reporting all interactions in a pair-wise fashion.
Our findings have several implications for our understanding of the RB-E2F switch as well as a variety of other multitasking networks (Table 1). First, our generic framework provides additional criteria to assess model selection, sometimes favoring choices that are not intuitive. In the case of the RB-E2F network, the relatively high tension between hysteretic and adaptive tasks suggests a critical need for additional mechanisms able to delay negative feedback (i.e. CYCA) or buffer parameter changes (i.e. E2F duplication). Another example involves a study by Ashall et al. [38] concerning how different pulsatile TNF-α input patterns encode unique NF-κB nuclear translocation dynamics (Figure S5A). The authors were prompted to propose an alternative network model wiring when they were unable to find common parameter sets that could satisfy all NF-κB tasks using a traditional model. Using their data, we calculated that the tension between two tasks (“Continuous” and “60 minute”) was reduced from TCon/60 = 1.69 to T′Con/60 = 0.84 in the alternative model along with a corresponding increase in accessibility from ACon/60 = 0 to A′Cont/60 = 0.29 (Figure S5B). What system-wide values of tension and accessibility are across all model parameters remains to be seen. Nonetheless, our study suggests that dynamic shifting of parameters is more desirable from the perspective of robustness.
Second, tension has the potential to affect network evolvability [6]. In particular, coopting additional functions could interfere with pre-existing network dynamics (i.e. partial overlap of solution space), thereby reducing the ability of the network to tolerate additional alterations. For example, Meir et al. [39] modeled the ability of the Notch-Delta signaling network [40] to generate three spatial cell fate patterns – “2-cell”, “7-cell” and “Line” –attributed to the pathway during animal development. They showed that the solution spaces for these tasks were only partially overlapping (Figure S5 C and D): Only 25% of solutions for the “2-cell” tasks could accommodate a “7-cell” pattern while nearly 80% of “7-cell” solutions could also produce “2-cell” patterning corresponding to an accessibility of A2–7 = 0.51 for dual solutions. Also, parameters for “Line” overlap to an even lesser extent with solutions for the two other tasks. From this, the authors speculated that existence of universal parameter sets represent an evolutionarily feasible route towards the goal of achieving novel functions. On the other hand, these same observations offer direct support for our argument that tension reduces robustness and constrains a network's capacity to adopt additional tasks. An intriguing possibility is that dynamic shifting, increased complexity, or other strategies may enable a network such as this to increase its workload [41].
Third, by focusing on the coordination of different tasks, our methodology can provide novel, experimentally testable hypotheses concerning what mechanisms are tied to potential conflicts between dynamics and how they are resolved. For example, Santos et al. [42] showed that the MAPK cascade, consisting of RAF, MEK1/2, and ERK1/2, demonstrates distinct dynamics and contrasting phenotypes in response to EGF and NGF (Figure 1 and Figure S6A). Importantly, EGF stimulated negative feedback between ERK and RAF whereas NGF stimulated positive feedback. The growth-factor context-dependent MAPK topologies are a clear example of tension between dynamics and the functional role of network reconfiguration. Analogous to our results showing that modulating NFB strength could impact hysteresis (Figure 3D), small-molecules used to constitutively suppress and sustain positive feedback could swap ERK dynamics and physiological effects of NGF and EGF. Furthermore, the authors showed that partial activation of positive feedback via interfering RNA (RNAi) generated an intermediate ability of EGF to induce differentiation, suggesting a quantitative relationship between tension and phenotypic outcome.
Another example involves the multifunctional response of the p53 tumor suppressor. Batchelor et al. [4] demonstrated that repeated, digital pulses stimulated by γ-radiation (γ-IR) required WIP1-mediated negative feedback whereas UV radiation generated a single, graded p53 response (Figure S6B). Importantly, suppression of negative feedback by RNAi against WIP1 was sufficient for γ-IR to generate a p53 response characteristic of UV [43]. These observations represent a clear demonstration of tension between dynamics attributed to negative feedback, and its reconciliation through duplication and diversification of the network (i.e. ATM and ATR). In retrospect, our framework provides a rational means to identify such network tension, which may not easily arise from intuition alone or even a deep knowledge of the network, especially when tension arises from subtle and/or multiple parameter shifts.
For the RB-E2F network, our detailed examination of tension and robustness provide experimentally testable hypotheses. First, our results suggest that the strength of negative feedback acting upon E2F is inversely related to the extent of hysteresis. The strength and timing of NFB could be realized through a small-molecule inducible Cyclin A expression construct. Alternatively, premature Cyclin A activity could be achieved through introduction of an N-terminal deletion mutant resistant to APC/C-mediated destruction [44]. The effect of this on the E2F dose-response to serum could be readily achieved using a previously devised fluorescent reporter for E2f1 [45]. Second, this same experimental system could be used to test the hypothesis that additional copies of E2F insulate the hysteretic response from premature or intensified NFB. Finally, our results lead directly to the hypothesis that strong NFB will reduce the robustness of networks able to accommodate both dynamics. Such a question would be best suited using a synthetic biology approach and predicts that circuits with both bistable and adaptive dynamics would arise with relatively mild NFB.
If tension places a fundamental constraint on the operation and architecture of multifunctional networks, it has implications for engineering of synthetic biological systems. To date, most efforts have focused on engineering of gene circuits with limited, dedicated functions. More complex functions can then be realized by integrating well-defined modules [46], [47]. For those functioning in individual cells, however, this strategy is limited by the ability to insulate different modules as well as the inevitable burden they impose upon cells which can undermine desired functionality [48]. As such, it may be more effective to explore strategies that include dynamic network reconfiguration to perform multiple functions in a robust manner. In this case, synthetic biology may take a cue from nature: Rather than attempting to generate an ever-expanding toolkit of biological components, an emphasis will be placed back upon the vast potential in differential regulation of existing entities.
Simulations were performed with Matlab, version R12 (Mathworks, Natick MA) employing the ode15 solver.
|
10.1371/journal.ppat.1006674 | Enterovirus 71 protease 2Apro and 3Cpro differentially inhibit the cellular endoplasmic reticulum-associated degradation (ERAD) pathway via distinct mechanisms, and enterovirus 71 hijacks ERAD component p97 to promote its replication | Endoplasmic reticulum-associated degradation (ERAD) is an important function for cellular homeostasis. The mechanism of how picornavirus infection interferes with ERAD remains unclear. In this study, we demonstrated that enterovirus 71 (EV71) infection significantly inhibits cellular ERAD by targeting multiple key ERAD molecules with its proteases 2Apro and 3Cpro using different mechanisms. Ubc6e was identified as the key E2 ubiquitin-conjugating enzyme in EV71 disturbed ERAD. EV71 3Cpro cleaves Ubc6e at Q219G, Q260S, and Q273G. EV71 2Apro mainly inhibits the de novo synthesis of key ERAD molecules Herp and VIMP at the protein translational level. Herp differentially participates in the degradation of different glycosylated ERAD substrates α-1 antitrypsin Null Hong Kong (NHK) and the C-terminus of sonic hedgehog (SHH-C) via unknown mechanisms. p97 was identified as a host factor in EV71 replication; it redistributed and co-exists with the viral protein and other known replication-related molecules in EV71-induced replication organelles. Electron microscopy and multiple-color confocal assays also showed that EV71-induced membranous vesicles were closely associated with the endoplasmic reticulum (ER), and the ER membrane molecule RTN3 was redistributed to the viral replication complex during EV71 infection. Therefore, we propose that EV71 rearranges ER membranes and hijacks p97 from cellular ERAD to benefit its replication. These findings add to our understanding of how viruses disturb ERAD and provide potential anti-viral targets for EV71 infection.
| Understanding of viral-host interactions is important for learning about viral pathogenesis and providing potential anti-viral targets. The protein quality control system, which consists of ERAD and autophagic degradation, is necessary for cellular homeostasis. Our previous studies and others have demonstrated that autophagy is involved in the EV71 lifecycle, but the role of ERAD remains unclear. In this study, we found that EV71 infection also significantly inhibits physiological ERAD at multiple points, causing ERAD substrates to remain tethered in the ER lumen. When exploring the mechanism of EV71-induced ERAD inhibition, data revealed that EV71-encoded viral proteases 2Apro and 3Cpro are involved, and they target different molecules with different mechanisms. We also found that ERAD component p97 was essential for the EV71 lifecycle, and it redistributed and co-exists with viral protein and other known replication-related molecules in EV71-induced replication organelles. Thus, we found a novel viral-host interaction that provides new potential anti-viral targets for EV71 infection.
| Enterovirus 71 (EV71), which belongs to the Picornaviridae family Enterovirus genus, is a single-stranded positive-sense RNA virus [1]. This pathogen is the causative agent of hand, foot and mouth disease (HFMD), and is especially the major cause of severe HFMD. Since the first report in the United States in 1974, EV71 outbreaks have been reported around the world, particularly in the Asia-Pacific region in recent years [2]. In China, EV71 caused a severe HFMD outbreak in Fuyang, Anhui province in 2008, and has since become an epidemic problem [3]. The frequency and severity of HFMD have shown an increased annual trend and pose a serious threat to children’s health and social stability in China [4]. However, no effective therapy is currently available for the treatment of this infection and more studies are needed to elucidate the pathogenesis of EV71.
The genome of EV71 encodes eleven proteins, including four viral capsid proteins (VP1–VP4) and seven non-structure proteins (2A–2C, 3A–3D) [1,5]. Among these viral proteins, viral proteases 2Apro and 3Cpro have been demonstrated play important roles in virus-host interaction and EV71 pathogenesis. EV71 2Apro has been reported to hijack host cell gene expression by cleaving the eukaryotic initiation factor 4G (eIF4G) and poly(A)-binding protein (PABP) [4,6–8]. It has also been reported to antagonize host innate immunity by down-regulating interferon receptor 1 (IFNAR1) and cleaving mitochondrial antiviral signaling protein (MAVS) [4,9]. EV71 3Cpro has been reported to mediate viral immune-evasion by targeting many key components in host innate immunity, including TRIF, IRF7, IRF9, and the RIG-I/IPS-1 complex [5,10–13]. Moreover, 3Cpro has also been reported to disrupt host cell gene expression by cleaving CstF-64 [14]. In general, previous studies concerning EV71 viral proteases have focused on innate immunity and gene expression.
In mammalian cells, approximately one-third of the proteins are assembled into mature proteins in the ER [15,16]. This process is tightly monitored by the ER protein quality control (ERQC) system, which is a comprehensive maintenance mechanism for the highly crowded proteins in the ER. This system ensures that only correctly folded and assembled proteins reach their ultimate destination [15,17]. The ERQC system achieves its function via several molecular chaperones and two degradation pathways: the autophagy-lysosome-mediated autophagic degradation and ubiquitin-proteasome-mediated ER-associated degradation (ERAD) pathways [17–20].
Autophagy removes protein aggregates and damaged organelles in double membrane vesicles and degrades them in autolysosomes [21,22]. Several viruses utilize and alter cellular autophagy to facilitate their own replication, including hepatitis C virus (HCV), coronavirus, Dengue virus, influenza A virus, poliovirus (PV), and coxsackievirus B3 (CVB3) [21,23–25]. Huang et al. and our previous studies demonstrated that EV71 can also induce cellular autophagy and exploit autophagy for its own replication [23,26,27]; however, it remains unknown whether ERAD is also affected and involved in EV71 replication.
ERAD is a process that facilitates the degradation of terminally misfolded, misassembled, and metabolically regulated proteins in the ER by retro-translocating them to the cytosol for degradation by the ubiquitin-proteasome system [16,17,28,29]. This process consists of four coupled steps: (i) substrate recognition; (ii) retro-translocation; (iii) ubiquitination; and (iv) 26S proteasome-mediated degradation [16,30,31]. Since ERAD is a key cellular machinery for ensuring correct cell function, it is unsurprising that viruses can manipulate this process for their own benefit. Previous studies demonstrated that different viruses can affect and exploit the ERAD process in different manners [28,31–34]. There are four reasons why viruses exploit ERAD. First, to escape the immune surveillance system by eliminating immune molecules; examples include herpes virus and human immunodeficiency virus (HIV) [28,32]. Second, viruses take advantage of ERAD to achieve the membrane penetration of intact viruses from the ER to the cytosol, such as simian virus 40 (SV40), human BK virus, and murine polymavirus [28,32]. Third, viruses promote ERAD tuning and hijack EDEMosomes to support their replication, including coronaviruses such as severe acute respiratory syndrome coronavirus (SARS-CoV) and mouse hepatitis virus (MHV) [32,35–37]. Finally, viruses can activate ERAD to degrade viral glycoproteins and thereby reduce the viral particle and maintain a chronic infection status; examples include hepatitis B virus (HBV) and hepatitis C virus (HCV) [32,38,39]. However, despite the numerous studies mentioned above, no reports have investigated the relationship between picornaviruses and ERAD.
Here, we demonstrated that EV71 infection inhibits cellular ERAD processes at multiple key ERAD molecules via its proteases 2Apro and 3Cpro, and ERAD component p97 is involved in EV71 replication. This study reveals a novel relationship between EV71 and cellular ERAD and thus sheds light on the pathogenesis of EV71.
ERAD and autophagic degradation are two facets of the host protein quality control system [18]. These biological processes are exploited by various infectious pathogens as survival and proliferation strategies [21,28]. Our previous study also demonstrated that EV71 can take advantage of host autophagy for its own proliferation [27]. However, it remains unknown whether EV71 can affect and modulate the host ERAD machinery. To address this question, we first categorized the ERAD substrates according to their different chaperone systems and established stable cell lines ectopically expressing these substrates. There are two types of ERAD substrate according to their varied chaperone system: calnexin (CNX)/calreticulin (CRT)-dependent substrates and BiP-dependent substrates [16,40,41]. The different types of substrate may be disposed by distinct ERAD sub-pathways and different molecules may be involved.
We first investigated the degradation of two well-characterized CNX/CRT-dependent ERAD substrates: the C-terminus of SHH (SHH-C) and α-1 antitrypsin Null Hong Kong (NHK) [42–47]. Rhabdomyosarcoma (RD) cells stably expressing SHH and NHK were mock infected or infected with EV71 and treated with the protein synthesis inhibitor cycloheximide (CHX) for different times (according to their pre-tested half-life). Western blotting was then used to measure the expression of the substrates. The levels of both SHH-C and NHK were gradually decreased in CHX-treated mock-infected cells in a time-dependent manner. However, the decrease under CHX chase was significantly inhibited in EV71-infected cells (Fig 1A and 1B), indicating that EV71 may inhibit ERAD of SHH-C and NHK.
To further confirm these inhibitory effects, SHH and NHK stable cell lines were treated with tunicamycin (Tun), an inhibitor of N-glycosylation modification, and the fate of already synthesized glycosylated substrates was monitored. In mock-infected cells stably expressing SHH and NHK, treatment with Tun led to a time-dependent decrease in glycosylated SHH-C and NHK, which was accompanied by an increase in the levels of their newly synthesized non-glycosylated forms. However, in EV71 infected cells the degradation of glycosylated SHH-C and NHK was dramatically inhibited (S1A and S1B Fig), confirming the inhibitory effects of EV71 on the ERAD of glycosylated SHH-C and NHK. However, it is worth noting that the de novo synthesized non-glycosylated forms of SHH-C and NHK were detectable in mock-infected cells, but not EV71 infected cells (S1A and S1B Fig), indicating that the de novo synthesis of SHH and NHK was inhibited by EV71 infection [6–8].
Next, the ERAD of BiP substrates during EV71 infection was evaluated. Transthyretin D18G (the 18th D amino acid mutated to G; TTR D18G) and non-secretory immunoglobulin kappa-type light chain (NS1 κ LC) were selected as representative. TTR is a non-glycosylated soluble secretory protein and TTR D18G is the most destabilized mutant of TTR that is subject to ERAD [48,49]. NS1 κ LC is an unassembled immunoglobulin light chain that is degraded through ERAD [50]. RD cells stably expressing TTR D18G and NS1 κ LC were treated with CHX and their degradation was assessed. The degradation of both substrates was remarkably inhibited in a time-dependent manner during EV71 infection (Fig 1C and 1D), suggesting that EV71 infection also inhibits the ERAD of BiP substrates. It is worth noting that high molecular weight bands of TTR D18G were visible (Fig 1C), which we demonstrated were glycosylated forms of TTR D18G by treating cell lysates with glycosidase PNGase F (S2 Fig). In EV71-infected cells, the molecular weight of glycosylated TTR D18G was decreased, and we speculate that is due to extensive mannose trimming.
Since the above experiments were all performed using stable cell line ectopically expressing different substrates, we next assessed whether EV71 infection could inhibit the degradation of endogenous substrates. Therefore, the degradation of core-glycosylated CD147 (CG), a reported constitutive endogenous substrate was examined during EV71 infection [51]. The CHX chase assay showed that CD147 (CG) was gradually degraded in a time-dependent manner in CHX-treated mock-infected RD cells, However, the degradation was partially inhibited by EV71 infection (Fig 1E), suggesting that EV71 infection also inhibited the ERAD of cellular endogenous substrates.
Considering that EV71 can induce cell apoptosis and that the above experiments were performed with EV71 infection or infection combined with CHX treatment up to 17 h, we checked for cell apoptosis in RD cells under these conditions. The results showed that the proportion of apoptosis was 28.6% in cells infected with EV71 for 17 h, and the combined treatment with CHX for the last 8 h only slightly upregulated the proportion to 32.2% (S1C Fig).
Taken together, the above results demonstrate that EV71 infection inhibits the ERAD of different types of substrates, including both CNX/CRT-dependent glycosylated substrates and BiP-dependent non-glycosylated substrates and cellular constitutively and endogenously expressed substrates.
Since ERAD is a process that detects misfolded proteins in the ER and extracts them to the cytosol for proteasomal degradation [16,32], we next assessed the location of ERAD substrates in EV71-infected cells. First, SHH-C was used as the substrate. SHH-C is a glycosylated protein that undergoes deglycosylation when retro-translocated to the cytosol. The accumulation of deglycosylated SHH-C when the proteasome is inhibited reflects its retro-translocation degree [43–45]. In cells not treated with the proteasome inhibitor MG132, the deglycosylated SHH-C was not detected in either mock- or EV71-infected cells. However, when cells were treated with MG132, deglycosylated SHH-C was visible as a low-molecular weight band in mock-infected cells, but was barely detectable in EV71-infected cells (Fig 2A). This suggests that the retro-translocation of SHH-C was inhibited by EV71 infection and that SHH-C could be trapped inside the ER lumen.
To further confirm the above conclusion, a previously reported dislocation-dependent reconstituted GFP (drGFP) assay was used to evaluate different substrates, and the principle of this system is illustrated in Fig 2B. Briefly, the GFP molecule is split into two fragments: the C-terminal β-strand (S11) and the remaining 10β strands (S1–10). The S11 is linked to ERAD substrates expressed in the ER lumen, and S1–10 is expressed in the cytosol. When S11-tagged substrates are retro-translocated into the cytosol, S11 is reassembled with S1–10, and the resulting GFP signal is detected when the coupled proteasome degradation is inhibited (Fig 2B) [52]. First, the drGFP assay was used to determine the location of SHH-C in EV71-infected cells. The results showed that in cells not treated with MG132, no GFP signal was observed in either mock-infected or EV71-infected cells. However, the reconstituted GFP signal could be detected in MG132-treated mock-infected cells, but not in EV71-infected cells (Fig 2C). This suggests that EV71 inhibited the retro-translocation activity of SHH-C and that this substrate was trapped inside the ER during infection. The same method was also used to test the retro-translocation activity of NHK and TTR D18G during EV71 infection, and similar results were obtained (Fig 2D and 2E). The drGFP assay was not used to evaluate substrate NS1 κ LC since previous studies reported that NS1 κ LC was retained in the ER lumen and not translocated to the cytosol when cells were treated with a proteasome inhibitor [53]. The above experiments were performed with RD cells expressing different substrates infected with EV71 up to 18 h, and we also used flow cytometry to monitor cell apoptosis under this situation. The results showed that EV71 infection for 18 h caused apoptosis in 30.4% cells, and combined treatment with MG132 in the last 8 h slightly downregulated this proportion (S3 Fig).
Taken together, the above results demonstrated that both CNX-dependent glycosylated substrates and BiP-dependent non-glycosylated substrates were trapped inside the ER during EV71 infection. Therefore, they could not be retro-translocated to the cytosol to undergo subsequent proteasomal degradation.
The above results demonstrated that EV71 inhibits the degradation of different ERAD substrates. Since different substrates are degraded through distinct sub-pathways, it is likely that EV71 inhibits ERAD at multiple targets or at one shared key point. To clarify the specific molecular mechanisms, ERAD-related molecules were categorized by their different functions and a kinetic study was performed to measure their expression during EV71 infection. The molecules examined were classified into four categories: (1) recognition factors (calnexin, calreticulin, BiP, EDEM1, OS9, and XTP3-B); (2) retro-translocation factors (SEL1L, Herp, Derl1, and Derl2); (3) ubiquitination factors (Hrd1, gp78, RNF5, Ubc6e/UBE2J1, and Ubc7/UBE2G2); and (4) proteasomal degradation factors (VIMP, UBXD8, p97, Ufd1, and Npl4) [16,30,31] (Fig 3A). As expected, EV71 infection downregulated the expression of several molecules in a time-dependent manner, including Herp, Hrd1, Ubc6e, VIMP, and UBXD8 (Fig 3B). Among these, two bands that seemed like cleavage products of Ubc6e could be detected around 25–26 kDa in molecular weight, and their intensity increased as the infection proceeded. Cleavage bands were also detected in the blot of UBXD8, but not in the blots of Herp, Hrd1, and VIMP (Fig 3C).
Previous studies reported that three E2 ubiquitin-conjugating enzymes function in mammalian ERAD: Ubc6e/UBE2J1, UBE2J2, and UBE2G2 [54]. Ubc6e forms an E2-E3 pair with Hrd1 and is considered the principal E2 in cellular ERAD [54–56]. To further clarify the precise mechanism by which Ubc6e is cleaved, we investigated whether EV71-encoded viral proteases 2Apro and 3Cpro participate in this process. These viral proteases are responsible for cleaving poly-protein precursors to obtain mature viral proteins, and increasing amounts of evidence have demonstrated that they could cleave various host factors to facilitate viral replication [1,4,5,13]. First, 293T cells were transfected with plasmids encoding EV71 3Cpro or a protease-dead mutant of 3Cpro (C147S), and Ubc6e cleavage was detected by western blotting. Overexpressed 3Cpro, but not 3Cpro(C147S), could cleave Ubc6e in a dose-dependent manner, and the cleavage bands were the same molecular weight as in EV71-infected cells (Fig 4A). This result was also achieved by in vitro cleavage assay with recombinant 3Cpro and its protease-dead mutant 3Cpro(E71A) (Fig 4B). We also tested the role of 2Apro in Ubc6e cleavage, and the result showed that no cleavage bands were detected in 2Apro-transfected cells (S5 Fig). This suggests that EV71 3Cpro cleaves Ubc6e during infection.
Next, we assessed whether 3Cpro is the causes of ERAD inhibition. RD cells stably expressing SHH were transfected with increasing doses of plasmids encoding EV71 Flag-tagged 3Cpro and western blotting were performed to monitor the degradation of SHH-C under CHX chase. The results showed that SHH-C degradation was inhibited gradually with the increasing expression of EV71 3Cpro (Fig 4C). However, the inhibitory effects were not as potent as EV71 infection. This might be due to the limited transfection efficiency of the overexpressed EV71 3Cpro, or it could be because Ubc6e cleavage is only one of the factors that inhibit ERAD, and other factors could also be involved. Overall, this result suggests that EV71 3Cpro-induced Ubc6e cleavage could be a crucial mechanism by which EV71 inhibits ERAD.
Then we identified the EV71 3Cpro cleavage sites on Ubc6e. 293T cells were co-transfected with plasmids encoding GFP-3C and C-terminal FLAG-tagged Ubc6e, and Ubc6e cleavage was monitored using antibodies against Ubc6e and FLAG (Ubc6e antibody is a mouse monoclonal antibody with unknown epitope). The results showed that the ~25–26 kDa cleavage bands could be recognized by the Ubc6e antibody but not the FLAG antibody (Fig 4D), indicating that these cleavage bands were N-terminal cleavage fragments from Ubc6e. The molecular weight of the cleavage fragments suggested that the cleavage sites might be located at the region of amino acids 200–300 of Ubc6e (Fig 4E). Since picornavirus 3Cpro preferentially cleaves glutamine-glycine (Q-G), glutamine-alanine (Q-A), and glutamine-serine (Q-S) bonds in viral polyproteins and cellular targets [11,57], a panel of site-directed mutants with Q mutated to A within the 200–300 amino acid (aa) region was constructed. Then, 293T cells were co-transfected with the GFP-3Cpro and Ubc6e/Ubc6e mutants to map the cleavage sites. As illustrated in Fig 4F, all the mutants could be cleaved by 3Cpro as before. However, the cleavage bands from mutant Q219A were shifted to a molecular weight of ~30 kDa (Fig 4F). This indicates that the 219th glutamine of Ubc6e (Q219) is one of the cleavage sites and that another cleavage site(s) exists that is located closer to the C-terminus of Ubc6e. To further identify the remaining cleavage sites, a new panel of double-site mutants (Q to A) was constructed within the 219–300 aa region based on the Q219A mutant. These series mutants revealed that all the double-site mutants could be cleaved by 3Cpro; however, the cleavage bands were changed in cells transfected with mutant Q219Q260A and Q219Q273A. This indicates that both Q260 and Q273 are the cleavage sites of 3Cpro on Ubc6e (Fig 4G). To further confirm this result, a third round of screening was conducted using triple-site Q to A mutants that were generated based on Q219Q260A including Q219Q260Q262A and Q219Q260Q273A. The results showed that triple-site mutant Q219Q260Q273A was totally resistant to 3Cpro cleavage; thus, Q260 and Q273 were identified as the second and third cleavage sites (Fig 4H). Taken together, the above results demonstrate that EV71 3Cpro cleaves Ubc6e at Q219G, Q260S, and Q273G. These sites are all located on the cytoplasmic side of Ubc6e, and cleavage at these points will lead to the release of the E2 catalytic domain of Ubc6e into the cytosol and inactive Ubc6e at the ER membranes. When all the cleavage sites were identified, we re-analyzed the results of Fig 4E–4G and speculated that each cleavage fragment was accompanied by a phosphorylated form with a lower migrating speed [58,59]. We also confirmed this by treating cell lysates with phosphatase and detected the cleavage of Ubc6e; the results showed that both the dual-band of Ubc6e and cleaved Ube6e changed to a single-band when treated with phosphatase (Fig 4I).
Next, the mechanism behind the decreased expression of Herp and VIMP was examined. Herp, whose expression is strongly induced by the unfolded protein response (UPR), is involved in the turnover of ERAD substrates [60–62]. It participates in building the dislocation machinery around Hrd1 and itself has a very fast turnover [44,63–65]. To further explore the reason behind the Herp reduction, we first assessed whether the reduction of Herp was caused by accelerated degradation or blocked synthesis. Mock- and EV71-infected RD cells were treated with the ER stress inducers thapsigargin (Tg) and tunicamycin (Tun) to induce Herp expression, and the degree of upregulation was compared between mock- and EV71-infected cells. Tg and Tun could induce Herp expression efficiently in mock-infected cells, especially when the cells were treated with ER stress inducers combined with the proteasome inhibitor MG132. However, the same treatment induced less Herp expression in EV71-infected cells, implying that EV71 infection may inhibit the de novo synthesis of Herp and that Herp is degraded via proteasomal degradation (Fig 5A). Cell apoptosis in EV71-infected cells and infected cells combined with different chemical treatments as shown in Fig 5A were also checked, and no obvious differences were observed between different groups (S6 Fig).
Since EV71 can inhibit host gene expression at the protein translation level [4], we next assessed whether the EV71-induced Herp reduction was caused by blocked protein translation. To test this hypothesis, the impact of EV71 on mRNA transcription was checked. Mock- and EV71-infected RD cells were treated with Tg and Tun, and Herp mRNA expression was assessed using real-time PCR. Tg and Tun induced Herp expression efficiently in mock-infected cells, but the extent of upregulation was reduced by 2/3 in EV71-infected cells. However, there were no differences of basal Herp expression between mock- and EV71-infected cells (Fig 5B). This suggests that EV71 might inhibit basal Herp synthesis at the translational level and inhibit Tg- and Tun-induced Herp synthesis at both the transcriptional and translational levels.
EV71 2Apro can inhibit protein translation in host cells by cleaving many host factors related to the translation process, including eukaryotic translation initiation factor (eIF4GI), eIF4GII, and poly-A binding protein (PABP) [4,6–8]. Therefore, we assessed whether EV71 2Apro blocked the protein synthesis of Herp in a previously reported BSRT7 cell line that stably expresses T7 RNA polymerase and effectively avoids the difficulties in expressing 2Apro in eukaryotic cells [4,8,66]. Using GFP and 2A protease-dead mutant 2A(C110A) as the control, Tg and Tun induced Herp expression to a much weaker extent in 2A-transfected cells compared with control cells (Fig 5C), suggesting that EV71 2Apro could inhibit Herp synthesis.
Next, the mechanism of VIMP reduction during EV71 infection was investigated. VIMP is a valosin-containing protein (VCP)-interacting membrane protein [65]. It is an important component of the ERAD complex because it can recruit p97 and other cofactors to the ER membrane for retro-translocation [67–69]. A CHX chase assay revealed that VIMP is a short-lived protein, similar to Herp (S4 Fig). Therefore, it is possible that EV71 inhibits VIMP and Herp expression via a similar mechanism. To confirm this, the BSRT7 cell line was transfected with EV71 2A, and VIMP expression was monitored at both the mRNA and protein levels. The results showed that there were no differences of VIMP mRNA expression between GFP and 2A-transfected cells (Fig 5D). However, the protein expression of VIMP was significantly downregulated in 2A-transfected cells, but not GFP and 2A(C110A)-transfected cells at the protein level, in a dose-dependent manner (Fig 5E), suggesting that the EV71 2Apro also inhibits the protein synthesis of VIMP.
Taken together, the above results demonstrate that EV71 2Apro inhibits protein synthesis of Herp and VIMP. Although 2Apro of picornaviruses is well known to inhibit cap-dependent host cell translation by cleaving several host factors related to the translation process [4,5,7,8], we thought that its effect on host cell protein expression mainly embodied on short-lived proteins, like Herp and VIMP in this study. We also included c-MYC, another reported short-lived protein [70,71], and p97, which is not short-lived, as controls and detected their expression in 2A- and 2A(C110A)-transfected cells. As expected, the expression of c-MYC but not p97 was downregulated in 2A-transfected cells (Fig 5E).
We next evaluated of Herp and Ubc6e in the degradation of different ERAD substrates. First, their role in the degradation of SHH-C was examined. RD cells stably expressing SHH were transfected with siRNA to silence Ube6e and Herp. The cells were then chased with CHX under mock- or EV71-infection conditions and the degradation of SHH-C was assessed by western blotting. Knocking down Ubc6e but not Herp significantly inhibited the degradation of SHH-C under CHX chase in RD cells (Fig 6A), suggesting that the ERAD of SHH-C was Ubc6e-dependent and Herp-independent. It is worth noting that silencing Ubc6e substantially upregulated the expression of Herp, suggesting that Ubc6e is a critical modulator of Herp degradation. Next, the same method was used to evaluate the role of Ubc6e and Herp in NHK degradation. The results differed to those of SHH-C, since knocking down either Ubc6e or Herp inhibited the degradation of NHK (Fig 6B). This suggests that both Ubc6e and Herp are required for its degradation. Moreover, since the E2 responsible for NHK ERAD has not yet been identified, this study strongly suggests that Ubc6e is also the E2 for NHK degradation.
Next, the contribution of Ubc6e and Herp to the ERAD of TTR D18G and NS1κ LC was validated. The depletion of Ubc6e significantly inhibited the degradation of non-glycosylated TTR D18G and upregulated the basal intracellular expression of glycosylated TTR D18G (Fig 6C), suggesting that Ubc6e is a key element for the degradation of TTR D18G. Since knocking down Herp also inhibited the degradation of non-glycosylated TTR D18G (Fig 6C), these data suggest that Herp is also involved in the ERAD of non-glycosylated TTR D18G. Regarding NS1κ LC, the results showed that knocking down either Ubc6e or Herp inhibited the degradation of NS1κ LC in CHX-chased cells and obviously upregulated the basal expression of NS1κ LC (Fig 6D). This suggests that both Ube6e and Herp were also essential to the ERAD of NS1κ LC.
Finally, we determined whether the degradation of the endogenous substrate CD147 (CG) requires Ubc6e and Herp. Similar experiments as those described above were performed. As illustrated in Fig 6E, knocking down either Ubc6e or Herp attenuated the degradation of CD147 (CG), suggesting that Ubc6e and Herp were also involved in the ERAD of the endogenous ERAD substrate CD147 (CG).
Taken together, these data demonstrate that Ubc6e functions as a key E2 to play a role in the ERAD of all the substrates tested in this study. In addition, Herp is required for the ERAD degradation of NHK, TTR D18G, NS1κ LC, and endogenous CD147, but not SHH-C. Herp differentially participates in the degradation of different glycosylated ERAD substrates SHH-C and NHK, suggests that factors other than the binding chaperone and glycosylation determine the substrate specificity of Herp.
In this study, we extensively investigated the impact of EV71 on ERAD and reported a novel relationship between the virus and cellular ERAD. These findings are summarized in Fig 10.
In previous studies, all the reported relationships between viruses and cellular ERAD are common in that the viruses use the ERAD machinery for their own benefit [28,31–33]. However, EV71 uses a totally different strategy to influence cellular ERAD. It inhibits ERAD via viral proteases, thoroughly suppresses physiological ERAD at multiple points. EV71 then hijacks host factor p97 and utilizes ER-derived membrane components from the disabled ERAD machinery to form its own replication organelles. In this study, we found that EV71 viral protein expression and viral RNA replication did not change in Ubc6e or Herp knockdown cells, although degradation of ERAD substrates was inhibited. We thought this was because the damage to ER caused by Ubc6e or Herp knockdown was relatively mild when compared to EV71 infection, and the ER membranes were not rearranged under this situation. However, in EV71 infected cells, ERAD was inhibited at multiple points, and ER homeostasis suffered serious damage. This will inevitably cause changes to ER and lead to membrane rearrangement, allowing viruses to take advantage of ER membranes and ER-related host factors more conveniently [76,85].
This study revealed novel functions for the EV71 proteases 2Apro and 3Cpro. The picornavirus protease influences many cellular processes and previous studies have mainly focused on innate immunity and gene expression [5,13,86]. To our knowledge, this is the first study to reveal a role for viral proteases in the ERAD process. However, future studies are needed to determine whether other viral proteases have the same function.
During ERAD, the E2 ubiquitin-conjugating enzyme and E3 ligase work cooperatively to ubiquitinate substrates. To date, three E2s (Ubc6e/UBE2J1, UBE2J2, and UBE2G2) and five principal E3s (Hrd1, gp78, RMA1, TEB4, and TRC8) have been identified in the mammalian ERAD system [29,54–56,87]. Among these, Hrd1 forms an E3-E2 pair with Ubc6e and gp78 forms an E3-E2 pair with UBE2G2. These two E3-E2 pairs are considered the main executors of substrate ubiquitination in ERAD and they have distinct substrate specificities that are determined by the E3s [56]. The current results revealed that EV71 infection caused both Ubc6e cleavage and Hrd1 downregulation and thus totally destroyed the Hrd1-Ubc6e E3-E2 pair. We also investigated the gp78-UBE2G2 pair, but there were no obvious changes in these two molecules during EV71 infection (Fig 2B), indicating that the Ubc6e-Hrd1 E2-E3 pair is the key ubiquitination element targeted by EV71 during infection. We attempted to reconstitute a functional ERAD system in EV71-infected cells by co-transfecting cells with a 3Cpro-resistant mutant of Ubc6e (Q219Q260Q273A) and WT Hrd1. However, this could not rescue the inhibited ERAD process (S8 Fig), suggesting that EV71 inhibits ERAD in a comprehensive manner that involves multiple factors. The decrease of Hrd1 and the cleavage of UBXD8 found in our study are examples. We also checked the role of EV71 2Apro and 3Cpro in the cleavage of UBXD8 but the results excluded their role (S4 Fig) and our future work will focus on the specific mechanisms.
In this study, we identified Ubc6e as the key E2 ubiquitin-conjugating enzyme in EV71-disturbed ERAD. Knockdown of Ubc6e by siRNA inhibited the degradation of all ERAD substrates we tested. However, in one recent study, Hagiwara et al. reported accelerated degradation of NHK in Ubc6e−/− cells because Ubc6e could downregulate ERAD enhancers [88]. This discrepancy might be caused by the different silencing efficiencies of the methods used by the two groups. We used siRNA to knock down Ubc6e while Hagiwara et al. used Ubc6e knock-out MEFs. As described by Hagiwara et al. themselves, very little UBC6e suffices to properly regulate ERAD enhancers [88]. When using siRNA, knockdown of Ubc6e was not complete, and the remaining Ubc6e was sufficient to maintain the proper expression level of ERAD enhancers. Moreover, in Ubc6e−/− cells, other E2s may compensate in the Hrd1 complex and substitute Ubc6e to degrade ERAD substrates.
Okuda-Shimizu et al. and Kny et al. reported conflicting results about the role of Herp in the degradation of different substrates [42,50]. The current study explored the involvement of Herp in different ERAD pathway and the results are consistent with Kny and colleagues. Specifically, Herp was involved in the ERAD of the BiP substrates NS1 κ LC and TTRD18G and the calnexin substrate NHK, but not another calnexin substrate SHH-C. This suggests that other signatures but not the binding chaperone and glycosylation determine the substrate specificity of Herp. As a critical constituent of ERAD, its turnover is mediated by ERAD tuning, and this is a strategy by which Herp can regulate ERAD [64,89]. A previous study demonstrated that UBE2G2-gp78 is responsible for Herp degradation during the recovery from ER stress [63]. However, the mechanism by which Herp is degraded under physiological conditions is unknown. In the current study, Ubc6e was identified as the key E2 that participates in Herp degradation since silencing Ubc6e increased the intracellular levels of Herp (Fig 6A–6E) and silencing UBE2G2 had no effect (S9A Fig). We also tried to identify the E3 partner of Ubc6e that participates in Herp degradation. Although the candidate E3s included Hrd1, gp78, and RNF5, we unfortunately could not draw a definite conclusion because Herp was not upregulated when any of them were silenced (S9A–S9C Fig). Nevertheless, we speculate that Hrd1 is involved in the degradation of Herp because overexpressing a dominant-negative mutant of Hrd1 (C329S) increased Herp expression compared with WT Hrd1 (S9D Fig). However, it is unclear why silencing Hrd1 had little impact on Herp whereas overexpressing Hrd1 dramatically increased Herp levels. We postulate that Hrd1 and its association with Herp is a key determinant of intracellular Herp expression. Specifically, when Hrd1 is silenced Herp is downregulated, but when Hrd1 is overexpressed Herp is upregulated. Future studies are needed to verify this hypothesis and clarify the specific mechanism by which Herp is degraded.
Previous studies demonstrated that p97 is a host factor for viral replication. Arita et al. demonstrated that p97 is required for poliovirus replication and is involved in the cellular protein secretion pathway [73]. Panda et al. identified p97 as a conserved regulator of Sindbis virus entry [90]. A recent study demonstrated that p97 is essential to HCV replication and proposed that p97 is involved in the assembly of HCV replicase [91]. In addition, Wu et al. identified p97 as a cellular factor involved in EV71 replication using a genome-wide RNAi screen [92]. The current study proposed that EV71 hijacks p97 from the disabled ERAD machinery, identified p97 as a host factor for EV71 replication, and demonstrated that p97 co-localized with the viral protein 2C on EV71-induced ROs. This study supports the role of p97 in the viral lifecycle and discloses the origin of the p97 that participates in the formation of ROs. However, it remains unclear how p97 functions in ROs. We propose that p97 might participate in EV71 replication complex assembly and membrane rearrangement via its AAA+ATPase activity.
In this study, we tried to distinguish EV71-induced ROs from EDEMosomes. According to previous reports, virus-induced ROs originating from EDEMosomes have some common features: they are double-membraned vesicles coated with EDEM1 and nonlipidated LC3; silencing LC3 significantly inhibits virus propagation [32,35,37,93,94]. However, EV71-induced ROs we observed under electron microscopy included both single- and double-membrane vesicles, and most were single-membrane vesicles. We also checked the distribution of EDEM1 in EV71-infected cells, and the results showed that it did not change during EV71 infection (S10 Fig). Therefore, we conclude that EV71-induced ROs differ from EDEMosomes and that EV71 influences the cellular ERAD machinery via a totally novel and different mechanism.
Previous studies revealed that picornavirus formed replication organelles at ER-Golgi interface and took advantages of membranes and host factors of both ER and Golgi origin [74–79,82,83,95–99], but it was unclear why picornavirus tend to choose these specific elements of the host cells for their replication and how they are utilized. The current study found that EV71 inhibited cellular ERAD, a critical function through which the ER maintains homeostasis. Inhibiting ERAD inevitably causes ER deformation, which may lead to membrane rearrangement and allows viruses to take advantage of host membranes and host factors more conveniently.
Taken together, this study demonstrated a novel relationship between EV71 and the cellular ERAD system. p97 was identified as a new host factor that is essential for EV71 replication; it redistributed and co-localizes with the EV71 viral proteins in EV71-induced ROs. These findings provide potential targets for the anti-viral treatment of EV71 infections.
Rhabdomyosarcoma (RD) cells and human embryonic kidney 293T (HEK-293T) cells were purchased from ATCC. They were cultured in MEM (Modified Eagle’s medium) and DMEM (Dulbecco’s modified Eagle’s medium), respectively, supplemented with 10% fetal bovine serum (FBS) and penicillin (100 units/ml)/streptomycin (100 mg/ml). BSRT7 cells were described in our previous study [4] and cultured in DMEM supplemented with 10% FBS and 1 mg/ml G418. All cells were maintained at 37°C in a humidified atmosphere of 5% CO2 and 95% air. EV71 is a Fuyang strain (GenBank accession no. FJ439769.1) and was propagated in RD cells. RD cells stably expressing SHH, NHK, TTR D18G, and NS1 κ LC were obtained by transfecting RD cells with the corresponding plasmids followed by selection using 1 mg/ml G418.
The following antibodies were used in this study: anti-β-actin (A5441), anti-GFP (GSN24), anti-HA (H6908), anti-FLAG (A8592), anti-V5 (V8012), anti-RNF5 (SAB2701502), anti-EDEM1 (E8406), anti-OS9 (SAB4200021), anti-Derl1 (D4443), anti-Derl2 (D1194) and anti-calnexin (C4731) were purchased from Sigma-Aldrich; anti-Hrd1 (12925S), anti-Npl4 (13489S), anti-BiP (3183S), anti-calreticulin (2891S), anti-VIMP (15160S), anti-eIF4GI (2858S), and anti-gp78 (9590S) were purchased from Cell Signaling Technology; anti-EV71 3C (GTX630191) was purchased from Genetex; anti-CD147 (11989-1-AP), anti-c-MYC (10828-1-AP), anti-PI4KB (13247-1-AP), anti-ARF1 (20226-1-AP), anti-GBF1 (25183-1-AP), and anti-RTN3 (12055-2-AP) were purchased from Proteintech; anti-UBE2G2 (ab174296), anti-XTP3-B (ab181166), anti-UBXD8 (ab154064), anti-Ufd1 (ab181080), and anti-p97 (ab11433) were purchased from Abcam; anti-SEL1L (sc-48081), anti-Herp (BML-PW9705), and anti-Ubc6e (TA504988) were obtained from Santa Cruz Biotechnology, Enzo Life Sciences, and Origene, respectively; anti-EV71 (MAB979) and anti-EV71 VP1 (MAB1255-M05) were purchased from Millipore and Abnova, respectively. The corresponding IRDye 680- or 800-labeled secondary antibodies were obtained from LI-COR Biosciences. The fluorescence-labeled secondary antibodies used in immunostaining were purchased from Jackson ImmunoResearch. EV71 2C antibodies were generated in rabbits or mice using recombinant protein as the immunogen. DBeQ (SML0031), CHX (C4859), DTT (D9779), and Tg (T9033) were purchased from Sigma-Aldrich. DAPI (D1306), Lipofectamine 2000, and Lipofectamine RNAiMAX were purchased from Invitrogen. Tun (12819) and MG132 (474790) were purchased from Cell Signaling Technology and Calbiochem, respectively. G418 (E859) was purchased from Amresco. Lambda protein phosphatase (P0753) was purchased from New England Biolabs.
pCMV6-SHH, pCMV6-TTR and pCMV6-Ubc6e were purchased from Origene. NHK cDNA were kindly provided by Dr. Richard N. Sifers (Baylor College of Medicine, USA). NS1 κ LC plasmid was a gift from Dr. Linda M. Hendershot (St. Jude Children’s Research Hospital, USA). The pCMV-SP-S11-NHK-HA and pRRL-S1–10 constructs were gifts from Dr. Shengyun Fang (University of Maryland School of Medicine, USA). pcDNA3.1-IRES-2A was a generous gift from Dr. Shih-Yen Lo (Tzu Chi University, Taiwan, China). To construct pCMV6-NHK and pCMV6-NS1 κ LC, cDNA templates were amplified by PCR and then cloned into the SgfI and MluI sites of pCMV6-Entry vector. To construct pCMV-SHH-S11-HA and pCMV-TTR D18G-MycFlag-S11, SHH, TTR D18G, and linker-S11 sequences were amplified by PCR and then cloned into the EcoRI and SalI sites of pCMV-SP-S11-NHK-HA. To construct the pVRC-p97-FLAG plasmid, RNA from RD cells was reverse transcribed into cDNA, and p97 cDNA was amplified by PCR and cloned into the SalI and XbaI sites of pVRC. To construct pCMV6-p97-GFP, the p97 sequence was amplified from pVRC-p97-FLAG by PCR and then cloned into the SgfI and MluI sites of pCMV6-AC-GFP. To construct pVRC-2C-HA and pVRC-3C-FLAG, the 2C and 3C cDNAs were amplified by PCR from pEGFPC1-2C/3C and then cloned into the SalI and XbaI sites of pVRC vector. pCMV6-TTR D18G, pVRC-p97QQ-FLAG, and plasmids expressing Ubc6e mutants were generated by site-directed mutagenesis. pcDNA3.1-IRES-2A(C110A) and pEGFPC1-3C/3C(C147S) were described previously [4,100].
Cells were lysed on ice for 30 min in lysis buffer (25 mM Tris-Cl, 150 mM NaCl, 1 mM EDTA, 1% NP-40 or 1% Triton X-100, pH 7.4) supplemented with protease inhibitor cocktail (Roche, 04693132001). Then, the cell lysates were centrifuged at 20,000× g at 4°C for 15 min to remove insoluble materials. Equal amounts of total proteins (20–100 μg) were separated by 8%–15% SDS-PAGE and then transferred to nitrocellulose membranes (Pall, 66485). After blocking with 5% nonfat dry milk solution in TBS at room temperature for 1 h, the membranes were incubated with different primary antibodies overnight at 4°C. The following day the membranes were washed three times in TBST and then incubated with IRDye 680 or 800-labeled secondary antibodies at room temperature for 2 h. They were then scanned using the Odyssey Infrared Imaging System (LI-COR Biosciences), and immunoblot bands were quantified using Image J software (National Institutes of Health).
RD cells or RD cells stably expressing SHH, NHK, TTR D18G, and NS1 κ LC were incubated with 100 μg/ml CHX in culture medium at 37°C for the indicated times to inhibit de novo protein synthesis. Then, the cells were collected and lysed in ice-cold lysis buffer for further analysis.
The recombinant EV71 3Cpro and 3C protease-dead mutant 3Cpro(E71A) were previously described [101,102] and provided by Dr. Sheng Cui (Institute of Pathogen Biology, Chinese Academy of Medical Sciences & Peking Union Medical College, China). In the in vitro cleavage assay, 293T cell lysates were incubated with indicated amounts of 3Cpro or 3Cpro(E71A) in reaction buffer (50 mM Tris-HCl, pH 7.0, 200 mM NaCl) at 30°C for 2 h. Then, the mixtures were analyzed by immunoblotting.
Cells were transfected with plasmids or siRNA duplexes using Lipofectamine 2000 (Invitrogen, 11668019) and Lipofectamine RNAiMAX (Invitrogen, 13778100), respectively, according to the manufacturer’s instructions. siRNAs were transfected at a final concentration of 40 nM. siRNAs were purchased from Guangzhou RiboBio and the sequences are shown in S1 Table. Herp siRNA (sc-75245) was purchased from Santa Cruz Biotechnology with unknown sequences.
RD cells were fixed with 4% paraformaldehyde in PBS for 15 min and permeabilized using 0.3% Triton X-100 in PBS containing 3% BSA (Sigma-Aldrich, A7906) for 30 min. After washing three times with PBS, the cells were incubated with primary antibodies overnight at 4°C. The following day, cells were incubated with the appropriate fluorescence-conjugated secondary antibodies (Jackson ImmunoResearch) for 1 h at room temperature followed by DAPI for nuclear staining. Images were captured and analyzed using a TCS SP5 laser-scanning confocal microscope and LAS AF software (Leica Microsystems).
Total cellular RNA was extracted using Trizol reagent (Life Technologies, 15596–018) according to the manufacturer’s protocols. One microgram of total RNA was reverse transcribed in a 20 μl volume using Improm-IITM reverse transcriptase (Promega, A3500) following the manufacturer’s instructions. Real-time PCR was performed using 1 μl cDNA as the template with PowerUp SYBR Green Master Mix (Applied Biosystems, Life Technologies, A25742). The reaction was run on a 7900HT Fast Real-Time PCR machine (Applied Biosystems, Life Technologies), and levels of gene mRNAs were normalized to GAPDH mRNA. Results were analyzed as fold-change using the ΔΔCT method. The primers for EV71 RNA quantification were described previously [100], and the sequences of other primers were as follows:
Herp F for RD cells: 5´-CCGGTTACACACCCTATGGG-3´
Herp R for RD cells: 5´-TGAGGAGCAGCATTCTGATTG-3´
GAPDH F for RD cells: 5´-AAAATCAAGTGGGGCGATGCT-3´
GAPDH R for RD cells: 5´-GGGCAGAGATGATGACCCTTT-3´
VIMP F for BSRT7 cells: 5´-TGAGATTCCTGCACGTCACAGT-3´
VIMP R for BSRT7 cells: 5´-AGTGCCCTCAGTCGGACTGTA-3´
GAPDH F for BSRT7 cells: 5´-GACCTCAACTACATGGTCTAC-3´
GAPDH R for BSRT7 cells: 5´-TCGCTCCTGGAAGATGGTGAT-3´
RD cells grown on 10-cm dishes were mock-infected or infected with EV71 (MOI = 10) for 12 h. The cells were harvested and washed three times with PBS and then frozen using a high-pressure freezer (Leica EM HPM100). Freeze substitutions were performed in a substitution unit (Leica EM AFS2) in dry acetone with 2% osmium tetroxide at −90°C for 72 h followed by gradual warming to −20°C for 8 h and 0°C for 2 h. After washing with dry acetone at 0°C three times, the samples were warmed to room temperature. After infiltration with SPI812 resin for several hours, the samples were embedded into SPI812 resin, placed in a 60°C oven for 24 h, and then sectioned using a microtome (Leica EM UC6). Ultrathin sections were stained with uranyl acetate and lead citrate and then examined by transmission electron microscopy (FEI Tecnai Spirit 120 KV).
|
10.1371/journal.pgen.1005722 | Escape from Lethal Bacterial Competition through Coupled Activation of Antibiotic Resistance and a Mobilized Subpopulation | Bacteria have diverse mechanisms for competition that include biosynthesis of extracellular enzymes and antibiotic metabolites, as well as changes in community physiology, such as biofilm formation or motility. Considered collectively, networks of competitive functions for any organism determine success or failure in competition. How bacteria integrate different mechanisms to optimize competitive fitness is not well studied. Here we study a model competitive interaction between two soil bacteria: Bacillus subtilis and Streptomyces sp. Mg1 (S. Mg1). On an agar surface, colonies of B. subtilis suffer cellular lysis and progressive degradation caused by S. Mg1 cultured at a distance. We identify the lytic and degradative activity (LDA) as linearmycins, which are produced by S. Mg1 and are sufficient to cause lysis of B. subtilis. We obtained B. subtilis mutants spontaneously resistant to LDA (LDAR) that have visibly distinctive morphology and spread across the agar surface. Every LDAR mutant identified had a missense mutation in yfiJK, which encodes a previously uncharacterized two-component signaling system. We confirmed that gain-of-function alleles in yfiJK cause a combination of LDAR, changes in colony morphology, and motility. Downstream of yfiJK are the yfiLMN genes, which encode an ATP-binding cassette transporter. We show that yfiLMN genes are necessary for LDA resistance. The developmental phenotypes of LDAR mutants are genetically separable from LDA resistance, suggesting that the two competitive functions are distinct, but regulated by a single two-component system. Our findings suggest that a subpopulation of B. subtilis activate an array of defensive responses to counter lytic stress imposed by competition. Coordinated regulation of development and antibiotic resistance is a streamlined mechanism to promote competitive fitness of bacteria.
| Antibiotics are one mechanism among many that bacteria use to compete with each other. Bacteria in the environment and in host organisms likely use networks of competitive mechanisms to survive and to shape the composition and function of diverse communities. In this study, we cultured two species of soil bacteria to observe the outcome of competition and to identify competitive functions that dictate the outcome. We show that one organism, Streptomyces sp. Mg1, produces antibiotic linearmycins that cause cellular lysis and degradation of a competing colony of Bacillus subtilis. In turn, the B. subtilis activate a resistance mechanism, either transiently or through mutation of a two-component signaling system. Activation of the signaling system produces a suite of identified responses, which include resistance to linearmycins, altered colony morphology that resembles biofilms, and enhanced motility of B. subtilis. This work identifies a unified, multifaceted survival response that is induced by a subpopulation of bacteria to escape lethal consequences of antibiotic-mediated competition.
| Bacteria are communal organisms. As such, bacteria have mechanisms to interact with other species that range from cooperative to antagonistic. Antibiotics are a classic example of molecules produced by bacteria that probably function in shaping microbial communities due to their bioactive function, including growth inhibitory and stimulatory activities [1–4]. The study of antibiotics has revealed a great deal about the cellular functions they target, mechanisms of resistance, and uses in treating disease. The traditional approach to discovery of antibiotics typically begins with extraction of metabolites from culture media, followed by direct screening of culture extracts to identify growth inhibitory agents [5]. While this approach has had tremendous success for antibiotic discovery, it has left great gaps in our understanding of competitive dynamics between bacteria. Approaches to bacterial competition that rely on culture of two or more organisms together are emerging as a powerful tool to discover new bioactive molecules and reimagine mechanisms of competition between diverse species of bacteria [6,7]. For instance, microbial competitive functions include secreted enzymes, type VI secretion systems, and specialized metabolism, including developmental signals and antibiotics [1,4,8]. In addition, changes in community functions such as biofilm formation or motility are recognized increasingly as important competitive strategies for bacteria [9,10].
Specialized metabolism and developmental functions are common features among soil bacteria, including the actinomycetes, bacilli, and myxobacteria [11–17]. In these bacteria, antibiotic production and cellular development are often intertwined and co-regulated processes, which is thought to provide fitness benefits to the organisms [18–20]. For example, during typical development Streptomyces species differentiate and develop spores [14]. During Streptomyces sporulation the substrate mycelium is cannibalized, which is thought to provide the cells with necessary nutrients to complete sporulation [21,22]. Cannibalization of the substrate mycelium is concurrent with production of many antibiotics, which are thought to protect the nutrient resources from opportunistic competitors [23]. Use of simple, tractable assays of two or more competing bacteria is one approach to identify new specialized metabolites, enzymes, and bacterial functions that determine the outcomes of competitive interactions. Indeed, interaction assays reveal not only growth inhibitory metabolites, but also changes in development and colony morphology that expose abundant and poorly understood survival mechanisms for bacteria. Dynamic patterns of interaction based on models of competition are producing new insights into bacterial competitive mechanisms [9,18,24–28].
As a model for competitive interactions, we use different species of Bacillus and Streptomyces. This competition model has led to identification of new functions for known molecules, including bacillaene and surfactin [18,24]. In the case of surfactin, a secreted hydrolase was identified from Streptomyces sp. Mg1 (S. Mg1) and shown to be a resistance mechanism that specifically degrades surfactin and plipastatin produced by Bacillus subtilis [25]. The current study stems from observing colonies of S. Mg1 and B. subtilis placed side by side on agar media. In this format, cellular lysis occurs along with progressive degradation of the B. subtilis colony [29]. Previously, imaging mass spectrometry revealed the loss of the polyglutamate component of colony extracellular matrix in the area of lysis, indicating degradation of both cellular and extracellular materials [30,31]. Streptomyces sp. Mg1 encodes production of many specialized metabolites with potential to participate in lysis and degradation [32]. One gene cluster encodes the biosynthetic enzymes for chalcomycin A, which inhibits the growth of B. subtilis but does not cause lysis and colony degradation [29]. Here we report both the identification of a lytic degradative activity (LDA) from S. Mg1, as well as a mechanism of resistance to LDA for B. subtilis. We show that resistant mutants of B. subtilis have a complex phenotype, which includes LDA resistance and visible changes in colony morphology and motility. We show the LDA resistance and the changes in colony morphology and motility are genetically separable functions, all regulated by a two-component system of previously unknown function. Our results indicate that a subpopulation of B subtilis cells in a colony trigger a complex mechanism for competitive fitness when challenged by the streptomycete.
When cultured next to S. Mg1, Bacillus subtilis colonies are progressively degraded and the underlying cells are lysed (Fig 1A) [29]. Progressive degradation of the cells and the extracellular matrix is visible as a translucent patch that develops on a formerly opaque colony (Fig 1A, S1 Movie). Our initial interest was to identify causes of lysis and colony degradation. To identify candidate lytic agents, we chose a direct approach to isolate S. Mg1 metabolites or enzymes that contribute to the lytic and degradative activity (LDA). Initially, we found active material present in whole plate butanol extracts from S. Mg1 grown on agar. To improve yields and decrease complexity of LDA extracts, we cultured S. Mg1 in liquid medium in the presence of non-polar HP-20 resin for adsorption of metabolites. Adsorbed metabolites were eluted using methanol to generate the crude extract. Bacillus subtilis colonies exposed to the crude extract lysed, indicating the presence of the activity. To isolate the active agent, we fractionated the crude extract, first using a stepwise (10%) methanol gradient followed by time-based HPLC fractionation, and tested for active fractions (see methods for a detailed description). The Δpks strain of B. subtilis was used for enhanced sensitivity in these assays, because the mutant is hypersensitive to lysis in co-culture with S. Mg1 [29]. We isolated a single peak from a HPLC fraction that caused lysis and colony degradation similar to S. Mg1 (Fig 1A and 1C). The similarity between the effects of isolated LDA and a competing S. Mg1 colony suggested that lysis and colony degradation of B. subtilis may result from the action of a single compound.
To identify the active molecule, we analyzed the HPLC-purified sample by UV absorbance and ESI-mass spectrometry. The molecule showed strong UV absorbances at 319, 333, and 351 nm, indicative of a conjugated pentaene moiety [33]. We found that the observed mass, 1166.7 [M+H]+ (S1A Fig) and the UV absorbance profile are consistent with linearmycin B (Fig 1D), a linear polyene antibiotic produced by Streptomyces sp. no. 30 [34,35]. We used 1D and 2D NMR of the active sample to confirm the presence of linearmycin B. Previously reported chemical shifts for linearmycin B accorded with those we obtained for the LDA sample (S1B Fig, S1 Table) [35]. The linearmycins were originally identified as a pair of compounds, linearmycin A and B [34,35]. We examined the crude extracts and found them to also contain linearmycin A, which is also active for lysis of B. subtilis (m/z 1140) (S2 Fig). Furthermore, the S. Mg1 genome (GenBank Accession CP011664) [32] includes a polyketide gene cluster predicted to be responsible for linearmycin biosynthesis. We tested a mutant strain, S. Mg1-Δ37, which contains a chromosome truncation that removes the linearmycin biosynthetic cluster, and found the mutant failed to lyse B. subtilis or produce linearmycins (S2 Fig). In a parallel study, a targeted deletion of the acyl-transferase encoding gene in the linearmycin biosynthetic gene cluster disrupts linearmycin production specifically and blocks all lytic activity from the strain (personal communication, B. Chris Hoefler). Taken together, we conclude that S. Mg1 produces linearmycins, which are sufficient for LDA against B. subtilis. For simplicity, we collectively refer to these molecules as LDA.
No mechanism is known for either growth inhibition or the lytic effect that we observe with LDA. Linearmycin A was originally shown to inhibit growth of Escherichia coli and Staphylococcus aureus in addition to three fungal species, but no antibacterial mechanism of action was reported [35]. Structurally related polyene antibiotics include antifungal agents such as amphotericin B [36], nystatin [37], and ECO-02301 [38]. Amphotericin B and nystatin inhibit fungal growth specifically by interactions with ergosterol and the fungal plasma membrane [39–42]. However, bacterial membranes lack ergosterol, suggesting a different mechanism of action against bacteria for LDA. Nystatin was found to induce biofilm formation by B. subtilis grown in LB media [43], demonstrating that antifungal polyenes are biologically active in the absence of ergosterol. The lytic activity of LDA indicates a mechanism of action for the linearmycins that differs from nystatin. In the absence of a known target, we sought an approach to better understand the lysis and degradation of B. subtilis.
When plated next to extracts of LDA or S. Mg1 colonies, small B. subtilis colonies emerge in the region of lysis and appear to be resistant to LDA [29]. We wanted to identify mechanisms of resistance as an approach to better understand the lytic process caused by LDA [44]. Direct comparison of Δpks and wild-type strains of B. subtilis, either in culture with S. Mg1 or when treated with LDA, showed that the Δpks strain is hypersensitive to lysis but has no other observable phenotype in these assays [29]. Therefore as before, we used the Δpks strain of B. subtilis for these assays, because the LDA hypersensitivity provided an expanded area of lysis in which we could scan for potential resistant mutants. We challenged colonies of the Δpks strain of B. subtilis with extracts from S. Mg1 cultures and observed small colonies appearing in the degraded portion of the parent colony after lysis occurred (e.g. Fig 1B). We isolated 60 small colonies from several lysed colonies and tested them for resistance to LDA in co-culture with S. Mg1. The majority of the isolates lysed when cultured again with S. Mg1, indicating only transient resistance to LDA. However, ten isolates were stably resistant to LDA (LDAR), potentially having acquired mutations leading to resistance (Fig 2A). Notably, all of the stable LDA resistant colonies developed a rough, wrinkled colony morphology that is distinct from the parental strain (Fig 2A). Due to a biofilm-like appearance of the LDAR colonies, we suspect the mutations have pleiotropic effects on growth mode and development, as well as resistance to LDA.
To identify the mutant alleles in the LDAR isolates, we sequenced six of the ten mutant genomes and compared the sequences to the parental genome (Δpks strain). Surprisingly, all six isolates had point mutations in either of the two genes in the yfiJK operon. In addition, eleven non-overlapping mutations occurred in a subset of the spontaneous LDAR mutants (S2 Table). Three spontaneous LDAR mutants possessed point mutations only in yfiJ, which prompted our focus on the yfiJK operon. Using PCR and Sanger sequencing we found that the other four LDAR isolates also contained point mutations in yfiJ. In total, nine of ten mutations were found in yfiJ and one in yfiK (Fig 2B, Table 1). The yfiJ gene encodes a membrane-bound sensor histidine kinase (HK), and the yfiK gene encodes its cognate cytoplasmic response regulator (RR) [45]. Together these proteins comprise a two-component system (TCS). In a canonical TCS, a HK dimer senses a signal and autophosphorylates on a conserved histidine residue [46]. The phosphate is subsequently transferred to the cognate RR, which then effects a response, most commonly through DNA binding and regulation of gene expression [46]. In the case of YfiK, the effector domain is a helix-turn-helix domain that likely binds DNA to modulate changes in gene expression [45]. A role in LDA resistance is the first indication of a native function for this two-component system.
To determine whether LDA resistance requires active YfiJK, we deleted yfiJ and yfiK independently, or yfiJK together, in otherwise wild-type genetic backgrounds, and co-cultured these mutants with S. Mg1. In all three cases the mutants lysed and were indistinguishable from wild-type B. subtilis (Fig 2C). The absence of any observable phenotype for the yfiJK deletion mutations suggested that resistance arises from gain-of-function alleles that activate the two-component system. As a test for gain-of-function alleles, we genetically complemented the deletion strains of yfiJ or yfiJK with PCR-amplified alleles from the spontaneous LDAR strains. Control strains complemented with native alleles were wild type with respect to lysis and colony morphology (Fig 2D). Conversely, complementation with the mutant alleles caused B. subtilis to be resistant to LDA when cultured with S. Mg1, and the mutants developed a more wrinkled colony surface than wild type (Fig 2D, Table 1). Based on these observations, we concluded that each LDAR allele is likely activating YfiJK to stimulate both abnormal colony development and LDA resistance.
We next investigated how YfiJK may relate to the mechanism of lysis and colony degradation. We considered the results of a previous microarray study to define the regulon of each known RR in B. subtilis [47]. In that study, overexpression of yfiK repressed expression (≥ 4-fold) of 29 different genes, the majority involved in amino acid biosynthesis [47]. The reported regulon also includes skfF, which encodes the ATP-binding cassette (ABC) transporter necessary for release of spore-killing factor (SKF), and iseA, a cell wall-associated protein that inhibits two major autolysins [48–50]. We hypothesized that SKF and autolysis might be involved in linearmycin-induced lysis, and that yfiJK may regulate the expression of those functions. We tested sensitivity to LDA using four strains of B. subtilis. First, we tested a strain unable to produce SKF (ΔskfA-H) to determine if the cannibalism peptide functions as a lytic agent. Second, we tested whether a strain deficient in iseA would show enhanced lysis in the absence of an autolysin inhibitor. Third, because iseA regulates autolysins, we tested whether a strain deficient in production of three major autolysins (ΔlytABC ΔlytD ΔlytF) may show diminished lysis when exposed to LDA. Fourth, we tested a strain with a deletion of the major motility/autolysin regulator (ΔsigD) [51]. All four strains lysed when cultured with S. Mg1, indicating that SKF and autolysis do not likely contribute to the lysis mechanism (S3 Fig). In a parallel approach, we used transposon mutagenesis to identify genes in B. subtilis that may cause lysis under linearmycin-induced stress. We obtained a single, stable LDAR mutant, however LDA resistance was unlinked to the site of transposon insertion in this strain. We sequenced the mutant genome and identified an additional point mutation in yfiJ (yfiJL254P) (Table 1, S1 Methods). Thus, using multiple approaches to identify functions conferring LDA resistance, we have found only apparent gain-of-function alleles in yfiJK.
Two-component signaling systems require conserved phosphoacceptor residues for activation and downstream signaling [46]. We identified the phosphoacceptor histidine (H201) in YfiJ and the phosphoacceptor aspartate (D54) in YfiK using multiple sequence alignment to experimentally characterized TCS. Using site-directed mutagenesis we disrupted the phosphoacceptor residues and created the new alleles yfiJH201N and yfiKD54A. As anticipated based on the ΔyfiJK phenotype, both phosphoacceptor mutants were sensitive to LDA when cultured with S. Mg1 (Table 2). Next, we constructed new alleles that combined the phosphoacceptor disruptions with substitutions found in LDAR alleles to generate the new, double mutant alleles yfiJA152E, H201N and yfiKD54A, T83I. When cultured with S. Mg1 these mutant strains lysed, which confirmed the disruption of the gain-of-function LDAR phenotype in the absence of functional a TCS (Table 2). As a final test that LDA resistance results from specific downstream signaling of YfiJK, we constructed a pair of double mutants: (i) combining the LDAR allele yfiJA152E with the phosphoacceptor disruption yfiKD54A and (ii) combining the phosphoacceptor disruption yfiJH201N with the LDAR allele yfiKT83I. When these strains were cultured with S. Mg1 they were sensitive to LDA (Table 2). These results suggest that LDA resistance is due to specific downstream signaling of YfiJK, leading us to conclude that LDA resistance is due specifically to activation of the TCS.
Amphotericin B and nystatin are cyclic polyene antifungals [36,37]. The structurally related linear polyene, linearmycin A, is also antifungal but has also been shown to have antibacterial activity as well [35]. We tested amphotericin B, nystatin, and ECO-02301, a polyene structurally related to the linearmycins [38], for activity against B. subtilis. ECO-02301 caused lysis similar to linearmycins, but the macrocyclic polyenes amphotericin B and nystatin were not lytic (Fig 3). When tested against purified ECO-02301, the LDAR mutant (yfiJA152E) strain appeared to be partially resistant in this assay (Fig 3). We next sought a quantitative measure of the difference between LDA resistance and sensitivity to LDA and ECO-02301. First, we measured the minimum lytic concentration (MLC) for ECO-02301 using a quantitative agar diffusion assay and determined that a LDAR strain of B. subtilis was 3.65-fold more resistant to ECO-02301 (Table 3). We applied the same assay to LDA, containing both linearmycin A and B, isolated from S. Mg1 cultures and quantified the fold difference in resistance between LDAR and wild-type B. subtilis. The LDAR mutant was nearly ten-fold more resistant to LDA compared to the sensitive strain (Table 3). The difference in relative resistance to ECO-02301 and LDA may be in part due to structural differences in the molecules. The synthesis of ECO-02301 includes tailoring reactions that glycosylate the polyketide backbone and condense an amidohydroxycyclopentenone moiety onto the terminal carboxylic acid group [38,52]. The structural differences may affect target affinity, solubility, or other properties of the molecule, leading to differences in overall activity.
Because polyene antibiotics typically exert their effects on the cellular membrane, we wanted to determine if LDA resistant alleles of yfiJK provide B. subtilis with a generalizable cross resistance to membrane-active antibiotics. Daptomycin is a lipopeptide antibiotic that targets the cell membrane [39,40,53]. The killing mechanism of daptomycin is not lytic, although lysis follows prolonged exposure [54]. We found that daptomycin caused a morphologically similar lysis and degradation of B. subtilis when spotted on a filter paper disc adjacent to a colony (Fig 3). A LDAR strain of B. subtilis also lysed when exposed to daptomycin. In comparison to the LDA sensitive strain, the LDAR strain showed some residual opacity following lysis, suggesting that LDAR alleles might provide cross-protection to daptomycin (Fig 3). However, we found the MLC of daptomycin was identical between the LDA resistant and sensitive strains (Table 3). Our results suggest that YfiJK signaling provides resistance either specifically to linear polyene molecules related to linearmycins, or commonly to the type of lytic cell damage caused by linearmycins.
Immediately downstream of the yfiJK operon are three genes, yfiLMN, predicted to encode an ABC transporter [45,55]. This genetic architecture is similar to peptide-antibiotic resistance systems previously characterized in B. subtilis and other Firmicutes [56]. In these systems, a TCS and an ABC transporter are functionally linked and required for antibiotic resistance. We hypothesized that YfiJK-LMN may function similarly to confer LDA resistance. Thus, we were interested in determining if YfiLMN is necessary for LDA resistance. We engineered a strain with all five genes, yfiJKLMN, deleted. The ΔyfiJKLMN strain was lysed in co-culture with S. Mg1 (Fig 4). We inserted resistant alleles of yfiJK at the non-essential amyE locus to generate strains unable to produce YfiLMN but possessing LDAR alleles of yfiJK. When cultured with S. Mg1, these strains were sensitive to LDA (Fig 4). We then complemented the loss of yfiLMN in these strains by inserting the yfiLMN genes, including the intergenic region between yfiK and yfiL, at the non-essential lacA locus. A predicted terminator exists downstream of yfiK (-8.9 kcal/mol) (genolist.pasteur.fr/SubtiList) [57]. Our initial yfiLMN complementation construct included sequence immediately downstream of the terminator. However, this construct failed to complement the loss of resistance (S4 Fig). Upon further investigation, we found no recognizable promoter elements in the intergenic region between the yfiK terminator and yfiL (143 bps). We hypothesized that yfiJKLMN may constitute a single operon with some level of terminator read-through resulting in yfiLMN expression. To circumvent the lack of an independent promoter, we placed the expression of yfiLMN under a constitutive Pspac(c) promoter and inserted these constructs at the non-essential yhdG locus [58]. Under constitutive expression, the yfiLMN-complementation strains were LDA resistant, showing only minimal lysis in co-culture with S. Mg1 (Fig 4). This effect was observed even in a strain complemented with wild-type yfiJK and in a strain lacking yfiJK entirely. These results demonstrate that YfiLMN is necessary for LDA resistance, and that constitutive expression bypasses the need for YfiJK. We speculate that YfiLMN either removes linearmycins from B. subtilis cells to provide resistance, or alternatively, functions in cell envelope processes or regulatory functions that control LDA resistance. Determining the mechanism of YfiLMN-mediated LDA resistance will require further investigation.
In our study of the different LDAR alleles, we observed some variation in the degree of wrinkled, motile phenotype in competition with S. Mg1 or under LDA exposure. To separate effects of the competitor from inherent LDAR phenotypes, we plated colonies of LDAR strains in isolation to view morphological features. All of the yfiJK mutant strains displayed a pattern of increased colony wrinkling and spreading across the agar surface and were distinct from the wild-type strain (Fig 5). We asked if differences in LDAR morphology would be visible on the biofilm-inducing medium, MSgg [59]. The mutant B. subtilis colonies developed a wrinkled appearance similar to wild type, indicating that traditional biofilm morphology and development are not disrupted in the mutant strains (Fig 5). We also noted that B. subtilis strains, either wild type or ΔyfiJ, formed smooth colonies in the absence of S. Mg1 when grown on rich media (Fig 5). In contrast, the same B. subtilis colonies in competition assays tend so show a somewhat wrinkled morphology, regardless of the yfiJK alleles present. Thus, the morphology of the B. subtilis colonies appears to be influenced by a combination of both the LDA alleles and the presence of the competitor species.
To directly compare colony morphology in isolation and with the competitor, the wild type and LDAR strains were cultured at different distances to S. Mg1. We inoculated colonies of LDAR B. subtilis and S. Mg1 in a perpendicular, cross-wise pattern on 1.5% agar MYM medium to provide a format for increasing distances between colonies of each species (Fig 6). The growth of B. subtilis with the wild-type yfiJ allele showed smooth colony formation with lysis proximal to the S. Mg1. In contrast, the B. subtilis strain with the yfiJA152E allele revealed different effects based on its proximity to S. Mg1 (Fig 6). The LDAR colonies distant from S. Mg1 had the expected wrinkled morphology and spreading outgrowths, as was observed when cultured in isolation (Fig 5). However, the S. Mg1-proximal colonies were morphologically different with a flattened surface and more uniform spreading pattern. The observed changes in colony morphology associated with LDAR suggest that the YfiJK TCS regulates both specific resistance and developmental functions that coordinate a survival response to the competitor species.
To gain insight into possible connections between colony phenotypes and resistance to lysis, we considered that changes to extracellular matrix (ECM), the associated biofilm-like colony morphology, and changes in motility, may be responsible for LDA resistance [10,60]. For instance, the ECM may impede access of linearmycins to their target, possibly through overproduction of EPS or other matrix components [60]. To test whether LDA resistance is dependent on known components of biofilm ECM, we sought to separate the two processes. We generated an ECM-defective strain, which was unable to produce exopolysaccharide (EPS) due to an epsH deletion [59], in an otherwise LDA resistant background (yfiJA152E). This strain developed as a flat, mucoid colony, but remained LDA resistant in co-culture with S. Mg1 (Fig 7). Based on this result, we concluded that, while EPS production is necessary for the wrinkled colony morphology, intact biofilm matrix in the LDAR strains is not responsible for the LDA resistance mechanism. However, LDA resistance may require other biofilm matrix components [61]. We asked whether hyperactivation of biofilm production would mimic the LDA resistance phenotype. We deleted the gene encoding sinR, the master biofilm repressor [62], in a LDA sensitive strain. When sinR is deleted, B. subtilis overproduces biofilm matrix and the colonies grow with a profoundly wrinkled appearance [62]. If biofilm formation is responsible for LDA resistance, then a ΔsinR mutant should be resistant in co-culture with S. Mg1. However, the ΔsinR strain was sensitive to lysis (Fig 7). LDA resistance was observed in a ΔsinR strain only in the presence of the mutant yfiJ (yfiJA152E).
Biofilm formation is controlled not only by SinR but also by the TCS DegSU. This TCS is responsible for control of the production of biofilm extracellular matrix components. Among these components are BslA and γ-poly-DL-glutamate (γ-PGA) [63–65]. To test if matrix functions provided by DegU may contribute to LDA resistance, we deleted degU in a LDA resistant background (yfiJA152E) and cultured the strain with S. Mg1. This mutant developed as a flat colony that was LDA resistant, suggesting that resistance to LDA does not require functions provided by DegU (Fig 6). Based on this finding and our SinR and EpsH experiments, we conclude that the changes in colony morphology of LDAR mutants are not the principle cause of LDA resistance.
In addition to wrinkled colony morphology, the enhanced motility of LDAR strains may be linked to resistance. For instance, swarming motility has been associated with elevated antibiotic resistance in multiple bacteria [10]. Previously, we observed lysis in a ΔsigD strain, which is deficient in swarming and autolysin production [51,66,67] (S3 Fig). We tested whether a ΔsigD, yfiJA152E double mutant strain would undergo lysis despite the presence of the LDAR allele (Fig 7). This strain maintained both LDA resistance and morphological changes, including colony spreading. The spreading phenotype in the absence of sigD is consistent with LDAR mutants exhibiting sliding motility when cultured with S. Mg1 [68]. In sum, the combined phenotypes of LDAR support a model wherein activation of YfiJK leads to LDA resistance through YfiLMN activation coordinated with separable changes in colony motility and morphology that promote survival during competition.
In an effort to identify a regulatory network for YfiJK, we sought to identify genes differentially expressed in a LDAR mutant that may contribute to colony phenotypes. To perform differential expression analysis, we isolated and sequenced RNA from yfiJK+ (PDS0627) and yfiJA152EK (PDS0685) strains cultured on agar plates. In our analysis, we identified six genes with statistically significant changes in expression between the two strains (Table 4). Expression of yfiLMN was increased on average ~18-fold in the LDAR mutant. To corroborate this result we used qRT-PCR and observed a ~20-fold increase in yfiL expression from the yfiJA152EK strain (S6 Fig). We did not observe a change in expression of yfiJK in our RNA-seq experiments, which is consistent with an additional control element between yfiK and yfiL. Three other differentially expressed genes were all decreased in the LDAR mutant: des, which encodes a phospholipid desaturase responsible for cold shock adaptation [69,70] and yvfRS, which encodes an ABC transporter of unknown function [55]. Surprisingly, no genes in the eps operon or other known biofilm-related genes were identified as differentially expressed between the LDA sensitive and LDAR strains. Also of note, we found no correspondence between the YfiJK-regulated genes we identified by RNA-seq and the regulon previously defined by microarray study of yfiK overexpression [47]. In the absence of a clear connection to established biofilm and motility functions, the RNA-seq results suggest that the morphological changes observed in LDAR colonies may arise directly from activation of YfiLMN function combined with repression of des (phospholipid content) and yvfRS (unknown function) by an unknown mechanism. Alternatively, the morphological changes may occur only in a subpopulation of cells insufficient to be detected during our analysis.
One of our initial goals was to identify mechanisms of resistance in an attempt to expose mechanistic aspects of linearmycin activity. We considered that LDA resistance may only exist under aberrant conditions, which arise through mutations that hyperactivate the YfiJK signaling system. In the absence of a clear phenotype for deletion of the genes, we sought an approach to identify wild-type YfiJK function in colony morphology and LDA resistance. We returned to an early observation that small colonies resistant to LDA emerge in the lysed region of a B. subtilis colony. The majority of the isolated LDA resistant colonies isolated were only transiently resistant (50/60). We reasoned that if the natural function of YfiJK is to provide temporary resistance to LDA-induced damage, the emergence of transient resistance would depend upon the function of YfiJK. Therefore, we tested 6 independent colonies, each in triplicate, of wild type and ΔyfiJK versus S. Mg1 to determine if resistant colonies would emerge in the absence of YfiJK (Figs 8 and S5). The resulting cultures showed many small colonies arising in the lysed areas of the wild-type B. subtilis colonies. By contrast, the few small colonies observed with the ΔyfiJK strain did not grow appreciably and lacked the morphological features of the yfiJK+ colonies (Fig 8). This result is consistent with the natural function of YfiJK providing transient resistance to LDA-induced stress. In the case of yfiJK gain-of-function alleles, the substitutions in YfiJK may lock the TCS into an active state wherein every cell becomes resistant to LDA in contrast to the subpopulations observed among wild-type cells. Intriguingly, the transient resistance appears in only a subset of cells in the colony. Variable antibiotic resistance among a clonal population of cells has been described as heteroresistance, and is thought to be advantageous for survival of bacteria during antibiotic treatment [71–73]. The ability to activate YfiJK in a subset of cells may constitute a mechanism of transient heteroresistance to linearmycins and related molecules, but defining the mechanism and limitation to a subpopulation of cells will require further investigation. The observed pattern of YfiJK-dependent LDA resistance highlights that this TCS, and possibly many TCS, may transiently serve a subset of cells in a population during times of competitive crisis.
In this study, we used a two-species culture model of bacterial competition to identify functions that contribute to bacterial competitive fitness. The present study stemmed from an earlier observation of lysis and degradation of B. subtilis colonies when cultured adjacent to S. Mg1 [29]. Here, we first identified linearmycins, produced by S. Mg1, as the primary cause of progressive lysis and colony degradation. The culture format used for competition revealed small B. subtilis colonies spontaneously resistant to lysis. When isolated, the resistant colonies showed a biofilm-like appearance with increased wrinkled colony morphology and aberrant motility. We sequenced whole genomes of the resistant colonies and identified mutations that confer resistance. Genomic analysis revealed alleles of the yfiJK operon, which encodes a two-component system of previously unknown function. Based on our observations, we define yfiJK as a regulator of yfiLMN, encoding an ABC transporter, and possibly other target genes that govern modes of colony growth and motility (Fig 9).
We show that the LDA resistance is not dependent upon known biofilm-specific functions, suggesting that colony morphology and LDAR are separable processes, unified under YfiJK regulation. Two-component systems are well established as regulators for cellular responses to environmental stresses, including antibiotics [74,75]. The significance of the current work is the use of model interspecies competition to reveal both the agent of aggression, linearmycins, and a multifaceted survival response from genes with no prior functional assignment, yfiJKLMN.
Only gain-of-function mutations in yfiJK were identified in this study to cause LDA resistance. The resistance alleles of yfiJK were due to missense mutations causing changes to four regions of YfiJK: (i) the third TM helix in YfiJ, (ii) the cytoplasmic linker between the fifth TM helix and the dimerization and histidine phosphotransfer (DHp) domain in YfiJ, (iii) the C-terminal end of the DHp, and (iv) the regulator domain in YfiK. We hypothesize that each of these amino acid substitutions are responsible for conformational changes in YfiJK, leading to a constitutively active state. A previous study described a similar phenotype caused by point mutations of pmrAB in Pseudomonas aeruginosa. Gain-of-function alleles in pmrB lead to polymyxin B resistance via increased signaling through the histidine kinase [76]. We also considered an alternative mechanism, wherein point mutations in yfiJK could lead to non-cognate interactions of YfiJ or YfiK and aberrant signal transduction [77]. However, we view this mechanism as unlikely because only one of the affected residues (L254) lies in the DHp domain, which is predicted to be involved in specificity [78], and LDA resistance required the presence of the phosphoacceptor residue in the cognate partner. Thus, we conclude that gain-of-function alleles cause LDA resistance and changes in both colony morphology and motility, and that the signaling is specific to YfiJK. Although the specific defects caused by each allele will require further investigation, we note that many of the mutations we observed are responsible for amino acid changes in the cytoplasmic linker of YfiJ. The cytoplasmic linker domain of HKs has been best characterized in periplasmic-sensing histidine kinases. In these kinases, the linker may contain conserved PAS or HAMP domains that are necessary for signal transduction from the sensory machinery to the kinase domains [46,79–81]. YfiJ has neither of these conserved domains, suggesting that the short linker in this protein is the sole signal-transducing domain. The mutations in the yfiJ linker, through fixing the protein in activated state, may be very informative for determining the mechanism of signal transduction via the YfiJ intramembrane histidine kinase.
Two-component systems are commonly involved in sensing antibiotic and environmental stress [74,75]. Among Firmicutes, a conserved mechanism for resistance to peptide antibiotics pairs genes for two-component systems and ABC transporters [56,82,83]. The identification of mutations in yfiJK suggests the cell envelope is the linearmycin target, based on comparison to other TCS-ABC transporter pairs in B. subtilis [56]. Immediately downstream of yfiJK are three genes, yfiLMN, that are predicted to encode an ABC transporter. We found that when B. subtilis was unable to produce YfiLMN, the colonies were LDA sensitive and failed to develop altered colony morphology, regardless of the presence of a LDAR allele of yfiJK. Furthermore, expression of yfiLMN under a constitutive promoter resulted in LDA resistance, even in the absence of yfiJK. Thus, the YfiLMN transporter is necessary and sufficient for LDA resistance. We hypothesize that YfiLMN may act as an exporter either for linearmycin or for cell envelope remodeling factors that lead to LDA resistance.
We used RNA-seq to identify genes that may be regulated by YfiJK. As expected we identified that yfiLMN expression was increased in a LDAR mutant. We also identified yvfRS, encoding an ABC transporter of unknown function, and des as genes downregulated by YfiJK. The des gene encodes a fatty acid desaturase that is responsible for altering membrane fluidity in response to cold shock [69,70]. Intriguingly, B. subtilis strains with des deletions are more susceptible to daptomycin-treatment, potentially due to their altered membrane fluidity [84]. Antifungal polyenes structurally related to linearmycins target ergosterol in fungal membranes [39–42]. The decreased expression of des in LDAR mutants may contribute to resistance by affecting interactions between linearmycins and the cell membrane. Characterization of the cell envelopes of LDA sensitive and LDAR strains may provide insight into the mechanism of linearmycin-induced lysis.
Mutants with LDAR alleles of yfiJK grow as rugose colonies that resemble some aspects of biofilm development on rich media, which does not support traditional biofilm development. We demonstrated that we could functionally divorce this colony morphology phenotype and LDA resistance by expressing yfiLMN constitutively and by introducing deletions of genes specifically required for biofilm development (epsH, sinR, degU). In so doing, we found that changes to the biofilm extracellular matrix are not responsible for resistance. LDA resistance may be modulated by specific matrix or cell envelope modifications activated by YfiJK-LMN, but such modifications remain to be identified. Although we found no obvious candidates in our RNA-seq data to explain colony morphological changes, the decreased expression of des or yvfRS may contribute to alterations in colony development. We also observed that LDAR mutants respond to S. Mg1 by inducing motility, whereas wild type B. subtilis colonies are lysed. The pleiotropic phenotypes of yfiJK LDAR alleles differentiate this coupled TCS-ABC transporter system from the BceRS-AB, PsdRS-AB, YxdJK-LM systems in B. subtilis, which appear to be dedicated antibiotic resistance systems [56,85–88]. To our knowledge, there are no phenotypes associated with development that have been attributed to these TCS-ABC transporter pairs, suggesting that YfiJK holds a specialized role in providing specific LDA resistance and in activating biofilm development and motility, both of which are known to increase resistance to antimicrobials [10,60]. We propose that activation of YfiJK-LMN promotes competitive fitness of B. subtilis by coupling a specific resistance mechanism (LDAR) with generalized-resistance that occurs as a consequence of altered development and motility. A recent study using strains of Pseudomonas aeruginosa demonstrates that biofilm formation is stimulated in response to competition, as opposed to a cooperative function of different strains or cell types [9]. The identification of YfiJK as a regulator of biofilm and motility functions is consistent with a model wherein competition with S. Mg1 induces developmental responses, including biofilm and colony spreading, among a subpopulation of resistant cells of B. subtilis.
Using microbial competition we assigned resistance and developmental functions to a previously uncharacterized TCS in B. subtilis. Without imposing the conditions of competition on B. subtilis, these TCS functions may be difficult to identify, because the yfiJ, yfiK, and yfiJK deletion mutants have no phenotype when compared to wild type. The B. subtilis genome encodes 36 histidine kinases and 34 response regulators [89]. The functions of at least 11 of these TCS are currently unknown. Bacteria use these systems to sense and respond to their environment, which include stresses and nutrient conditions, but also include other bacteria and their antagonistic enzymes and specialized metabolites. Many TCS of unknown function may have a role in the context of microbial competition, despite having no distinct phenotype under laboratory conditions. Thus, microbial competition studies provide an effective approach to identify functions for TCS and other genes that promote competitive fitness of bacteria. By expanding our knowledge of individual competitive functions, a more comprehensive view of bacterial competitive fitness will emerge.
The strains of B. subtilis we used in this study are listed in S3 Table. We cultured B. subtilis strains at 37°C in lysogeny broth (LB) [1% tryptone (Bacto), 0.5% yeast extract (BBL), 0.5% sodium chloride (Sigma)] or on LB agar plates [1.5% Agar (Bacto)]. We maintained Streptomyces sp. Mg1 (PSK0558) as a spore stock in water at 4°C. Unless otherwise stated all co-cultures were grown on MYM [0.4% malt extract (Bacto), 0.4% yeast extract (BBL), 0.4% D-(+)-maltose monohydrate (Sigma)] with 2% agar (Bacto). We used chloramphenicol (5 μg/mL), kanamycin (5 μg/mL), MLS (1 μg/mL erythromycin, 25 μg/mL lincomycin), spectinomycin (100 μg/mL), and tetracycline (20 μg/mL) as needed. The primers we used in this study are listed in S4 Table. We used Escherichia coli DH5α or XL-1 blue for plasmid maintenance and manipulation. We prepared All B. subtilis genetic manipulations in either the 168 or PY79 strain background and then transduced them to NCIB3610 using SPP1 phage transduction as previously described [90].
We wetted 1 g Diaion HP-20 resin in 25 mL methanol (MeOH) followed by washes with 25 mL of water five times while shaking. Next, we removed the bulk liquid and resuspended the resin in 250 mL of MYM. We sterilized the media and resin by autoclaving the mixture. We inoculated cultures using 1 mL of S. Mg1 that was grown overnight (107 spores in 3% tryptone soy broth). We cultured the S. Mg1 for 6 d at 30°C while shaking at 225 RPM in the dark. We performed all culture growth and extractions in low ambient light, because the activity of extracts was diminished or lost if manipulated in the light. We separated the HP-20 resin from the bulk of the S. Mg1 by repeatedly washing the resin with water until all visible filaments were removed. To extract resin-bound molecules, we washed the resin with successive 25 mL volumes of MeOH until the solvent was clear. To generate our crude extract, we pooled the washes and removed MeOH using a rotary evaporator. The crude extract was dissolved to 100 mg/mL in 50% acetonitrile (ACN) and fractionated over a C8 solid-phase extraction (SPE) column eluted with a MeOH/water stepwise gradient. We removed solvent from our fractions using a rotary evaporator and suspended the fractions to 50 mg/mL in 50% ACN. We tested the fractions for lytic activity against B. subtilis by spotting 10 μL on a filter paper disc adjacent to a B. subtilis colony that had been pre-grown for 24 h and observing lysis over a period of 48 h. The 70% and 80% MeOH fractions were active in the lysis assay and pooled for further fractionation.
Using an Agilent 1200 HPLC system, we further fractionated the active extract fractions over a semi-preparative (10 x 250 mm, 5 μm) Phenomenex Luna C18 column and eluted with an ACN/20 mM ammonium acetate pH 5 (NH4OAc) gradient running at 5 mL/min. The elution program was as follows: 1) 5 min at 40% ACN then 2) a gradient up to 50% ACN over 10 min then 3) a gradient up to 75% ACN over 5 min, and 4) a gradient diminishing to 40% over 5 min. We injected 35 μL of pooled active fraction per injection. We collected time based fractions and tested them for lytic activity as above. Active fractions were analyzed by mass spectrometry using a Bruker microTOF mass spectrometer. For NMR analysis, the sample was dried and resuspended in 300 μL deuterated dimethylsulfoxide (DMSO-d6). We collected spectra on a Bruker Avance III 500 MHz spectrometer equipped with a cryoprobe.
We diluted overnight cultures inoculated with a single colony of B. subtilis Δpks (PDS0067) into 5 mL of LB at OD600 = 0.08 with no antibiotics. We cultured the cells to early stationary phase (OD600 = 0.9–1.5) at 37°C and spotted 2 μL on MYM7 plates [as above with 100 mM MOPS and 25 mM potassium phosphate buffer pH 7, 1.5% agar (Bacto)]. After 24 h incubation, we placed 6 mm filter paper discs next to the B. subtilis colonies and added 10 μL of extract from S. Mg1. We returned the plates to the incubator and observed lysis and colony degradation over the next 48 h. After incubation, small colonies were observed in the region of lysis. We isolated 60 small colonies and passaged them on LB plates. We tested each isolate for LDA resistance using co-culture, as described below.
LDAR mutants that were stable through passage in isolation and the parental Δpks strain were used for whole genome sequencing. Sequencing libraries were prepared using the PCR-free TrueSeq Kit from Illumina. 250 bp paired-end reads were sequenced using an Illumina MiSeq. We mapped reads from the LDAR mutants onto the parental Δpks strain using MIRA and identified mutations by consensus discrepancy between the sequences [91,92].
We used long-flanking homology PCR to delete yfiJK and yfiJKLMN. Briefly, to delete yfiJK we amplified the upstream sequence using primers 13 and 14, the downstream sequence using primers 15 and 16, and the kanamycin resistance cassette from pDG780 using primers kn-fwd and kn-rev. We mixed the three PCR products together and used primers 13 and 16 to amplify a product, which we directly transformed into PDS0312 to generate PDS0546.
To delete yfiJKLMN we used primers 76 and 77 to amplify the upstream sequence of yfiJ and primers 78 and 79 to amplify the downstream sequence of yfiN. We combined these fragments with the kanamycin resistance cassette and amplified a product using primers 76 and 79, which we directly transformed into PDS0312 to generate PDS0652.
To test alleles of yfiJ, we complemented the ΔyfiJ deletion. We amplified yfiJ with primers 25 and 26 from wild type and spontaneous LDAR mutants. These primers include a BamHI and EcoRI site, which we used to clone the product into plasmid pDR183 (lacA::mls). We transformed the plasmids into PDS0559 and verified insertion into the lacA locus by PCR. We moved these constructs into PDS0555 using SPP1 phage transduction.
We tested alleles of yfiK by complementing both yfiJK together into a ΔyfiJK strain. We complemented both genes together because yfiK is the second gene in the operon. We amplified yfiJK using primers 54 and 75 from wild-type or spontaneous LDAR mutant and the plasmid backbone of pDR111 (amyE::spc, without the IPTG-inducible system) using primers 59 and 74. We combined these products together using Gibson assembly [93], transformed the plasmid into PDS0546, and verified insertion into the amyE locus by PCR. We moved these constructs into PDS0554 using SPP1 phage transduction.
To complement yfiLMN we first amplified Pspac(c) from BJH157 using primers 112 and 113. These primers included an EcoRI and SpeI site, which we used to clone the Pspac(c) fragment into pBB275 to generate pRMS1. We amplified yfiLMN using primers 120 and 121 and the backbone of pRMS1 using primers 118 and 119. We assembled these fragments using Gibson assembly and transformed them directly into PDS0652 to generate PDS0717.
We used primer-mediated site-directed mutagenesis to generate phosphoacceptor residue changes. To generate yfiJH201N alleles we used primers 42 and 43. To generate yfiKD54A alleles we used primers 50 and 51. Briefly, we PCR amplified plasmids containing yfiJ or yfiJK using overlapping primer pairs that included a single nucleotide change, DpnI-digested the reactions, and transformed E. coli. We isolated the plasmids and sequenced them to verify the mutation. We used plasmids containing the mutations to transform B. subtilis as above.
To observe lysis, we grew cultures of B. subtilis as above and spotted 1 μL of B. subtilis on 20 mL MYM plates. We then spotted 5 μL of a 109 spores/mL stock of S. Mg1 ~6 mm from B. subtilis. These plates were incubated at 30°C and monitored every 24 h.
To observe the effect of yfiJ alleles on motility we used a modified version of a motility assay we previously described [18]. We plated 2.5 μL of a 107 spores/mL stock of S. Mg1 on a 25 mL MYM plate and incubated the plate at 30°C. After 12 h of growth, we spotted 1.5 μL of B. subtilis, grown as above, perpendicularly to S. Mg1, returned the plates to the 30°C incubator, and monitored the plates every 24 h.
To measure MLC values we used an agar diffusion assay. We grew cultures of a LDA sensitive strain (PDS0571) and a LDAR strain (PDS0572) in 25 mL of MYM to OD600 = 2. When the cultures reached this density, we centrifuged the cultures at 3220 x g for 5 min and resuspended the cell pellet in half the volume to reach OD600 = 4. We mixed 1.5 mL of resuspended cells with 4.5 mL of MYM agar (0.67%) and poured the layer over a MYM plate to generate an overlay with OD600 = 1 and 0.5% agar. We placed 6 mm filter paper discs onto the overlay and added 10 μL of 2-fold serial dilutions of daptomycin, ECO-02301, and LDA to the discs. Afterwards we incubated the plates for 4 h at 30°C and then photographed the plates. We measured haloes of lysis using ImageJ [94] and determined MLC values by plotting natural log-transformed antibiotic concentrations versus the area of lysis, and calculated the intercept to determine MLC values for the lytic agents [95].
We grew two independent cultures each of yfiJK+ (PDS0627) and yfiJA152EK (PDS0685) strains as above. When the cultures reached early stationary phase we diluted them 10−3 in LB and spread plated 100 μL on MYM plates. After 24 h we scraped the lawns of B. subtilis into RNAprotect Bacteria Reagent (Qiagen) and isolated RNA using an RNeasy mini kit (Qiagen). We removed trace DNA from the RNA samples using a Turbo DNA-free kit (Applied Biosystems). The ribosomal RNA was removed from RNA samples using a Ribo-Zero rRNA Removal Kit (Gram-Positive Bacteria) (Illumina). 50-bp single-end reads libraries were prepared using a TruSeq Stranded Total RNA Kit (Illumina) and sequenced on an Illumina HiSeq 2500. We mapped reads to each open reading frame (ORF) in the B. subtilis 168 genome (GenBank: NC_000964.3) with kallisto [96] and used edgeR [97] for differential gene expression analysis. We filtered out lowly expressed ORFs (<1 count per million and only represented in one of the four samples) and used trimmed mean of M-value normalization to calculate effective library sizes before analysis [98]. We used the single-factor exact test and reported differentially expressed genes with a false discovery rate cutoff of < 1−4 [99]. The raw reads for this experiment are accessible from NCBI BioProject Accession PRJNA29593.
We isolated RNA from PDS0627 and PDS0685 as above and preformed qRT-PCR similarly as previously described [18]. Briefly, we used 100 ng of total RNA as template for cDNA synthesis using a High-Capacity RNA-to-cDNA Kit (ThemoFisher Scientific). We used an SsoAdvanced Universal SYBR Green Supermix Kit (Bio-Rad) for and preformed quantitative PCR with a CFX96 Touch real-time PCR thermocycler (Bio-Rad). We used the following cycling parameters: denaturation at 95°C for 30 s; 40 cycles of denaturation at 95°C for 15 sec, annealing at 58°C for 30 s, and extension at 72°C for 30 s; and a final melting curve from 60°C to 95°C for 6 min. We used gyrB as our reference gene. We amplified yfiL using primers q1 and q2 and gyrB using primers gyrB qPCR-fwd and gyrB qPCR-rev (S3 Table). We ran each reaction in triplicate. Using LinReg [100] we calculated primer efficiency and quantification cycle values. We normalized yfiL abundance to gyrB and report fold difference relative to PDS0627.
|
10.1371/journal.ppat.1007701 | Mutational pathway maps and founder effects define the within-host spectrum of hepatitis C virus mutants resistant to drugs | Knowledge of the within-host frequencies of resistance-associated amino acid variants (RAVs) is important to the identification of optimal drug combinations for the treatment of hepatitis C virus (HCV) infection. Multiple RAVs may exist in infected individuals, often below detection limits, at any resistance locus, defining the diversity of accessible resistance pathways. We developed a multiscale mathematical model to estimate the pre-treatment frequencies of the entire spectrum of mutants at chosen loci. Using a codon-level description of amino acids, we performed stochastic simulations of intracellular dynamics with every possible nucleotide variant as the infecting strain and estimated the relative infectivity of each variant and the resulting distribution of variants produced. We employed these quantities in a deterministic multi-strain model of extracellular dynamics and estimated mutant frequencies. Our predictions captured database frequencies of the RAV R155K, resistant to NS3/4A protease inhibitors, presenting a successful test of our formalism. We found that mutational pathway maps, interconnecting all viable mutants, and strong founder effects determined the mutant spectrum. The spectra were vastly different for HCV genotypes 1a and 1b, underlying their differential responses to drugs. Using a fitness landscape determined recently, we estimated that 13 amino acid variants, encoded by 44 codons, exist at the residue 93 of the NS5A protein, illustrating the massive diversity of accessible resistance pathways at specific loci. Accounting for this diversity, which our model enables, would help optimize drug combinations. Our model may be applied to describe the within-host evolution of other flaviviruses and inform vaccine design strategies.
| The spectrum of viral mutants that exists in infected individuals defines the diversity of drug resistance pathways accessible to any virus. Drug combinations that block these pathways the most effectively are likely to elicit the best responses. The mutants may lie below detection, rendering treatment optimization difficult. We constructed a multiscale mathematical model to estimate the pre-treatment frequencies of the entire spectrum of hepatitis C virus mutants at specific resistance loci. We described intracellular evolution stochastically and extracellular dynamics deterministically, gaining accuracy without escalating computational costs. Model predictions quantitatively captured experimental observations, explained confounding inter-subtype differences, and unraveled the massive diversity of accessible resistance pathways. Our study would help describe viral evolution more accurately, optimize drug treatments and design vaccines.
| Direct acting antiviral agents (DAAs) have revolutionized the treatment of chronic hepatitis C virus (HCV) infection, eliciting nearly 100% cure rates in clinical trials with oral treatments often lasting as short as 8 weeks [1]. Efforts are now focused on identifying DAA combinations that prevent the development of drug resistance more effectively and can reduce treatment durations further [2–8]. Mutations that confer resistance to individual DAAs, termed resistance-associated amino acid variants (RAVs), have been identified [9]. The frequencies with which RAVs are likely to exist in individuals before treatment are important to the identification of optimal DAA combinations; DAAs must effectively block the growth of these pre-existing drug resistant strains during treatment [10–12]. Triple-DAA combinations were found recently to lower the likelihood of the development of resistance significantly compared to double-DAA combinations [5]. Current assays are inadequately equipped to estimate the frequencies of minority strains. The assays can detect mutants with frequencies up to ~0.1% [13, 14]. With typical baseline viral loads of 106 copies/ml in chronic infection [15], a mutant frequency of 0.01% would imply ~100 mutant copies/ml, which would go undetected but can be sufficient to cause treatment failure. Indeed, a recent study has argued, using phylogenetic analysis, that resistance to a new DAA observed in a longitudinal study was due to undetected pre-existing RAVs [16]. Mathematical modelling may provide an alternative route to estimating the frequencies of such minority variants and aid the identification of optimal DAA combinations.
Mathematical models have played a crucial role in describing hepatitis C viral kinetics and drug action and have guided treatments [17]. Following the advent of DAAs, the models have been extended to describe the development of drug resistance and to define optimal drug combinations [5, 18–20]. The models, however, are adaptations of models of HIV dynamics [21, 22] and therefore present approximate descriptions of HCV evolution and DAA treatments. Two key challenges must be overcome to develop an accurate model of within-host HCV evolution and estimate the pre-existing frequencies of RAVs.
First, HCV evolution is a multiscale phenomenon, with selection both at the intracellular and extracellular levels. This represents a departure from HIV evolution: An HIV infected cell typically carries a single integrated provirus and produces identical virions [23]. Selection therefore occurs largely at the extracellular level. In contrast, HCV undergoes continuous replication, mutation, and selection within each infected cell [24–26], resulting in potentially diverse progeny virions from each infected cell. Further, each infected cell carries a few hundred HCV RNA copies [27], which makes this evolutionary process strongly stochastic. Finally, infected cells have short lifespans (a few days [28]), which may not allow intracellular evolution to achieve a steady state. Mutation-selection balance, which underlies most current models [18, 21], where the frequency of resistant strains is determined by the balance between mutation of the wild-type yielding the mutant and selection against the wild-type eliminating it, is thus unlikely to hold and founder effects may dominate. Extracellular dynamics, however, is expected to be like HIV, captured by current HCV kinetics models [18, 29–31]. Accurate integration of intracellular and extracellular evolution has been an outstanding challenge [16, 25].
Second, although the positions where mutations confer resistance to DAAs are well defined, the mutations at those positions are not unique [9, 12]. For instance, at the position 155 on the NS3 gene, any of the mutations R155K/I/G/M/T/Q/C/W/N could confer resistance to several NS3/4A protease inhibitors, namely, boceprevir, telaprevir, simeprevir, asunaprevir, paritaprevir, grazoprevir, glecaprevir, and voxilaprevir [9, 32]. An entire spectrum of mutations at the R155 position, thus, can lead to treatment failure, with each mutation representing a potentially independent resistance pathway. Similarly, the mutations Y93H/C/N/R/W/S/T all lead to resistance to the NS5A inhibitors daclatasvir, ledipasvir, ombitasvir, elbasvir, velpatasvir, and pibrentasvir [9, 32]. While R155K is often detected pre-treatment, the other RAVs at this position are not [33]. Accurate estimation of the likelihood of the development of resistance to different DAAs would require quantification of the frequencies of the entire spectrum of RAVs that may exist in a chronically infected individual. Current models have not been designed for this; they are restricted to either the most prominent or the fittest few RAVs or lump all the RAVs into a combined mutant species [18–20].
Here, we constructed a model that overcame both these challenges. Our model could thus estimate the frequencies of the entire spectrum of variants at chosen loci, defining accessible resistance pathways and presenting a framework for the comparative evaluation of DAA combinations.
We constructed a multiscale model of HCV kinetics with stochastic intracellular viral replication and evolution coupled with deterministic extracellular population dynamics (Fig 1A). We represented the viral genome as a string of nucleotides (Fig 1B). We restricted the string to loci where mutations can give rise to resistance to a DAA. We considered genomes carrying all possible mutations at these loci. For instance, for a hypothetical string of two loci, 6 genomes carrying single mutations and 9 carrying double mutations were possible (Fig 1B), all of which were considered in our model. When a single codon associated with resistance to a DAA was considered, a total of 43−1 = 63 different genomes carrying different single, double, and triple mutations became possible. Virions carrying each of these genomes could exist in the viral population in an infected individual. The distribution of these genomes in the population would define the spectrum of mutations at the locus. We quantified this spectrum as follows (see Methods for details). We first performed stochastic simulations of intracellular evolution with each one of the possible genomes as the infecting strain and estimated the probability that the strain established productive infection and, when it did, the distribution of different genomes in progeny virions. Performing a million realizations with every infecting strain, we estimated the mean relative infectivity, λj, of each strain j and the specific release rate, pij, of virions containing genomes i from cells infected with strain j for all combinations of i and j. The simulations involved replication of positive- to negative-strand RNA and vice versa, mutations, distinguished into transitions and transversions, fitness selection, and progeny virion production. The quantities λj and pij provided inputs to our deterministic model of extracellular dynamics. These quantities modified the standard model of viral kinetics by accounting for the effects of mutations on viral infectivity and the distribution of genomes in progeny virions. Solving the resulting equations, using parameters representative of HCV infection in vivo (Table 1), we obtained the within-host frequencies of all variants, quantifying the spectrum of mutants at any chosen loci.
We first considered the position 155 in the NS3 protease of HCV, where mutations yield resistance to NS3/4A protease inhibitors, such as telaprevir [9, 12]. The wild-type HCV genotype 1a contains the amino acid arginine (R) represented by the codon AGG at this position [12]. We performed stochastic simulations of intracellular evolution with the infecting strain containing the codon AGG. Mutations could yield different amino acids, such as lysine (K) and threonine (T). The relative fitness of the RAVs at this position has been estimated previously; only the RAVs K, T, and methionine (M) had non-zero fitness [19]. We employed these fitness values in our simulations. (These fitness values were found to correlate well with estimates from in vitro studies [19, 34]. Using the latter in vitro values, which were available for a wider set of RAVs, made the computations more complex because of the presence of the additional mutational pathways, but did not change our estimates significantly (S1 Fig).) We examined individual realizations and found that in most realizations the population of the infecting genome rose from one to nearly the carrying capacity of the cell, where it stabilized (Fig 2A). Other genomes were rarely present. The time when the population began to rise, indicating the onset of viral replication, varied significantly across cells, with some cells seeing the rise soon after infection whereas others seeing it as late as 40 h after infection. The initiation of replication was thus subject to strong stochastic fluctuations. If the infecting genome were to be degraded before the initiation of replication, the cell would cease to be productively infected (see below).
In some realizations, where the infecting genome experienced a mutation early on, the population came to be dominated by the mutant, which reached the carrying capacity and stabilized. In a small minority of realizations, where the mutant population was on the rise, a reverse mutation leading to the infecting genome occurred. The infecting genome then grew at the expense of the mutant because of its higher relative fitness. Eventually, the infecting genome came to dominate the population and the mutant died down, a pattern akin to the replacement of a less fit strain following superinfection with a fitter strain [35].
Thus, three patterns of intracellular evolution were evident (Fig 2A). The first, which occurred in a vast majority of the realizations, was where the infecting genome dominated the population; the second, which occurred in a minority, was where the mutant dominated; and the third, which occurred in a smaller minority, was where the mutant dominated initially but was eventually outcompeted by the infecting genome.
We examined next the average evolution across a large number (106) of realizations. We found that the intracellular population was dominated by the infecting strain, which existed at levels close to the carrying capacity of ~200 genomes per cell (Fig 2B). The mutants were present in a small minority, ranging on average from 10−1 to 10−4 genomes per cell; i.e., one mutant-dominated cell in 10 to 10000 infected cells. The types of mutants present and their frequencies again indicated strong founder effects. All the mutants present were single mutants; double and triple mutants were hardly observed. Further, even the mutations that were synonymous, such as AGA, which did not lead to a fitness penalty, were present in extremely small numbers. This implied that mutations occurred rarely, as expected [25], and cells predominantly carried viral genomes of the type that infected them. Simulations with a two-locus/two-allele model, which were simpler but easier to visualize, corroborated these results (S2 Fig).
For the infecting strain AGG, five single mutants with non-zero fitness were possible: CGG, AGA, AAG, ACG, and ATG. Of these, CGG and AGA were synonymous–encoding R–and so introduced no fitness penalty. Yet, they were present at different frequencies, with CGG several orders of magnitude lower than AGA (Fig 2B). This was because CGG required a transversion from AGG, whereas AGA could be produced by a transition. The higher probability with which the latter could be produced thus resulted in the different frequencies. The other three single mutants encoded the amino acids K, T, and M, respectively, which had fitness decreasing in that order (Fig 2B inset). Further AAG required a transition, whereas ACG and ATG required transversions. Thus, AAG was present in higher frequencies than the other two. It was also present at a higher frequency than CGG, which had a higher fitness but required a transversion. CGG, however, was present at a frequency higher than ACG and ATG, the latter present at similarly low frequencies, dictated by their low fitness and the low transversion rate.
The distribution of replication complexes too followed the same trends, with the wild-type dominant and single mutants alone present in small minorities with the ordering of the mutant frequencies defined by the relative fitness and whether a transition or transversion to the infecting strain was required (Fig 2C). Accordingly, the progeny virions released were also predominantly of the type that contained the wild-type genomes (Fig 2D). This transition-transversion bias is consistent with previous studies [36].
Together, these findings implied that strong stochastic and founder effects resulted in the dominance of the infecting strain within cells. The mutation-selection balance, often invoked to describe the frequencies of mutant strains [18, 21], did not hold. Had the mutation-selection balance been achieved, the population would have been dominated by the fittest strain, the wild-type, regardless of the infecting strain. The small intracellular carrying capacity, the low mutation rate, and the short lifespan of infected cells together precluded the mutation-selection balance from being established.
We repeated the simulations above with every strain, 13 in all, that had a non-zero relative fitness as the infecting strain and estimated the relative infectivity, λj, and the specific release rate, pij, which provided the necessary inputs to the extracellular model. We found that λj was dependent on the amino acid of the infecting strain and not the codon and decreased as the fitness of the infecting strain decreased (Fig 3A).
When productive infection did occur, pij increased overall with the fitness of the infecting strain (Fig 3B; S3 Fig; S1 Table). Thus, more virions were produced from an infected cell on average when the infecting strain was AGG than ATG. The virions produced, however, were predominantly of the type that contained the infecting genome regardless of fitness (Fig 3B); i.e., for any infecting strain j, pjj>pij. For instance, even with ATG as the infecting genome, which had the least relative fitness (Fig 2B inset), the dominant progeny virion type was the one containing ATG (Fig 3B). Further, pij dropped to zero for all i removed from j by more than one mutation; i.e., no genomes containing more than one mutation in the infecting strain were produced. Finally, pij was lower for values of j that required a transversion from i compared to those that required a transition, reiterating the transition-transversion bias. pij (Fig 3B and S3 Fig) are collated in a heat map (Fig 3C).
Using λj and pij estimated thus, we solved our model of extracellular dynamics.
We let infection begin with a founder virion containing the wild-type genome with the codon AGG. The viral population quickly rose and, in a few weeks, reached a set point of approximately 1011 virions in the infected individual (Fig 4A), consistent with observed viral loads in chronically infected individuals [15]. The population consisted predominantly of virions containing AGG. 12 different mutants, corresponding to amino acids with non-zero fitness, were also present but in much lower numbers. The mutant numbers ranged from ~103 to ~109 virions in the individual, yielding frequencies of approximately 10−8–10−2, during the first few months of the infection.
To understand this wide distribution of mutant frequencies, we constructed a map of mutational pathways (Fig 4B). The map grouped codons separated by the same number of mutations from the wild-type into distinct layers, indicated with increasingly lighter shades of gray. Thus, the single mutants, CGG, AGA, AAG, ACG, and ATG, formed the first layer next to the wild-type. These were all the mutant codons that could be produced from a cell infected with the wild-type. The codons are colored based on the amino acids they encode. They are connected to the wild-type by solid or dashed lines depending on whether the mutation involved is a transition or transversion. Accordingly, codons connected with solid lines were more likely to be produced than those with dashed lines. In the same way, we connected codons in the second layer, the double mutants, with those in the first layer. Codons within a layer were also connected when they were removed from each other by a single mutation. The resulting map provided a complete set of accessible mutational pathways at the locus in consideration. Using the map, we understood the distribution of mutants predicted by our model as follows.
Strong founder effects at the intracellular level implied that most infected cells would carry and produce virions containing the wild-type. A small fraction of the virions produced would be single mutants, determined by the specific release rate estimated above (Fig 3C). These mutants would in turn infect other cells with probabilities determined by their relative infectivity (Fig 3A). The latter cells would produce virions predominantly containing the respective single mutants. A small percentage of the progeny would yield double mutants, which would in turn infect cells and expand their population. Among the single mutants, the easiest to produce were the ones that required transitions, viz., AGA and AAG. Of these, AGA encodes R–it has a synonymous mutation–and therefore was as fit as the wild-type, whereas AAG encodes K and was less fit. Among the mutants, we thus found AGA present in the highest numbers. Next in numbers were CGG and AAG. CGG involved a transversion and was therefore harder to produce than AAG, but was synonymous and therefore fitter than AAG. Well below these numbers were the other single mutants, ACG and ATG, which required transversions and were much less fit.
The ordering of the double mutants can again be understood following the above arguments. CGA had the highest fitness (R). It was also produced by a transition and a transversion from the single mutants CGG and AGA, respectively, which had the highest numbers among single mutants. CGA therefore occurred in the highest numbers among the double mutants. The other double mutants with the highest fitness (encoding R), CGC and CGT, both required transversions from their single mutant parent CGG, and were therefore much less represented than CGA. Much higher than them was the double mutant AAA, which could be produced by transitions from two single mutants, AGA and AAG, the former present in high numbers. Yet, it was less prevalent than CGA because it encoded K and was thus less fit. Of the three other double mutants possible, ACA, ACT and ACC, only the former was observed, in low numbers, because it was produced by a transition from its parent single mutant ACG, which in turn was present in small numbers because of its low fitness. The latter two double mutants, although they were as fit as ACA, were not observed at all (their numbers were below one) because they had to be produced by transversions from ACG, which given the strong founder effects was not realized.
Thus, the spectrum of mutants was determined not by the fitness of the mutants alone, but also by founder effects and the mutational pathways involved. Mutants that could be produced by transitions via multiple pathways involving well represented ancestral mutants were present in significant numbers even if their fitness was low.
The long-term dynamics was dictated by fitness effects. Gradually, all the mutants encoding the same amino acid converged to the same frequency. The frequencies were then ordered according to the fitness of the amino acids. This long-term evolution, however, was slow and required many tens of years because of the absence of a fitness difference between the codons encoding the same amino acids (Fig 4A).
We applied our model to estimate the frequency of the R155K mutant and compared it with experimental observations. No experiments have thus far measured the entire spectrum of mutants at any locus in vivo. Measurements in infected individuals sample a few genomes (e.g., [16]), which may leave estimates of frequencies subject to uncertainties. Besides, the frequencies are typically below current assay detection limits. We therefore considered the frequency of mutants in public HCV sequence databases, which we expected to be representative of the frequency in typical HCV infected individuals. From 3328 sequences of HCV genotype 1a in public databases, the position 155 was found to have the wild-type amino acid R in 99.82% of the sequences and the mutant K in 0.03% of the sequences [33]. The study was published in 2008, before the advent of DAAs, so that transmitted drug resistance may be ignored.
To compare with these findings, we grouped our codon distributions into their respective amino acid populations and computed the frequencies. We found that although the populations of genomes carrying different codons encoding the same amino acid were gradually varying (Fig 4A), their total populations had nearly reached steady levels (Fig 4C). We found from these steady populations that the frequency of the R155K mutant was 0.03% (Fig 4D), in excellent agreement with the frequency in the public databases [33], giving us confidence in our model predictions. The other mutants, R155T/M, were present at far lower frequencies of ~0.0001%. Previous models underpredict mutant frequencies (S4 Fig), highlighting the improved accuracy of our model.
We applied our model next to two clinically relevant questions, which also highlighted the novelty, scalability, and the wider applicability of our approach.
An intriguing clinical observation has been the significantly lower detection rate of the R155K RAV in HCV genotype 1b infected individuals compared to genotype 1a infected individuals and therefore better responses to NS3/4A inhibitors in the former [10, 12]. To understand this difference, we performed calculations that mimic infection with HCV genotype 1b. The calculations above mimicked genotype 1a infection. We could use the intracellular results above (Fig 3) because the same codons were involved in genotype 1b infection. We assumed that the relative fitness was the same as genotype 1a. The extracellular dynamics had to be recomputed using the appropriate founder strain. The wild-type codon for R in genotype 1b at the position 155 is CGG [12]. The mutational pathway map with this wild-type indicated that no viable single mutants were non-synonymous (Fig 5B). The mutant AAG, which encodes K, required two mutations to CGG. Further, the first of these mutations was a transversion to AGG, which was followed by a transition to AAG. Accordingly, the viral population was not only dominated by the wild-type, the mutant spectrum was dominated by the single mutants which were all synonymous (Fig 5A). The resistant mutant AAG was present in very low numbers, ~104−105 virions, in contrast to the ~108 virions in HCV genotype 1a infection (Figs 4A and 5A). Aggregating the codons into their respective amino acids (Fig 5C), we found that the frequency of the R155K mutant was ~0.0001%, nearly 100-fold lower than the corresponding frequency in the case of genotype 1a (Fig 5D). The other mutants (R155T/M) were present at far lower frequencies. This significantly lower presence of the RAVs in HCV genotype 1b compared to 1a presents a plausible explanation of the less frequent detection of RAVs in individuals infected with the former and may contribute to their better response to NS3/4A inhibitors.
To demonstrate the scalability and wider applicability of our model, we considered another class of DAAs, NS5A inhibitors. Unlike NS3/4A inhibitors, which fail predominantly due to the RAV R155K, the NS5A inhibitor daclatasvir has been observed to fail due to the growth of different RAVs in different individuals [37]. The RAVs Y93H/C/F/N have all been associated with daclatasvir resistance [9, 37]. Further, the RAVs are rarely detected pre-treatment but grow rapidly during treatment, indicating that they are present pre-treatment below detection limits [37]. To understand these observations, we applied our model to define the spectrum of mutants at the position 93 in the NS5A region of HCV.
In a comprehensive study recently, the fitness of all single mutants, carrying every one of the 20 amino acids at every position of the NS5A protein, have been estimated experimentally [38]. We employed the fitness data pertaining to position 93 (S5 Fig). Based on amino acids with non-zero fitness, we found that 44 different codons could potentially exist at this locus in an infected individual. The problem is thus of a much larger scale than the R155 case above, where only 13 codons existed. To estimate the mutant frequencies, we first performed our stochastic intracellular simulations with each of the 44 codons as the infecting strain (S6 Fig) and estimated the relative infectivity (A) and the specific release rates (Fig 6B and S7 Fig). With these values, we solved our extracellular model and estimated the populations and frequencies of each of the variants at steady state. We constructed the mutational pathway map, involving 6 single mutants, 18 double mutants, and 19 triple mutants, connected via transitions and transversions (Fig 7A). The pathways explained the frequencies of the mutants we observed (Fig 7B).
Of the 6 single mutants, one was synonymous and thus contributed to the frequency of the wild-type amino acid, Y. Of the remaining 5 single mutants, 2, encoding the amino acids H and C, were produced by transitions of the wild-type codon, whereas the other three, encoding the amino acids F, N, and S, were produced by transversions. Indeed, H and C were the variants with the highest frequencies, ~0.01%. The next highest were F, N, and S, with frequencies of ~0.001%. The rest of the variants, involving 7 different amino acids, were present at lower frequencies ranging from 10−8–10−12. (One of the mutants, M, had a frequency far below 10−12, resulting in a mean virion number much less than 1 in a typical individual; the mutant is thus expected to occur rarely.) This distribution of frequencies thus defined the spectrum of mutants at the position 93 of NS5A within an HCV infected individual. It indicated that the Y93H and Y93C were most likely to be detected pre-treatment because of their high pre-treatment frequencies. The frequencies, however, were below detection limits of current assays, explaining why they are not typically detected. Similarly, RAVs containing each of the 13 amino acids are expected to exist in an infected individual below detection limits. The RAV that would lead to viral breakthrough during treatment would depend on the fitness of the RAVs in the presence of the drug, defined by the extent of resistance or increase in IC50 values relative to the wild-type [38]. The RAV with the most increase in IC50 may drive treatment failure. Thus, the wide spectrum of mutants renders a variety of resistance pathways accessible to the virus in vivo.
Treatment options for chronic hepatitis C are increasing rapidly as many new DAAs have been approved for clinical use recently and many are in advanced stages of development [11]. At the same time, the demand for DAAs is set to rise sharply with growing evidence of their success in the real world [39] and with >98% of the ~150 million chronic hepatitis C patients worldwide yet to receive DAA-based treatments [40, 41]. Efforts are therefore underway to develop rational strategies to identify the best combinations of the available DAAs, which would ensure cure while minimizing the treatment duration, cost, and side effects [2–7, 42, 43]. Our study informs these timely efforts. The success of DAAs relies on their ability to prevent the growth of resistance associated viral variants in patients [9]. In this study, we developed a multiscale mathematical model that quantifies the spectrum of such variants that may exist in chronically infected individuals, often below detection limits, before treatment initiation, and thus defines the possible pathways of the growth of drug resistance due to pre-existing variants. DAA combinations that most effectively preclude the realization of these pathways in vivo are likely to elicit the best responses.
Describing within-host HCV evolution has been an outstanding challenge, with many recent studies constructing multiscale models to integrate intracellular and extracellular dynamics [19, 20, 25, 31, 36, 44–46]. The complexity increases manifold because the evolution is strongly stochastic, given the mutation rate of approximately 10−5 per site per replication [25] and the small number of viral RNA, typically a few hundred, that an infected cell carries [27]. Stochastic models of HCV evolution have been constructed [25, 36]. The computational cost of such models increases prohibitively as the genome size or the viral and cell populations considered increases. Concepts such as the effective population size [47, 48] are then invoked to keep the simulations tractable, but this restricts the applicability of the models [48, 49]. Our study presents a novel strategy to overcome this limitation. We performed intracellular simulations fully stochastically and comprehensively, considering every possible genomic variant as the infecting strain. We thus obtained all possible expected “input-output relationships” for individual cells in an infected individual. These input-output relationships for all cells in the individual were coupled by the exchange of free virions through the plasma. Given that the population of free virions in a chronic hepatitis C patient is estimated to be over 1010 [18], the resulting extracellular dynamics could be solved deterministically. Our model thus gains accuracy over current models without a prohibitive escalation of computational cost.
The complexity in our model, resulting from the consideration of all possible genomic variants, is in keeping with recent advances in high throughput and single molecule experimentation. For the first time, a sizeable portion of the fitness landscape of HCV has recently been determined: In a tour de force, the fitness of every mutant of HCV in the NS5A region, obtained by replacing the amino acid at every residue in the protein with every one of the remaining 19 amino acids, one at a time, was estimated experimentally [38]. Further, advances in amplification, detection and sequencing technologies are allowing the identification of every genomic variant produced from an infected cell [50]. Our model is designed to efficiently exploit such data. Using a codon level description of amino acids, combinations of transitions and transversions that lead from any amino acid to each of the other 19 alternatives, a corresponding fitness landscape, and the input-output relationships above, we could predict the frequencies of all possible mutants at given loci, presenting a measure of the scale of the diversity of accessible mutational pathways. Thus, we estimated that 13 different amino acid variants encoded by 44 different codons would exist in the viral quasispecies in an infected individual at the residue 93 of the NS5A protein, presenting 44 different potential routes to NS5A inhibitor resistance. Our model estimated the frequencies of each of these variants and found them all to be below detection limits, highlighting the limitation of current assays and the importance of mathematical models in providing realistic estimates of RAV frequencies. Indeed, in a recent study using ultradeep sequencing, at the residue 31 of the NS5A protein, which is another locus of NS5A inhibitor resistance, although no resistance was detectable pre-treatment in one individual, 3 different RAVs, L31V, L31I and L31M, were detectable in the individual within a week of starting therapy that included the NS5A inhibitor daclatasvir [37].
We compared our estimate of the frequency of the RAV R155K, resistant to NS3/4A protease inhibitors, with corresponding database frequencies [33] and found good agreement, giving us further confidence in our formalism. We recognize that database frequencies are representative of sequences prevalent across patients and may be subject to selection pressures at the population level including transmission bottlenecks. Because we have considered databases collected before DAA treatments commenced, we expect transmitted drug resistance not to be a confounding factor. Further, transmission bottlenecks are expected to influence the viral envelope proteins much more strongly than nonstructural proteins. The database frequencies, which are estimated by sampling a large number of sequences across patients (here ~3000), are thus expected to broadly mimic the pre-treatment mutant frequencies at corresponding loci on nonstructural proteins in a typical patient. Future studies that may employ deeper sequencing techniques than currently available may provide a more direct test of our formalism.
Interestingly, we found that the rank ordering of the frequencies of the various mutants was not dictated by fitness effects alone, in contrast to the classical mutation-selection balance [21]. Strong founder effects offset the influence of fitness in our simulations. Combining the founder effects and the fitness landscape, we could create a map of mutational pathways accessible to any founder strain. Importantly, the maps were different for different founder strains containing the same amino acid but represented by different codons. Thus, HCV genotypes 1a and 1b both contain the amino acid R at the position 155 of the NS3 protein but have different mutational pathway maps because they are encoded by different codons. NS3/4A inhibitor resistance was thus predicted to be far more prevalent with genotype 1a than 1b, which is consistent with the rare detection of RAVs and the better response of the latter to NS3/4A inhibitor treatments [12, 51]. That the difference arises because genotype 1a requires a single transition whereas genotype 1b requires a transversion followed by a transition for the R155K mutation has been recognized earlier [36, 51]. Our model makes quantitative predictions of the frequencies of the mutant in the two cases, which is consistent with observations [33], facilitating more accurate tailoring of treatments for the two cases. Such tailoring may have to account also for the genetic backgrounds in which the RAVs arise, which may be different across the two genotypes, as has been recognized, for instance, with NS5A inhibitor resistance [8, 52, 53].
Currently, systematic resistance testing is not recommended before the start of DAA treatments, due possibly to the ability of DAAs to cure patients regardless of pre-existing RAVs [54]. Only RAVs with frequencies above ~10–15%, which are detectable using population sequencing techniques, have been found to influence treatment outcomes [55]. Our interest in estimating minority RAV frequencies is in optimizing treatments without compromising outcomes. We expect that dosages and/or treatment durations may be reduced beyond current guidelines if RAVs can be ensured to remain responsive with the altered protocols. Indeed, current guidelines do recommend resistance testing, where such testing is reliable and accessible, before the use of NS5A inhibitors [54]. Interestingly, a comprehensive analysis extending over 50 clinical trials showed recently that DAA treatments elicited better responses in treatment naïve individuals than in previous null responders to the combination of interferon and ribavirin [20]. A model based on the premise that greater responsiveness to interferon suppressed the replication space available to HCV and therefore prevented the growth of RAVs was able to quantitatively describe the clinical observations [20], reiterating the importance of RAVs in treatment optimization. By accurately estimating RAV frequencies, our model aids such optimization.
Many recent studies have detected RAVs in a significant fraction of patients pre-treatment [55–57]. This is not in conflict with our predictions of minority RAVs typically lying below detection limits. Where the fitness penalties associated with specific RAVs are not significant, it is possible that they exist well above detection limits. Thus, for instance, while RAVs were detected at the position Q30, no RAVs were detectable at the positions Y93 or L31, all associated with NS5A inhibitor resistance, in 41 HCV genotype 1a infected individuals or 77 HCV genotype 1a infected individuals coinfected with HIV [57]. RAV frequencies may increase in treatment experienced patients, given the weaker interferon responses expected in such individuals [20]. Further, transmitted resistance may also contribute to the observed pre-existence, especially with RAVs to NS5A inhibitors, which are known to last years in patients even in the absence of treatment [55].
We recognize that the identification of optimal DAA combinations requires additional inputs. In particular, the dynamics of the growth of RAVs during treatment must be accounted for. Remarkably, the extent of resistance, in terms of the fold change in IC50 relative to the wild-type, for every single amino acid variant in the NS5A region has been experimentally identified [38]. Extending our model by incorporating the latter data would present an understanding of the most likely pathways of the growth of pre-existing RAVs. A combination of high pre-existing frequency and high level of resistance would decide the most likely pathways. Drug combinations would then be designed to prevent those pathways. Such extensions of our model would also require knowledge of epistatic effects that define the fitness of viral genomes with multiple mutations, which is currently lacking for HCV. Techniques from statistical physics are being applied to develop more comprehensive fitness landscapes [58]. Further, resistance may often arise from new mutations that occur during treatment and not from the growth of pre-existing strains, in which case, either fully stochastic models [25] or models that estimate the waiting times for the emergence of such mutants [59] may have to be developed.
The dynamics during and post-treatment can be complex. In a recent study, the rate of viral load decline during treatment with a second-generation protease inhibitor, MK-5172, and the turnover of drug resistant variants post-treatment were found to be far more rapid than previously expected [16]. The study attributed the rapid decline to the cure of infected cells by the DAA. The rapid turnover of mutants post-treatment was argued to be due to cellular superinfection and the ensuing replacement of less fit strains by more fit ones within superinfected cells. This allowed a new, more fit strain to become dominant swiftly even when a less fit strain had established infection with maximal viremia leaving little “replication space” for the new mutant. HCV is thought to induce a superinfection block [60, 61], which renders such superinfection rare, although strains that exhibit enhanced ability to superinfect can be selected in vitro [35]. The mechanism of the replacement of the less fit strain by a more fit strain is less well understood [35]. Previous studies have speculated that the replacement may occur during cell division, when new replication space is created, and the more fit strain has an advantage in terms of establishing infection in the daughter cells [35, 62]. Models considering the partitioning of viral variants into daughter cells are yet to be constructed. Other immune mechanisms may also influence the dynamics during and post-treatment. For instance, the reduction in viral load due to treatment may reverse immune exhaustion and rejuvenate CD8+ T cell responses [43, 63–65]. This has been argued to contribute to the post-treatment cure of HCV infection in some patients despite detectable viremia at the end of treatment [43]. Whether this leads to responses against temporally dominant viral variants and contributes to the observed rapid turnover of variants remains to be examined. Further, cells that are cured by the treatment are likely to be exposed to interferon secreted when they were infected [66, 67]. Cells exposed to interferon may enter an antiviral state that renders their productive infection less likely [66–68]. HCV subverts this interferon response by introducing a block in translation [68]; the block is released when HCV is cleared and the cell cured [31, 68]. It is conceivable that fitter viral strains are more likely to overcome the interferon response in such cells and reestablish infection [31], which again may contribute to the rapid turnover of viral variants observed. Our study has focused on the frequencies of mutants before the onset of treatment, which are less likely to be influenced by these latter mechanisms.
We envision broader implications of our study. The prevalent paradigm for describing within-host viral evolution is the molecular quasispecies theory [69, 70]. The theory, built originally to describe the origin of life, has shaped the modern view of viral evolution by describing the error-prone self-replication of molecules such as RNA, which constitute viral genomes. The theory, however, assumes a well-mixed milieu of such genomes subjected to common selection forces, which ceases to hold for viruses such as HCV where intracellular and extracellular selection are segregated. Our model thus goes beyond models based on the molecular quasispecies theory [19, 20, 45] by accurately describing and integrating intracellular and extracellular evolution. The resulting formalism may be useful in describing the within-host evolution of other important human viruses, such as dengue, West Nile and Zika, which have a lifecycle similar to HCV. A second implication of our formalism is in vaccine design. Although we have focused here on loci leading to drug resistance, our model can be readily applied to sites of immune escape, allowing estimation of the genetic diversity that vaccine candidates must target [58].
In summary, our study presents a novel approach to estimating the entire spectrum of mutants present in infected individuals, explains several clinical observations associated with chronic hepatitis C, and presents a framework that would aid the rational optimization of modern DAA-based treatments.
We present details here of our multiscale model of within-host HCV dynamics and evolution (Fig 1).
We obtained most parameter values from previous studies (Table 1). We estimated the replication rates, k+ and k-, and the carrying capacity, K, to ensure consistency with the overall population dynamics of viral RNA in cells (S1 Text, S9 Fig). We performed simulations of intracellular dynamics using the Stochastic Simulation Algorithm (SSA) in the software Stochkit 2 [76]. We ensured that 106 simulations were adequate to obtain reliable predictions (S10 Fig). We solved our equations of extracellular dynamics in MATLAB using initial conditions where the target cells were in their uninfected steady state, infected cells were absent and a single virion of the wild-type existed.
|
10.1371/journal.pgen.1004694 | Functional Interaction between Ribosomal Protein L6 and RbgA during Ribosome Assembly | RbgA is an essential GTPase that participates in the assembly of the large ribosomal subunit in Bacillus subtilis and its homologs are implicated in mitochondrial and eukaryotic large subunit assembly. How RbgA functions in this process is still poorly understood. To gain insight into the function of RbgA we isolated suppressor mutations that partially restored the growth of an RbgA mutation (RbgA-F6A) that caused a severe growth defect. Analysis of these suppressors identified mutations in rplF, encoding ribosomal protein L6. The suppressor strains all accumulated a novel ribosome intermediate that migrates at 44S in sucrose gradients. All of the mutations cluster in a region of L6 that is in close contact with helix 97 of the 23S rRNA. In vitro maturation assays indicate that the L6 substitutions allow the defective RbgA-F6A protein to function more effectively in ribosome maturation. Our results suggest that RbgA functions to properly position L6 on the ribosome, prior to the incorporation of L16 and other late assembly proteins.
| Ribosomes are complex macromolecular machines that carry out the essential function of protein synthesis in the cell. The assembly of ribosomal subunits is a multistep process that involves the accurate and timely assembly of 3 rRNA molecules and>50 ribosomal-proteins. In recent years many ribosome assembly factors have been identified in bacterial and eukaryotic cells; however, their precise functions in ribosome biogenesis are poorly understood. We have previously shown that the GTPase RbgA, a protein conserved from bacteria to humans, is essential for ribosome assembly in Bacillus subtilis. Here, we show that growth defect caused by a mutation in RbgA is partially suppressed by mutations in ribosomal protein L6. The suppressor strains accumulate novel ribosomal intermediates that appear to suppress the RbgA defect by weakening the interaction of L6 for the ribosome and facilitating RbgA dependent assembly. Our work provides evidence for a functional interaction between ribosome assembly factor RbgA and ribosomal protein L6 during assembly, a function that is likely important for mitochondrial, chloroplast, and eukaryotic ribosome assembly as well.
| The assembly of the 30S and 50S ribosomal subunits is a complex and tightly coordinated series of events that consists of the synthesis, processing and modification of 5S, 16S and 23S rRNA and the addition of more than 50 ribosomal proteins (r-proteins) [1], [2], [3]. The in vitro reconstitution of a mature 50S subunit has been extensively studied in Escherichia coli and the formation of a mature 50S subunit from its constituent r-proteins and rRNA is a multi-step process that requires non-physiological conditions such as high ionic concentration, high temperatures and long incubation times [4], [5], [6], [7]. Relatively fewer studies focused on ribosome assembly in other bacterial species, such as Geobacillus stearothermophilus, and these demonstrated that the intermediates formed in this system are different than those in E. coli, however similar non-physiological steps are required for formation of a functional ribosomal subunit [5], [8]. Moreover, recent studies have utilized biophysical techniques to study ribosome assembly in vivo and demonstrated that assembly of the ribosome subunits is a multistage process that appears to follow multiple parallel pathways in which the accumulation of assembly intermediates identified in vitro do not accumulate in vivo [9], [10], [11]. The slow kinetics and attenuated efficiency of in vitro assembly strongly suggest that assembly factors are involved in vivo and indeed, several classes of assembly factors such as GTPases, RNA helicases, RNA modification enzymes and chaperone proteins have been implicated in in vivo ribosome assembly in bacterial and eukaryotic cells [2], [12], [13], [14], [15]. However, while studies show that these factors are functionally significant and play a critical role in ribosome assembly, the molecular functions of these factors remain elusive. RbgA (ribosome biogenesis GTPaseA) is an essential GTPase that is required for a late step in assembly of the 50S subunit in Bacillus subtilis [16], [17]. RbgA is a widely conserved protein and its eukaryotic homologs such as Mtg1, Lsg1, Nug1 and Nog2 have also been implicated in assembly of the large ribosomal subunit [18], [19], [20], [21]. RbgA depleted cells do not form mature 50S subunits but instead accumulate a 45S complex. Quantitative mass spectrometry analysis of this particle shows that the 45S completely lacks ribosomal proteins L16, L28, and L36 and contains severely reduced amounts of L27, L33, and L35 [16], [22]. Proteins L16 and L27 are crucial components of the peptidyltransferase center in 50S subunit and directly contact the A-site and the P-site respectively [23], [24]. Functional studies have shown that both proteins play a role in stabilization of the peptide bond formation, the positioning of tRNA on their respective sites and are required for optimal functioning of the ribosome [25], [26], [27]. While there have been no reports of deletion of L16, the deletion of L27 in E. coli causes a severe growth defect [28]. However, studies in B. subtilis indicate that both proteins are essential and deletion mutants could not be obtained for either protein [29]. In vitro assembly experiments have demonstrated that incorporation of L16 into the growing complex occurs at a late stage in the assembly process and is accompanied by a large conformational change [30]. In yeast, the RbgA homolog Lsg1 has been proposed to play a role in the incorporation of the L16 homolog Rpl10 into the large ribosomal subunit, suggesting that RbgA and its homologs regulate an evolutionarily conserved step during biogenesis [31], [32]. RbgA has been shown to interact directly with both the 45S complex and the 50S subunits and the GTPase activity of RbgA is enhanced ∼60 fold in the presence of the mature 50S subunit [33]. Mutational analysis of RbgA has shown that a stretch of 15 amino acids in the N-terminal domain, which is largely conserved among all bacterial RbgA homologs as well as eukaryotic homologs, plays a crucial role in this GTPase activity [34]. Mutations that affect GTP hydrolysis result in the accumulation of the 45S complex similar to RbgA depleted cells indicating that GTP hydrolysis plays a key role in maturation of the 50S subunit [34].
To further investigate the role of RbgA in the assembly of the 50S subunit we constructed a B. subtilis strain that expressed a mutated RbgA protein that results in a severe growth defect and screened for suppressors that alleviated this growth defect. We isolated and characterized eight independent suppressor strains and found they contained six distinct mutations in the rplF gene, which encodes for ribosomal protein L6. Analysis of ribosome assembly in these strains led to discovery of a novel ribosomal intermediate that differs from the 45S complex observed in the parental strain and also in RbgA-depleted cells. We discuss the implications of these results and present a model for the role of RbgA in assembly of the 50S subunit.
To generate a strain that displayed a strong growth defect that would be amenable to suppressor analysis, we analyzed the phenotypes of over 40 site-directed mutations in the rbgA gene [34]. We were interested in identifying substitutions in RbgA that displayed reduced GTPase activity upon association with the ribosome and were still able to bind to the ribosome. One such mutation, rbgA-F6A, was identified as meeting both of these criteria. Our results showed that GTPase activity of RbgA-F6A was reduced ∼12 fold, however the mutation did not prevent stable association with the 45S complex and the 50S subunit [34]. Therefore we constructed a strain in which rbgA-F6A was the only functional copy of rbgA in the cell expecting that cells harboring rbgA-F6A would be viable but display reduced growth. To achieve this we constructed strain RB1043 by cloning the rbgA gene (containing a mutation that results in a F6A substitution) fused to its native promoter into the plasmid pAS24 and inserted this construct at the amyE locus (Table 1). A control strain (RB1006) that contains a wild-type copy of the rbgA gene at the amyE locus was constructed in similar manner as a control. The native rbgA gene was inactivated in both strains by the insertion of a MLS cassette by marker replacement, which led to the complete removal of the native rbgA gene. Comparison of the two strains showed that the strain expressing RbgA-F6A (RB1043) was severely growth compromised and exhibited a growth rate ∼7 fold slower than the RB1006 strain (Figure 1A). This severe growth defect was utilized to isolate suppressor mutations that allowed this strain to grow more rapidly.
To isolate independent, spontaneous suppressor mutations we inoculated a single colony of the RB1043 (rbgA-F6A) strain per flask into a total of 50 flasks and isolated suppressors that exhibited faster growth at 37°C (only one per flask). We identified eight independent suppressor strains that partially alleviated the growth defect of RB1043 (Figure 1 and Table 2). Individual suppressors were grown in liquid medium and their growth rates were compared to the parental RB1043 strain and the control strain RB1006. The wild-type control strain RB1006 and the parental RB1043 strains exhibited a doubling time of 23 minutes and 173 minutes, respectively, whereas the growth rate of the suppressor strains ranged from 46 to 77 minutes (Table 2). Next, we sequenced the rbgA-F6A gene to check for reversion mutations and found that all eight strains lacked intragenic suppressor mutations. We then proceeded to backcross each suppressor strain with the wild-type RB247 strain and inactivated the native rbgA gene. The reappearance of RB1043 phenotype (∼7-fold increase in doubling time) in each backcrossed strain indicated that the suppressor mutation was unlinked to the rbgA-F6A mutation.
To identify the genetic changes responsible for the partial suppression of the growth defect we obtained the whole genome sequence of all eight suppressor strains, RB247 (wild-type background) and the parental RB1043. The sequence reads from the parental RB1043 (rbgA-F6A) strain were compared with each suppressor strain sequentially. After accounting for mutations that have arisen in our genetic background or were sequencing errors in the original B. subtilis sequencing project [35], we found that each suppressor strain bore a single point mutation in the ribosomal protein L6 encoding gene rplF gene. Three suppressor strains had the same mutation (Table 2) and thus we obtained six unique suppressor mutations that caused single amino acid substitutions in L6; R3C, G5C, G5S, H66L (3 isolates), T68R and R70P. Alignment of L6 proteins from phylogenetically diverse bacteria indicates that these residues are conserved in bacterial L6 proteins, with T68 demonstrating the most conservation when compared to L6 homologs from archaea and eukaryotes (Figure S1). We constructed a homology model of the B. subtilis L6 protein based on the structure of the L6 protein from Geobacillus stearothermophilus and mapped the suppressor mutations onto the modeled structure of the protein. Our analysis shows that all of the six suppressor substitutions reside in close vicinity in the protein structure (Figure 2) and are contained within the N-terminal structural domain.
To assess the status of ribosome assembly in the suppressor strains, we analyzed the ribosome profiles using 10–25% sucrose density gradients. Our results showed that all of the suppressor strains accumulated a novel ribosomal intermediate that migrated at ∼44S and was distinct from the 45S complex that accumulates in RbgA-depleted cells and RB1043 strain expressing RbgA-F6A (Figure 3). In addition, each suppressor strain exhibited an increased 70S ribosome peak compared to RB1043, corresponding to the increased growth rate of the suppressor strains.
Given the changes in these suppressor strains' intermediate particle migration patterns, we set out to identify compositional differences between the 44S and 45S intermediates by isolating the particles on a sucrose gradient and measuring their r-protein composition using quantitative mass spectrometry (qMS). As described in materials and methods, our SILAC-like approach resulted in multiple independent peptide measurements for the ribosomal proteins. Additionally, standard curves measured with our technique exhibited a linear dose-response between 0.1 and 1.6 r-protein equivalents (Figure S2), providing confidence in the precision of the approach.
Whereas most proteins were present at stoichiometric levels in the intermediates, we found that these particles were severely lacking in proteins L16, L28, L35, and L36 (occupancy<0.4), and were significantly depleted of proteins L27 and L33 (occupancy<0.8) (Figure 4A). These latter depletion effects were more pronounced in the parental RB1043 strain than in any of the suppressor strains, suggesting that these proteins are more efficiently incorporated as a result of the suppressor mutations. Protein L6 showed the greatest variability in protein occupancy across the strains with the parental RB1043 exhibiting full protein incorporation whereas the suppressors RB1051 and RB1068 largely lacked L6 (occupancy<0.3). Interestingly, the intermediates from suppressor strains RB1055 and RB1057 showed more mild L6 depletion effects (occupancy ∼0.8 and 0.6, respectively) despite migrating similarly to the other suppressor strain intermediates. This result suggested that the difference in migration between the 44S and the 45S intermediates arouse from conformational changes in the intermediates and was not a direct result of the extent of L6 incorporation. Finally, we measured protein occupancy in these intermediates from three independent biological replicates and consistently identified that the aforementioned proteins were depleted from the intermediate particles (Figure 4B), confirming the significance of the observed effects.
We then determined the protein composition of each 70S particle from these strains to test whether the protein depletion effects we observed in the intermediate particles persisted into the 70S fractions. The depletion effects observed in the intermediates were largely absent from the mature particles, with only strain RB1057 exhibiting relatively mild occupancy defects (occupancy>0.7) in proteins L6, L16, L27, L28, L35, and L36 (Figure 4A, B). Whether these subtle effects result from instability of the RB1057 70S particles during purification or from a subpopulation of particles that lack these r-proteins in vivo remains to be investigated. In all other strains tested, each r-protein was present at equal stoichiometry with the exception of the rapidly exchanging protein L7/L12 [36], which likely dissociates during particle purification. Taken together, these data affirmed that the mutant L6 proteins, along with the full complement of other large subunit proteins, are integrated during the late assembly stages of the 70S particles.
To test if the L6 levels in the complex influenced the GTPase activity of RbgA, we incubated RbgA with each 44S intermediate and measured the rate of GTP hydrolysis. Our results show that GTPase activity of RbgA is stimulated ∼4–6 fold in the presence of the each 44S intermediate, with no correlation between the hydrolysis rate and L6 occupancy (Figure S3). This increase in GTPase activity is similar to the fold change observed in the presence of the 45S complex and highly reduced compared with the ∼60 fold stimulation observed in the presence of the mature 50S subunit [33].
We were interested in studying the effects of the alterations in ribosomal protein L6 on cell growth and ribosome assembly in an otherwise wild-type background. To do this we created strains in which the rplF mutations were linked to an antibiotic resistance marker and moved into the wild-type background RB247. Once each mutation was transferred into a wild-type background, the antibiotic resistance marker was easily removed by passage on media without selection, resulting in strains that only contained mutations in rplF (see materials and methods for details). We successfully constructed strains in which mutations in rplF resulting in the R3C (RB1125), G5S (RB1131), T68R (RB1133), and R70P (RB1123) substitutions were the only alterations in the chromosome (Table 1). Each L6 mutant strain's growth rate was indistinguishable from the congenic wild-type RB247 strain, demonstrating that the partial suppression of the RbgA-F6A growth phenotype was not due to an impairment of growth due to defects in L6.
Although the rplF mutations did not have an effect on cell growth, we were interested to identify if they had any impact on ribosome maturation. Ribosome profiling of strains RB1123, RB1125, RB1131 and RB1133 through 10–25% sucrose gradients was performed and in each case L6 substitutions resulted in abnormal ribosome profiles (Figure 5, S4). The mutants had increased levels of individual ribosomal subunits when compared to wild-type cells, indicating that the L6 substitutions impacted subunit joining or maintenance of 70S ribosome stability. We further analyzed the 50S subunits that accumulated in these strains and found that both ribosomal proteins L6 and L16 were present in levels similar to wild-type 50S subunits indicating that these substitutions in RplF (R3C, G5S, T68R and R70P) do not impact the association of L6 or L16 with the 50S subunit in the context of wild-type RbgA.
One possible mechanism for how L6 substitutions may suppress the RbgA-F6A defect is that 44S particles may be more easily matured into 50S subunits than 45S particles. To address this possibility we concentrated cellular lysates from RB1043 (rbgA-F6A) and RB1055 (rbgA-F6A, rplF-R3C) and incubated them for 1 hour at either 37°C or, as a negative control, at 0°C. After incubation, these lysates were centrifuged over 18–43% sucrose gradients in the presence of 20 mM Mg2+ (to facilitate mature subunit joining since L6 mutants show subunit association defects in 10 mM Mg2+, see Figure 5) [37]. After incubation of the RB1055 lysate at 37°C, we found that many of the 44S particles in the RB1055 lysate were converted into 50S subunits that subsequently partnered with 30S subunits to form 70S ribosomes (Figure 6A). Indeed, 70S ribosomes showed a more than 100% increase during 37°C incubation with a concomitant decrease in 44S and 30S subunits. The RB1043 lysate yielded a much lower level of 70S formation, with only a 10% increase in 70S ribosomes when incubated at 37°C and similar small reductions in 45S and 30S subunits (Figure 6B). These data were consistent with the hypothesis that 44S particles mature into 50S subunits more quickly than 45S particles in vitro.
Previously, we found that cells depleted of RbgA had very small precursor pools for the r-proteins L16, L27, L28, and L35 [22]. This result suggested that upon RbgA depletion the cell down-regulated synthesis of these proteins through an as-yet uncharacterized mechanism, resulting in very low cytosolic levels of free (unbound) copies of these proteins. To determine if these suppressor strains also lacked free equivalents of these proteins, we directly measured their whole-cell stoichiometry using qMS, specifically, an isotope-label based selective reaction monitoring protocol (SRM) focused on ribosomal peptides (see materials and methods). As predicted by our precursor pool measurements, we found depressed whole-cell protein levels for L16, L27, L28, and L35 in strain RB301 starved for RbgA (Figure 7, S5; RB301:6 µM). In contrast, cells grown with near wild-type levels of this factor exhibited significantly greater levels of each of these proteins (Figure 7, S5; RB301:1 mM IPTG). Assays with strain RB1043, and the L6 suppressors also revealed significant cellular depletion of this set of proteins, indicating that this same regulatory mechanism is activated the RbgA F6A strain and the suppressor mutants. In contrast, protein L36, which is also depleted from the intermediate particles, was abundant in the whole cell lysate. This result is consistent with unregulated synthesis or degradation that results in significant free (unbound) quantities of protein L36.
To determine if the suppressor mutations affected translation or degradation of protein L6, we next inspected the abundance of this protein in each lysate. Interestingly, its level varied greatly between the different suppressor strains with RB1055 lysates bearing ∼2.5 more L6 than those of RB1051. Indeed, the low cellular L6 abundance in RB1051 may explain the low protein occupancy observed in its 44S intermediate particle. Notably, however, with the exception of strain RB1051, L6 abundance in the whole cell lysates correlated poorly with occupancy in the intermediate particles (Pearson's r = 0.37). This result argues that the variable L6 occupancy observed in the 44S particles is not strictly a result of altered translation or degradation of the mutant proteins but, rather, results at least in part from effects of the mutations on protein incorporation or complex stability.
Given the depressed levels of proteins L16, L27, L28, and L35 in the suppressor strain lysates, we were curious if these proteins were in fact incorporated into the 70S particles during our in vitro maturation assays. Using qMS, we measured the protein composition of both the precursor 44/45S and product 70S particles from RB1043 and the suppressor strain RB1055 at 0 or 37°C. Whereas the precursor particles were depleted of L16, L27, L28, L33, L35, and L36 (Figure 8; light circles), we found that the 70S particles contained nearly stoichiometric quantities of each of these proteins (Figure 8; dark circles). Although the extent of maturation in strain RB1055 showed a strong temperature-dependence (Figure 6A), the protein occupancy patterns were effectively temperature-independent indicating that the 70S particles formed during our 37°C in vitro maturation assay were indistinguishable from those formed in vivo and maintained during the 0°C incubation (Figure 8; dark orange, dark red).
We provide evidence that mutations causing substitutions in the N-terminal domain of L6 can suppress ribosome assembly defects associated with a mutation that impairs the function of RbgA. Formally, these L6 suppressor mutations could be acting to facilitate assembly either by allowing the defective RbgA-F6A protein to function more effectively in assembly or by allowing ribosomes to assemble in an RbgA-independent pathway. In testing for an RbgA-independent assembly pathway, we repeatedly attempted to generate an rbgA null mutation in the background of several of the rplF suppressor mutations but were unsuccessful. If the L6 substitutions were able to completely bypass the need for RbgA during maturation then we should have easily isolated null mutations in rbgA in the rplF mutant backgrounds. Additionally, if significant flux where flowing through an RbgA-independent assembly pathway in these mutant L6 strains, we would expect to find 44S particles in strains bearing wild-type RbgA and L6 mutations. Instead, we only identified 50S particles. Critically, these 44S particles, which require RbgA-F6A, can be matured into 70S ribosomes in vitro. Taken together, we propose that the partial suppression of growth and ribosome assembly defects observed are not due to the L6 substitutions completely bypassing the requirement for RbgA in the cell, but rather, that RbgA function is still required for maturation.
L6 is a two-domain protein that is located on the L7/L12 side of the 50S subunit and forms an L-like structure that appears to bridge between the front and the back of the subunit (Figure 9) [38], [39]. The N-terminus of the protein interacts with helix 97 (h97) of the 23S rRNA, while the C-terminus of L6 interacts with the sarcin/ricin loop (SRL) [38], [40], [41]. All of the L6 substitutions that suppress the RbgA-F6A defect map to a small region in the N-terminus of the protein and, in some cases, appear to disrupt direct interactions between L6 and h97 (Figure 9). Although some of the suppressor mutations cause L6 to unstably associate with the 44S intermediate (T68R, R70P), the ability to suppress the RbgA-F6A defect does not seem to correlate with L6 binding as 44S particles isolated from the other suppressors contain near wild-type levels of L6 (Figure 4B). The consequence of the L6 substitutions in a wild-type background appears to be at the level of 70S stability. Individual 50S subunits that contain mutant L6 proteins appear to have normal amounts of both L6 and L16, indicating that once matured, these proteins are stably incorporated. However, clearly there is some disruption of 50S subunit structure that causes decreased stability of 70S ribosomes. This is possibly due to improper positioning of the intersubunit bridge helix 89, which is located between and makes direct contacts with L16 and L6.
What effect might mutations in L6 have in suppressing ribosome assembly defects associated with reduced function of RbgA? L6 binds prior to L16 and has been implicated in setting up the binding site for L16 [4], [42]. In E. coli, the expression of ribosomes that are deleted for the SRL, which interacts with the C terminus of L6, are dominant-lethal and result in the accumulation of 50S subunits that lack L16 [43]. Lancaster et al. propose that L6 binds to the assembling subunit via initial interactions between the N-terminus of L6 and h97, which then results in the subsequent assembly of the functional core of the 50S subunit [43]. This includes the formation of several key interactions between h97, h42, h89, h91, and h95, which are predicted to be initiated by the binding of L6 with h97. When the SRL is deleted these interactions are disrupted and the L16 binding site, along with other functional regions of the large subunit, are improperly assembled and non-functional.
While the precise role that RbgA plays during ribosome assembly is still unknown, the identification of the second-site suppressors in L6 supports a model in which RbgA participates in facilitating the correct association of L6 with the ribosome to allow the subsequent maturation events to take place (Figure 10). Recent studies have postulated that ribosomal subunits can be formed via multiple parallel pathways [44]. We suspect that the large subunit pathways converge on a late assembly intermediate (LAI50-1) and GTPases, such as RbgA, act on LAI50-1 to complete maturation. We envision two scenarios in which RbgA could act on LAI50-1 to facilitate maturation. In scenario 1, RbgA binds to an undefined late assembly intermediate (LAI50-2), and promotes the rearrangement or movement of helix 97 to facilitate the correct incorporation of L6. In scenario 2, L6 binds to the ribosome prior to RbgA (resulting LAI50-3, equivalent to the 45S complex) in an unproductive interaction and the role of RbgA binding is to promote the correct interaction of L6 with the helix 97 [43]. We suspect in RbgA mutants the interaction of L6 with LAI50-2 is reversible and the suppressor mutations in L6 enhance this reversible step by weakening the interaction with the ribosome. Recently, we have shown that the 45S particle is not a dead end particle and can be fully matured into a 50S particle in vivo. The fact that L6 is not fully visible in the cryo-EM structures of the 45S complex provides support that L6 is not in its proper conformation. In both scenarios, correct positioning of L6 and h97 allows for proteins L16, L27, L28, L33, L35, and L36 to be stably incorporated into the large subunit. Once RbgA senses that incorporation of these proteins has taken place GTP hydrolysis occurs, a final maturation event takes place, and RbgA leaves the subunit. Because we have not been able to isolate RbgA mutants that are deficient in GTPase activity that form 50S subunits, we predict that the GTP hydrolysis plays a dual role in both promoting conformational changes in the ribosome while also resulting in RbgA dissociation. Support for this latter step stems from the fact that 50S subunits lacking only ribosomal proteins L16 and L28 do not stimulate the GTPase activity to levels observed with wild-type 50S subunits [22].
Although we do not know the order of binding of L6 and RbgA, in both scenarios the proposed role of RbgA is to properly position L6 and helix 97 to facilitate assembly. This interaction between L6 and h97 is evolutionarily conserved (see Figure S6) and, given that RbgA homologs are present in archaea and eukaryotes, the role of RbgA proteins in ribosome assembly is likely to be conserved as well. Thus it appears that in small subunit and large subunit ribosome biogenesis, one function of assembly factors is to prevent binding of late binding ribosomal proteins until the subunit is ready to receive them [22], [45], [46]. Whether or not these potential checkpoints are related to quality control mechanisms that insure only functional ribosomes enter into translation remains to be seen [22], [47]. Interestingly, E. coli and many other proteobacteria lack RbgA, a function that was present in the last common ancestor and subsequently lost in this lineage of bacteria. We are currently using a comparative genomics approach to identify differences between E. coli and B. subtilis ribosomes in an attempt to further localize the precise site and mechanism of RbgA function.
All strains were grown at 37°C in LB medium and cultures were shaken at 250 rpm. Antibiotics were added at the following concentrations when required: chloramphenicol (5 µg/ml), erythromycin (5 µg/ml), lincomycin (12.5 µg/ml), spectinomycin (100 µg/ml) and ampicillin (100 µgml). IPTG was added to a final concentration of 1 mM when required for strain growth.
Plasmid pMA1 was derived from pSWEET, an amyE insertion vector with a chloramphenicol resistance cassette, by placing the rbgA gene under the control of a xylose inducible promoter. Plasmid pAS24, an amyE insertion vector with a spectinomycin resistance, was used to construct pMG28 by inserting a wild-type copy of rbgA under the control of its native promoter. Plasmid pMG29 bearing a F6A mutation in the rbgA gene (accomplished by a TTC to GCC codon change) was constructed from pMG28 using the QuikChange II XL kit (Stratagene) by following the manufacturer's instructions. Plasmid pJCL87 was derived from pDR111 and contains a chloramphenicol resistance cassette and the IPTG inducible Phyperspank promoter. Plasmid pMG30 was constructed from pJCL87 by cloning the first 330 bp of the map gene under the control of the Phyperspank promoter.
All strains used in this study are derived from the wild type strain JH642 (RB247) and listed in Table 1. The construction of strain RB301 and RB418 has been described previously [16]. RB395 was constructed by transforming RB247 with pMA1 and knocking out the native rbgA gene by using a MLS cassette. Strain RB1006 was constructed by transforming RB247 with plasmid pMG28 at the amyE locus and knocking out the native rbgA gene by using a MLS cassette. The strains were checked for interruption of amyE by growth on starch plates. Strain RB1043 was constructed by transforming RB247 with plasmid pMG29 and knocking out the native rbgA gene by using chromosomal DNA from RB395. Independently, strain RB1044 was constructed in a manner identical to RB1043 to serve as a biological duplicate. All strains discussed in this study were confirmed for desired change using PCR to amplify the region of interest followed by sequencing.
Strains RB1043 and RB1044 were used for suppressor analysis. A single colony from each of these strains was inoculated per flask (25 colonies per strain, total of 50 colonies) and grown at 37°C for 16 hours. The undiluted culture from each flask as well as two serial dilutions (10-, and 100-fold) were plated on LB plates and incubated overnight at 37°C. The parental strains RB1043 and RB1044 were also plated along with RB1006 carrying wild-type RbgA to serve as controls. Isolated colonies from eight strains-RB1051, RB1055, RB1057, RB1059 (from RB1043) and RB1061, RB1063, RB1065, and RB1068 (from RB1044) that grew faster than parental strains were identified and characterized further.
Genomic DNA from RB247, RB1043, RB1051, RB1055, RB1057, RB1059, RB1061, RB1063, RB1065 and RB1068 was isolated using the Wizard genomic DNA isolation kit (Promega). The genomic DNA was analyzed on a 0.8% agarose gel to ensure that the quality was suitable for sequencing. Whole genome sequencing was performed on a Genome Analyzer II instrument equipped with a paired end module (Illumina) at the MSU Research Technology Support Facility. The sequencing reads obtained were quality tested using FASTQC and trimmed if needed. Next we aligned sequence reads from RB247 and RB1043 against the reference B. subtilis strain 168 genome using R2R software. We identified the insertion of pMG29 in RB1043 when compared with RB247 reads and the insertion of the MLS cassette in RB1043 at the native rbgA locus. The sequence of suppressor strains RB1051, RB1055, RB1057, RB1059, RB1063, RB1065 and RB1068 was then compared to RB1043 (the parental strain). In addition to the expected insertions found in RB1043 and each suppressor strain (corresponding to pMG28 at the amyE locus and the MLS cassette at the native rbgA locus) we identified only a single change in each suppressor strain in the rplF gene. The suppressor mutations that were identified utilizing the R2R platform were confirmed by PCR amplification of the rplF gene and sequencing the amplified product.
Homology model of L6 from B. subtilis was obtained by using Modeller 9.12 [48], utilizing the crystal structure of L6 (PDB code: 1RL6) from G. stearothermophilus as a template. Out of 20 models constructed, the model with lowest energy (molpdf) was chosen for further analysis. All structural analysis for figure 9 were carried out in Chimera using the 50S structure (PDB: 2AW4) [48], [49].
Strain RB1095 was constructed by transforming RB247 with pMG30 such that that the expression of the map gene (at the end of the operon that contains the rplF gene) was placed under the control of the IPTG inducible Phyperspank promoter. RB1102 was constructed by transforming suppressor strain RB1051 with chromosomal DNA from RB1095 and selecting cells on IPTG, chloramphenicol and MLS (lincomycin and erythromycin) such that the rbgA-F6A gene at amyE locus was selected and the mutated rplF gene operon was linked to the chloramphenicol marker. RB1103, RB1106 and RB1107 were constructed similarly by using RB1055, RB1065 and RB1068 as the parental strains, respectively. RB1117 was constructed by transforming RB247 with chromosomal DNA from RB1102 and selecting cells on IPTG and chloramphenicol, thus ensuring that this strain had a wild type rbgA gene at the native locus and the mutated rplF gene (operon was tagged with the chloramphenicol marker). RB1118, RB1121 and RB1122 were constructed similarly by utilizing chromosomal DNA from RB1103, RB1106 and RB1107 respectively. RB1123 was constructed by growing RB1117 on LB plates without chloramphenicol and IPTG such that the plasmid pMG30 was excised out leaving the mutated rplF gene in a wild type background. RB1125, RB1131and RB1133 were constructed similarly from RB1118, RB1121 and RB1122 respectively.
Ribosomal subunits were prepared by sucrose density centrifugation. 50S and 45S complexes were isolated from lysates of RB418 and RB301 cells, respectively as previously described [34]. RB1051, RB1055, RB1057, RB1063, RB1065 and RB1068 were grown to OD600 of 0.5 at 37°C in LB medium. Chloramphenicol (Sigma) was added to a final concentration of 100 µg ml−1 5 minutes prior to harvest. Cells were harvested by centrifugation at 5000 g for 10 min and resuspended in lysis buffer [10 mMTris-HCl (pH 7.5), 60 mMKCl, 10 mM MgCl2, 0.5% Tween 20, 1 mM DTT, 1× Complete EDTA-free protease inhibitors (Roche) and 10 U ml−1RNase-free DNase (Roche)]. Cells were lysed by three consecutive passes through a French press set at 1400 to 1600 psi and clarified by centrifugation at 16000×g for 20 minutes. Clarified cell lysates were loaded on top of 10–25% sucrose density gradients equilibrated in buffer B (10 mMTris-HCl, pH 7.6, 10 mM MgCl2, 50 mM NH4Cl) and centrifuged using a SureSpin 630 rotor (Sorvall) for 4.5 hours at 30,000 rpm. Gradients were then fractionated on a BioLogic LP chromatography system (BioRad) by monitoring UV absorbance at 254 nm. Fractions corresponding to ribosomal subunits of interest were pooled, concentrated using 100 kDa cutoff filters (Millipore) and stored in buffer A (10 mM Tris-HCl, pH 7.6, 10 mM MgCl2, 60 mM KCl and 1 mM DTT) at −80°C. For qMS, we followed the protocol as described above except that we used 18–43% sucrose gradient that was centrifuged at 21000 rpm for 14 hours.
Cell lysates from RB1043 and RB1055 were obtained as described above. Lysates were concentrated using 4 mL Amicon ultra-4 centrifugal filters with 10 kDa cutoff (Millipore). An equal volume of lysate was incubated at 37°C or 0°C for 1 hour, then loaded onto 18–43% sucrose gradient made in buffer C (10 mM Tris-HCl, pH 7.6, 20 mM MgCl2, 50 mM NH4Cl) followed by centrifugation at ∼82000 g for 14 hours at 4°C in SureSpin 630 rotor (Sorvall). Gradients were fractionated on BioLogic LP system (BioRad) monitoring absorbance at 254 nm.
The assay was performed as described [34]. Briefly, for measuring GTPase activity in the presence of ribosomal subunits/intermediates 100 nM RbgA protein was incubated with 100 nM 50S subunit or 45S subunit or 44S subunit and 200 µM GTP at 37°C for 30 minutes and for measuring intrinsic GTPase activity 2 µM RbgA protein was incubated with 200 µM GTP at 37°C for 30 minutes. We predetermined that under these conditions the values were in the linear range of the assay. The phosphate released was measured using the Malachite Green Phosphate Assay Kit (BioAssaySystems).
Experimental samples were prepared for quantitative mass spectrometry as described previously [22], with the following noteworthy modifications. First, each 14N-labeled sample to be analyzed (20 pmol) was mixed with a “double spike” internal reference standard mixture of 14N (10 pmol) and 15N (30 pmol) labeled 70S particles. The addition of the 14N standard still allowed for independent quantitation of the 15N isotope distribution for each peptide, but simultaneously ensured that each 15N peak bore a corresponding 14N peak pair irrespective of the concentration in the experimental sample. After precipitation, reduction, alkylation and tryptic digestion, peptides were analyzed on an Agilent G1969A ESI-TOF mass spectrometer. 14N:15N peak pairs were identified in the raw data, assigned to unique ribosomal peptides using a theoretical digest and the quantities of 14N and 15N species were calculated by fitting each isotope envelope using a Least Square Fourier Transform Convolution algorithm [50].
The contribution of the 14N material in the reference spike was eliminated from each experimental measurement by first analyzing the double spike mixture in isolation (Figure S2A; red). To account for variations in sample preparation and ionization efficiency between experiments, the 14N fitted amplitude of each peptide in each sample was normalized using the 15N internal standard amplitude (derived from a fixed concentration of 15N 70S ribosomes in each sample). Once normalized, the contribution of the double spike to the measured 14N amplitude could be eliminated from each experimental sample by simple subtraction of the 14N amplitude of the double spike sample measured in isolation (Figure S2C).
To test the efficacy of the approach and to establish a detection limit, we first measured a standard curve using 0, 2, 4, 8, 16, or 32 pmol of 14N-labeled 70S ribosomes mixed with the double spike (10 pmol 14N, 30 pmol 15N 70S). As our analysis pipeline depends on the identification of peak pairs, this double spike approach greatly increased the number of peptides detected for low-abundance proteins (J.H. Davis unpublished observation). Indeed, we consistently identified multiple peptides for each ribosomal protein, even at the low end of the standard curve (Figure S2D). By comparing the measured 14N/15N ratio to the known ratio added we found the approach to be linear over the range of this standard curve (Figure S2E). Moreover, this experiment established a quantitation limit of 2 pmol of the experimental sample, corresponding to occupancy = 0.1 when 20 pmol of the 44/45S or 70S particles were analyzed.
For each experimental sample, relative protein levels were calculated as the 14N corrected isotope distribution amplitude divided by the 15N isotope distribution amplitude. Isotope distributions and local chromatographic contour maps were examined and peptides with low signal-to-noise were excluded. Finally, to account for differences in the total amount of sample added, each relative protein level was normalized to that of the primary binding protein L20, which did not vary in occupancy between samples.
To improve our quantitation accuracy in more complex samples such as the cell lysates, we developed a selective reaction monitoring (SRM) protocol focused on ribosomal proteins. First, 14N-labeled tryptic peptides were generated from 70S particles as described above. These peptides were eluted from an analytical C18 nano-column across a 90 min concave 5–50% acetonitrile gradient at 300 nL/min. Mass spectrometry was performed on an AB/Sciex 5600 Triple-TOF instrument with an information dependent acquisition method utilizing 250 ms MS1 scans followed by 20 successive MS2 scans, each lasting 50 ms (cycle time of 1.3 sec, 4150 cycles/run). Each precursor was excluded from the MS2 target list for 12 seconds after observation. Using the fragmentation data and a theoretical digest of the B. subtilis proteome, precursor peptides were assigned using Mascot (Matrix Science). After generating a spectral library from the Mascot identifications, 8 SRM methods each targeting ∼110 ribosomal precursor ions each were generated in Skyline [51]. An equal mixture of 14N and 15N labeled peptides derived from 70S ribosomes were analyzed using these methods on the Triple-TOF and transitions with low signal-to-noise were eliminated. Using the measured retention times bracketed by a 7.5 min window, Skyline was used to generate a single scheduled MRM method targeting 310 precursors and ∼10 product ions per precursor. MS1 and MS2 scans lasted 200 ms and 30 ms, respectively. To measure ribosomal protein levels in cellular lysates, 0.5 OD*mL of each culture was mixed with 20 pmol 15N-labeled 70S ribosomes and prepared for qMS as described above. Each sample was then analyzed using the scheduled MRM method. Transition chromatograms were extracted from the raw data using Skyline and 14N/15N peak areas were calculated, filtered to exclude those with low signal-to-noise, and plotted using a series of Python scripts.
The Pearson correlation coefficient, r, was calculated between the whole cell protein level and immature particle datasets for strains RB1043, RB1055, RB1057, and RB1068 using the SciPy library.
|
10.1371/journal.pntd.0002014 | Perceptions on the Effectiveness of Treatment and the Timeline of Buruli Ulcer Influence Pre-Hospital Delay Reported by Healthy Individuals | Delay in seeking treatment at the hospital is a major challenge in current Buruli ulcer control; it is associated with severe sequelae and functional limitations. Choosing alternative treatment and psychological, social and practical factors appear to influence delay. Objectives were to determine potential predictors for pre-hospital delay with Leventhal's commonsense model of illness representations, and to explore whether the type of available dominant treatment modality influenced individuals' perceptions about BU, and therefore, influenced pre-hospital delay.
130 healthy individuals aged >18 years, living in BU-endemic areas in Benin without any history of BU were included in this cross-sectional study. Sixty four participants from areas where surgery was the dominant treatment and sixty six participants from areas where antibiotic treatment was the dominant treatment modality were recruited. Using a semi-structured interview we measured illness perceptions (IPQ-R), knowledge about BU, background variables and estimated pre-hospital delay.
The individual characteristics ‘effectiveness of treatment’ and ‘timeline acute-chronic’ showed the strongest association with pre-hospital delay. No differences were found between regions where surgery was the dominant treatment and regions where antibiotics were the dominant treatment modality.
Individual characteristics, not anticipated treatment modality appeared predictors of pre-hospital delay.
| Delay in seeking treatment for Buruli ulcer (BU) is a major challenge in current BU control. Research to date shows that several factors relate to delay, including a lack of knowledge about BU and its treatment, beliefs in a supernatural cause of the disease, feelings of fear and worry regarding the treatment, fear of surgery, direct and indirect costs, social isolation as a consequence of unbearable costs to the patients' family, a lack of confidence in the treatment, and stigma. This study focused upon the relationship between Illness perceptions and pre-hospital delay by using the Illness Perceptions Model of Moss-Morris et al in a sample of healthy community members living in 3 endemic areas for Buruli ulcer in Benin. We found that a chronic timeline perspective on Buruli ulcer and a higher perceived effectiveness of the treatment were independently associated with pre-hospital delay. The available dominant treatment modality in endemic areas (surgery or antibiotics) did not influence pre-hospital delay, a finding contrary to the previous suggestion that a fear of surgery would be related to delay in presenting to the hospital. This study has identified several individual characteristics which can form the basis of future interventions.
| Buruli ulcer (BU) is the third most important mycobacterial disease in humans today, with one of the highest prevalence rates reported in southern Benin (21.5/100,000 per year), although these numbers are probably underestimated due to the focal distribution of the disease [1]. The mode of transmission is largely unknown although skin injuries and insect bites have been suggested to play a role. Human to human transmission is considered negligible. The most important risk factor is living in tropical climates and wading in rivers or streams [2], [3]. The majority of patients in West-Africa are children [4], [5]. Buruli ulcer often starts as a firm, non-tender nodule, as plaques, or as edema. The skin breaks down into a painless ulcer with undermined edges, with the risk of complications such as osteomyelitis. After a varying period of time, a granulomatous healing response occurs, resulting in fibrosis, scarring, calcification, and contractures with residual disabilities [6]. Antimicrobial treatment (intramuscular streptomycin and oral rifampicin) has been standard since 2004; it is highly effective in the early stages of the disease, has low recurrence rates, and access to it is free of charge or at minimal costs [7]. Today, surgery is only considered for those who do not respond to antibiotics or to patients with extensive lesions. Surgery is especially important for the treatment of contractures and large skin defects where reconstructive surgery is needed [8]. Other treatment options perceived as conventional by people in Benin are self-medication and traditional (or herbal) treatment [9]–[11]. A recent study conducted in Ghana emphasizes the preference for herbal treatment, especially of patients with Buruli ulcer in an early pre-ulcerative stage [12]. Some of the major treatment centers located in the endemic south of Benin adhere to the WHO guidelines on antibiotics while others use surgery as the dominant treatment modality.
Irrespective of the treatment modalities (antimicrobial or surgical) offered, early presentation at a hospital or health care center is advantageous, because this minimizes trauma and pain, shortens admission to a treatment center, lowers the costs involved, and reduces the risk of amputation and functional limitations [6], [7], [10], [11], [13]. Despite these advantages, and the fact that people recognize BU and perceive the disease as threatening [10], [11], delay in visiting the health care center is the major challenge for national programs to fight BU. Pre-hospital delay is the time from onset of symptoms to arrival at the hospital to receive the recommended treatment [14]. Although literature is limited, previous explorative studies indicate several internal and external factors related to delay, including a lack of knowledge about BU and its treatment, beliefs in a supernatural cause of the disease, feelings of fear and worry regarding the treatment, fear of surgery, direct and indirect costs, social isolation as a consequence of unbearable costs to the patients' family, a lack of confidence in the treatment, and stigma [6], [10], [15], [16].
National BU control programs from several endemic countries initiated awareness raising campaigns followed by active case-finding strategies, which have shown to be effective in Benin [17], Côte d'Ivoire [18], Ghana [19], [20] and the Democratic Republic of the Congo [21]. Such programs seem to diminish delay, however, the question remains why some patients do while others do not delay in presenting to the hospital.
Based on previous psycho-social research on delay in Buruli ulcer,we hypothesize that despite economic, social factors and a lack of knowledge about BU and its treatment, cognitive and emotional factors play a role in delay, e.g. beliefs in a supernatural cause of the disease, feelings of fear and worry regarding the treatment, fear of surgery and a lack of confidence in the treatment. The studies that addressed the psycho-social factors in Buruli ulcer thus far, did not go into detail about the cognitive and emotional representations people have of Buruli ulcer.
Leventhal's commonsense model of illness (CSM) is a self-regulation model which describes how people respond to a health threat or illness. One assumption of the CSM is that people form cognitive and emotional representations to the illness. People will try to manage these emotions and cognitions by coping efforts. These coping efforts lead the person to take action e.g. visiting a doctor, taking medication. The coping strategy used, depends on their representation of the illness. This model is advantageous over other models such as the Access framework of health care utilization [22] and Anderson's model of health care utilization [23] in that it is unique to the individual and disease-specific.
The model describes a process in which the cognitive and emotional responses to an illness occur in parallel. The cognitive part is made up of perceptions on the identity, timeline (acute/chronic or cyclical), causes, consequences, coherence, control (treatment control/personal control) of the disease, the emotional part comprises emotional representations (figure 1). Perceptions on the identity of the disease contain ideas on symptoms attached to BU, for example, induration of the skin and underlying tissue. Timeline perceptions are ideas on the acuteness, chronicity or cyclical timeframe of the illness. Causal beliefs comprises views on the factors that caused Buruli ulcer, which are divided into biological, emotional, environmental and psychological causes. Consequences refer to the impact BU has on daily life, social and occupational functioning. The extent to which people feel they understand their disease is measured by the coherence subscale, and the control subscale is the extent to which people think they can influence the course of their disease and perceive treatment to be effective.
Basic knowledge about the disease and cultural factors (stigma) are suggested to influence illness perceptions and therefore, influence pre-hospital delay indirectly, while practical factors (time and distance to the hospital) are known to directly influence pre-hospital delay. We expect the regionally determined differences in dominant treatment modality to influence individuals' perceptions about BU, and therefore, influence pre-hospital delay indirectly (figure 1).
Illness perceptions have shown to be related to delay for different chronic illnesses [24], as well as infectious diseases such as tuberculosis [25] and they appear to exist in healthy individuals as well [26]. Illness perceptions are especially relevant in Buruli ulcer, because literature describes BU as being associated with cognitions such as worries about (economic) consequences of the treatment, ideas on causal mechanisms, major social (stigma) as well as general consequences on ones lives. Therefore, it seems important to ask questions about how an individual thinks about Buruli ulcer, how they would cope and the sense they make of it. Illness perceptions can quantitatively be measured by the widely used and validated Illness Perceptions Questionnaire [24], which can be adapted to a specific illness.
Our target group is healthy community members living in high risk areas for BU, in order to capture the perceptions and the future pre-hospital delay of those who are at risk for contracting the disease. The reason for choosing this group instead of actual cases, is that we are interested into the beliefs of potential patients irrespective of their treatment choice. Since actual cases already made their decision, namely; going to the hospital, we would only have captured the beliefs of this group while we are especially interested in the beliefs of those who would not present to the hospital, because this group is at risk for pre-hospital delay.
The first aim was to explore to what extent individual characteristics were related to future pre-hospital delay. The second aim was to explore whether the type of available dominant treatment modality influenced individuals' perceptions about BU, and therefore, influenced pre-hospital delay (figure 1).
Questionnaires were translated and back-translated from French to Fon, the local language in this area of Benin.
Analysis was performed using SPSS Statistics 17.0 and MLwiN 2.24. Means and standard deviations were calculated per region and for the total group of participants. Inter-relationships were indicated by Spearman's rho for ordinal variables and Chi-square statistics for the binary outcome; pre-hospital delay. A logistic regression analysis was performed with pre-hospital delay as outcome variable. All psychological, cultural and practical factors significantly related to pre-hospital delay, as well as knowledge on BU were entered. A second logistic regression analysis was performed with only the most important predictors for delay, starting with cognitive illness representations, followed by emotional illness representations, practical factors and knowledge on BU.
A multilevel logistic regression model with individuals (level 1) nested in villages (level 2) was used to examine the effect on pre-hospital delay. The factors shown to be related to pre-hospital delay in the logistic regression analysis were selected as individual-level characteristic. Dominant treatment modality was added as village characteristic. The final model selection was based on the estimate of the variables (p≈(approx.) .05) Continuous variables were centered around 1 in order to ease interpretation. Wald statistics were used to test the significance of the coefficients. High correlations between independent variables were examined on multicollinearity.
An overview of sample characteristics and health-related factors for the total group (n = 130) and per region (Lalo, Zagnanado, Pobè) are presented in table 1. No significant regional differences were found on background variables.
Forty five percent of the participants (n = 58) responded to the pictures in such a way that they were classified as ‘delayers’, while 55% (n = 72) were classified to be ‘non-delayers’. There were no regional differences (χ2 degrees of freedom (df) = 1) = 2.65, p = .26) or differences between the antibiotics and the surgery group (χ2 (df = 1) = 2.46 p = .12) on pre-hospital delay.
32% of the respondents believed inaccurately that BU was transmissible from person to person and almost half of our respondents (48.4%) believed inaccurately that there were direct costs involved in treatment. No regional or treatment differences were found on knowledge on BU (resp. χ2 (df = 1) = 1.95 p = .16 and χ2 (df = 2) = 1.95 p = .38).
Table 2 presents inter-correlations between explanatory variables, and their relationship with pre-hospital delay. There was a significant association between delay and eight subscales of the illness perception – i.e., timeline acute/chronic, illness coherence, effectiveness of treatment, effectiveness of alternative treatment, personal control, emotional representations and ‘chance’ as a cause of BU. Knowledge on BU was significantly associated with delay. Effect sizes for the mean difference on delay were large for the perceived effectiveness of treatment, personal control and timeline cyclical (table 2, bottom row).
The results of the logistic regression analysis (table 3) presents the model with the best fit, Model X2 (df = 5) = 47.64, p<.001; (Hosmer & Lemeshow test X2 (df = 8) = 7.11; p = 5.25). The most important predictors for the outcome ‘pre-hospital delay’ were personal control and timeline acute-chronic. If personal control increased by one unit (people perceive 1 unit more control over the illness; scale range 1–5) people are 2.10 times more likely to show pre-hospital delay. If timeline increased by one unit (people perceive the illness 1 unit more chronic in timeline; scale range 1–5), the probability to be delayed increased twice (Cox & Snell = .29, Nagelkerke R2 = .43.)
Table 4 presents the multilevel model with three dimensions of the illness perceptions as level 1 variables and dominant treatment as level 2 variable. The coefficient for dominant treatment was not significant, which means there is no difference between the region where surgery is the dominant treatment modality and regions where antibiotics are the dominant treatment on pre-hospital delay.
Two level 1 dimensions - the effectiveness of treatment and timeline acute-chronic, had significant coefficients. This reflects an increased probability of pre-hospital delay when the score of one of these dimensions increases, adjusted for the other level 1 and 2 variables. The adjustment means that the effect on pre-hospital delay is consistent, regardless of the other variables in the model.
We take an example of a person who lives in a region where surgery is the dominant treatment, perceives the treatment as not effective and thinks BU is acute in timeline (which is what most people believe). If this persons' score on the personal control dimension is low (e.g. There is nothing which I can do to control my symptoms -item 12), the probability of delay is 28%. When this persons' score on the personal control dimension is high (e.g. I have the power to influence my illness - item 16) the probability of delay is 92%. The proportion variance present at level 2 is 0.10 (0.374/0.374+3.29 [29]), which is twice as large as the proportion of unexplained variance in a model with no variables included (0.173/(0.173+3.29) = 0.05; data not presented). The explained level 2 variance by the full model presented in table 4 is 5%.
The multilevel model indicates that there were non-significant, small dominant-treatment modality differences (level 2) on pre-hospital delay. However, the illness perceptions dimensions: effectiveness of treatment, and timeline acute-chronic (level 1) were more important.
The results of this study suggest that psychological factors were predictors for pre-hospital delay, and not factors related to the dominant treatment available for BU (surgery or antibiotics). People who perceived BU as chronic in timeline, perceived treatment as effective or perceived higher personal control over the disease had a higher probability of delay. The dominant treatment available (surgery or antibiotics) in endemic regions in Benin did not show any effect on pre-hospital delay or on the individual characteristics related to pre-hospital delay.
Limited research is performed on the relationship between the type of treatment offered in a certain region and the amount of pre-hospital delay of individuals living in these regions. In a systematic review on factors related to treatment adherence in tuberculosis patients by Munro et al [30] reviewing studies from Asia, Africa, Europe and the USA, no influence of geographic location or type of treatment program was observed on treatment adherence. Instead, a number of structural, social, health service and personal factors correlated with treatment adherence. It is plausible that in our study, despite knowing someone with BU and living in highly endemic areas, respondents were not aware of the treatment modality provided in their region and that this was the reason that treatment modality was not related to pre-hospital delay.
In our study, illness perceptions were important for pre-hospital delay. People who believed the illness to have a chronic timeline, were more likely to delay. It is known that people who believe an illness to be chronic are more likely to attribute it to causes such as health habits, while people who believe an illness to be acute, are more likely to see a virus or bacterial agent as the cause. Our results are supported by a meta-analysis of Figureas and Alves on illness perceptions in healthy individuals [26]. They report a chronic timeline perception to account for a significant proportion of variance in attitudes towards preventive health behavior, irrespective of the experience of the illness.
Participants in our study who believed more in the effectiveness of treatment were more likely to delay, a finding which is in line with a recent study of Peeters et.al (2012) [31], who describe the length and complexity of patients treatment choices as result of their determined search for effective treatment. Some patients in their study experienced financial and professional loss and social isolation due to their search for effective treatment. They conclude that the overall difficulty of finding successful treatment is an important factor for late arrival at treatment centers. A similar explanation might be at stake here. An alternative explanation is that due to the high inter-correlation between the dimensions ‘Effectiveness of treatment’ and ‘effectiveness of alternative treatment’, people interpreted the word ‘treatment’ as alternative treatment and the individuals who found this effective were more likely to delay. This is in line with previous findings of Stienstra et al (2002), and Brienza et al (2002) [6], [32].
A stronger believe in the controllability of the illness by one's own behavior was related to more pre-hospital delay. A review of Brienza et al [32], claims that personal control is related to adaptive outcomes such as changing health behavior. An explanation for the relationship found in our study could be that individuals who perceive more personal control are more likely to take a situation into their own hands, decide to seek help in alternative treatment or engage in self-medication and therefore, delay in presenting in the hospital.
A strength of the CSM is that it is a dynamic model which is unique to the person. The model is applicable to specific illnesses, as opposed to more generalised health behaviour models such as the Access framework of health care utilization [22] and Anderson's model of health care utilization [23]. Furthermore, it has proven to predict a number of health behaviors such as adherence to treatment, treatment attendance, delay and recovery from illness. A strength of using the IPQ-R in measuring illness perceptions is that items relevant to specific illnesses can be added while maintaining psychometric validity.
Limitation of the use of the CSM in our study is that it describes illness perceptions and coping as being a dynamic process, however, when using the IPQ-R in a cross-sectional study, this process view is not taken into account. Although our study is useful in identifying factors that may impact on delay, they do not test for causal relationships. Another limitation is that knowledge about the psychometric properties of the IPQ-R in African populations is limited, since the instrument was developed and used mostly in European populations. To our knowledge, there is one study reporting on the IPQ-R in a from Africa origin diabetes population [33]. We found similar patterns of inter-correlations between subscales of the IPQ –R with this study. Timeline cyclical was positively related to consequences, illness coherence and personal control. Illness coherence correlated positively with personal control and negatively with emotional representations. Timeline acute-chronic correlated positively with consequences and emotional representations and negatively with personal control and illness coherence. The positive relationships between timeline and consequences, the negative relationship between timeline and cure/control (personal control and effectiveness of treatment) and cure/control and consequences were also similar to previous psychometric studies [34]. There were also discrepancies which could be due to differences across varying disease types, implying that outcomes are specific to specific diseases.
The relationships established in this cross-sectional study were based upon an estimated measure of pre-hospital delay of healthy people. Forty five percent of our participants expected to show up in a late stage (picture 3, 4, 5 or 6 in figure 3), while 55% expected to show up at the hospital in an early phase (picture 1 and 2 in figure 3). When considering the accuracy of this estimation, some attention should be paid to the possibility of a self-serving bias [35]. This concept assumes that people tend to overestimate their own behavior, while accurately predicting others' behavior [36]. Therefore, the proportion of people expecting to show pre-hospital delay in our study might be an underestimation of the real proportion, strengthening the relevance of factors we found to be related to pre-hospital delay.
A strength of this study is that above stated relationships were established by quantitative, individually derived data with standardized instruments, adding information to the results of previous studies using different approaches. Furthermore, the geographical distribution where participants resided, and the high (99%) response rate, contributed to the representativeness of the sample. Finally, an approach in which native-looking and native-speaking interviewers performed the interviews contributed to the quality of the data.
A limitation was the relatively small sample size (n = 130) which might have contributed to the lack of a statistically significant impact of the dominant treatment modality differences on pre-hospital delay, although the required sample size of 56 participants in each group should have given sufficient power to detect a meaningful difference on delay. Furthermore, the cross-sectional design of the study restricts the results to associations, and although predictors for delay are suggested, these are potential predictors for which no causal interference can be made. Another limitation is our choice for not taking previous medical history, which is often a contributing factor for delay [37] into account. Another limitation is the relatively low number of level 2 entities (17 villages). Scherbaum et al recommend a minimum of 30 level 2 units in for performing a multilevel analysis [38].
Our findings add to literature the importance of individual characteristics in explaining pre-hospital delay, above and beyond practical (means of transportation and the self-perceived time from home to hospital), socio-demographic and economic factors, and knowledge on the disease. The measures (IPQ-R) in this study are new for this population, and further research is needed to explain some of the counterintuitive findings such as the relationship between personal control and pre-hospital delay.
Further research is also needed in order to explain whether the illness perceptions of healthy individuals predict delay using longitudinal designs. This cross-section study cannot explain why people report late, because this would need an approach in which a solid qualitative method is used that departs from these predictors and systematically seeks for reasons.
We suggest studying this with a multilevel design which incorporates sufficient level 2 entities and by using neighborhoods instead of villages as the aggregated level. We expect that neighborhoods reflect a more appropriate macro level, because differences in and between neighborhood with respect to beliefs on the effectiveness of treatment are more relevant than differences on a village level. Such studies may help in identifying factors to focus upon in community programs aiming at reducing pre-hospital delay.
|
10.1371/journal.pntd.0003589 | Uptake of Rabies Control Measures by Dog Owners in Flores Island, Indonesia | Rabies has been a serious public health threat in Flores Island, Indonesia since it was introduced in 1997. To control the disease, annual dog vaccination campaigns have been implemented to vaccinate all dogs free of charge. Nevertheless, the uptake rate of the vaccination campaigns has been low. The objective of this paper is to identify risk factors associated with the uptake of rabies control measures by individual dog owners in Flores Island.
A total of 450 dog owners from 44 randomly selected villages in the Sikka and Manggarai regencies were interviewed regarding their socio-demographic factors, knowledge of rabies, and their uptake of rabies control measures. The majority of dog owners surveyed (>90%) knew that rabies is a fatal disease and that it can be prevented. Moreover, 68% of the dog owners had a high level of knowledge about available rabies control measures. Fifty-two percent of the dog owners had had at least one of their dogs vaccinated during the 2012 vaccination campaign. Vaccination uptake was significantly higher for dog owners who resided in Sikka, kept female dogs for breeding, had an income of more than one million Rupiah, and had easy access to their village. The most important reasons not to join the vaccination campaign were lack of information about the vaccination campaign schedule (40%) and difficulty to catch the dog during the vaccination campaign (37%).
Dog owners in Flores Island had a high level of knowledge of rabies and its control, but this was not associated with uptake of the 2012 vaccination campaign. Geographical accessibility was one of the important factors influencing the vaccination uptake among dog owners. Targeted distribution of information on vaccination schedules and methods to catch and restrain dogs in those villages with poor accessibility may increase vaccination uptake in the future.
| Rabies has been a serious public health threat in Flores Island since its introduction in 1997, with significant economic consequences for the government due to the cost of providing vaccines for humans and dogs. Vaccination of dogs against rabies offers a safe and effective means to prevent rabies in humans in the long term. The local government of Flores Island has implemented dog vaccination campaigns throughout the island, free of charge for all dog owners. However, the uptake rate of vaccinating dogs is too low to control the disease. Identifying factors associated with the dog owners’ decisions to vaccinate their dogs provides information to policy makers to increase vaccination uptake in the future. This study evaluated the impact of rabies knowledge and socio-demographic factors of dog owners on the uptake of the 2012 dog vaccination campaign. Overall, the uptake rate of the 2012 vaccination campaign was low. Levels of knowledge about rabies and its control were high, but not associated with uptake of vaccination. Geographical accessibility was significantly associated with vaccination uptake. Targeted distribution of information in those villages with poor accessibility may increase vaccination uptake in the future. Information should cover vaccination schedules and methods to restrain dogs.
| Rabies still poses a significant health problem in many countries of the world, despite it being a vaccine-preventable disease in dogs and humans [1]. Approximately 55,000 people around the world die each year due to rabies, with 45% of these cases occurring in the South East Asian region [2]. Within this region, Indonesia has the fourth largest number of human rabies cases after India, Bangladesh and Myanmar, with 150–300 cases reported per year [2].
The first occurrence of rabies in Indonesia was reported in 1889 [3]. Since its introduction, rabies has posed a serious public health threat with significant economic consequences to society [4]. The national strategic plan of Indonesia emphasizes the control of rabies as a policy priority, aiming for eradication by the year 2020 [5].
Flores Island is located in the eastern part of Indonesia and covers an area of 15,624 km2 [6]. The island is divided into eight regencies, with a human population of more than 1.8 million [7] and a dog population greater than 0.2 million [4]. Many of the rural areas on the island are only accessible by foot or with high-clearance vehicles, motor bikes, or horses [8]. The main socio-economic activity on the island is agriculture (production of coconut, corn, groundnut, cocoa, coffee, potato, and paddy), in which dogs are used to guard the crops [9]. Most dogs are owned and roam freely day and night. Although Indonesia is predominantly Muslim (practicing the Islamic principles in which it is prohibited to eat dog meat or to keep dogs inside the house), the majority of people in Flores are Catholic. Dogs have a high cultural and economic value in Flores Island, as they provide a source of animal protein in addition to their guarding capacities. Dog meat is a popular menu item in certain traditional ceremonies of the island [9].
On Flores Island, the first cases of dog rabies were officially confirmed in April 1998 in the regency of East Flores [10]. The introduction of the disease was traced back to three suspected rabid dogs that were brought from Buton Island by a fisherman in September 1997 [10]. Despite the initial control measures applied, which entailed the culling of the entire dog population in and around the affected villages (1998–1999), rabies spread to other regencies of the island [11]. In response, the Flores Island government have implemented a comprehensive control campaign since 2000. This control campaign is based on a combination of control measures, including mass vaccination of dogs, culling of roaming dogs, placing imported dogs in quarantine, and giving pre- and post-exposure treatment to humans. Complementary control measures include investigation of dog bites, diagnostic testing of suspected rabid dogs, and tracing of human contacts with rabid dogs. However, this campaign has not yet been successful in eliminating rabies from Flores Island.
Thousands of people bitten by dogs are looking for post-exposure treatment each year, resulting in a large economic cost for both government and local communities. The annual cost of rabies control efforts in Flores Island has been estimated to exceed US$ 1.0 million [4]. The impact on public health is difficult to measure. Until 2012, 96 human cases of rabies were officially registered by the Public Health Department, with the highest number of cases in Manggarai regency (27 cases), followed by Sikka (22 cases), Ngada (16 cases), West Manggarai (11 cases), East Manggarai (8 cases), East Flores (6 cases), Nagakeo (4 cases), and Ende (2 cases) [12]. However, these numbers do not reflect the real human rabies burden in Flores, as the data only capture the number of rabies patients who visited hospitals or public health centers during the period that rabies was clinically manifest. The number of human cases reported by the Husbandry Department of East Nusa Tenggara Province was more than two times higher (228 cases) [13] than the 96 cases officially recorded by the Public Health Department [12].
During the last thirteen years, the regencies on Flores Island have implemented annual dog vaccination campaigns using Rabivet Supra 92 [4]. Although vaccination is compulsory for all dogs (Manggarai Regency Law, number 6, year 2003), it is difficult to enforce due to the absence of a proper registration system and the lack of resources to catch and restrain dogs. Vaccination is therefore only feasible with the support of the dog owner, who presents and restrains the dogs for vaccination. To increase vaccination coverage, regencies have offered dog owners the vaccination of their dogs for free. Moreover, the vaccine has been delivered using a ‘house-to-house’ approach, undertaken by the local authority [4] to directly persuade dog owners to vaccinate their dogs. The ‘house to house’ vaccination approach is a method in which the vaccination teams visit the dog owners at their own homes. As the vaccination team are not equipped to handle roaming dogs, dog owners need to catch and restrain their dogs themselves. Because the dogs in Flores Island are not used to being restrained, it is expected that ‘house-to-house’ vaccination campaigns result in a higher vaccination coverage than central point vaccination campaigns [14]. Moreover, ‘house-to-house’ campaigns put more social pressure on dog owners to vaccinate than central point campaigns. Nevertheless, the uptake rate of the dog vaccination measure has been low, with an average vaccination coverage of around 53% of the registered dogs during the 2000–2011 vaccination campaigns [4]. This value is lower than the 70% coverage of the complete dog population, which is recommended to maintain the control of rabies between annual vaccination campaigns [1].
Although there are publications describing the uptake of dog vaccination campaigns in developing countries [14,15,16,17,18,19,20], none of these studies have focused on the situation of rabies in Flores Island, nor evaluated the impact on the uptake of vaccination of the socio-demographic characteristics of dog owners and their knowledge of rabies. An understanding of this impact is essential to support policy decisions about rabies control in the future. The objective of this paper is to identify risk factors associated with the uptake of rabies control measures by dog owners in Flores Island, Indonesia. This is achieved by undertaking an extensive survey among dog owners in the regencies of Sikka and Manggarai. Risk factors concern socio-demographic factors and the level of rabies knowledge of dog owners. Special emphasis is given to risk factors associated with the uptake of the vaccination campaign in 2012.
An extensive survey was conducted among dog owners in the regencies of Sikka and Manggarai during January and February 2013. The regencies were selected because of the high prevalence of human rabies and the control legislation in place. Sikka relies on the national rabies control campaign, whereas Manggarai has a local control legislation in place [4]. Based on this local legislation, Manggarai has been applying additional control measures (such as culling) alongside the nationally recommended vaccination control campaign. The regencies have similar sized populations, 300,301 inhabitants in Sikka and 292,037 inhabitants in Manggarai (census data of 2010) [7]. There are no officially registered data available on the number of villages and the number of households owning dogs, nor on the size of the dog population. During the ‘house-to-house’ rabies vaccination campaign in 2012, 351 villages in Sikka and 162 villages in Manggarai were involved. All households in these villages were visited by the local authorities to vaccinate the dogs for free. Given the number of dogs registered during this vaccination campaign, the number of dogs is estimated at 37,000 in Sikka and 6,675 in Manggarai. The difference in the number of dogs per regency is a result of the culling measures that were implemented in Manggarai.
The minimum sample size required to estimate the proportion of dog owners vaccinating their dogs was based on the conservative assumption that 50% of dog owners vaccinated at least one of their dogs (pexp), with a 5% error in estimate (d) and a 95% confidence interval. Given the standard power calculation:
n=1.962×pexp×(1−pexp)d2
(1)
The required sample was a minimum of 385 dog owners. The sample size was increased to 450 dog owners to account for incomplete interviews, with 300 dog owners sampled in Sikka and 150 in Manggarai. The relative size of the samples in Sikka and Manggarai reflected the difference in the number of villages involved in the 2012 vaccination campaign, which was 351 villages in Sikka and 162 villages in Manggarai.
Dog owners were selected from the villages included in the 2012 vaccination campaign. A random order of villages for the survey was obtained for each regency by randomly ranking the villages involved in the 2012 vaccination campaign. Subsequently, villages were visited in the order of this ranking until the predefined sample sizes (300 dog owners in Sikka and 150 dog owners in Manggarai) were reached. A total of 44 villages were visited, with 27 villages in Sikka and 17 villages in Manggarai. The following process was carried out in each of the 44 villages in the survey. Firstly, the village leader was approached to inform him about the study and to seek permission to carry out the study in his village. Subsequently, 7 to 15 dog owners (respondents) aged 18 years or older were selected per village. The number of dog owners to be interviewed was predefined for each village and proportional to the number of dog owners involved in the rabies vaccination campaign of 2012. Due to the lack of registration data on households, the first respondent in each village was chosen by chance by spinning a pen at the center of the village [20]. The direction of the pen tip determined the first household/respondent to be interviewed. In case there were no dogs or adult persons present in the house, the next household was selected [21]. Subsequent respondents were selected from the closest neighboring households that owned dogs.
The questionnaire interviews were conducted by two survey teams (one team per regency) assisted by local people with knowledge of the local languages (Sikka, Lio, and Manggarai) as well as Bahasa Indonesia. Prior to each individual interview, the purpose of the survey was explained to the respondents. A verbal informed consent (permission to carry out the interview) was obtained from the dog owners before the interview was conducted. The interviews generally took place between 8 a.m. and 6 p.m., from Mondays to Saturdays. When there were not enough participants available in a village due to the absence of dog owners, the interviews were subsequently administrated in the early morning or evening of the next day. Daily evening briefings among the survey team members ensured interview consistency. Obtained data were entered in Data Editor of SPSS software version 19.
The questionnaire was designed after an extensive literature review of previous survey studies, which focused on either the level of rabies knowledge or the uptake of rabies control measures in dogs and humans [18,20,22,23].
The questionaire contained open and closed questions, which were divided into four sections: (1) socio-demographics of dog owners, (2) knowledge of rabies and its control measures (3) uptake of rabies control measures, and (4) reasons for joining or not joining the rabies vaccination campaign of 2012.
1) Socio-demographics. Questions on the socio-demographics of dog owners covered personal characteristics, characteristics of the household, characteristics of the dogs and reasons for keeping them, and one characteristic of the village. Personal characteristics included gender, age, education level, occupation, income, and religion. Characteristics of the household included the size of the household and presence of children in the household. Characteristics of the dogs included whether female for breeding or male dogs were kept, reasons for keeping dogs, and the economic value of dogs. The characteristic of the village concerned the accessibility of the village. This was assessed for each village according to the type of road infrastructure between the village and the main road. Three categories were used to stratify accessibility. Villages located along the provincial road connecting West to East Flores Island were categorized as villages with ‘good accessibility’, as these villages have easy access to frequent public transportation. Villages located more than 3 km from the main road, which have less frequent access to public transportation, were categorized as villages with ‘average accessibility’. Villages that can only be reached by foot or motorcycle, which have no public transportation facility, were categorized as villages with ‘poor accessibility’.
2) Knowledge of the risk, prevention in humans, and control of rabies. The knowledge of the dog owners about the risk of rabies to humans was assessed using the following two ‘yes’ or ‘no’ questions, as modified from Tenzin et al. [22]: (1) “do you know that rabies is a fatal disease in humans?” and (2) “do you know that rabies in humans can be prevented?”. To assess the level of knowledge of rabies control measures, a subsequent question was posed to those dog owners who responded positively to the second question: “which measures are known to you in order to prevent rabies in humans?”. The intention of this question was to evaluate the level of knowledge about the range of control measures that could prevent rabies in humans. This knowledge does not necessarily reflect an understanding of the efficacy of the control measures. The respondents’ answers to the question were classified as either correct or incorrect based on scientific evidence [24,25]. Answers were considered scientifically correct if the control measures mentioned by the respondents were in line with those recommended by the WHO [1,26] and OIE [27]. These recommendations consist of: (1) vaccination injections before exposure (pre-exposure treatment), (2) cleaning wound after being bitten, (3) injection of human rabies vaccines and/or immunoglobulin after exposure (post-exposure treatment), (4) vaccination of dogs, (5) dog movement restrictions, and (6) leashing of dogs. Each corresponding answer was given a score of 1. Answers that were not based on scientific evidence (e.g., prayer and traditional medicine) were given a score of 0. In addition, as culling of dogs is a control measure within Manggarai regency law (number 6, year 2003), an additional score of 1 was given to those respondents who reported culling as a control measure. Within Mangarai regency law, this culling refers to the culling of dogs that are aggressive and tend to bite and culling of roaming dogs in newly infected villages and public areas regardless of their health status. The total score per respondent (range of 0–7) was subsequently categorized into a binary variable, by defining one category to indicate total scores lower than the median score of all answers and another category to indicate total scores equal to or higher than the overall median score [22].
3) Uptake of rabies control measures. To obtain information about the uptake of rabies control measures, dog owners were asked about the measures they had adopted during the period 1999–2012. With respect to the control measures in dogs, dog owners were specifically questioned about the uptake of vaccination and culling. In addition, the uptake of a complementary rabies control measure was also specified, namely castration of male dogs as part of dog-population management. We did not consider the option to sterilize female dogs, assuming that very few female dogs will be sterilized due to the lack of animal health facilities and the costs involved.
Concerning the control measures in humans, dog owners who indicated that they had experienced a bite incident in their family were asked about their uptake of post-exposure treatments (wound cleaning, rabies immunoglobulin injection, and series of rabies vaccine injections). The uptake items were recorded into dichotomous variables for subsequent data analysis (1 = ‘Uptake’ and 0 = ‘No uptake’).
4) Reasons for (not) joining the rabies vaccination campaign of 2012. In the final item, motives for joining or not joining the 2012 dog vaccination campaign were explored with open questions. The question about the motives for joining was posed to the dog owners who had vaccinated their dogs, allowing them to mention multiple reasons. The question about motives for not joining was posed to those who had not vaccinated their dogs and allowed them to indicate only the main reason.
The questionnaire was developed in English and translated to Bahasa Indonesia. It was pre-tested by a focus group, consisting of 4 veterinarians, 4 veterinarian assistants/vaccinators, a husbandry officer of the Animal Health and Husbandry Department of Sikka, and by five pilot interviews [22] with dog owners from a village in Sikka. The final questionnaire was revised based on the pre-test and pilot interviews to improve clarity and interpretation [22]. An interview took approximately 45 minutes to complete the questionnaire.
Descriptive statistics were used to summarize the dog owners’ responses for each of the four survey items. Differences in the proportions of dog owners with knowledge of rabies (yes/no), knowledge of control measures (high/low), and uptake of the 2012 vaccination campaign (yes/no) were tested using the Chi-square test. The associations of the socio-demographic factors with the levels of knowledge (i.e. about the risk of the disease, prevention in humans, and control measures) and the uptake of the 2012 vaccination campaign were assessed using univariable logistic regression analyses. The effect of the knowledge level of dog owners on the uptake of the 2012 dog vaccination campaign was also explored using an univariable logistic regression analysis. Four multivariable logistic regression analyses were conducted to determine the independent contribution of each of these variables to the outcome (e.g. uptake of the 2012 dog vaccination campaign) after adjusting for other variables [28,29]. All independent variables, which had p-values of less than 0.25 in the univariable analyses were subsequently included in the initial models for the multivariable analyses [30].[22,31]. Prior to the multivariable analyses, Spearman’s rank correlation coefficient (ρ) was calculated to check for multicollinearity between the independent variables selected from the univariable analyses. Multicollinearity was considered to be present at ρ>0.7. The final multivariable logistic models were derived by backward stepwise elimination of variables with a p-value greater than 0.05. The Hosmer-Lemeshow goodness-of-fit test was performed to determine the fit of the final models with the data [30].
The multivariable models are represented by the logit formula:
ln(v1−v)=β0+β1x1+β2x2+.....+βpxp,
(2)
where ln(v1−v) is the log of the odds of the outcomes (i.e., having knowledge about rabies and the risk it poses for humans or not, having a high or low level of knowledge about rabies control measures, and having participated in the 2012 vaccination campaign or not, represented by (v) and (1 − v), respectively), β0 is the estimated intercept, and β(1) ,.., β(p) represent the regression coefficients of each independent variable included in the model. Exponentiation of these regression coefficients (eβ(1),...,β(p)) gives the odds ratios (OR) for each independent variable. SPSS version 19 was used for the analysis of all the data.
A total of 463 households were visited. Of these households, 5 dog owners (in Sikka) refused to be interviewed and 8 dog owners (3 in Sikka and 5 in Manggarai) were not at home during the time of the interviews. The socio-demographic characteristics of the 450 respondents are shown in Table 1. The majority of the respondents was male (67%), aged between 18–45 years (56%), and had children in the household (84%). Most respondents were farmers (79%) and of catholic religion (99%). Of the 450 respondents, almost 50% had attended or graduated from elementary school. The median numbers of humans and dogs per household were 5.0 humans (mean 5.3; range: 1–11) and 2.0 dogs (mean 2.2; range: 1–12). The majority of dog owners (68%) indicated that they kept dogs to guard their house and property, and to chase away wild animals that destroy their crops.
The majority of the dog owners surveyed in Sikka and Manggarai regencies agreed with the statement that “rabies is a fatal disease in humans” (92%). Agreement with this statement was significantly different (p<0.05) for regency, income, presence of children in the household, primary function of dogs, and economic value of dogs (Table 1).
Four hundred and three dog owners (90%) agreed with the statement “rabies in humans can be prevented”. Agreement was significantly different (p<0.05) for regency, education level of dog owners, having a family member previously bitten by dogs, having female dogs for breeding, primary function of dogs, and economic value of dogs (Table 1). The preventive measures that were most frequently known by the dog owners, who agreed with the statement that rabies in humans can be prevented (n = 403), were vaccination and/or immunoglobulin injection (81%) and wound cleaning (79%) (Fig. 1). Other indicated control measures included dog vaccination (77%), leashing of dogs (36%), traditional treatment (31%), and prayer (15%). The total number of scientifically correct measures indicated per dog owner varied between 0 and 6, with a median of 3. The majority of the dog owners (68%) mentioned 3 or more measures, indicating a relatively high level of knowledge about rabies control measures (Table 1). Only four respondents (1%) indicated traditional treatment and prayer as the only means to prevent rabies in humans. The level of knowledge of control measures differed significantly (p<0.05) by income, religion, number of people per household, and primary function of dogs. The level of knowledge was not significantly different between regencies (p>0.05).
Factors related to knowledge of the risk, prevention, and control of rabies. The factors regency, having male dogs, and economic value of dogs were significantly related with knowledge of the risk of rabies in humans (Table 2). Dog owners living in Sikka were more aware about the risk of rabies in humans (OR = 5.55; 95% CI = 2.33–13.18) compared to dog owners living in Manggarai. Dog owners having male dogs had lower odds of having knowledge of the risk of rabies in humans (OR = 0.41; 95% CI = 0.18–0.96) compared to their counterparts. Dog owners who kept dogs with an average economic value between Rp250,000 and Rp500,000 per dog were more likely (OR = 2.74; 95% CI = 1.14–6.59) to have knowledge of the risk of rabies in humans compared to a value of less than Rp250,000 per dog. The final model had a good fit with the data (Hosmer-Lemeshow goodness-of-fit test p-value was 0.76) and no multicollinearity was found between the independent variables (highest Spearman’s rank correlation coefficient (ρ) was 0.36).
Knowledge of the prevention of rabies in humans was significantly associated with regency, economic value of dogs, and education level (Table 3). The odds of having knowledge of rabies prevention were higher among dog owners living in Sikka (OR = 3.44; 95% CI = 1.68–7.05), having a high educational level (OR = 4.64; 95% CI = 1.50–14.33), and having dogs with an average economic value between Rp250,000 and Rp500,000 per dog (OR = 2.94; 95% CI = 1.40–6.16) compared to a value of ≤Rp250,000 per dog. The Hosmer-Lemeshow goodness-of-fit test p-value for this model was 0.48, which indicates an adequate fit of the model to the data. There was no multicollinearity between the independent variables (the highest Spearman’s rank correlation coefficient (ρ) was 0.53).
The results of the logistic multivariable regression analysis (Table 4) on the level of knowledge of rabies control measures showed a significant association with the following factors: primary function of dogs, the level of dog owners’ income, and the geographical accessibility of the village. The odds of having a high level of knowledge of rabies control measures was higher in the following situations: for dog owners who lived in villages with good accessibility (OR = 2.14; 95% CI = 1.07–4.27) compared to poor accessibility, for dog owners who kept dogs as a source of income (economy) (OR = 3.18; 95% CI = 1.20–8.44) or as a guard of the house or property (OR = 3.44; 95% CI = 1.39–8.51) compared to the function of traditional ceremony, and for dog owners who had an income of more than Rp1,000,000 (OR = 3.02; 95% CI = 1.36–6.71) compared to an income of Rp500,000–1,000,000. The model fitted the data well (the Hosmer-Lemeshow goodness-of-fit test p-value was 0.80) and no multicollinearity was found between independent variables (the highest ρ was 0.37).
Control measures in dogs. Respondents’ uptake of rabies control measures during the last fourteen years (1999–2012) are shown in Table 5. Fifty-six percent of respondents reported that at least one of their dogs had been vaccinated during the last fourteen years, the majority of which (92%) had had their dogs vaccinated during the 2012 campaign. Regarding culling as a control measure, 33% of the dog owners reported that at least one of their dogs had been culled during the last fourteen years. The majority of the culling was carried out in 1999 (45%) and 2010 (25%). Only 1% of the respondents reported that at least one of their dogs had been culled in 2012. Most often, these dogs were culled after having bitten someone or showing unusual behavior. In total, 12% of the dog owners had castrated at least one of their dogs. The main purpose given by the dog owners for castration was to keep the dogs close to home (prevent roaming away) in their function as guard of the house and property or as hunter to chase away wildlife. Castration of male dogs was carried out by the dog owners themselves or by their family.
The uptake of vaccination and culling differed among regencies. The proportion of dog owners who had vaccinated their dogs was significantly (p<0.001) higher in Sikka (65%) than in Manggarai (39%) (Table 5). In contrast, the proportion of dog owners that had culled their dogs was significantly higher in Manggarai (45%) than in Sikka (27%) (p<0.001).
Control measures in humans. Of the 450 dog owners interviewed, 89 (20%) reported that at least one of their family members had been bitten by a suspected rabid dog during the last fourteen years. Of these 89 bite cases, 75 (84%) cleaned the wound, and 50 (56%) received vaccination. The level of uptake of these measures did not differ between regencies (p>0.05). Approximately 87% of the reported bite cases occurred during the period 2009–2012, of which 34% in 2012 alone.
Uptake of the 2012 vaccination campaign. During the 2012 vaccination campaign, 52% (234/450) of the dog owners had vaccinated at least one of their dogs (Table 1). This uptake proportion was significantly higher in Sikka (63%; 189/300) than in Manggarai (30%; 45/150) (p<0.001). The proportion of vaccination uptake was also significantly higher for owners of female dogs (59%; 87/147) than male dogs (49%; 147/303) (p<0.05). The vaccination uptake was significantly associated with the knowledge of the dog owners about rabies (p<0.001) and its control measures (p<0.05) (Table 6). The proportion of dog owners who had vaccinated dogs was higher for those who considered rabies a fatal disease (54%; 225/415) than for those who did not (26%; 9/35). Similarly, dog owners with a high level of knowledge of rabies control measures (59%; 161/272) tended to vaccinate their dogs compared to their counterparts (46%; 60/131). We found no significant association of vaccination uptake with the age and education level of the dog owner, presence of children in the household, or having male dogs (Table 1).
Multivariable model for the 2012 vaccination uptake. Of the 13 independent variables that had an association (p-value less than 0.25) with the uptake of the 2012 vaccination campaign in the univariable analyses (Table 1 and Table 6), only five variables were retained in the final multivariable model. Regency, having female dogs for breeding, economic value of dogs, income of dog owners, and accessibility of the village were significantly associated with the uptake of vaccination (Table 7). Dog owners from Sikka were more likely to vaccinate their dogs (OR = 4.07; 95% CI = 2.30–7.20) than those from Manggarai. The dog owners who held female dogs for breeding had significantly higher odds to vaccinate their dogs (OR = 2.07; 95%CI = 1.31–3.27) compared with those who did not. Similarly, the dog owners who owned dogs with an economic value ranging between Rp250,000–500,000 tended to vaccinate their dogs (OR = 2.38; 95%CI = 1.36–4.17) compared with owners who valued their dogs at less than or equal to Rp250,000. Moreover, the uptake of vaccination was higher if the dog owners had a monthly income of more than 1 million Rupiah (OR = 2.39; 95%CI = 1.10–5.20) and lived in a village with a good accessibility (OR = 3.84; 95%CI = 1.92–7.67) compared with those having a yearly income less than Rp500,000 and who lived in a village with poor accessibility. The Hosmer-Lemeshow goodness-of-fit test p-value for this model was 0.85, which indicates the model fitted the data well. There was no multicollinearity between independent variables (the highest Spearman’s rank correlation coefficient (ρ) was 0.49).
Motivation to adopt the vaccination control campaign. Reasons for having their dogs vaccinated against rabies in 2012 were given by the 234 dog owners who indicated that at least one of their dogs was vaccinated during this vaccination campaign. The most common reasons for the dog owners to vaccinate their dogs were to protect their own health and that of their family (97%) and to protect the children in their community (77%) (Fig. 2). For those dog owners who had not vaccinated their dogs (216 of 450 dog owners), the most important reasons for not joining were the lack of information about the schedule of the vaccination campaign (40%) and the difficulty to catch their dogs during the vaccination campaign (37%) (Fig. 3). Other reasons, of minor importance, included the lack of belief in the vaccine efficacy (13%), and the young age of the dog at the time of the vaccination campaign (6%).
The knowledge of dog owners in Flores Island about the risk of rabies to human health and about the possibilities to prevent the disease was generally high. This positive result might have been overestimated or biased due to the structure of the questions posed (“do you know that rabies is a fatal disease” and “do you know that rabies in humans can be prevented”). However, given the dog owners’ prompt responses on the subsequent open questions, e.g., “which control measures are known to you in order to prevent rabies in humans”, to which all dog owners provided a response, we expect that the structure of the questions did not influence the result substantially. This high knowledge level is comparable with findings of other studies conducted in South East Asian countries [22,32,33]. The high level of knowledge about the risk of rabies and its control might be due to the long history of rabies in these countries and the frequent coverage of human rabies cases by the mass media.
Vaccination of dogs against the rabies virus offers a safe and effective means to prevent rabies infection in humans [34]. Vaccination coverage should be at least 70% [1] to maintain the control of rabies between annual vaccination campaigns. Mass vaccination of dogs (70% of the estimated total number of dogs) in Bali Island, Indonesia, successfully decreased the human rabies incidence on that island by 74% [35]. In our study, around 52% of the dog owners had vaccinated at least one of their dogs in 2012. Real vaccination coverage (number of dogs vaccinated divided by the size of the total dog population) will be lower than the estimated 52%, as most households own multiple dogs, which makes it hard to handle them all at a single time during a vaccination campaign. The Sikka Regency estimated the vaccination coverage during the 2012 campaign to be around 58%. This recorded coverage rate, however, overestimated the real coverage as it did not account for the dogs and their owners that were not at home during the ‘house-to-house’ campaign. These non-registered dogs were estimated to represent approximately 30% of the total dog population [4]. These findings indicate that the real rate of vaccination coverage for the dog population in Flores Island is still far below the WHO recommended rate of 70% [1]. A targeted vaccination coverage of 70% is very important to maintain the overall herd immunity between campaigns above the threshold immunity coverage (e.g. 20–45%; [36]). This is especially relevant for Flores Island, where the dog population is characterized by a high turn-over rate (>45%) [37] and the vaccine used has a short duration of immunity. High quality, cell-culture vaccines are recommended for rabies control, such as Rabisin, which was used to effectively reduce the prevalence of dog and human rabies in Bali [35].
Relatively more dog owners in Sikka vaccinated their dogs than in Manggarai, whereas in Manggarai the proportion of dog owners that had culled dogs was higher than in Sikka. This difference reflects the different approaches used by the regencies to implement control measures. The animal health authority of Sikka has focused on vaccination of dogs as the main approach to control rabies in the regency, which is in line with the national campaign. Whereas the authority in Manggarai implemented culling of roaming dogs as an additional control measure alongside the national campaign. In 2010, for example, Manggarai conducted mass culling of 2,440 dogs (24% of the estimated total number of dogs in the regency), which were free roaming in the public area, regardless of the vaccination status of these dogs [38]. As a consequence, the size of the dog population in Manggarai reduced considerably [4]. The number of registered dogs during the vaccination campaign of 2012 was six times lower than in Sikka, even though the size of the human population in both regencies was comparable. A positive impact of culling is the removal of all potentially exposed dogs in infected villages, thereby reducing the transmission of rabies between dogs and decreasing the risk of rabies for humans [10]. The culling of free roaming dogs was, however, less acceptable for the local community in comparison to the vaccination campaign [4]. This resulted in unintended negative consequences [39] in which the dog owners hid or moved their dogs to another village during the incubation phase of rabies [9,40]. In this context, the OIE and other international animal health related organizations (e.g. WHO, WSAVA, and GARC) do not recommend culling as a rabies control measure [27,41,42,43]. Culling (i.e. the killing of dogs regardless of their health status) is not effective in controlling rabies [44] and can be counterproductive [43], as previously vaccinated dogs may also be culled.
The difference in vaccination uptake between the regencies might also be due to the intensity of local community support for the control campaign [45]. Religious and village leaders in Sikka participated actively in encouraging dog owners to vaccinate their dogs, whereas this was not the case in Manggarai. The encouragement of community leaders may have stimulated dog owners to increase their efforts to catch and restrain the dog to be vaccinated. In addition, the vaccination campaign in Sikka coincided with the national celebration of World Rabies Day on 8th October 2012 in Maumere. During this celebration, religious and village leaders were invited to join the event to share experiences on rabies control measures in their villages. Our findings suggest that the involvement of local communities in rabies control activities can be important to implement rabies control measures successfully in a regency. Moreover, a good collaboration among sectors, such as public health and veterinary authorities, is also important, as was reported from Latin America [1].
Theoretical and empirical evidence suggests that reducing population density through sterilization, which includes castration of male dogs, does not reduce disease transmission [36,44]. There is limited, equivocal empirical evidence that in open, dynamic populations mass sterilization extends vaccination by reducing the number of new susceptible dogs entering the population through reducing local births [46,47,48]. Nonetheless, castration of males has been encouraged by the local authority since 2000, as a method to limit the number of free roaming dogs within villages in Flores and restrict male dog behavior (such as dispersal and fighting) that facilitates the spread of rabies [1]. Humane dog population management, which includes sterilization and the provision of basic dog health care, is currently recommended as a supplementary measure to mass vaccination programs [41], and for this reason we have estimated the prevalence of castration. Only 12% of dog owners reported that at least one of their male dogs had been castrated during the period 1999–2000. This low uptake might be due to the lack of skill to castrate dogs as castration of male dogs was carried out by the dog owners themselves or by their family. Another reason could be attributed to the dog owners’ preference, especially those who keep female dogs for breeding, to have a male dog without castration.
Dog owners, both in Sikka and Manggarai, had a high level of knowledge about the preventive measures to be taken after being bitten by a suspected rabid dog. However, a high level of knowledge of rabies and its prevention does not guarantee a high uptake of proper treatment after exposure. In our study, less than 60% of patients (89 reported bite cases in humans) went to the medical center to seek proper medical treatment, even though the majority of the community knew that rabies in humans is fatal and can be prevented by a series of vaccine injections after exposure. Most of the people in Flores Island that died after being bitten by a rabid dog did not receive any post exposure treatment (Purnama, personal communication, 2014). In addition to the level of knowledge, socio-economic factors such as income level and distance to the nearest rabies-treatment center can contribute significantly to the decision to adopt the appropriate treatment [49]. People living in rural areas, far from any rabies-treatment center, may not have access to prompt and appropriate treatment [49]. Even if the treatment is provided for free (e.g. costs of human rabies vaccines and physician costs are paid by the individual Regency Governments of Flores Island), costs associated with travel to and from the rabies-treatment center and income loss during the treatment [50] could prevent dog-bite victims from seeking medical care.
The distribution of reported control measures in dogs and bite cases in humans over the previous fourteen years may have been influenced by recall bias, reflecting the extensive time frame posed in the research questions. However, the results give an overview of the uptake of rabies control measures over the previous fourteen years and could be used for better planning of rabies control in the future.
In the univariable analysis, knowledge of rabies and its control measures was significantly associated with the 2012 vaccination uptake by dog owners. However, in the multivariable analysis this association was no longer significant. This indicates that the level of rabies knowledge did not have a direct effect on the uptake of the 2012 vaccination campaign. An important factor associated with vaccination uptake was the accessibility of the village. Uptake of the vaccination campaign was four times more likely for dog owners living in a village with good infrastructure than for those living in more remote villages with a poor road infrastructure. This is an interesting finding, as this factor has not been studied before. The accessibility of a village might be related to the transfer of knowledge from animal health authorities (especially the distribution of vaccination schedule information). Informal discussions during the survey with dog owners in the less accessible villages revealed that many of these dog owners became aware of the vaccination campaign only on the day it was conducted. This is in line with our survey results, which showed that one of the main reasons for not joining the vaccination campaign was the lack of information about the vaccination campaign. As a consequence, dog owners were not at home when the vaccinators arrived. This corresponds with the study results of Durr et al [16], in which 26% of the surveyed dog owners did not join the vaccination campaign as they were not aware of the time schedule. In Flores Island, it is common practice for the notification letter about the vaccination schedule to be sent to the village leader through public transportation. Given the poor accessibility of the more remote villages due to natural barriers, (especially in the rainy season, September-April) these notifications do not always reach the villages on time. As a result, dog owners in less accessible villages are less informed about the vaccination campaign. This finding suggests that targeted distribution of vaccination campaign information within these villages is an effective and practical way to increase the uptake of rabies vaccination in the future. Effective channels for the distribution of information about the vaccination schedule, prior to the visit of the vaccination team, could be through elementary school teachers [51], and church and village leaders.
The second important reason given by dog owners for not joining the vaccination campaign was the inability to handle and restrain their dogs (37%). This reason was given more frequently in our study than reported in studies from other endemic rabies countries [18,20,52]. The difference could be due a different relationship between humans and dogs in those countries. Dogs in Flores are never restricted and roam freely within the village, so the interaction between owners and their dogs is very low. Dogs in Flores Island have a primary function as guard dogs. This type of dog is more aggressive and difficult to handle compared with companion dogs. This suggests that educating the vaccinators and dog owners about dog behavior and the safe handling of dogs might improve vaccination coverage whilst limiting the risk of being bitten. Alternatively, training teams of government dog catchers, similar to Bali [35], may be required to increase vaccination coverage if owners are unable to catch and restrain their dogs.
Our study highlighted the association between keeping female dogs for breeding and the uptake of vaccination. The dog owners who had female dogs for breeding purposes were more likely to join the vaccination campaign. The perceived value of the dog may have increased the dog owners’ effort to catch and restrain the dog to be vaccinated. Other reasons that might have contributed to this association are related to the relatively longer life span and lower turn-over rate of female dogs. In the event of a traditional ceremony in which dog meat is needed, dog owners in Flores prefer to cull male dogs and keep female dogs for breeding. A reproductive female can produce puppies until the age of 11 years (reproductive lifespan) [53]. Therefore, dog owners get more benefit from vaccinating reproductive female dogs.
Furthermore, our study found that dog owners who valued their dogs at less than Rp250,000 were less likely to join the vaccination campaign compared to dog owners with dogs valued between Rp250,000 and Rp500,000. This result might be due to the age of the dogs. The majority of the dogs valued at less than Rp250,000 were younger than one year old. It is well documented that dogs younger than one year are less likely to be vaccinated by their dog owners [14,18,54,55], putting this cohort at a higher risk for contracting rabies [56]. This is the result of a common perception that this cohort is too young to be vaccinated [20,55]. Vaccine manufacturers indicate that rabies vaccine can be effectively administered to dogs at as early as 3 months of age. Therefore, the vaccination campaign in the future should place additional emphasis on this unvaccinated cohort of dogs, aged between 3 and 12 months. In addition, community education efforts should be focused on dog owners who have a female dog for breeding, encouraging them to vaccinate their young dogs before selling them. Dog owners with a high income (more than 1 million Rupiah per year) had a higher probability of having their dogs vaccinated than dog owners with lower income levels. This finding is similar to the result of a study by Beran [40]. It is common practice in Flores Island for people with a good income to hire other adults (particularly from their family) to take care of their children and home. As a consequence, there will always be an adult person present at the home to handle the dogs during vaccination. This may explain our finding that dog owners with a higher income were more likely to vaccinate their dogs.
The level of knowledge of dog owners in Flores Island about rabies and its control measures was high, but not associated with the uptake of the vaccination campaign of 2012. Overall, the uptake rate of the 2012 vaccination campaign was relatively low (52%) and differed between regencies. Geographical accessibility is one of the important predictors of vaccination uptake among dog owners. Targeted interventions in those villages with poor accessibility may increase the vaccination uptake in the future. These interventions should include: (1) an effective system for distributing information so that dog owners are provided with timely information on the vaccination schedule, for instance through elementary school teachers, and church and village leaders; and (2) the provision of dog owners and vaccinators with a technique or skill to catch and restrain dogs.
|
10.1371/journal.ppat.1001016 | Rabies Virus Infection Induces Type I Interferon Production in an IPS-1 Dependent Manner While Dendritic Cell Activation Relies on IFNAR Signaling | As with many viruses, rabies virus (RABV) infection induces type I interferon (IFN) production within the infected host cells. However, RABV has evolved mechanisms by which to inhibit IFN production in order to sustain infection. Here we show that RABV infection of dendritic cells (DC) induces potent type I IFN production and DC activation. Although DCs are infected by RABV, the viral replication is highly suppressed in DCs, rendering the infection non-productive. We exploited this finding in bone marrow derived DCs (BMDC) in order to differentiate which pattern recognition receptor(s) (PRR) is responsible for inducing type I IFN following infection with RABV. Our results indicate that BMDC activation and type I IFN production following a RABV infection is independent of TLR signaling. However, IPS-1 is essential for both BMDC activation and IFN production. Interestingly, we see that the BMDC activation is primarily due to signaling through the IFNAR and only marginally induced by the initial infection. To further identify the receptor recognizing RABV infection, we next analyzed BMDC from Mda-5−/− and RIG-I−/− mice. In the absence of either receptor, there is a significant decrease in BMDC activation at 12h post infection. However, only RIG-I−/− cells exhibit a delay in type I IFN production. In order to determine the role that IPS-1 plays in vivo, we infected mice with pathogenic RABV. We see that IPS-1−/− mice are more susceptible to infection than IPS-1+/+ mice and have a significantly increased incident of limb paralysis.
| Rabies virus (RABV) is a neurotropic RNA virus responsible for the deaths of the at least 40,000 to 70,000 individuals globally each year. However, the innate immune response induced by both wildtype and vaccine strains of RABV is not well understood. In this study, we assessed the pattern recognition receptors involved in the host immune response to RABV in bone marrow derived dendritic cells (DC). Our studies revealed that Toll like receptor (TLR) signaling is not required to induce innate responses to RABV. On the other hand, we see that IPS-1, the adaptor protein for RIG-I like receptor (RLR) signaling, is essential for induction of innate immune responses. Furthermore, we found that RIG-I and Mda-5, both RLRs, are able to induce DC activation and type I interferon production. This finding is significant as we can target unused pattern recognition receptors with recombinant RABV vaccine strains to elicit a varied, and potentially protective, immune response. Lastly, we show that IPS-1 plays an important role in mediating the pathogenicity of RABV and preventing RABV associated paralysis. Overall, this study illustrates that RLRs are essential for recognition of RABV infection and that the subsequent host cell signaling is required to prevent disease.
| Type I interferon (IFN) was first identified as a “factor” that rendered cells resistant to viral infection [1]. It is now known that following viral infection, cells induce type I IFN, which in turn upregulates the expression of numerous antiviral proteins [2]. This class of cytokines is comprised of several genes including multiple IFN-α genes, a single IFN-ß gene, and the less well-defined IFN-ω, -ε, -τ, -δ, and -κ (for review [3]). In addition to having antiviral functions, type I IFNs play a part in activating the adaptive immune response following infection [4], [5], [6]. For instance, IFN-α/ß can strengthen the innate immune response by activating antigen presenting cells (APC). Additionally, following maturation in the presence of type I interferon and GM-CSF, monocyte-derived DCs more effectively stimulate an antigen-specific CD8+ T cell response than when differentiated with GM-CSF alone [7].
Viral infection can trigger the type I IFN response via various pattern recognition receptors (PRR), namely Toll-like receptors (TLR) and RIG-I-like receptors (RLR). In the case of negative stranded RNA viruses, the members of the TLR family that are generally involved in viral recognition, TLR-3 and TLR-7, are found on the endosomal membrane. To initiate the signaling cascade, TLR-3 binds double stranded RNA molecules [8], whereas TLR-7 recognizes immunomodulatory compounds (ie-imiquimod) [9] or single-stranded RNA molecules [10]. Although negative stranded RNA viruses do not produce double stranded RNA as part of their normal replication cycle, it is likely that abnormal replication products resulting from errors by viral RNA-dependent RNA polymerases give rise to some level of double stranded RNA in virus-infected cells [11]. TLR-3 and TLR-7 initiate signaling though different adaptor molecules, Trif and MyD88, respectively; however, the pathways converge on the phosphorylation of IRF-3. Following phosphorylation, IRF-3 forms protein dimers, which allow for its transport into the nucleus where it can bind to the IFN-ß promoter [12].
Alternatively, RNA viruses can be recognized in the cytoplasm by RLRs, namely RIG-I and Mda-5 [13]. These helicase-like proteins recognize double-stranded RNA and 5′ tri-phosphate groups [14]. In the case of rabies virus (RABV), the negative stranded RNA virus of interest in this study, the leader RNA remains unmodified [15], [16] and thus provides a potential ligand for these RLRs. Signaling by RIG-I and Mda-5 is mediated through the mitochondria-bound protein IPS-1, which is also referred to as MAVS, Cardif, or VISA [17], [18], [19], [20]. Similar to what is seen in TLR signaling, RLR signaling culminates with the activation and nuclear translocation of IRF-3 [19].
Rabies virus is a member of the Rhabdoviridae family. RABV has a relatively simple genome, comprised of just 5 proteins: the nucleoprotein, phosphoprotein (P), matrix protein, glycoprotein and the RNA dependent RNA polymerase. Infection with RABV can induce IFN-α/ß production rapidly in vivo. Furthermore, it was seen that a mouse's ability to induce type I IFN, as measured by serum concentrations 4 days post infection, positively correlates to the animal's resistance to RABV [21]. The type I IFN response is also important in driving immunity, as mice injected with anti-mouse IFN-α/ß antibody prior to infection with RABV were more sensitive to the virus than mice injected with a control antibody [22]. However, RABV has the ability to antagonize type I IFN induction [23]. Thus, shortly after infection of fibroblast cells, RABV-P prevents IRF-3 phosphorylation in order to suppress IFN-α/ß production [23].
Although IFN is induced after RABV infection, RABV is able to suppress the IFN response shortly after infection. Therefore, in order to study the receptors responsible for the initial induction of IFN, several groups have used recombinant viruses with lower levels of RABV-P. Using this method, it was determined that IFN-ß promoter activity was seen following recombinant RABV infection of VERO cells transfected with wildtype RIG-I, but not in cells transfected with dominant-negative mutant RIG-I [24], thus indicating a role for RIG-I in mounting an innate immune response to RABV. Additionally, following infection of human postmitotic neurons with RABV, Prehaud et al. saw an increased production of IFN-ß and TLR-3 mRNAs [25]. Furthermore, the expression of TLR-3 on cerebellar cortex tissues of individuals that had died of rabies, but not on an individual that died of cardiac arrest, verify the viral induced expression of TLR-3 in human brains in vivo [26]. This upregulation of TLR-3 following infection suggests a possible role for TLR-3 signaling in the innate recognition of RABV; however, TLR-3 activation needs to be further studied to conclusively define such a role. Although these results hint at the receptors responsible for interferon expression, there is no evidence that other PRR receptors, such as TLR-7 and Mda-5, do not also play a role. Furthermore, since the recombinant viruses used in some of these studies exhibit decreased pathogenicity, it is possible that a wildtype virus may act differently following infection.
In order to study the IFN-inducing pathways triggered by RABV, we needed to identify a cell type in which RABV-P is unable to antagonize type I IFN signaling. Of note, it has been seen that following infection of dendritic cells (DCs) with influenza, another negative stranded RNA virus, the DCs become infected, but this infection is non-productive [27]. Here, we sought to determine whether APCs were productively infected with RABV. Similar to previous reports that human DCs are susceptible to RABV infection [28], [29], we saw that mouse DCs became infected; however we also observed that very little viral progeny was released due to limited viral replication. Due to the overall suppression of viral transcription in RABV infected DCs there are presumably low levels of RABV-P that may not be able to inhibit interferon induction. Thus, we decided to utilize infection of DC to study the IFN-inducing capabilities of RABV and found that RLRs are responsible for viral recognition in DCs.
It has been previously shown that RABV-P can inhibit the phophorylation of IRF-3 in fibroblast cells [23], thus crippling the induction IFN-α/ß. However, RABV is able to infect a variety of cells including neurons [30] and antigen presenting cells (APC) [28], [29] in addition to fibroblasts. Thus, we wanted to determine whether RABV is able to inhibit IFN signaling in other cell types including DCs, which are known to induce the adaptive immune response.
In order to check for type I IFN production, we first infected a variety of cell types including fibroblasts (BSR), neuronal cells (NA), macrophages (Raw264.7) and DCs (JAWSII) with a RABV vaccine strain-based vector, SPBN. Following infection with RABV, cell supernatants were collected and subsequently UV-treated in order to deactivate any infectious virus but retain secreted cellular proteins, such as type I IFN. We then transferred the supernatants to reporter cells, which are sensitive to IFN. Twenty-four hours after supernatant transfer, reporter cells were infected with recombinant vesicular stomatitis virus expressing GFP (VSV-GFP, [31]) for 5–8h. VSV replication is highly sensitive to type I IFN [32], and thus, in the presence of type I IFN, the replication of VSV is suppressed [4]. Following infection with RABV, macrophages as well as DCs, but not fibroblasts or neuronal cells, produce type I IFN that inhibits VSV-GFP replication, as indicated by the lack of GFP expression (Figure 1A). Of note, when BSR, NA, Raw264.7, or JAWSII cells are originally treated with UV-deactivatecd RABV, the supernatants from these cells are unable to block VSV replication (Figure 1B); therefore, IFN is secreted only after RABV replication.
In order to account for the increased amounts of type I IFN produced following RABV infection of macrophages and DCs when compared to the amount produced by fibroblast and neuronal cells, we did a one-step growth curve following infection of the various cell types. Supernatants from infected cells was titered on BSR cells, which are insensitive to type I IFN [4]. We detected that, although all four cell types were infected, only BSR and NA cells produce infectious virus (Figure 2A). There are two possible explanations for the defect in viral production observed here: either a block in viral replication or a defect in viral assembly. In order to compare viral transcription and replication in fibroblast and dendritic cell lines we used quantitative PCR. In fibroblast cells, we saw that the amount of RABV-N messenger RNA (mRNA) transcripts increased an average of 1.95 logs from 8 hours post infection (hpi) to 48 hpi. Similarly, the quantity of RABV-N genomic RNA transcripts (gRNA) increased an average of 1.2 logs from 8 hpi to 48 hpi. This data indicates that following infection of fibroblast cells both viral transcription (mRNA) and replication (gRNA) occurs. On the other hand, when looking at the quantity of RABV-N found in dendritic cells following infection we see that there was no increase in the number of mRNA or gRNA viral transcripts when comparing 8hpi to 48 hpi. Thus, it appears that RABV is able to enter APCs, but only limited viral transcription occurs following entry.
It is reasonable to assume that decreased levels of transcription might result in low levels of RABV-P. It has been previously shown that recombinant RABV expressing low amounts of RABV-P is unable to inhibit type I IFN induction [23]. Furthermore, we show by Western blotting that cell lysate from RABV infected JAWSII cells contained undetectable levels of RABV P 48hpi (Figure 2B). On the other hand, we were able to detect RABV P in lysates from infected BSR cells as early as 12 hpi after infection. This results support the conclusion that a very low level of RABV P within infected APCs is not able to block the induction of type I IFN and therefore is responsible for the increase in type I IFN production by these cells following infection.
In order to better understand the interaction of RABV with host cells following infection, we sought to identify the pathway(s) responsible for type I IFN induction in infected cells. Since we detected that DCs make large amounts of IFN following RABV infection we decided to use bone marrow derived DCs (BMDC) in our studies. To differentiate BMDCs, we cultured the cells in the presence of 10 ng/ml GM-CSF. After 7 days the majority of cells have matured to DCs as shown by the expression of CD11b+CD11c+ (Figure 3). In order to identify the PRR that recognizes RABV, we isolated BMDC from mice deficient in various signaling components of PRR pathways. In each experiment, cells were stimulated, and the CD11c+ cell population (Figure 3) was analyzed for production of type I IFN and expression of CD86, a co-stimulatory molecule that is upregulated on activated DCs.
First, we analyzed the role that TLR signaling plays in BMDC activation and type I IFN production following a RABV infection. It has been previously reported that following infection of human postmitotic neurons with RABV, there is an increased production of IFN-ß and TLR-3 mRNAs. In addition, treatment of neurons with poly(I:C), a TLR-3 agonist, generated a similar cytokine profile to that which was seen following RABV infection [25]. Thus, we differentiated BMDCs from TLR-3−/− and congenic wildtype mice and infected the cells with RABV. We then analyzed the infected cells for the presence of CD86 (Figure 4A). As shown in Figure 4C, there is no significant difference in the expression of CD86 on the cell surface of RABV infected BMDCs derived from TLR-3−/− or wildtype mice. As expected, TLR ligands that signal via other TLR receptors, namely TLR-4 (LPS), TLR-9 (ODN1826), and TLR-7/8 (R848), equally activate BMDCs derived from wildtype (wt) or TLR-3−/− mice. Interestingly, poly(I:C), a known ligand for TLR-3, was able to activate BMDC isolated from TLR-3−/− mice as well as wt mice. However, it has been previously shown that poly(I:C) can also signal through Mda-5 and that Mda-5 is the dominant receptor for mediating type I IFN induction following poly(I:C) stimulation in BMDCs [33], [34]. As the RLR pathway remains intact in TLR-3−/− mice, BMDC activation in TLR-3−/− cells following poly(I:C) stimulation is not inexplicable, but rather highlights the need for a better TLR-3 agonist. Taken as a whole and based on the fact that that RABV infection activated BMDC derived from both wt and TLR-3−/− mice equally, we conclude that TLR-3 signaling is not required for the activation of BMDCs following a RABV infection.
To our knowledge, TLR-7 has never been investigated in the context of a RABV infection and thus the role that it plays in type I IFN induction and DC activation following RABV infection is unknown. To analyze the function that TLR-7 has in the induction of type I IFN and DC activation, we isolated BMDCs from MyD88−/− and C57BL/6 mice. We detected an equal upregulation of CD86 on BMDCs from MyD88−/− and wildtype mice (Figure 4B). As expected, activation of MyD88−/− BMDCs is significantly reduced following stimulation with ODN1826 and R848, ligands for TLR-9 and TLR-7/8 respectively, both of which signal via MyD88 [35] (Figure 4D). Thus we conclude that, similar to TLR-3 signaling, the activation of BMDCs following a RABV infection occurs independently of MyD88 signaling.
In order to determine if TLR-3 and MyD88 signaling might have an impact on type I IFN production, supernatant from infected BMDCs was collected at various times post infection, and a VSV-sensitivity assay was performed. As seen with BMDC activation, both TLR-3 and MyD88 are dispensable in the induction of type I IFN (Table 1). We did not detect any VSV-GFP replication on reporter cells following pre-treatment with supernatant from TLR-3−/−, MyD88−/−, or wildtype BMDC, indicating the presence of type I IFN in the supernatant.
Having excluded TLRs as the required receptors mediating BMDC activation and type I IFN production, we next looked at the potential role for RLR signaling. Hornung et al. showed that a recombinant RABV expressing low levels of RABV-P signals via RIG-I to induce IFN-ß promoter activity following infection. Furthermore, it was shown that the 5′-triphosphate on the leader sequence of RABV was the ligand for RIG-I [24]. To determine whether the RIG-I pathway is also activated in DCs following RABV infection, we isolated BMDCs from IPS-1+/+, +/−, or −/− mice. Our results indicate that following infection with RABV, IPS-1+/+ and IPS-1+/− BMDCs express high levels of CD86 on their surface (Figure 5A–B). Of note, IPS-1+/− BMDCs are slightly less activated then IPS-1+/+ cells. On the other hand, IPS-1−/− BMDCs express significantly lower levels of CD86 on their surface at all time points (Figure 5A–B). The TLR ligands LPS, ODN1826, and R848 equally activated all IPS-1 BMDC samples, indicating that the defect in the IPS-1 −/− BMDCs is specific to the RLR pathways (Figure 5B). As such, when cells are stimulated with RLR agonists, there is a defect in the activation of IPS-1−/− BMDCs when compared to IPS-1+/+ or +/− BMDCs. We see a low CD86 upregulation following both poly(I:C) stimulation and infection with a NS1-deficient strain of influenza (ΔNS1/PR8) (Figure 5B). It has been reported previously that poly(I:C) can signal via Mda-5 [33] and ΔNS1/PR8 signals exclusively via RIG-I [34]. Taken together this data indicates that BMDC activation is dependent on IPS-1 signaling following a RABV infection.
In order to determine whether type I IFN production by BMDC is also dependent on IPS-1 mediated signaling, we assayed for the presence of type I IFN in the supernatants of infected IPS-1 BMDCs by VSV-GFP sensitivity assays and quantified the amount of IFN-ß by ELISA. It was seen that supernatant obtained from IPS-1+/+ and IPS-1+/− BMDCs infected with RABV was able to inhibit VSV-GFP replication, and thus contained type I IFN. On the other hand, the VSV-GFP replication on reporter cells was not inhibited by pre-treatment with supernatants from RABV infected IPS-1−/− BMDCs (Figure 5C). Likewise, IPS-1 +/+ BMDCs produce on average 250 pg/ml IFN-ß while the IPS-1−/− BMDCs produced less than 16.7 pg/ml, if any, IFN-ß (Figure 5C). These results indicate that RABV infected IPS-1−/− BMDCs do not secrete type I IFN. Also consistent with the results seen for BMDC activation, IPS-1−/− cells stimulated with RLR agonists produced less type I IFN compared to IPS-1+/+ or +/− BMDCs (Table 2).
It has been shown that IPS-1 mediated pathways are also capable of activating the NF-κB signaling cascade [36]. Thus, we quantified the amount of IL-6 in the supernatant of RABV infected BMDC isolated from IPS-1+/+, +/− and −/− mice (Figure 5D). We see that there is a significant decrease in IL-6 produced by IPS-1−/− BMDCs compared to IPS-1+/+ BMDCs. However, IPS-1−/− cells do secrete some IL-6 following infection with RABV, and thus, the use of IPS-1 independent pathways to induce NF-κB activation, in contrast to type I IFN activation, seems to be utilized.
Mda-5 mediated induction of IFN-ß has been described to occur in response to plus-stranded RNA viruses like picornaviruses, whereas it is reported that RIG-I is responsible for type I IFN induction in response to rhabdovirus infection [34]. However, the function of Mda-5 in the innate immune response to rhabdoviridae has not yet been elucidated. Furthermore, the role of these PRRs following a RABV infection in DCs remains unknown. Therefore we wanted to determine which of the two receptors recognizes RABV. For this approach, BMDCs from Mda-5−/− mice and RIG-I−/− mice were isolated. As shown in Figure 6A, Mda-5−/− BMDCs express high levels of CD86 on their surface at 24 and 48 hpi. Of note, there is a significant reduction of CD86 surface expression on Mda-5−/− BMDCs at 12 hpi when compared to wildtype cells. Likewise, RIG-I −/− BMDCs also have a defect in BMDC activation at 12 hpi, while CD86 expression at 24 and 48 hpi is equal for RIG-I−/− and RIG-I+/+ cells (Figure 6B). In addition, it appears that while Mda-5−/− cells are able to induce type I IFN expression 12 hpi, RIG-I−/− cells have an early defect in type I IFN induction. Importantly, by 48hpi, RIG-I−/− BMDC do produce enough type I IFN to suppress VSV-GFP replication (Figure 6C). This indicates that RABV can induce BMDC activation and type I IFN via both Mda-5 and RIG-I ligation. Furthermore, any perturbation in IPS-1 mediated signaling cascades seems to affect the early response (12hpi) to RABV.
Once type I IFN is produced, it will further activate the infected cell via autocrine signaling through IFNAR. Ligation of the IFNAR initiates the Jak/STAT signaling cascade, which culminates in the upregulation of antiviral genes. In addition to antiviral genes, Jak/STAT signaling also upregulates proteins required for type I IFN induction, thus providing a positive feedback for the type I IFN pathway [2]. In order to determine how much IFN induction is directly related to RABV infection and how much is due to positive feedback that is driven by IFN-α/ß production, we infected BMDCs derived from IFNAR−/− mice, which eliminates the contribution of positive feedback. Interestingly, BMDC isolated from IFNAR−/− mice produce enough type I IFN to block VSV-GFP replication on reporter cells after 12, 24, and 48 h (Figure 7A). However, we detected a significant decrease in the CD86 cell surface expression of IFNAR−/− BMDC when compared to wt BALB/c mice (Figure 7B–C). Thus, although RABV infection is sufficient to induce type I IFN, the cells need an amplification signal in order to undergo maturation. Additionally, we see a significantly greater infection by RABV in IFNAR−/− cells, presumably due to their inability to induce antiviral gene expression (Figure 7D).
Lastly, we wanted to determine the impact that the RIG-I and Mda-5 pathways play in the in vivo response to RABV utilizing IPS-1 −/− mice. Interestingly, we detected that IPS-1−/− BMDC, which do not produce type I IFN, have significantly more RABV-N expression post infection (Figure 8A). This indicates that in the absence of IFN-α/ß induction, viral replication in DCs occurs at a faster rate, which should also increase viral pathogenicity. Therefore, we infected IPS-1 −/−, +/−, and +/+ mice, intramuscularly with SPBN-N2c, a recombinant RABV that is modestly pathogenic after peripheral inoculation [37]. Figure 8B shows that about 60% of the IPS-1+/+ or +/− mice lived, while only 45% of the IPS-1 −/− mice survived infection. More dramatically, nearly 90% of the IPS-1−/− mice had hind limb paralysis 11 days post infection while the IPS-1+/+ and +/− mice exhibited only about 45% paralysis (Figure 8C). This data indicates that RABV infection of IPS-1−/− is more pathogenic than RABV infection in wildtype mice.
It has been previously seen that RABV can infect APCs [28], [29]; however, the impact of the infection on generating an innate immune response to RABV had not been delineated. We show here that following RABV infection of APCs, unlike fibroblasts or neuronal cells, are able to produce copious amounts of type I IFN. We also determined that infected APCs do not produce novel viral progeny. A similar phenotype has also been seen following influenza infection of DCs. BMDCs become infected by the influenza strain, PR8, as seen by co-expression of influenza HA and DC marker N418 on 72% of cells. However, infected BMDC do not release viral progeny, as seen by a failure of infected DC supernatants to induce hemagglutination of chicken red blood cells [27]. Non-productive infection of DCs may have significant biological relevance over the course of an infection. Since RABV infection within APCs is easily controlled, the cells become a source of viral antigen, with little risk of spreading infection to neighboring cells. Taken together, APCs seem to be of critical importance during a RABV infection both for the prolonged production of type I IFN as well as a source of viral antigen.
In this study, we used APCs as a tool to study the PRRs used to recognize RABV following infection. Interestingly, we see that TLR-3 has no role in inducing a type I IFN response or DC activation despite its previously recognized upregulation following RABV infection [25]. However, recent publications may explain this potential discrepancy. It was reported that TLR-3 is required for the formation of Negri bodies in RABV infected cells and that these bodies are the site of viral replication [38], [39]. Furthermore, TLR-3 −/− mice are less susceptible to infection with pathogenic RABV, as seen by increased survival and lower viral titers in the brains of TLR-3 −/− animals compared to wt mice [39]. Thus, the requirement for TLR-3 by RABV may explain why it is upregulated following infection despite the fact that it is not required for a type I IFN response.
We next sought to identify whether TLR-7 was critical for DC activation and type I IFN production. To our knowledge, no one has directly examined the role of TLR-7 following a RABV infection. Of note, TLR-7 signaling does play a role in the cellular recognition of a closely related Rhabdovirus, VSV. Infection of wild type plasmacytoid DCs (pDC) with VSV induced the production of IFN-α. However, infection of pDCs from TLR7−/− or MyD88−/− mice resulted in no cytokine production [40]. Thus, indicating that single-stranded RNA derived from VSV is able to trigger TLR-7 signaling. However, in the case of RABV it appears that MyD88-dependent signaling, and thus TLR-7, is dispensable for IFN-α/ß production following infection.
On the other hand, RLR signaling via IPS-1 is critical for both the activation of DCs and production of type I IFN by infected DCs. It was shown previously that RIG-I signaling is necessary for IFN-ß promoter activity in VERO cells following recombinant RABV infection [24]. However, we show here using RIG-I−/− derived DCs that Mda-5 is also able to induce DC activation and type I IFN production. This is interesting, as Mda-5 is generally recognized as a receptor for positive stranded RNA viruses, not negative stranded RNA viruses like RABV. Of note, another negative stranded RNA virus of the Paramyxoviridae family, Sendai virus, requires MDA-5 signaling for the sustained expression of type I IFN [41].
Our data indicates that RABV can be recognized by either RIG-I or Mda-5 following infection. The use of both RIG-I and Mda-5 receptors has also been observed following infection with West Nile virus (WNV). Following WNV infection, RIG-I−/− cells had a delayed upregulation of host anti-viral genes; however, the ability to respond was conserved. Thus, indicating that another receptor was involved in recognition of WNV, this receptor was identified as Mda-5 [42], [43]. Following RABV infection in the absence of either RIG-I or Mda-5, there is a delay in the activation of BMDCs. Furthermore, RIG-I−/− BMDCs have an early defect in type I IFN production. Thus, it appears that in response to a RABV infection, both RIG-I and Mda-5 are utilized in order to rapidly induce high levels of IFN-α/ß production and DC activation.
Of note, IPS-1 +/− cells exhibited a phenotype that was intermediate to IPS-1 −/− and IPS-1 +/+ mice. This observation also supports the requirement for rapid induction of IFN-α/ß following a RABV infection. The heterozygous cells are lacking one of the IPS-1 alleles, and this may result in less functional protein in the heterozygous mice compared to homozygous wildtype mice. This again highlights the importance of a rapid response following viral infection in order to control viral replication and spread.
The type I IFN response occurs in two phases after infection: the induction of IFN-α/ß following recognition of the pathogen by a PRR and then autocrine or paracrine signaling by IFN-α/ß through the IFNAR to induce upregulation of many other genes. Included among the genes that are upregulated in response to IFNAR signaling are several genes required for PRR signal transduction [2]. In this manner, the infected cell undergoes positive feedback to increase both the host response and PRR signaling. We wanted to identify which arm of the IFN response was responsible for the effects we observed following RABV infection of DCs, viral induction or IFN-α/ß amplification. We saw that both wt and IFNAR −/− mice are able to induce type I IFN production, thus highlighting the host's ability to rapidly induce IFN-α/ß following infection with RABV and indicating that the amplification of IPS-1 signaling by IFNAR signaling is not a critical factor in the induction of type I IFN. Surprisingly, we see that in the absence of IFNAR signaling, there is very little BMDC activation. Thus, it appears that DC activation occurs via IFN-α/ß signaling and is not a direct consequence of viral infection. This fact highlights the importance of a type I IFN response in initiating the adaptive immune response following infection with RABV.
Lastly, we sought to determine the biological relevance of IPS-1 mediated PRR signaling following infection with a pathogenic strain of RABV. Although this experiment did not focus specifically on type I IFN production by DCs, it indicates how IPS-1 signaling, and thus IFN-α/ß production and DC activation, impacts the prognosis of infected animals. We saw that 87% of the IPS-1−/− mice in the study became paralyzed, whereas only about 45% of the IPS-1 +/+ or +/− mice exhibited signs of paralysis. There is some data that suggests paralysis following a RABV infection is an early symptom of disease. In humans who present with the less common paralytic rabies, their survival time is slightly longer [44]. Although not significant, this data supports the fact that an early, rapid type I IFN response is an important factor mediating RABV disease outcome. Of note, opposed to vaccine strain of RABV used in the BMDC experiments, the pathogenic RABV strain, SPBN-N2c, infects mostly neurons [37] and we showed here that RABV is able to suppress the type I IFN response in neurons by 12hpi (Figure 1). Despite this limitation, there is no other model to study RABV pathogenicity. The role that antigen presenting cells play in initiating the immune response to RABV in vivo should also be investigated further. It is known that pathogenic RABV is less immunogenic than vaccine strains of RABV [45] thus it is likely that pathogenic RABV avoids or alters infection of DC in order to elicit a lesser immune response.
In summary, we show here that RABV replication is cell type dependent; namely, RABV is able to antagonize the induction of type I IFN in fibroblast and neuronal cells but is unable to inhibit IFN-α/ß induction in APCs. Furthermore, in APCs RABV infection is non-productive due to a defect in viral transcription, and no viral production is observed. Infection of BMDCs allowed us to delineate that RABV is exclusively recognized by either RIG-I or Mda-5 and both receptors are required for a rapid type I IFN response to RABV. This finding has significant implications for the development of a RABV-based vaccine vector. In light of these results, a recombinant RABV expressing a TLR agonist may allow for RABV recognition via TLRs. Such a response may potentiate the type I IFN response and induce better protection in a vaccine setting. We also show here that BMDC activation is secondary to IFN-α/ß induction and requires IFNAR. In addition, IPS-1 mediated signaling does have a role in vivo, as it seems to play a critical role in preventing RABV pathogenesis following RABV challenge.
All animals were handled in strict accordance with good animal practice as defined by the relevant international (Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC) (Accreditation Status TJU: Full)) and national (TJU Animal Welfare Assurance Number: A3085-01), and all animal work was approved by the Institutional Animal Care and Use Committee (IACUC) at Thomas Jefferson University TJU. Animal use protocols are written and approved in accordance with Public Health Service Policy on Humane Care and Use of Laboratory Animals, The Guide for the Care and Use of Laboratory Animals. TJU IACUC protocol number 414A, 414G, and 414I were utilized in this study.
The fibroblast cell line used in these studies is a cell clone of BHK-21 (ATCC: CCL-10), BSR. The neuronal cell line used in these studies is a neuroblastoma cell line referred to as NA [46]. The antigen presenting cell lines used here were JAWSII (ATCC: CRL-11904) and Raw264.7 (ATCC: TIB-71).
Mice used in this work are as follows: B6/129S1-Tlr3tm1Flv/J (TLR-3 −/−, Jackson Laboratory, stock 005217); B6129SF1/J (Jackson Laboratory, stock 101043); MyD88−/− [47]; C57BL/6 (NIH); IPS-1 [48]; Mda-5 −/− [33]; RIG-I [49]; BALB/c (NIH); and IFNAR−/− mice [50].
For pathogenicity study, IPS-1 mice were genotyped as described previously [48]. IPS-1 +/+ (n = 7), IPS-1 +/− (n = 13), and IPS-1 −/− (n = 15) mice were infected intramuscularly with 106 ffu SBPN-N2c [37]. The weight of the mice was monitored daily, and the animals were euthanized after losing 25% of their body weight, which indicates a severe rabies infection.
Cellular supernatants were assessed for the ability to inhibit vesicular stomatitis virus (VSV) replication as described previously [4]. Briefly, the cell line of interest was infected with the vaccine strain of RABV, SPBN, at a multiplicity of infection (MOI) of 10, and supernatant was collected at various time points post infection. Alternatively, supernatant from infected BMDCs was used. The supernatants were UV-deactivated with a 254nm UV light source for 15 min. UV-deactivated viral supernatant was then diluted 1∶10 in RPMI-1640 and added to a reporter cell line (either NA, for cell line experiments or 3T3 cells, for BMDC experiements). Following the 24 h pre-treatment, reporter cells were infected with VSV- expressing GFP at a MOI of 5 for 5–8 h. VSV replication was determined by fluorescence under a UV light source.
For IFN-ß ELISA (PML Laboratories) the manufacturer's protocol was followed with the following modification: 50 µl of sample or standard was loaded into the 96-well plate. For IL-6 ELISA (eBioscience) the manufacturer's protocol was followed. Briefly, 5 µg/ml coating antibody was added to MaxiSorb (Nunc) plates and kept at 4°C over night. Wells were then washed with 0.05% Tween-20/PBS and blocked with Assay Buffer (eBioscience) for 2 hours. Plates were again washed with 0.05% Tween-20/PBS and then 100 µl standard or sample and 50 µl Biotin-Conjugate was added to the plate. Plates were incubated at room temperature for 2 hours, on a microplate shaker set at 200 rpm, and then washed with 0.05% Tween-20/PBS. Subsequently, wells were incubated with Streptavidin-HRP at room temperature for 1 hour, on a microplate shaker set at 200 rpm. The wells were washed and developed with 100 µl of Substrate Solution for 10 min followed by the addition of 100 µl of Stop Solution. Absorbance at 450 nm was recorded for each well. For both ELISAs a fourth-order non-linear regression curve (Prism software, GraphPad version 4.00) was fit to the standard curve and used to determine the concentration of the unknown samples.
BSR, NA, JAWSII and Raw264.7 cells were infected with SPBN at a MOI of 10. Following 60 min incubation at 37°C, the virus was aspirated, and cells were washed twice with PBS to remove any virus that had not yet infected the cells. Media was then added to the cells, and, at indicated time points, 0.3ml of supernatant was removed and stored at 4°C. The aliquots were titered in duplicate on BSR cells.
Messenger and genomic RABV-N RNA in SPBN (MOI-10) infected BSR and JAWSII cells was determined by TaqMan probe-based real-time PCR as described previously [4], [37].
Western blotting was performed as described previously [51].
BMDCs were differentiated as described previously [52]. Briefly, bone marrow (BM) was obtained from the mouse's tibia and femur. Following red blood cell lysis using ACK lysis buffer (Invitrogen), the BM cells were cultured in 24-well costar plates at a density of 1 million cells per ml in the presence of 10ng/ml GM-CSF (Peprotech). During the 7 day culture, the cells were washed once by aspirating 600µl of media from the wells and adding 1ml of fresh media supplemented with 10ng/ml GM-CSF. On the seventh day of culture, the non-adherent and semi-adherent cells were collected and used as the BMDCs population.
BMDCs were plated in 12-well plates (Nunc) at a maximum density of 1 million cells per ml of media. BMDCs were infected with SPBN at an MOI of 10 or ΔNS1/PR8 [53] at an MOI of 1. SPBN was harvested from BSR cells grown in serum free Opti-Pro media 4 and 7 dpi. Viral supernatant was pooled and spun at 1600 rpm for 10 min to remove cell debris. Alternatively, cells were left uninfected or stimulated with UV-deactivated SPBN, LPS (2µg/ml, Sigma), ODN1826 (2µM, InvivoGen), R848 (1 µg/ml, InvivoGen), or poly(I:C) (50 µg/ml, InvivoGen). Infected BMDC were kept at 37°C with 5% CO2 for 12, 24, or 48 h.
Following differentiation of BMDC, cells were characterized for expression of DC markers. Briefly, cells were washed in FACS buffer (2% BSA/PBS) and blocked at 4°C for 30–60 m with 2µl rat anti-mouse CD16/CD32 (Fc block) (BD Biosciences Pharmigen) in 100µl FACS. Cells were then stained with APC-CD11b, PerCP-B220, and FITC-CD11c (BD Biosciences Pharmingen) for 30 min at RT. After staining, cells were washed with FACS buffer and fixed with Cytofix (BD Biosciences) for 16–18 hours at 4°C. Samples were washed and resuspended in 300 µl of FACS buffer. Samples were analyzed on BD FACS Calibur and a minimum of 50,000 events were counted.
Following infection, BMDCs were analyzed for the expression of activation markers. At each given timepoint, BMDCs were removed from wells with cell scrappers and spun at 1600rpm for 5 min. Cells were then blocked at 4°C for 30–60 min with Fc block in 100µl FACS. Cells were then stained with APC-CD11c and PE-CD86 (BD Biosciences Pharmingen) for 30 min at RT. After staining, cells were washed with FACS buffer and fixed with Cytofix (BD Biosciences) for 16–18 hours at 4°C. Cells were then washed twice in Perm/Wash Buffer (BD Bioscience) and then stained with FITC-anti RABV-N (Centacor, Inc) for 30 min at RT. After staining, cells were washed with Perm/Wash buffer and then resuspended in 300 µl of FACS buffer. Samples were analyzed on BD FACS Calibur and 20,000–30,000 APC+ events were counted.
All data were analyzed by Prism software (GraphPad, version 4.00). To compare two groups of data we used an un-paired, two-tailed T-test. For all tests, the following notations are used to indicate significance between two groups: *p<0.05, **p<0.01, ***p<0.001.
|
10.1371/journal.ppat.1006167 | Epigenetic Landscape of Kaposi's Sarcoma-Associated Herpesvirus Genome in Classic Kaposi's Sarcoma Tissues | Kaposi's sarcoma-associated herpesvirus (KSHV) is etiologically related to Kaposi's sarcoma (KS), primary effusion lymphoma (PEL) and multicentric Castleman's disease (MCD). It typically displays two different phases in its life cycle, the default latency and occasional lytic replication. The epigenetic modifications are thought to determine the fate of KSHV infection. Previous studies elegantly depicted epigenetic landscape of latent viral genome in in vitro cell culture systems. However, the physiologically relevant scenario in clinical KS tissue samples is unclear. In the present study, we established a protocol of ChIP-Seq for clinical KS tissue samples and mapped out the epigenetic landscape of KSHV genome in classic KS tissues. We examined AcH3 and H3K27me3 histone modifications on KSHV genome, as well as the genome-wide binding sites of latency associated nuclear antigen (LANA). Our results demonstrated that the enriched AcH3 was mainly restricted at latent locus while H3K27me3 was widespread on KSHV genome in classic KS tissues. The epigenetic landscape at the region of vIRF3 gene confirmed its silenced state in KS tissues. Meanwhile, the abundant enrichment of LANA at the terminal repeat (TR) region was also validated in the classic KS tissues, however, different LANA binding sites were observed on the host genome. Furthermore, we verified the histone modifications by ChIP-qPCR and found the dominant repressive H3K27me3 at the promoter region of replication and transcription activator (RTA) in classic KS tissues. Intriguingly, we found that the TR region in classic KS tissues was lacking in AcH3 histone modifications. These data now established the epigenetic landscape of KSHV genome in classic KS tissues, which provides new insights for understanding KSHV epigenetics and pathogenesis.
| Epigenetic modifications are thought to determine the fate of KSHV infection. The epigenetic landscape of KSHV genome in in vitro cell culture systems was well studied previously. However, the physiologically relevant scenario in clinical KS tissues is unclear. In this study, we performed ChIP-Seq experiments in classic KS tissues and mapped out the AcH3 and H3K27me3 histone modifications on KSHV genome, as well as the genome-wide LANA binding sites. The results revealed a similar H3K27me3 landscape but distinct AcH3 patterns on the KSHV genome compared to the results from in vitro cultured PEL and KSHV infected SLK cells. Intriguingly, there were different LANA binding sites seen on the host genome and a reduced number of AcH3 histone modifications at the TR region of KSHV genome were found. The established epigenetic landscape of KSHV genome in classic KS tissues provides new insights towards our understanding of KSHV epigenetics, which is important for future studies on the mechanism of KSHV infection and pathogenesis.
| Kaposi's sarcoma-associated herpesvirus (KSHV) was first identified in Kaposi's sarcoma (KS) biopsies by Chang and Moore in 1994 and has been proven to be the etiological agent of several human cancers including KS, primary effusion lymphoma (PEL) and multicentric Castleman's disease (MCD) [1–3]. KSHV is a double stranded DNA virus with a large genome about 170 Kb, belonging to the gamma herpesvirus subfamily [4, 5]. It typically displays two different phases in its life cycle, the default latency and occasional lytic replication [5]. During latency, the viral genomes persist as episomes with limited latent gene expression in the nucleus of the infected cell and no virion is produced [6, 7]. The latent genes are grouped at one locus in the genome, including ORF73 (latency-associated nuclear antigen, LANA), ORF72 (vCyclin), ORF71/K13 (vFlip), K12 (Kaposin) and a cluster of miRNAs [8, 9]. Under specific conditions such as hypoxia, cell stress and valproic acid or butyrate stimulation, the virus will reactivate from latency with an orchestrated expression of lytic genes, leading to the massive production of mature virions [10–12]. Replication and transcription activator (RTA) encoded by ORF50 is the key switch regulator that controls KSHV reactivation [13, 14]. Adding inhibitors of DNA methyltransferases or histone deacetylases to KSHV infected cells can effectively induce the expression of RTA, which promotes the virus entering lytic replication from latency [11, 15, 16].
Since the epigenetic modifications are thought to determine the fate of KSHV infection, it is important to understand the epigenetic status of viral genome during latency and reactivation [17–20]. Previous studies have elegantly depicted the genome-wide histone modifications on KSHV genome in in vitro cell culture systems [21, 22]. It has been demonstrated that activating histone modifications like acetylation of histones (AcH3) are only enriched in several loci while repressive histone modifications like H3K27me3 are widespread across the viral genome, which well explained the expression pattern of viral genes during latency [21, 22]. While most studies are established on the usage of in vitro cultured PEL and KSHV-infected endothelial cell lines, the physiologically relevant scenario in clinical KS samples is unclear.
KS is a highly vascular sarcoma on the skin originated from endothelial cells, which is characterized by the infiltrated inflammatory cells and neo-angiogenesis [8, 23]. According to the geographical distribution and clinical outcomes, KS can be classified into four subtypes, which are classic, endemic, iatrogenic and AIDS-related KS. All the subtypes of KS lesions share a common histological characteristic but are substantially different in disease progression [23–26]. The expression pattern of latent genes in KS is not exactly the same as the one in PEL [27, 28]. A previous study has shown that the expression of vIRF3 gene is only detected in PEL samples, which suggests a distinct epigenetic status in KS [28]. Typically, cells derived from KS tissues will loss episomes very quickly during the in vitro culture, thus it is difficult to obtain cell lines reflecting the physiologically relevant scenario in KS tissues [29, 30]. Therefore, it is important to directly determine the epigenetic landscape of KSHV genome in KS tissues.
In the present study, we established a protocol of ChIP-Seq for clinical KS samples and directly mapped out the epigenetic landscape of KSHV genome in classic KS tissues which are only associated with KSHV infection and derived from Xinjiang area of China. Specifically, we examined AcH3 and H3K27me3 histone modifications on KSHV genome, as well as the genome-wide LANA binding sites. Our results demonstrated that the enriched AcH3 histone modifications were mainly restricted at latent locus while H3K27me3 histone modifications were widespread on KSHV genome in classic KS tissues. The epigenetic landscape at the region of vIRF3 gene confirmed its silenced gene expression in KS tissues. Meanwhile, the abundant enrichment of LANA at the terminal repeat (TR) region was also validated in the classic KS tissues, however, different LANA binding sites were observed on the host genome. Furthermore, the dominant repressive H3K27me3 histone modifications at RTA promoter region were verified by ChIP-qPCR. Intriguingly, we found that the TR region in classic KS tissues was lacking in AcH3 histone modifications with abundant LANA accumulation. Moreover, the established epigenetic landscape in KS tissues was further confirmed in new cases of classic and AIDS-related KS tissues. By analyzing histone modifications and LANA binding sites in classic KS tissues, our study provides new insights for the understanding of KSHV epigentics.
Previous studies on the epigenetic landscape of KSHV genome using in vitro cell culture systems have systemically determined genome-wide distributions of four well-known histone modifications by ChIP-on-ChIP, including AcH3, H3K4me3, H3K27me3 and H3K9me3 [21, 22]. These results supported the predicted activating role of AcH3 and predominant repressive role of H3K27me3 on the KSHV genome. To obtain a physiologically relevant map of the epigenetic landscape of KSHV genome in classic KS tissues, we performed ChIP-Seq experiments on classic KS tissues originated from two different patients. The specific protocol of ChIP in clinical KS tissues was summarized in the Fig 1 and material and methods section. Each experiment was divided into five experimental groups which are input, IgG, LANA, AcH3 and H3K27me3. The purified ChIP product from input, LANA, AcH3 and H3K27me3 groups were subjected to next generation-sequencing. Sequence reads for each sample were aligned to the KSHV genome (HQ404500+35TR) and human genome (Hg19) using Bowtie2 [31]. The results of alignment was presented in Table 1. The overall alignment rate to the KSHV genome was around 0.02%. The relatively high rate in the H3K27me3 group suggested a possible enrichment in KSHV genome. The aligned files were subjected to peak calling and generation of genome-wide maps by using Model-based Analysis of ChIP-Seq (MACS) and Hypergeometric Optimization of Motif EnRichment (Homer) software [32, 33]. The general maps of AcH3 and H3K27 histone modifications on the KSHV genome are illustrated in Fig 2. The enlarged maps are presented in Fig 3 (AcH3) and Fig 4 (H3K27me3). The peak panels illustrated in Figs 2, 3 and 4 demonstrates the most potentially and significantly enriched signals on the KSHV genome. As shown in Fig 2, the enriched AcH3 histone modifications were mainly restricted to the latent locus while H3K27me3 histone modifications were widespread on the KSHV genome. The comparison between AcH3 and H3K27me3 panels showed mutually exclusive signals (high AcH3 level with low H3K27me3 level) on the latent locus and several loci, including the promoter region of vIRF3 gene, regions around ORF8 gene and K5 gene. The coding region of the vIRF3 gene was dominated by the repressive H3K27me3 histone modifications (Figs 3 and 4), which suggests that the vIRF3 gene is silenced but may be easily activated in classic KS tissues. Meanwhile, we also validated that the expression of vIRF3 was restricted at extremely low level in classic KS samples (S1 Fig). This result was in line with previous findings that the vIRF3 expression is not detected in KS tissues by immunohistochemistry analysis [28]. The epigenetic landscape of the KSHV genome in these two cases of classic KS tissues were different from each other in several loci. The first case showed higher level of AcH3 histone modifications on the KSHV genome than the second case in general (Figs 2 and 3), although there was similar trend. To be noted, the promoter and coding regions of the K15 gene were enriched with AcH3 histone modifications only in the first case (Fig 3). In the meantime, we observed a unique peak of H3K27me3 at the promoter region of the LANA gene only in the first case (Fig 4). The higher level of AcH3 and the unique peak of H3K27me3 at the promoter region of the LANA gene in the first case provides supporting evidence for the repressive role of LANA on viral gene expression as previously reported [34–37]. The difference in the overall epigenetic landscape between these two cases indicates different states of KSHV infection in these two patients, which might have clinical relevance to KS progression (S1 File). By comparing with previously published epigenetic maps of the KSHV genome [21, 22], we found that the enriched AcH3 signals at multiple loci (e.g. 10 Kb; 20–30 Kb; 87 Kb) in in vitro cell culture systems (BCBL-1 and KSHV infected SLK) were not observed in classic KS tissues while the landscape of H3K27me3 histone modifications in in vitro cell culture systems was much similar to the one in KS tissues. It has been reported that KSHV may have different latency programs in different tissues or cell lines and the expression pattern of viral genes could be affected by the cytokines present in the local cellular milieu [27, 38]. The difference in AcH3 histone modification of classic KS tissues might indicate a relatively mild environment with less cytokines for classic KS tissues. Files for the generation of genome-wide maps were provided in the S1 Supporting Information section.
LANA protein is critical for the maintenance of KSHV episome [39–41]. The genome-wide LANA binding sites in the in vitro cell culture systems were well described in several studies [42–47]. Yet its footprint is not known in KS tissues which also shows consistent LANA expression. Therefore, it is important to investigate the behavior of LANA in KS tissues. We designed the experimental group of LANA in the ChIP-Seq experiments. The genome-wide LANA binding sites on the KSHV genome in classic KS tissues are illustrated in Fig 5A. The abundant enrichment of LANA at the terminal repeat (TR) region was validated in both of the two cases of classic KS tissues whereas previously reported the enrichments at the latent locus of LANA was only found in the first case, which further confirmed different states of KSHV infection in these two patients. The panels presented several small peaks across the genome, but was not consistent in these two cases, and was difficult to distinguish from the background noise. Although several weak binding sites of LANA were found on KSHV genome in addition to the TR region and latent locus in the previous studies [42–47], the results in classic KS tissues did not show these peaks. This difference may be a result of the reduced sensitivity of ChIP-Seq for the weaker protein binding sites and the smaller size of classic KS tissues.
In the meantime, we also analyzed LANA binding sites on the host genome. The identified peaks of LANA binding by MACS were subjected to Peak Annotation and Visualization (PAVIS) analysis [48]. The PAVIS result illustrated the relative distribution of LANA peaks in relation to genes (Fig 5B). A dramatic reduction in LANA peaks were annotated at 5' UTR region in KS tissues as compared to PEL cell lines by 3–5 fold. By analyzing the relative distance from peaks to TSS (transcription start site), we found a completely different distribution pattern of LANA binding peaks at the promoter regions in KS tissues as compared to previous studies in PEL (Fig 5C). By cross-comparing the identified LANA peaks in these two KS tissues, we found very few overlapped peaks as shown in Fig 5D and S1 Dataset. The representative overlapped peak was illustrated and validated in Fig 5E. Further comparing with LANA binding sites in PEL, KSHV infected SLK and endothelial cell lines, we found almost no common sites in KS tissues (S1 Dataset). The difference in LANA binding sites on the host genome in KS tissues might suggest a different role for LANA in KSHV pathogenesis, but could not rule out the possibility that the results arose from the cellular heterogeneity in KS tissues.
RTA protein encoded by ORF50 is the key switch regulator that controls KSHV reactivation [13, 14]. The epigenetic status in RTA region may reflect the state of KSHV infection [21, 22]. To validate the epigenetic landscape in the classic KS tissues, we carefully examined and verified the histone modifications at the RTA promoter region by ChIP-qPCR. As shown in Fig 6A, the repressive H3K27me3 histone modifications dominated the RTA promoter region (69–71 Kb) in KS tissues of both patients. The results of ChIP-qPCR also confirmed the enrichment of H3K27me3 at the promoter region of RTA and reflected the differences between these two cases (Fig 6B). Moreover, the enrichment of H3K27me3 could be detected at different regions of RTA promoter in the ChIP-qPCR assay (S2 Fig). The GAPDH region exhibited enrichments of AcH3 and little or no H3K27me3 histone modifications, which was the experimental control for the specificity of LANA, AcH3 and H3K27me3 antibodies (Fig 6B). Previous studies in in vitro cell culture systems described a very similar epigenetic landscape at the RTA promoter region as compared to the results in KS tissues [21, 22], implicating that the conclusion from in vitro studies about RTA regulation could well support and apply to the in vivo scenario.
The TR region of KSHV genome consists of highly repeated sequences of 801 bp with multiple copies which can range over 20 Kb [4]. Previous studies have proved abundant AcH3 histone modifications with LANA accumulation at the TR region in in vitro cell culture systems [21, 22, 49]. However, we found that the TR region in classic KS tissues was lacking in AcH3 histone modifications with abundant LANA accumulation (Figs 3 and 5A). To analyze the epigenetic landscape at the TR region without consideration of sequence repetition, we re-aligned the sequence reads for each sample to the TR sequence using Bowtie2. The reanalyzed epigenetic landscape at the TR region did not change as illustrated in Fig 7A. To verify the results of the ChIP-Seq, we examined the TR region with the same samples by ChIP-qPCR. As shown in Fig 7B, the enrichment of LANA binding and absence of AcH3 histone modifications were confirmed by ChIP-qPCR. To validate the distinct results in KS tissues, we performed the ChIP experiments in Doxycycline inducible recombinant KSHV.219 harboring SLK (iSLK.219) and body-cavity-based lymphoma (BCBL-1 and BC3) cell lines using the same protocol. The results were shown in Fig 7C. Very strongly enriched signals of LANA binding and AcH3 histone modifications were observed at the TR region as previously reported. The hyperacetylation of histone H3 at TR region was presumably thought to be involved in the assembly of DNA replication factors, yet the significance remained unknown [49]. TR region contains the latent replication origin of KSHV genome, thus hypoacetylation of histone H3 at the TR region might affect the latent replication of KSHV genome, hampering the maintenance of KSHV episomes [40, 50, 51]. This needs further investigation to determine whether losing episomes during the process of in vitro culture of cells derived from KS tissues was related to the absence of AcH3 histone modifications at TR region. The hyperacetylation of histone H3 can introduce a loosened chromatin structure at TR region, which may facilitate a poised chromatin structure at the long unique region for the topological speculation. However, the hypoacetylation of histone H3 at the TR region in classic KS tissues was shown to correlate a silenced state of viral genome, thus the acetylation of histone at the TR region may not be related to KSHV gene expression according to the results.
Since KS tumor cells are originated from endothelial cells, we also examined histone modifications and LANA binding sites of KSHV genome in KSHV infected lymphatic endothelial cells (LEC.KSHV). However, the result in LEC.KSHV was also different from the established results in KS tissues. The strong enrichment of LANA binding and AcH3 histone modifications could be observed at the TR region as the same with other in vitro cultured cell lines (S3 Fig).
To further confirm the established epigenetic landscape in KS tissues, we examined the epigenetic histone modifications of KSHV genome in new cases of KS tissues (two classic and one AIDS-related). As shown in Fig 8A and 8B, the results of ChIP-qPCR in new cases of classic KS tissue kept good consistency with the previously examined two cases. The established epigenetic landscape at multiple sites were validated, including the TR, RTA promoter, miR-cluster and vIRF3 regions. The general maps of these two cases are illustrated in S4 Fig. Since other subtypes of KS share a common histological characteristic [23], we wondered whether they would have a similar epigenetic landscape as the classic KS tissues. We additionally determined the epigenetic histone modifications of KSHV genome in one case of AIDS-related KS tissue by ChIP-qPCR. Intriguingly, we found the results in AIDS-related KS tissue are similar to the established one in classic KS tissues, but more enrichment of AcH3 histone modifications were observed at the TR and vIRF3 coding region (Fig 9). The difference between classic and AIDS-related samples made us speculate that acetylation of histone H3 at the TR and vIRF3 coding region might correspond to the progression of KS disease. The lesions in classic KS cases are generally localized at the extremities with slow or limited progression while the lesions in AIDS-related KS cases usually spread to the whole body and lead to significant mortality [52–56]. The speculation of a relationship between histone acetylation at the TR and vIRF3 coding region and the progression of KS disease will need a larger subset of cases to be examined to support this hypothesis although the data so far is highly suggestive.
By analyzing histone modifications and LANA binding sites in classic KS tissues, our study established the epigenetic landscape of KSHV genome in clinical KS samples for the first time. The established epigenetic landscape of KSHV genome in classic KS tissues provided direct evidence to support distinct latent programs in KS tissues. A similar epigenetic landscape was observed at the RTA promoter region in KS tissues as compared to the results from in vitro cell culture systems, which supported the physiological significance of in vitro studies regarding RTA regulation. The distinct AcH3 histone modifications at the TR and vIRF3 coding regions in classic KS tissues provided important clues about the progression of KS disease, which would be a helpful reference for doctors to diagnose clinical patients using epigenetic targeted strategies. We have analyzed the epigenetic landscape of the KSHV genome in classic KS tissues and demonstrated similarities and differences which provide new insights towards understanding KSHV epigenetics, which is important for future studies on the mechanism of KSHV infection and pathogenesis.
Experiments in the present study were conducted according to the principles in the Declaration of Helsinki. The usage of clinical Kaposi's sarcoma (KS) tissues was reviewed and ethically approved by the Institutional Ethics Committee of the First Teaching Hospital of Xinjiang Medical University (Urumqi, Xinjiang, China; Study protocol no.20082012). Written informed consent was obtained from all participants, and all samples were anonymized.
The clinical KS tissues were collected from four patients who had received a pathological diagnosis of KS, including three classic KS and one AIDS-related KS. The patients were of Uygur and Kazak ethnicities from the local region. All samples were collected from Xinjiang province, northwestern China. Details about the patients/specimen were described in the S1 File.
Body-cavity-based lymphoma (BCBL-1) cell line was derived from KSHV positive primary effusion lymphoma patients [57]. BCBL-1 and BC3 was maintained in RPMI 1640 medium (Hyclone) containing 10% fetal bovine serum (FBS) and 5% antibiotics (penicillin and streptomycin, Hyclone). Doxycycline inducible recombinant KSHV.219 harboring SLK (iSLK.219) cell line was established by J. Myoung and D. Ganem [58], and was kindly provided by Fanxiu Zhu (Florida State University). iSLK.219 cell line was cultured in DMEM (Hyclone) supplemented with 10% FBS (Hyclone) and 5% antibiotics (penicillin and streptomycin, Hyclone). Lymphatic endothelial cells (LEC) were purchased from PromoCell (C-12216) and cultured with Endothelial Cell Growth Medium MV2 kit (C-22121, PromoCell).
The collected fresh clinical KS samples (0.1–0.2 g) were stored at -80℃ before usage. The protocol of ChIP-Seq in clinical KS samples is described below:
(Optional) Enhanced Cross-link: Add formaldehyde to a final concentration of 1.5%. Incubate at room temperature for 3 min. Quench the cross-linking by adding 0.15 mL glycine (1.25 M). Incubate at room temperature for 10 min. Centrifuge at 200 × g for 5 min to remove the supernatants and wash the pellet with PBS twice.
*Note the pretreated beads should not be blocked with sperm DNA.
Antibodies in the ChIP-Seq experiments: Anti-Trimethyl-histone H3 (Lys27) (H3K27me3) rabbit polyclonal antibody (07–449) was purchased from Merck Millipore. Anti-acetyl-histone H3 (AcH3) rabbit polyclonal antibody (06–599) was purchased from Merck Millipore. Anti-LANA mouse monoclonal antibody produced by 1B5 hybridoma was made in our laboratory (Antigen source for immunization: LANA 900-1162aa) [45].
The ChIP-Seq data (data quality parameters were described in the S2 File) were aligned to human genome (hg19) and KSHV genome (HQ404500 plus 35 copies of TR [U75699.1]) using Bowtie2 [31]; only one mismatch was allowed. The output files were subjected to peak calling and generation of genome-wide maps with MACS (Model-based Analysis of ChIP-Seq) and Homer2, as described previously [32, 33]. The Input group was used as control. For the analysis of histone modifications with MACS, the parameters were set according to the protocol as followed:—nomodel,—shiftsize = 73. The default P value cutoff for the peak detection was 10−5.. For the analysis of histone modifications with Homer2, the adjusted parameters were set as followed: -style = histone, -size = 100 or 150, -minDist = 300. The default P value cutoff for the peak detection was 10−4. The final result of identified peaks was generated by the combination of MACS and Homer2 analysis. For the analysis of LANA binding sites with MACS, the parameters were set as previously reported:—nomodel,—shiftsize = 50. The Input group was used as control. The P value cutoff for the peak detection was 10−3. Results were visualized by IGV software [59]. Normalization factors were calculated according to the depth of sequencing and formulated as followed: Normalization Factor = Total reads of sample / Total reads of Input. Values on the y axis of each panel in IGV software was adjusted according to the calculated normalization factors (S3 File). The peak information was annotated with Peak Analyzer. The distribution of peaks in relation to genes was calculated by PAVIS [48].
Real-time RT-PCR was performed with a SYBR green Master Mix kit (Toyobo). Reaction mixtures contained 5 μl Master Mix plus Rox, 1 μM each primer, and 4 μl diluted sample. All primers are listed below:
The program set on the 7900HT sequence detection system (Life Technologies) was 95℃ for 5 min, followed by 40 cycles at 95℃ for 15 s, 58℃ for 20 s and 72℃ for 30 s. Melting curve analysis was performed to verify the specificity of the products and each sample was tested in triplicate.
The original data have been submitted to SRA (Sequence Read Archive) in NCBI website. The accession number of this project is SRP081036. KSHV genome: HQ404500. TR: NC_009333, 137169–137969.
|
10.1371/journal.pntd.0001315 | Iquitos Virus: A Novel Reassortant Orthobunyavirus Associated with Human Illness in Peru | Oropouche (ORO) virus, a member of the Simbu serogroup, is one of the few human pathogens in the Orthobunyavirus genus in the family Bunyaviridae. Genetic analyses of ORO-like strains from Iquitos, Peru, identified a novel reassortant containing the S and L segments of ORO virus and the M segment of a novel Simbu serogroup virus. This new pathogen, which we named Iquitos (IQT) virus, was first isolated during 1999 from a febrile patient in Iquitos, an Amazonian city in Peru. Subsequently, the virus was identified as the cause of outbreaks of “Oropouche fever” during 2005 and 2006 in Iquitos. In addition to the identification of 17 isolates of IQT virus between 1999 and 2006, surveys for neutralizing antibody among Iquitos residents revealed prevalence rates of 14.9% for ORO virus and 15.4% for IQT virus. Limited studies indicate that prior infection with ORO virus does not seem to protect against disease caused with the IQT virus infection. Identification of a new Orthobunyavirus human pathogen in the Amazon region of Peru highlights the need for strengthening surveillance activities and laboratory capabilities, and investigating the emergence of new pathogens in tropical regions of South America.
| Oropouche (ORO) virus is one of the few human pathogens in the Orthobunyavirus genus in the family Bunyaviridae. Phylogenetic analyses of ORO-like strains isolated from febrile patients in Iquitos, Peru, identified a novel ORO reassortant virus, which we named Iquitos (IQT) virus based on the location of the isolation of the virus. This novel pathogen was first isolated during 1999 from a 13-year-old boy who had an illness that included symptoms of fever, headache, eye pain, body pain, arthralgias, diarrhea, and chills. Subsequently, the virus was identified as the cause of outbreaks of “Oropouche fever” during 2005 and 2006 in Iquitos. Limited serological studies indicate that prior infection with ORO virus does not seem to protect against disease caused with the IQT virus infection. In summary, we identified a new Orthobunyavirus human pathogen in the Amazon region of Peru; these results highlight the need for strengthening surveillance activities and investigating the emergence of new pathogens in tropical regions of South America.
| Viruses in the family Bunyaviridae are classified into five genera: Orthobunyavirus, Hantavirus, Nairovirus, Phlebovirus and Tospoviruses. The orthobunyaviruses are enveloped, negative sense RNA viruses whose genome comprises three segments: small (S), medium (M) and large (L). The S segment encodes for the nucleocapsid and a nonstructural protein, NSs. The M segment encodes for the glycoproteins Gn and Gc, whereas the L segment encodes the viral polymerase. Some members of the Orthobunyavirus genus are known to cause clinical disease in humans, including Oropouche (ORO) virus, a member of the Simbu serogroup, which causes a febrile disease often associated with headache, dizziness, weakness, myalgias and arthralgias [1]. Bunyamwera virus, which is considered the prototype member of the family, causes a febrile illness with headache, arthralgias, rash and infrequent central nervous system involvement [2]. Hemorrhagic manifestations associated with some orthobunyavirus infections have also been reported recently [3], [4].
Genetic reassortment among members of the same serogroup within the Orthobunyavirus genus occurs in nature and has led to the emergence of new viruses, occasionally with increased pathogenicity. This appears to be the case with Ngari virus, which has been associated with hemorrhagic fever in Kenya and Somalia [3], [5]. This virus is comprised of the S and L segments of Bunyamwera virus and the M segment of Batai virus [4]. On the basis of genetic and antigenic analyses, we previously reported that Jatobal virus (JAT), a member of the Simbu serogroup, is a reassortant virus that contains the S segment of ORO virus and the M and L segments of a still unrecognized Simbu serogroup virus [6].
Within the Orthobunyavirus genus, 18 serogroups have been recognized on the basis of the results of cross-hemagglutination inhibition (HAI) and antibody neutralization relationships [7]. The Simbu serogroup contains at least 25 members, and recent phylogenetic analyses demonstrated that the genetic relationships amongst these viruses are consistent with the results of serological relationships [7], [8]. ORO virus was originally isolated in 1955 from blood of a febrile forest worker who lived in Vega de Oropouche, Trinidad [9]. Outbreaks involving thousands of human cases continue to be reported in Brazil [10], [11], [12], [13], and circulation of the virus has been reported in Panama, Peru, and Trinidad [14], [15], [16], [17]. Based on the S segment, three genotypes of ORO virus were distinguished phylogenetically: genotype I includes the isolates from Brazil and Trinidad, genotype II includes isolates from Brazil and Peru, and genotype III is represented by the isolates from Brazil and Panama [18], [19].
ORO virus has been isolated from mosquitoes (Coquillettidia venezuelensis in Trinidad [9]; Ochlerotatus serratus and Culex quinquefasciatus in Brazil [20]) and frequently from the midge Culicoides paraensis [9], [21], [22]. High population densities of this midge have been found during epidemics of ORO virus [23]. Transmission of the virus has been demonstrated via C. paraensis, from infected to susceptible hamsters, and from infected humans to susceptible hamsters [20], [24]. ORO virus has also been isolated from sloths (Bradypus tridactylus [20]), and from a monkey (Callithrix sp.) in Brazil [10].
In 1995, the U.S. Naval Medical Research Unit Six (NAMRU-6) in Lima, in collaboration with the Ministry of Health of Peru, initiated a passive surveillance study to investigate etiology of febrile illnesses. As part of these surveillance activities, several ORO-like strains were obtained from febrile patients in Iquitos, Peru. In this study, we describe the identification of a novel reassortant virus which we named Iquitos (IQT) based on the location of the isolation of the virus. Specifically, we: 1) demonstrated that IQT virus causes clinical disease in humans similar to ORO fever, 2) described the genetic relationship of IQT virus to other members of the Simbu serogroup, 3) identified the risk factors associated with human infection by IQT virus in an urban setting of the Amazon region of Peru, and 4) describe the clinical manifestations associated with infection. Significantly, immunity to ORO virus does not appear to protect against infection by IQT virus, and both ORO and IQT viruses appear to have similar antibody prevalence in Iquitos over the past 10 years.
The IQT1690 ORO strain used in this study was isolated in 1995 from a 50 year-old male resident of Iquitos, Peru. The first strain of IQT virus (IQT9924) was isolated in Iquitos from a 13 year-old boy (Table 1). The origin of the ORO strains BeH379693, BeH544552, MD023, GML444477 and GML 445252 were previously described [18], [19].
Iquitos is a city of about 380,000 inhabitants located 120 meters above sea level in the Amazon Basin in northeastern Peru.
The human use study protocols were approved by the Ministries of Health of Peru and by the NAMRU-6 Institutional Review Board (protocol NMRCD.2000.0006). The study subjects were patients (≥ 5 years of age) who presented with a diagnosis of an acute, febrile undifferentiated illness at military or civilian outpatient clinics in Iquitos. The criteria for inclusion in the program was fever ≥ 38°C of no more than five days duration, headache, myalgia and other nonspecific symptoms. Demographic and clinical information were obtained from each patient at the time of voluntary enrollment and a signed consent form was obtained from patients 18 years of age and older. In addition, written assent was obtained from patients between 8 and 17 years of age. For patients younger than 18 years, written consent was obtained from parents or legal guardians. Paired-blood samples were collected, one during the acute phase of illness and a second sample 2-4 weeks after onset of symptoms.
Acute serum samples were tested for virus by cell culture, and both acute and convalescent serum samples were assayed for IgM antibody to a variety of arboviruses by an enzyme-linked immunosorbent assay (ELISA), as described previously [25].
For virus isolation attempts, serum samples were diluted 1∶5 in Eagle's minimum essential medium (EMEM) supplemented with 2% fetal bovine serum, 200 µg of streptomycin, and 200U/ml of penicillin. Two hundred µl of the diluted samples were inoculated into flasks containing either confluent monolayers of African green monkey kidney (Vero) cells or Aedes albopictus mosquito (C6/36) cells. Vero and C6/36 cell cultures were maintained at 37°C and 28°C, respectively, and were examined daily for 10 days for evidence of viral cytopathic effects (CPE). Spot-slides of C6/36 and Vero cells were subsequently prepared on days 10 post-inoculation of the samples (or sooner if CPE developed) and an immunofluorescence assay (IFA) was performed using polyclonal antibody against arboviruses endemic to Peru, including ORO virus [17], [25], [26], [27], [28], [29], [30].
A total of 1037 human serum samples collected in Iquitos during 2006 were tested for IgG antibody to the ORO strain IQT1690 and IQT9924 virus using a previously described ELISA [25]. The samples were collected as part of a cross-sectional antibody prevalence study conducted in Iquitos after an outbreak of febrile illness associated with Venezuelan equine encephalitis virus (VEEV) infection (protocol PJT.NMRCD.001). Three neighborhoods where VEE cases were reported and a control neighborhood where VEE cases were not reported were included in the study [31]. Serum samples from a subset of the original study participants who agreed to future use of their samples were tested by ELISA IgG antibody. ELISA positive samples were further tested using an 80% plaque reduction neutralization assay (PRNT) each for the ORO strain IQT 1690 and IQT9924 virus. Briefly, sera were heat-inactivated at 56°C for 30 minutes and 2-fold serum dilutions were prepared and mixed with 100 plaque forming units (PFUs) of each virus and incubated at 4°C overnight. The virus-serum dilutions mixtures were inoculated into confluent monolayer of Vero cells propagated in microplates and incubated at 37°C for 1 hour before adding an overlay of 0.4% of agarose in EMEM. After 72 hours of incubation at 37°C, the plates were stained with 0.25% crystal violet in 20% methanol and plaques were counted. All IgG antibody positive samples were tested at an initial concentration of 1∶20 and all positive samples were further titrated to determine endpoint titers. Neutralization titers were considered as the highest serum dilution that reduced plaque formation by ≥ 80%.
Viral RNA was extracted using the QIAamp viral RNA mini kit (Qiagen, Valencia, CA) or Trizol reagent (Invitrogen, Carlsbad, CA) following the manufacturer's protocols. The reverse transcription reaction (RT) was done using 1X RT buffer, 0.2 mM dNTPs, 1 µM of primers, 80 units of RNAsin ribonuclease inhibitor (Promega, Madison, WI), 1mM of dithiothreitol, 200U of SuperScript reverse transcriptase (Invitrogen) and 5 µl of RNA. The reactions were incubated at 42°C for 1 hour. The PCR included 1X PCR buffer, 0.25 mM dNTPs, 1 µM of primers, 3 mM of MgCl2, 2.5 units of GoTaq DNA polymerase (Promega, Madison, WI) and 5 µl of cDNA. The conditions for the PCRs included incubation at 94°C for 2 minutes, 35 cycles of 94°C for 30 seconds, 50°C for 1 minute, 72°C for 1.5 minutes and a final extension of 72°C for 10 minutes to ensure complete double-stranded DNA synthesis. The primers used for the PCR amplification have been previously described and included ORO N3 (GTGAATTCCCACTATATGCCAATTCCGAATT) and ORO N5 (AAAGAGGATCCAATAATGTCAGAGTTCATTT) that amplifies the S segment, M14C (CGG AAT TCA GTA GTG TAC TACC) and M619R (GAC ATA TG (CT) TGA TTG AAG CAA GCA TG) that amplifies the M segment and M13CBunL1C (TGTAAAACGACGGCCAGTAGTGTACTCCT), and BunL605R (AGTGAAGTCICCATGTGC) that amplifies the L segment [3], [10], [19].
Partial sequences of the S, M, and L segments were obtained and compared to published ORO virus sequences. Purified PCR products were sequenced directly and sequencing analyses of the PCR products were performed using an Applied Biosystems Prism automated DNA sequencing kit (Foster City, CA) according to the manufacturer’s’ protocols. Sequences were aligned using the Clustal program in the Mac Vector software package (MacVector Inc., Cary, NC) and phylogenetic analyses were performed using the maximum parsimony, neighbor joining, and maximum likelihood methods implemented in the PAUP software [32], [33]. For the neighbor joining analyses, the HKY85 distance was used. Bootstrap values, to place confidence values on grouping within trees, were calculated based on 500–1000 replicates.
To prepare ORO virus hyperimmune ascitic fluid for classical cross-neutralization tests, mice received 4 weekly intraperitoneal (IP) injections of virus-infected newborn mouse brain suspension with Freud's adjuvant. To investigate antigenic differences among the viruses, cross-neutralization assays were then performed, using a previously described PRNT [34]. To further evaluate the antigenic differences between the ORO strain IQT1690 and IQT9924 virus, acute and convalescent sera from patients infected with these viruses were tested against both strains using PRNT.
Three to 4 week-old golden Syrian female hamsters were inoculated (IP) with serial 10-fold dilutions of virus. The LD50 was calculated by the Reed and Muench method [35].
To determine viremia levels in infected patients, serum samples were prepared in 10-fold dilutions and 100 µl of each dilution was inoculated onto confluent monolayers of Vero cells in 12 well-plates. Viruses were adsorbed to the monolayers for 1 hour at 37°C. A 3-ml overlay consisting of EMEM with 0.4% agarose was added, and the cells were incubated at 37°C for 72 hours. Agar plugs were removed, and the cells were stained with 0.25% crystal violet in 20% methanol. The sensitivity of the assay corresponded to a detection limit of 100 PFU/ml.
The analysis was performed using SPSS 17.0 for Windows (SPSS Inc, Chicago, IL, USA). The proportions of positive results (PRNT for IQT1690 and IQT9924 antibody ≥ 20) were calculated with their respective 95% confidence intervals (95% CI). The proportions were compared using the Pearson Chi-Square and Fisher exact tests. A 2-tailed p-values < 0.05 was used for all statistical analyses. Multivariate analysis was performed using logistic regression (enter method) adjusting by gender, age (adult, child), occupation, type of house, contact with domestic animals, travel history and neighborhoods.
In 1999, IQT virus strain IQT9924 was isolated in Iquitos, Peru, from a 13-year-old boy who had an illness that included symptoms of fever, headache, eye pain, body pain, arthralgias, diarrhea, and chills. These clinical symptoms were typical of ORO virus infections in Peru [17]. The strain was provisionally identified as ORO virus based on results of serological reactivity in an indirect immunofluorescence test with virus-infected cells.
Sixteen additional IQT9924 virus isolates (provisionally identified as ORO virus) were obtained from febrile patients living in Iquitos: 1 in 2003, 7 in 2005, and 8 in 2006. Of the 17 IQT 9924 virus positive patients, detailed clinical information was available for 16 (Table 2). The most common general symptoms were: chills (16), headache (15), arthralgia or loss of joint function (15), general malaise (14), diminished appetite (13), myalgias (13), retro-orbital pain (11), bone pain (9), and pallor (7). Gastrointestinal symptoms were very common, and 12 patients had at least one of the following symptoms: nausea (10), abdominal pain (7), vomiting (4), or diarrhea (2). No patients had jaundice, hepatomegaly, splenomegaly, abdominal distension, or ascites. Six patients had at least one respiratory symptom: cough (6), rhinorrhea (3), pharyngeal congestion (2), or expectoration (1); no patient had dyspnea. Only one patient complained of a rash, and this was described as erythematous and affecting both the central trunk and distal extremities. With the exception of one patient who had petechiae, none of the patients had hemorrhagic manifestations, including epistaxis, bleeding gums, melena, hematochezia, vaginal bleeding, petechiae, purpura, and ecchymosis. All patients had a fever as this was one of the inclusion criteria.
Based on virus isolation and serological analyses, the peak incidence for ORO virus was reported in Iquitos as occurring in 2004, 2005 and 2006 [30]. Because the circulation of IQT9924 virus only was detected serologically and genetically (see below) during that time period, we believe that the study most likely reported the incidence of IQT9924 virus.
Genetic analyses of the S, M, and L segments were conducted to determine the relationships of the isolates to ORO virus and other members of the Simbu serogroup. Phylogenetic trees were constructed for both the S- and M-RNA segments based on members of the Simbu serogroup that included an example from each of the three known ORO genotypes [19]. The phylogenetic tree based on the S segment placed the S-RNA of IQT9924 virus among isolates of ORO virus genotype II (Fig. 1), while the M-RNA phylogenetic tree (Fig. 2) indicated that the strain IQT9924 had an M-RNA that was unique from any other Simbu serogroup virus identified to date, but was most closely related to ORO virus. Significantly, IQT9924 virus had an M-RNA distinct from JAT virus, a reassortant with a non-ORO virus M-RNA [6]. The L-RNA phylogenetic tree showed that IQT9924 virus had an L-RNA of ORO virus (data not shown). The M segment fragment sequence (∼590 nucleotides) of IQT9924 displayed only 68% nucleotide and 65% amino acid identity to the prototype ORO strains BeAn 19991 (Brazil) and Tr9760 (Trinidad) whereas the S and L segment fragment sequences exhibited 95% nucleotide and 96% amino acid identity to ORO virus. Based on partial sequences of the S, M and L segments, all ORO-like viruses isolated from patients living in Iquitos after 1999 had similar genetic characteristics and would appear to be examples of IQT9924 virus. Genetic analyses of ORO virus isolates obtained in Peru before 1999, did not reveal the circulation of IQT9924 virus prior to 1999 [19].
We investigated the serological relationships of IQT9924 virus using cross-neutralization tests. Mouse antisera were prepared against ORO virus strain Trinidad55, IQT9924 virus, and JAT virus. These antisera displayed a 4-fold or greater difference in neutralization titer among these viruses (Table 3), indicating that IQT9924 virus is serologically distinct from ORO and JAT viruses.
Next, convalescent sera from patients in Iquitos who had been infected with either OROV or IQT9924 virus were assayed by PRNT to confirm the antigenic differences. A 4-fold or greater difference in neutralization titer was observed between these viruses (Table 4) corroborating our initial findings with mouse antisera. Of particular importance, neutralizing antibody to ORO virus were detected in the acute serum samples of 2 febrile patients from whom the IQT9924 virus had been isolated (Table 5). Furthermore, a boost in ORO neutralizing antibody titers was observed in both patients after infection with IQT9924 (Table 5). Overall, the results suggest that prior ORO virus infection does not protect against febrile disease caused by the new reassortant virus (IQT 9924) and is consistent with the cross-neutralization data indicating that IQT9924 and ORO are distinct viruses, and that there is limited cross-protective immunity such that individuals can be infected and have clinical diseases caused by both viruses.
The hamster is used as an animal model of ORO virus infection [36]. Thus, we evaluated the hamster virulence phenotype of IQT9924 virus and compared the results to those obtained with representatives of the three ORO genotypes. The IQT 9924 virus was found to be poorly virulent with a LD50 of 4220 PFU and was similar to two Panamanian strains (GML 444477 and GML445252) belonging to ORO genotype III that were non-lethal at the highest doses inoculated (>2,500 and >2,000 PFU, respectively) (Table 6). In comparison, ORO genotype II strain MD023 from Madre de Dios, Peru, was virulent with LD50 of 45 PFU. Finally, two genotype I strains, BeH379693 and BeH544552 from Brazil, displayed different virulence phenotypes (>7,000 PFU vs 1 PFU, respectively).
Since Simbu serogroup viruses are arboviruses, we undertook preliminary studies with a limited number of serum samples to determine viremia titers among ORO and IQT9924 virus-infected patients. Serum infectivity titers were obtained from a total of 19 patients. Two ORO virus-infected patients had viremias of 6×103 and 7×105 PFU/ml whereas 17 patients infected with IQT9924 virus had levels that were below the limit of detection of the assay (<100) to 1.8×105 PFU/ml (Table 1).
Examination of 1037 human serum samples from acute febrile infections in Iquitos revealed that the overall neutralizing antibody prevalence to ORO virus was 14.9% (154/1037) (95% CI 12.8–17.1) whereas prevalence to IQT9924 virus was 15.4% (160/1037) (95% CI 13.3–17.7). Only 3.4% (35/1037) of the serum samples had neutralizing antibodies (>20) to both viruses. Neutralizing antibody prevalence to both ORO and IQT9924 viruses was higher among females than males (17.3% vs 10.2% and 17.5% vs 11.3%, respectively; p<0.05).
The neutralizing antibody prevalence to ORO virus in persons living in the neighborhoods of San Juan (24.1%) and Bellavista Nanay (16.5%) was higher compared to other neighborhoods (Fig. 3). Similarly, the PRNT antibody prevalence to IQT9924 virus was also higher in the neighborhoods of San Juan (33.9%) and Bellavista Nanay (12.2%). In San Juan, PRNT antibody prevalence to IQT9924 virus was significantly higher than to ORO virus (p<0.05). Neutralizing antibody prevalence to both viruses in the study population increased with age after adulthood (0% in 5–9 years old to 32.3% in 40-49 years old for ORO virus and 3.6% in 5–9 years old to 25% in >70 years old for IQT9924 virus) (Table 7). The antibody prevalence to ORO and IQT9924 viruses in adults was 20.1% and 18.8%, respectively, compared to 2.8% and 7.9% in the younger group (<20 years old). The univariate analysis did not detect an association between ORO or IQT9924 virus antibody prevalence and occupation, type of housing, travel or contact with chickens or rodents.
Logistic regression models were used to test the significance of each factor. The variables predictive of ORO virus neutralizing antibody prevalence in the model were gender, age and neighborhood, whereas factors such as gender, age, neighborhood, and contact with pigs were predictive of IQT9924 virus antibody prevalence.
The first confirmed ORO fever cases in Peru were reported in 1992 involving a small outbreak of eight febrile cases living in Iquitos [37]. Subsequently, only sporadic cases of ORO fever were reported in Peru, mainly in the Amazon region. The situation in Peru differs from Brazil, where ORO fever outbreaks are usually associated with hundreds of human cases [11], [18], [38]. The reasons for this apparent difference in ORO transmission rates remain unknown. To date, three ORO genotypes have been described based on the nucleoprotein gene (S segment): genotype I among viruses circulating in Brazil and Trinidad, genotype II among viruses from Brazil and Peru, and genotype III among viruses from Brazil and Panama [10], [18], [19]. In this study, we sought to genetically characterize strains from Peru that were provisionally identified as ORO virus. Sequence analyses based on the S and L segments placed the strain IQT9924, isolated in Iquitos in 1999, within ORO genotype II while phylogenetic analyses of the M segment revealed that IQT9924 virus contained a M-RNA segment of a still unidentified Simbu-serogroup virus. Thus, IQT9924 was identified as a Simbu serogroup reassortant virus. Serological characterization of IQT9924 virus confirmed that the virus was distinct from ORO virus. More importantly, we obtained serological evidence that prior ORO virus infection does not protect against clinical disease caused by this new reassortant virus, providing additional confirmation that ORO and IQT9924 are two distinct viral entities. Because most arbovirus laboratories in South America identify ORO virus without detailed serological or genetic characterization, it is uncertain whether IQT9924 circulates in other South American countries. Further genetic and antigenic characterization of ORO isolates from South America is necessary to fully determine the geographic distribution of this newly identified human pathogen.
Studies conducted in Brazil reported viremia titers higher than 3log10 suckling mice LD50 (SMLD50)/ml among ORO- infected patients, including approximately 10% of the patients who developed viremia titers ranging from 5.0 to 5.3 log10 SMLD50/ml during the first two days of illness [20]. These levels of viremia were high enough to infect C. paraensis and transmission was demonstrated to susceptible hosts. Consequently, it has been postulated that humans are the primary amplifying host during ORO fever epidemics [20], [24]. In contrast, viremia levels lower than 5.3log10 SMLD50/ml were not sufficient to infect C. paraensis [24]. In this study, we measured viremia by plaque assay (instead of SMLD50) and found that the average viremia levels detected among ORO and IQT9924 virus-infected patients were 4.8±0.7 log10 PFU/ml and 3.4±1.1 log10 PFU/ml, respectively with at least one ORO-infected patient developing a viremia titer of 5.9 log10 PFU/ml (Table 1).
It is not known if C. paraensis is the vector of IQT9924 virus. However, studies conducted in and around Iquitos during 1996–1997 identified C. paraensis as the most common biting midge of all host-seeking ceratopogonids at 16 sites and C. insinuatus being the second most common biting midge [39], [40]. Subsequent studies conducted in 2001, 2002, and 2003 showed that the peaks in biting activities in the Punchana district, located near Iquitos, occurred between October and December for C. paraensis. Likewise, peaks in biting activities were observed between October and April for C. insinuatus in Santa Clara, near Iquitos [40]. It has not been possible to determine peak incidence rates of ORO fever due to lack of identification of cases. In contrast, data from our febrile surveillance study suggest that the number of IQT9924 virus-infected patients peak from December to April indicating a possible overlap with the biting activities of C. insinuatus [30]. Felippe-Bauer et al [41] recently identified two new morphological species in the C. paraensis complex from the Department of Amazonas and Loreto, Peru. Thus, future studies should investigate the role of this midge in ORO and IQT9924 virus transmission.
In Iquitos, the overall neutralizing antibody prevalence for ORO virus was 14.9% whereas prevalence for IQT9924 virus was 15.4%. These numbers are lower than those from previous studies that reported antibody prevalence to ORO virus (based mostly on IgG antibodies) as high as 35% in certain neighborhoods near Iquitos [14], [15]. Thus, it would appear that ORO virus prevalence varies among neighborhoods and this was demonstrated in our study where neighborhoods, such as San Juan and Bellavista Nanay, have higher ORO virus prevalence rates. Antibody prevalence to ORO and IQT9924 viruses was higher among females and is consistent with previous studies on ORO virus prevalence [15], [42]. Another study carried out in Brazil after an ORO virus outbreak also detected higher antibody prevalence among women [43]; however, a subsequent study in Brazil failed to detect gender differences in attack rates [20].
Our clinical data indicate that the IQT9924 virus shares many of the same clinical manifestations as ORO virus. Headache and chills affected the vast majority of IQT9924-infected patients, similar to previous studies of ORO virus infection [18], [42]. Bone, muscle, and joint pain also were observed commonly for both viruses. Rash and hemorrhagic manifestations have not been classically described with ORO virus infection and were only found in one of 16 IQT9924-infected patients in this study. Interestingly, respiratory complaints—primarily cough—were found in 38% of our IQT9924-infected patients, a finding not commonly reported with ORO virus infection.
Among the Orthobunyavirus genus, studies have shown that the pathogenicity of these viruses is multigenic with the M segment being a major determinant [44]. Thus, it is very likely that the donor of IQT9924 virus M segment may also cause human illness in the Amazon region of Peru. It is worth noting that this new virus was first isolated in 1999 and, subsequently in 2005 and 2006, when it was the cause of outbreaks of febrile illness in Iquitos. Some of the patients infected with the IQT9924 virus in 2005 and 2006 resided within Quistococha and Zungarococha, which are considered rural areas near Iquitos. Thus, it is possible that a spill-over of cases occurred from rural to urban areas as was suggested during the 2005-2006 VEEV outbreaks in Iquitos [31]. Unusual high annual river levels occurred in early 2006, which may had an impact on arthropod density and geographic distribution leading to the observed VEEV and IQT9924 virus outbreaks. Finally, it appears that this IQT9924 virus has emerged and possibly replaced ORO virus in Iquitos because ORO virus cases have not been identified in the area since 1999.
In summary, this study identified a new pathogen, IQT9924 a reassortant of ORO virus that was associated with febrile illness in the Amazon region of Peru. We propose the name Iquitos virus for this newly identified Orthobunyavirus member of the Simbu-serogroup because there have been cases of disease caused by this virus in the Iquitos area over several years. While reassortment among members of the same serogroup of the Bunyaviridae has been identified, few reassortants have been associated with human disease. It will be important to determine the geographic distribution of Iquitos virus and to evaluate its potential as a major public health problem as ORO virus has been done in Brazil.
|
10.1371/journal.pmed.1002356 | Ultrasound non-invasive measurement of intracranial pressure in neurointensive care: A prospective observational study | The invasive nature of the current methods for monitoring of intracranial pressure (ICP) has prevented their use in many clinical situations. Several attempts have been made to develop methods to monitor ICP non-invasively. The aim of this study is to assess the relationship between ultrasound-based non-invasive ICP (nICP) and invasive ICP measurement in neurocritical care patients.
This was a prospective, single-cohort observational study of patients admitted to a tertiary neurocritical care unit. Patients with brain injury requiring invasive ICP monitoring were considered for inclusion. nICP was assessed using optic nerve sheath diameter (ONSD), venous transcranial Doppler (vTCD) of straight sinus systolic flow velocity (FVsv), and methods derived from arterial transcranial Doppler (aTCD) on the middle cerebral artery (MCA): MCA pulsatility index (PIa) and an estimator based on diastolic flow velocity (FVd). A total of 445 ultrasound examinations from 64 patients performed from 1 January to 1 November 2016 were included. The median age of the patients was 53 years (range 37–64). Median Glasgow Coma Scale at admission was 7 (range 3–14), and median Glasgow Outcome Scale was 3 (range 1–5). The mortality rate was 20%. ONSD and FVsv demonstrated the strongest correlation with ICP (R = 0.76 for ONSD versus ICP; R = 0.72 for FVsv versus ICP), whereas PIa and the estimator based on FVd did not correlate with ICP significantly. Combining the 2 strongest nICP predictors (ONSD and FVsv) resulted in an even stronger correlation with ICP (R = 0.80). The ability to detect intracranial hypertension (ICP ≥ 20 mm Hg) was highest for ONSD (area under the curve [AUC] 0.91, 95% CI 0.88–0.95). The combination of ONSD and FVsv methods showed a statistically significant improvement of AUC values compared with the ONSD method alone (0.93, 95% CI 0.90–0.97, p = 0.01). Major limitations are the heterogeneity and small number of patients included in this study, the need for specialised training to perform and interpret the ultrasound tests, and the variability in performance among different ultrasound operators.
Of the studied ultrasound nICP methods, ONSD is the best estimator of ICP. The novel combination of ONSD ultrasonography and vTCD of the straight sinus is a promising and easily available technique for identifying critically ill patients with intracranial hypertension.
| Intracranial pressure (ICP) monitoring is necessary in many clinical scenarios.
Invasive ICP methods are the gold standard, but have many contraindications. Nevertheless, non-invasive ICP (nICP) measurement is a poorly developed technique.
We present a study comparing ultrasound-based nICP measurement techniques with the gold standard.
In a cohort of 64 patients with brain injury, we compared invasive ICP measurement with 3 different ultrasound-based measurements of nICP: optic nerve sheath diameter (ONSD) ultrasonography, arterial transcranial doppler (aTCD)–derived methods, and straight sinus systolic flow velocity (FVsv).
We found that both optic nerve sheath diameter ultrasonography (ONSD) and straight sinus systolic flow velocity (FVsv) are strongly correlated with invasive ICP. In addition, the combination of these 2 nICP parameters (ONSD and FVsv) resulted in stronger correlation with ICP.
A novel nICP monitoring method based on combined ONSD ultrasonography and venous transcranial Doppler has shown promising results for the measurement of intracranial pressure in patients with brain injury.
This ultrasound-based method is low-cost, quick, and based on technology widely available.
Future prospective studies will be needed to validate these results.
| Intracranial hypertension is a frequent and harmful complication of brain injury; it is an important contributing factor for secondary brain injury, and its severity and duration have been correlated with a fatal outcome [1,2].
A recent trial comparing an invasive intracranial pressure (ICP) monitoring protocol with a protocol based on imaging and clinical examination found no significant differences in patient outcome [3]. However, the trial has been criticised for being underpowered and for the methodology used to measure and treat ICP. Thus, invasive monitoring and treatment of intracranial hypertension is still widely recommended in the management of severely brain-injured patients despite a paucity of randomized evidence [4].
Invasive ICP monitoring through an intraventricular catheter or intraparenchymal microtransducer continues to be the standard of care after severe traumatic brain injury, and should be performed when indications are met [5]. Because the use of invasive transducers can cause complications including infection or haemorrhage [6–8], reliable non-invasive ICP (nICP) estimation would be helpful, especially in clinical situations where the risk–benefit balance of invasive ICP monitoring is unclear or when ICP monitoring is not immediately available or is even contraindicated [4]. Several non-invasive methods based on transcranial Doppler and optic nerve sheath diameter (ONSD) ultrasound are gaining clinical popularity due to their safety, availability, and reliability [8–13]. At present, the best accuracy for a non-invasive method reported in the literature [14,15] has been demonstrated by 2-depth high-resolution transcranial Doppler insonation of the ophthalmic artery. This method does not need calibration and is based on the measurement of the balance point when the measured parameters of blood flow velocity waveforms in the intracranial segment of the ophthalmic artery (which reflect ICP) are identical to extracranial segments (which are mechanically compressed by an externally applied pressure). Other authors [16,17] have proposed different methods for continuous nICP monitoring based on the waveform analysis of cerebral blood flow velocity from the middle cerebral artery (MCA) and arterial pressure. However, despite these promising results, non-invasive techniques remain of insufficient accuracy and temporal resolution to replace invasive ICP monitoring [18,19].
The aim of this study was to compare the accuracy of different ultrasound-based methods for nICP measurement in patients with severe traumatic brain injury undergoing invasive ICP monitoring. Such methods included the ultrasound measurement of the ONSD, venous transcranial Doppler (vTCD), and derived indices obtained from the straight sinus (such as straight sinus systolic flow velocity [FVsv]), and arterial transcranial Doppler (aTCD)–derived indices such as middle cerebral artery (MCA) pulsatility index (PIa) and diastolic flow velocity (FVd).
This is a single-centre, prospective observational study conducted from 1 January 2016 to 1 November 2016. Recruited patients were admitted at the Neurosciences Critical Care Unit, Addenbrooke’s Hospital, Cambridge, UK. The protocol was approved by the research ethics boards at the University of Cambridge (REC 15/lo/1918), and written consent was obtained from all participants’ next of kin. The article is reported as per Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines (S1 Text). Patients older than 18 years requiring sedation, mechanical ventilation, and ICP monitoring with an admission diagnosis of severe traumatic brain injury, aneurysmal subarachnoid haemorrhage, intraparenchymal haemorrhage, or stroke were considered for inclusion. Exclusion criteria were the following: the absence of an informed consent, a known history of ocular pathology or optic nerve trauma, skull base fracture with a cerebrospinal fluid (CSF) leak, inaccessible ultrasound windows (temporal for aTCD and occipital for vTCD), and clinical or radiological suspicion of cerebral venous thrombosis or vasospasm.
Patients were sedated with a continuous infusion of propofol or midazolam, fentanyl, and, when necessary, the muscle relaxant atracurium besylate. Mechanical ventilation was targeted to maintain adequate oxygenation (SaO2 > 90%) and normocapnia (PaCO2 = 35–40 mm Hg). Intravenous fluids, vasopressors, and inotropic support (norepinephrine and/or epinephrine) were administered to achieve and maintain an adequate cerebral perfusion pressure (CPP > 60 mm Hg). Clinical management was in accordance to international guidelines [20–22].
Treatment of intracranial hypertension was based on a protocol-driven strategy, which included optimisation of arterial blood pressure (ABP) and volaemia, sedation, and infusion of hyperosmolar fluids according to our institutional guidelines.
After a decision to place an ICP monitoring device (by the neurosurgical and intensive care physician in charge), patients were enrolled in the study. ICP was measured via an intraparenchymal probe (Codman & Shurtleff, Raynham, Massachusetts, US) or a catheter inserted into the brain ventricles and connected to an external pressure transducer and drainage system (Codman, Johnson & Johnson, Raynham, Massachusetts, US).
For each patient, we collected the following characteristics: admission Glasgow Coma Scale (GCS), age, sex, height, weight, comorbidities, mechanism and severity of brain injury, and discharge Glasgow Outcome Scale (GOS). The Rotterdam and Marshall scores as well as the Fisher scale were calculated using the admission computer tomography head scan reports [22].
Ultrasound measurement was performed by a selected group of experienced operators (TT, JP, MB) using a standardised insonation technique to reduce inter-operator variability. The operators were blinded to the patient’s admission diagnosis, demographics, baseline characteristics, and clinical and physiological background. Operators were not blinded to the actual ICP, but they were blinded to the final formulae to obtain a nICP estimation from the different measurements. Mean arterial blood pressure, end-tidal carbon dioxide partial pressure (ETCO2), MCA flow velocities (diastolic [FVd], mean [FVm], and systolic [FVs]), straight sinus flow velocities (diastolic [FVdv], mean [FVmv], and systolic [FVsv]), and ONSD were recorded twice daily from day 1 to day 5 after ICP insertion. Additional measurements were performed in case of acute increases in ICP (above 20 mm Hg). In cases where ICP mean values changed more than ±2 mm Hg during any of the 3 studies (ONSD ultrasound, vTCD, and aTCD), the measurements were excluded from the analysis.
On the basis of previous reports [25–27], we hypothesized that ICP is linearly associated with ONSD, systolic flow velocity on the straight sinus (FVsv), PIa, and ABP × (1 − FVd/FVm), and verified this hypothesis in 64 patients. A multivariable linear regression model was obtained from the relationship among ICP, ONSD, and FVsv to derive an nICP estimator based on the combination of ONSD and FVsv (nICPONSD+FVsv).
Deviations from the initial statistical plan (S2 Text) were based on reviewers’ requests and consisted of inclusion of linear mixed effects model analysis for the determination of the estimation formulas for nICP and exclusion of Bland–Altman analysis.
Statistical analysis of the data was conducted with RStudio software (R version 3.1.2). Initially, multiple measurement points were averaged for each patient; therefore, every patient was represented by one single value for all variables assessed. Then, the correlations between ICP and the variables of interest—ONSD, PIa, ABP × (1 − FVd/FVm), and FVsv—were verified using the Pearson correlation coefficient (R, with the level of significance set at 0.05).
To provide prediction models for ICP estimation, the relationships between ICP and the correlated variables were expressed as linear mixed effects models (R package lme4 [28]). As fixed effects, we entered ICP and the non-invasive estimators into the model. As random effects, we had intercepts and slopes for the repeated measurement points for each patient (N = 445 measurements). A mixed effects multiple regression between ICP and 2 correlated variables, ONSD and FVsv, was also performed. Chi-square (χ2) values and p-values for the comparison of the models were obtained by likelihood ratio tests of the full model with random intercepts and slopes against the null model with random intercepts only.
The area under the curve (AUC) of the receiver operating characteristic (ROC) curve was calculated to determine the ability of the non-invasive methods to detect raised ICP (using a threshold of 20 mm Hg; N = 445 measurements). Moreover, we also performed an analysis to determine the best ONSD and FVsv cutoff values for prediction of ICP ≥ 20 mm Hg. In ROC analysis, these are the values presenting the best sensitivity and specificity for prediction of a given threshold. The predicting ability is considered reasonable when the AUC is higher than 0.7 and strong when the AUC exceeds 0.8 [29]. Statistical differences between ROC curves were verified using the DeLong’s test for 2 correlated ROC curves (R package pROC [30]).
An analysis of variance (ANOVA) was performed to verify whether any of the variables assessed were associated with mortality in the patient cohort.
In all, 80 patients with intracranial pathology requiring invasive ICP monitoring were initially considered for enrolment in this study. Among these, 3 were excluded because of the absence of written consent, 2 because it was not possible to find a temporal window, 6 because the occipital window was inaccessible (cervical collar or patient position), 3 because the straight sinus could not be insonated, and 2 because of ocular lesions that precluded the assessment of ONSD.
A total of 445 recordings from 64 patients (each one including ONSD ultrasound, aTCD, and vTCD) were included in the final analysis. The percentage of measurements presenting ICP ≥ 20 mm Hg was 19.3% (N = 86). The characteristics of the patients are shown in Table 1. In Table 2, we present the median (interquartile range [IQR]) values of the variables analysed.
The correlation analysis between patients revealed good correlation between ICP and ONSD (R = 0.76) and between ICP and FVsv (R = 0.72), averaged per patient (N = 64)—both were statistically significant (p < 0.001) and without any influential outliers. The p-values for the correlations between ICP and PIa and ABP × (1 − FVd/FVm) were both non-significant (p = 0.63 and p = 0.36, respectively). Thus, the regression formulas adopted in this work considered only ONSD and FVsv, and the combination of both in a multiple regression model (Table 3).
Considering the variability in slopes between patients, full models allowing for random intercepts and slopes were significantly better at fitting the data than null models for a nICP estimator based on FVsv (nICPFVsv) and for nICPONSD+FVsv (χ2 = 44.19, p < 0.001, and χ2 = 40.92, p < 0.001, respectively). The inclusion of random slopes in the model describing a nICP estimator based on ONSD (nICPONSD) did not produce a significant difference in comparison to the model with random intercepts only (χ2 = 2.41, p = 0.30). The formulas of the derived models that best fitted the data are described below:
nICPONSD=5.00×ONSD−13.92(mmHg)
(1)
nICPFVsv=0.38×FVsv+0.0005(mmHg)
(2)
nICPONSD+FVsv=4.23×ONSD+0.14×FVsv−14.51(mmHg)
(3)
The correlation coefficient between the ONSD method and ICP averaged per patient (N = 64) was R = 0.76; the FVsv method showed a correlation with ICP of R = 0.72. The combination of the 2 methods presented a correlation coefficient of 0.81 (Table 4; Fig 1). S1 Fig displays the regression plots between ICP and the non-invasive estimators (ONSD and FVsv) for each patient, demonstrating the slope variability between patients with multiple measurement points.
Table 5 summarises the 95% prediction and confidence intervals for the linear regressions between ICP and all non-invasive estimators. The 95% prediction interval for ONSD ranged from 5.05 ± 4.04 to 19.32 ± 4.17 mm Hg; the 95% confidence interval ranged from 11.03 ± 3.95 to 13.34 ± 4.30 mm Hg. The 95% prediction interval for FVsv ranged from 4.57 ± 3.83 to 19.79 ± 3.96 mm Hg; the 95% confidence interval ranged from 10.94 ± 3.71 to 13.43 ± 4.12 mm Hg. For the combination of the 2 methods (nICPONSD+FVsv), the 95% prediction interval ranged from 5.65 ± 4.30 to 18.72 ± 4.45 mm Hg; the 95% confidence interval ranged from 10.90 ± 4.19 to 13.47 ± 4.60 mm Hg.
Results of ROC analysis are showed in Table 6 and Fig 2. ONSD had the best AUC among all methods for discriminating cases with intracranial hypertension (ICP ≥ 20 mm Hg) from cases without it (AUC = 0.91, 95% CI 0.88–0.95). The best ONSD and FVsv cutoff values for prediction of intracranial hypertension were 5.85 mm and 38.50 cm/s, respectively. The method based on the combination of ONSD and FVsv showed a statistically significant improvement of AUC values compared with the ONSD method alone (0.93, 95% CI 0.90–0.97, p = 0.01 [DeLong’s test]).
The outcome assessed at discharge revealed that 13 patients died (20%) and 51 survived. Mean ICP showed a tendency to be greater in patients who died; mean ONSD was greater in patients who died than in those who survived (Table 7; Fig 3). FVsv was not significantly different between survivors and non-survivors (p = 0.28).
In this study, we present and compare new models for ultrasound-based non-invasive estimation of ICP, based on ONSD ultrasonography, aTCD, and vTCD. Our results show that nICP derived from ONSD has the strongest correlation with invasive ICP. Moreover, ONSD measured through ultrasound was correlated with mortality at discharge. Finally, we demonstrated that a method based on the combination of the 2 best correlated parameters in our cohort (ONSD and FVsv—nICPONSD+FVsv) performed even better across all measurement points (R = 0.78; AUC for prediction of ICP ≥ 20 mm Hg was 0.93).
Measuring ONSD and FVsv using a duplex Doppler machine is fast and does not require probe fixation or specific dedicated hardware [13]. Furthermore, ultrasonography devices are available in most emergency departments and intensive care units, and are used for many other purposes. Therefore, ultrasonography could be very useful for nICP assessment.
The optic nerve is surrounded by subarachnoid space [10,11,31]; hence, the intraorbital part of the subarachnoid space is distensible and can therefore expand if the CSF pressure increases, with the maximum ONSD fluctuations occurring in the anterior compartment. Although the diameter of the optic nerve is narrower in the anterior than in the posterior segment, increased ICP in the perioptic CSF causes a greater enlargement of the retrobulbar segment of the optic nerve sheath, 3 mm behind the globe, than of the posterior segment [32]. This is probably related to the asymmetrical distribution of the arachnoidal trabeculae and the lower density of the arachnoidal trabeculae in the retrobulbar space. ONSD has been investigated in different clinical scenarios [10,33–35], showing a good correlation with ICP measured invasively and low inter- and intra-observer variability [10,11,27,36]. Our results agree with these findings. Among the studied methods, ONSD was the most accurate in the assessment of ICP; moreover, it is a safe and quick method, as the orbital window is easily available and has no complications.
vTCD for the assessment of ICP is a poorly developed technique. It is known that increasing ICP leads to venous haemodynamic changes, as the part of the cerebral vasculature most sensitive to elevated ICP is the subarachnoid bridging veins. According to the Monro–Kellie doctrine, cerebral compliance strongly depends on the compressibility of the low-pressure venous compartment, and stasis in the pial veins occurs early as a compensatory mechanism in case of increased ICP [37,38]. Consequently, venous blood may be pooled toward larger venous vessels (straight sinus and Rosenthal vein), causing an increase in venous flow velocity. An alternative explanation may be that straight sinus can be compressed by rising ICP, and, with constant volume flow, flow velocity may increase.
Schoser et al. applied vTCD for the estimation of ICP in 30 control volunteers and 25 patients with elevated ICP and found a linear relationship, with strong correlation between mean ICP and FVsv [25]. Similarly to Schoser et al., we found that FVsv is strongly correlated with ICP, whereas other vTCD parameters (venous pulsatility index and FVdv) were not good estimators of ICP.
Although the measurement of FVsv seems promising, this technique has some limitations: it can be impossible in polytraumatic patients because of the presence of a cervical collar (6 cases in our cohort). Moreover, the insonation of the straight sinus is feasible in just 72% of cases because of anatomical variations in cerebral veins and transcranial insonation difficulties [25] (even though we had just 3 unsuccessful cases in our cohort, 3.7%).
Our method has several potential clinical applications: it could be useful when invasive monitoring is contraindicated or unavailable, or in many “borderline” situations in which the insertion of invasive monitoring is questioned but a nICP measurement could be useful [20,21]. It can also be applied in patients at risk of intracranial hypertension for causes that are not primarily neurosurgical (such as liver transplantation and intraoperative settings at risk of intracranial hypertension [23,24]) or as screening tool in the emergency department in patients where there is doubt about the need for invasive ICP monitoring.
There are several limitations that deserve to be mentioned. First, transcranial Doppler (and ONSD) measurements were intermittent, and continuous measurements remain more feasible with invasive techniques. Second, the mixed cohort of enrolled patients, including different types of acute brain injury, may represent a bias, as the ICP and cerebral perfusion pressure thresholds for subarachnoid haemorrhage, intracerebral haemorrhage, and stroke are not as well defined as for traumatic brain injury. However, this heterogeneity increases the applicability of the study in many clinical scenarios. Other major limitations are the small number of patients included in this study, the need for specialised training to perform and interpret the ultrasound tests, and the variability in performance among different ultrasound operators.
Finally, most our measurements were obtained in patients with relatively well-controlled ICP. Although a strong correlation between nICP and invasive ICP within the range investigated supports the assumption of validity beyond the range investigated, larger validation studies will need to be performed before non-invasive techniques will be able to substitute for invasive ICP monitoring. In addition, despite our findings showing that mortality has a stronger association with ONSD than with ICP, this does not imply that it would be clinically better to monitor and manage ONSD than ICP.
A novel nICP monitoring method based on combined ONSD ultrasonography and vTCD was shown to have promising value for the diagnosis of intracranial hypertension, and a strong correlation with invasive ICP monitoring. This ultrasound-based method is quick, low-cost, and based on technology widely available in emergency departments and intensive care units. Whilst we still advocate the superiority of invasive ICP monitoring when this is clearly indicated, the non-invasive methodology presented here may be of potential benefit for ICP assessment in several clinical scenarios where invasive measurement is not immediately available or is contraindicated. However, this method has several limitations, and further studies are needed to confirm and validate our findings.
|
10.1371/journal.pbio.1000041 | Neto1 Is a Novel CUB-Domain NMDA Receptor–Interacting Protein Required for Synaptic Plasticity and Learning | The N-methyl-D-aspartate receptor (NMDAR), a major excitatory ligand-gated ion channel in the central nervous system (CNS), is a principal mediator of synaptic plasticity. Here we report that neuropilin tolloid-like 1 (Neto1), a complement C1r/C1s, Uegf, Bmp1 (CUB) domain-containing transmembrane protein, is a novel component of the NMDAR complex critical for maintaining the abundance of NR2A-containing NMDARs in the postsynaptic density. Neto1-null mice have depressed long-term potentiation (LTP) at Schaffer collateral-CA1 synapses, with the subunit dependency of LTP induction switching from the normal predominance of NR2A- to NR2B-NMDARs. NMDAR-dependent spatial learning and memory is depressed in Neto1-null mice, indicating that Neto1 regulates NMDA receptor-dependent synaptic plasticity and cognition. Remarkably, we also found that the deficits in LTP, learning, and memory in Neto1-null mice were rescued by the ampakine CX546 at doses without effect in wild-type. Together, our results establish the principle that auxiliary proteins are required for the normal abundance of NMDAR subunits at synapses, and demonstrate that an inherited learning defect can be rescued pharmacologically, a finding with therapeutic implications for humans.
| The fundamental unit for information processing in the brain is the synapse, a highly specialized site of communication between the brain's multitude of individual neurons. The strength of the communication at each synapse changes in response to neuronal activity—a process called synaptic plasticity—allowing networks of neurons to adapt and learn. How synaptic plasticity occurs is a major question in neurobiology. A central player in synaptic plasticity is an assembly of synaptic proteins called the NMDA receptor complex. Here, we discovered that the protein Neto1 is a component of the NMDA receptor complex. Neto1-deficient mice had a dramatic decrease in the number of NMDA receptors at synapses and consequently, synaptic plasticity and learning were impaired. By indirectly enhancing the function of the residual NMDA receptors in Neto1-deficient mice with a small molecule, we restored synaptic plasticity and learning to normal levels. Our findings establish the principle that inherited abnormalities of synaptic plasticity and learning due to NMDA receptor dysfunction can be pharmacologically corrected. Our discoveries also suggest that synaptic proteins that share a molecular signature, called the CUB domain, with Neto1 may be important components of synaptic receptors across species, because several CUB-domain proteins in worms have also been found to regulate synaptic receptors.
| In the mammalian central nervous system, excitatory transmission at synapses is mediated primarily by the amino acid glutamate, acting through the postsynaptic α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors (AMPARs) and N-methyl-D-aspartic acid receptors (NMDARs) [1]. Basal synaptic transmission is principally mediated by AMPARs, which are rapidly activated by glutamate, while the more slowly activated NMDAR primarily mediates various forms of synaptic plasticity. A large body of evidence indicates the NMDAR is essential for a prominent form of synaptic plasticity, long-term potentiation (LTP) at Schaffer collateral-CA1 synapses, and for hippocampal-dependent spatial learning and memory [2,3].
The core NMDAR is a heterotetramer comprised of two obligate NR1 subunits and two NR2(A-D) subunits [1]. These core subunits are embedded in a multiprotein complex that includes more than 70 NMDAR-associated proteins [4]. An emerging theme in NMDAR biology is that proteins associated with the core NMDAR may have important roles in the trafficking, stability, subunit composition, or function of NMDARs and may therefore be critical for synaptic plasticity, learning, and memory [5]. However, proteins that function to specifically maintain synaptic NMDARs, which are well-known for AMPARs, have been elusive for NMDARs.
We investigated the complement C1r/C1s, Uegf, Bmp1 (CUB) domain protein neuropilin tolloid-like 1 (Neto1) [6,7], which we have discovered to be an NMDAR-associated protein [8]. The CUB domain is an extracellular motif of approximately 110 amino acids originally identified in the complement subunits Clr/Cls, sea urchin epidermal growth factor, and bone morphogenetic protein 1 (BMP1). Comprised of 10 β-strands forming a “jellyroll” topology [9], CUB domains mediate protein-protein interactions [10]. Notably, the CUB domain protein SOL-1 in Caenorhabditis elegans has been shown to be a component of the GLR-1 glutamate receptor [11], required for its gating [12], and another C. elegans CUB-domain protein LEV-10 has been found to regulate the clustering of acetylcholine receptors at the neuromuscular junction [13]. Whether CUB domain proteins are significant components or regulators of neurotransmitter receptor complexes at vertebrate synapses is unknown despite the presence of ∼100 identified or predicted CUB domain proteins in the vertebrate genome [14].
To investigate the role of Neto1 in the biology of mammalian excitatory synapses, we determined the molecular basis of the Neto1:NMDAR interaction and defined the nonredundant functions of Neto1 in synaptic plasticity, learning, and memory using Neto1 protein null mice. We found that Neto1 interacts with the core NMDAR subunits, NR2A and NR2B, and with the scaffolding protein postsynaptic density-95 (PSD-95). The complete loss of Neto1 reduced the abundance of NR2A but not NR2B subunits in the PSD of the hippocampus, leading to a decrease in the amplitude of synaptic NMDAR currents and a switch from the normal predominance of NR2A- to NR2B-containing NMDARs at Schaffer collateral-CA1 synapses. In Neto1-null mice, LTP at these synapses was reduced and spatial learning and memory was impaired. By indirectly enhancing NMDAR synaptic currents in the Neto1-null mice using the ampakine CX546 [15], we rescued the deficits in both LTP and spatial learning and memory.
Neto1 encodes a 533 amino acid polypeptide with an N-terminal ER signal sequence, two CUB domains, one low-density lipoprotein receptor domain class A (LDLa) motif, a transmembrane domain, and a cytoplasmic tail terminating in a class I PDZ binding tripeptide ligand (TRV-COOH) (Figure 1A). We designated this protein neuropilin tolloid-like 1 (Neto1) [8], because the first CUB domain is most similar (∼40% identity) to the CUB domains of neuropilins [16,17] and tolloid [18].
To elucidate the role of Neto1 in the brain, we first examined the expression pattern of its mRNA. In adult mice, Neto1 mRNA was present throughout the central nervous system (Figure 1B–1D and Figure S1), with strong expression in cerebral cortex, hippocampus, olfactory bulb, olfactory tubercle, and caudate putamen.
To identify the subcellular compartments in which Neto1 is localized, we performed subcellular fractionation and immunoblotting experiments of whole mouse brain lysates. Because the C-terminal sequence of Neto1 suggested that it localized to the PSD (see below), we employed a cell fractionation strategy that separated synaptic subcompartments [19]. Neto1 was prominently expressed in the crude synaptosomal (Figure 2A, lane S2) and PSD fractions, but was absent from the synaptic vesicle fraction (Figure 2A, lane LP2). To visualize the cellular distribution of Neto1, we examined immunostained hippocampal sections by confocal microscopy. We found that Neto1 immunostaining decorated MAP2 positive dendritic arbors and co-localized with that of PSD-95 (Figure 2B) and NR1 (Figure 2C). We also found that Neto1 co-localized with actin (Figure 2D), which is highly enriched in dendritic spines in the hippocampus [20]. The immunostaining for Neto1 was not detected in hippocampus from Neto1-null (Neto1tlz/tlz, see below) mice (Figure 2D, right), indicating that the staining was not nonspecific. Together, these findings demonstrate that Neto1 is a component of the PSD of excitatory synapses.
The sequence of the C-terminal tripeptide of Neto1, TRV, suggested that it is a PDZ ligand, predicted to bind preferentially to the third PDZ domain (PDZ3) of PSD-95 [21,22]. Using the yeast two-hybrid system, we established that the cytoplasmic domain of Neto1 (Neto1-cd) associated with the full-length PDZ proteins PSD-95 (Figure S2), PSD-93, and SAP102, but not with SAP-97 or NIP [23,24] (unpublished data). Furthermore, using crude synaptosomal fractions, we determined that anti-PSD-95 antibodies co-immunoprecipitated Neto1 from wild-type (Neto1+/+) mouse brain (Figure 2E). Conversely, anti-Neto1 antibody co-immunoprecipitated PSD-95 from Neto1+/+ (Figure 3A, lane 1) but not from Neto1-null crude synaptosomal fractions (Figure 3A, lane 2). Negative control antibodies did not immunoprecipitate either Neto1 or PSD-95 (Figures 2E and 3A). The Neto1 cytoplasmic domain bound most strongly to PDZ3 of PSD-95, binding that was completely dependent on the C-terminal TRV of Neto1, in both the two-hybrid system (Figure S2) and in HEK293 cells (Figure 2F, lane 2). Moreover, the Neto1 cytoplasmic domain bound to a truncated PSD-95 polypeptide (PDZ1–3) containing only the three PDZ domains (Figure 2F, lane 3). Altogether, these findings indicate that Neto1 associates with PSD-95 in brain synapses through the binding of its C-terminal tripeptide with the PDZ domains of PSD-95.
Because PSD-95 is a prominent NMDAR scaffold protein [25,26], we asked whether Neto1 associates with NMDARs. We found that anti-Neto1 antibodies co-immunoprecipitated the NR1, NR2A, and NR2B NMDAR subunits from crude synaptosomal fractions of wild-type but not Neto1-null mice (Figure 3A and 3B, lanes 1 and 2), whereas pre-immune antibodies did not (Figure 3A and 3B, lane 3). Reciprocally, anti-NR1, anti-NR2A, and anti-NR2B antibodies co-immunoprecipitated Neto1 from wild-type synaptosomal fractions (Figure 3C, lanes 2, 4, and 5, respectively). In contrast, we were unable to co-immunoprecipitate Neto1 and GluR2 (Figure 3A, lane 1 and Figure 3C, lane 3), a major subunit of the AMPAR [27]. We therefore conclude that Neto1 is a component of the NMDAR complex but is not a general component of ionotropic glutamate receptor complexes.
To determine whether the association of Neto1 with NMDARs was entirely dependent upon the binding of its C-terminal PDZ ligand to PSD-95, we examined the binding of PSD-95 to an hemagglutinin (HA)-tagged Neto1 protein lacking the C-terminal 20 amino acids (Neto1-Δ20HA). As predicted both by the interaction between PSD-95 and the NR2 subunits of the NMDAR [28], and by the binding of Neto1 to PSD-95 described above, we found that Neto1 was co-immunoprecipitated by anti-NR1 antibodies from lysates of cells co-expressing Neto1, PSD-95, NR1, and NR2B (Figure 4A, lane 1). Unexpectedly, however, anti-NR1 antibodies co-immunoprecipitated Neto1-Δ20HA (Figure 4A, lane 3). Moreover, Neto1 or Neto1-Δ20HA co-immunoprecipitated with both NR1 and NR2B, even in the absence of PSD-95 (Figure 4A, lane 2 and 3, respectively). These results indicate that the binding of Neto1 to PSD-95 was not required for Neto1 to interact with the NMDAR, and that Neto1 interacts with NMDARs through a PSD-95-independent mechanism.
To identify the region of Neto1 that mediates the PSD-95-independent association between Neto1 and NMDARs, we examined the ability of a series of C-terminally deleted Neto1 proteins to co-immunoprecipitate with NMDARs from HEK293 cell lysates. Removal of the cytoplasmic tail and transmembrane domain of Neto1 did not abolish the Neto1:NMDAR interaction (Figure 4B, lanes 2 and 3), suggesting that it was mediated by the ectodomain of Neto1. Moreover, a construct expressing only the signal sequence and N-terminal CUB domain of Neto1 was sufficient to mediate the NMDAR association (Figure 4B, lane 5). In contrast, no binding was observed between NMDARs and the ectodomain of CSF-1 (Figure 4B, lane 6), or between NMDARs and the CUB domains of neuropilin-1 (Figure 4B, lane 7). These results indicate that the Neto1:NMDAR extracellular interaction is dependent on the first CUB domain of Neto1.
We next asked which NMDAR subunit mediates the Neto1:NMDAR interaction, using heterologously expressed proteins in HEK293 cells. Full-length Neto1 or Neto1-Δ20HA co-immunoprecipitated with both NR2A (Figure 5A, lanes 1 and 2, and Figure 5C, lane 4) and NR2B (Figure 5B, lanes 1 and 2, and Figure 5C, lane 5) expressed in the absence of NR1 and PSD-95. In contrast, in the absence of NR2, no association was observed between Neto1-Δ20HA and NR1 (Figure 5C, lane 3; Figure 5D, lane 2). Consequently, we conclude that the PSD-95-independent Neto1:NMDAR interaction is mediated through NR2 subunits, and that the first extracellular CUB domain of Neto1 is sufficient for this binding. The simplest model consistent with our findings is that Neto1 interacts with the NMDAR bivalently, with one Neto1:NMDAR interaction mediated through the binding of the C-terminal tripeptide of Neto1 to PSD-95, and the second through the extracellular domains of Neto1 and NR2 subunits.
To determine whether Neto1 is required for normal brain function in the mouse, we disrupted the Neto1 locus by homologous recombination in mouse embryonic stem (ES) cells. We generated a protein null allele by simultaneously introducing a tau-lacZ (tlz) reporter gene [29] in-frame into the initiation codon of the Neto1 gene (Figure 6A–6C). Both Neto1+/tlz and Neto1-null animals were normal in overall appearance with no gross morphological abnormalities in the brain. Hematoxylin and eosin staining revealed no histological abnormalities in any brain region examined in Neto1-null mice (unpublished data), and Nissl (Figure 6D and 6E), MAP2 immunostaining (Figure 6F and 6G), and Golgi staining of the hippocampus showed no morphological defects in Neto1-null mice (Figure 6H–6K). The absence of Neto1 had no effect on the overall abundance of NR1, NR2A, NR2B, PSD-95, GluR2, VAMP2, or GABAAR1 proteins (Figure 6L) in whole brain extracts, or of NR1, NR2A, NR2B, or PSD-95 in crude synaptosomes (Figure 6M). Moreover, the amount of NR2A, NR2B, and PSD-95 that co-immunoprecipitated with NR1 from crude synaptosomes was normal in Neto1-null mice, indicating that the lack of Neto1 did not alter the overall abundance of the NMDAR:PSD-95 holocomplex (Figure 6N).
Having shown that Neto1 is a component of the NMDAR complex, we asked whether glutamatergic synaptic transmission and plasticity are altered in the absence of Neto1. Given that Neto1 is expressed in the CA1 region of the hippocampus (Figure 1C), we studied synaptic transmission and plasticity at Schaffer collateral-CA1 synapses, which are widely used to investigate glutamatergic synaptic physiology [30]. We recorded field excitatory postsynaptic potentials (fEPSPs) in acute hippocampal slices from adult animals and used theta-burst pattern stimulation to induce long-term potentiation (tbLTP), a robust form of NMDAR-dependent synaptic plasticity [31]. Basal fEPSPs, afferent fiber volley, and paired-pulse facilitation in slices from Neto1-null mice were not different from those of wild-type littermate controls (Figure 7A–7C). In contrast, we found that tbLTP was reduced in Neto1-null mice (Figure 7D): the magnitude of the potentiation in the mutant animals was approximately 50% of that in wild-type controls 60 min and longer after theta-burst stimulation. Because paired-pulse facilitation, a measurement of presynaptic function [32], was not different in Neto1-null mice versus wild-type controls, the reduction in tbLTP is not the result of a deficit in presynaptic function. We therefore conclude that basal synaptic transmission at Schaffer collateral-CA1 synapses appears intact, whereas LTP is significantly impaired in Neto1-null mice.
tbLTP at Schaffer collateral-CA1 synapses is NMDAR-dependent [31]. We investigated NMDAR excitatory postsynaptic currents (EPSCs) evoked by Schaffer collateral stimulation, by using whole-cell recordings from CA1 pyramidal neurons (Figure 8). In order to examine NMDAR EPSCs in relationship to synaptic activation, we recorded both NMDAR and AMPAR EPSCs in the same neurons in wild-type and Neto1-null slices. We found that the NMDAR:AMPAR EPSC ratio was significantly less in Neto1-null neurons (Figure 8A) regardless of the size of AMPAR EPSCs examined (Figure 8B). Because basal synaptic transmission (Figure 7) and AMPAR-EPSCs (Figure S5) in Neto1-null neurons were not different from wild-type, we interpret the decrease in NMDAR:AMPAR EPSC ratio as indicating that synaptic NMDAR currents were reduced in Neto1-null neurons. The current-voltage relationship for NMDAR EPSCs in Neto1-null mice was comparable to that of wild-type animals, demonstrating that the Mg2+ blockade of the NMDARs was not altered by the lack of Neto1 (Figure 8C). Furthermore, we observed no abnormalities in the current-voltage relationship for AMPARs in Neto1-null mice (Figure 8D). Thus, basal NMDAR-mediated, but not AMPAR-mediated, synaptic responses are impaired in CA1 pyramidal neurons in the absence of Neto1. These findings suggest that the impairment in NMDAR EPSCs may account for the reduced tbLTP in Neto1-null mice.
The reduction in tbLTP and NMDAR EPSCs at Schaffer collateral-CA1 synapses suggested that there might be a decrease in the abundance or function of synaptic NMDARs. We found that the abundance of NR2A in the PSD fraction from whole hippocampal lysates from Neto1-null mice was reduced by approximately one-third compared with that of wild-type littermates (Figure 9A and 9B). Consistent with this reduction, the number of NR2A puncta in stratum radiatum of the CA1 region was also reduced, by approximately 60%, in Neto1-null mice (Figure 9C and 9D). In contrast, no significant differences were observed in the abundance of PSD-95, NR1, NR2B, or GluR2 between Neto-1 null versus wild-type mice (Figure 9A and 9B). Similarly, there were no differences in the number of NR2B or PSD-95 puncta in CA1 stratum radiatum of Neto1-null mice (Figure 9D and Figure S3A and S3B). These findings indicate that Neto1 is required to establish or maintain the normal abundance of NR2A-containing NMDARs in the PSD.
To determine whether there was an overall decrease in cell surface expression of NR2A-containing NMDARs in Neto1-null mice, we quantified the abundance of biotinylated cell surface proteins in wild-type and Neto1-null hippocampal slices. No differences in the level of biotinylated NR1, NR2A, or NR2B were found in Neto1-null compared with wild-type mice (Figure S4A), indicating that the overall cell surface expression of NMDARs is normal in the hippocampus in the absence of Neto1. Similarly, total NMDA-evoked current density and the fractional current carried by NR2A-receptors were also normal in acutely isolated CA1 pyramidal neurons from Neto1-null mice (Figure S4B and S4C). Collectively, these findings indicate that lack of Neto1 does not alter the total surface expression of NMDARs, but rather decreases the targeting or stability of NR2A-containing NMDARs at synapses.
To determine whether the decreased synaptic abundance of NR2A subunits leads to a reduction in NR2A-mediated synaptic currents we examined the relative contribution of NR2A versus NR2B to NMDAR EPSCs at CA1 synapses. In the adult hippocampus, NR2A-containing NMDARs make a larger contribution to basal NMDAR-mediated synaptic transmission than those containing NR2B subunits [33]. Consequently, if the decrease in NMDAR EPSCs was due to the reduced level of NR2A-NMDARs, the relative contribution of NR2B-NMDARs to synaptic NMDAR currents would be predicted to be increased in Neto1-null mice. We therefore compared the effect of blocking NR2B-NMDARs using the NR2B-selective antagonist, Ro25–6981 [34], in wild-type and Neto1-null mice. Because Ro25–6981 is a use-dependent NMDAR blocker, we continued the regular synaptic activation (0.1 Hz) during Ro25–6981 application and calculated its effect only after NMDAR EPSCs had stabilized, 20–30 min after the start of Ro25–6981 administration. In wild-type synapses, Ro25–6981 (2 μM) reduced NMDAR EPSCs by ∼30% (Figure 9E and Figure S6). In contrast, in Neto1-null synapses the reduction was ∼70% (p < 0.001) (Figure 9E and Figure S6), indicating that basal NMDAR EPSCs in Neto1-null synapses are mediated primarily by NR2B-containing NMDARs. Moreover, in Neto1-null synapses, but not in those of wild-type mice, the component of the NMDAR EPSC resistant to Ro25–6981 (2 μM) decayed more rapidly than did the component sensitive to Ro25–6981 (Figure S6). Thus, the absence of Neto1 decreases the relative contribution of NR2A-containing receptors to NMDAR EPSCs at Schaffer collateral-CA1 synapses.
We investigated the impact of the decrease of synaptic NR2A-mediated currents on tbLTP at Schaffer collateral synapses. Because basal NMDAR EPSCs in Neto1-null mice were mediated primarily by NR2B-containing NMDARs, we examined the effect of blocking NR2B-NMDARs on the induction of tbLTP in wild-type and Neto1-null mice using Ro25–6981. In wild-type slices, Ro25–6981 (2 μM) had no effect on tbLTP (Figure 9F). In contrast, in Neto1-null slices Ro25–6981 led to a ∼60% reduction in tbLTP (Figure 9F and 9G). These findings indicate that tbLTP in Schaffer collateral-CA1 synapses of adult Neto1-null mice is mediated primarily by NR2B-containing NMDARs. Taken together, these findings demonstrate that Neto1 is required for the normal abundance of synaptic NR2A-containing NMDARs and, as a result, for the normal contribution of NR2A-NMDARs to synaptic transmission and plasticity in CA1 hippocampus.
We reasoned that the decrease in NMDAR abundance and function in the hippocampus of Neto1-null mice might disrupt NMDAR-dependent learning and memory [3], and therefore tested wild-type and Neto1-null littermate mice in the Morris water maze task, with two acquisition phases [35]. We found no difference between wild-type and Neto1-null mice in latency to find a platform marked with a visible cue (Figure 10A, pretraining), indicating that the lack of Neto1 had no detectable adverse effects on the visual and motor functions required for this task. Moreover, there were no differences between groups in the first acquisition phase (Figure 10A, days 1–6), nor in the first probe trial (Figure 10B and Figure S7A).
In contrast, when the platform was relocated in the second acquisition phase, Neto1-null mice failed to reduce their escape latency during training and were impaired in the second probe trial as compared with the wild-type controls (Figure 10A, days 7–9). The differences could not be explained by a deficit in motor performance because swim speed, measured in every trial, was not different between the two genotypes (Figure S7B). In the second probe trial [35], wild-type mice showed a strong preference for the new target quadrant whereas Neto1-null mice showed no preference for this quadrant (Figure 10C). In addition, the mutant mice crossed the new platform location less frequently than their wild-type littermates (Figure S7C) and did not persevere in crossing the original platform location (Figure S7C). Neto1-null mice predominantly used nonspatial search strategies, such as scanning and chaining, as compared with the spatial strategies such as focal searching and direct swims [36] used by wild-type mice (Figure S7D and S7E). Altogether, the above findings establish that Neto1-null mice are impaired in hippocampal-dependent spatial learning.
To further characterize the hippocampal-dependent learning abnormalities in Neto1-null mice, we used two other spatial learning tests—the delayed matching-to-place version of the Morris water maze task [37] and the displaced object (DO) task [38]—and a nonspatial test, the novel object recognition task [39]. Neto1-null mice were impaired in both the delayed matching-to-place task (Figure 11A–11C and Figure S8) and the DO task (Figure 11D). In contrast, the performance of Neto1-null mice was the same as wild-type littermates in the novel-object recognition task (Figure 11E, Figure S9A, and Table S1). Taken together, our findings from the behavioural studies indicate that Neto1-null mice have broad deficiencies in spatial learning whereas the nonspatial task examined did not require Neto1.
We considered that the deficits in LTP and learning might be restored by enhancing the residual NMDAR function in Neto1-null mice. Our strategy was to increase NMDAR-mediated currents preferentially at active synapses using the ampakine CX546. CX546 decreases the desensitization of AMPARs [15], thereby prolonging AMPAR EPSPs and secondarily increasing current through NMDARs by reducing the Mg2+ blockade. We found that at a concentration of 25 μM, CX546 had no effect on tbLTP in wild-type slices but restored tbLTP in Neto1-null slices to the wild-type level (Figure 12A and 12B). At this concentration, CX546 prolonged AMPAR-mediated EPSCs (Figure 12C) and this effect was similar in both wild-type and Neto1-null neurons (wild-type 160 ± 16%; Neto1-null 154 ± 21%). In contrast, CX546 (25 μM) had no effect on the amplitude, decay, or voltage-dependence of pharmacologically isolated NMDAR EPSCs (Figure 12D and 12E). CX546 (25 μM) also had no effect on paired-pulse facilitation (Figure S10A), indicating that presynaptic function was not altered by CX546. Corresponding to the prolongation of AMPAR EPSCs CX546 caused an increase in the duration of the fEPSPs (Figure S10B) and CX546-prolonged fEPSPs showed an NMDAR-component (Figure S10C). Moreover, we found that the fully rescued LTP in Neto1-null hippocampal slices was suppressed by more than 65% by Ro25–6981, at a dose that was without effect on LTP in wild-type slices (GMP, DN, RRM, MWS, unpublished data). These findings indicate that by prolonging AMPAR EPSCs, CX546 secondarily increases current through NMDARs in CA1 hippocampus in Neto1-null mice, thereby restoring tbLTP to wild-type levels.
Finally, we asked whether the strategy of using CX546 to indirectly enhance NMDAR function restores learning and memory in Neto1-null mice. In the Morris water maze task we used a dose of CX546 (15 mg/kg) that had no effect on learning in wild-type mice but that restored the escape latency and probe trial impairments in Neto1-null mice to normal (Figure 13A–13C and Figure S11). Moreover, in the DO task, Neto1-null mice treated with the same dose of CX546 spent the same amount of time investigating the DO as wild-type mice (Figure 13D). All test groups had a similar habituation profile (Figure S9B). In summary, tbLTP and spatial learning in Neto1-null mice were pharmacologically rescued by CX546, at doses that were without effect in wild-type animals.
We have established that Neto1 is a critical component of the NMDAR complex, and that loss of Neto1 leads to impaired hippocampal LTP and hippocampal-dependent learning and memory. We have shown that Neto1 interacts with NMDARs through the extracellular domain of their NR2 subunits, as well as intracellularly through PSD-95. Although Neto1 binds to both NR2A and NR2B, the loss of Neto1 leads to a reduction in the abundance of NR2A, but not NR2B, in the PSD fraction from hippocampus and a reduction in NR2A puncta in the CA1 region. Consistent with the reduction in NR2A protein in the PSDs of Neto1-null mice, which had no change in total NR2A abundance in whole brain, we identified a decrease in NMDAR EPSCs at Schaffer collateral-CA1 synapses, which are normally dominated by NR2A-containing receptors [40]. Blockade of NR2B-containing NMDARs in Neto1-null neurons caused a dramatic decrease in NMDAR-mediated EPSCs, indicating that the majority of NMDAR-mediated EPSCs in Neto1-null hippocampal neurons are contributed by NR2B-containing NMDARs and not NR2A-NMDARs. These findings indicate that Neto1 plays a critical role in maintaining the delivery or stability of NR2A-containing NMDARs at CA1 synapses.
The preferential effect of the loss of Neto1 on the abundance of synaptic, but not total, NR2A-containing NMDARs would not have been predicted from studies on the basis of the disruption of other NMDAR-interacting proteins. Rather than having a specific regulatory role on synaptic targeting of NMDARs like Neto1, loss of function of the other NMDAR-interacting proteins studied to date affects the overall cellular trafficking, function, or downstream signaling of NMDARs [41–43]. In its role in targeting NR2A-NMDARs to the synapse, Neto1 may be comparable to the TARPs, which control targeting of AMPARs to synapses [44,45]. Our identification of Neto1 as a critical auxiliary protein for NR2A-NMDARs raises the possibility that other proteins, perhaps other CUB domain proteins, may be required, like Neto1, to maintain non-NR2A-NMDARs at synapses. Thus, Neto1 represents a new protein that functions to specifically maintain synaptic NMDARs, a protein that has been elusive for NMDARs.
The loss of synaptic NR2A-containing receptors in the Neto1-null mice implies that the molecular events regulating the delivery or stability of NR2A-NMDARs at the synapse differ from those regulating NR2B-NMDARs. Despite the ability of Neto1 to bind to both NR2A and NR2B subunits in vitro, the differential effect of Neto1 on NR2A- versus NR2B-containing NMDARs in vivo, might be mediated by the extracellular, membrane or cytoplasmic domains of these NR2 subunits. The membrane domains of NR2A and NR2B, however, are over 95% identical and are therefore unlikely to be responsible for the differential effect of loss of Neto1. The extracellular domains of NR2A and NR2B are 54% identical, being dominated by the S1 ligand-binding region and the amino terminal domain, with the extreme N-terminal sequence being the most divergent. The cytoplasmic domains of NR2A and NR2B are the most divergent, having only 29% sequence identity. Differences in motifs within the extracellular or cytoplasmic domains may thus be responsible for the differential effect on synaptic NR2A NMDARs in the Neto1-null mice.
The functional consequences of the differences between NR2A and NR2B have been most clearly delineated for their cytoplasmic domains. For example, the endocytic motifs in the distal C termini of NR2A and NR2B, LL and YEKL, respectively, have been demonstrated to interact with clathrin adaptor complexes with different affinities [46]. After endocytosis, NR2A and NR2B sort into different intracellular pathways, with NR2B preferentially trafficking to recycling endosomes. Other studies indicate that the cytoplasmic domains of NR2A and NR2B preferentially associate with unique sets of proteins. For example, NR2B but not NR2A interacts with Ras-guanine nucleotide-releasing factor 1 (Ras-GRF1), which is critical for NMDAR-mediated activation of ERK [47]. NR2B also binds preferentially to CaMKII [48–51] allowing CaMKII to remain active after the dissociation of Ca2+/calmodulin. NR2 subunit-specific signalling mechanisms can therefore be dictated, in part, by the properties and context conferred by the different associated proteins. Thus, the Neto1-dependent subunit-specific regulation may reflect differences in NR2-NMDAR associated proteins.
The loss of Neto1, while having no effect on basal AMPAR-mediated synaptic transmission, suppresses LTP to a degree comparable to that observed in mice lacking NR2A [52] or its C-terminal tail [53]. In NR2A-null mutant mice, as in Neto1-null mice, LTP at Schaffer collateral-CA1 synapses is mediated by NR2B-NMDARs [54]. Moreover, the spatial memory deficit of Neto1-null mice in the Morris water maze task is comparable to that of NR2A-null mice: the initial acquisition is normal, but other tests of spatial memory are impaired including, for example, the “spontaneous spatial novelty preference test” [55]. Similarly, in mice lacking the C terminus of NR2A, the initial acquisition in the Morris water maze is normal but, like the NR2A-null, these mice also have impaired spatial working memory [55]. The deficits in the Neto1-null mice indicate that Neto1 may have specific roles in the acquisition of spatial memory. The deficit in the delayed matching-to-place indicates that Neto1 is crucial for rapid spatial learning as described by Nakazawa and colleagues [37].
Our discovery that Neto1 in vertebrates is a component of the NMDAR complex, together with the previous identification of SOL-1 [11] and LEV-10 [13] in C. elegans as CUB domain-containing proteins associated with the GLR-1 and ACh receptors, respectively, suggests that the CUB domain may be an evolutionarily conserved molecular signature of a significant subset of the proteins associated with neurotransmitter receptors. Loss of function of these three CUB domain proteins has no impact on the overall abundance of the associated receptor complexes. Rather, loss of Neto1 and LEV-10 each leads to a reduction in synaptic localization of the cognate receptors, whereas loss of SOL-1 leads to a loss of function of normally distributed GLR-1. Both Neto1 and SOL-1 interact with ionotropic subunits by an extracellular CUB domain. Binding of a soluble CUB domain of SOL-1 partially rescues the function of GLR-1 ionotropic receptors [12]. It is not yet known whether soluble Neto1 CUB domains can rescue the impaired LTP or the reduced number of NR2A-containing receptors at hippocampal excitatory synapses in Neto1-null mice. Because Neto1, SOL-1, and LEV-10 are associated with neurotransmitter receptors of different classes, our work suggests that a critical interaction with a CUB domain-containing protein may be a general characteristic of ligand-gated ion channels throughout nature.
In Neto1-null mice, the impairments in LTP and spatial learning were rescued by the ampakine CX546, administered acutely by bathing hippocampal slices in the drug prior to LTP-inducing stimulation, or by administering it systemically prior to each training session, respectively. Importantly, CX546 was used at doses that we demonstrated to have no effect on synaptic plasticity or learning in wild-type mice. This is the first report of a pharmacological rescue of an NMDAR impairment, and consequently, our results extend the principle that in vertebrates, an inherited defect in synaptic plasticity and spatial learning can be corrected in the adult [56]. We showed that CX546 prolongs AMPAR-mediated EPSCs and that the prolongation is the same in wild-type and Neto1-null mice, but that it does not affect NMDAR-mediated EPSCs or paired pulse facilitation. Consequently, the most parsimonious explanation of the CX546-mediated rescue (Figure 14) is that it indirectly facilitates NMDAR-mediated synaptic responses by prolonging AMPAR EPSCs, extending the temporary relief of the Mg2+ blockade and thereby increasing Ca2+ influx through NMDARs to the wild-type level required for full expression of the LTP signaling cascade [43,57]. A comparable strategy of modulating non-NMDARs to secondarily facilitate NMDAR currents has also been used, but with a genetic approach, in C. elegans: the disruption of foraging behaviour by mutant NMDARs was restored by a slowly desensitizing variant of the non-NMDARs [58]. Thus, we expect that a slowly desensitizing AMPAR variant would rescue LTP in the Neto1-null mice. The recovery of LTP or learning by CX546 could be explained by facilitation of either NR2A- or NR2B-NMDAR mediated responses. However, we found that the fully rescued LTP is suppressed by more than 65% in Neto1-null hippocampal slices by Ro25–6981, at a dose that is without effect on LTP in wild- type slices indicating that NR2B-NMDARs, and not only NR2A-NMDARs, are required for the rescue of LTP. Hence, the rescue of spatial learning observed in Neto1-null mice may also be dependent on NR2B-NMDARs.
In summary, in addition to the rescue of synaptic plasticity mediated by CX546, we have discovered that the CUB domain protein Neto1 is a component of the NMDAR complex and that it plays a central role in the normal function of NMDARs at hippocampal excitatory synapses. Mice lacking Neto1 have a normal abundance of NR2B-containing NMDAR receptors but a reduction of NR2A-containing receptors at hippocampal excitatory synapses. The reduction of NMDAR-mediated synaptic currents, impaired synaptic plasticity at hippocampal Schaffer collateral-CA1 synapses, and impaired spatial learning observed in the Neto1-null animals can be attributed to the decreased levels of NR2A-containing receptors at hippocampal excitatory synapses. Altogether, our findings establish that Neto1 is an important regulator of the NMDAR complex required for normal NMDAR-mediated synaptic plasticity and learning. Our results, together with the identification of the CUB domain proteins SOL-1 and LEV-10 as regulators of ionotropic receptors in nematode, suggest that a critical interaction with a CUB domain protein may be a common feature of different types of ligand-gated ion channels across species. Moreover, our studies establish the principle that inherited abnormalities of synaptic plasticity and spatial cognition due to NMDAR dysfunction can be pharmacologically corrected.
Human UniGene clusters were analyzed using the BLAST algorithm [59] to identify proteins with motifs suggestive of a neurodevelopmental function. One retinal UniGene cluster, Hs.60563, a partial cDNA predicting a CUB-domain ORF related to neuropilins and tolloids, which we designated NETO1 [8], was selected for further study. Full-length mouse Neto1 cDNAs were obtained by reverse transcription (RT)-PCR from adult mouse brain cDNA. To disrupt the Neto1 gene by homologous recombination, we generated a targeting construct with a tau-lacZ-loxP-pgk-neo-loxP cassette cloned in-frame with the Neto1 start codon (Figure 6A). Mouse R1 embryonic stem (ES) cells were electroporated, and positive clones were identified by Southern blotting. Two independent mouse lines were generated by blastocyst injection, and transmitting male chimeras were mated with C57BL/6J mice. A proportion of F2 Neto1tlz/tlz mice were observed to have infrequent myoclonic seizures commencing at the age of weaning [8]. However, no F3 Neto1tlz/tlz mice or subsequent generations exhibited seizure activity either by behavioural observation or by EEG recording. Therefore, we used only F3 and later generation Neto1+/tlz and Neto1tlz/tlz mice in the present study.
The generation of guinea pig anti-Neto1 antibodies is described elsewhere [60]. Rabbit antibodies to Neto1 were raised to the C-terminal 86 amino acids of Neto1 and prepared as described by Chow and colleagues [60], except that the antigen was further purified by electroelution from a SDS-polyacrylamide gel. Other antibodies were purchased from commercial sources. See Table S2 for details.
Immunostaining was adapted from Schneider Gasser et al. [61]. Briefly, fresh 300-μm vibratome-cut hippocampal slices, trimmed from sagittal brain slices, were fixed in 2% PFA/PBS on ice for 20 min, washed three times in PBS, and incubated “free-floating” in blocking solution (10% goat serum, 0.1% triton-X, PBS) for 1 h. Primary antibodies (see Table S2) in blocking solution were incubated with slices for 48 h under gentle agitation at 4 °C. Slices were washed three times in PBS, and incubated with appropriate secondary antibodies for 24 h under gentle agitation at 4 °C. Following incubation, slices were washed three times with PBS, transferred, and mounted on to glass slides with Immun-Mount (Thermo Scientific). Images were acquired using a Zeiss LSM 510 confocal microscope. For quantitative studies, three age-matched (2-mo-old) pairs of wild-type and Neto1-null littermates were examined. In each littermate pair, brain slices from each genotype were combined into the same well, and subsequently processed together under identical conditions, as described above. Slices were double-labeled with antibodies against Neto1 and either NR2A, NR2B, or PSD-95. All slices from the same well were mounted onto the same glass slide, and images were acquired with fixed exposure settings. Puncta from stratum radiatum in CA1 of Neto1-null and control slices were quantified using ImageJ software with identical parameters.
The yeast two-hybrid system was initially used to determine whether the cytoplasmic tail of Neto1 could interact with PSD-95 and the related proteins PSD-93, SAP-102, and SAP-97. Fragments encoding the cytoplasmic region of Neto1 (amino acids 345–533) and the C-terminal mutant ΔTRV (comprising amino acids 345–530) were amplified by PCR from mouse whole brain cDNA and subcloned into the yeast vector pBD-GAL4 (Stratagene) containing the GAL4 DNA-binding domain. Full-length PSD-95, PSD-93, SAP-102, and SAP-97 cDNAs, and cDNAs encoding different parts of PSD-95 were derived from mouse brain by RT-PCR using primers designed from published DNA sequences. The cDNAs were subcloned into the yeast vector pAD-GAL4 (Stratagene). The controls used were the cytoplasmic domain of mouse neuropilin-1 [16] cloned into the pBD-GAL4 vector, and full-length NIP [24] cloned into the pAD-GAL4 vector. The yeast vectors were sequentially transformed into the Saccharomyces cerevisiae strain YRG-2 (Stratagene) and the interactions scored by growth in the absence of leucine, tryptophan, and histidine, and using a β-galactosidase filter assay.
Full-length Neto1 cDNA (encoding amino acids 1–533) and deletion mutants Neto1-ΔTRV (1–530), Neto1-Δ20HA (1–513), Neto1-Δ20-eGFP (1–513), Neto1-Δcyto-eGFP (1–363), Neto1-ΔcytoTM-eGFP (1–340), Neto1 CUB12-eGFP (1–290), Neto1 CUB1-eGFP (1–162), Nrpn1 CUB12-eGFP (1–270) (from neuropilin-1) [16], and CSF-1 EC-eGFP (1–294) (from macrophage colony-stimulating factor 1 receptor) [62] were generated by PCR and subcloned into a variant of pcDNA3.1mycHisA(+) (Invitrogen) containing two copies of the influenza hemagglutinin (HA) epitope tag or the eGFP coding sequence, and sequence verified. GW1-PSD-95 (full-length human PSD-95) and pM18S-PDZ1–3 (containing PDZ domains 1, 2, and 3 of human PSD-95) have been described [63]. The NR1 construct used expresses the NR1-1a isoform, which lacks the PDZ binding motif [64].
For co-immunoprecipitation experiments, HEK293 cells were transfected using SuperFect (Qiagen). Cells transfected with NR1 and NR2 subunits of the NMDA receptors were grown in the presence of 300 μM DL-2-amino-5-phosphonovaleric acid (Sigma). 48 h after transfection, cells were washed with PBS and lysed in RIPA buffer (1 ml/100-mm plate), containing 50 mM Tris/HCl (pH 7.4), 150 mM NaCl, 1 mM EDTA, 1% Nonidet P-40, 0.1% SDS, 0.5% deoxycholate (DOC) supplemented with protease inhibitors. Lysed cells were incubated on ice for 30 min, and centrifuged at 14,000g for 15 min at 4 °C.
Cell lysates (∼1.5 mg of protein) were incubated directly in the presence or absence of antibodies (2 μg) for periods ranging from 1 h to overnight at 4 °C on a rotating platform. Lysates were subsequently incubated with either 30 μl protein A-agarose beads (GE Healthcare) or 30 μl anti-mouse IgG beads (Sigma) for 1–5 h at 4 °C on a rotating platform. After centrifugation, beads were washed three times with RIPA buffer. Bound proteins were eluted with SDS sample buffer and subjected to SDS-PAGE and immunoblotting. For immunoprecipitation from crude synaptosomal fractions, prepared as previously described [65], 1 mg of synaptosomal protein was incubated in the presence or absence of antibodies (2 μg) or pre-immune IgGs overnight with rotation at 4 °C, and further incubated with either 30 μl protein A-agarose beads or 30 μl anti-mouse IgG beads for 3–5 h with rotation at 4 °C. After centrifugation, beads were washed three times with RIPA buffer. Bound proteins were eluted with SDS sample buffer and subjected to SDS-PAGE and immunoblotting.
Subcellular fractionation of mouse brains was performed as described [66,67]. All buffers contained a cocktail of protease inhibitors (Roche). The PSD fraction was prepared from whole brains or pooled hippocampi from 2–4-mo-old mice as described previously [19], except that PSDs were extracted only once with Triton X-100. Crude synaptosomal fractions were prepared as previously described [65] from wild-type or Neto1-null brains. For protein quantification, proteins were solubilized by boiling in 1% SDS and quantitated using a detergent-compatible assay (Bio-Rad).
Biotinylation studies were performed as previously described with modifications [68]. Briefly, 200-μm hippocampal slices from age-matched wild-type and Neto1-null littermate mice were incubated in ACSF saturated in 95% O2 5% CO2 at room temperature for at least 1 h. Ten slices from each genotype were incubated in 2 ml of ACSF containing 500 μg/ml biotin (Pierce), on ice, bubbled in 95% O2 5% CO2, with gentle agitation for 1 h. Slices were washed three times in ACSF and homogenized with 1 ml of RIPA buffer with a protease inhibitor cocktail (Roche) and incubated on ice for 30 min. The homogenate was centrifuged and supernatant was collected, and quantified using the BioRad Dc protein quantification kit. 50 μg of total protein in a total volume of 300 μl was mixed with 200 μl of a 50% slurry of Neutravidin beads (Pierce) and rotated for 1 h at 4 °C. The beads (first bound fraction) were harvested by centrifugation and washed three times in RIPA buffer. The remaining supernatant was subjected to a second binding of 200 μl of 50% slurry of Neutravidin beads and rotated for 1 h at 4 °C. The beads (second bound fraction) were then centrifuged and washed three times with RIPA buffer. Samples were resolved by SDS PAGE and blotted with appropriate primary antibodies.
A DNA fragment corresponding to the first CUB domain (CUB1) of mouse Neto1 was used to hybridize RNA blots using standard procedures. For in situ hybridizations, mouse embryos and mature tissues were fixed in PBS/4% paraformaldehyde (PFA) overnight, rinsed in PBS, and equilibrated in PBS/30% sucrose at 4 °C. In situ hybridization was adapted from an established protocol [69].
Two-month-old animals were perfused with 4% PFA in PBS and brains were sectioned and stained using hematoxylin and eosin or cresyl violet using standard methods. Brains used for Golgi staining were processed according to manufacturer's directions (FD Neurotechnologies, Inc). Serial coronal and saggital brain sections were examined.
Hippocampal slices prepared from 8–12-wk-old littermate mice were placed in a holding chamber for at least 1 h prior to recording. A single slice (300 μm) was then transferred to a recording chamber and superfused with artificial cerebrospinal fluid (ACSF) at 2 ml/min composed of 132 mM NaCl, 3 mM KCl, 1.25 mM NaH2PO4, 2 mM MgCl2, 11 mM D-glucose, 24 mM NaHCO3, and 2 mM CaCl2 saturated with 95% O2 (balance 5% CO2) at 28 ± 2 °C (pH 7.40; 315–325 mOsm). fEPSPs were evoked using bipolar tungsten electrodes located approximately 50 μm from the cell body layer in CA1 and were recorded using glass micropipettes filled with ACSF placed in the stratum radiatum 60–80 μm from the cell body layer. Stimulation of Schaffer collateral afferents consisted of single pulses (0.08-ms duration) delivered at 0.1 Hz. In LTP experiments, theta-burst stimulation (TBS) consisted of 15 bursts of four pulses at 100 Hz, delivered at an interstimulus interval of 200 ms. Stimulus intensity was set to 30%–35% of that which produced maximum synaptic responses. fEPSP slope was calculated as the slope of the rising phase between 10% and 60% of the peak of the response. Whole-cell EPSC recordings were done using the visualized method (Zeiss Axioskop 2FS microscope) with patch pipettes (3–5 MΩ) containing intracellular solution composed of: 132.5 mM Cs-gluconate, 17.5 mM CsCl, 10 mM HEPES, 10 mM BAPTA, 2 mM Mg-ATP, 0.3 mM GTP, 5 mM QX-314, (pH 7.25; 290 mOsm) placed in the cell body layer in the CA1. Synaptic responses were evoked with a bipolar tungsten electrode placed approximately 50 μm from the CA1 cell body layer. ACSF was supplemented with bicuculline methiodide (10 μM). AMPAR EPSCs were recorded with cells held at −70 mV. Stimulation to evoke AMPAR EPSCs consisted of single pulses (0.08-ms duration) delivered to Schaffer collateral-CA1 synapses at 0.1 Hz with increasing strength (Figure 8 and Figure S5). For each cell at each stimulus intensity tested, six consecutive EPSCs were recorded and the peak amplitudes averaged. NMDAR EPSCs were recorded from the same CA1 pyramidal neurons (Figure 8) but held at +60 mV in order to remove the NMDAR-voltage-dependent Mg2+ block and perfused with ACSF containing DNQX (5 μM) or CNQX (10 μM). The same stimulation protocol used to evoke AMPAR EPSCs was used to evoke NMDAR EPSCs. Current-voltage relationships for AMPAR and NMDAR EPSCs were also performed. Raw data were amplified using a MultiClamp 700A amplifier and a Digidata 1322A acquisition system sampled at 10 KHz, and analyzed with Clampfit 9.2 (Axon Instruments) and Sigmaplot 7 software. Recordings were performed with the experimenter blind to the genotype. ACSF was supplemented as indicated with Ro25–6981 (2 μM; Tocris), which was made fresh immediately before the experiment. ACSF was also supplemented as indicated with CX546 (25 μM; dissolved in H2O; Cortex Pharmaceuticals), which was made fresh immediately before the experiment. CX546 caused no change in the initial slope of the fEPSP but prolonged the decay phase. Data are presented as mean (±SEM). Student's t-test or two-way ANOVA with the Tukey test were used for statistical comparison.
Acutely dissociated hippocampal CA1 neurons were obtained from Neto1+/+ and Neto1tlz/tlz mice as previously described [70]. At 20–22 °C, pyramidal CA1 neurons were voltage-clamped at −60 mV in the whole cell configuration using borosilicate micropipettes (series resistance 3–8 MΩ) filled with intracellular solution that contained (in mM): CsF 140, HEPES 10, MgCl2 2, ethylene glycol-O-O'-bis(2-aminoethyl)-N,N,N′,N′-tetraacetic acid (EGTA) 10, magnesium adenosine 5”-triphosphate (MgATP) 4, buffered to a pH of 7.4 using CsOH and adjusted to an osmolality of 290–300 mOsm. The CA1 neurons were then lifted into the stream of extracellular perfusion solution containing (in mM): NaCl 140, CaCl2 1.3, KCl 5.4, N-2-hydroxyethylpiperazine-N′-2-ethanesulphonic acid (HEPES) 25, glucose 33, tetrodotoxin 0.0003, and glycine 0.01, buffered to a pH of 7.4 with NaOH and adjusted to an osmolality of 320–325 mOsm. Rapid solution exchanges were accomplished by a motor-stepped fast perfusion system. NMDA-evoked current were recorded using the Multiclamp 700A amplifier with data filtered at 2 kHz, digitized using the Digidata 1322A, and acquired on-line at a sampling frequency of 10 kHz using the pCLAMP8 program. Prior to agonist exposure, a capacitance transient resulting from a 10-mV hyperpolarizing step was also recorded and used to estimate neuron size and current density in response to NMDA 1 mM. The concentration of NMDA that produced 50% of the maximal peak responses (EC50) and the respective Hill coefficient (nH) were determined according to the equations: where Imax is the maximal response observed at a saturating concentration (1 mM) of NMDA (using Graphpad Prism version 4). In experiments using ifenprodil 10 μM to inhibit NR2B-containing NMDA receptors, the ifenprodil was preperfused for 2 min before its co-application with NMDA 1 mM. Data are represented as mean ±SEM.
For the Morris water maze task, mice tested were the 12–16-wk-old Neto1-null and wild-type F3 progeny of intercrossed Neto1+/tlz heterozygotes having a mixed genetic background averaging 50% C57BL/6J, 25% 129S1/SvImJ, and 25% 129X1/SvJ. Pink-eyed mice were excluded from behavioural testing to minimize variation in visual acuity. The water maze consisted of a 185-cm diameter cylindrical tank that contained a 15-cm circular platform and water (26 ± 1 °C) rendered opaque by the addition of white nontoxic paint. The training regime consisted of three phases: pretraining to a visible (V) platform in the northeast quadrant (NE) for 1 d (four trials; maximum duration, 90 s; inter-trial interval [ITI], 30 min); acquisition training to a hidden platform in the southeast (SE) quadrant for 6 d (day 1–6; six trials per day; maximum duration, 90 s; ITI, 40 min); second acquisition training to a hidden platform in the northwest (NW) quadrant for 3 d (day 7–9; six trials per day; maximum duration, 90 s; ITI, 30 min). Probe trials (90 s duration) were administered 18 h after the last acquisition and reversal trials, respectively.
The same cohort of mice was further trained in a delayed matching-to-place task, in which mice had to repeatedly learn a new spatial location of a hidden platform within six training trials of a daily session [71]. In this test, each mouse was given six 90 s training trials (ITI = 40 min) every day for 12 d, with the hidden platform placed in a novel location at the start of each day. The scores of each trial were averaged across the last 4 d of the 12-day training period. Behavioural data for escape latency were analysed using a two-way ANOVA. For the probe trials, statistical comparisons between genotypes for the number of crossings over the former platform location were done using one-way ANOVA with the critical α level set to 0.05 for all statistical analyses.
Swim paths of Neto1-null and wild-type mice in each trial of the second acquisition phase (Figure S7D and S7E) and delayed matching-to-place version of the Morris water maze task (Delayed Matching-to-Place [DMP] days 9–12, Figure S8A and S8B) were categorized according to their swim search strategies, as described [71,72]. Thigmotaxis: swimming along the edge of the wall or wall-hugging. Random search: randomly swimming over the entire area of the pool. Scanning: adopting a more systematic and efficient way of swimming in the central area of the pool. Chaining: memorizing a specific distance between the platform and the wall and swimming in wide circles to all possible platform locations at that distance. Focal search: restricted swimming to a specific area of the pool. Focal search signifies the beginning of spatial navigation and it could be separated into focal search in the correct target quadrant and focal search in the incorrect quadrants. The highest level of precision in spatial navigation is reached when the animal employs direct swims to the platform, independent of its release point. Swim strategies were characterized according to the predominant swim strategy used during the entire length of each trial and overall swim strategies were presented as the percentage of time spent on the strategy of choice. The experimenter classifying the swim search strategies was blind to the genotype or trial sequence within the experiment. The chaining parameter in the Wintrack computer software [36] was used to statistically verify qualitative swim search strategies of Neto1-null and wild-type mice during the second acquisition phase of the Morris water maze task and days 9–12 of the DMP task. The chaining score comparisons between genotypes were analyzed using ANOVA.
The modified open field procedure was performed as described [73], with slight modifications, using a second cohort of Neto1-null and wild-type littermate mice. The open field apparatus consisted of a cubical box (41 × 41 × 33 cm) made of clear Perspex (Ugo Basile) that was connected to horizontal and vertical infrared sensors. All behavioural events were video recorded and analyzed using Observer 5.0 software (Noldus Information Technology). The test consisted of four sessions with intertrial intervals of 2 min during which mice were returned to their home cage. During the open field session, each mouse was placed into the center of the empty, brightly lit open field for 5 min and the baseline level of locomotion (horizontal and vertical activity) and other behavioural parameters were recorded. The behavioural parameters were latency to escape the center; time of freezing (remaining in one place with only slight movement of the head); time of self-grooming; number of risk assessments (behaviour involving the mouse stretching its body from the corners/wall towards the center). Exploratory activity and walking were recorded separately for the central and peripheral field of the open arena, and the ratio between duration of central and peripheral activity was calculated.
During the habituation session, four different plastic objects were presented in the open field: cube (5 × 5 × 5 cm); hollow cylinder (6 cm height and 4 cm diameter); solid cylinder (3 cm height × 6 cm diameter); and prism (3.5 × 4.5 × 6 cm). Exploration of the four different plastic objects in the open field were measured every 5 min for 15 min under dim lighting (habituation profile). In the spatial object recognition session, the four objects, initially placed in a square arrangement, were reconfigured into a polygon-shaped pattern by moving two DOs. The remaining two objects were left at the same location (nondisplaced objects [NDOs]). Times of exploration of the DO and NDO were recorded for 5 min and expressed as a percentage of the total time of objects investigated. In the novel object recognition session, one of the familiar NDOs was replaced with a new object (NO) at the same location and the two familiar DOs were removed. The time examining a NO or a familiar object (FO) was recorded for 5 min and was expressed as a percentage of the total time of objects investigated. Data were analyzed with ANOVA with genotype as a between-subjects factor, and object rearrangement or object replacement as a repeated measures factor. The Tukey test was used for post hoc comparisons when ANOVA yielded statistically significant main effects or interactions.
To examine the effects of CX546 in spatial learning, new cohorts of Neto1-null and wild-type littermate mice were used for the water maze and displaced-object tasks. In the water maze task, a single daily intraperitoneal injection of CX546 (15 mg/kg, dissolved in 25% cylcodextran) or vehicle (25% cyclodextran) was administered 30 min prior to training. No injection was given on probe trial days. For the displaced-object task, a single intraperitoneal injection of CX546 (15 mg/kg) or vehicle was administered 30 min prior to displaced-object recognition testing.
All animal procedures were conducted in accordance with the requirements of the Province of Ontario Animals for Research Act, 1971 and the Canadian Council on Animal Care (CCAC 1984, 1995).
GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession numbers discussed in this paper are: PSD-95 (D50621); PSD-93 (AF388675); SAP-102 (D87117); and SAP-97 (NM_007862).
|
10.1371/journal.pcbi.1002353 | Residual Structures, Conformational Fluctuations, and Electrostatic Interactions in the Synergistic Folding of Two Intrinsically Disordered Proteins | To understand the interplay of residual structures and conformational fluctuations in the interaction of intrinsically disordered proteins (IDPs), we first combined implicit solvent and replica exchange sampling to calculate atomistic disordered ensembles of the nuclear co-activator binding domain (NCBD) of transcription coactivator CBP and the activation domain of the p160 steroid receptor coactivator ACTR. The calculated ensembles are in quantitative agreement with NMR-derived residue helicity and recapitulate the experimental observation that, while free ACTR largely lacks residual secondary structures, free NCBD is a molten globule with a helical content similar to that in the folded complex. Detailed conformational analysis reveals that free NCBD has an inherent ability to substantially sample all the helix configurations that have been previously observed either unbound or in complexes. Intriguingly, further high-temperature unbinding and unfolding simulations in implicit and explicit solvents emphasize the importance of conformational fluctuations in synergistic folding of NCBD with ACTR. A balance between preformed elements and conformational fluctuations appears necessary to allow NCBD to interact with different targets and fold into alternative conformations. Together with previous topology-based modeling and existing experimental data, the current simulations strongly support an “extended conformational selection” synergistic folding mechanism that involves a key intermediate state stabilized by interaction between the C-terminal helices of NCBD and ACTR. In addition, the atomistic simulations reveal the role of long-range as well as short-range electrostatic interactions in cooperating with readily fluctuating residual structures, which might enhance the encounter rate and promote efficient folding upon encounter for facile binding and folding interactions of IDPs. Thus, the current study not only provides a consistent mechanistic understanding of the NCBD/ACTR interaction, but also helps establish a multi-scale molecular modeling framework for understanding the structure, interaction, and regulation of IDPs in general.
| Intrinsically disordered proteins (IDPs) are now widely recognized to play fundamental roles in biology and to be frequently associated with human diseases. Although the potential advantages of intrinsic disorder in cellular signaling and regulation have been widely discussed, the physical basis for these proposed phenomena remains sketchy at best. An integration of multi-scale molecular modeling and experimental characterization is necessary to uncover the molecular principles that govern the structure, interaction, and regulation of IDPs. In this work, we characterize the conformational properties of two IDPs involved in transcription regulation at the atomistic level and further examine the roles of these properties in their coupled binding and folding interactions. Our simulations suggest interplay among residual structures, conformational fluctuations, and electrostatic interactions that allows efficient synergistic folding of these two IDPs. In particular, we propose that electrostatic interactions might play an important role in facilitating rapid folding and binding recognition of IDPs, by enhancing the encounter rate and promoting efficient folding upon encounter.
| It is now widely recognized that many functional proteins lack stable tertiary structures under physiological conditions [1]–[5]. Importantly, such intrinsically disordered proteins (IDPs) are highly prevalent in proteomes [6], play crucial roles in cellular areas such as signaling and regulation [7], [8], and are often associated with human diseases such as cancers [9]–[11]. The concept that intrinsic disorder can confer functional advantages has been discussed extensively [12]–[16]. For example, the disordered nature of IDPs could offer several unique benefits for signaling and regulation, including high specificity/low affinity binding, inducibility by posttranslational modifications, and structural plasticity for binding multiple partners. The last property appears to be particularly advantageous, and could support one-to-many and many-to-one signaling [16], [17]. Nonetheless, the physical basis of these proposed phenomena remains largely elusive. Specifically, how IDP recognition and regulation are supported by the interplay of residual structures, conformational fluctuations and other physical properties as encoded in the peptide sequence is poorly understood.
The current limit in mechanistic understanding of how intrinsic disorder supports function might be attributed to two key challenges in characterizing IDPs. These challenges are broadly shared by mechanistic studies of protein folding, misfolding, and aggregation in general [18]–[21]. The first one is related to the difficulty in deriving detailed structural information of the disordered unbound states [22]–[25]. In general, only ensemble-averaged properties can be measured for disordered proteins except with single-molecule techniques (which have their own limitations in spatial resolution, labeling need, and protein size [26]–[28]). Recovering the underlying structural heterogeneity using averaged properties is a severely underdetermined problem [29]–[33]. It is generally not feasible to construct a unique disordered structure ensemble that is consistent with the available data. This fundamental limitation leads to significant ambiguity in the current knowledge of the conformational nature of unbound IDPs. The second challenge is to further clarify the functional roles of any putative conformational sub-states or other properties of an IDP in its recognition and regulation (i.e., “function”). In particular, whereas some IDPs remain disordered in complexes [34], [35], many fold into stable structures upon binding to specific targets [36]. The roles of intrinsic disorder vs. residual structures in such coupled binding and folding interactions have been under much debate [36]. On one hand, residual structures have been observed frequently in unbound IDPs, and intriguingly, such residual structures often resemble those in the folded complexes [37]–[41]. These observations have led to an attractive hypothesis that preformed structural elements might provide initial binding sites to facilitate efficient recognition (i.e., conformational selection-like mechanisms) [12], [37]. On the other hand, evidence has accumulated in recent years, from computation as well as experimentation, to support a central role of nonspecific binding and emphasize the importance of disordered nature itself in promoting facile IDP recognition [41]–[51]. In fact, all published studies that extend beyond examining the unbound states alone have suggested induced folding-like mechanisms, at least at the baseline level.
Precisely how the disordered nature contributes to binding, however, is less clear. One proposal is that nonspecific binding of unstructured and presumably more extended conformations can increase the capture radii to enhance the binding kinetics [52], [53]; however, such “fly-casting” effects is small with a theoretical maximum of ∼1.6-fold acceleration. Recent studies have shown that unbound IDPs tend to be much more compact than previously assumed [54]–[57], further reducing the proposed fly-casting affects. In addition, the rate-enhancing affect due to increased size is likely offset by slower diffusion [58]. Alternatively, the unbound state of IDPs is presumed heterogeneous and strongly fluctuating. More specifically, conformational sub-states in the unbound IDPs should be marginally stable and separated by small free energy barriers (e.g., a few kcal/mol or less). These conformational fluctuations could contribute to efficient IDP recognition by allowing the peptide to fold rapidly upon (nonspecific) binding [50], [58], which is required for achieving the diffusion-controlled maximum binding rate (otherwise folding becomes rate-limiting) [59]. It should be noted that cellular events frequently modify the folding of IDPs to modulate their activities, such as through phosphorylations or by binding of other proteins [60]. Therefore, in contrast to globular proteins where folding often serves only to achieve the native structures, folding and unfolding appears to be direct and inherent aspects of IDP function. This underpins the importance and biological relevance of obtaining a mechanistic understanding of binding-induced folding of IDPs beyond a subject of theoretical curiosity.
The challenge in detailed characterization of IDPs represents a unique opportunity for molecular modeling to make critical contributions [5]. In particular, atomistic simulations could provide the ultimate level of detail necessary for understanding the structure and interaction of IDPs. At the same time, the dynamic and heterogeneous nature of IDPs also pushes the limits of both the force field accuracy and conformational sampling capability. So-called implicit solvent is arguably an optimal choice for de novo simulations of IDPs because of its necessary balance of accuracy and speed [61]–[64]. The basic idea of implicit solvent is to capture the mean influence of water by direct estimation of the solvation free energy, therefore reducing the system size about 10-fold. Important advances have been made to greatly improve the efficiency and achievable accuracy of implicit solvent, such as via the popular generalized Born (GB) theory [64]. With reduced system size, implicit solvent is also particularly suitable for replica exchange (REX) simulations [65]–[67], an enhanced sampling technique that has proven highly effective in sampling protein conformational equilibria [68]. Importantly, improved efficiency with implicit solvent also allows careful optimization to suppress certain systematic biases that have plagued explicit solvent approaches [69], [70]. For example, we have previously optimized the generalized Born with smooth switching (GBSW) model [71], [72] together with the underlying CHARMM22/CMAP protein force field [73]–[76]. The resulting GBSW protein force field not only recapitulates the structures and stabilities of helical and β-hairpin model peptides with a wide range of stabilities [77], [78], but also allows calculation of the conformational equilibria of small proteins under stabilizing and destabilizing conditions [79]–[81]. Although inherent and methodological limitations remain in implicit solvent [82], initial applications of implicit solvent to modeling small IDPs have been reasonably successful [41], [46], [55], [83]–[86], substantiating the notion that it is a viable approach for atomistic simulations of IDPs.
The current work focuses on the nuclear-receptor co-activator binding domain (NCBD) of the transcription coactivator CREB-binding protein (CBP) and its interaction with the p160 steroid receptor co-activator ACTR. CBP and its paralogue p300 are general transcriptional coactivators that play critical roles in transcriptional regulation and participate in cell cycle control, differentiation, transformation, and apoptosis [87], [88]. The NCBD domain (residues 2059–2117 in mouse CBP) is also known as interferon regulatory factor (IRF) binding domain (iBID) or the SRC1 interaction domain (SID). It mediates the interaction of CBP with a number of important proteins, including steroid receptor coactivators, p53 and IRFs [2], [89]. The interaction of CBP with p160 coactivators in particular is important for recruitment of CBP/p300 to transmit the hormonal signal to the transcription machinery [90]. Besides the biological and medical significance, the NCBD/ACTR interaction also offers unique opportunities for understanding the molecular principles of IDP recognition. Both NCBD and the activation domain of ACTR that it interacts with (residues 1018–1088 in human ACTR; hereafter referred to as ACTR) are IDPs. Their interaction is an example of the “synergistic folding” mechanism [91] (the other known example also involves NCBD, but with the p53 transactivation domain, TAD [92]). In addition, four folded structures of NCBD have been solved in complex with various protein targets besides ACTR [91]–[94]. In these complexes, NCBD adopts two distinct tertiary folds that involve three similar helices, represented by the NCBD/ACTR and NCBD/IRF3 complexes (see Figure 1). Therefore, NCBD represents one of the few experimentally validated examples of structural plasticity, which is believed to be a key functional advantage of intrinsic disorder [16].
Interestingly, although free ACTR is largely devoid of residual structures, free NCBD contains one the highest levels of residual structures with folded-like helical content and molten globule characteristics [95], [96]. In addition, even though nuclear magnetic resonance (NMR) relaxation analysis has established that free NCBD is highly dynamic on picosecond (ps) to nanosecond (ns) timescales [96], it appears to have a strong tendency to adopt marginally stable tertiary folds, allowing two NMR structures of the unbound state determined to date [40], [89]. These structures are presumably obtained by stabilizing various conformational sub-states under specific solution conditions. Particularly intriguing is that the latest NMR structure of free NCBD turns out to be similar to the folded conformation observed when bound to ACTR [40]. Although such pre-existence of folded-like conformations should be considered only as a necessary but insufficient condition for conformational selection-like mechanisms, the unusually high level of residual structures of NCBD strongly suggests a functional role of pre-folding in its coupled binding and folding interactions. In this work, we first exploit implicit solvent-based atomistic simulations and REX enhanced sampling to characterize the conformational properties of free NCBD and ACTR. The roles of preformed structures vs. conformational fluctuation in the NCBD/ACTR interaction are then directly probed using high-temperature unfolding and unbinding simulations in both implicit and explicit solvents. Combined with our recent coarse-grained simulations and existing experimental data, we aim to obtain a detailed mechanistic picture of how residual structures, conformational fluctuations, and electrostatic interactions contribute to efficient synergistic folding of NCBD and ACTR.
De novo calculation of the disordered ensembles for IDPs is challenging [5], especially for NCBD that is both of moderate size and apparently with a complex, solution condition-sensitive conformational equilibrium. Our previous works have suggested that implicit solvent coupled with REX enhanced sampling could generate reasonably accurate disordered ensembles for small IDPs, including a 28-residue segment of the kinase inducible domain (KID) of transcription factor CREB [55]. In Figure S1, we first test the convergence of the calculated disordered ensembles by examining the dependence of residue helicity on REX simulation time and by comparing results from independent simulations initiated from dramatically different conformations (folding vs. control; see Methods). The sequences of both domains are provided in Methods. Free ACTR appears to be highly disordered with marginal residual helicity. The calculated residual helicity profiles from the control and folding runs converge to similar ones (data not shown). For NCBD, while the time evolution of the calculated residual helicity appears to stabilize over the course of 100 ns in either the control or folding REX simulation, the final profiles from these two independent calculations differ substantially, suggesting that the actual convergence is rather limited. Nonetheless, both the folding and control simulations clearly suggest significant residual helicity in all three helical segments that become stably folded upon binding to various specific targets. Detailed analysis of the conformational ensemble (see below) demonstrates that free NCBD is compact and contains substantial tertiary contacts. These conformational properties of NCBD, coupled with the larger size, contribute to the difficulty of achieving better convergence using the REX/GB protocol. In addition, the current surface area-based treatment of nonpolar solvation can over-stabilize non-specific collapsed states [82], [97]. This problem further limits the ability to sufficiently sample accessible tertiary organizations of free NCBD and their inter-conversions, which is required for achieving good convergence.
Given the limited convergence achieved in the REX simulations of free NCBD and apparent difficulties in substantially improving the level of convergence, we focus on semi-quantitative or qualitative analysis of the conformational properties of NCBD. That is, although significant conformational sub-states sampled by REX may be genuine, the relative stability (population) is not likely to be reliable. Considering that NCBD is experimentally known to be highly helical, the folding simulations (initiated with a fully extended conformation) should take longer to converge, and the disordered ensemble calculated from the control simulation is likely more realistic. Therefore, all the subsequent analysis is based on the ensemble of conformations sampled during the last 60 ns of the 100 ns control REX simulation. In Figure 2, we compare the residue helicity of NCBD and ACTR in the free and bound states. The results appear to be fully consistent with the previous NMR secondary chemical shift analysis (Figure 2 of Ref. [96]), showing that all three NCBD helices are largely formed in the unbound state and ACTR is largely free of residual helices. Interestingly, the poly-Q segment of NCBD (residues 2082–2086), although disordered in the NCBD/ACTR complex, is largely helical in the unbound state and extends Cα2. This is fully consistent in the NMR chemical shift analysis [96]. Recent sequence correlation analysis has revealed a link between sequence order and binding promiscuity [98], [99]. One might expect that the length of the poly-Q stretch might affect conformational flexibility, and furthermore, the ability to interact with diverse targets. We also have analyzed the ensemble distribution of the radius of gyration of free NCBD. The results, shown in Figure S2A, confirm that free NCBD is highly compact. Despite a clear lack of convergence, the control and folding simulations appear to sample a set of conformation sub-states with similar characteristic sizes. Direct comparison of the calculated size profiles to one derived from a recent small-angle X-ray scattering (SAXS) study [40] is complicated by the different constructs used and uncertainty in proper inclusion of the solvation shell for a heterogeneous ensemble. Nonetheless, one can estimate that including the disordered N- and C-terminal tails (13 residues total) truncated in the current simulations would increase the radius by 2–3 Å, and that the solvation shell may add another 2–3 Å (estimated by comparing results from HydroPro [100] and CHARMM). These corrections together bring the calculated radius of the gyration profile close to the SAXS-derived profile that centers around 15.2 Å under “native-like” conditions [40]. Apparent agreement between NMR and SAXS on these ensemble-averaged properties is not sufficient to validate the reliability of the simulations, but it suggests that the simulated ensemble may offer a qualitative or even semi-quantitative characterization of the conformational properties of free NCBD.
Because all three NCBD helices are largely formed in the unbound state, the conformational fluctuation of free NCBD mainly involves tertiary packing of these helices. For example, as shown in Figure 1, when aligned using the central helix Cα2, the two representative folded conformations of NCBD differ mainly in the orientation of Cα1 and slightly less so in that of Cα3. Therefore, all conformations of the calculated ensemble first re-oriented by aligning Cα2 (to the −z axis) before the orientations of Cα1 and Cα3 were calculated. Note such analysis also provides an effective description of the tertiary packing even when one or more of the three NCBD segments are not in helical states. The results, shown in Figure 3, illustrate that NCBD is strongly fluctuating and samples a large number of helix configurations, as expected for a molten globule. Intriguingly, free NCBD appears to substantially sample all three distinct conformations that have been observed experimentally so far, either in complexes or in isolation. These folds are represented by PDB structures 1kbh, 1zoq, and 1jjs, respectively. The Cα1 orientation of 1kbh and Cα3 orientation of 1jjs appear to be least sampled. Nonetheless, conformational sub-states exist with similar orientations, as marked by arrows in panels c) and d) of Figures 3. Specifically, for 1kbh-like Cα1 orientation, the adjacent sub-state contains more parallel (with smaller helix cross angles), and thus tighter, packing of Cα1 with Cα2, but with a helix interface similar to that of 1kbh. Further structural analysis (see the following paragraph) suggests that such tighter packing is likely a result of helix formation in the poly-Q segment (e.g., see Figure 2), which shortens the Cα1-Cα2 loop and promotes tighter packing.
Clustering analysis was performed to further analyze the structural properties of the major conformational sub-states of free NCBD. The average structures of the six most populated clusters identified using K-means clustering with a 3.0 Å radius are shown in Figure 4. Helix configurations for all members of these clusters are shown in Figure S3. Interestingly, even though one of the clusters (Figure 4D) is similar to the fold observed in 1kbh, most clusters are different from either 1zoq or 1kbh on the whole domain level, as suggested by the large RMSD values. Therefore, even though both individual Cα1-Cα2 and Cα2-Cα3 helix pairs sample all three distinct PDB folds, these folded-like configurations of individual helix pair generally do not occur at the same time. Notably, the folded conformations of NCBD in 1kbh and 1zoq have relatively similar Cα2-Cα3 helix packing (see Figure 1C). The packing of Cα2 and Cα3 also appears to be more restricted in free NCBD compared to that of Cα1 and Cα2 (e.g., as indicated by a larger “inhibited” red area in Figure 3D compared to Figure 3C). NCBD has a strong inherent propensity to adopt Cα2-Cα3 configurations analogous to those in 1kbh and 1zoq. Such persistent folded-like conformations of free NCBD could contribute to recruitment of specific targets such as ACTR and IRF3, allowing NCBD to adopt different final structures by docking the more flexibly linked Cα1 into different positions. Another interesting observation is that the poly-Q segment appears to be capable of readily switching between helical and coil states. Such conformational fluctuations could allow NCBD to adapt to different substrates, extending the Cα2 helix when bound with IRF3 but becoming more disordered when in complex with ACTR (see Figure 1).
Although the REX simulations provide intriguing insights into the possible residual structures of free NCBD, how these conformational properties contribute to synergistic folding of NCBD with ACTR is not obvious based on these equilibrium simulations alone. For this, one could calculate the coupled binding and folding free energy surfaces [41], [51] or transition paths [46] to more directly clarify the recognition mechanism and probe the roles of residual structures vs. pre-folding in specific finding. However, given the moderate size and relatively complex topology, such calculations can be extremely demanding using an atomistic physics-based force field for the NCBD/ACTR complex. Instead, temperature-induced unfolding and unbinding simulations may be used to effectively infer the molecular processes of coupled binding and folding. A key assumption is that binding/folding is largely a reverse of unbinding/unfolding. An important concern is that the transition states or the most probable transition paths might depend on temperature [101]. Nonetheless, high-temperature unfolding simulations have so far proven quite successful for studying folding and interaction of many proteins, including IDPs [44], [102]–[104].
A 100 ns equilibrium simulation of the complex was first performed at 300 K, which confirms that the native fold (model 1 of PDB:1kbh) is very stable in the GBSW/MS2 implicit solvent (see Figure S4). Subsequent pilot simulations suggest 475 K to be optimal for simulating unbinding and unfolding of the NCBD/ACTR complex in GBSW/MS2 (e.g., see Figure S5). In Figure 5, we compare the time evolution of various fractions of native contacts computed from 50 independent unfolding simulations at 475 K. The fraction of native intermolecular interactions (Qinter) is used to describe binding, and the fraction of native tertiary intramolecular interactions (QNCBD) is used for folding of NCBD. As shown in Figure S6, ACTR is completely devoid of any inter-helix tertiary contacts in the NCBD/ACTR complex. Because ACTR is largely free of residual structures in the unbound state, the overall helicity (αACTR) is used to effectively monitor (binding-induced) folding of ACTR. On the baseline level, all unfolding and unbinding kinetics appear to be reasonably well represented by single exponential functions. The fitted kinetic data is summarized in Table 1. The secondary (helix) unfolding of NCBD is predicted to be the slowest process (αNCBD; green traces in Figure 5), which is expected given the high level of residual structures in unbound NCBD; however, both the ACTR (helix) and NCBD tertiary unfolding appear to be significantly faster than unbinding. This result suggests that binding occurs prior to the folding of both ACTR and NCBD; that is, both ACTR and NCBD follow induced-folding-like mechanisms on the baseline level in the GBSW/MS2 implicit solvent. Considering the apparent tendency of NCBD to pre-fold (see above), this result is somewhat surprising, but it highlights the importance of conformational fluctuations and nonspecific binding in specific recognition of IDPs, even for IDPs with significant residual structures like NCBD. Significant heterogeneity is apparent in the unfolding/unbinding pathways of NCBD/ACTR and is partially reflected in substantial ruggedness that remains in the curves shown in Figure 5 (e.g., compared with a previous explicit solvent unfolding simulation of the p53-MDM2 complex, where 10 10-ns simulations at 498 K were sufficient to yield much smoother curves [103]). The complex fully disassociates within 10 ns in only 6 out of the 50 independent runs. In examining the unbinding/unfolding characteristics at a lower temperature of 450 K (see Figure S7), we found the heterogeneity of unfolding/unbinding pathways to be even more evident. In addition, the complex appears trapped in some intermediate states and does not fully unfold/unbind even after 20 ns. Nonetheless, unfolding of either ACTR or NCBD appears to lag behind unbinding, which is consistent with the induced-folding baseline mechanisms predicted at 475 K.
Indications are that binding-induced folding of NCBD and ACTR is not simply 2-state-like. For example, decay of QNCBD and αACTR appears to pause at ∼2 ns (red and blue traces in Figure 5), which could suggest a common intermediate state where ACTR and NCBD are partially bound and folded. The decay curves are too noisy (partially due to underlying heterogeneity) for reliable kinetic fitting using double exponential functions. Therefore, we constructed (pseudo) unbinding and unfolding free energy surfaces based on statistics collected from the first 5 ns of the unfolding simulations. Note that the system is not at equilibrium during this time frame, so the resulting free energy profiles are not equilibrated (and thus strongly dependent on initial conditions). Nonetheless, the profiles provide qualitative approximations of the true free energy surfaces [105]. As shown in Figure 6A, an intermediate state is evident at Qinter∼0.25 and QNCBD∼0.15. Interestingly, a similar key intermediate state also has been predicted in our recent topology-based modeling of the NCBD/ACTR complex [50]. A strong resemblance between the free energy surface is shown in Figure 6A and the result derived from topology-based modeling (Figure 5A of reference [50]). Both the atomistic simulations (see further analysis detailed in the following paragraph) and topology-based modeling predict that the intermediate state mainly involves the C-terminal segments of NCBD and ACTR. Such a prediction appears highly consistent with a recent H/D exchange mass spectrometry (H/D-MS) study [106], where peptide segments within the C-terminal regions of both NCBD and ACTR were found to have much larger protection factors compared with those mapped into other folded regions of the complex.
In Figure 7, we further examined the binding kinetics of individual NCBD and ACTR helices. The kinetic data derived from fitting to single exponential functions is summarized in Table 1. The analysis shows that Aα3 and Cα3 unbind with the largest half times, τ = 2.93 ns and 2.20 ns, respectively, which are greater than that of the overall intermolecular interaction formation (τ = 1.61 ns). This result indicates that binding is mainly initiated by the C-terminal helices. In contrast, the first helices of NCBD and ACTR unbind much faster then the second and third helices. In fact, unbinding of Aα1 and Cα1 occurs even faster than folding of either NCBD or ACTR (as described by QNCBD and αACTR, see Table 1). These kinetic rates are consistent with a multi-stage synergistic folding process, where NCBD and ACTR first bind rapidly through the C-terminal segments, forming intermediates that are mainly stabilized by native-like interactions between α2 and α3 helices. This first step appears to be highly cooperative (e.g., see Figure 6A), although indications are that both induced folding and conformational selection might contribute [50]. Interestingly, the transition between the intermediate and bound states appears largely conformational selection-like where NCBD and ACTR folding precedes Aα1 and Cα1 binding. Formation of the partially folded core appears to facilitate the rest of NCBD to fold into native-like conformations, allowing Cα1 and Aα1 to rapidly form native intermolecular interactions en route to the fully folded bound state. Taken together, even though the synergistic folding of NCBD and ACTR follows an induced folding-like baseline mechanism (where binding precedes folding on the overall level), detailed analysis reveals multiple stages of induced folding and conformational selection. Such a mechanism closely resembles an “extended conformational selection” recently proposed by Csermely et al. [107], [108] and is remarkably consistent with our recent topology-based modeling of the NCBD/ACTR complex [50].
One of the most notable features of the NCBD/ACTR complex is a buried salt-bridge between NCBD R2105 and ACTR D1068 [91] (see Figure 1A), which is also conserved in the interaction of NCBD with p53 TAD [92]. Interestingly, this buried salt-bridge is part of a local network of salt-bridges that could form between multiple complementary charges, including R2105 and K2108 of NCBD and D1060, E1065, and D1068 of ACTR (see Figure 1A). This network of native and non-native salt-bridges appears to play a significant role in stabilizing the putative intermediate state, either thermodynamically or kinetically. Although most individual salt-bridges frequently break and reform during individual unfolding simulations (see Figure S8), on average they largely persist throughout the 10 ns unfolding simulations at 475 K and hinder the transition from the partially bound intermediates to fully disassociated ones (see Figure 8). Out of the 50 unfolding simulations at 475 K, the complexes fully dissociate only by the end of 10 ns simulations in six cases. The native salt-bridges, between NCBD R2105 and ACTR D1068 and D1060, are the most protected. As shown in Figure 8, they are the most preserved and remain formed over 80% of the time throughout the simulations (blue and black traces in Figure 8A). NCBD K2108 is adjacent to R2015 and close enough to interact with ACTR D1068 and D1060, but these salt-bridges are more solvent-exposed and thus slightly less preserved during high-temperature simulations. The side chain of ACTR E1065 is positioned away from NCBD in the native structure. Partial unfolding of Aα2 allows E1065 to rotate and participate in the salt-bridge network with 10–30% probability by the end of the 10 ns simulation at 475 K (purple and red traces in Figure 8A).
The conformational heterogeneity of the intermediate state does not permit reliable free energy calculations to quantify the contribution of salt-bridge interactions to stability. Nonetheless, previous mutagenesis studies have suggested that the buried salt-bridge between NCBD R2105 and ACTR D1068 contributes minimally to binding affinity [95]. The salt-bridge network likely could not significantly stabilize the intermediate state thermodynamically, either, which raises a concern that the observed persistence of the local salt-bridge network is artificial, such as due to over-stabilization of charge-charge interactions in the GBSW/MS2 implicit solvent. To address this concern, we first examine the potential of mean forces (PMFs) between Arg and Asp side chain analogs in TIP3P and GBSW/MS2. The results, summarized in Figure 8, show that GBSW/MS2 actually slightly under-stabilizes the Arg-Asp interaction compared with TIP3P, either in a constrained head-to-head configuration (which was used in the force field optimization [72]) or when fully unconstrained. In particular, configurationally unconstrained Arg-Asp interaction is unstable in GBSW/MS2 (Figure 9B). Therefore, the observed stabilization effects of salt-bridges on the intermediates are likely of a kinetic nature. Such kinetic stabilization arises from substantial desolvation barriers in disassociation of salt-bridges, particularly in partially folded protein environments where the side chain configurations are restricted (e.g., see Figure 9A). With a concentrated local network of salt-bridges, very large desovaltion barriers can be expected for complete dissociation of NCBD and ACTR, which explains why only a small fraction of the high-temperature simulations (6 out of 50) successfully reached the fully unbound state in 10 ns.
To further confirm that the observed salt-bridge network is not an artifact of implicit solvent, a set of 10 unfolding simulations was performed in TIP3P explicit solvent at 500 K. Most simulations were terminated between 3 to 4 ns when the complex size exceeded the periodic box dimensions. The lengths of these simulations are insufficient to capture degrees of unfolding and unbinding similar to implicit solvent simulations, and the number of trials is insufficient to obtain smooth curves for kinetic fitting. Nonetheless, visual inspection of simulation trajectories as well as examination of the evolution of various contact fractions support an unbinding and unfolding mechanism that is consistent with the one derived from implicit solvent simulations (see Figure S9). The same set of native and non-native interactions, particularly the buried one between NCBD R2105 and ACTR D1068 (blue trace in Figure S9B), persist and appear to stabilize the partially unbound and unfolded intermediates. Note that the helical secondary structures are substantially over-stabilized in these explicit solvent simulations (e.g., see the blue trace in Figure S9A). This is a known artifact of the current version CHARMM22/CMAP explicit solvent force field [78], [109], [110].
A control simulation of the double-Leu mutant complex, NCBD:R2105L/ACTR: D1068L, at 300 K suggests that the native fold remains stable in the GBSW/MS2 implicit solvent (data not shown). A set of 50 unfolding simulations was carried out at 450 K to further investigate the role of the buried salt-bridge in synergistic folding. The heterogeneity of the unfolding/unbinding pathway observed in the wild-type complex (e.g., see Figure 5) is even more pronounced without the buried salt-bridge. All averaged time traces of contact fractions remain very noisy (e.g., see Figure S10). Most traces cannot be satisfactorily fitted to either single or double exponential functions, preventing quantitative analysis of unfolding and unbinding kinetics. Nonetheless, the pseudo binding and folding free energy surface computed from the first 5 ns of the unfolding trajectories appears to resemble that from simulations of the wild-type complex (see Figure 6). In particular, a similar intermediate state exists at Qinter∼0.2 and QNCBD∼0.15; however, the small free energy barrier separating the intermediate and fully unbound states in Figure 6A is largely absent in Figure 6B. Removal of NCBD:R2105L largely disrupts the local salt-bridge network. The intermediate state appears to have much shorter resident times, and can quickly fluctuate to the fully unbound state. Importantly, examination of the evolution of intermolecular contact factions of individual NCBD and ACTR helices, shown in Figure S10, supports that the mutant complex largely follows a similar, albeit more heterogeneous, unbinding and unfolding mechanism, with the N-terminal α1 helices disassociated first (black traces in Figure S10B–C). These results suggest the local salt-bridge network does not appear to fundamentally modulate the recognition mechanism. Instead, it mainly augments a productive synergistic folding mechanism inherent in (the topology of) the NCBD/ACTR complex, by transiently stabilizing a key on-pathway intermediate state to facilitate complete folding en route to the specific complex.
With one of the highest levels of residual structures, NCBD is an intriguing model system for understanding the roles of residual structure vs. conformational fluctuations in coupled binding and folding of IDPs. We have combined equilibrium and non-equilibrium simulations using physics-based, atomistic protein force fields to characterize the conformational properties of unbound NCBD and ACTR and to understand how these properties facilitate efficient synergistic folding of these two IDPs. The calculation recapitulates that free NCBD has folded-like helical content, is strongly fluctuating, and samples a wide range of tertiary configurations, which is consistent with the previous notion that free NCBD is a molten globule [96]. Intriguingly, the calculated disordered ensemble of NCBD contains significant populations with helical packings that are highly similar to all those previously observed experimentally in isolation and in complex with various targets. Observations of such pre-folded conformations, especially for IDPs with significant residual structures like NCBD, could be considered strong evidence for conformational selection-like mechanisms, where such preformed structural elements provide initial binding sites. Direct examination of the unfolding and unbinding pathways in high-temperature simulations, however, shows that both ACTR and NCBD tend to unfold first before unbinding, suggesting an induced folding-like baseline mechanism for their synergistic folding. This seemingly surprising result appears to be consistent with the observation that, although individual Cα1/Cα2 and Cα2/Cα3 helical pair samples folded-like packing with substantial probability, these configurations rarely occur simultaneously. Therefore, population of folded-like tertiary conformations on the whole domain level is insufficient to support conformational selection-like mechanisms on the baseline level. Further analysis reveals an on-pathway intermediate state that mainly involves the C-terminal helices of ACTR and NCBD, which also has been predicted by a recent coarse-grained simulation study using topology-based models [50]. Importantly, existence of such a major intermediate state also appears to be consistent with a recent H/D-MS experiments showing that peptide segments within the C-terminal regions of NCBD and ACTR have much larger protection factors compared with those mapped into other regions of the complex [106]. Our kinetic analysis suggests that, once the initial mini folding core is formed, the N-terminal helix of NCBD folds rapidly (Table 1), allowing subsequent facile binding and folding the ACTR N-terminal helix en route to the final specific complex. Therefore, although the baseline mechanism is induced folding-like, conformational selection actually occurs at local levels. Together with our recent topology-based modeling study [50], the atomistic simulations strongly support the prediction that synergistic folding of NCBD and ACTR follows the “extended conformational selection” mechanism [107]. Our topology-based modeling of the NCBD/ACTR interaction [50] has revealed a separate, albeit less prevalent, pathway where binding is initiated by the N-terminal α1 helices. These mechanistic insights on synergistic folding of NCBD and ACTR, derived from the atomistic and coarse-grained simulations, are summarized in Figure 10.
An intriguing interplay appears to exist among residual structures, conformational fluctuations, and electrostatic interactions to facilitate the rate-limiting step of forming the partially folded intermediates. The NCBD Cα2/Cα3 helix-turn-helix motif appear to be conformationally more restricted (Figure 2D), whereas the C-terminus of Cα3 retains the least amount of helical content and is considerably more heterogeneous (Figure S2B). Both features were also observed in the previous NMR chemical shift and relaxation analysis [96]. Such a balance of residual structures and conformational fluctuations is likely important for the NCBD C-terminal to act as a key initiation point for coupled folding and binding to ACTR and other proteins. Another novel insight provided by the current atomistic simulations is the role of a local network of native and non-native salt-bridges in transiently stabilizing the intermediates. These salt-bridge interactions likely do not contribute substantially to the thermodynamic stability of either the intermediates or the final specific complex [95], but substantial desolvation barriers involved in breaking up these interactions in a conformationally restricted protein environment (e.g., Figure 9A) can extend the resident time of the intermediates to allow the rest of the complex to fold with higher efficiency. As demonstrated using a dual-transition state kinetic model [59], efficient folding upon encounter is necessary for achieving facile binding at or near the diffusion-limited basal binding rate, a highly desirable property for signaling and regulatory IDPs that need to constantly evade protein degradation machinery in cell. IDPs are known to be enriched with charges [6]. NCBD and ACTR are no exceptions, with +6 and −8 net charges, respectively (including the flanking loops that remain disordered in the complex [91]). These enriched charges hinder (independent) folding and can protest against aggregation. In addition, long-range electrostatic interactions between these large numbers of complementary charges on NCBD and ACTR could dramatically enhance the encounter rate, similar to electrostatic steering, which is known to be important in interactions of globular protein [111]. Furthermore, the complementary pattern of charge, especially within the predicted mini folding core involving the C-termini (Figure 1), suggests that long-range electrostatic interactions could further promote folding-competent encounter complexes before transiently stabilizing the on-pathway intermediates via formation of short-range salt-bridge network. These effects can enhance the efficiency of folding upon encounter to promote facile recognition.
The current study also reveals important limitations in both the protein force field accuracy and sampling capability, especially for modeling IDPs of moderate sizes and with complex residual structures. These limitations underscore the importance of continual development of the protein force field, with increased focus on balancing various competing interactions to allow an accurate description of not only a few (native) folds but also the whole conformational equilibrium [82], [112]. Sampling methodologies clearly need to improve. The standard temperature REX-MD has failed to achieve convergence for the disordered ensemble of NCBD within 100 ns. Besides limited simulation timescale, certain limitations of the implicit solvent protein force field also contributed. In particular, current empirical protein models have been shown to contain a systematic bias to over-stabilize protein-protein interactions [113], [114]. Furthermore, simple surface area-based estimation of the nonpolar solvation free energy employed in most current implicit solvent models also tends to over-stabilize nonspecific compact protein states [82]. The standard temperature REX-MD clearly has limited ability to sample alternative deeply trapped low energy states with high efficiency. These limitations together have also prevented us from more directly investigating the proposed mechanistic roles of electrostatic interactions using atomistic simulations. Despite these outstanding limitations, the key mechanistic features derived from atomistic physics-based simulations, coarse-grained topology-based modeling, and various biophysical measurements are remarkably consistent, which suggests that an integration of multi-scale modeling and experimentation can provide a viable approach for understanding the functional and control of IDPs.
Only segments of the NCBD and ACTR domains that are structured in the complex are included in the current simulations, which include residues 2066–2112 for NCBD (in mouse CBP numbering; SALQD LLRTL KSPSS PQQQQ QVLNI LKSNP QLMAA FIKQR2105 TAKYV AN) and residues 1040–1086 for the ACTR domain (in human ACTR numbering; E GQSDE RALLD QLHTL LSNTD ATGLE EID1068RA LGIPE LVNQG QALEP K). The peptide termini are neutralized using with either acetyl (Ace) or amine (NH2) groups. A previously optimized GBSW/MS2 model was used in all implicit solvent simulations unless otherwise noted [72]. This model adopts an effective approximation of the molecular surface for defining the solute-solvent boundary, which is believed to be more physical compared to the van der Waals-like surface used in the original GBSW model [115], [116]. Importantly, the GBSW/MS2 model has also been carefully optimized to balance solvation and intramolecular interactions and can reasonably capture the competition between α and β secondary structures. Specifically for NCBD/ACTR, the structure of the complex (PDB: 1kbh [91]) remains stable in the GBSW/MS2 force field for over 100 ns, but substantially deviates from the native conformation in the original GBSW protein force field (see Figure S4).
REX was used to enhance the sampling of the accessible conformational space of free NCBD and ACTR. For this, the Multiscale Modeling Tools for Structural Biology (MMTSB) toolset [117] (http://www.mmtsb.org) was used in conjunction with CHARMM [118], [119]. The basic idea of REX is to simulate multiple non-interacting replicas at different temperatures simultaneously. Periodically, one attempts to exchange the simulation temperatures between pairs of replicas based on a Metropolis criterion derived from the detail balance principle. As such, not only the resulting random walk in the temperature space facilitates the system to cross the energy barriers and exploit the conformational space more efficiently, but proper canonical ensembles are also generated at all temperatures, allowing direct calculation of thermodynamic properties for comparison with experiments. We performed two independent REX simulations for each peptide, initiated from the folded structure extracted from the complex (control) and a fully extended conformation (folding), respectively. Comparison of the calculated structure ensembles from these independent control and folding runs with dramatically different initial conditions allows rigorous assessment of the convergence. In each REX simulation, 16 replicas were simulated at temperatures exponentially distributed from 270 to 500 K. SHAKE [120] was applied to fix the lengths of all hydrogen-related bonds, allowing a 2.0 fs molecular dynamics (MD) time step. Temperature exchanges between neighboring replicas were attempted every 2 ps, and the total length of each REX simulation was 100 ns (50,000 REX cycles). Similar REX/GBSW protocols have proven effective in calculating the disordered structural ensembles for other IDPs (albeit of smaller sizes than NCBD and ACTR studied in the current work) [41], [55]. All analysis was performed based on the conformations sampled during the last 60 ns of the control simulation at 305 K (where most existing experimental data were acquired), unless otherwise noted. The orientations of helical segments (1044–1058, 1063–1071, 1072–1080 in ACTR; 2067–2076, 2086–2091, 2095–2110 in NCBD) were calculated using the Chothia-Levitt-Richardson algorithm [121] as implemented in CHARMM. The K-means clustering algorithm as implemented in the MMTSB toolset was used to cluster the calculated disordered ensembles based on mutual Cα RMSD distances. Various clustering radii ranging from 1.5 to 4.5 Å were tested before an optimal radius of 3.0 Å was used for the final clustering results presented. All molecular visualizations were generated using the VMD software [122].
The same peptide segments defined above were included the simulations of the complex. The model 1 from the NMR ensemble (PDB: 1kbh) was first equilibrated in the GBSW/MS2 implicit solvent using energy minimization and short MD with weak harmonic positional restraints imposed on all backbone heavy atoms. Subsequently, a 160 ns unrestrained simulation was performed at 300 K to examine the structural stability and dynamics of the complex near its native basin. The native structure of the NCBD:R2105L/ACTR:D1068L double-Leu mutant complex was prepared by computational mutagenesis and then equilibrated using a similar protocol as described above. To identify the optimal temperatures for unbinding/unfolding simulations, a series of pilot simulations was performed at temperatures ranging from 350 K to 500 K (e.g., see Figure S5). At the optimal temperature, the complex should unfold/unbind within tractable time scales (e.g., 10–20 ns) while retaining important details of the unfolding/unbinding pathways. Once such optimal temperatures were chosen (450–475 K for the wild-type and 450 K for the mutant), 50 independent high-temperature simulations of 10–20 ns in length were initiated from the equilibrated native structures with different initial velocities. The results presented in this work are averages computed from 50 unfolding simulations unless otherwise noted. For native fraction analysis, a list of native tertiary contacts (shown in Figure S6) was first identified using the equilibrated native structure based on side chain minimal heavy atom distances with a 4.2 Å cutoff. The native contacts were then divided into inter-molecular and intra-molecular categories. In analysis of the high-temperature simulation trajectories, a contact was considered formed when the minimal heavy atom distance between two side chains was no greater than 4.5 Å. Helicity of various helical segments was calculated based on the hydrogen bonding patterns using the COOR SECS module of CHARMM.
Additional high-temperature unfolding and unbinding simulations of the wild-type complex were performed in TIP3P water to examine the unfolding/unbinding pathway and in particular the putative role of the buried salt-bridge between NCBD:R2105 and ACTR:D1068 in (transiently) stabilizing the intermediate state(s). For this, the equilibrated NCBD/ACTR complex was placed in a cubic water box with periodic boundary conditions imposed. The final solvated system contains 9176 TIP3P water molecules and the box size is ∼65 Å. Two potassium ions were added to neutralize the total charge. The proteins were described by the CHARMM22/CMAP protein force field [73]–[76]. The particle mesh Ewald method was used for long-range electrostatic interactions [123], and the van de Waals interactions were smoothly switched off from 12 to 13 Å. Lengths of all hydrogen-related bonds were kept constant with SHAKE [120], and the MD time step was 2 fs. After 10 ps of NPT equilibration at 300 K, a set of 10 independent NVT productions was carried out at 500 K up to 10 ns until the dimensions of the proteins exceed those of the periodic box. The dynamic time step was reduced to 1 fs in the NVT production simulations for numerical stability.
An umbrella sampling protocol [77] was used to compute the PMFs between the side chains of Asp and Arg, either constrained in a head-to-head configuration [77] (see Figure 9) or allowed to freely rotate. In the constrained setup, the side chains were allowed to move only in fixed orientations along the reaction coordinate (indicated by a dashed line in Figure 9), enforced using the MMFP module in CHARMM. For explicit solvent simulations, solutes were solvated by either ∼710 TIP3P waters in a rectangular box (for the constrained PMF) or by ∼1040 TIP3P waters in a truncated octahedral box (for the unconstrained PMF). Periodic boundary conditions were imposed. Non-bonded and other setups are identical to those described above for explicit solvent high-temperature simulations. Harmonic restraint potentials were placed every 0.5 Å along the reaction coordinate with a force constant of 5.0 kcal/mol/Å2. For each umbrella-sampling window, the system was first equilibrated for 60 ps, followed by 2 ns (constrained PMF) or 4 ns (unconstrained PMF) NPT production at 300 K and 1 atm. The final PMFs were calculated using the weighted histogram analysis method (WHAM) [124]. The constrained PMF in GBSW/MS2 was computed by direct translation of the side chains along the reaction coordinate, and the unconstrained PMF in GBSW/MS2 was computed in the same umbrella sampling protocol except that implicit solvent was used instead of TIP3P waters. Convergence of the PMFs was examined by comparing results from the first and second halves of the data and was shown to be on the order of 0.2 kcal/mol.
|
10.1371/journal.pbio.1001647 | Analysis of the RelA:CBP/p300 Interaction Reveals Its Involvement in NF-κB-Driven Transcription | NF-κB plays a vital role in cellular immune and inflammatory response, survival, and proliferation by regulating the transcription of various genes involved in these processes. To activate transcription, RelA (a prominent NF-κB family member) interacts with transcriptional co-activators like CREB-binding protein (CBP) and its paralog p300 in addition to its cognate κB sites on the promoter/enhancer regions of DNA. The RelA:CBP/p300 complex is comprised of two components—first, DNA binding domain of RelA interacts with the KIX domain of CBP/p300, and second, the transcriptional activation domain (TAD) of RelA binds to the TAZ1 domain of CBP/p300. A phosphorylation event of a well-conserved RelA(Ser276) is prerequisite for the former interaction to occur and is considered a decisive factor for the overall RelA:CBP/p300 interaction. The role of the latter interaction in the transcription of RelA-activated genes remains unclear. Here we provide the solution structure of the latter component of the RelA:CBP complex by NMR spectroscopy. The structure reveals the folding of RelA–TA2 (a section of TAD) upon binding to TAZ1 through its well-conserved hydrophobic sites in a series of grooves on the TAZ1 surface. The structural analysis coupled with the mechanistic studies by mutational and isothermal calorimetric analyses allowed the design of RelA-mutants that selectively abrogated the two distinct components of the RelA:CBP/p300 interaction. Detailed studies of these RelA mutants using cell-based techniques, mathematical modeling, and genome-wide gene expression analysis showed that a major set of the RelA-activated genes, larger than previously believed, is affected by this interaction. We further show how the RelA:CBP/p300 interaction controls the nuclear response of NF-κB through the negative feedback loop of NF-κB pathway. Additionally, chromatin analyses of RelA target gene promoters showed constitutive recruitment of CBP/p300, thus indicating a possible role of CBP/p300 in recruitment of RelA to its target promoter sites.
| The NF-κB family of transcription factors regulate the expression of numerous genes involved in the immune response, cell survival, differentiation, and proliferation. The interaction of the RelA subunit of NF-κB with the general co-activator protein CBP/p300 is vital for RelA-dependent gene transcription. Although the recruitment of RelA to its cognate genomic κB sites for target gene activation is well-established, the involvement of CBP/p300 in this process remains unclear. Through our structure/function-based approach we provide the molecular and functional details of the RelA:CBP/p300 interaction and its contribution to the regulation of distinct subsets of target genes. We also show that disruption of this interaction deregulates the NF-κB pathway by interfering with its negative feedback loop. Furthermore, our study indicates a possible reciprocal role for CBP/p300 in the recruitment of RelA to its target gene promoters.
| The NF-κB family of inducible transcription factors has emerged as a major player in immune response [1], activating a plethora of immunoregulatory genes upon stimulation. In vertebrates, the family is comprised of five members, namely, RelA (also known as p65), RelB, c-Rel, p50, and p52. They exist in various combinations of homo- and hetero-dimers, with RelA:p50 being the most abundant NF-κB dimer present in the cell. All of the family members share a structurally conserved N-terminal rel homology region (RHR), which is responsible for DNA binding and also contains the dimerization domain (DD). RelA, RelB, and c-Rel distinguish themselves from p50 and p52 by possessing the transcriptional activation domain (TAD), which is vital for the transcriptional regulation of the NF-κB-regulated genes. In unstimulated cells, NF-κB is maintained in a latent state in the cytoplasm by the family of Inhibitor-κB (IκB) proteins, which strongly bind to the RHR domain of NF-κB dimers, thereby masking its nuclear localization signal. Upon stimulation, a cascade of signaling events led to the degradation of the IκB proteins, thus releasing the NF-κB dimers, which then enter the nucleus to regulate the transcription of NF-κB-dependent genes. In the nucleus, NF-κB activates transcription by binding to its cognate κB sites in the promoters/enhancers of its target genes [2],[3]. However, mere binding of the NF-κB to its cognate κB sites does not ensure transcriptional initiation, and further requires the assembly of the “basal transcription machinery” on the transcription start site. Assembly of the transcription initiation complex requires NF-κB to interact with the mediator complex, various transcriptional adaptors, and co-activator proteins like CREB-binding protein (CBP) and its paralog p300 [4]–[7]. CBP/p300 are general transcriptional co-activators that help NF-κB bridge with the basal transcription machinery [8]. CBP/p300 interact with a large array of transcription factors to integrate multiple cellular signaling pathways [9],[10] and also possess chromatin-remodeling capabilities owing to their histone- and transcription factor–acetylating properties.
To date, it is widely believed that in the nucleus NF-κB recruits CBP/p300 to its target promoter sites, and in this process it must compete with various other cellular transcription factors for the limiting amounts of CBP/p300 [11]. However, in recent genome-wide analyses of gene promoters/enhancers using ChIP-chip and ChIP-seq technology, p300 has emerged as a prominent marker for enhancers [12]. p300 is also pre-loaded in most of the promoters and enhancers of NF-κB-regulated genes marked with histone-H3 lysine 4 trimethyl (H3K4me3) and histone H3 lysine 4 monomethyl (H3K4me1), respectively, in unstimulated THP-1 and HeLa cells [13]. In light of these new findings, the question arises that if p300 is already preloaded on the promoter/enhancer regions of the NF-κB target genes, then what is the role of NF-κB:CBP/p300 binding in the NF-κB activated transcription.
Among the NF-κB proteins, RelA is the most potent, ubiquitously expressed and well-studied family member. It contains a TAD, which can be further divided into two regions, TA1 and TA2, as depicted in Figure 1A [14]. Although the activation of the NF-κB pathway is extensively studied [15], mechanistic details of the transcription initiation process by NF-κB at the promoter site are limited [16]. This can partly be attributed to the lack of any structural information and limited overall knowledge of RelA–TAD [17],[18]. The interaction of RelA with CBP/p300 represents a key step in the initiation of transcription of a subset of RelA-activated genes [19]. RelA interacts with CBP/p300 in a bipartite manner—the RHR domain of RelA contacts the KIX domain of CBP/p300 and the TAD of RelA contacts the TAZ1 (also known as the CH1) domain of CBP/p300 [20]. The former interaction requires RelA-phosphorylation at a well-conserved Ser276 residue. Although both the RHR:KIX and TAD:TAZ1 interactions contribute to transcriptional activation of RelA-dependent genes, phosphorylation of Ser276 is considered critical for the RelA:CBP/p300 interaction [20]. It remains unclear whether the interaction between the RelA–TAD and the TAZ1 domain of CBP/p300 can activate transcription of RelA-target genes independent of RelA(Ser276) phosphorylation or vice versa. In addition to CBP/p300, RelA also interacts with a number of other proteins for the transcriptional activation of its target genes through its TAD, and hence a detailed knowledge of the TAD:TAZ1 binding is a necessary prelude to any mutations in the RelA–TAD, which can selectively abrogate RelA:CBP/p300 interaction [16].
As an initial step towards elucidating the role of NF-κB:CBP/p300 interaction in transcription of NF-κB-regulated genes, we determined the solution structure of the mouse RelA–TA2 (a subdomain of TAD) in complex with the TAZ1 domain of mouse CBP. The structure reveals the high affinity binding of RelA–TA2 through its well-conserved hydrophobic sites (Figure 1B) in a series of grooves on the TAZ1 surface. The structure enabled us to design point mutants of RelA, which selectively abrogated the RelA–TAD:CBP–TAZ1 binding, thereby allowing us to gain detailed insight into the molecular determinants of the RelA:CBP/p300 interaction. Using these RelA mutants defective in CBP/p300 binding, we performed a genome-wide analysis of the genes influenced by the RelA:CBP/p300 interaction to confirm the differential role of the two interaction sites on the RelA regulated transcriptome. We found genes like nfkbia and ptgs2, which were previously considered independent of RelA:CBP/p300 interaction [19], were actually dependent on it. The dependence of nfkbia expression on RelA:CBP/p300 interaction explains how this interaction regulates the temporal profile of nuclear NF-κB (nNF-κB) following TNFα stimulation. Furthermore, we revisited the model of CBP/p300 recruitment by RelA using Chromatin Immunoprecipitation (ChIP) assay. Our study confirmed that CBP/p300 is preloaded on the promoter regions of our set of RelA target genes, thus indicating that RelA:CBP/p300 interaction might be vital for recruitment of RelA to its target promoter sites.
The RelA–TAD:CBP–TAZ1 interaction was previously mapped to RelA(477–504) region [20] (Figure 1A). We found only a weak interaction between this RelA fragment and TAZ1 as compared to the whole RelA–TAD [RelA(425–549)] and therefore scanned the entire TAD for TAZ1 binding using GST pull-down assays (Figure S1). A RelA construct spanning residues Lys425–Pro490 (in the TA2 region) emerged as the minimal fragment with maximal affinity to TAZ1. This fragment (henceforth referred to as RelA–TA2) was also optimal for structural analysis by NMR spectroscopy (Figure S1B,C). This highly acidic RelA–TA2 fragment is unstructured in its free state, as observed from the narrow dispersion of peaks in the [15N-1H]–HSQC spectrum, but exhibits a well-dispersed spectrum indicative of a folded structure when in complex with TAZ1 (Figure 1C).
The structure of the complex between RelA–TA2 (residues 425–490) and CBP–TAZ1 (residues 340–439) was determined using distance and dihedral angle restraints derived from heteronuclear NMR experiments (see Table 1 for NMR statistics). TAZ1 is a scaffolding domain that interacts with intrinsically disordered TADs of various transcription factors, which fold upon complex formation [21]–[23]. The overall structure of TAZ1 in the complex (Figure 2) resembles that of TAZ1 in its free state (RMSD 1.6 Å) and in complex with other protein targets like HIF-1α (RMSD 1.8 Å), CITED2 (1.4 Å), and STAT2 (RMSD 1.3 Å) [21],[23]–[25]. As observed in our structure, the TAZ1 fold comprises four α-helices (α1, α2, α3, and α4) arranged in a roughly tetrahedral shape and is stabilized by three zinc atoms, each of which is bound to three Cys residues and one His residue. A prominent surface feature of the TAZ1 domain is a series of interlinked hydrophobic grooves that bind intrinsically disordered TADs (Figure S2).
In the RelA–TA2:TAZ1 structure, the N- and C-terminal regions of the RelA–TA2 fragment used for structural analysis (Lys425–Thr433 and Ser482–Pro490, respectively) are dynamically disordered, with zero or negative [1H]-15N heteronuclear NOEs (Figure S3A), and hence are omitted from the structures shown in Figure 2A,B. Our structure shows that RelA–TA2, essentially the Leu434–Val481 region, entirely wraps around TAZ1 in a predominantly extended conformation by docking itself through its well-conserved bulky hydrophobic residues into the interlinked hydrophobic grooves of TAZ1 (Figure 2 and Figure S4A). The C-terminal region (Glu471–Asn477) of RelA–TA2 folds into one short helix αC that is anchored to the hydrophobic pocket formed by packing of α1, α2, and α3 helices of TAZ1 (Figure 2B). Likewise, the N-terminal region (Leu434–Leu439) of RelA–TA2 also forms a short helix αN, which is docked into the shallow hydrophobic groove formed at the junction of α1 and α4 of TAZ1. The αN helix is dynamically disordered with only about 30% of the helical population, as calculated from the magnitude of the C-alpha and carbonyl secondary chemical shifts [26] (Figure S3B), and the small value of the [1H]-15N heteronuclear NOE (∼0.3–0.5 for the αN residues) confirms that this region is highly flexible on the nanosecond time scale (Figure S3A). RelA–TA2 also forms two additional helical turns, Leu449–Leu452 and Leu465–Val468, which make hydrophobic contacts in the α1–α2–α3 and α1–α3 interfaces of TAZ1, respectively. The hydrophobic interactions are complemented by electrostatic interactions between the highly acidic RelA–TA2 region and the strongly electropositive surface of TAZ1. The region from Asp444–Asp448 in RelA–TA2 is particularly acidic, with four out of five residues being negatively charged, and passes through a deep cleft in the TAZ1 surface that is lined with basic residues (Lys366, Arg368, Arg369, Lys419, Lys438, Arg439) (Figure 2C). Our structural analysis was further corroborated by ITC experiments carried out at two different salt concentrations to study the electrostatic contribution of the binding event. Increasing the NaCl concentration from 50 to 150 mM weakened RelA–TA2:TAZ1 binding by 4-fold, consistent with a significant electrostatic contribution to binding from the complementary charges on TAZ1 and the RelA–TA2 region (Table 2).
The structure of the RelA–TA2:TAZ1 complex revealed that several hydrophobic residues in RelA–TA2 mediate binding within the exposed hydrophobic grooves of TAZ1. The RelA–TA2 amino acid residues, which extensively interact with the hydrophobic pockets of TAZ1, belong to two ψXXψψ motifs and a ψψXXψXXψ sequence (Figure 3A), where ψ is a bulky hydrophobic residue (Leu, Val or Phe in RelA–TA2) and X can be any residue (Figure 1B). The ψXXψψ motif is a generalization of the LXXLL motifs that are known to mediate protein–protein interactions [27]. Furthermore, Val481, Phe443, and the Leu residues belonging to the αN helix of RelA–TA2 also contribute to complex formation as observed from the intermolecular NOEs (Figure S4B).
To investigate the role of the ψXXψψ motifs, we substituted alanine for Leu449 and Phe473 of RelA, each of which represents the first ψ residue of the motif. Both Leu449 and Phe473 are deeply buried in the molecular interface and participate in extensive hydrophobic interactions (Figures 1B and 3A). We also replaced the third ψ residue of ψψXXψXXψ sequence with Ala–Leu465Ala (Figure 3A, middle panel and Figure S4A). All three RelA mutants showed diminished to completely abrogated TAZ1 binding in the GST-pulldown assays (Figure 3B) and isothermal calorimetric (ITC) experiments (Figure S5, Table 2, Figure 3C), confirming the critical role of hydrophobic interactions in stabilizing the complex. The Leu449Ala and Leu465Ala substitutions each led to a 7-fold decrease in the TAZ1 binding affinity (Figure 3C and Table 2). As seen in the structure, the side chain of Leu465 is fully buried into a deep narrow hydrophobic pocket formed by Ile353, Leu357, Gln413, Ile414, His417, Trp418, Cys426, and Val428 of TAZ1 and Val468 of RelA (Figure 3A, middle panel and Figure S4A). This interaction contributes significantly to enthalpic stabilization of the complex, since the Leu465Ala substitution greatly decreases the enthalpic contribution to TAZ1 binding (Figure 3C). Leu449 binds in a shallower hydrophobic pocket, and accordingly substitution by alanine causes far less enthalpic destabilization (Figure 3A, left panel). The Phe473Ala mutation led to a complete loss of TAZ1 binding. In the structure, the side chains of Phe473, Leu476, and Leu477 (all the ψ residues of this ψXXψψ motif) are accommodated in the hydrophobic groove of TAZ1 (Figure 3A, right panel and Figure S4A). This arrangement is destabilized in the Phe473Ala mutant, leading to complete loss of TAZ1 binding. We speculate that a similar effect would be observed for Leu476Ala and Leu477Ala mutants.
Apart from the above-mentioned residues of RelA–TA2, the contacts between TAZ1 and the disordered αN helix of RelA–TA2 contribute slightly to formation of the RelA–TA2:TAZ1 complex (Figure S4B). N-terminal truncation to remove residues Lys425–His440, thereby eliminating the entire αN helix, decreases the affinity by only 1.6-fold (Figure 3C and Table 2). Likewise, deletion of residues Ser486–Pro490 from the disordered C-terminal end of RelA had no effect on TAZ1 binding (Figure S1D). However, further truncation to remove Val481–His485 greatly impaired binding by eliminating hydrophobic contacts between Val481 of RelA(TA2) and the side chains of Leu359, Leu381, and Pro382 in TAZ1. Replacement of Met483 by alanine had no effect on binding (Figure 3B), confirming that RelA residues beyond Val481 do not contribute to the interaction.
In summary, our structure shows that RelA–TA2 spirals through the exposed hydrophobic pockets of TAZ1 and anchors itself on TAZ1 at a number of points. As observed from the mutation studies, disruption of any of these anchoring residues of RelA–TA2 leads to the destabilization of the entire binding architecture of the entire complex.
The RelA–TA2 region also contains a well-conserved Ser467 (Ser468 in human RelA), which is a known phosphorylation site. Depending upon the stimulus, this site can be phosphorylated by GSK3β, IKKε, or IKKβ and plays a critical role in transcriptional regulation of NF-κB dependent genes [28]–[31]. Ser467 resides in the ψψXXψXXψ sequence of RelA–TA2 (Figure 1B), and in the complex it is close to the cluster of basic residues at the N-terminus of helix α1 of TAZ1. Our ITC experiments show that a phosphomimetic mutant, RelA(Ser467Asp), binds TAZ1 with about 1.4-fold higher affinity than wild type RelA–TA2 and exhibits the same dependence on ionic strength, suggesting an enhanced but still nonspecific electrostatic interaction (Table 2 and Figure 3C). Thus, the increased negative charge of phosphoserine would likely enhance TAZ1 binding even further over nonphosphorylated RelA.
In addition to the RelA–TA2:TAZ1 interaction, the RelA–RHR interacts with CBP–KIX via p-Ser276–RelA [20]. We compared the binding of RHR to KIX and TA2 to TAZ1 by GST-pulldown assay using nuclear extracts (NEs) of wild-type 3T3 cells stimulated with TNFα for 30 min. KIX and TAZ1 were purified as GST-fusion proteins and used to pull down nuclear-RelA (nRelA). As seen in Figure 4A, significantly lower amounts of nRelA were pulled down by GST–KIX compared to GST–TAZ1, which can be due to either of the two possibilities or both: first, a lower binding affinity of RelA–RHR:KIX compared to RelA–TA2:TAZ1 complex, and second, presence of only a small fraction of p-Ser276–RelA in the total pool of nRelA. In this context, we compared the binding affinities reported in the literature for KIX with other target proteins [32],[33] and found the reported binding affinities for KIX complexes typically 10–100-fold lower than those observed for various TAZ1 complexes [32],[33].
Overall, in either possible scenarios the total amount of RelA–TA2 bound to TAZ1 is significantly higher than that of RelA–RHR bound KIX, which should be reflected in a subset of RelA-activated genes that is regulated through RelA–TA2:TAZ1 interaction and is independent of p-Ser276–RelA:KIX binding.
Next, we studied the role of the critical hydrophobic residues of RelA–TA2 (Leu449, Leu465, and Phe473) on TAZ1 binding in the context of full-length RelA. We used in vitro purified GST-tagged TAZ1 to pull down full-length wild-type (wt) or mutant RelA (Figure 4B) from the NEs of TNFα treated rela−/− cells (see Materials and Methods) reconstituted with RelA(wt/mutants). We found that the RelA mutants, namely RelA(Leu449Ala+Phe473Ala) and RelA(Leu449Ala+Phe473Ala+Ser467Ala), were completely defective in binding to TAZ1 (Figure 4B). As expected, the Ser276Ala mutation had no effect on TAZ1 binding. Interestingly, the RelA(Leu449Ala) mutant did not show any noticeable decrease in binding affinity, contrary to the impaired TAZ1 binding observed in our in vitro GST pulldown and ITC experiments. A possible explanation is that in vivo posttranslational modifications like the phosphorylation of the well-conserved Ser467 could potentially mitigate the effects of the Leu449 mutation, thus increasing binding affinity of the RelA–TA2 to TAZ1, as observed by ITC experiments (Figure 3C and Table 2). Additionally, in the context of full-length RelA, the other domains of RelA could further influence the RelA–TA2:TAZ1 binding.
The above experiments established the minimal subset of mutations required in RelA to selectively prevent its interaction with the TAZ1 domain of CBP. These RelA mutants along with the RelA(Ser276Ala) were further tested for their interaction with endogenous CBP by coimmunoprecipitation (co-IP). Binding of all three RelA mutants [RelA(Leu449Ala+Phe473Ala), RelA(Leu449Ala+Ser467Ala+Phe473Ala), and RelA(Ser276Ala)] to endogenous CBP was impaired relative to RelA(wt) (Figure 4C). However, the disruption of either component of the RelA:CBP interaction—RHR:KIX or TA2:TAZ1—diminished but could not completely abrogate the entire RelA:CBP interaction. The RelA(Leu449Ala+Phe473Ala) and RelA(Ser276Ala) mutants were equally defective in binding to p300 (Figure 4D), whose TAZ1 domains share high sequence identity with CBP [22].
To further elucidate the role of Ser467 in RelA:CBP/p300 binding, we first studied the RelA(Ser467) phosphorylation event both spatially and temporally following TNFα stimulation. As shown in Figure S6A, the amount of p-Ser467–RelA is negligible in the unstimulated cells. Phosphorylation increases within 5 min of TNFα stimulation, reaching its maximum at 10 to 15 min followed by gradual dephosphorylation starting at 20 min after stimulation. We confirmed the presence of p-Ser467–RelA exclusively in the nucleus [34] (Figure 4E). The timing of RelA nuclear entry coincides with the p-Ser467–RelA maximum, indicating a possible correlation between the Ser467 phosphorylation event and an enhanced early RelA:CBP/p300 interaction. This correlation holds true only if Ser467 phosphorylation contributes towards increasing the affinity. To confirm this, co-IP experiments with α-CBP and α-p300 using NEs of RelA(wt) and RelA(Ser467Ala) mutants at three different time points after TNFα stimulation (Figure 4E) in addition to unstimulated condition were performed. Interestingly, in the RelA(wt) reconstituted cells, despite lower amounts of nRelA at 15 min than at 30 min of TNFα treatment, equivalent amounts of RelA co-immunoprecipitated with CBP/p300. The amount of nRelA bound to CBP/p300 at 45 min after TNFα stimulation diminished significantly. The nonlinearity of the amounts of nRelA co-immunoprecipitated with CBP/p300 with respect to the nRelA concentration is attributed to the enhanced binding affinity of p-Ser467–RelA for CBP/p300, as this nonlinearity was not observed for the identical experiment performed using RelA(Ser467Ala) mutant reconstituted rela−/− cells (Figure 4E, Figure S6B).
Despite the role of p-Ser467–RelA in RelA:CBP/p300 binding, RelA(Leu449Ala+Phe473Ala) mutant [henceforth known as the RelA(TA2) mutant] was selected for further gene expression experiments due to two reasons: First, it had the minimal number of mutations required to impair the RelA–TA2:TAZ1 interaction. Second, RelA(Ser467) is a key determining factor for various other processes in the NF-κB pathway [28],[35],[36], which could further complicate and lead to misinterpretation of our experimental observations. Therefore, the RelA(Leu449Ala+Ser467Ala+Phe473Ala) mutant was not used further in this study.
Next we investigated the effect of RelA:CBP/p300 interaction on the inducible expression profile of NF-κB target genes by qRT-PCR. Expression of the target genes was tested in the rela−/− cells reconstituted with RelA(wt/mutants) following TNFα stimulation. Fibroblast cells were used because in this cell type the NF-κB-driven transcription is mainly carried out by the RelA subunit [16],[37]. rela−/− cells reconstituted with an empty vector were used as a control, which also ensured that the genes tested were RelA dependent. The set of nine RelA-dependent genes are well-studied early TNFα responsive genes (Figure 5A) [16],[19],[38]. In this gene set, ptgs2 and nfkbia are NF-κB target genes believed to be independent of RelA:CBP/p300 interaction [19].
Our qRT-PCR data demonstrated inducible expression for all of the genes in the chosen set following TNFα stimulation in the RelA(wt) reconstituted rela−/− cells. However, the expression of all the nine genes was significantly diminished in the cells reconstituted with RelA(TA2) mutant (Figure 5), suggesting that genes like ptgs2 and nfkbia, which are believed to be independent of RelA:CBP/p300 interaction, are actually dependent on RelA:CBP/p300 binding through the RelA–TA2:TAZ1 component (Figure 5A).
On the other hand, for the RelA(Ser276Ala) reconstituted cells, only seven of the nine genes displayed reduced gene expression. The relative mRNA levels of the remaining two genes, namely vcam1 (NM_011693) and ptgs2, were not affected by the RelA(Ser276Ala) mutation following TNFα stimulation (Figure 5). In our study, expression of nfkbia, which is believed to be independent of Ser276 phosphorylation, appeared to be partially dependent on it. The lower levels of nfkbia mRNA in the RelA(Ser276Ala) reconstituted cell line following TNFα stimulation as compared to those observed for RelA(wt) reconstituted cells could possibly be due to the lower expression of RelA(Ser276Ala) mutant in the cell relative to RelA(wt) and RelA(TA2) mutant (Figure S7).
To ascertain the effect of the RelA:CBP/p300 interaction on the RelA-regulated transcriptome, we carried out RNA-seq analysis of total mRNA isolated from RelA(wt/mutants) reconstituted cells prior to TNFα stimulation as well as at two different times points after TNFα stimulation (Figure 5B). The set of genes activated upon TNFα stimulation was classified into four groups: Group A contains genes that are dependent on both the interaction components of the RelA:CBP/p300 complex. Group B contains those dependent only on the RelA–TA2:TAZ1 interaction. Group C contains genes that are dependent only on p-Ser276–RelA, and Group D contains TNFα activated genes, which are RelA regulated but are altogether independent of RelA:CBP/p300 interaction, or the genes that are activated by transcription factors other than RelA. The RNA-seq data revealed that a majority of the RelA:CBP/p300 target genes belong to Group B (25 genes) followed by Group A (20 genes) and Group C (six genes). A sizable number of RelA target genes belong to RelA:CBP/p300 independent Group D (48 genes) (Figure 5B). Based on these experiments, it is clear that the RelA–TA2:TAZ1 interaction regulates a subset of RelA target genes independent of p-Ser276–RelA. Thus, the subset of RelA-activated genes dependent on RelA:CBP/p300 interaction is much larger than previously believed.
The gene expression analysis revealed nfkbia (IκBα) and tnfaip3 (A20), which control the negative feedback loop of the NF-κB pathway, are dependent on RelA:CBP/p300 interaction. IκBα essentially is the main NF-κB inhibitor, which strongly controls the negative feedback and is responsible for the fast turnoff of the NF-κB response after TNFα treatment in fibroblast cells [39]. Therefore, we asked whether reduced expression of nfkbia due to defective RelA:CBP/p300 interaction had any effect on the NF-κB pathway.
To study the effect of the RelA:CBP/p300 interaction on nfkbia expression in detail, we investigated the temporal profile of the IκBα protein in rela−/− cells reconstituted with RelA(wt/mutants) following TNFα stimulation. This study was also performed to ensure that the RelA mutations did not interfere with the upstream NF-κB pathway in the cytoplasm or with the DNA binding property of RelA. nfkbia is a prominent RelA target gene, which has a robust NF-κB responsive promoter. IκBα is a strong negative regulator of nuclear NF-κB (nNF-κB) and hence a key determinant of the nNF-κB temporal profile [39]. The canonical NF-κB pathway stimulated by TNFα has been extensively studied and focuses on the activation of the IKK-IκB-NF-κB signaling module. A mathematical model, which accounts for experimentally observed nNF-κB activity and IκB expression profiles, has been established [40],[41]. This model depends on the abundance and the transcriptional activation potential of NF-κB (RelA) [42],[43], among other parameters.
We observed that the normalized protein levels of RelA in the rela−/− cells reconstituted with RelA-mutants varied relative to that in RelA(wt)-reconstituted cells (Figure 6A). This prompted us to simulate the effect of NF-κB abundance on the temporal profile of nNF-κB activity following TNFα stimulation (Figure 6B) before studying the transcriptional activation potential of the RelA-mutants. The simulations show that a decrease in the total NF-κB concentration leads to sustained nNF-κB activity and lower levels of IκBα synthesis following TNFα stimulation (Figure 6B).
We next compared the nuclear translocation of RelA and the temporal profile of IκBα protein following TNFα stimulation in rela−/− cells reconstituted with wild-type or mutant RelA. To detect nNF-κB in NEs, the NF-κB DNA binding activity was measured by Electrophoretic Mobility Shift Assay (EMSA) (Figure 6C). nRelA levels were also monitored directly by immunoblotting (Figure S8). The time of entry of RelA(TA2) and RelA(Ser276Ala) mutants into the nucleus after TNFα stimulation is identical to that observed for RelA(wt). However, a prolonged nuclear residence time is observed for RelA(TA2) mutant relative to RelA(wt) reconstituted cells (Figure 6C and Figure S8). We also performed similar time course experiments using the whole cell extracts (WEs) of the respective cells and monitored the IκBα protein levels by immunoblot (Figure 6D). While the time of IκBα degradation and RelA entry into the nucleus are identical in the RelA(wt), RelA(Ser276Ala) and RelA(TA2) mutant reconstituted rela−/− cells (Figure 6C,D), the temporal profile of IκBα regeneration and the nuclear residence time of RelA differed significantly (Figure 6C–E). For the RelA(TA2) mutant reconstituted cells, IκBα regeneration is delayed and does not reach its initial levels within the time course of the experiment. The RelA(Ser276Ala) mutant on the other hand did not show any significant delay in IκBα regeneration, although the IκBα level was reduced.
In order to address our observed results, we applied the available mathematical model of NF-κB regulation [40],[41]. We assumed that all parameters other than the RelA protein abundance and its transcriptional activation potential remained unchanged in all of the RelA reconstituted rela−/− cell lines. Identical IκBα degradation and nuclear translocation of RelA for the three cell lines validated our assumption (through 30 min after TNFα stimulation in Figure 6C,D and Figure S8). Further, when the cell-line-specific RelA concentration estimates were incorporated into the model, the calculated profile for nNF-κB activity and IκBα abundance matched the experimental data of the RelA(Ser276Ala) mutant to a significant extent (Figure 6E, lowest panel), but did not resemble the observed profiles for the RelA(TA2) mutant (Figure 6E, middle panel). The root mean square deviation (RMSD) was used as a measure of agreement between calculated and the observed data. The results suggest that the observed partial expression of nfkbia in RelA(Ser276Ala) mutant reconstituted cells in our qRT-PCR data (Figure 5) was to a great extent due to lower expression of this RelA mutant in the cell line used here.
For the RelA(TA2) mutant, we investigated whether agreement between the experimentally observed and calculated data could be obtained by reducing the NF-κB-dependent IκBα mRNA production rate (kmRNA) parameter to 25%, 12.5%, or 6.25% of its value for RelA(wt) (Figure 6F). We found that the lowest kmRNA of 6.25% of RelA(wt) produced the best superimposition of the calculated and experimentally determined temporal profiles of the nNF-κB activity and IκBα regeneration as determined by the RMSD score (Figure 6F, top panel). It should be noted that the kmRNA is directly proportional to the transcriptional activation potential of nNF-κB. Our study suggests that the Leu449Ala and Phe473Ala mutations in RelA lower the transcriptional activation of NF-κB by impairing the RelA–TA2:TAZ1 component of RelA:CBP/p300 interaction.
From the above results we conclude that while nfkbia is dependent on RelA–TA2:TAZ1 interaction, it is independent of RelA–RHR:KIX interaction. Thus, our data show that disrupting the RelA–TA2:TAZ1 component of the RelA:CBP/p300 interaction disturbs the temporal profile of the nuclear activity of RelA by interfering with the negative feedback loop of the NF-κB pathway. This could further deregulate the expression of the subset of NF-κB-dependent genes, which are independent of the overall RelA:CBP/p300 interaction.
To further understand the nature of RelA:CBP/p300 interaction on the chromatin, ChIP-qPCR assay was performed to analyze the chromatin at six RelA target genes in unstimulated and TNFα-stimulated RelA(wt/mutants) reconstituted rela−/− cells. We examined the chromatin for RelA occupancy on its target promoter set (Figures 7A and S9). As expected, we found that RelA(wt) was recruited to the promoters of all the six genes (although to a varied extent) at 30 min after TNFα treatment. This enrichment of RelA(wt) was diminished at 60 min after stimulation in congruence with its nuclear exit. The recruitment of RelA(TA2) and RelA(Ser276Ala) mutants to the promoter sites of highly expressed tnfaip3, nfkbia, and ptgs2 was significantly reduced at 30 min following TNFα treatment compared to RelA(wt), while for the mutants there was no significant RelA signal detected on the promoters with low mRNA expression levels (cxcl2, csf2, and tnf). In agreement with the prolonged nuclear retention of RelA(TA2) mutant due to a defective negative feedback loop, reduced amounts of the RelA(TA2) mutant remained bound to the tnfaip3, nfkbia, and ptgs2 promoters even at 60 min after stimulation. As indicated earlier, expression of tnfaip3 and nfkbia were reduced but not completely eliminated. Therefore, the mRNA levels of the RelA target genes correlate well with the amounts of RelA recruited to their respective promoters. These results clearly indicate defects in promoter recruitment of RelA mutants despite their uncompromised DNA binding potential (Figure 6C). This led us to hypothesize that CBP/p300 might be responsible for recruitment of RelA to its target gene promoters. If our hypothesis were true, then CBP/p300 should be constitutively loaded on the promoters and there should exist a direct correlation between the amounts of CBP/p300 present on the promoters to that of RelA recruited upon stimulation. To test our hypothesis, the chromatins were analyzed for CBP/p300 enrichment for our promoter set. The results show CBP/p300 enriched on all of the promoters to a varied extent, with no significant change upon TNFα stimulation (Figures 7B and S9). A comparison of CBP/p300 and RelA signals in this assay revealed a direct correlation between enrichment of RelA upon stimulation with that of CBP/p300 (Figure 7C). Similar RelA recruitment was observed for RelA(wt), RelA(TA2), and RelA(Ser276Ala) mutants on RelA target promoters independent of RelA:CBP/p300 interaction (Figure S10).
Since acetylation of histone-3 (H3) at Lys27 by CBP/p300 is a characteristic of active genes [44], we further examined the levels of H3 occupancy and compared them with that of acetylated histone H3 at Lys 27 (H3K27ac). We expected to observe increased H3K27ac signals on the promoters only after TNFα treatment as a mark of activated promoters as compared to those in unstimulated cells. To our surprise we found H3K27ac enriched in the unstimulated promoters, whose levels remained constant even after TNFα stimulation. Additionally, the H3K27ac levels were directly proportional to the amount of CBP/p300 present on the specific promoters (Figure 7C). This suggests that CBP/p300 bound to promoters of RelA targets is poised for transcription initiation. Upon nuclear entry RelA is recruited to its respective promoter targets by CBP/p300, which is preloaded on these sites, leading to transcription activation.
Upon TNFα stimulation in RelA(TA2) or RelA(Ser276Ala) mutant reconstituted rela−/− cells, the mRNA expression levels of the highly expressed genes like tnfaip3 and nfkbia registered only partial reduction (Figure 5A). This is because of a reduced but significant amount of RelA recruited to their respective promoters (Figure 7A). Both the promoters also maintain high levels of CBP/p300, thereby leading to recruitment of the RelA(TA2) and RelA(Ser276Ala) mutants, which possess reduced but significant CBP/p300 binding affinity (Figure 4C,D). This further proves that the RelA-dependent gene expression is regulated by the total amount of RelA recruited on the promoter by CBP/p300.
Overall our study indirectly shows that CBP/p300, which marks the promoter/enhancers of genes, also aids in recruiting RelA to the promoters of its target genes.
In this study, we have provided the structural basis and the functional role of the RelA:CBP/p300 interaction in RelA-driven transcription. In the RelA–TA2:TAZ1 structure, the RelA–TA2 region binds within a series of interlinked hydrophobic grooves that spiral almost entirely around the globular TAZ1 domain. Due to the intrinsically disordered nature of the RelA–TA2 region, binding of an elongated ∼50 amino acid region to TAZ1 in a relatively extended configuration can be accomplished without the energetic cost that would be incurred if unfolding of preexisting globular structure was required. The interaction is mediated primarily by hydrophobic residues interspersed with acidic residues located in the RelA–TA2 region, which dock in the hydrophobic pockets in the surface of TAZ1. While RelA becomes more ordered upon interaction with TAZ1, only limited, highly localized secondary structure is formed. Contacts between the N-terminal part of the RelA–TA2 and TAZ1 are dynamic, contributing little to binding affinity, and are an example of what is frequently termed a “fuzzy” interaction [45]. The hydrophobic contacts are complemented by electrostatic interactions involving the many acidic residues in the RelA–TA2 region. Disruption of any of the hydrophobic contacts leads to a significant decrease in binding affinity toward TAZ1. This is due to the destabilization of the ordered RelA–TA2 structure held primarily by the exposed hydrophobic grooves of TAZ1. Upon disruption of any of the anchor points, RelA–TA2 in absence of any highly stabilized secondary structural elements regains its highly dynamic unstructured high-energy state. This leads to destabilization of the complex and a decrease in TAZ1 binding affinity. Disruption of the hydrophobic anchors of RelA–TA2 can also destabilize the hydrophobic core of TAZ1, which is otherwise stabilized to a significant extent on complex formation with RelA–TA2 (Figure S11). This interlinked core of TAZ1 holds the entire RelA–TA2 region in the complex as seen through the contacts of the core elements with the anchoring residues of RelA–TA2.
The mode of binding of the TAZ1 domain to RelA–TA2 is closely analogous to its interactions with the transactivation domains of the HIF-1α, CITED2, and STAT2 proteins [21]–[24],[46]. In particular, RelA binds in the same hydrophobic grooves as the HIF-1α C-terminal activation domain but in the reverse orientation [21], despite the fact that the two proteins share no common sequence motifs and do not interact with TAZ1 through common secondary structures. A salient property of TAZ1 is its ability to interact with long, negatively charged 40–50 residue regions of intrinsically disordered proteins that dock in the narrow hydrophobic grooves in its surface. This interaction is further enhanced by the phosphorylation of Ser467, thereby adding another layer of regulation for a subset of early response target genes of RelA following TNFα stimulation. Similar phosphorylation events are observed in other transcription factors, which influence their interaction with CBP/p300 [47],[48].
To date, inducible phosphorylation of Ser276 is considered to be the decisive factor for the recruitment of CBP/p300 by RelA for the transcriptional activation of RelA-dependent genes [19],[20]. Earlier studies have shown that in the absence of Ser276 phosphorylation, histone deacetylases (HDACs) could remain bound to nNF-κB and lead to the repression of NF-κB-dependent genes [19],[49]. Upon phosphorylation at Ser276, RelA readily recruits CBP/p300, which in turn acetylates RelA at Lys310, which is further required for the transcription of a subset of RelA-dependent genes [50],[51]. Our experiments show that a major subset of TNFα-inducible genes is dependent on the RelA–TA2:TAZ1 interaction. Importantly, a significant number of these genes (Figure 5B, Group B genes) are activated independently of Ser276 phosphorylation. For these genes the requirement of p-Ser276–RelA could be compensated by a synergy between multiple protein factors (including transcription factors) and RelA. A similar mechanism was previously observed for cxcl2 and nfkbia transcription, where the requirement for RelA:TRAP-80 interaction was waived for the expression of these genes through the recruitment of secondary transcription factors [16]. Similarly, for tnfaip3 transcription, the constitutive transcription factor Sp1 plays an important role in synergy with RelA [52]. Moreover, the location of Ser276 in the DD of RelA could potentially influence the selectivity of its dimer formation with other NF-κB family members [53]. Additionally, the proximity of Ser276 to DNA binding residues of RelA could further impact the κB site selectivity [54], thereby leading to differential gene expression and/or compensating for the RelA–RHR:KIX interaction.
In this study, for the first time we show that the negative feedback loop is directly controlled by RelA:CBP/p300 interaction through the expression of nfkbia. Similarly, A20 protein (tnfaip3 gene product), another regulator of the negative feedback loop of the NF-κB pathway, is also directly controlled by RelA:CBP/p300 interaction. The impaired negative feedback loop can impact the expression profile of TNFα-induced genes. The longer residence time of RelA in the nucleus can prolong the activation of the genes that are independent of RelA:CBP/p300 interaction.
RelA is believed to recruit co-activator CBP/p300 to the gene promoters for transcription activation of its target genes. On the other hand, recent ChIP-seq and ChIP-chip studies have established p300 as a mark for the enhancers/promoters of active/poised genes throughout the genome including those for RelA target genes [12],[13]. However, the mechanism of its recruitment by RelA to the promoters remains to be established. In this study, we found that CBP/p300 was constitutively enriched on the RelA target promoters as against the TNFα-induced recruitment of RelA to these sites. This points towards a role played by CBP/p300 in recruitment of RelA to its target promoters. Although we have tested only a small set of RelA target promoters, our observation is consistent with the report by Jin et al. [13], where they could accurately predict the target selection by NF-κB in a cell-type-specific manner only when experimentally determined p300 enrichment of promoters was considered along with κB motif [55],[56] along the human genome.
The GST tagged mouse RelA fragments were expressed in E. coli [BL21(DE3) cells] and purified by affinity chromatography using Glutathione sepharose beads (GE Life Sciences). The purified protein fragments were then dialyzed in the binding buffer of the respective experiments.
For NMR experiments, the mouse RelA fragments were expressed as fusion proteins with an N-terminal His6-tag followed by the 58 amino acid residue B1 domain of streptococcal protein G (GB1) as per previous reports [57],[58]. The protein was isotopically labeled by growth in M9 minimal media supplemented with 15N ammonium chloride and 13C-glucose. The fragments were purified under denaturing conditions on a Ni-NTA (SIGMA) column followed by dialysis against cleavage buffer [20 mM Tris (pH 7.5), 200 mM NaCl] at 4°C. The His6–GB1 peptide was cleaved by thrombin at room temperature for 18 h. The cleaved RelA peptides were further purified by reverse-phase HPLC (RPHPLC) using a C18 column on ÄKTA purifier system. The purified peptides were lyophilized immediately. For NMR and ITC experiments, the TAZ1 domain (residues 340–439 of mouse CBP) was expressed and refolded as described previously [24]. For ITC experiments, the His6-tag was cleaved from the GB1–RelA peptides by TEV protease (in house purified).
The GST-pulldown assay was carried out in binding buffer (20 mM Tris pH 7.5, 150 mM NaCl, and 0.5 mM DTT). GST-RelA fragments (20 µg) were mixed with purified TAZ1 (10 µg) and 20 µl of a 50% slurry of washed glutathione sepharose beads and incubated at 4°C for 30 min in a rotator. The beads were then washed 5 times with 500 µL of binding buffer containing 0.3% TritonX-100, followed by elution and analysis by 15% SDS-PAGE.
ITC experiments were performed at 25°C using a MicroCal Omega VP-ITC instrument. For ITC experiments, GB1-tagged RelA-TA2 fragments were used. The proteins were dialyzed overnight in the ITC buffer [20 mM Tris (pH 8.0), 50 mM or 150 mM NaCl and 1 mM DTT]. The protein concentrations were in the range of 10 to 18 µM of RelA–TA2 peptide in the cell and 169 to 185 µM TAZ1 in the syringe. Protein concentrations were determined by absorbance at 280 nm. The GB1-tag was used for accurate concentration determination of the RelA–TA2 peptides, which have a low extinction coefficient. The GB1-tag did not interact with TAZ1. A typical ITC experiment consisted of a total of 25 injections, using 5–6 µl of TAZ1 for the first injection followed by 24 injections of 11–12 µl of TAZ1 into the cell containing either blank or RelA peptide. Data were analyzed using a single-binding site model in the MicroCal Origin Software. The stoichiometry of binding ranged from 0.9 to 1.2. Errors for the Kd values were estimated from duplicate measurements. All the buffers used were made using water purged with nitrogen.
NMR experiments were carried out at 25°C on Bruker DRX600 MHz and Avance 900 MHz spectrometers. The typical protein concentration for NMR experiments was 0.7–1.2 mM. All spectra were referenced to external DSS. Refolded TAZ1 and RelA fragments were mixed to form the complex followed by overnight dialysis in NMR buffer (20 mM Tris pH 6.5, 40 mM NaCl, and 2 mM DTT). We added 5% D2O to the sample prior to NMR experiments. NMR data were processed using NMRPipe [59] and analyzed using CARA [60]. Backbone and side-chain resonances were assigned using standard triple resonance experiments [61]. Distance restraints were obtained from 3D 15N-edited NOESY-HSQC (τm = 100 ms) and 13C-edited NOESY-HSQC (τm = 100 ms) spectra. Intermolecular NOEs were derived from 12C-filtered-13C-edited NOESY-HSQC (τm = 200 ms) [62] experiments. Unambiguous intermolecular NOES were assigned manually and were used in the initial structure calculations, which were performed using CYANA [63] with CANDID [64]. Chemical shift-based torsion angles were obtained using TALOS+ [65]. Distance and torsion angle restraints were imposed to ensure tetrahedral geometry for the zinc atoms of TAZ1 [25]. A total of 200 structures were generated in CYANA and were further refined by restrained molecular dynamics simulated annealing using the AMBER 10 software package [66],[67].
These structures were subjected to 2,000 steps of energy minimization, followed by 20 ps of simulated annealing in vacuum and another 20 ps of simulated annealing using a generalized Born solvent model [68] and finally 2,000 steps of energy minimization. During the simulated annealing, the system was heated to 1,000 K for the first 2 ps, followed by 4 ps at constant temperature, and final cooling to 0 K for the remaining 14 ps. Force constants were 30 kcal mol−1 Å−2 for NOE restraints and 100 kcal mol−1 rad−2 for dihedral angle restraints. The 20 lowest energy structures were analyzed using PROCHECK-NMR [69].
Immortalized rela−/− fibroblast cells [70] were reconstituted with retroviral vectors pBABE-puro with mouse RelA(wt/mutants) inserted or with empty vector controls. A final concentration of 5 ng/ml TNFα (Roche Diagnostics) was used for stimulation. Antibodies used for IB and IP in this study were from Santa Cruz: anti-RelA (sc-372g), anti-IκBα (sc-371), anti-CBP (sc-369), anti-Actin (sc-1615) and anti-USF2 (sc-861), anti-α-Tubulin (sc-5286), anti-p300 (sc-584 and sc-585), and Cell-Signaling Technology: p-Ser468-RelA (#3039). cOmplete, EDTA-free Protease Inhibitor cocktail was from Roche Diagnostics. Protein G-beads were from GE Life Sciences.
For the NE, 3T3 cells were cultured in 10 cm3 plates and harvested using PBS buffer. Cell pellets were prepared by 5 min centrifugation at 500 g at 4°C and then resuspended in CE buffer (0.32 M Sucrose, 10 mM Tris HCl pH 8.0, 3 mM CaCl2, 2 mM MgOAc, 0.5% NP-40, 1 mM DTT, 0.5 mM PMSF) followed by centrifugation at 500 g for 5 min at 4°C. The supernatant (CE) was flash frozen and stored at −80°C. The nuclear pellet was washed twice with CE buffer without NP-40 followed by centrifugation as above. The washed pellet was resuspended in hypotonic buffer [20 mM HEPES (pH 7.9), 1.5 mM MgCl2, 20 mM KCl, 25% glycerol, 0.5 mM DTT, 0.5 mM PMSF] followed by gradual addition of high salt buffer [20 mM HEPES (pH 7.9), 1.5 mM MgCl2, 800 mM KCl, 25% glycerol, 1% NP-40, 0.5 mM DTT, 0.5 mM PMSF, and 1× complete EDTA-free protease inhibitor]. The samples were incubated with mild agitation for 45 min at 4°C followed by centrifugation at 12,000 g at 4°C for 15 min. The supernatant was collected as NE and was further subjected to quantification.
For WE, the cells were harvested as above followed by addition of RIPA lysis buffer containing protease inhibitor cocktail. The samples were centrifuged at 12,000 g at 4°C. The supernatant was collected as the WE. All of the buffers contained phosphatase inhibitor cocktail.
For immunoblotting, WEs and NEs were resolved on SDS-PAGE and immunoblotted with respective antibodies. For co-immunoprecipitation (co-IP) assay, NEs were incubated with anti-RelA or anti-CBP antibodies for 1 h followed by addition of prewashed protein-G beads and incubation for 4 h. The beads were then washed thoroughly and resolved on SDS-PAGE and immunoblotted with anti-RelA and anti-CBP antibodies. For quantitative analysis of the immunoblots, the loading controls were analyzed from the same gel as for the protein monitored.
RNA was extracted and purified as per the manufacturer's protocol using the RNeasy kit (QIAGEN). We reverse transcribed 1 µg of the quantified RNA using superscript II RT system (Invitrogen) and poly-dT primers. cDNA fragments generated were analyzed by qPCR using the KAPA system and Realplex thermocycler (EPPENDORF). PCR amplification conditions were 95°C (4 min) and 40 cycles of 95°C (15 s), 56°C (30 s), and 72°C (30 s). Primer pairs used were from previous studies [38],[71] and were tested by melt-curve analysis. Δ(ΔCt) method was used for data analysis with gapdh as normalization control to derive fold induction over basal levels. For each set of time-course experiment (0 min, 10 min, 60 min, and 120 min) RNA was collected from empty vector, RelA(wt), RelA(TA2), RelA(Ser276Ala) mutant reconstituted rela−/− cells stimulated with 5 ng/ml TNFα. For qRT-PCR experiments, the data were collected for three set of biological replicates.
Next generation sequencing was performed at the UCSD Biogem core facility. Reads were obtained from Illumina HiSeq 2000 sequencing system. Sequencing experiments were performed on the total mRNA extract from RelA(wt), RelA(TA2), and RelA(Ser276Ala) reconstituted rela−/− cells at two different time points after TNFα (5 ng/ml) treatment in addition to unstimulated cells. RNA-seq was performed on nine different experimental conditions in total. Two independent replicates were used for each experimental condition. Reads were mapped to the mm10 mouse genome using TopHat [72], and the transcriptome assembly was generated using Cufflinks [73] followed by detection of differentially expressed genes using Cuffdiff.
The ordinary differential equation (ODE)-based model for NF-κB regulation [40] was adapted (Table S1) by increasing IκB mRNA degradation rates by a factor of 1.8 in line with recent measurements [74] and replacing the terms describing transcription-factor-dependent mRNA production for IκBα and IκBε by those used in [41] (Model 3, additive Pol II recruitment). Model simulations were carried out with the “Stiffness Switching” method of the NDSolve function in the package Mathematica 8 (Wolfram Research, Champagne, IL) using a numerically defined IKK activity induced by exposure to TNFα Table S2 [40].
Samples for ChIP experiments were prepared using SimpleChIP Plus Enzymatic chromatin IP kit (Cell Signaling Technology #9005) according to the manufacturer's protocol with some modifications. Primer pairs used were from previous studies [38],[71],[75] and were tested by melt-curve analysis. ChIP experiments were performed with anti-H3 (Cell Signaling #4620), anti-acetyl H3K27 (Cell Signaling #8173), anti-RNA Pol II (Santa Cruz sc-900), anti-p300 (Santa Cruz sc-585), anti-CBP (Santa Cruz sc-369), anti-RelA (Santa Cruz sc-372), and anti-IgG (Cell Signaling #2729)
Coordinates and structural restraints for the RelA–TA2:CBP–TAZ1 complex have been deposited in the Protein Data Bank under the accession number 2LWW. The chemical shifts have been deposited in the BioMagResBank, accession number 18650. The RNA-seq data discussed in this publication have been deposited in NCBI's Gene Expression omnibus [76] and are accessible through GEO Series accession number GSE46213. (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE46213).
|
10.1371/journal.pntd.0004980 | West Nile Virus Spreads Transsynaptically within the Pathways of Motor Control: Anatomical and Ultrastructural Mapping of Neuronal Virus Infection in the Primate Central Nervous System | During recent West Nile virus (WNV) outbreaks in the US, half of the reported cases were classified as neuroinvasive disease. WNV neuroinvasion is proposed to follow two major routes: hematogenous and/or axonal transport along the peripheral nerves. How virus spreads once within the central nervous system (CNS) remains unknown.
Using immunohistochemistry, we examined the expression of viral antigens in the CNS of rhesus monkeys that were intrathalamically inoculated with a wild-type WNV. The localization of WNV within the CNS was mapped to specific neuronal groups and anatomical structures. The neurological functions related to structures containing WNV-labeled neurons were reviewed and summarized. Intraneuronal localization of WNV was investigated by electron microscopy. The known anatomical connectivity of WNV-labeled neurons was used to reconstruct the directionality of WNV spread within the CNS using a connectogram design. Anatomical mapping revealed that all structures identified as containing WNV-labeled neurons belonged to the pathways of motor control. Ultrastructurally, virions were found predominantly within vesicular structures (including autophagosomes) in close vicinity to the axodendritic synapses, either at pre- or post-synaptic positions (axonal terminals and dendritic spines, respectively), strongly indicating transsynaptic spread of the virus between connected neurons. Neuronal connectivity-based reconstruction of the directionality of transsynaptic virus spread suggests that, within the CNS, WNV can utilize both anterograde and retrograde axonal transport to infect connected neurons.
This study offers a new insight into the neuropathogenesis of WNV infection in a primate model that closely mimics WNV encephalomyelitis in humans. We show that within the primate CNS, WNV primarily infects the anatomical structures and pathways responsible for the control of movement. Our findings also suggest that WNV most likely propagates within the CNS transsynaptically, by both, anterograde and retrograde axonal transport.
| West Nile virus (WNV) is a mosquito-borne neurotropic flavivirus that has emerged as a human pathogen of global scale. During recent WNV outbreaks in the US, half of the reported human cases were classified as neuroinvasive disease. Although much research has been done, there are still gaps in our understanding of WNV neuropathogenesis. While WNV neuroinvasion is proposed to occur by the hematogenous route and/or by axonal transport along the peripheral nerves, how virus spreads once within the central nervous system (CNS) remains unknown. In this study, we examined the expression of viral antigens in the CNS of monkeys that were intrathalamically inoculated with WNV. Next, we mapped the localization of WNV-infected neurons to specific anatomical structures, identified the intraneuronal localizations of WNV particles and investigated the role of neuronal connectivity in the spread of WNV within the CNS. Our results revealed that all structures containing WNV-labeled neurons belonged to the pathways of motor control. Virions were found in close vicinity to the axodendritic synapses, strongly indicating transsynaptic spread of the virus. Neuronal connectivity-based reconstruction of the directionality of transsynaptic virus spread suggests that, within the CNS, WNV can utilize both anterograde and retrograde axonal transport to infect connected neurons.
| West Nile virus (WNV) is a mosquito-borne neurotropic flavivirus that has emerged as a human pathogen of global scale [1]. During the latest and largest outbreak of human WNV disease in US history during 2012–2013 [2,3], over half of the reported cases (51%) were classified as WNV neuroinvasive disease. Although much research has been done, there are still gaps in our understanding of WNV neuropathogenesis [4]. There is no consensus on how WNV infects the central nervous system (CNS). Two hypotheses of neuroinvasion are being considered: (i) hematogenous route and (ii) transneural entry through peripheral nerves. However, regardless of the mode of virus entry from the periphery, it is not clear how virus spreads once within the CNS.
Cell-to-cell spread of viruses contributes significantly to the pathogenesis of viral infections by facilitating virus dissemination and immune evasion [5]. The ability of several neurotropic viruses to spread between neurons via neuronal synapses and use axonal transport (anterograde, retrograde, or both) is well recognized and has been successfully exploited to trace neuronal connectivity. Of these viruses, the most studied are the alpha herpes viruses and rabies virus (reviewed in [6–9]). A study using an in vitro system of compartmentalized neuronal cultures showed that WNV can spread between neurons in both anterograde and retrograde directions via axonal transport [10]. However, how WNV spreads in vivo, especially within the CNS, is less clear. Transneuronal WNV spread was reported as a putative route of neuroinvasion after sciatic nerve inoculation in hamsters [10,11], but the brain was not studied in this model, and it was speculated that in the CNS, both anterograde and retrograde neuronal transport contributes to “centrifugal” spread of WNV among neurons in the brain [10]. It is also not clear how well rodent models reproduce all aspects of WNV neuropathogenesis in humans [4]. Nonhuman primates (NHP) represent a more suitable model due to their natural susceptibility to a wide range of human pathogens and the high degree of genetic similarity to humans [12]. However, primates do not develop neurological WNV disease after peripheral (either natural or experimental) infection [13–17]. On the other hand, the NHP model of neuroinfection, in which animals are inoculated intracerebrally, remarkably recapitulates the features of WNV encephalomyelitis seen in humans [15–19].
In humans with WNV neuroinvasive disease, the CNS structures that are often involved include (listed by an increasing gradient of the severity of infection): cerebral cortex (least severe), basal ganglia, thalamus, brainstem, cerebellum, and spinal cord (most severe) [20–24]. The possibility of WNV propagation via neuronal processes within the CNS was suggested from autopsy findings in an immunosuppressed patient with a fatal WNV encephalitis [25]. However, the connectivity between affected structures and its possible role in virus spread within the CNS have not been studied. Here, we used the CNS tissues from our previous NHP study [19], in which animals were inoculated intrathalamically with WNV, and investigated the role of the neuroanatomical connectivity in the spread of WNV within the brain and spinal cord.
The rhesus macaques (Macaca mulatta) used for this study were housed in a BSL-3 facility in compliance with the National Institute of Allergy and Infectious Diseases (NIAID), Division of Intramural Research (DIR) Animal Program Policy on Social Housing of Non-Human Primates, and Comparative Medicine Branch NHP enrichment programs. Animals were provided with commercial food pellets supplemented with appropriate treats. Drinking water was provided ad libitum. All steps were taken to minimize suffering. The experimental procedures requiring anesthesia were performed using ketamine hydrochloride or other anesthetics at the discretion of the attending veterinarian. For euthanasia, ketamine hydrochloride pre-anesthesia and sodium pentobarbital were used. The NIAID DIR Animal Care and Use Committee approved the animal study proposal (#LID 7E). The NIAID DIR Animal Care and Use Program, as part of the NIH Intramural Research Program (IRP), complies with all applicable provisions of the Animal Welfare Act (http://www.aphis.usda.gov/animal_welfare/downloads/awa/awa.pdf) and other Federal statutes and regulations relating to animals. The NIAID DIR Animal Care and Use Program is guided by the "U.S. Government Principles for the Utilization and Care of Vertebrate Animals Used in Testing, Research, and Training" (http://oacu.od.nih.gov/regs/USGovtPrncpl.htm). The NIAID DIR Animal Care and Use Program acknowledges and accepts responsibility for the care and use of animals involved in activities covered by the NIH IRP’s PHS Assurance #A4149-01, last issued 11/24/2014. As partial fulfillment of this responsibility, the NIAID DIR Animal Care and Use Program ensures that all individuals involved in the care and use of laboratory animals understand their individual and collective responsibilities for compliance with that Assurance, as well as all other applicable laws and regulations pertaining to animal care and use. The NIAID DIR Animal Care and Use Program has established and will maintain a program for activities involving animals in accordance with the most recent (2011, 8th edition) of “The Guide for the Care and Use of Laboratory Animals” (ILAR, NRC) (http://oacu.od.nih.gov/regs/guide/guide_2011.pdf). The policies, procedures and guidelines for the NIH IRP are explicitly detailed in NIH Policy Manual 3040–2, “Animal Care and Use in the Intramural Program” (PM 3040–2) and the NIH Animal Research Advisory Committee Guidelines (ARAC Guidelines). Those documents are posted on the NIH Office of Animal Care and Use public website at: http://oacu.od.nih.gov.
Our animal model of WNV neuropathogenesis in NHPs (Macaca mulatta; WNV-seronegative; 2–3 year old) that were inoculated intrathalamically (bilaterally) with a dose of 5.0 log10 PFU of wild-type WNV strain NY99-35262 (hereafter WNV) has recently been described [19]. We performed a systematic collection of all major CNS regions from these animals for downstream analyses. In this study, CNS tissues were examined by immunohistochemistry and electron microscopy. CNS tissues were from twelve WNV-infected monkeys (3 days post infection (dpi) [n = 3]; 7 dpi [n = 3]; and 9/10 dpi (9 dpi [n = 5]; 10 dpi [n = 1]) and two mock-inoculated monkeys (7 dpi [n = 1] and 10 dpi [n = 1]). WNV-infected animals developed a fulminant encephalomyelitis by 9/10 dpi (details of the clinical course, CNS virus burden, and histopathological scores can be found in our prior publication [19]). Brains and spinal cords were collected immediately after euthanasia and cardiac perfusion with sterile saline. After a parasagittal cut, the right brain hemisphere was fixed in 10% buffered formalin. Rhesus Mon-key Brain Matrix (Ted Pella, Redding, CA) was used to make 4 mm coronal brain slices that were further cut to facilitate mounting of subsequent sections onto standard 1 x 3 inches slides. Slices were routinely processed and embedded in paraffin. Two 5 μm sections (1st and 4th) from each paraffin block were mounted onto single slides and processed for immunohistochemistry. Spinal cord was dissected transversely and sections from cervical, thoracic, and lumbar regions were mounted onto single slides and also processed for immunohistochemistry.
Immunohistochemical detection of WNV antigens in the CNS of rhesus monkeys was performed using WNV-specific primary antibodies in hyperimmune mouse ascitic fluid (ATCC VR-1267 AF; 1:1000) and subsequent steps were according to previously described procedures [26]. Diaminobenzidine was used for colorimetric detection of WNV antigens. Sections were counterstained with hematoxylin.
Whole tissue section imaging was performed at 20x magnification using ScanScope XT (Aperio, Vista, CA). Aperio Spectrum Plus and ImageScope software was used for digital slide organization, viewing, and analysis. We analyzed all major CNS regions including: cerebral cortex, basal ganglia, thalamus, midbrain, pons, medulla oblongata, cerebellum (cerebellar cortex and deep cerebellar nuclei), and spinal cord (cervical, thoracic, and lumbar regions). The “Primate Brain Maps: Structure of the Macaque Brain” [27] were used for neuroanatomical orientation and mapping. To examine the WNV-immunoreactivity and to add to the visualization of WNV-antigen positive cells in the cerebellar cortex, a custom “WNV-labeled cell segmentation” image analysis algorithm was developed based on the ImageScope nuclear algorithm.
For ultrastructural analysis, core tissue samples (2 mm in diameter; 4 mm thick) were extracted using sterile Harris Uni-Cores (Ted Pella, Redding, CA). Samples that included the gray matter (wherever possible) were extracted from the following CNS regions: cerebral cortex, basal ganglia, thalamus, pons, medulla oblongata, cerebellar cortex, and spinal cord (cervical and lumbar regions). For the cerebellar cortex, core samples were extracted from the folia in a manner that included the molecular layer, Purkinje cell layer, and granule cell layer. For the spinal cord, core samples were extracted from the ventral horns.
Collected core tissue samples were fixed in 2.5% glutaraldehyde and 2% paraformaldehyde (Electron Microscopy Sciences, Hatfield, PA), then washed in Millonig’s sodium phosphate buffer (Tousimis Research, Rockville, MD), post-fixed in 1% osmium tetroxide (Electron Microscopy Sciences), stained en bloc with 2% uranyl acetate (Fisher Scientific, Waltham, MA), dehydrated in increasing concentrations of ethanol, and then infiltrated and embedded in Spurr plastic resin (Electron Microscopy Sciences). Embedded tissue samples were sectioned using a Leica UC7 Ultramicrotome (Leica Microsystems, Buffalo Grove, IL). Ultra-thin sections (60–80 nm in thickness) were collected, mounted onto 200 mesh copper grids, and contrasted with lead citrate (Fisher Scientific). The grids were then examined and imaged using a transmission electron microscope (FEI G2 Tecnai).
The method of circular representation, named a “connectogram”, is an intuitive and suitable approach for the visualization and interpretation of neuroanatomical connectivity using magnetic resonance imaging [28,29]. This type of representation is also highly suitable for visualization of complex neuroanatomical connections with an attempt to reconstruct virus spread between the infected CNS structures in this study. For this purposes, we adopted the connectogram idea and manually created our connectograms using Adobe Illustrator. The information used to create the connectograms is based on the literature review of established connectivity only between neuroanatomical structures relevant to this study.
Our first goal was to identify WNV-labeled cells using immunohistochemistry and then map their distribution to specific anatomical structures within the CNS. We did not detected WNV antigens at 3 dpi in any CNS region. WNV-labeled neurons became readily detectable in the CNS at 7 dpi and 9/10 dpi. WNV-infected CNS regions, anatomical structures/types of neurons, reference virus titers [19], extent/intensity and timing of WNV-labeling, as well as references to representative images in this report are summarized in Table 1. A general pattern of anatomical localization and extent of neuronal WNV-labeling closely followed the distribution and amounts of infectious virus at the same time points. To add to these comparisons, the changes in WNV loads within each major CNS region during the time course of neuroinfection are provided by radar charts in S1 Fig. One caveat of this study is the fact that we were unable to detect WNV antigens by immunohistochemistry at 3 dpi. This could be explained by the immunohistochemical limit of flavivirus detection of approximately 3 log10 PFU/g (personal observation and compare mean virus titers and average WNV-labeling in Table 1). Whether other neuronal cells could be infected at the levels below our limit of detection remains an open question.
WNV-labeling was detected in the neuronal cytoplasm and processes of the following anatomical structures and/or neuronal types: motor cortex (corticospinal motor neurons [Betz cells]); subcortical regions (neurons in the motor [ventrolateral] thalamus and basal ganglia); midbrain (substantia nigra pars compacta and red nucleus magnocellular); and pons/medulla oblongata (pontine nuclei, vestibular nuclei, medullary reticular formation, inferior olivary nuclei, and accessory cuneate nucleus) (Figs 1 and 2).
In the cerebellum, WNV-labeling was unambiguously detected in neurons of the deep cerebellar nuclei (Fig 3) and in the Purkinje cells (Fig 4 and S2 Fig). WNV-labeling of the granule neurons was much less frequent. To better appreciate the differences in WNV immunoreactivity between the infected cells of the cerebellar cortex, we developed a custom “WNV-labeled cell segmentation” image analysis algorithm, which produced markup images highlighting the intensity of WNV immunoreactivity in the Purkinje cells (Fig 4B and 4D) and granule neurons (Fig 4D).
Of note, despite the fact that only a few small groups of granule neurons were infected, we observed a substantial focal rarefaction of the granule cell layer at 9 dpi. This phenomenon cannot be explained by virus-induced cell death since only a small number of these cells were WNV-positive, nor can it be ascribed to a known artefact of granule cell dissolution due to postmortem autolysis. The latter is because necropsy and tissue fixation were performed immediately after euthanasia and also because our experiments were well controlled by inclusion of mock-inoculated animals that were euthanized at the same time points as WNV-infected animals (compare anti-NeuN immunostaining highlighting the rarefaction of the granule cell layer in WNV-infected animals and normal granule cell layer in mock-inoculated animals, S4 Fig). This phenomenon therefore remains unexplained and deserves further investigation.
In the spinal cord, WNV-labeling was detected in the lower motor neurons residing in the Rexed’s laminae IX of the ventral horns, spanning cervical, thoracic, and lumbar regions (Fig 5). However, in addition to this, an intriguing finding was that many spinocerebellar relay neurons that occupy a discrete nucleus known as Clarke’s column also contained large amounts of WNV antigens in their cytoplasm and transverse axonal profiles (Fig 5E).
These mapping results, when taken together, provided a detailed picture of WNV infection of the CNS. Interestingly, all structures identified as harboring WNV-labeled neurons are thought to participate in the control of movement (Table 2)
Of note, during the terminal stage of neuroinfection, WNV also infected the structures that relay proprioceptive signals from the upper parts of the body (accessory cuneate nucleus) and from the lower parts of the body (Clarke’s column) to the cerebellum. Clarke’s column (medial portion) also integrates the corticospinal inputs with relevance to motor planning and evaluation [46].
Remarkably, all WNV-infected structures and/or neuronal groups identified in this study were also reported to be affected in humans with WNV encephalomyelitis (i.e., cerebral cortex, basal ganglia, thalamus, substantia nigra, red nucleus, pons, vestibular nuclei, medulla, inferior olive, cuneate nucleus, Purkinje cells, dentate nucleus, Clarke’s column, and ventral horns of spinal cord [20–25,47–51]. Our findings are also in line with the pioneering studies of WNV encephalitis in intracerebrally inoculated nonhuman primates [15–18]. These early studies, although not well equipped to precisely detect specific groups of infected neurons, clearly showed infection of brainstem, cerebellum and spinal cord. Our findings provide the evidence that, within the primate CNS, WNV preferentially infects specific neuroanatomical structures responsible for the control of movement.
We next used electron microscopy (EM) to determine intraneuronal localization of WNV particles. It is generally accepted that in order to detect virus particles in tissue culture by electron microscopy (EM), the virus titers have to be high (i.e., 105 to 106 particles per milliliter) [52]. The same is true for detection of viruses in tissues. A major limitation of virus detection in tissues by EM is that the sampling might miss the areas containing viruses. With this in mind, we focused on the most heavily infected CNS regions with highest virus loads (i.e., cerebellum and spinal cord; see Table 1) from animals which developed fulminant encephalitis (at 9/10 dpi). In order to maximize the probability of virion detection, we used a targeted small-volume sampling of the gray matter for the ultrastructural analysis (see Materials and Methods). Tissue-core samples of the cerebellar cortex included the molecular layer, Purkinje cell layer, and granule cell layer. Tissue-core samples of the spinal cord contained the gray matter from the ventral horns.
Ultrastructural analysis confirmed WNV infection of the Purkinje cells in the cerebellar cortex and motor neurons in the ventral horns of the spinal cord. We also examined many other CNS regions with lower virus burden (i.e., cerebral cortex, basal ganglia, thalamus, pons, and medulla oblongata) but, as expected, were unable to detect virions.
Similar to well-defined structures that can be seen in non-polarized cells and have been linked to sites of virus replication [53–55], many infected neurons in this study showed smooth-membrane structures, convoluted membranes, and tubular structures that are characteristic of flavivirus infection. We also observed the formation of prominent paracrystalline arrays (S3 Fig). However, there were several unique findings related exclusively to WNV infection of neurons in this study:
It is also important to note that during EM examination of neurons, the cell organelles such as neuropeptide-containing dense-core vesicles [56] should not be mistaken for virions. This has been also emphasized by other investigators [57]. In this study, we often observed large dense-core vesicles in axons and axon terminals. Two features helped to distinguish between the dense-core vesicles and WNV particles: (i) each dense-core vesicle had a single membrane and (ii) dense-core vesicles were larger in diameter (75–120 nm) compared to WNV virions (40 nm) (compare Fig 7C and 7D, insets).
In human cases of WNV encephalitis, the visualization of WNV particles in neurons by electron microscopy is very rare, likely due to the difficulties in performing a sufficient sampling of particular tissue areas with high virus loads. Interestingly, when found, WNV particles were present in the cerebellar neurons (type of neurons was not specified) [22]. We found virions grouped in the vesicular structures within the dendrites (shafts and spines) as well as within axon terminals in very close vicinity to the synaptic clefts. To our knowledge, this is a first in vivo electron microscopy evidence suggesting transsynaptic spread of WNV between synaptically connected neurons in the primate CNS.
Another intriguing ultrastructural finding in this study was that WNV virions in the axon terminals were enclosed in the double-membrane vacuoles indicative of autophagosomes. The role of autophagy in WNV-infected cells in vitro is not clear [reviewed in [58]]; however, a recent study in a neonatal mouse model of WNV infection of the CNS showed that pharmacological activation of autophagy by a pro-autophagic peptide can protect against WNV-induced neuronal cell death and improve the clinical outcome [59]. To this end, our finding of WNV virions within the autophagosomes that were positioned pre-synaptically might indicate sequestration of virions for degradation [60] to prevent their transsynaptic release and infection of post-synaptic neuron. Alternatively, since the autophagosomes in neurons are initiated distally at axon terminals and fuse with late endosomes to form the amphisomes that are then transported retrogradely to reach acidic lysosomes in the cell soma [61,62], WNV particles encapsulated within the autophagosomes/amphisomes might take advantage by using retrograde axonal transport to the neuronal perikarya as a way of transneuronal spread. Whether this transport would result in the virion degradation upon fusion with lysosomes or would deliver the virions to the neuronal perikarya for successful subsequent replication remains to be investigated.
Intraneuronal movement of WNV most likely involves microtubules [63–65] and their associated motors for anterograde, retrograde, and/or bidirectional transport [66–69]. In agreement with this scenario, the analysis by electron microscopy in this study revealed virions inside the vesicles that were adjacent to microtubule structures (Fig 6B). Assuming that WNV is transported in neurons within the vesicular structures, it is possible that already established mechanisms for vesicular intraneuronal transport are in use [8]. For example, neuropeptide-containing dense core vesicles can move bidirectionally, switching between anterograde and retrograde axonal transport motors in a conveyor belt-like manner for continuous circulation [70,71]. As mentioned above, it is possible that WNV particles are captured by autophagosome formations at distal axons (whether delivered there by anterograde transport from neuronal soma or transmitted transsynaptically from a post-synaptic neuron) and then delivered by retrograde transport to the neuronal soma for degradation or release and replication.
Since the directionality of virus spread cannot be determined based on the “snapshots” revealed by immunohistochemistry and EM, we next attempted to reconstruct the directionality of transsynaptic spread of WNV based on the neuroanatomical connectivity between identified infected structures. For this purpose, we compiled known connectivity information and designed connectograms to capture and visualize the complex neuroanatomical connections between WNV-infected structures in an intuitive and concise way. We created two connectograms (Fig 8) showing a proposed directionality of WNV spread in our model based on the neuroanatomical connectivity and time of immunohistochemical virus detection (i.e., 7 and 9/10 dpi). Reference connectivity information is provided in the S1 Table.
Each connectogram contains two concentric rings and a black core. The names for each major CNS region are given on the outside periphery of outermost ring (i.e., subcortical regions, motor cortex, midbrain, spinal cord, cerebellum, medulla oblongata, and pons) in no particular order. For the outermost ring, each CNS region was assigned a spectrum domain color counterclockwise as follows: subcortical regions (red), motor cortex (orange), midbrain (yellow), spinal cord (green), cerebellum (light blue), medulla oblongata (blue), and pons (purple). The next ring toward the center of the connectogram is divided in sixteen segments. Each segment representing a specific structure/type of neurons includes an abbreviation and is assigned a unique color within the spectrum domain color of the corresponding larger anatomical CNS region. Within the black core, the directions of proposed spread of WNV between the structures/types of neurons are depicted by arrows (anterograde virus spread—solid arrows; retrograde virus spread—dashed arrows; anterograde/retrograde virus spread [due to existence of reciprocal connections]—white lines). Each arrow has the same color as the segment representing a structure/type of neurons from which it originates. The direction of the arrow indicates a proposed direction of virus spread between one neuronal order to the next. The proposed orders of neurons are indicated by circled numbers.
Assuming that the neurons of motor thalamus (Mthal) represent the neuronal order “0” (starter cells), the next neuronal order “1” will be neurons of the deep cerebellar nuclei (DCN), basal ganglia (BG), and corticospinal motor neurons (CSMN) (Fig 8A). The arrows connecting these neuronal groups show that there are three possibilities of the retrograde axonal virus spread (Mthal → DCN; Mthal → BG; and Mthal → CSMN) and one possibility of the anterograde axonal virus spread (Mthal → CSMN). The possibility of both retrograde and anterograde virus spread between Mthal and CSMN exists because of reciprocal connections between these structures. From the neurons of the order “1”, the virus spread to the next neuronal order “2” could also occur by the anterograde axonal transport (BG → SNC; CSMN → SMN; CSMN → Pn; DCN → MeRF; DCN → IO) and by the retrograde axonal transport (BG → SNC; DCN → Pn; DCN → IO; DCN → Purkinje cells).
By the terminal stage of neuroinfection (9/10 dpi; Fig 8B), the directionality of virus spread to the next neuronal orders could occur by anterograde axonal transport (CSMN → RnM; CSMN → CC; DCN → RnM; DCN → Ve; Pn → Granule cells) and by retrograde axonal transport (SMN/Cervical → RnM; SMN → Ve; Purkinje cells → Granule cells; Granule cells → ACu; Granule cells → CC). The maximal neuronal order reached by virus in this model is “4” (ACu and CC).
WNV infection of the neurons of accessory cuneate nucleus (ACu) in the medulla oblongata and Clarke’s column (CC) in the spinal cord is intriguing. These structures relay proprioceptive information from the upper (ACu) and lower (CC) parts of the body to the cerebellum. The axons of these relay neurons terminate as mossy fibers on the granule neurons. It is conceivable that these structures may have been infected by retrograde spread of the virus along the dorsal spinocerebellar tract from granule neurons. However, we found very few WNV-labeled granule neurons in the cerebellum (Fig 4C and 4D). Similarly, the granule cells do not appear to be infected in humans with WNV encephalomyelitis [20]. This is consistent with the recently reported enhanced antiviral response in the granule neurons and might explain their relatively low permissiveness to WNV infection [72]. These considerations suggest a limited contribution of granule neurons to the retrograde spread of WNV along the dorsal spinocerebellar tract to the proprioceptive relay neurons. Recently identified inputs from the descending corticospinal axons [46], better explain the infection of Clarke’s column. By analogy, the infection of the accessory cuneate nucleus could probably be also explained by the cortical inputs (yet to be identified) since this nucleus is the anatomical and functional correlate of Clarke’s column in the medulla.
In summary, our reconstruction of the directionality of WNV spread within the CNS of intrathalamically inoculated NHPs suggests both anterograde and retrograde axonal transport. The connectograms (Fig 8) show eleven possible routes of anterograde and twelve possible routes of retrograde axonal spread (with three unidentifiable directionalities between the same neuronal orders).
Collectively, our results of the anatomical and ultrastructural mapping of WNV neuronal infection in the primate CNS, together with the connectivity-based reconstruction of the directionality of virus spread strongly suggest the following:
Progression of WNV neuroinfection transsynaptically along specific pathways governing motor control in both, anterograde and retrograde directions suggested by this study may open a way for future therapeutic approaches. Although it seems unlikely that antiviral interventions would be justified for asymptomatic or self-limiting WNV cases in humans, we cannot neglect the necessity for development of rational treatments of flavivirus neurological disease. Our findings imply that the focus of such treatments should not only be on limiting virus replication but also on blocking neuron-to-neuron virus transmission, thus preventing further damage to the CNS.
|
10.1371/journal.pgen.1004816 | The Evolution of Fungal Metabolic Pathways | Fungi contain a remarkable range of metabolic pathways, sometimes encoded by gene clusters, enabling them to digest most organic matter and synthesize an array of potent small molecules. Although metabolism is fundamental to the fungal lifestyle, we still know little about how major evolutionary processes, such as gene duplication (GD) and horizontal gene transfer (HGT), have interacted with clustered and non-clustered fungal metabolic pathways to give rise to this metabolic versatility. We examined the synteny and evolutionary history of 247,202 fungal genes encoding enzymes that catalyze 875 distinct metabolic reactions from 130 pathways in 208 diverse genomes. We found that gene clustering varied greatly with respect to metabolic category and lineage; for example, clustered genes in Saccharomycotina yeasts were overrepresented in nucleotide metabolism, whereas clustered genes in Pezizomycotina were more common in lipid and amino acid metabolism. The effects of both GD and HGT were more pronounced in clustered genes than in their non-clustered counterparts and were differentially distributed across fungal lineages; specifically, GD, which was an order of magnitude more abundant than HGT, was most frequently observed in Agaricomycetes, whereas HGT was much more prevalent in Pezizomycotina. The effect of HGT in some Pezizomycotina was particularly strong; for example, we identified 111 HGT events associated with the 15 Aspergillus genomes, which sharply contrasts with the 60 HGT events detected for the 48 genomes from the entire Saccharomycotina subphylum. Finally, the impact of GD within a metabolic category was typically consistent across all fungal lineages, whereas the impact of HGT was variable. These results indicate that GD is the dominant process underlying fungal metabolic diversity, whereas HGT is episodic and acts in a category- or lineage-specific manner. Both processes have a greater impact on clustered genes, suggesting that metabolic gene clusters represent hotspots for the generation of fungal metabolic diversity.
| Fungi are important primary decomposers of organic material as well as amazing chemical engineers, synthesizing a wide variety of natural products, some with potent toxic activities, including antibiotics and mycotoxins. In fungal genomes, the genes involved in these metabolic pathways can be physically linked on chromosomes, forming gene clusters. This extraordinary metabolic diversity is integral to the variety of ecological strategies that fungi employ, but we still know little about the evolutionary processes involved in its generation. To address this question, we analyzed 247,202 enzyme-encoding genes participating in hundreds of metabolic reactions from 208 diverse fungal genomes to examine how two major sources of gene innovation, namely gene duplication and horizontal gene transfer, have contributed to the evolution of clustered and non-clustered metabolic pathways. We discovered that gene duplication is the dominant and consistent driver of metabolic innovation across fungal lineages and metabolic categories; in contrast, horizontal gene transfer appears highly variable both across organisms and functions. The effects of both gene duplication and horizontal gene transfer were more pronounced in clustered genes than in their non-clustered counterparts suggesting that metabolic gene clusters are hotspots for the generation of fungal metabolic diversity.
| As one of the primary decomposers of organic material in nature, fungal species catabolize a wide diversity of substrates [1], including cellulose and lignin, the two most abundant biopolymers on earth [2]. Fungi are also superb chemical engineers, capable of synthesizing a wide variety of metabolites, including amino acids, small peptides, pigments and other natural products with potent toxic activities, such as antibiotics and mycotoxins [3]–[6].
Fungal metabolites have historically been divided into primary, that is metabolites essential for growth and reproduction, and secondary, which include ecologically important metabolites not essential to cellular life [7], [8]. However, this distinction is arbitrary when applied to metabolic pathways rather than their products not only because the essentiality of a given pathway is species-specific [9] but also because the pathways that generate primary and secondary metabolites are not mutually exclusive [10], [11]. Perhaps more informatively, pathways can be divided into those shared by most organisms, which can be considered as belonging to general metabolism, and those specialized pathways that have evolved in response to the specific ecologies of certain lineages and, as a result, are more narrowly taxonomically distributed.
An intriguing feature of specialized metabolic pathways in fungi is that constituent genes are often physically linked on chromosomes forming what are known as gene clusters [12], [13]. Fungal metabolic gene clusters are distinct from the developmental gene clusters typically found in animal genomes, such as the Hox gene clusters; whereas animal gene clusters are composed of tandemly duplicated genes [14], [15], fungal metabolic gene clusters comprise genes that are evolutionarily unrelated. Fungal metabolic gene clusters participate in diverse activities including nitrogen [16], [17], carbohydrate [18], amino acid [19], and vitamin [12] metabolism as well as in xenobiotic catabolism [11], [20] and the biosynthesis of secondary metabolites [e.g.], [ 21]–[28].
Although this extraordinary metabolic diversity, whether in the form of clustered or non-clustered pathways, is integral to the entire spectrum of fungal ecological strategies (e.g., saprotrophic, pathogenic and symbiotic), we still know little about the evolutionary processes involved in its generation. Gene duplication (GD), a major source of gene innovation, is often implicated in the evolution of fungal metabolism [e.g.], [ 29]–[31], especially in the context of whole genome duplication (WGD) [32]–[34] and gene family expansion [35], [36]. Notable examples include the GD of enzymes involved in organic decay [30], starch catabolism [37], degradation of host tissues [31], [38], [39] and toxin production [36]. Repeated rounds of GD, followed by divergence and differential gene loss, have also been invoked to explain the evolution of the gene clusters that generate the diverse alkaloids produced by plant symbiotic fungi [4]. A second key source of metabolic gene innovation in fungi is horizontal gene transfer (HGT) [40]–[44]; significant cases include the transfer of genes involved in xenobiotic catabolism [45], [46], toxin production [45], [47], degradation of plant cell walls [48], [49], and wine fermentation [50]. More recently, HGT has been shown to be responsible for the transfer of entire metabolic gene clusters between unrelated fungi [11], [51]–[58].
Although both GD and HGT have been extensively studied in fungal genomes, how these two major sources of gene innovation have interacted with clustered and non-clustered metabolic pathways and sculpted their evolution is largely unknown. To address this question, we analyzed 247,202 enzyme-encoding genes from 208 diverse fungal genomes whose protein products participate in hundreds of metabolic reactions. We found that both GD and HGT were more pronounced in clustered genes than in their non-clustered counterparts. On average, 90.0% of clustered metabolic genes underwent GD and 4.8% underwent HGT, whereas 88.1% and 2.9% of non-clustered metabolic genes experienced GD and HGT, respectively. Remarkably, some genera appear to have undergone a larger number of HGT events than entire subphyla. While the effect of GD was largely stable across metabolic categories, HGT varied extensively. These results suggest that GD is the dominant and stable process underlying fungal metabolic diversity, whereas HGT's impact is more pronounced in specific lineages and metabolic categories. The disproportionate effect of GD and HGT on clustered genes renders metabolic gene clusters into hotspots of metabolic innovation and diversification in fungi.
Analysis of 208 fungal genomes identified 247,202 Enzyme Commission (EC)-annotated metabolic genes (ECgenes for short), which encoded proteins catalyzing 875 distinct enzymatic reactions in 130 metabolic pathways (Figure 1; Table S1; Table S2). The percentage of the fungal proteome dedicated to metabolism was 15.4% in Saccharomycotina, 12.6% in Pezizomycotina and 8.9% in Agaricomycetes (Table S3; Figure S1).
Examination of fungal metabolism for the presence of metabolic gene clusters revealed that 3.0% (7,409) of ECgenes belonged to 3,408 distinct gene clusters, with the average genome containing 16.7 metabolic gene clusters and 36.3 clustered ECgenes (Table S3). The percentage of clustered ECgenes was highly variable across the major lineages, being more than two-fold greater in the two Ascomycota lineages, namely Pezizomycotina (3.6% of ECgenes) and Saccharomycotina (3.7%), than in Agaricomycetes (1.6%) (Figure 1, Table S3). For example, the plant pathogen Fusarium solani species complex species 11 (a.k.a., Nectria haematococca, Sordariomycetes) had 152 clustered ECgenes (representing 6.2% of its ECgenes), the most of any genome analyzed, the yeast Torulaspora delbrueckii (Saccharomycotina) had 59 clustered ECgenes (7.3%), whereas the ectomycorrhizal fungus Laccaria bicolor (Agaricomycetes) had only 14 clustered ECgenes (1.1%).
To test whether clustering was variable across fungal metabolism, we used the Kyoto Encyclopedia of Genes and Genomes (KEGG) metabolism hierarchy [10] to assign all ECgenes to 12 overlapping, higher-order metabolic categories (carbohydrate, energy, lipid, nucleotide, amino acid, glycan, cofactor/vitamin, terpenoid/polyketide, other secondary metabolite, xenobiotics, biosynthesis of secondary metabolites, and microbial metabolism in diverse environments). We found that the proportion of clustered ECgenes varied significantly across metabolic categories (Figure 2, Table S4). For example, clustered ECgenes from all lineages were significantly overrepresented in the KEGG categories carbohydrate and terpenoid/polyketide and underrepresented in the glycan category. In addition, the proportion of clustered ECgenes in a given category often varied significantly between lineages. For example, clustered ECgenes in the nucleotide and xenobiotic categories were only significantly overrepresented in Saccharomycotina and Agaricomycetes; clustered ECgenes in the same categories were underrepresented in Pezizomycotina (Figure 2). Similarly, clustered ECgenes in the amino acid and lipid categories were underrepresented in Saccharomycotina, whereas clustered ECgenes in these same categories were overrepresented in Pezizomycotina and Agaricomycetes (Figure 2).
To evaluate the impact of GD and HGT on fungal metabolism, we inferred GD and HGT events by reconciling the gene tree of each ECgene to the fungal species phylogeny [59]–[61]. Specifically, we assigned costs to GD, HGT, gene loss, and incomplete lineage sorting (ILS) and determined the most parsimonious combination of these four events to explain the ECgene tree topology given the consensus species phylogeny. Therefore, HGT events were inferred only when an ECgene tree topology was contradictory to the species phylogeny and could not be more parsimoniously reconciled using a combination of differential GD and gene loss. We evaluated multiple HGT costs and ultimately implemented a cost four times greater than the GD cost because it was the lowest HGT cost that recovered three published cases of HGT without any additional (e.g., potentially spurious) cases of HGT in the corresponding ECs (Table S5).
On average, 88.7% of ECgenes per genome were inferred to have undergone one or more GD events (Table S3). This percentage was lower in early diverging lineages; this was the case for both taxa with typical gene densities (e.g., Chytridiomycetes) as well as for the extremely reduced microsporidians, which displayed the lowest percentages of duplicated metabolic genes (49.0% and 49.5% of ECgenes in E. cuniculi and E. intestinalis, respectively). While the low percentages of GD in microsporidians are likely explained by genome streamlining, the low percentages observed in other early diverging lineages are harder to explain, although we note that their current sparse representation in the set of sequenced fungal genomes increases the uncertainty associated with estimating GD and HGT. In contrast, 93.7% of ECgenes underwent GD in the Agaricomycetes (Figure 1), with the button mushroom, Agaricus bisporus, having 97.0% of its ECgenes affected by GD (704 to 722 ECgenes depending on the strain). GD percentage was also high in the Saccharomycotina (91.4%; Figure 1), including in species belonging to the Saccharomyces sensu stricto group, where the average increased to 95.3%, most likely as a consequence of an ancient whole genome duplication [33], [62].
Our analysis also identified that on average 2.8% of ECgenes per genome had undergone one or more HGT events (Table S3), which could be traced back to 823 unique HGT events. The Pezizomycotina showed the highest percentage of HGT of all the major lineages, with an average 4.1% of ECgenes transferred per genome, and Saccharomycotina the lowest, with an average 1.8% of ECgenes transferred (Table S3; Figure 1). Remarkably, some Pezizomycotina genera showed nearly as many or more HGT events than the entire Saccharomycotina subphylum (Figure 3; Figure S2). For example, we identified 111 HGT events since the last common ancestor of the 15 Aspergillus species, the largest for any genus included in our analysis, but only 60 HGT events since the last common ancestor of the 48 Saccharomycotina genomes. Notwithstanding the fact that genome coverage and age are not the same across fungal genera, several other Pezizomycotina genera showed an abundance of HGT events including Cochliobolus (53 HGTs; 8 genomes), Fusarium (52 HGTs; 4 genomes), and Trichoderma (50 HGTs; 6 genomes). Within the Agaricomycetes, the highest concentration of HGT events was observed in the two Agaricus bisporus genomes (23 HGTs).
Examination of the degree to which GD and HGT have differentially impacted clustered and non-clustered metabolic genes revealed significant differences (Figure 4; Table S6). On average, 90.0% of clustered ECgenes and 88.1% of non-clustered ECgenes underwent GD (P = 4.58×10−4). Similarly, 4.8% of clustered ECgenes underwent HGT compared to 2.9% of non-clustered ECgenes (P = 4.02×10−12). Examination of the impact of GD and HGT in the three major lineages shows that only in the Pezizomycotina was the percentage of GD and HGT significantly higher for clustered ECgenes than for non-clustered ECgenes (GD: 93.3% for clustered ECgenes versus 89.5% for non-clustered, P = 1.74×10−11; HGT: 6.6% for clustered ECgenes versus 4.0% for non-clustered, P = 2.77×10−10), suggesting that the trend is largely driven by Pezizomycotina. In fact, in both Saccharomycotina and Agaricomycetes GD was more common in non-clustered ECgenes than in clustered ECgenes (P = 0.02 and P = 0.01, respectively; Figure 4). HGT was more common in Saccharomycotina non-clustered ECgenes than in clustered ones, whereas in Agaricomycetes a higher incidence of HGT events was observed in clustered ECgenes, although neither of these associations was statistically significant (P = 0.54 and P = 0.16, respectively; Table S6).
To test whether GD and HGT prevalence varied across fungal metabolism, we examined the rates of the two processes in each of the 12 KEGG metabolic categories across our three major lineages. We found that the effect of GD was generally consistent across metabolic categories, with 9/12 categories showing the same pattern of under/overrepresentation of duplicated ECgenes across the three lineages (Figure 2, Table S4). Specifically, the categories carbohydrate, glycan, and biosynthesis of secondary metabolites were overrepresented, the categories lipid, nucleotide, cofactor/vitamin, other secondary metabolites, and xenobiotics were underrepresented, whereas energy was not differentially represented in duplicated and non-duplicated ECgenes in all three lineages.
Unlike GD, HGT differentially affected metabolic categories in a lineage-specific fashion, with 10/12 categories differing in the pattern of under/overrepresentation of duplicated ECgenes across lineages (Figure 2, Table S4). For example, ECgenes in biosynthesis of secondary metabolites were overrepresented for HGT events in Pezizomycotina and Saccharomycotina, but not in Agaricomycetes. In contrast, ECgenes were overrepresented for HGT in lipid and terpenoid/polyketide in Agaricomycetes but underrepresented in the Pezizomycotina. Only 2 categories, amino acid and microbial metabolism in diverse environments, were overrepresented in transferred ECgenes across all three lineages.
Determining the relative role of GD and HGT with clustered and non-clustered metabolic pathways is important for understanding the evolution of the fungal metabolic repertoire. Examination of the synteny and evolutionary history of 247,202 ECgenes from 875 metabolic reactions across fungal diversity showed that GD is the dominant source of metabolic gene innovation in fungi, whereas HGT is variable across metabolic categories and fungal lineages. Both GD and HGT are more pronounced in clustered genes than in their non-clustered counterparts, suggesting that metabolic gene clusters can act as hotspots for the generation of fungal metabolic innovation.
On average 88.7% of fungal ECgenes retain the signature of one or more GD events in their ancestry compared to only 2.8% for HGT (Table S3). Even though these percentages are not directly comparable because reconciliation of ECgene histories with the species phylogeny requires that costs are assigned for every inferred GD or HGT event [60], our finding that nearly nine out of every ten metabolic genes have undergone GD suggests that this is the dominant source of gene innovation underlying fungal metabolism. These results are consistent with the hypothesis that specialized metabolic pathways evolve via GD from general metabolic precursors. Support for this hypothesis has come from phylogenetic analysis of single gene families [63], [64] such as the polykeytide synthases, which share a common evolutionary origin with the fatty acid synthases of general metabolism [65]. Further diversification of genes involved in specialized pathways may occur through additional duplication, functional divergence and differential loss in response to variable ecological pressures as has been proposed for polyketide, nonribosomal peptide and alkaloid biosynthesis genes [4], [66]–[68].
Our analysis showed that certain lineages in the Pezizomycotina and Agaricomycetes have increased HGT rates. Interestingly, bacteria-to-fungi HGT events are also elevated within Pezizomycotina, particularly in Fusarium and Aspergillus genomes [43]. HGT of entire chromosomes has been reported in Fusarium [69], [70], a genus in our analysis, which in addition to Aspergillus, Cochliobolus and Magnaporthe, appears not only receptive to HGT but also includes highly virulent plant and animal pathogens, ecological lifestyles associated with many known cases of HGT [11], [45], [47], [51], [69]–[71]. Similarly, mycoparasitism in the genus Trichoderma may also provide ecological opportunities for fungal-to-fungal HGT.
GD alone or in combination with HGT affected nearly every reaction in fungal metabolism (727, 95.7% of ECs that passed the phylogenomic analysis; Figure 5). The effect of both GD and HGT varied between metabolic categories, suggesting that some pathways may tolerate the introduction of new genes better than others. One possible explanation for this variation is that the metabolic networks associated with the different functional categories have different degrees of connectivity. Genes whose products make up large protein complexes or that have many interacting partners exhibit less variation in copy number [35], perhaps because unbalanced increases in gene dosage can lead to malformed protein complexes and a buildup of toxic intermediates in metabolic pathways [72]–[74], and might be less likely to undergo GD [75], [76] as well as HGT [77]. In addition to gene dosage effects, deleterious interactions between native and horizontally acquired proteins that function as parts of multi-protein complexes, and as a consequence have distinct co-evolutionary histories, are likely also important barriers to HGT [77], [78].
Another possible explanation is that the source of the variation of GD and HGT lies in the differing functions encoded by these metabolic categories. Gene innovation is often correlated with molecular function, with informational genes such as those involved in DNA replication, transcription and translation duplicated and transferred less often than metabolic genes [35], [76], [78]. Within metabolism, one might expect that widely distributed pathways involved in universal metabolic functions, such as oxidative phosphorylation and the citric acid cycle, are more likely to be functionally constrained and, as a consequence, less likely to tolerate GD or HGT of their constituent genes. In contrast, GD and HGT might be more advantageous for specialized metabolic pathways that are under strong selection in fluctuating environments [11].
33 EC reactions are associated with 332 ECgenes that are never duplicated or transferred in our analysis; 31 of these 33 reactions (93.9%) are also never clustered (Table S7a). For the majority of these ECs, the reason for the apparent lack of GD or HGT is because they are represented by only a few ECgenes in our analysis; therefore, their ECgene trees consist of few taxa with topologies in agreement with the consensus species phylogeny. For other EC reactions in this set, strong selection pressure to maintain a single, native gene copy could explain the lack of GD and HGT. Only three genes annotated with EC reaction numbers and which were never duplicated or transferred in our analysis were present in the Saccharomyces cerevisiae genome (YNL219C [2.4.1.259], YBR003W [2.5.1.83], and YPR184W [3.2.1.33]). When examined against the yeast phenotype and interaction data from the Saccharomyces Genome Database (http://www.yeastgenome.org), these three genes displayed a variety of phenotypes and all their null mutants were viable (Table S7b). Interestingly, overexpression of two of the ECgenes (YNL219C [2.4.1.259] and YBR003W [2.5.1.83]) resulted in reduced rate of vegetative growth in S. cerevisiae (Table S7b), suggesting that the acquisition of additional gene copies through GD or HGT could be disadvantageous. Furthermore, one S. cerevisiae ECgene, a glycosyltransferase (YNL219C [2.4.1.259]) involved in the biosynthesis of asparagine-linked glycans, has a very complex interaction network of 315 described physical and genetic interactions (Table S6a), which could serve as an additional barrier to GD and HGT.
3.0% of fungal genes examined in our study lie within gene clusters. This is likely a conservative estimate because ECgene annotation is better for general rather than specialized metabolism. Although our analysis includes many specialized pathways (Table S2), such as biotin production (KEGG map00780), nitrate assimilation (map00910) and terpenoid backbone biosynthesis (map00900), and the fraction of enzymatic reactions encoded by clustered ECgenes is extensive (441 reactions, 50.4% of ECs; Figure 5), lineage-specific genes involved in specialized metabolic pathways are less likely to be included. In addition, fungal metabolic gene clusters are often identified through the presence of one or more conserved synthesis genes (e.g., genes encoding polyketide synthase or nonribosomal peptide synthase enzymes); proper demarcation of associated genes encoding modifying enzymes (e.g., oxidases and transferases) is challenging because they often lack functional annotation and are lineage-specific, leading to underestimates of gene cluster size.
Gene clustering in fungi is positively associated with both GD and HGT, but this pattern appears to be driven by Pezizomycotina ECgenes (Figure 4). Saccharomycotina ECgenes cluster more often than the global fungal average but are less often affected by HGT, whereas Agaricomycetes display the opposite trend; they experience more HGT but less gene clustering (Figure S3). GD affects nearly all ECgenes, and this large sample size undoubtedly contributes to the statistical significance of its association with gene clustering, even though the fold increase in the percentage of GD events observed in clustered versus non-clustered ECgenes is only 1.02. In contrast, the effect of HGT on clustered genes is 1.66 fold greater than its effect on non-clustered genes.
The uniqueness and wide distribution of fungal metabolic gene clusters has given rise to many models that attempt to explain their formation and maintenance [53], [79]–[83]. For example, the selfish gene cluster model proposes that HGT allows gene clusters to avoid being lost by facilitating colonization of new genomes [84], [85]. Although several instances of HGT of fungal gene clusters have been discovered in recent years [11], [51]–[58], clustered pathways are also more likely to be lost than non-clustered ones [53]. The small percentage of clustered genes affected by HGT in our analysis (4.8%), albeit larger than the background percentage of transferred un-clustered genes (2.9%), suggests that selfishness is unlikely to be the predominant mechanism driving gene cluster formation and maintenance in fungi. Nevertheless, the association between metabolic gene clusters and GD/HGT suggests that gene clustering can facilitate the duplication and transfer of entire metabolic pathways. This is consistent with the view that the barriers to gene innovation acting on gene clusters may be lower than those acting on single genes because the latter undergo GD or HGT in the absence of their functional partners.
A custom enzyme classification pipeline assigned EC numbers to protein-coding genes from the genomes of 208 fungi and 9 stramenopiles (five oomycetes and four algal relatives), which were included in this analysis because of published reports of HGT between oomycetes and fungi [44]. Each gene was queried against a database of KEGG orthology (KO)-annotated proteins from 53 KEGG Organisms (Table S8) using ublast (http://drive5.com/usearch) with an accel setting of 0.7 and minimum identity cutoff of 0.3. A KO term was assigned to the query for ublast hits with greater than 80% sequence identity and no more than 10% difference in length. In cases where highly similar matches were not recovered, KO terms were assigned to query sequences with respect to the ublast hits showing the lowest e-values; all ublast hits that followed the first e-value increase of 10−50 or greater were excluded. EC numbers were assigned according to KO term (http://www.genome.jp/kegg-bin/get_htext?ko00001.keg).
Fungal proteomes were screened for metabolic gene clusters as described [81]. Briefly, two ECgenes were considered clustered if they were separated by no more than 6 intervening genes according to published annotation and their EC numbers were nearest neighbors in one or more KEGG pathways. Gene clusters were inferred by joining overlapping metabolic gene pair ranges that were separated by no more than 6 intervening genes; the cutoff of 6 intervening genes was determined empirically with reference to previous analyses of both primary [52], [53] and secondary [54] metabolism clusters.
We constructed a draft fungal species phylogeny using protein sequences of the widely used DNA-directed RNA polymerase II subunit RPB2 marker, which were aligned with mafft using the E-INS-i strategy [86]. The resulting alignment was trimmed with trimal using the automated1 strategy [87], and the topology was inferred using maximum likelihood (ML) as implemented in raxml version 7.2.8 [88] using a PROTGAMMALGF substitution model and rapid bootstrapping (100 replications). Branches with bootstrap support less than 50 were collapsed using the Consense module in the phylip program [89]. The final bifurcating and consensus (multifurcating) species phylogenies (File S1) were constructed by making targeted corrections to the RPB2 topology based on published literature (Table S9).
ECgene trees were constructed using a custom phylogenomic pipeline (Figure S4). Guide trees were first constructed for each ECgene family with mafft using the scores of pairwise global alignments [86] and rooted with the notung rooting optimization algorithm using event parsimony. This distance-based guide tree and the consensus species phylogeny were used to delineate groups of homologs by aiming to maximize taxonomic diversity while minimizing the number of paralogs in each gene tree. The ECgene sequences from each one of these groups of homologs were then extracted in FASTA format for phylogenomic analysis. FASTA files of ECgenes with less than 4 or more than 1000 sequences were excluded. Sequences were aligned in mafft using the auto strategy selection [86]. Alignments were trimmed in trimal using the automated1 trimming strategy [87], and trimmed alignments shorter than 150 amino acid residues were discarded. Phylogenetic trees were constructed using fasttree [90] with a WAG+CAT amino acid model of substitution, 1000 resamples, four rounds of minimum-evolution subtree-prune-regraft moves (-spr 4), and the more exhaustive ML nearest-neighbor interchange option enabled (-mlacc 2 –slownni).
Gene tree-species phylogeny reconciliation was performed in notung using its duplication, transfer, loss and ILS aware parsimony-based algorithm [59]–[61], [91]. Ambiguity in the fungal species phylogeny and low branch support in ECgene trees were handled through a multi-step approach. First, ECgene tree branches with less than 0.90 SH-like local support were collapsed using treecollapsercl v4 (http://emmahodcroft.com/TreeCollapseCL.html). This collapsed ECgene tree was rooted and its polytomies resolved against the bifurcating species phylogeny. This resolved ECgene tree was then reconciled to the multifurcating, consensus species phylogeny using a duplication cost of 1.5, loss cost of 1 and ILS cost of 0. Transfer costs of 2, 4, 6, 8, 10 and 12 as well as the option to prune taxa not present in the gene tree from the species phylogeny were evaluated. A transfer cost of 6 with the prune option enabled best recovered published cases of HGT between fungi (Table S5). Percent GD and HGT were expressed over the 152,835 fungal ECgenes that passed this reconciliation pipeline. Because a single ancestral HGT event could be recorded in multiple ECgene trees, we defined unique HGT events as all cases where ECgenes assigned to the same EC number were inferred to have undergone HGT to/from the same recipient/donor nodes in the species phylogeny.
Fisher's exact tests were performed using the R function fisher.test with a two-sided alternative hypothesis [92]. P values were adjusted for multiple comparisons using the R function p.adjust with the Benjamini & Hochberg (BH) method [93]. Box-and-whisker plots were created using the R plotting system ggplot2 [94].
|
10.1371/journal.pgen.1003278 | Regulation of Neurod1 Contributes to the Lineage Potential of Neurogenin3+ Endocrine Precursor Cells in the Pancreas | During pancreatic development, transcription factor cascades gradually commit precursor populations to the different endocrine cell fate pathways. Although mutational analyses have defined the functions of many individual pancreatic transcription factors, the integrative transcription factor networks required to regulate lineage specification, as well as their sites of action, are poorly understood. In this study, we investigated where and how the transcription factors Nkx2.2 and Neurod1 genetically interact to differentially regulate endocrine cell specification. In an Nkx2.2 null background, we conditionally deleted Neurod1 in the Pdx1+ pancreatic progenitor cells, the Neurog3+ endocrine progenitor cells, or the glucagon+ alpha cells. These studies determined that, in the absence of Nkx2.2 activity, removal of Neurod1 from the Pdx1+ or Neurog3+ progenitor populations is sufficient to reestablish the specification of the PP and epsilon cell lineages. Alternatively, in the absence of Nkx2.2, removal of Neurod1 from the Pdx1+ pancreatic progenitor population, but not the Neurog3+ endocrine progenitor cells, restores alpha cell specification. Subsequent in vitro reporter assays demonstrated that Nkx2.2 represses Neurod1 in alpha cells. Based on these findings, we conclude that, although Nkx2.2 and Neurod1 are both necessary to promote beta cell differentiation, Nkx2.2 must repress Neurod1 in a Pdx1+ pancreatic progenitor population to appropriately commit a subset of Neurog3+ endocrine progenitor cells to the alpha cell lineage. These results are consistent with the proposed idea that Neurog3+ endocrine progenitor cells represent a heterogeneous population of unipotent cells, each restricted to a particular endocrine lineage.
| Diabetes mellitus is a family of metabolic diseases that can result from either destruction or dysfunction of the insulin-producing beta cells of the pancreas. Recent studies have provided hope that generating insulin-producing cells from alternative cell sources may be a possible treatment for diabetes; this includes the observation that pancreatic glucagon-expressing alpha cells can be converted into beta cells under certain physiological or genetic conditions. Our study focuses on two essential beta cell regulatory factors, Nkx2.2 and Neurod1, and demonstrates how their genetic interactions can promote the development of other hormone-expressing cell types, including alpha cells. We determined that, while Nkx2.2 is required to activate Neurod1 to promote beta cell formation, Nkx2.2 must prevent expression of Neurod1 to allow alpha cell formation. Furthermore, the inactivation of Neurod1 must occur in the earliest pancreatic progenitors, at a stage in the differentiation process earlier than previously believed. These studies contribute to our understanding of the overlapping gene regulatory networks that specify islet cell types and identify the importance of timing and cellular context for these regulatory interactions. Furthermore, our data have broad implications regarding the manipulation of alpha cells or human pluripotent stem cells to generate insulin-producing beta cells for therapeutic purposes.
| The destruction or dysfunction of the insulin-producing beta cells of the pancreas contributes to a family of metabolic diseases known as diabetes mellitus. Given that the specification of the three major cell types in the pancreas, endocrine, exocrine and ductal cells, occurs in the embryo, understanding the normal course of pancreas development will ultimately facilitate the generation of insulin-producing beta cells from alternative cell sources for beta cell replacement therapies [1], [2], [3]. Single knockout mouse models have determined the relative importance of many transcription factors in the process of endocrine cell specification and differentiation. Of particular significance, deletion of the basic helix-loop-helix transcription factor Neurogenin3 (Neurog3; Ngn3) results in the loss of the hormone-producing cell types [4]. Subsequent lineage tracing experiments confirm that hormone-expressing endocrine cell types, including alpha cells (expressing glucagon), beta cells (insulin), delta cells (somatostatin), epsilon cells (ghrelin), and PP cells (pancreatic polypeptide), are Neurog3-derived [5], [6].
A recent study suggested that each Neurog3+ endocrine progenitor cell within the population is destined to become a single hormone+ cell type [7]. The idea that endocrine progenitor cells are unipotent implies that the transcription factor code responsible for the differentiation of each hormone+ cell type may be delineated before endocrine progenitors are specified. In support of this hypothesis, forced expression of factors within the Pdx1+ pancreatic progenitor cells can affect the resulting complement of differentiated endocrine cells [8], [9], [10]. Ultimately, the proper timing and location of transcription factor expression and function during pancreas development is essential for the appropriate differentiation of all the hormone-expressing endocrine cells.
The homeobox transcription factor Nkx2.2 is a particularly interesting pancreatic regulatory protein due to its dynamic expression pattern and cell-specific regulatory activities. Nkx2.2 is widely expressed throughout the early undifferentiated pancreatic epithelium, but gradually becomes restricted to beta cells and a large subset of alpha and PP cells [11], [12]. Despite its early and widespread expression, deletion of Nkx2.2 specifically affects later endocrine lineage specification: beta cells do not form, alpha and PP cell numbers are decreased, and there is a significant increase in the ghrelin cell population. Furthermore, while Nkx2.2 is expressed in both glucagon+ alpha cells and insulin+ beta cells [13] and the physical interaction of Nkx2.2 with the co-repressor Groucho3 (Grg3; Tle3) occurs in both cell types, the recruitment of a repressor complex to the promoter of the homeobox transcription factor Arx occurs in beta, but not alpha cells [14], presumably due to cell-specific and/or promoter-specific protein interactions. Disruption of the Nkx2.2/Grg3 interaction results in the mis-specification of islet cell types and the subsequent trans-differentiation of beta cells into alpha cells [14]. Studies of other developmental systems, including muscle and CNS, have also provided examples of how a single transcription factor can differentially regulate cell specification [15], [16], [17], [18]. Altogether these studies demonstrate that cell-specific transcription factor regulation plays a fundamental role in cell fate determination and the maintenance of cell identity.
While single knockout mouse models can uncover the role of a specific factor in the process of cell fate determination [19], [20], [21], compound deletion mutants demonstrate how multiple transcription factors work together to permit or restrict the differentiation of specific lineages. Whereas the deletion of Arx results in the loss of alpha cells and an increase in beta and delta cells [19], [22], deletion of Nkx2.2 affects all islet cell types in the pancreas except the delta cell population [12]. Interestingly, simultaneous deletion of these two factors revealed for the first time that Nkx2.2 was required to repress somatostatin in the ghrelin-expressing epsilon cell lineage [23], [24]. Furthermore, the simultaneous deletion of Nkx2.2 and the beta cell transcription factor Neurod1 identified an unexpected epistatic relationship between these factors that regulates the formation of the non-beta cell types [25]. While deletion of Neurod1 does not affect the formation of alpha or beta cells, alpha cells are reduced late in development and beta cells undergo catastrophic apoptosis by birth [26]. In contrast, the null mutation of Nkx2.2 results in a severe reduction in alpha cells, and beta cells are completely absent [12], [27]. Despite the expression of Nkx2.2 and Neurod1 in beta cells [13], [26], [28] and the severe phenotypes associated with beta cells in both single knockout mice [12], [26], the simultaneous deletion of Neurod1 and Nkx2.2 did not alter the beta cell phenotype but rather restored alpha cell and PP cell formation, while simultaneously reducing the ghrelin-expressing epsilon cells, which are over abundant in the Nkx2.2 null pancreas [25]. These examples demonstrate that deciphering the complex pancreatic gene regulatory network will provide valuable insight into the cellular processes required to generate each islet cell type, and will facilitate the in vitro differentiation of functional insulin-producing cells for therapeutic purposes.
The Nkx2.2−/−;Neurod1−/− (Nkx2.2null;Neurod1null) compound mutant provides a useful model for how two transcription factors coordinately regulate the specification of multiple endocrine cell types. Our study aimed to dissect the cooperative roles of Nkx2.2 and Neurod1, and determine specifically where and how these factors work together to permit endocrine cell formation in the pancreas. The result of this analysis demonstrated that in the absence of Nkx2.2, deletion of Neurod1 in the Pdx1+ pancreatic progenitors resulted in restoration of the alpha, PP and epsilon cells; however, deletion of Neurod1 from the Neurog3+ endocrine progenitor cells restored the PP and epsilon cells, but only a small population of alpha cells. Using in vitro reporter assays we also showed that Nkx2.2 repressed Neurod1 in certain cellular contexts. Consistent with the idea that Neurog3+ cells are unipotent [7], we hypothesize that Nkx2.2 must repress Neurod1 in the Pdx1+ pancreatic progenitors early in development to appropriately prime the Neurog3+ endocrine progenitor cells to become alpha cells.
To determine the precise cell type in which the genetic interaction between Nkx2.2 and Neurod1 is required for endocrine cell specification, we conditionally removed Neurod1 from different pancreatic cell populations in the absence of Nkx2.2. We generated a pancreas-specific deletion of Neurod1 in the Nkx2.2 null background using Pdx1-cre [29] (Nkx2.2−/−;Neurod1flox/flox;Pdx1-cre, denoted as Nkx2.2null;Neurod1Δpanc). We first confirmed that the single deletion of Neurod1 in the Pdx1+ cells (Neurod1Δpanc) phenocopied the Neurod1null mouse (Figure 1B, 1F, 1J; Figure S1), displaying the expected reduction in insulin and glucagon mRNA levels at P0 (Figure 1M; Figure S1) [26], [30]. We also demonstrated that when Neurod1 was deleted from Pdx1+ cells in the absence of Nkx2.2, the pancreas phenotype was identical to the Nkx2.2null;Neurod1null mouse [25] (Figure S1). Specifically, all beta cells were absent, alpha and PP cells were restored, and epsilon cells, which were overabundant in the Nkx2.2null, were significantly reduced (Figure 1A–1L; Figure S1). The partial rescue of the epsilon cells is likely due to the inability of Neurod1 deletion to restore the balance between the epsilon and beta cell populations, similar to the Nkx2.2null;Neurod1null mice (Figure 1N; Figure S1; [25]). Hormone expression was quantified using real time PCR and cell numbers were determined with morphometric analysis; these analyses confirmed that the observed gene expression and cellular changes were equivalent between the Nkx2.2null;Neurod1Δpanc and the Nkx2.2null;Neurod1null (Figure 1M–1O; Figure S1). Moreover, we confirmed that Neurod1 was appropriately deleted in mutants and controls (Figure 1P). These data demonstrate that in an Nkx2.2 null background the deletion of Neurod1 in the pancreas progenitors phenocopies the Nkx2.2null;Neurod1null.
Given that all hormone-producing endocrine cells are Neurog3-derived [4], [5], [6], we hypothesized that the genetic interaction between Nkx2.2 and Neurod1 would be required within the Neurog3+ endocrine progenitors to allow for the specification of particular hormone+ cell types. Using the Neurog3-cre allele [31], we generated an endocrine progenitor cell-specific deletion of Neurod1 in the Nkx2.2 null background (Nkx2.2−/−;Neurod1flox/flox;Neurog3-cre, denoted as Nkx2.2null;Neurod1Δendo), and assessed the pancreatic endocrine cell phenotype. To achieve optimal recombination in the Neurog3-expressing precursor population, we used the BAC-derived Neurog3-cre allele; Cre is highly co-expressed with Neurog3 in the embryonic pancreas and Cre activity is sufficient to lineage-label all pancreatic endocrine cells in the islet [31]. Importantly, despite the short half-life of Neurog3 protein, we can detect Cre activity in approximately 75% of Neurog3-expressing cells (Figure S2B). Similar to the Nkx2.2null;Neurod1Δpanc and Nkx2.2null;Neurod1null mice, we observed rescue of PP cells (Figure 2A, 2B), and a large reduction of ghrelin+ epsilon cells in the Nkx2.2null;Neurod1Δendo compared with the Nkx2.2null mice (Figure 2C–2H). As seen in the Nkx2.2null;Neurod1Δpanc and Nkx2.2null;Neurod1null mice, there was no rescue of the insulin-producing beta cell population (Figure 2C–2F; Figure S3). Given this similar phenotype between the Nkx2.2null;Neurod1null, Nkx2.2null;Neurod1Δpanc and Nkx2.2null;Neurod1Δendo we conclude that the genetic interaction between Nkx2.2 and Neurod1 is required in the Neurog3+ cells to permit specification of the PP and epsilon cell populations.
Changes in the beta, PP and epsilon cell populations were identical when Neurod1 was deleted from either the pancreatic or endocrine progenitors in the absence of Nkx2.2. However, in contrast to the Nkx2.2null;Neurod1Δpanc and the Nkx2.2null;Neurod1null, the glucagon-expressing alpha cell population was only minimally restored in the Nkx2.2null;Neurod1Δendo (Figure 3A–3D). Morphometric analysis (Figure 3E) and real time PCR for glucagon expression (Figure 3F) confirmed this observation. We also established that the partial rescue was not due to incomplete deletion of Neurod1 by Neurog3-cre, as Neurod1 was reduced at an early stage of Neurog3 expression; becoming almost undetectable in the mutant pancreata by P0 (Figure 3G; Figure S4). Taken together, these data suggest that the genetic interaction between Nkx2.2 and Neurod1 in Pdx1+ progenitors, prior to Neurog3+ endocrine progenitor formation, is required for complete alpha cell formation.
Data from the Nkx2.2null;Neurod1Δpanc and Nkx2.2null;Neurod1Δendo clearly demonstrate that Neurod1 must be deleted from the Pdx1+ progenitor population and not the Neurog3+ endocrine progenitor population to allow for complete rescue of alpha cell formation. Furthermore, the simultaneous loss of Nkx2.2 and Neurod1 was able to rescue even the earliest glucagon-expressing cell population; the number of glucagon-expressing cells was equivalent between the Nkx2.2null;Neurod1null and wildtype littermate controls at E10.5 (Figure 4A–4D; data not shown), Interestingly, the early glucagon-expressing cells are known to express low levels of Pdx1 (Figure S5; [24]). To determine whether the alpha cell restoration was due to deletion of Neurod1 specifically from this glucagon+ (Pdx1low) population in the absence of Nkx2.2, we deleted Neurod1 in the glucagon-expressing cells using Glu-cre [32] (Figure S2C, S2D). In the Nkx2.2−/−;Neurod1flox/flox;Glu-cre (denoted as Nkx2.2null;Neurod1Δalpha), the complement of all hormone-expressing cells in the pancreas was phenotypically identical to the Nkx2.2null, as determined by immunofluorescent analysis of islet cell markers (Figure 5A–5L; data not shown) and real time PCR for quantitative hormone expression (Figure 5M–5O; Figure S6). These results suggest that restoration of alpha cells requires the deletion of Neurod1 in Pdx1+ progenitors that have not yet committed to the glucagon-expressing lineage. We hypothesize that Nkx2.2 represses Neurod1 in the Pdx1+ cells to give rise to Neurog3+ endocrine progenitor cells that are primed to differentiate into the alpha cell fate.
Since Neurod1 is a downstream target of Neurog3 [33], [34] and the Neurod1 single knockout phenotype does not manifest until the end of gestation [26], it was surprising that manipulation of Neurod1 within the Neurog3+ endocrine progenitors was not sufficient to rescue the alpha cell fate in the Nkx2.2 null background. To begin to reconcile these unexpected results, we re-examined when and where Neurod1 was expressed during pancreatic development. It was previously reported that Neurod1 is expressed at E9.5 in the earliest islet precursors, and is often co-expressed with glucagon [26]. Using the Neurod1 null mouse, which has a LacZ insertion into the Neurod1 locus [35], we confirmed the presence of Pdx1+/Neurod1(beta-gal+) cells and glucagon+/Neurod1(beta-gal+) cells in the earliest pancreatic domain (Figure 6A; Figure S7A); however, not all glucagon+ cells were Neurod1+ (Figure 6A, 6E). Consistent with previous reports [28], this pattern was also evident at E13.5 (Figure 6B, 6E) during the stage of pancreas development marked by a major wave of endocrine cell differentiation referred to as the “secondary transition” [36].
Neurod1 is expressed throughout the epithelial cord region, overlapping extensively with the Neurog3+ precursor cells (Figure S7B, S7C). We used expression of the Neurod1:LacZ allele to identify Neurod1 (beta-gal+) cells that co-expressed Neurog3 at E9.5 (Figure 6C) and at E13.5 (Figure 6D). Interestingly, the overlap of Neurog3 and Neurod1 was not exclusive at either age, and a subset of Neurog3+ cells did not express Neurod1 (Figure 6F). We also detected Neurod1 (beta-gal+) expression in a small population of Sox9low cells (Figure S7D–S7F), indicating that Neurod1 expression can be found in cells that are transitioning into Neurog3 precursor cells [37]. Taken together these expression analyses identified heterogeneous populations of Neurog3+ cells and glucagon+ cells based on their expression of Neurod1, and may suggest that the presence or absence of Neurod1 could influence downstream cell fate decisions.
Our cumulative data suggest that Nkx2.2 may function to repress Neurod1 in a subset of Pdx1+ pancreatic progenitor cells to promote specification of the alpha cell fate. We had previously determined that Nkx2.2 directly activates the Neurod1 promoter in beta cells, which is consistent with the beta cell phenotypes of the single and double knockout mice [12], [26], [28] (Figure 7A). To determine whether Nkx2.2 could also repress Neurod1 expression in other (non-beta) cell contexts, we analyzed the effect of Nkx2.2 on Neurod1 expression in alpha cells in vitro. Utilizing previously described Neurod1 promoter deletion constructs [28] we determined that Nkx2.2 repressed the Neurod1 promoter in alphaTC1 cells, which express Nkx2.2 [28] (Figure 7). Specifically, the repressive activity of Nkx2.2 mapped to the proximal region of the Neurod1 promoter, which is retained in the NDΔ2 promoter construct (Figure 7B). We also determined that, similar to Nkx2.2-dependent activation of the Neurod1 promoter in beta cells, Nkx2.2 repression required the presence of at least one of the three Nkx2.2 binding sites; deletion of either region containing these consensus elements (promoter constructs NDΔ3, NDΔ4) resulted in a loss of Nkx2.2 repression (Figure 7B).
To begin to understand how Nkx2.2 mediates differential cell context-specific regulatory activities through the same set of promoter elements, we assessed the ability of Nkx2.2 to recruit specific cofactors and/or modified histones to the Neurod1 promoter in alpha versus beta cell lines. We previously demonstrated that Nkx2.2 preferentially recruits Grg3 and a large co-repressor complex to the inactive Arx promoter in beta cells, but this complex was not present on the same promoter region in alpha cells, where Arx was actively transcribed [14]. Surprisingly, neither Grg3 nor HDAC1 were recruited to the Neurod1 promoter in either alpha or beta cell lines (data not shown), suggesting that Nkx2.2 mediates Neurod1 regulation through an alternative mechanism. Interestingly however, we determined that histone H3K4me3 preferentially occupied the Neurod1 promoter in beta cells, and this differential binding was dependent upon the phosphorylation state of Nkx2.2 (Figure 7C). Histone H3K27me3 was not significantly present at the Neurod1 promoter in either alpha or beta cell lines (Figure 7D). These results suggest that while Nkx2.2 promotes activation of Neurod1 in beta cells [28], Nkx2.2 appears to prevent the activation of the Neurod1 promoter in alpha cells. This finding is consistent with the idea that Nkx2.2 is required to prevent expression of Neurod1 in a subset of Pdx1+ progenitor cells and then maintain this repression in “alpha-cell competent” Neurog3-expressing cells, and subsequently mature alpha cells.
Single deletion mutants have identified the importance of a number of transcription factors for the process of endocrine cell differentiation (reviewed in [38]). Interestingly, very few factors when deleted affect only one islet cell type. Therefore we can deduce that each regulatory protein has multiple roles during development and it is likely that different combinations of these factors must be simultaneously present or absent within the endocrine progenitor cells to permit the specification of alpha, beta, delta, epsilon or PP cells. The generation of compound deletion mutants would assist in deciphering this combinatorial transcription factor code. One such example is the regulatory interaction between Nkx2.2 and the alpha cell transcription factor Arx; simultaneous deletion revealed that these factors differentially cooperate to affect the specification of several islet cell lineages [23], [24]. In this current study, we explore the relative roles of Nkx2.2 and the beta cell transcription factor Neurod1. The single deletion mutants for Nkx2.2 or Neurod1 display alterations in several islet cell types [12], [26]; however, these mutants are noted for their severe beta cell phenotypes. In particular, Nkx2.2 and Neurod1 are necessary for beta cell specification and maintenance, respectively [12], [26]. Interestingly, simultaneous deletion of Nkx2.2 and Neurod1 did not affect the respective beta cell phenotypes of the single mutants, but rather identified complex genetic interactions between these factors for the specification of alpha, PP and epsilon cells [25]. In this set of experiments, we have determined the cellular locations of the genetic interactions between Nkx2.2 and Neurod1, and have uncovered a possible mechanism for how these transcription factors contribute to the process of alpha cell specification. Given the increasing number of studies identifying transdifferentiation between alpha cells and beta cells [10], [14], [39], refining our understanding of alpha cell development may provide insight into the unique relationship between alpha and beta cells, and ultimately aid in understanding how beta cells develop in both the normal and diseased state.
Knowing that all endocrine cell types are derived from Neurog3-expressing cells [5], [6], we hypothesized that the genetic interaction between Nkx2.2 and Neurod1 would be required in the Neurog3+ endocrine progenitors to specify islet cell fates. In support of this hypothesis, deletion of Neurod1 from the Neurog3+ endocrine progenitor cells in an Nkx2.2 null background (Nkx2.2null;Neurod1Δendo) was sufficient to rescue the relative ratios of the ghrelin-expressing epsilon cells and pancreatic polypeptide-expressing PP cells when compared to the Nkx2.2 null phenotype. This demonstrates that the genetic interaction between Nkx2.2 and Neurod1 is required within the Neurog3+ endocrine progenitor population to permit appropriate specification of the PP and epsilon cell populations. In contrast, although alpha cells were completely rescued in the Nkx2.2null;Neurod1Δpanc, we observed only a minimal restoration of glucagon+ cells in the Nkx2.2null;Neurod1Δendo, suggesting that alpha cell recovery requires the genetic interaction between Nkx2.2 and Neurod1 to occur within the Pdx1+ pancreatic progenitors, prior to Neurog3+ endocrine progenitor cell formation. This finding would support the concept proposed by Degraz and Herrera [7] that the Neurog3+ endocrine progenitors represent a heterogeneous population of unipotential cells that are already committed to become a single hormone-producing cell fate.
If all Neurog3+ progenitors are indeed unipotent, then how do we explain rescue of the PP and ghrelin cell ratios that resulted from manipulating gene expression after the Neurog3+ cells are formed? It is possible that there are both unipotential and multipotential endocrine progenitor populations. Alternatively the “pro-PP” or “pro-ghrelin” Neurog3+ populations may retain more plasticity throughout development. The latter explanation is consistent with the findings of Johansson et al., [9], which demonstrated that as development proceeds the progenitor cells are less competent to produce alpha cells and instead favor the generation of other endocrine cell types. This would suggest that although the alpha cell fate decision can be made at multiple points during development, the ability to generate alpha cells is most robust in the earliest pancreatic progenitors and becomes restricted over time. Alternatively, it is possible that later born progenitors retain a certain degree of plasticity that accounts for their ability to respond to lineage manipulations after Neurog3+ cell specification has occurred.
The inability to rescue alpha cells by simultaneously removing Nkx2.2 and Neurod1 from the Neurog3+ precursor population, suggests that the genetic interaction between Nkx2.2 and Neurod1 is required in the Pdx1+ progenitor population, prior to acquisition of Neurog3 expression. However, it remains possible that there is a spectrum of Neurog3-cre activity within a Neurog3+ precursor cell, with Cre-based inactivation reaching its peak in the middle or late in the lifespan an individual cell. If this were the case, and the genetic interaction between Nkx2.2 and Neurod1 is required only early in the lifespan of a Neurog3+ precursor to rescue alpha cells, then Neurog3-cre activity may occur too late within this population to affect its differentiation potential. Although we are unable to resolve the kinetics of Cre activity in the lifespan of a single cell, we can demonstrate co-expression of Neurog3, Cre and R26R reporter activity, suggesting that although Neurog3 protein expression is transient, Cre is present and active in most of the Neurog3+ population during the time window when Neurog3 is expressed (Figure S2B). Furthermore, published lineage studies using this Neurog3-cre allele demonstrated that all endocrine cells of the islet, including the glucagon-expressing alpha cells, are labeled by a Cre-dependent R26R:LacZ reporter [31]. This would suggest that even if alpha cells can only be differentiated from “young” Neurog3+ precursors, there is sufficient Cre activity at this earliest stage during the lifespan of a Neurog3+ cell to genetically label the alpha cell population.
Our failure to recover alpha cells by deleting Neurod1 in a glucagon-expressing population may also be due to the inefficiency of the Glu-cre allele, especially in Nkx2.2null embryos that have a severe reduction in alpha cell numbers. However, we detected similar levels of Glu-cre activity in wildtype and Nkx2.2null pancreata, which should have been sufficient to permit any possible alpha cell rescue (Figure S2C–S2D; see Materials and Methods). Although caveats exist with the use of Cre/lox technologies, these are currently the best tools available to assess spatial and temporal protein function.
Interestingly, we do observe some rescue of alpha cells in the Nkx2.2null; Neurod1Δendo embryos. This could be due to deletion of Neurod1 in a subset of Neurog3+ progenitors that have not yet become restricted in their ability to differentiate into alpha cells. Alternatively, the glucagon-expressing cells recovered in the Nkx2.2null;Neurod1Δendo may represent alpha cells that form independent of Neurog3 function; such an alpha cell population has been previously documented [40], [41]. On the other hand, the recovered alpha cells may actually represent a distinct subpopulation of glucagon-expressing cells that express Neurod1, which would be consistent with our identification of a subpopulation of glucagon+/Neurod1+ cells. While these explanations are not mutually exclusive, the identification of unique alpha cell markers and the generation of genetic tools utilizing these markers, would be necessary to clarify the existence of subpopulations of alpha cells, as well as the factors involved in the generation of these distinct populations.
Our findings also suggest that Nkx2.2 must regulate Neurod1 differentially in the Pdx1+ progenitor population in the early pancreatic epithelium in order to initiate the specification of different populations of Neurog3-expressing cells. In particular, the prevention of Neurod1 activation by Nkx2.2 would result in alpha cell formation, while the activation of Neurod1 by Nkx2.2 results in beta cell formation (Figure 8). This is compatible with our discovery that not all Neurog3+ cells express Neurod1, and further supports the idea that the Neurog3+/Nkx2.2+/Neurod1+ cells most likely become beta cells, whereas Neurog3+/Nkx2.2+/Neurod1− cells would become alpha cells. Ideally, we would test this hypothesis by quantifying the increase in the number of Pdx1+/Neurod1+ pancreas progenitors and/or Neurog3+/Neurod1+ endocrine progenitors expected to be observed in the Nkx2.2null pancreas; however, this analysis is confounded by the simultaneous loss of the Neurod1+ pro-beta cell progenitor populations in the Nkx2.2null pancreas. Instead, we used an in vitro approach to determine whether it was possible for Nkx2.2 to differentially regulate the Neurod1 promoter in different cellular contexts. We had previously demonstrated that Neurod1 is activated by the cooperative binding of Nkx2.2 and Neurog3 specifically in beta cells [28]. Given the lack of availability of an appropriate pancreatic progenitor cell line, we reasoned that a genetic interaction between Nkx2.2 and Neurod1 that was initiated in a “pro-alpha cell” progenitor would be maintained in the mature alpha cell. We utilized alphaTC1 cells, which express Nkx2.2 [28], to demonstrate that Nkx2.2 prevents activation of Neurod1 in alpha cells. Highlighting the complexity of gene regulation, the cell type specific regulation of Neurod1 by Nkx2.2 appears to function through a mechanism that is different from Nkx2.2 regulation of the Arx gene [14]. This may reflect the mechanism by which Nkx2.2 functions as an activator and a repressor in the same cell type and/or the presence or absence of cell-specific co-regulatory proteins. As we gain the molecular tools to study transcriptional and epigenetic mechanisms in purified primary pancreatic cell populations, we hope to elucidate the complex regulatory interactions that are required to form and maintain appropriate islet-cell specific gene expression.
While the process of endocrine specification likely requires the concerted action of many factors, our data suggest a mechanism that involves the differential regulation of Neurod1 by Nkx2.2 in the Pdx1+ pancreatic progenitor cells to direct the subsequent endocrine progenitors to become specific islet cell types. The generation of tools to identify, separate and analyze different subpopulations of Neurog3+ progenitor cells would conclusively determine whether each hormone+ endocrine cell type is derived from a specific unipotent subpopulation of endocrine progenitor cells, each bearing a unique gene profile.
Using the pancreas as a model system, our study has provided a prime example of how lineage decisions are established in the developing epithelium. The cooperative action of multiple transcription factors within the early progenitor cells can dictate the fate of subsequent cell lineages. Altering the regulation or complement of this set of factors within the progenitor populations can ultimately skew cell lineage specification. These data have important implications for the current efforts to generate pancreatic cells in vitro for therapeutic use in diabetic patients. Understanding the cooperative transcription factor code will make it possible to initiate the appropriate program in the Pdx1+ pancreatic progenitor cells necessary to correctly prime the Neurog3+ endocrine progenitor cells and generate pools of functional, single hormone-expressing islet cell types in vitro.
All experiments involving mice were approved by the Columbia University Institutional Animal Care and Use Committee and performed in accordance with the National Institutes of Health guidelines for the care and use of animals. All mouse strains were previously generated, and were bred and maintained on an outbred Black Swiss background (NTac:NIHBS, Taconic). Cell-specific Neurod1 null mice were generated by intercrossing Neurod1tm1Kan (Neurod1flox/flox; [42]) and either Tg(Ipf1-cre)1Tuv (Pdx1-cre; [29]), Tg(Neurog3-cre)C1Able (Neurog3-cre; [31]), or Glu-cre ([32]) mice. Neurod1flox/flox;Pdx1-cre and Neurod1flox/flox;Neurog3-cre mice died postnatal, similar to the Neurod1 null (data not shown; [30]). Certain experiments required the use of either Gt(ROSA)26Sortm9(CAG-tdTomato)Hze (R26R:Tomato; [43]) or Gt(ROSA)26Sortm1Sor (R26R:LacZ; [44]) reporter alleles. The Pdx1-cre will delete Neurod1 in all pancreatic progenitor cells; however, the Pdx1 expression domain also includes a portion of the stomach and the duodenum [45], [46]. We and others have previously reported the early and relatively non-mosaic activity of the Pdx1-cre allele ([29], [47]; Figure S2A). Previous characterization of the Neurog3-cre allele demonstrated almost complete co-expression of Neurog3 and Cre and sufficient Cre activity to lineage label all endocrine cells within an islet [31]. Consistent with this published analysis, quantification of cells co-expressing Neurog3 and the LacZ reporter in a Neurog3-Cre;R26R:LacZ E15.5 embryo indicated 74.82% Cre efficiency (268 Neurog3+ beta− gal+/349 total Neurog3+ cells; calculations were performed as described below (Figure S2B). Similar assessment of the Glu-cre mice demonstrated that the Glu-cre allele is active in approximately 30–35% of alpha cells; notably this degree of activity is unchanged in the Nkx2.2null background, despite the overall reduction in alpha cell numbers (Figure S2C, S2D).
The heterozygous mice (Neurod1flox/+;Pdx1-cre) were crossed to Nkx2-2tm1Jlr knock-in mice [12] to generate compound heterozygotes. Embryos were collected from timed matings between Nkx2.2+/−;Neurod1flox/+;Pdx1-cre and Nkx2.2+/−;Neurod1flox/flox or Nkx2.2+/−;Neurod1flox/+;Neurog3-cre and Nkx2.2+/−;Neurod1flox/flox or Nkx2.2+/−;Neurod1flox/+;Glu-cre and Nkx2.2+/−;Neurod1flox/flox mice. Noon on the day of appearance of a vaginal plug was considered embryonic day (E) 0.5. The experimental genotypes of wildtype, Nkx2.2−/− (Nkx2.2null), Neurod1flox/flox;Pdx1-cre (Neurod1Δpanc), Nkx2.2−/−;Neurod1flox/flox;Pdx1-cre (Nkx2.2null;Neurod1Δpanc), Neurod1flox/flox;Neurog3-cre (Neurod1Δendo), Nkx2.2−/−;Neurod1flox/flox;Neurog3-cre (Nkx2.2null;Neurod1Δendo), Neurod1flox/flox;Glu-cre (Neurod1Δalpha), and Nkx2.2−/−;Neurod1flox/flox;Glu-cre (Nkx2.2null;Neurod1Δalpha) were studied. Litters were assessed at postnatal day (P) 0. For expression studies, the Neurod1tm1Jle LacZ knock-in (Neurod1LacZ/+ or Neurod1null) [35] was used (also in combination with the Nkx2.2null thereby producing Neurod1null;Nkx2.2null double knockout embryos; DKO), and embryos were assessed at E9.5, E10.5, E13.5 and P0. All embryo dissections were carried out in cold PBS, using a dissecting microscope (Leica MZ8). A portion of each embryonic tail or yolk sac was detached from the embryo, digested with proteinase K, and DNA extracted for genotyping purposes. Genotyping was carried out with standard conditions and primers as previously described [12], [29], [31], [32], [35], [42].
Pancreas was dissected from each embryo and stored in RNAlater (Ambion) until RNA was extracted using the NucleoSpin RNAII Kit (Clontech). Subsequently, cDNA was made with equal amounts of RNA for each sample (Superscript III Kit, Invitrogen, CA). Real time PCR was performed using TaqMan gene expression assays (Applied Biosystems) for glucagon (Mm00801712_m1), ghrelin (Mm00445450_m1), somatostatin (Mm00436671_m1), insulin1 (Mm01950294_s1), insulin2 (Mm00731595_gH), pancreatic polypeptide (Mm00435889_m1) and Neurod1 (Mm01280117_m1). CyclophilinB was used as a control housekeeping gene, and was assayed using a probe and primer set previously described [25]. A standard two-step real time PCR program was used for all genes assessed, with an annealing temperature of 61°C and 40 cycles of amplification (CFX96 RealTime System C1000 Thermal Cycler, Biorad). All gene expression values were normalized to the internal control gene, cyclophilinB, and relative quantification was performed using a standard curve from embryonic age-matched cDNA. Statistical analyses were conducted with Prism Software (GraphPad Software, La Jolla, CA) using both the Mann-Whitney test and the Student t-test. Equivalent results were obtained; t-test results were reported in all Figures.
Immunofluorescence was performed according to standard protocols, on E9.5, E10.5, E13.5, E15.5 and P0 whole embryos that were embedded in OCT, after fixation with 4% PFA and cryopreservation in 30% sucrose. Transverse frozen sections (8 µm) were cut and mounted on glass slides. Sections were stained with rabbit α-ghrelin (1∶800; Phoenix Pharmaceuticals, CA), goat α-ghrelin (1∶800; Santa Cruz), guinea pig α-glucagon (1∶1000; Linco/Millipore, MA), guinea pig α-insulin (1∶1000; Millipore), rabbit α-insulin (1∶1000; Cell Signaling Technology), rabbit α-somatostatin (1∶200; Phoenix Pharmaceuticals), rabbit α-pancreatic polypeptide (1∶200; Zymed), rabbit α-amylase (1∶1000; Sigma), rabbit α-Pdx1 (1∶1000; Millipore), guinea pig α-Pdx1 (1∶500; BCBC), rabbit α-Neurog3 (1∶500; BCBC), goat α-Neurog3 (1∶500; BCBC), goat α-FoxA (1∶1000; Santa Cruz), rabbit α-sox9 (1∶500; Chemicon), and chicken α-beta-galactosidase (1∶250; Abcam). Donkey α-guinea pig-Cy2, -Cy3 or -Cy5, α-rabbit-Cy2 or -Cy3, α-chicken-Cy3, and α-goat Cy2 or -Cy5 secondary antibodies were used (1∶400, Jackson ImmunoResearch). DAPI (1∶1000; Invitrogen) was applied for 30 minutes following secondary antibody incubation. Images were acquired on a Leica DM5500 or Leica 510 confocal microscope. Morphometric analysis was performed by immunostaining every 10th section throughout each embryo (N = 3 or 4 for each genotype). For quantification of individual hormone-expressing cells at P0, cell number was assessed versus total pancreas as defined by amylase area. For quantification of hormone-expressing cells at E10.5, cell number was assessed versus total pancreas as defined by Pdx1 area. Pancreas area was calculated using ImagePro software.
RNA in situ hybridization was performed on 8 µm sections mounted on glass slides as previously described [25] using an antisense riboprobe transcribed from linearized plasmid. The riboprobe for Neurod1 was generated from the plasmid pCS2:MTmNeuroD1 (J. Lee). RNA in situ hybridization was performed on pancreas tissue sections from Neurod1Δendo and wildtype littermate controls at E10.5 and Neurog3-cre;R26RLacZ at E15.5.
The Neurod1-2.2 kb minimal promoter was fused to the firefly luciferase open reading frame in the pGL3 Basic vector (Promega). The alphaTC1 cells were grown in 12-well plates. The design of all Neurod1 promoter deletion constructs and the transfection conditions were previously described [28]. Firefly luciferase readings were normalized to Renilla luciferase values. A Student t-test was performed to determine significance.
Point mutations were made to 3xmyc-tagged Nkx2.2 cDNA using the QuickChange II Site Directed Mutagenesis kit (Agilent Technologies) with the following primers S-11-A: (FWD) CAACACAAAGACGGGGTTTGCTGTCAAGGACATCTTGGAC, (REV) GTCCAAGATGTCCTTGACAGCAAACCCCGTCTTTGTGTTG; S-11-D: (FWD) CAACACAAAGACGGGGTTTGATGTCAAGGACATCTTGGAC, (REV) GTCCAAGATGTCCTTGACATCAAACCCCGTTTTGTGTTG. Wild type or mutated Nkx2.2 cDNA encoding a triple myc epitope tag (250 ng) was transfected into betaTC6 or alphaTC1 cells using X-treme gene HP (Roche) according to manufacturer's protocol. Chromatin was prepared using the ChIP-IT express kit (Active Motif). Immunoprecipitation protocol was modified from Tuteja et al. [48]. In brief, immunoprecipitation was performed using the isolated chromatin diluted in ChIP dilution buffer with 5 micrograms of either mouse anti-H3K27me3 (Abcam) or mouse anti-H3K4me3 (Abcam) antibodies while rotating overnight at 4°C. The following day antibody/chromatin complexes were pulled down using ChIP grade protein G magnetic beads (Cell Signaling). After washing, antibody/chromatin complexes were eluted from the beads and allowed to rotate at room temperature for 15 minutes. NaCl (5 micromolar) was added to the eluate and incubated at 65°C overnight. The following day Tris-HCl (1 M, pH 7.5), EDTA (0.5 M) and proteinase K (10 mg/mL) were added and allowed to incubate at 37°C for 1 hour. Samples were then purified using the QIAquick PCR purification kit (Qiagen). Quantitative analysis of ChIP products was performed using SYBR Green fluorescence with primers for Gapdh (FWD – CTCCACGACATACTCAGCACC; REV – TCAACGGCACAGTCAAGGC) or Neurod1 (FWD – AAAGGGTTAATCTCTCCTGCGGGT; REV - CATGCGCCATATGGTCTTCCCGGT).
|
10.1371/journal.pgen.1004363 | Predicting the Minimal Translation Apparatus: Lessons from the Reductive Evolution of Mollicutes | Mollicutes is a class of parasitic bacteria that have evolved from a common Firmicutes ancestor mostly by massive genome reduction. With genomes under 1 Mbp in size, most Mollicutes species retain the capacity to replicate and grow autonomously. The major goal of this work was to identify the minimal set of proteins that can sustain ribosome biogenesis and translation of the genetic code in these bacteria. Using the experimentally validated genes from the model bacteria Escherichia coli and Bacillus subtilis as input, genes encoding proteins of the core translation machinery were predicted in 39 distinct Mollicutes species, 33 of which are culturable. The set of 260 input genes encodes proteins involved in ribosome biogenesis, tRNA maturation and aminoacylation, as well as proteins cofactors required for mRNA translation and RNA decay. A core set of 104 of these proteins is found in all species analyzed. Genes encoding proteins involved in post-translational modifications of ribosomal proteins and translation cofactors, post-transcriptional modifications of t+rRNA, in ribosome assembly and RNA degradation are the most frequently lost. As expected, genes coding for aminoacyl-tRNA synthetases, ribosomal proteins and initiation, elongation and termination factors are the most persistent (i.e. conserved in a majority of genomes). Enzymes introducing nucleotides modifications in the anticodon loop of tRNA, in helix 44 of 16S rRNA and in helices 69 and 80 of 23S rRNA, all essential for decoding and facilitating peptidyl transfer, are maintained in all species. Reconstruction of genome evolution in Mollicutes revealed that, beside many gene losses, occasional gains by horizontal gene transfer also occurred. This analysis not only showed that slightly different solutions for preserving a functional, albeit minimal, protein synthetizing machinery have emerged in these successive rounds of reductive evolution but also has broad implications in guiding the reconstruction of a minimal cell by synthetic biology approaches.
| In all cells, proteins are synthesized from the message encoded by mRNA using complex machineries involving many proteins and RNAs. In this process, named translation, the ribosome plays a central role. The elements involved in both ribosome biogenesis and its function are extremely conserved in all organisms from the simplest bacteria to mammalian cells. Most of the 260 known proteins involved in translation have been identified and studied in the bacteria Escherichia coli and Bacillus subtilis, two common cellular models in biology. However, comparative genomics has shown that the translation protein set can be much smaller. This is true for bacteria belonging to the class Mollicutes that are characterized by reduced genomes and hence considered as models for minimal cells. Using homology inference approach and expert analyses, we identified the translation apparatus proteins for 39 of these organisms. Although striking variations were found from one group of species to another, some Mollicutes species require half as many proteins as E. coli or B. subtilis. This analysis allowed us to determine a set of proteins necessary for translation in Mollicutes and define the translation apparatus that would be required in a cellular chassis mimicking a minimal bacterial cell.
| Mollicutes constitute a monophyletic class that share a common ancestor with Gram-positive bacteria of low G+C content or Firmicutes but have adopted a parasitic life style (Figure S1) [1]. During their coevolution with their eukaryotic hosts, mollicutes progressively lost the genes coding for cell-wall synthesis enzymes and for enzymes involved in the synthesis of small metabolites, such as amino acids, nucleotides and lipids that were available in the host. As a result, mollicute genomes are much smaller (580–1,840 Kbp; eg: about 482–2,050 CoDing Sequences or CDSs, Table S1) than those of model bacteria such as Escherichia coli or Bacillus subtilis (4,639–4,215 Kbp; eg: 4,320–4,176 CDSs respectively). These bacteria have nevertheless retained the full capacity to synthesize DNA, RNA and all the proteins required to sustain a parasitic life-style. In addition most of them are still able to grow in axenic conditions in rich media usually containing 20% serum (see [2] for review); only the hemoplasmas and the Candidatus phytoplasma species have yet to be cultured in vitro. Mollicutes are therefore considered as the smallest and simplest known bacteria capable of autonomous multiplication [3], [4]. ‘Simple’ does not mean ‘simplistic’. One should not underestimate the elaborate solutions that mollicutes have used to solve problems related to their peculiar macromolecular organization and cellular compactness (discussed in [3], [5], [6] and references therein). From an evolutionary point of view, mollicutes should be considered as some of the most evolved prokaryotes that still have retained ability to perform the complex reactions that encompass DNA, RNA and protein synthesis, with possibly new tricks and inventions to make the most of their limited genetic capacities [7], [8]. For these reasons, specific Mollicutes strains have been used as a test bench to improve our understanding of the basic principles of a cell and for reconstructing a microbe that would function with a synthetic minimal genome (see [3], [4], [9], [10], [11] for examples).
Identification of essential proteins is a long-standing problem that is directly linked to the concept of a minimal cell [12]. The approaches used in Mollicutes to identify the set of essential genes have been: i) comparative genomic analyses to create an overview of the protein content in model mycoplasmas (notably Mycoplasma genitalium and Mycoplasma pneumoniae) [5], [13], [14], [15], ii) identification of genes that cannot be individually inactivated [16], [17], [18], [19], iii) reconstruction of synthetic genomes and transplantation into a recipient cell [10]. Depending on the Mollicutes species considered and the method of analysis, the number of essential genes varies from 256 to 422. For M. genitalium, 256 were identified by in silico comparative genomics analysis [15] but over 382 were found by saturation transposon mutagenesis experiments [16], [19]. For Mycoplasma pulmonis and Mycoplasma arthritidis, saturation transposon mutagenesis identified 422 and 417 essential genes respectively [17], [20].
Messenger-RNA-dependent protein synthesis is one of the most complex cellular processes both in its biogenesis and its function. For a cell with a reduced genome such as M. genitalium, more than 25% of the genome encoding capacity is mobilized to build this complex machinery [2]. The bacterial ribosome is a giant multicomponent complex of several millions of daltons, composed of 3 RNA species (5S, 16S and 23S rRNA) and many structural proteins (60–70). Together with other RNAs (tRNAs, tmRNA and RNA-P) and a large repertoire of enzymes and protein factors, this protein synthesis machinery allows translation of mRNAs into polypeptides according to precise rules. Comparative analysis of bacterial genomes reveals that the majority of genes coding for the ribosomal proteins, aminoacyl-tRNA synthetases, translation factors and several ribosome biogenesis/maturation enzymes are universal [7], [21] and essential [22], [23], [24]. Genes coding for enzymes involved in rRNA and protein processing, RNA or protein modification, and ribosome maturation RNases appear less important, as deleting these does not lead to severe growth defects, and are the most easily lost genes during genomic erosion in Mollicutes species (see below).
As the number of sequenced Mollicutes genomes has significantly increased, most of the phylogenetic sub-groups of this class of bacteria are now covered allowing for the analysis of the erosion of translation from an evolutionary perspective. This analysis defined the minimal set of proteins needed to sustain protein synthesis in various mollicutes. A major goal of this work was to identify the minimal set of proteins that can sustain ribosome biogenesis and translation of the genetic code in Mollicutes that are model organisms of choice for synthetic biology. Also, by careful analysis of the evolutionary pattern of gene losses and a few cases of gene gain in different individual Mollicutes species, light was shed on the progressive adaptation of an ancestral and complex cellular proteome towards a simpler, yet functional alternative one.
The major goal of this work is to identify the minimal set of proteins that can sustain ribosome biogenesis and translation of the genetic code in self replicating bacteria with reduced genomes (MPSM for Minimal Protein Synthesis Machinery). Comparative genomics of 39 Mollicutes species allowed the identification of 104 genes encoding ubiquitous translation proteins designed as the core set herein. The acronyms of these proteins are listed according to their main functions in Figure 4. The majority of these core proteins are present in both B. subtilis and E. coli, the exceptions are proteins that are found only in Gram-positive bacteria (indicated in red; Figure 4A). In M. genitalium and M. pneumoniae almost all (except 4) of these 104 proteins were experimentally demonstrated to be essential (Figure S2), attesting their primordial importance for ribosome biogenesis and function in the context of Mycoplasma metabolism.
This set of 104 core proteins might not be sufficient for ribosome biogenesis and translation to work. Indeed, extant culturable Mollicutes maintain a set of translation proteins above an apparent lower limit of 138 (Figure 2). An additional set of essential proteins, not necessarily the same in each species, are obviously required. Among them are the 17 persistent gene products discussed above that are absent only in one (usually M. suis) or several non-culturable Mollicutes (indicated with red asterisks in Figure 4B). Eight additional proteins that are notably persistent or can only be replaced by an alternate mechanism have been added in the MPSM. These are: i) r-protein L9 (RplI) absent only in M. penetrans and three non cultivable species, L9 interacts with tRNA in the P site and limits mRNA slippage during translation; ii) r-protein S21 (PpsU) that is essential in the absence of r-protein S1 (RpsA), particularly for translating leaderless mRNAs; iii) 2′-O-RNA methyltransferase RlmB2 or YqxC predicted to methylate a conserved G residue in the A-loop (helix 92) of the peptidyl-transferase center of 23S rRNA (counted for one protein); iv) one of the three paralogous double-stranded endonucleases (RNases HI, HII, HIII) as all mollicutes harbour at least one of these enzymes that possibly could have broad specificity; v) the essential lysidine-tRNA transferase (TilS) that can be lost only if compensatory mutations occur in the tRNA recognition domain of IleRS and the anticodon of tRNAIle; finally vi) the three subunits of the Gln-tRNA amidotransferase complex (GatA-GatB-GatC) of the Gln-tRNA amidotransferase complex essential for the formation Glutamine-tRNAGln in Mollicutes lacking the Glutamine-tRNA synthetase GlnRS (counted for 3 proteins).
Proteins that were easily lost during Mollicutes evolution were not included as essential elements of an MPSM (Figure S3A). However, some of these proteins may fine-tune ribosome biogenesis, improve efficiency of translation and/or display other side functions, such as coupling of translation with transcription and/or regulating protein expression. Finally, proteins that are absent in all Mollicutes were definitively discarded as elements of the MPSM, the majority of these are also absent in Gram-positive bacteria (Figure S3B).
Therefore, in absence of stress conditions that require specific proteins not discussed here, we propose that these 17+8 = 25 proteins, combined with the core of 104 proteins, comprise a theoretical MPSM of 129 proteins. This MPSM corresponds to a set of well characterized homologous proteins in our model bacterial systems and they are encoded by the most persistent genes in the Mollicutes analyzed. However, because some genes are still of unknown function in E. coli, B. subtilis and Mollicutes, we cannot exclude the possibility that a yet unidentified protein involved in the biosynthesis or function of the ribosome might have been missed.
Our evaluation of 129 minimal translation associated genes accounts for a large fraction of the total genes identified in mollicutes with reduced genomes (26% in the case of M. genitalium and 18% for M. pneumoniae). The protein synthesis factory is clearly the dominant and most energy consuming process in small cells such as Mollicutes [14].
The progressive reduction of the size of precursor RNAs (mainly mRNAs and tRNAs) by reducing their 3′ and/or 5′-tails is probably also part of the genomic size economization strategy. In Mollicutes, 18% of mRNA in average are leaderless mRNAs ([123], thus lacking the classical/canonical Shine-Dalgano (SD) sequence required for specific translation initiation on 30S subunit. Similarly precursor tRNAs have shorter 5′-leader sequence and no 3′-tail (see above). However, because of the constraint of maintaining canonical bacterial type of ribonucleoprotein 30S and 50S particles, the length of 16S and 23S rRNAs in Mollicutes is almost identical to those of other bacteria [124].
The best-studied extant Mollicutes with reduced genomes and capable of independent growth are the two phylogenetically related M. genitalium and M. pneumoniae. With a total of about 482 CDS, including 144 CDS for the translation machinery, for a 0.580 Mbp genome, M. genitalium is generally considered as the best representative of a minimal free-living cell. A schematic view of the translation machinery in M. genitalium is depicted in Figure 5, together with the list of all the elements required for ribosome biogenesis and mRNA translation. The 128 proteins classified above as belonging to the MPSM are in bold-black acronyms (only the putative r-RNA modification enzyme RlmB2/YqxC of the selected 25 additional proteins is missing), while the additional 16 proteins present in M. genitalium are in blue italic acronyms (see also Figure S4). These latter proteins include two DEAD- box helicases, one protein kinase (PrkC) and its associated protein phosphatase (PrpC), one r-RNA protein modification (RimK) and two chaperones (GroEL+GroES), all classified as proteins of ribosome assembly and protein maturation. In addition are found three ribonucleases of the RNA processing (RNase M5, RNase Y and a second nano-RNase), three tRNA modification enzymes (TruA, ThiI and TrmK) and three translation factors (DEF, FMT, SpoT/RelA). These proteins, especially GroEL/GroES, RNase MV and RimK are lacking in many other Mollicutes (Figure 1B, Table S3), RimK is even absent in B. subtilis and arose in both M. genitalium and M. pneumoniae probably by lateral gene transfer (see above). In M. genitalium, these proteins may have specific functions such as fine-tuning of RNA processing and ribosome assembly, mRNA translation and its regulation in response to specific physiological demands of the cell. Despite these differences, the translation apparatus in M. genitalium fits well with the MPSM concept developed above and closely resembles the classical scheme of translation in bacteria [125].
The most remarkable features of protein synthesis in M. genitalium and other Mollicutes with minimal genomes are: 1) almost all canonical r-proteins are present (however, as shown in the case of M. pneumoniae [126] not all r-proteins may be present in every ribosome, a certain degree of plasticity in r-protein composition may exist according to specific type of mRNA to be translated); 2) the GTP/ATPases involved in 30S/50S/70S assembly are identical in sequence and number to those found in other bacteria with larger genomes, attesting that the assembly process follows a path extremely conserved in bacteria; the frequent lack of DEAD-box helicases probably results from the A/T-rich RNA sequences; 3) the DnaK-dependent protein folding/quality control system is ubiquitous. However in only a few Mollicutes, including M. genitalium and M. pneumoniae, GroEL/GroES are present and therefore should not be considered as essential; 4) the multiplicity of genes coding for nano-RNases allowing to scavenge for mononucleotide building blocks is of clear advantage for Mollicutes that are devoid of nucleotide biosynthetic pathway; 5) among post-translational protein modification enzymes, only the methyltransferase PrmC (HemK) that methylates termination factor RF-1 is conserved in Mollicutes; 6) a repertoire of 19 aaRSs plus the GatA/GatB/GatC amidotransferase complex allowing to generate Gln-tRNAGln and a minimal set of 28 isoacceptor tRNAs are used to decode all 62 sense codons into 20 canonical aminoacids; 7) an extra tRNATrp harboring an anticodon U*CA reads UGA as Trp [55], the absence of termination factor RF-2 being consistent with this scheme; 8) the methionine residue attached to initiator tRNAMet is formylated in M. genitalium but in most mollicutes the formylation/deformylation enzymatic system (FMT/MAP) is absent and therefore not essential; 9) the majority of post-transcriptional enzymatic modifications in tRNA and rRNA are restricted to a few nucleotides located mostly in the anticodon loop of tRNA, the ribosomal decoding sites (h18, h44 and h45) of 30S subunit and the peptidyl transferase site (H90, H69) of 50S subunits; 10) the majority of the essential bacterial factors are needed, except the stress rescue and silencing factors TypA, AraFA and RsfA; 11) the SpoT/RelA alarmone system is present in M. genitalium and most species of the Pneumoniae sub-group but absent in all species of the Hominis sub-group; 12) tmRNA and its associate protein SmpB of the trans-translation system and the ribozyme RNaseP with only one associated protein RnP are preserved; 13) because of the use of numerous leaderless mRNAs in Mollicutes, an alternative mechanism of translation initiation exists beside the canonical Shine-Dalgano (SD)-depending mRNA initiation, translation initiation of SD-containing mRNA occurs on 30S subunit and is usually mediated by r-protein S1, while S1 but not S21 become dispensable for translation of leaderless mRNAs on intact 70S ribosome [39]; finally, 14) because of their small sizes, a Mollicutes species like M. pneumoniae contains only 140–200 ribosomes per cell volume of 0.067 µm3 [11], while an E. coli cell of about 1 µm3 usually contains several thousands of ribosomes [127].
This study shows that comparative genomics analyses can help define the minimal set of genes required for translation in Mollicutes. Translation genes that have not been lost in any of the species analyzed belong to a translation core that is most certainly needed to sustain protein synthesis. However, loss of a specific protein or enzyme in a given Mollicutes species does not necessarily translate in loss of the corresponding cellular function, as some cellular enzymes or proteins may display overlapping specificities or fulfill closely related, analogous functions. Occasional gene gains are also indicative of the need for compensation for the gene losses or acquiring new functionalities to maintain a reduced, but coherent functional protein synthesis machinery. The corollary of these premices is that different solutions to minimize translation machinery can evolve in different Mollicutes and it is illusory to try to define a universal minimal set of translation proteins that would be common to very distantly related bacteria (discussed in [28]).
The class of Mollicutes is particularly suited for defining a minimal translation apparatus. Not only do they include organisms that have eliminated many primordial metabolism genes (including translation genes), while retaining the capability to replicate and translating mRNAs in an axenic medium, but they also appear as some of the most evolved prokaryotes able to sustain complex metabolism with a minimum elements of its cellular chassis (discussed in: references [3], [4], [9], [10], [11]). Recent studies from independent laboratories have shown that two Mollicutes species (Mesoplasma florum and Mycoplasma gallisepticum) exhibit the highest known rate of base-substitutional mutation for any unicellular organism showing these are fast-evolving bacteria [69], [71]. Although Mollicutes species share a small genome size, our study indicates that there remains room for diversity even in a highly conserved apparatus such as translation. On one side of the spectrum, M. suis probably stands out as the most minimal organism with only 116 proteins dedicated to translation. At this stage, it is not understood how this uncultured organism that lives associated to red blood cells of its mammalian host is able to synthetize proteins with a machinery that appears so deficient. It is tempting to hypothesize that translation in M. suis requires factors from its host, but owing to the lack of general knowledge on hemoplasma biology, it is too speculative to further elaborate. On the other side of the spectrum, A. laidlawii has a much larger repertoire of proteins implicated in translation (183) than most other Mollicutes species, but still lower proteins than in our model bacteria E. coli (228) and B. subtilis (210). In fact, this species with other Acholeplasmatales also stands apart from other Mollicutes because it has larger metabolic capacities and is ubiquitous, being able to live as a saprophyte in soil, compost or wastewaters [128]. The reconstruction of the evolution of translation-related gene set in Mollicutes (Figure 3) indicated that A. laidlawii is probably the species among the Mollicutes that is the closest to the common ancestor with the Firmicutes.
Important aspects of genome downsizing in bacteria concern the accuracy, efficiency and regulation of the minimalist translation process. Recent works at studying aminoacylation of tRNA in vitro demonstrated that several aminoacyl-tRNA synthetases of M. mobile are prone to mistake the amino acid or the tRNA substrate to be charged (discussed above). Such mis-aminoacylations will lead to subsequent incorporation of wrong amino acids into proteins and consequently will reduce the global fitness of the proteome. The possibility that mis-incorporation of amino acids into the nascent polypeptide also occurs because of mis-functioning of the minimalist ribosome cannot be discarded [82]. Elimination of abnormal/misfolded proteins by the usually abundant cellular GroEL/GroES and/or DnaK-dependent chaperone/degradation system acting as promiscuous buffer of genetic variations should not be underestimated (see for example: [129]). As long as the remaining mutant proteins allow cell viability, a low quality of the proteome may even be of some advantage by contributing to the antigenic variation of the mycoplasma exposed to its host's immune response [70], [130].
The genome-scale analysis of soluble complexes in M. pneumoniae has revealed an unexpected high level of protein interaction leading to an estimate of some 200 molecular machines [11]. The ribosome assembly represents one of the most complex networks of interaction. Interestingly, among the 13 polypeptides for which a function was not yet attributed in this specific network, two of them were predicted in our analysis as DEAD-box RNA helicase (MPN623) and as endonuclease M5 (RnmV; MPN072); see Table S3. In fact, MPN623 was curated as an ATP-dependant RNA helicase in the work of Kuhner et al [11], which is consistent with our predictions.
The small number of proteins of the MPSM in Mollicutes is also reminiscent of the translation machinaries in mitochondria and bacterial endosymbionts [131]. However, in the case of mitochondria, a more massive gene and protein loss occurred, resulting in the loss or transfer to the nuclear host genome of majority of bacterial proteins encoding essential genes, including those related to protein synthesis machinery. Of the original bacterial machinery for translation, only genes coding for the structural RNA (t/r/mRNAs), have been preserved (only 16 Kbp in mammalian mitochondria). All the proteins required for the extant/modern mitochondrial ribosome assembly and translation are nuclear encoded, synthesized on the cytoplasmic ribosomes of the cell host, and subsequently imported into the mitochondria via several transport machineries. Despite this unique mitochondrial organization, translation in mitochondria is essentially bacterial-like. One major difference with Mollicutes, even with M. genitalium described above, is that only a small number of mito-mRNAs (mono- and di-cistronic) are translated, all coding for proteins that are part of the membrane reaction centers of the respiratory chain complexes. Consequently, all mito-ribosomes are permanently tethered to the inner membrane and its composition, especially around the polypeptide exit tunnel, is much different from bacterial ribosome. This peculiarity allows a better coordination of the synthesis of the highly hydrophobic mitochondrial proteins and their immediate assembly within the mitochondrial membrane [132]. The possibility exists that, beside the cytoplasmic ribosomes producing mainly soluble cellular proteins, a minor fraction of such specialized membrane-bound ribosomes also exists in Mollicutes, a cellular strategy that certainly allows better efficiency of certain membrane proteins. Another difference is that all mito-mRNAs are leaderless, while in Mollicutes the majority of mRNAs (80% in average [123]) harbor a Shine-Dalgano (SD) sequence that determines the translation initiation pathway followed (Figure 5). Beside these mitochondrial specifications, both organelles and mycoplasmas, uses UGA codon for Trp and the translation factors are essentially the same (except for the lack of mito IF-1), attesting for a very similar translation mechanism as depicted for M. genitalium in Figure 5 (reviewed in: [133], [134], [135]).
Bacterial endosymbionts like Wolbachia (range of genomesize: 958–1,482 Kbp), and Buchnera (422–1,502 Kbp) that infect arthropods and aphids respectively have also evolved in a parasitic life-style by reducing their genome sizes. In some species such as Carsonella ruddii, Candidatus Tremblaya and Nasuia deltocephalinicola, the genomes are even smaller (160–112 Kbp). These tiny bacteria originated about 200 My ago from independent lineages of diverse bacterial groups. At variance with majority of Mollicutes, they cannot be cultivated as free-living organisms and live in a close symbiosis within the host cell, like an organelle. Beside nutrient exchanges, possible protein exchanges between the endosymbiont, the cell host and often cohabiting additional distinct co-endosymbiont(s) remain a matter of debate [136], [137]. Therefore, insect endosymbionts represent a heterogeneous group of organisms and those with the smallest genomes are not ideal model organisms to identify minimal gene sets for autonomous replication. However, examination of the available information on translation genes from a selected set of endosymbionts [138], [139] reveals that most persistent translation machinery genes in these minimal organisms correspond to a large part of the MPSM defined in Mollicutes (see Figure S5). However, from the smallest sets of endosymbiotic proteins it is difficult to build a self-constructing ribosome and successful translation machinery. Evidently in these cases additional proteins from the co-symbiont(s), the host mitochondria or even the host cell would have to complement those translation proteins of the endosymbionts.
Owing to the minimal size of their genomes, Mollicutes have been chosen as the starting point in efforts aiming at building a minimal cell using tools from synthetic biology (for review see [140]). The ambitious goal of these studies is not only to decipher all the functions required for sustaining a minimal life but also for building a cell chassis that could be used in biotechnological processes. Following major progress in DNA assembly, genome engineering and transplantation, this goal seems to be within reach. However, building a minimal cell requires an in-depth knowledge of the cell machinery including of the translation apparatus. Our results should contribute to this goal by providing not only one scenario for the MPSM, but rather a series of possible sets based on the analysis of the different Mollicutes sub-groups. This prediction is now open to experimental verification using synthetic biology.
The phylogenetic tree required for the reconstruction of the ancestral gene sets at the different stages of Mollicutes evolution was generated using concatenated multiple alignments of selected 79 orthologous protein sequences. Proteins encoded by single copy genes present in the genome of all mollicutes were selected. This list is provided in the Figure S1. Multiple alignments were generated using MUSCLE [141], concatenatedusing Seaview [142] and curated from unreliable sites with GBlock [143]. The final concatenated alignment contained 10,686 sites. The phylogenetic tree was constructed by the Maximum Likelihood method using PhyML [144] available on the web server Phylogeny.fr [145]. The list of mollicutes analyzed with some of their genomic characteristics is given in Table S1.
The whole set of proteins of the of Escherichia coli str. K-12 substr. MG1655 and of Bacillus subtilis subsp. subtilis str. 168 translational apparatus were obtained from the Modomics [146], Biocyc [147], SEED [148], SubtiList [27] databases, and Kyoto Encyclopedia of Genes and Genomes [149], plus an extensive review of literature (Table S2).
Homology between E. coli and B. subtilis proteins was inferred by sequence similarity using a reciprocal BLAST search approach (bidirectional best hit). All E. coli and B. subtilis proteins were used as queries for BLAST searches in 39 selected genomes from distinct Mollicutes species included in the MolliGen genome database ([150]; http://www.molligen.org) (Table S3). In this database, initial annotated genomes were obtained from GenBank files. These genomes were further curated by expert annotation that resulted in changes in the functional annotation of specific CDSs and in adding CDSs that were missing in the initial Genbank file. This step of data curation was performed in the frame of the present project for all the homologs involved in translation. Multiple genomes from the same species were excluded from our dataset because initial analyses indicated that no intra-species differences are evident in the gene sets encoding proteins involved in a central process such as translation and ribosome biogenesis. They were nevertheless useful for confirming the presence or absence of a given gene or solving some abnormalities due to occasional sequencing errors in the dataset. BLASTp searches were first conducted with an e-value cutoff of e−8. However, proteins sequences retrieved with an e-value ranging from e−8 to e−3 were maintained in the dataset if a domain related to the considered query was detected using the Conserved Domain search engine [151]. When no hit could be found for a given protein query in one of the Mollicutes genomes, the protein of the closest species identified as a putative hit for this query was used as a query for additional BLASTp and tBLASTn searches. For each query, sequences of the putative Mollicutes homologs were aligned with Clustal W [152]. Subsequent phylogenetic analyses were conducted by using the Neighbour Joining method in Mega5 [153]. Annotation of paralogs was resolved, when possible, by analyzing the microsynteny in MolliGen and the topology of the corresponding phylogenetic trees.
The translation-related gene set at ancestral stages of Mollicutes evolution was inferred using probabilistic and parsimony approaches implemented in the COUNT software package [36]. We used the above described phylogenetic tree and a presence/absence matrix describing the occurrence of 210 genes over 39 Mollicutes genomes and one reference genome, B. subtilis. The posterior probabilities were calculated using a birth-and-death model. We maximized the likelihood of the data set using a gain–loss model with a Poisson distribution at the root. Gain rate for B. subtilis was fixed at 0 to avoid false prediction of many gene gains by this species. Several combinations of parameters were tested to maximize the likelihood. The best value was obtained with the edge length, loss and gain rates set at 4 gamma categories. Edge length and loss rate parameters had more impact than gain rate on the final likelihood of the optimized model. Wagner parsimony [37] was also used to infer ancestral gene sets. A gain penalty of 4 was used to minimize predicted gene gain events, in accordance with the massive genome reduction context of Mollicutes evolution.
|
10.1371/journal.pgen.1000628 | The Limits of Individual Identification from Sample Allele Frequencies: Theory and Statistical Analysis | It was shown recently using experimental data that it is possible under certain conditions to determine whether a person with known genotypes at a number of markers was part of a sample from which only allele frequencies are known. Using population genetic and statistical theory, we show that the power of such identification is, approximately, proportional to the number of independent SNPs divided by the size of the sample from which the allele frequencies are available. We quantify the limits of identification and propose likelihood and regression analysis methods for the analysis of data. We show that these methods have similar statistical properties and have more desirable properties, in terms of type-I error rate and statistical power, than test statistics suggested in the literature.
| It was shown recently by Homer and colleagues that it may be possible to determine whether a person with known genotypes at a number of markers was part of a pool of DNA from which only frequencies of alleles at the markers are known. In this study, we quantify how well such identification can work in practice. The larger the size of the sample from which the allele frequencies are available, the more independent genetic markers are required to allow individual identification.
| Homer et al. [1] showed that it was possible in some circumstances to identify whether a person with observed genotypes at multiple loci was part of a sample from which only estimated allele frequencies were known. Such identification would be particularly useful in forensic science if the presence or absence of a person's DNA in a mixture of DNA could be established. The authors also discussed the relevance of their findings when summary statistics such as allele frequencies were available in the public domain as part of genotype-phenotype studies, because it possibly could be established that individuals, or their close relatives, were part of a particular study. As a result of the publication of Homer et al., NIH and the Wellcome Trust added more restrictions to the access of such data to avoid potential identifiability (http://grants.nih.gov/grants/gwas/data_sharing_policy_modifications_20080828.pdf).
The approach taken by Homer et al. was to have two samples with estimated allele frequencies, here called the “test” and “reference” sample, and to ask whether an individual was ‘close to’ either of these samples, using a statistic that measured a distance to the sample. The properties of the test statistic were not investigated theoretically (although simulation studies were performed), and the difference between “sample” and “population” was not always clear.
In this note we take a best-case idealised setting in which there is a single population from which there is a test sample with allele frequencies at a number of loci and from which there is a single individual, called the proband, with full genotypes. The question is whether the person was part of this test sample from which allele frequencies are available. We use both likelihood and linear regression theory, which illustrate different approaches to the problem, to draw inference about the hypothesis that a proband was part of the test sample. We show that the power of identification of a proband as part of a test sample is, approximately, proportional to the number of independent SNPs divided by the size of the sample from which the allele frequencies are available. The power is reduced by a predictable magnitude if the frequencies in the population are themselves estimated imprecisely. Properties of likelihood-ratios and regression test statistics and a comparison with the statistic used by Homer et al. were verified by simulation.
There are m independent SNP markers with a population frequency of pi for allele B at the ith SNP. We assume Hardy-Weinberg equilibrium in the population, so that the genotype proportions for the ith SNP are (1−pi)2, 2pi(1−pi) and pi2 for genotypes AA, AB and BB, respectively. We have estimated allele frequencies based upon a test sample of N unrelated individuals. In the test sample of 2N alleles, ni is the number of B alleles at locus i. In this study we assume that N is known and individuals are equally represented in computing . Note that these conditions are unlikely to be fully met in forensic applications when the test sample may be a DNA pool and we consider the implications later.
The genotype for proband X at the ith SNP is gi, which can take values of 0, 1 and 2 for genotypes AA, AB and BB, and the expectation of yi = ½gi is the population frequency pi, i.e. E[½gi] = pi.
To simplify derivations, we shall first assume the population frequencies pi, are known. More generally, we assume we have prior unbiased estimates of the allele frequencies from the same population from a different finite sample (the “reference sample”) of size N*, in which there are n*i B alleles at locus i. As both the test and reference samples are drawn independently from the population, the best estimate of the frequency in the population is given by the pooled value, It is explained subsequently why this estimate, rather than say n*i/2N*, the estimate of the allele frequency from the reference sample, is used in the statistical analysis.
We show that the main results for the regression approach are based upon the expectation that the regression of the proband frequency, yi = ½gi, on , each expressed as deviations from population frequencies, is distributed about unity for all loci if the proband was part of the test sample, and about zero otherwise.
Population allele frequencies on m markers were drawn from a uniform distribution with lower bound 0.05 and upper bound 0.95 (i.e., minor allele frequency (MAF)>0.05). For the ith SNP, a genotype score (yi) of a proband was simulated from a binomial distribution with probability pi and sample size 2. Allele frequencies in the reference and test samples were simulated from a binomial distribution with probability pi and sample size 2N* and 2N, respectively. If the proband was part of the test sample then the test sample was simulated on N−1 individuals and the allele count from the proband was added to that from this sample to create a sample from N individuals. Linear regression was performed as described previously, for a type-I error rate of 0.05, and the Homer et al. [1] test statistic (see Text S2) was also implemented. 1000 simulations were performed for combinations of N = 100, 1000, 10000, N* = 100, 1000, 10000 and ∞ and m = 50,000, when the proband was either part or not part of the test sample.
The results are shown in Table 1. The regression type-I error rates are well controlled when the hypotheses tested are true. As predicted (Text S2), the type-I error rates for the Homer et al. test statistic are not well controlled. In many cases the probability of rejecting the null hypothesis when it is true is close to zero. Power to determine whether the proband is part of the test sample is good for test samples of 1000 if the reference sample size is large. Inference from the regression and likelihood-ratio approach is similar, as expected (Table S1).
Simple methods were proposed to test the hypothesis of whether a proband was part of a test sample. The expected likelihood ratio or the power to reject the null hypothesis when it is false were derived and shown to be a simple function of m/N, the ratio of the number of markers and test sample size. If allele frequencies in the population are well-estimated then there is good power to determine if a proband is part of a sample of ∼1000 individuals when using a whole genome scan of ∼50,000 independent markers.
There is a strong relationship between the logLR statistic and regression test statistics. The difference in the two regression test statistics, in or out of the test sample, is approximately equal to twice the logLR statistic. Hence, twice the logLR statistic is very similar to a test statistic from regression that also tests for the in vs out hypothesis (Table S1).
Could any inference be drawn in the case where there are no prior estimates of allele frequencies? The analyses indicate that, even with many marker loci, there is little power as N* approaches 0 unless the sample size N is also very small, and no larger than N*.
The parameter m was defined as the number of independent SNPs. When many SNPs are used, e.g. all common SNPs on a chip, then there is correlation (linkage disequilibrium) among the SNPs. Consequently, the y variables (allele numbers in the proband) are correlated and not taking this into account will inflate the test statistic because the true variance of the estimated regression coefficient is larger than appears from the total number of SNPs. Similarly, the variance of the likelihood statistic is increased if allele frequencies across SNPs are correlated. There are a number of ways to deal with this correlation structure. (i) Restrict the analyses to SNPs that are in linkage equilibrium. This seems wasteful because information is discarded. (ii) Take the correlated nature of y into account by fitting the covariance structure of y into the regression or likelihood analysis. The effect of LD on the variance of the log likelihoods is shown earlier, and appropriate corrections using the mean r2 given. In view of the correspondence of the likelihood and regression approaches, the same correction can be applied to the latter. The relevant quantity may be obtained from a separate data set (e.g. HapMap). (iii) Perform a theoretical adjustment on the test statistic, by calibrating the variance of the test statistic on the equivalent number of independent markers. According to population genetics theory, the number of independent loci (‘segments’) in a random population with effective size Ne and genome length L (Morgan) is approximately 2NeL/log(4NeL) [3]. For human populations, with Ne = 10,000 and L = 35, this implies a total of ∼50,000 SNPs. This number can also be estimated using a simulation approach, conditioning on the observed LD structure in a sample where individual-level genotype data are available. Such an application resulted in ∼55,000 independent SNPs for one genome-wide association study [4].
In our derivations we have assumed that all samples (proband, reference and test) are from the same population and that within the population there is random mating. What if these assumptions are violated?
If all samples are from the same population but there is deviation from HWE then the tests are somewhat biased because HWE is assumed in computing the likelihood and the variance of sample allele frequencies. Population differences are more serious and can lead to the wrong inference. There are a large number of possibilities because, in principle, the proband, reference and test samples can all come from different populations. However, population differences between the reference and test sample can be tested explicitly using standard tests for differences in gene frequency. There seems little point in testing whether a proband was part of a specific test sample when there is no reference sample from the same population. Nevertheless, what can we predict if the reference population is not actually from the same population, but is used as if it is? Then both the likelihood statistics for the hypothesis ‘in’ and ‘out’ are inflated, by essentially the same amount, so the problem is not the divergence between the two populations, but bias in the test statistic. If population frequencies are inappropriately or approximately estimated, the sample is more likely to be assigned as ‘in’ when it should not be. The reference sample is of little value if the divergence between the populations, expressed as Wright's FST, approaches 1/(2N).
Can we quantify the limits of identification in practical situations? This is hard, because there are (at least) three difficulties in addition to the theoretical sample m/N criterion:
For these reasons we cannot set a simple limit to identification without reference to other parameters (or speculation).
In the analysis we have not considered the possibility that the proband is not in the test sample, but is related to one or more persons who is. For example if a relative with relationship R (e.g. R = ½ for full sibs) is in the test sample, then the expectation of the regression coefficient is E(b) = R rather than 0 or 1. Similar calculations can be done if, for example, there are several relatives in the test or reference samples. If many markers are used, a value of b of approximately one-half would raise suspicions that in fact a full sib, parent or child is in the test sample. Lower, but non-zero values could be consequences of sampling or relationship. The simulation results in Table 1 illustrate how sensitive the methods can be, and hence there seems a real possibility of identifying not just the proband but also his/her relatives.
A problem frequently met in forensic applications is whether a particular individual's DNA appears in a mixture obtained at a crime scene, for example. In this case, it is usually unknown how many individuals' DNA is present in the sample (i.e., N is unknown), equal representation cannot be assumed, and there may be allelic drop out in the sample, although Homer et al. [1] showed empirically that probands could be detected even if their contribution to the DNA pool was small. We do not therefore consider the present results to be relevant for probabilistic inference in a forensic setting. However, exclusion of a proband from a pooled DNA sample is possible if many markers are used, the actual N is small and frequencies of alleles from the pool are estimated accurately. The likelihood framework is sensitive to genotyping errors in that false exclusions could occur, but the analysis could be adapted to model genotype counts with specified probability of errors or by assuming replacement sampling in computing P(in). The linear regression approach is likely to be robust to genotyping error.
In contrast to forensic applications, in the situation considered by Homer et al. in which the test sample is a database constructed using a specified number of individuals each with individual genotypes, and with the gene frequencies estimated as their average, our results support their conclusions. Probands that were part of a test sample could be identified even for samples sizes of 1000. If, for example, there are both diseased case and healthy control samples in the association test, each assumed to be sampled from the same population, then it is possible to test whether an individual is present in either the case or control group using the analysis we have described, but using each sample in turn as the test sample.
Current genome-wide association studies (and meta-analyses based upon multiple studies) are conducted on large samples, often of the order of 10,000 or so, and in this case our results show that the power to identify a proband who was part of such a large sample when the reference sample is of similar size is only about one-half (Table 1) assuming 50,000 independent loci, even under the ideal circumstances considered in this study.
|
10.1371/journal.pgen.1003311 | Estrogen Mediated-Activation of miR-191/425 Cluster Modulates Tumorigenicity of Breast Cancer Cells Depending on Estrogen Receptor Status | MicroRNAs (miRNAs), single-stranded non-coding RNAs, influence myriad biological processes that can contribute to cancer. Although tumor-suppressive and oncogenic functions have been characterized for some miRNAs, the majority of microRNAs have not been investigated for their ability to promote and modulate tumorigenesis. Here, we established that the miR-191/425 cluster is transcriptionally dependent on the host gene, DALRD3, and that the hormone 17β-estradiol (estrogen or E2) controls expression of both miR-191/425 and DALRD3. MiR-191/425 locus characterization revealed that the recruitment of estrogen receptor α (ERα) to the regulatory region of the miR-191/425-DALRD3 unit resulted in the accumulation of miR-191 and miR-425 and subsequent decrease in DALRD3 expression levels. We demonstrated that miR-191 protects ERα positive breast cancer cells from hormone starvation-induced apoptosis through the suppression of tumor-suppressor EGR1. Furthermore, enforced expression of the miR-191/425 cluster in aggressive breast cancer cells altered global gene expression profiles and enabled us to identify important tumor promoting genes, including SATB1, CCND2, and FSCN1, as targets of miR-191 and miR-425. Finally, in vitro and in vivo experiments demonstrated that miR-191 and miR-425 reduced proliferation, impaired tumorigenesis and metastasis, and increased expression of epithelial markers in aggressive breast cancer cells. Our data provide compelling evidence for the transcriptional regulation of the miR-191/425 cluster and for its context-specific biological determinants in breast cancers. Importantly, we demonstrated that the miR-191/425 cluster, by reducing the expression of an extensive network of genes, has a fundamental impact on cancer initiation and progression of breast cancer cells.
| MicroRNAs are small noncoding RNAs that act as posttranscriptional repressors of gene expression. A pivotal role for miRNAs in all the molecular processes driving initiation and progression of various malignancies, including breast cancer, has been described. Divergent miRNA expression between normal and neoplastic breast tissues has been demonstrated, as well as differential miRNA expression among the molecular subtypes of breast cancer. Over half of all breast cancers overexpress ERα, and several studies have shown that miRNA expression is controlled by ERα. We assessed the global change in microRNA expression after estrogen starvation and stimulation in breast cancer cells and identified that miR-191/425 and the host gene DALRD3 are positively associated to ERα-positive tumors. We demonstrated that ERα regulates the miR-191/425 cluster and verified the existence of a transcriptional network that allows a dual effect of estrogen on miR-191/425 and their host gene. We show that estrogen induction of miR-191/425 supports in vitro and in vivo the estrogen-dependent proliferation of ERα positive breast cancer cells. On the contrary, miR-191/425 cluster reprograms gene expression to impair tumorigenicity and metastatic potential of highly aggressive ERα negative breast cancer cells.
| MicroRNAs (miRNAs) are a class of evolutionarily conserved regulatory RNAs that pleiotropically suppress gene expression at post-transcriptional level [1]. MiRNAs control the expression of 10–30% of the human transcriptome and are crucial regulators of both physiologic and pathologic processes [2]–[4]. In cancer, the spectrum of miRNAs expressed in neoplastic cells differs dramatically from that found in normal cells and it is now well established that miRNAs play fundamental roles in essentially all aspects of tumor biology [5], [6].
In breast cancer, divergent miRNA expression between normal and neoplastic tissues has been demonstrated, as well as differential miRNA expression among the molecular subtypes of breast cancer, including luminal A, luminal B, Her2+ and basal-like [7], [8]. MiRNAs have been shown to play an important role in breast cancer initiation and progression. For example, overexpression of miR-21 in breast carcinomas has been shown to target important tumor-suppressor genes such as PTEN, PDCD4, and TPM1, and was associated with advanced clinical stage, lymph node metastasis, and poor patient prognosis [9], [10]. MiR-10a was reported to be overexpressed in about 50% of metastatic breast cancer and transcriptionally activated by the pro-metastatic transcription factor TWIST1 [11]. Reduced expression of miR-126 and miR-335 in the majority of primary breast tumors from relapsed patients was reported, and simultaneous loss of miR-126 and miR-335 expression was associated with poor distal metastasis-free survival [12]. Oncogene regulation by miRNAs has also been reported, including tyrosine kinase receptors HER-2 and HER-3 by miR-125b and miR-205, respectively [13], [14], and the miR-200 family, known to reduce cell migration and invasiveness by targeting ZEB transcription factor members, was suppressed in metastatic breast cancer [15], [16].
miRNA regulation by estrogen receptor-alpha (ERα), the most important prognostic and therapeutic indicator in breast cancer, has recently been described by us and others [17]–[20]. Specifically, the majority of miRNAs upregulated by ERα are key components of a negative feedback loop that restrict E2 action and thus play a tumor suppressive role. In this regard, ERα-activation of let-7 family members limits the expression of oncogenes, such as Ras and c-Myc, and promotes differentiation of cancer cells [18]; ERα-mediated activation of the miR-17/92 cluster functions as a tumor suppressing mechanism in breast cancer through the downregulation of cyclin D1 and AIB1 by the miR-17/20/106 family and the direct suppression of ERα mediated by miR-18 and miR-19 [19]. We and others have described a double-negative feedback loop involving E2-suppressed microRNAs that target ERα, specifically miR-206 and miR-221&222, resulting in upregulation of ERα expression and low miRNA level in luminal A-type breast cancers [17], [21].
Recent works from our group have shown that miR-191 is highly induced in several human solid tumors including colon, lung, pancreas, prostate, and stomach cancer [22], as well as acute lymphocytic leukemia (ALL)-associated hematopoietic malignancies [23]. We have also reported a strong positive correlation between miR-191 expression and ERα levels in breast tumors [7], suggesting an oncogenic function for this miR. A role for miR-191 in tumorigenesis is further strengthened by several findings, including that miR-191 is induced by a dioxin family carcinogen, the miR is hypomethylated and overexpressed in liver cancer [24], [25], and miR-191 inhibition decreases cell proliferation and tumor growth of hepatocellular carcinoma cells [24]. Furthermore, miR-191 overexpression promotes cell growth and suppresses apoptosis of gastric cancer cells [26]. However, in ovarian and thyroid follicular cancer, miR-191 represses MDM4 or CDK6 expression, respectively, thereby delaying cancer progression and tumor-related death [27], [28]. These contradictory findings indicate that the precise role for miR-191 in human neoplasia may be tumor-type specific and not well understood.
In this current study, we report a positive association between ERα expression and miR-191 and miR-425, two intronic miRNAs hosted by the putative protein coding gene DALR anticodon binding domain containing 3 (DALRD3), and further show direct control of the miR-191/425/DALRD3 transcriptional unit by the E2/ERα axis. We evaluated that the estrogen dependent activation of miR-191/425 induces proliferation in part by targeting the estrogen modulated tumor-suppressor gene, EGR1. We also demonstrated that, when constitutively expressed in highly aggressive ERα negative breast cancer cells, the miR-191/425 cluster reprograms gene expression to impair tumorigenicity and metastatic potential through the suppression of several different oncogenic proteins.
MiR-191 and miR-425 are highly conserved miRNAs found on human chromosome 3 within the first intron of DALRD3 (Figure S1). Given their genomic organization and proximity, we hypothesized that miR-191 and miR-425 are co-transcribed and transcriptionally dependent on the host gene DALRD3. We examined expression of mature miR-191, miR-425, and DALRD3 mRNA in 20 different normal human tissues using qRT-PCR (Figure S2A). Both miRNAs were detected in all tissues and, their levels of expression were highly correlated, as shown by scatter plot analyses, (R2 = 0.7351; p<0.001) (Figure S2B). However, only a partial correlation was observed between the host gene DALRD3 and miR-191 (R2 = 0.4058; p<0.001) or miR-425 (R2 = 0.2101; p<0.001) (Figure S2B), suggesting the existence of DALRD3-independent mechanism of miR-191/425 expression/accumulation in some tissues.
Based on the previous association between miR-191 and ERα and the miR-191 and miR-425 co-expression results (Figure S2A), it was of interest to examine ERα positive breast tumors for the expression of miR-191 and miR-425. qRT-PCR analysis of 44 human breast cancer specimens with different ERα status revealed that miR-191 and miR-425 expression was higher (p-value<0.01) in ERα positive than ERα negative tumors (Figure 1A). DALRD3 mRNA also showed a significant positive correlation with the ERα status (Figure 1A and Figure S3A). Next, to further verify the positive association between ERα levels and miR-191/425 expression, miRNA in-situ hybridization was performed on an independent set of 132 human breast cancer specimens. As anticipated, the majority of ERα positive breast tumors were also miR-191 (80%) and miR-425 (87%) positive, while only 23% and 15% of ERα negative specimens expressed miR-191 and miR-425, respectively (Figure 1B and Figure S4A). Furthermore, co-labeling of miR-191 and miR-425 by miRNA in situ-hybridization on the same ERα positive breast specimens showed co-localization of the two microRNAs in the majority of breast tumor cells (Figure S4B).
Finally, a set of 16 different breast cancer cells, clustered by ERα, progesterone receptor (PR) and HER2 expression was also analyzed for the expression of miR-191, miR-425 and the host gene DALRD3. Expression of both miR-191 and miR-425 was higher in the ERα positive cell lines, with the exception of MDA-MB-453 (a non-aggressive ERα negative/androgen receptor positive breast cancer cell line but with a gene expression profile that overlaps with ERα positive breast cancer cells [29]) (Figure 1C, 1D). DALRD3 expression correlated with the expression levels of the mature miRNAs (R2 = 0.725 for miR-191, p<0.01; R2 = 0.63 for miR-425, p<0.01) (Figure 1C, 1D and Figure S2C). Moreover, we assessed the expression levels of the two different alternative splicing variants of DALRD3 and confirmed that the two variants are both transcribed and their expression levels are higher in the ERα positive than ERα negative breast cancer cells (Figure S3B). Taken together, these data revealed for the first time that miR-191 and miR-425 are co-transcribed and preferentially expressed in ERα positive breast cancer cells and tumors.
Recently, various microarray approaches have been used to identify E2-induced miRNA expression in hormone-dependent breast cancer cells [17]–[20]. However, based on the lack of consensus on E2-regulated changes in miRNA expression [30], we investigated global changes in endogenous miRNA expression after E2 stimulation of breast cancer cells using the multiplexed Taqman microRNAs assay, a highly sensitive technology that allowed us to detect changes in 754 miRNAs (“miRNome”) with the same sensitivity of a Taqman realtime PCR. ERα positive MCF7 cells were hormone starved for 6 days and then exposed to 10 nM of E2 for 6 h. The miRNome was determined at 2, 4, 6 days of hormone deprivation and 6 h after E2 stimulation (Figure 2A and Table S1). After 6 days of E2 deprivation, downregulation of 146 and upregulation of 25 mature miRNAs, organized in 69 different miRNA genes, were observed (fold change 1.2, p-value<0.05) (Figure 2A). Of these 69 miRNA genes, 43 genes (85 mature miRNAs) were modulated after 6 h of E2 stimulation (Figure 2A). The miR-191/425 cluster showed a progressive downregulation during the 6 days of hormone deprivation (p-value<0.05; fold change miR-191: 2 d: 0.82; 4 d: 0.63; 6 d: 0.45; miR-425: 2 d: 0.81; 4 d: 0.78; 6 d: 0.35) followed by a significant induction by 6 h of E2 stimulation (p-value<0.05; fold change miR-191: 1.36; miR-425: 1.12) (Table S1). We assessed the reliability of the treatment by using qRT-PCR to evaluate the expression levels of the E2-regulated genes, TFF1/pS2 and miR-17 after 3, 6, 24, 48 and 72 h of E2 stimulation [19] (Figure S5A). Both genes showed a strong and stable induction over time after E2 treatment. Next, we performed qRT-PCR on miR-191 and miR-425 and both miRNA levels increased after E2 stimulation although with a different kinetic of induction compared to miR-17 (Figure 2B). Specifically, after 72 h of E2 treatment, we detected a 2- to 3.5-fold induction of miR-191 and -425 compared to untreated cells and the presence of a block in their induction at 24 h after E2 treatment (Figure 2B). Next, we assessed expression levels of the primary precursor of miR-191 and miR-425; the induction profile was similar to the mature miRNAs (Figure S5B). Despite the positive correlation between miR-191/425 and the host gene DALRD3 in breast cancer cells (Figure 1C, 1D), the expression level of the total DALRD3 mRNA was decreased of 35% after 72 h of E2 treatment compared to untreated cells (p-value = 0.053) (Figure 2B). qRT-PCR for the two different alternative splicing variants of DALRD3 also showed a repression of both variants after estrogen stimulation (Figure S5C). Moreover, total DALRD3 mRNA and both variants were also highly upregulated in hormone-deprived MCF7 cells (Figure S5D). To further confirm the ability of E2 to modulate miR-191/425, MCF7 were treated with fulvestrant, an ERα antagonist that induces ERα protein degradation (Figure S6A). We observed a consistent reduction in miR-191/425 levels and a constant increase in DALRD3 levels after fulvestrant treatment (Figure S6B, S6C). TFF1/pS2 expression was downregulated by hormone deprivation or fulvestrant treatment (Figure S5D; Figure S6C). Collectively, the data showed that miR-191/425 levels are positively regulated by ERα, and the increased levels of miR-191 and miR-425 after estrogen stimulation are associated with a reduction in the accumulation of the host gene DALRD3.
Next, we addressed the direct involvement of ERα in the regulation of miR191/425 cluster by performing chromatin immunoprecipitation (ChIP) experiments across nine different regions spanning miR-191/425 cluster and covering a region of 4200 bp (Figure 2C). MCF7 cells were E2 starved for 6 days (0 h) and then treated with E2 (10 nM) for 3 h, 6 h and 24 h. Enrichment of ERα after E2 treatment was identified at region 3 and 8 (Figure 2C). Region 3 showed a specific enrichment of ERα that reached the highest levels after 3–6 h of treatment and started to decrease at 24 h. Although ERα was also detected at region 8 after 3 h and 24 h of E2 treatment, this enrichment was considered to be non-specific since it was also detected for the ERα negative MDA-MB-436 cells (Figure 2C). We also examined the localization of the non-phosphorylated RNA polymerase II large subunit (polII) and the acetylation status of the histone H3 (AcH3) after E2 treatment (Figure 2C). Immunoprecipitation against polII showed the presence of two different areas of enrichment: region 3, with an E2-dependent recruitment of polII that decreased over time, and region 7–9 which showed a progressive reduction in polII recruitment during E2 treatment (Figure 2C). AcH3 ChIP showed a specific enrichment at region 1, 2 and 8 with a significant increase in H3 acetylation after 6 h of E2 treatment only for region 2 (Figure 2C). Taken together, these experiments show that ERα is recruited to the miR-191/425 genomic locus, in response to the estrogen stimulation.
Because of the presence of two sites of enrichment of polII and the presence of two CpG islands located at the 5′end of the two isoforms of DALRD3 (Figure 2C), we hypothesized the existence of two promoter regions: one responsible for the transcription of the longest isoform of DALRD3, which includes miR-191 and -425 and a second responsible only for the transcription of the short isoform of DALRD3. Computer-assisted analysis (Figure S7A) identified two distinct predicted regions as possible candidates for promoters regulating miR-191/425/DALRD3 gene transcription: 3900 bp (prom1) a marginal predicted region, located upstream of the long isoform of DALRD3 and also involved in the production of miR-191/425; 6500 bp (prom2) a highly likely predicted region, associated only to the transcription of the short isoform of DALRD3 mRNA (Figure S7A). To test the transcriptional activity of these two elements, both putative promoters (Figure 2D) were cloned individually in the promoter-less pGL3basic luciferase vector, and their expression was examined in HEK293 cells. Both vectors showed an increase in the luciferase activity, and as expected, the highly likely predicted region prom2 showed the strongest basal luciferase activity (Figure 2D). Next, we assessed the E2 responsiveness of the two identified promoter regions. We first tested the luciferase activity of both plasmids in five breast cancer cell lines with different ERα expression levels (Figure S7B). Both promoter elements showed higher levels of activity in the three ERα positive cell lines (MCF7, T47D, BT-474) compared to the ERα negative cells (BT-459, MDA-MB-436). Treatment with E2 for 6 h induced a 3-fold increase in luciferase activity for the prom1 element (Figure 2E); in contrast, luciferase activity for the prom2 region was repressed by E2 treatment (Figure 2E). Furthermore, silencing of ERα by siRNA reduced luciferase activity of the prom1 reporter vector by approximately 50% specifically in ERα positive cells, but no effect on prom2 activity was detected (Figure S7C).
Taken together, these experiments showed that (1) ERα directly regulated miR-191/425 cluster expression and (2) verified the existence of two promoter elements involved in the transcription of the two DALRD3 isoforms, allowing a differential accumulation of miR-191/425 and DALRD3 upon E2 stimulation.
To identify the functional role of the E2 mediated-induction of miR-191 and miR-425 in ERα positive breast cancer cells, both miRNAs were knocked down in estrogen dependent MCF7 cells in normal culture condition. A 33% reduction in cell proliferation rate was observed compared to a control oligonucleotide (Figure 3A). Indeed, enforced expression of miR-191/-425 in hormone deprived MCF7 cells, with low levels of endogenous miR-191/425 (Table S1), induced a 70% increase in cell proliferation (Figure 3A). To shed more light in the proliferative effects of miR-191/425 in ERα positive breast cancer cells, flow cytometric analyses of transiently-transfected cells were performed and revealed an increased number of cells in G1 and fewer cells in G2/M following knockdown of either miR-191 or miR-425 compared to control cells (Figure 3B and Figure S8A). Moreover, enforced expression of miR-191/425 in hormone deprived MCF7 cells protects cells from hormone starvation induced apoptosis (Figure 3C).
We next evaluated the in vivo effect of miR191/425 knockdown on tumor growth. Specifically, miR-191/425 were transiently inhibited in ERα positive MCF7 cells for 48 h and tumor growth was assessed after subcutaneous transplantation of the transfected MCF7 cells in nude mouse. A 50% reduction in tumor growth was observed (Figure 3D) in miR-191/425 knocked-down cells compared to control cells. Same results were also obtained after xenotrasplantation of miR-191/425 knocked-down ERα positive ZR-75-1 cells (Figure S8B).
To uncover the molecular players involved in the proliferative response of ERα positive breast cancer cells controlled by the E2 mediated activation of miR-191/425, published transcriptomic data set of E2 induced ERα positive MCF7 and ZR-75-1 cells were compared with the predicted miR-191/425 target genes [31], [32]. Specifically, the target genes of miR-191 and miR-425 obtained from the prediction program Targetscan v5.2 (Table S2) were compared with the pool of E2 downregulated genes. 43 and 23 miR-191 targets and 199 and 116 miR-425 targets were found in the E2 repressed gene lists of MCF7 and ZR-75-1, respectively (Figure S9A). Only 5 and 18 targets for miR-191 and miR-425 were repressed by estrogen in both cell lines respectively (Figure S9A). We focus our attention on the early growth response 1 (EGR1), a member of the early growth response (EGR) transcription factor family that has been implicated in breast cancer progression and antiestrogen resistance [33]–[35]. First, the expression levels of EGR1 were assessed after E2 stimulation in MCF7 cells. EGR1 expression showed a 50% induction after 30 minutes from the stimulation (Figure S9A) followed by a continuous repression (Figure S9B). To verify that miR-191 regulates the expression of EGR1, knockdown of miR-191 was performed in MCF7 cells and western blot analyses confirmed the upmodulation of EGR1 and its direct transcriptional target CDKN1A (p21) (Figure 3E) [34]. Next, to assess that miR-191 directly controls EGR1 in cells, a luciferase reporter assay was performed with a luciferase expressing plasmid containing the conserved miR-191 predicted binding site for EGR1 cloned after the luciferase reporter gene (Figure 3F). Co-transfection of miR-191 with the reporter plasmid significantly suppressed (p-value<0.01) the luciferase activity of the reporter, relative to transfection of the control oligonucleotide (Figure 3F). Disruption of the predicted binding site reduced the inhibitory activity of miR-191 overexpression on the luciferase activity (Figure 3F). To study in more depth the interaction miR-191/EGR1, hormone deprived MCF7 cells were transfected with miR-191 inhibitor and control oligonucleotide and 48 h later treated with estradiol. Western blot analyses showed that miR-191 inhibition prevents EGR1 degradation at 6 h and 24 h after E2 treatment compared to control cells (Figure 3G). qRT-PCR showed that EGR1 mRNA is also under the control of miR-191 but only in the early phase of E2 induction (Figure 3H). As expected, induction of p21 transcript was confirmed by qRT-PCR specifically in miR-191 knocked-down (Figure 3H). MiR-191 inhibition was also confirmed by qRT-PCR (Figure S9C). Taken together, these results highlight the proliferative effects of E2-induced miR-191/425 cluster in ERα positive breast cancer cells that are in part related to the miR-191 repression of the tumor-suppressor gene EGR1.
Approximately 75% of diagnosed breast tumors express ERα, and this ERα-positive status is associated with a better prognosis and response to hormonal treatment [36]. Several studies suggested that a fraction of ER-negative tumors arise from ER-positive precursors [37]. Moreover, restoration of functional ERα expression in ERα-negative human breast cancer cells can block their proliferation and aggressiveness, supporting the notion that ERα confers a less aggressive phenotype of breast cancer [38], . To determine if miR-191/425 cluster as a part of the ERα signaling can partially mediate the anti-proliferative effect that ERα showed in the aggressive breast cancer cells, a genome-wide expression analysis in aggressive MDA-MB-231 cells, which express low levels of miR-191/425, was performed 72 h after enforcing expression of both miR-191 and miR-425 and control oligonucleotide (Figure 4A). Unsupervised clustering analyses showed significant deregulation of gene expression by miR-191/425, with 753 upregulated and 1105 downmodulated genes (by>1.5 fold; p-value 0.001) (Table S2). Functional profiling of these genes defined that the greatest proportion of them is associated with cell adhesion, adherens junction followed by phosphatidylinositol signaling (Figure 4B). We used qRT-PCR to validate the modulation of over 20 genes identified in the microarray analyses or to their related molecular pathways in two different breast cancer cell lines (Figure 4C). Expression of many genes involved in promoting growth and metastasis of breast cancer cells was found to be downmodulated by miR-191/425 cluster: CCND1, CCND2, E2F1, CSDA and API5, regulatory proteins of the cell cycle progression and apoptosis [40]–[42]; FSCN1, TNC, VEGFA, CDC42 and SOX4, which have roles in angiogenesis and migration, and are involved in filopodia/invadopodia formation [43]–[48]; the protooncogene MYC, which initiates the transcription of a large set of genes involved in cell growth by stimulating metabolism and protein synthesis [49]; and SATB1, which reprograms gene expression to enhance aggressive histomorphological features and invasive capabilities [50]. We also found that miR-191/425 cluster represses cell-structure and adhesion genes typical of invasive breast cancer cells such as fibronectin, an ECM adhesive glycoprotein, and vimentin, the intermediate filament protein of mesenchymal cells, which together provide cellular integrity and resistance against stress [51]. Finally, miR-191/425 cluster upregulates zonula occludens-1 (ZO-1), a component of the tight junction barrier in epithelial and endothelial cells [52]; E-cadherin (CDH1), an important marker of epithelial tumor progression; and β-catenin (CTNN1) a component of wnt pathway that drives progression in various cancers [53].
To confirm that targets of the mir-191/425 cluster showed an enrichment signature in this dataset, we assessed the cumulative density function (cdf) plot comparing the expression changes of mir-191 and miR-425 targets based on TargetScan v5.1 gene list [54]. We found that the mir-191/425 targets set (targets) was more repressed than the control set of genes (control) matched for 3′UTR length, dinucleotide composition, and expression level (Figure 5A). Stronger repression was observed for the conserved miR-191/425 cluster targets (conserved targets), suggesting further enrichment of genuine targets in this set (Figure 5A). These observations supported the utility of this expression data for the discovery of novel miRNA targets based on miR-associated genes. Because the expression levels of target mRNAs tend to correlate negatively with the expression levels of their specific miRNAs [55], we next focused on the miR-191/425 downregulated genes. First, the target prediction program TargetScanv5.1 was used to search for predicted target genes of miR-191 and miR-425 in the pool of downregulated genes in miR-191/425-expressing MDA-MB-231 cells (Table S2). This list of genes was further compared with the list of target genes downregulated exclusively by the expression of miR-191 or miR-425 (Figure S10 and Table S2). A total of 37 and 346 downregulated targets were obtained for miR-191 and miR-425, respectively (Figure 5B). Among these large set of genes, we selected 12 genes (SATB1, CCND2, CTDSP2, SOX4, LRRC8A, SLC16A2, CSDA for miR-191 and FSCN1, TNC, SIAH2, CCND1, CSDA for miR-425) predicted to have at least one potential binding site for miR-191 and/or mir-425 in their 3′UTRs. Based on their reduction in miR-191/425-expressing cells (Figure 4C and Table S2), we tested whether these genes are direct targets of miR-191 and miR-425 constructing reporter plasmids containing the miRNA binding site in the 3′UTR of these genes downstream of a luciferase reporter gene (Figure S11A). Co-transfection experiments showed that the introduction of either miR-191 or miR-425 markedly suppressed the expression of a luciferase containing the 3′UTR of these downregulated genes (Figure 5C) but did not affect the luciferase activity of the 3′UTR-CCND1 plasmid, indicating that CCND1 is not a direct target of miR-425 (data not show). Mutations that disrupt base paring with miR-191 and miR-425 rescued the luciferase expression for all the target genes, further confirming that these genes are direct targets of miR-191 and miR-425 (Figure S11B).
We next focused our attention exclusively on SATB1, CCND2 and FSCN1 as mediators of miR-191 and miR-425 effects, respectively, because of their strong repression obtained after miRNA expression and their reported tumorigenic function in breast cancer [46], [50], [56], [57]. Western blot analyses on MDA-MB-231 expressing either miR-191 or miR-425 showed a strong suppression of SATB1 only after enforced miR-191 expression (Figure 5D). Because of SATB1 repression, we also detected marked repression of fibronectin and to lesser extent of vimentin (Figure 5D). Further, we also observed a ∼2 fold increase of the β-catenin protein (Figure 5D) and its sequestration at the cytoplasmic membrane due to the increased expression of e-cadherin (Figure 5D, 5E). Indeed, miR-191 over-expressing cells also showed a specific repression of CCND2 as well as CDK6 (Figure 5D), a previously demonstrated miR-191 target [28]. Furthermore, we observed a decrease in the levels of CCND1, E2F1 and a strong upmodulation of CDKN1A (p21) for both miR-191 and miR-425 (Figure 5D). In contrast, miR-425 over-expression specifically reduced expression of FSCN1, TNC and CDC42 (Figure 5D). Pathway analyses also revealed a repression of the PI3K-AKT pathway in miR-191/425 over-expressing cells. Western blot analyses against pERK1/2, pAKT and its direct targets pGSK3β confirmed the inhibition of PI3K-AKT signaling and highlighted that miR-191 is primarily responsible for the inhibition (Figure 5D). Moreover, we performed silencing of SATB1, CCND2 and FSCN1 in order to evaluate the specific contribution of each target to modulated miR-191/425 pathways. We found that only SATB1 knockdown, as well as miR-191 over-expression, were responsible for the up-modulation of β-catenin, whereas both CCND2 and FSCN1 silencing decreased β-catenin expression (Figure 5F). Finally, we found that SATB1 and CCND2 silencing controlled AKT pathway activation (Figure 5F). Taken together, these data indicate that miR-191/425 modify a number of genes that play critical roles in controlling the progression of highly invasive breast cancer.
Next, we assessed the in vitro biological effect of miR-191/425 on aggressive breast cancer cells. First, enforced expression of miR-191 or miR-425 in MDA-MB-231 and MDA-MB-436 cells induced an approximately 50% reduction in cell proliferation (Figure 6A and Figure S12A). Lentivirally-infected cells over-expressing either miR-191 or miR-425 were generated (Figure S12B), and cell proliferation was assessed using a (2D) colony formation assay (Figure 6B and Figure S12C). Cells over-expressing miR-191 not only showed a reduced number of colonies compared to control but also developed smaller colonies than control (Figure 6B and Figure S12D); in contrast, miR-425-expressing cells exhibited mainly a reduction in the number of colonies (Figure 6B and Figure S12D). Further, we tested the abilities of lentivirally-infected MDA-MB-231 cells to form colonies in soft agar. Compared to control cells, cells over-expressing either miR-191 or miR-425 formed significantly fewer colonies, indicating a decrease in anchorage-independent growth (Figure S12E). We then performed proliferation assays with cells cultured in three dimensions (3D) within Matrigel, and we observed that over-expression of either miR-191 or miR-425 impaired the formation of large filopodia/invadopodia-like structures at the periphery of the aggregates like in the control cells, thus resulting in the appearance of tightly adherent aggregates (Figure 6C). These results demonstrated that gain of cell adhesion and reduced migration are related to the degree of miR-191 and miR-425 expression in aggressive breast cancer cells. To more accurately quantify the anti-proliferative properties of miR-191/425 in aggressive breast cancer cells, flow cytometric analyses of transiently-transfected cells revealed fewer cells in S phase and an increased number of cells in G1 following over-expression of either miR-191 or miR-425 compared to scrambled transfected cells (Figure 6D and Figure S12F). To gain additional insight regarding the numbers of cells arrested in G1, we treated the cells with the microtubule-destabilizing agent nocodazole, which traps cycling cells in M phase. Cell populations with enforced miR-191 or miR-425 expression were characterized by significantly increased numbers of cells remaining in G1 (Figure S12G), confirming that both miRNAs caused cell-cycle arrest.
We next evaluated the in vivo effect of miR191/425 over-expression on tumor growth. First, we tested if over-expression of either miR-191 or miR-425 inhibits tumor growth of highly aggressive MDA-MB-231 cells. Lenti-miR-191 and lenti-miR-425 infected MDA-MB-231 were subcutaneously injected into the right flank of athymic nude mice and the tumor growth was monitored compared to control lenti-GFP infected and parental MDA-MB-231 cells. Tumors in the parental and GFP control groups were large, poorly differentiated, heavily necrotic and highly vascularized that formed within only 22 days post-implantation (5 out of 5 mice per group). In contrast, all five mice injected with either miR-191- or miR-425-infected cells exhibited greatly reduced tumor growth (Figure 6E). Interestingly, miR-191 and miR-425 over-expressing tumors were strictly non-invasive, as shown by their circumscribed profiles and confinement within dense fibrotic capsules (Figure 6F), in stark contrast to the spindle-like morphology of the parental and control tumors along with islands of cancer cells invading the fat pad and the muscle (Figure 6F and Figure S13A). Hence, ectopic expression of miR-191 and miR-425 in MDA-MB-231 cells impaired tumor growth and invasion in the surrounding tissue. To determine whether miR-191 and miR-425 expression in the primary tumors affects cell proliferation, we performed immunohistochemistry for the proliferation marker Ki-67. We found that the total number of Ki-67 positive cells in the tumors over-expressing miR-191 or miR-425 were significantly lower relative to the number observed in the control tumors (lenti-GFP control cells: 97.3%; lenti-miR-191: 81%; lenti-miR-425: 89%; p-value<0.05) (Figure 6F). High expression of miR-191 and miR-425 in the tumor cells was confirmed by qRT-PCR (Figure S13B). qRT-PCR revealed that miR-191 induced a reduction of mesenchymal (fibronectin) and acquisition of epithelial (e-cadherin and β-catenin) markers while miR-425 only a specific increase in e-cadherin (Figure S13C). Reduction of SATB1, CCND2 by miR-191 and FSCN1 by miR-425 over-expressing tumors was confirmed by western blot analyses (Figure 6G; and Figure S13D). Based on these numerous observations, we concluded that the impaired tumor growth of miR-191- or miR-425-over-expressing cells was a consequence of the reduced cell proliferation.
We then assessed the effects of miR-191/425 over-expression on migration and metastasis by using in vitro and in vivo experimental approaches. First, we evaluated the rate of cell migration by using the Boyden Chamber assay and found that miR-191- and miR-425-transfected cells migrated more slowly than control MDA-MB-231 cells (miR-191: p-value<0.05, ∼3-fold; miR-425: p-value<0.05, ∼6-fold) (Figure 7A). Further, we performed wound-healing assays on lenti-miR-191, lenti-miR-425 cells and GFP control (Figure 7B). By 16 hour post wounding, parental cells and GFP control cells migrated into the wound, resulting in 90% and 70% closure, respectively. In contrast, wound closure was significantly less in miR-191 and highly impaired in miR-425 (miR-191: 60% closed; miR-425: 25% closed) (Figure 7B). Migration and wound healing experiments were also performed using MDA-MB-436 cells, and the results were essentially similar (Figure S14A and S14B).
Finally, we tested the differential migratory abilities of miR-191 or 425-over-expressing cells by using an in vivo metastasis assay. Control lenti-GFP, lenti-miR-191, lenti-miR-425 infected-cells (2×10∧6 cells) were injected into the lateral tail vein of 6-week-old NOD-SCID mice, and their survival was evaluated in circulation, extravasation to and growth in lungs. After 8 weeks, histological analyses revealed that the number of micrometastasis was markedly reduced in the lungs of mice injected with miR-191 or miR-425 cells compared to the control tumor cells (Figure 7C). Of note, we also observed pneumonitis only in mice injected with the control GFP cells (Figure 7C). Collectively, all these data support the idea that sustained miR-191 and miR-425 activity impairs local invasion and metastatic colonization of breast cancer cells.
Defining the role of the differentially regulated miRNAs in breast cancer could lead to the development of new diagnostic tools and therapeutic approaches. In the present study, we provide new evidence for the role of miR-191 and miR-425 in breast cancer. We demonstrate that expression of miR-191 and miR-425 occurs as a part of the same transcriptional unit and strongly correlates with cellular ERα status. Moreover, we show that ERα directly regulates the expression of miR-191 and miR-425. Finally, our functional studies demonstrate that miR-191/425 cluster exerts a dual role in breast cancer cells depending on their ERα status: in ERα positive cells miR-191/425 work as oncogenes by inducing proliferation in part through the suppression of EGR1 during the E2 stimulation; in ERα negative cells, they impair tumor growth and invasiveness conferring a more epithelial phenotype to highly aggressive breast cancer cells.
We have demonstrated that miR-191 and miR-425 are co-expressed (Figure S4B) and, at least in part, transcriptionally dependent from the host gene DALRD3 in normal human tissues (Figure S2A, S2B). The identification of two distinct promoter regions responsible for the production of the two DALRD3 isoforms may allow the independent production of DALRD3 from the miRNAs and thus explain the partial correlation between miR-191/425 and DALRD3 found in some of the human tissues. Furthermore, the existence of the dual promoter for DALRD3 may contribute to “fine-tuning” of the estrogen-dependent regulation of miR-191/425 and DALRD3 gene transcription. We demonstrate that while E2/ERα signaling induces an increase in miR-191/425 expression ERα activation has a negative effect on the expression of the host gene DALRD3 (Figure 2B, Figure S5C and S5D, and Figure S6). qRT-PCR of the two different alternative splicing variants of DALRD3 showed that both variants are preferentially expressed in ERα positive cells and both reduced during E2 stimulation (Figure S3B and Figure S5C). These results highlight that E2 stimulation of the miR-191/425/DALRD3 transcriptional unit is essentially related to the production of miR-191 and miR-425. The reduction of the host gene isoform 1 may be explained with the mechanism proposed by Gromak et al. which showed that the cleavage of an intron can affect alternative splicing if it occurs between an alternatively spliced exon and its intronic regulatory elements [58]. Moreover, it has been demonstrated that ERα directly interacts with Drosha to modulate the processing of E2-regulated microRNAs [59]. In this scenario, we can hypothesize that the recruitment of ERα at the upstream promoter (Figure 2C) might improve the assembly of the Microprocessor complex at miR-191/425 locus and increase the cleavage of the intron for the production of the miRs, impairing the processing of the pre-mRNA. We further show that the increase of miR-191 and miR-425 upon E2 stimulation is associated with gradual reduction of polII accumulation on the downstream promoter (Figure 2C). Interestingly, this negative effect on DALRD3 promoter 2 is independent by ERα (silencing of ERα does not modify the downstream promoter activity), but is still related to E2 treatment, based on the strong reduction of promoter activity after E2 treatment (Figure 2E and Figure S6C). Both genomic and non-genomic estrogen actions may contribute to the regulation of miR-191/425-DALRD3 transcriptional unit [60], [61]: E2 treatment induces recruitment of ERα at the upstream promoter to improve only the accumulation of miR-191/425 (i.e., genomic regulation/processing activity), while estrogen-mediated effects, transmitted via enzymatic pathways or ion channels, induces repression of the downstream promoter (non-genomic regulation).
Next, we focused on the functional role of miR-191 and miR-425 in ERα signaling. Inhibition of miR-191 and miR-425 strikingly impairs cell proliferation and tumor formation in ERα positive cells (Figure 3A, 3D). Moreover, miR-19/425 overexpression in hormone deprived ERα positive cells, which have low levels of endogenous miR-191/425, reduces cell cycle arrest and apoptosis (Figure 3C). In silico analyses, based on the endonucleolytical activity of microRNAs, identify Early Growth Response 1 (EGR1) as a miR-191 target (Figure 3E, 3F and Figure S9A). EGR1 is involved in the regulation of cell growth and differentiation in response to signals, such as mitogens, growth factors, and stress stimuli [62], [63]. In most human tumors, such as breast cancer, fibrosarcoma, and glioblastoma, EGR1 is described to be a tumor suppressor gene [64]–[66]. In fact, re-expression of EGR1 in human tumor cells inhibits neoplastic transformation [63]. EGR1 represents also an important upstream gatekeeper of the p53 tumor suppressor pathway and many p53 downstream target genes, such as CDKN1A (p21), are dependent on EGR1 status. We demonstrate that during E2 stimulation, after an initial increase, the levels of EGR1 are repressed (Figure 3G and Figure S9B). Inhibition of miR-191 blocks the suppression of EGR1 and induces high levels of CDKN1A (p21) (Figure 3G, 3H) explaining at least in part the anti-proliferative activity of miR-191/425 cluster knockdown. However, the tumor-suppressive role of EGR1 seems to be tissue specific, because several studies implicated a tumor growth-promoting role of EGR1 in prostate cancer progression [67]–[69].
The loss of ERα expression causes tumor growth that is no longer under estrogen control, which leads to greater cancer aggressiveness and the failure of endocrine therapy. Therefore, restoration of ERα protein expression or signaling in ERα negative breast cancer cells represents an important key event to promote apoptosis and differentiation of aggressive breast cancer. Since miR-191 and miR-425 are players of the ERα signaling, we also inquire their role in ERα negative breast cancer. To this aim, we overexpressed both miRs in ERα negative cells and showed that miR-191 and miR-425 markedly alters the transcriptome of aggressive breast cancer cells, resulting in impaired tumor growth and metastasis (Figure 4 and Figure 5). Mechanistically, the effects of miR-191 and miR-425 on tumor growth and invasion require, at least in part, the suppression of SATB1, CCND2 and FSCN1. Specifically, miR-191-mediated SATB1 repression is associated with gain of epithelial markers (e.g., such as e-cadherin), and loss of mesenchymal markers (e.g., fibronectin and vimentin) (Figure 5C and Figure 6D). The increase of e-cadherin levels, mediated by miR-191/425, results in greater cell-cell adhesion, reduced detachment of cells, and cytoplasmic localization of β-catenin (Figure 5E). Mounting evidence indicates multiple reciprocal interactions of e-cadherin and cytoplasmic β-catenin with EMT-inducing transcriptional repressors to destabilize an invasive mesenchymal phenotype of epithelial tumor cells. Moreover, SATB1 and CCND2 repression by miR-191 are related to the suppression of the PI3K/AKT pathway and the corresponding reduced cell proliferation and tumor growth. We have also identified FSCN1, which is responsible for the reduced invasiveness and partial reversion to an epithelial morphology, as a target of miR-425 (Figure 6C).
All together our experiments demonstrate a duality in the biological role of miR-191/425 cluster in breast cancer: estrogen dependent-high levels of miR-191/425 induce proliferation in ERalpha positive cells by suppressing a strong tumor-suppressor gene, such as EGR1; low levels of miR-191/425 cluster are essential for the high expression of important modulators, such as SATB1, CCND2 and FSCN1, which confer a proliferative advantage to aggressive breast cancer cells.
Human breast cancer cell lines MCF10A, MCF10F, MCF7, T47D, BT474, BT483, ZR-75-1, MDA-MB361, HBL-100, SKBr3, MDA-MB-468, MDA-MB-453, BT549, MDA-MB-436, MDA-MB-231 as well as the Human Embryonic Kidney cell line HEK293, were purchased from the American Type Culture Collection (ATCC) and grown in accordance with ATCC recommendations. ERα, progesterone receptor (PGR) and HER2 status were confirmed for all cell lines by Western blot analyses. All transfections were carried out with Lipofectamine 2000 (Invitrogen, Carlsbad, CA) according to the manufacturer's instructions. For hormone depletion experiments, MCF7 cells were grown to 70% confluency in phenol red–free DMEM supplemented with 5% charcoal–dextran-stripped FBS for 6 days and collected every two days with the relative normal growth control. For estradiol (E2) treatments (Sigma Aldrich), MCF7 cells were hormone starved for 6 days and then treated with E2 (10 nM) at the indicated times. For Fulvestrant treatments, MCF7 cells were treated daily with fulvestrant (Sigma Aldrich) (100 nM) and collected at the reported time points. The 44 breast tumor tissue samples were provided from the Department of Pathology, The Ohio State University. All human tissues were obtained according to a protocol approved by the Ohio State Institutional Review Board.
Quantitative real-time PCR (qRT-PCR) was performed with the TaqMan PCR Kit (Applied Biosystems, Foster City, CA), followed by the detection with the Applied Biosystems 7900HT Sequence Detection System (P/N: 4329002, Applied Biosystems). PCR was carried out in 10 µL of reaction buffer containing 0.67 µL RT product, 1 µL TaqMan Universal PCR Master Mix (P/N: 4324018, Applied Biosystems), 0.2 mM TaqMan probe, 1.5 mM forward primer, and 0.7 mM reverse primer. The reaction mixture was incubated in a 96-well plate at 95°C for 10 minutes, followed by 40 cycles of denaturation (95°C for 15 seconds) and extension (60°C for 1 minute). All reactions were performed in triplicate. Simultaneous quantification of small endogenous nucleolar RNA U44/U48 was used as a reference for TaqMan assay data normalization. For quantification of DALRD3, trefoil factor 1 (TFF1/pS2), pri-miR-191, pri-miR-425, VEGFA, FSCN1, EGR1, TNC, CDC42, SATB1, SOX4, CCND1, VIM, CCND2, E2F1, SIAH2, API5, FIBR, CSDA, MYC, CTNN1 and CDH1 mRNAs, the appropriate TaqMan probes were purchased from Applied Biosystems. The TaqMan Array Human MicroRNA Card (Applied Biosystem) Set v3.0 is a two-card set containing a total of 384 TaqMan MicroRNA Assays per card that enables accurate quantification of 754 human miRNAs. Included on each array are three TaqMan MicroRNA Assays as endogenous controls to aid in data normalization and one TaqMan MicroRNA Assay not related to human as a negative control.
The hybridized Human Genome U133A 2.0 Array (Affymetrix) was scanned and analyzed with the Affymetrix Microarray Analysis Suite version 5.0. The average density of hybridization signals from three independent samples was used for data analysis, and genes with signal density less than 300 pixels were omitted from the analysis. P values were calculated with two-sided t-tests with unequal variance assumptions. To correct for multiple hypothesis testing, the false discovery rate was calculated. Differentially expressed genes were selected using both a false discovery rate of less than 0.01 and a fold-change greater than 1.5 or less than −1.5. A tree cluster was generated by hierarchical cluster analysis to classify the miR-transfected cells; for this analysis, we used average linkage metrics and centered Pearson correlation (Cluster 3.0). Java Treeview 1.1 (http://sourceforge.net/projects/jtreeview/) was used for tree visualization. The associations between gene modulations by two miRNAs were examined using a two-sided Fisher exact test. The association between modulations by any two miRNAs was statistically significant if P was less than .001. The online program Pathway-Express (http://vortex.cs.wayne.edu/Projects.html) was used to explore the most biologically relevant pathways affected by a list of input genes. Specific biological pathways were defined by the Kyoto Encyclopedia of Genes and Genomes database (Kanehisa Laboratories, Kyoto, Japan) (http://www.genome.jp/kegg/pathway.html). Pathways were considered statistically significant if the corrected gamma P was less than 0.01.
For cell-cycle analysis, MDA-MB-231 and MDA-MB-436 cells were plated in 6 cm dishes, transfected as indicated in the figures, trypsinized, washed in PBS, and fixed with ice-cold 70% ethanol while vortexing. Cells were rehydrated in PBS and stained 30 min at RT with propidium iodide (50 mg/ml PI, 0.5 mg/ml RNase in PBS) prior to flow-cytometric analysis. Lenti-GFP, lenti-191 and lenti-425 infected-cells were also analyzed by flow cytometry after 12 h treatment with nocodazole.
All mouse experiments were conducted following protocols approved by the institutional animal care and use committee at the Ohio State University. Parental MDA-MB-231, lenti-GFP, lenti-191 and lenti-425 infected-cells (5×106) were injected subcutaneously into the right flank of 6-week-old athymic nude mice. Tumor size was assessed twice per week using a digital caliper. Tumor volumes were determined by measuring the length (l) and the width (w) of the tumor and calculating the volume (V = lw2/2). Statistical significance between the control and treated mice was evaluated using Student's t test. We sacrifiedthe mice 35 days after injection and tumors were excided and processed for histology and for RNA and protein extractions. 4 µm sections of tumor tissues were stained with hematoxylin/eosin and with Ki-67 by immunohistochemistry. For MCF7 and ZR-75-1 xenografts, estradiol pellets (Innovative Research) were implanted in nude mouse and after two weeks mice were injected subcutaneously with one 10 cm plate of anti-miR191/425 transfected MCF7 or ZR-75-1 cells. Mouse experiments were conducted after approval by the institutional animal care and use committee at Ohio State University.
Transwell insert chambers with an 8-µm porous membrane (Greiner Bio One) were used for the assay. Cells were washed three times with PBS and added to the top chamber in serum-free medium. The bottom chamber was filled with medium containing 10% FBS. Cells were incubated for 24 h at 37°C in a 5% CO2 humidified incubator. To quantify migrating cells, cells in the top chamber were removed by using a cotton-tipped swab, and the migrated cells were fixed in PBS, 25% glutaraldehyde and stained with crystal violet stain, visualized under a phase-contrast microscope and photographed. Crystal-violet–stained cells were then solubilized in acetic acid and methanol (1∶1), and absorbance was measured at 595 nm. For the scratch assay, parental MDA-MB-231 cells, lenti-GFP, lenti-miR191 and lenti-miR425 infected-cells were plated in culture dishes and after 24 h the confluent monolayer was scratched. Images were acquired directly after scratching (0 h) and after 5 h, 9 h and 16 h. For quantification of migration distance Image J software was used. The distance covered was calculated by converting pixel to millimeters.
In situ hybridization (ISH) was carried out on deparaffinized human breast tissues using previously published protocol (Nuovo GJ, 2009), which includes a digestion in pepsin (1.3 mg/ml) for 30 minutes. The probes contained the dispersed locked nucleic acid (LNA) modified bases with digoxigenin conjugated to the 5′ end. The probe cocktail and tissue miRNA were co-denatured at 60°C for 5 minutes, followed by hybridization at 37°C overnight and a stringency wash in 0.2× SSC and 2% bovine serum albumin at 4°C for 10 minutes. The probe-target complex was seen due to the action of alkaline phosphatase on the chromogen nitroblue tetrazolium and bromochloroindolyl phosphate (NBT/BCIP). Negative controls included the use of a probe that should yield a negative result in such tissues (scrambled miRNA).
Total RNA isolation was performed with Trizol (Invitrogen, Carlsbad, CA) according to the manufacturer's instructions. For, acrylamide northern blotting 10 µg aliquots of total RNA were resolved on a 15% denaturing polyacrylamide gel (Bio-Rad, Hercules, CA) and were electrophoretically transferred to BrightStar blotting membrane (Ambion Inc, Austin, TX). The oligonucleotide encoding the complementary sequence of the mature miRNA annotated in the miRNA Registry (release 14: September 2009) was end-labeled with [γ32 P]-ATP by T4 polynucleotide kinase (USB, Cleveland, OH). RNA-blotted membrane was prehybridized in Ultrahyb Oligo solution (Ambion Inc) and subsequently hybridized in the same solution containing probe at a concentration of 106 cpm/mL at 37°C overnight. The membrane was washed at high stringency in the solution containing 2× standard saline citrate and 1% sodium dodecyl sulfate at 37°C. Northern hybridization signals were captured and converted to digital images with the Typhoon Scanner (GE Healthcare Biosciences, Piscataway, NJ).
Chromatin immunoprecipitation (ChIP) assays were performed with the ChIP assay kit (Upstate Biotechnology, Lake Placid, NY) with minor modifications. Briefly, MCF7 and MDA-MB-436 cells were hormone starved for 6 days and then treated with E2 (10 nM) for 3 h, 6 h and 24 h. The cross-linking was performed with 1% formaldehyde at 37°C for 10 minutes. Cells were then rinsed with ice-cold PBS and resuspended in 0.4 mL of lysis buffer containing 1% sodium dodecyl sulfate, 10 mM EDTA, 50 mM Tris–HCl, pH 8.1, 1× protease inhibitor cocktail (Roche Molecular Biochemicals), and sonicated. A 30 µL aliquot of the preparation was treated to reverse the cross-linking, deproteinized with proteinase K, extracted with phenol–chloroform, and the DNA concentration determined by Nanodrop 2000c (Thermo Scientific, Wilmington, DE) measurements. An aliquot of chromatin preparation containing 25 µg DNA was used per ChIP. The primary antibodies used for immunoprecipitation were rabbit polyclonal ERα (Bethyl Laboratories [Montgomery, TX] A300-498A), rabbit IgG control (Zymed, Carlsbad, CA), rabbit polyclonal acetyl-H3 (Upstate Biotechnology), rabbit polyclonal polIII (Upstate Biotechnology). ChIP-enriched DNA was subjected to SYBR green qPCR (Applied Biosystems). Primer sequences are listed in the Primer Table. Results were expressed as relative enrichment according to the following formula: 2−[(ctChIP−ctinput)−(ctIgG−ctinput)], where ctChIP, ctIgG, and ctinput indicate the cycle threshold for the specific antibody, IgG control, and input (5% of the total amount of immunoprecipitated material), respectively.
For miR-191 and -425 promoter prediction, a 9200 base pair (bp) DNA genomic region spanning miR-191 and-425 was used as input for the online software Promoter 2.0 (http://www.cbs.dtu.dk/services/promoter/).
To generate SATB1, CCN2, CTDSP2, SOX4, LRCC8A, SLC16A2, EGR1, CSDA, FSCN1, TNC, SIAH2 and CSDA luciferase reporter constructs, the 3′UTRs were amplified by polymerase chain reaction (PCR) and cloned downstream of the luciferase-coding sequence in the pGL3-control vector at the XbaI restriction site (Promega). Mutations were introduced into the miRNA-binding sites by using the QuikChange Mutagenesis Kit (Stratagene, La Jolla, CA). To map the miR-191-425 promoter, prom1 or prom2 genomic region (see schematic representation of miR-191/425-DARLD3 transcription unit, Figure 3, C) were amplified by PCR and cloned at the NheI and XhoI sites of the pGL3-basic vector (Promega). All constructs were sequenced to verify integrity.
To confirm that SATB1, CCN2, CTDSP2, SOX4, LRCC8A, SLC16A2, EGR1, CSDA, FSCN1, TNC, SIAH2, CSDA harbor responsive seed regions (complementary sequences) so that miR-191 and/or miR-425 can bind to their 3′UTRs, 250 ng of pGL3 reporter vector carrying the miR-191 or miR-425 binding site (see plasmid construct, Figure S9A), 25 ng of the phRL-SV40 control vector (Promega), and 100 nM miRNA precursors or scrambled sequence miRNA control (Ambion, Inc, Austin, TX) were cotransfected into HEK293 cells in 24-well plates. To map the miR-191 and miR-425 promoter, 250 ng of pGL3 reporter vector carrying prom1 or prom2 genomic region (see schematic representation of miR-191/425-DARLD3 transcription unit, Figure 3, C) and 25 ng of the phRL-SV40 control vector were cotransfected into HEK293 cells in 24-well plates. To asses estrogen responsiveness of the two promoter regions, same experiment was carried out in 5 breast cancer cell lines with different ERalpha status, in MCF7 cells after E2 (10 nM) treatment and in MCF7 after ERalpha silencing. Firefly luciferase activity was measured with a Dual Luciferase Assay Kit (Promega) 24 hours after transfection and normalized with a Renilla luciferase reference plasmid. Reporter assays were carried out in quadruplicate. Statistical significance was analyzed by the unpaired Student t test.
All cell lysates were prepared by using RadioImmuno Precipitation Assay Buffer (Pierce, Rockford, IL). Fifty micrograms of cell lysates was separated by sodium dodecyl sulfate–polyacrylamide gel electrophoresis and then electroblotted onto a polyvinylidene fluoride membrane (Hybond P; Amersham Biosciences, Piscataway, NJ). All primary antibodies used for western blot analyses are reported in Supplemental Materials and Methods (available online). Detection was performed with horseradish peroxidase–conjugated secondary antibodies (specific to rabbit and mouse) and enhanced chemiluminescence (Pierce).
Nuclear/Cytoplasmic differential protein extraction was performed by using the NE-PER Nuclear and Cytoplasmic extraction kit (Pierce) according to the manufacturer's instructions.
MDA-MB-231 cells were stably infected with the Human pre-microRNA Expression Construct Lenti-miR expression plasmid containing the full-length miR-191 or miR-425 and the GFP gene under the control of two different promoters (System Biosciences). An empty vector was used as control. Pre-miRs expression and control constructs were packaged with pPACKH1 Lentivector Packaging Plasmid mix (System Biosciences) in a 293TN packaging cell line. Viruses were concentrated using PEGit Virus Precipitation Solution, and titers were analyzed using the UltraRapid Lentiviral Titer Kit (System Biosciences). Infected cells were selected by FACS analysis (FACScalibur; BD Bioscience). Infection efficiency >90% was verified by fluorescent microscopy and confirmed by real-time PCR for miRs expression.
MDA-MB-231 cells, previously transfected with miR-191 or miR-425 precursors for 72 h, were plated (3000 per well) in 96-well plates and grown for 96 hours after transfection (final miRNA concentration of 100 nM) in normal culture conditions. MCF7 in normal culture conditions (+E2) transfected with anti-miR-191/425 and CTR oligonucleotide or in hormon deprivation conditions (−E2) transfected with miR-191/425 and CTR oligonucleotide were plated in 96-well plates and grown for 96 hours after transfection. Cell proliferation was documented every 24 hours for 4 days using a 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide assay kit (Promega, Madison, WI), and absorbance at 490 nm was evaluated by a SpectraMax 190 microplate reader (Molecular Devices, Sunnyvale, CA).
|
10.1371/journal.ppat.1004008 | Broadly Reactive Human CD8 T Cells that Recognize an Epitope Conserved between VZV, HSV and EBV | Human herpesviruses are important causes of potentially severe chronic infections for which T cells are believed to be necessary for control. In order to examine the role of virus-specific CD8 T cells against Varicella Zoster Virus (VZV), we generated a comprehensive panel of potential epitopes predicted in silico and screened for T cell responses in healthy VZV seropositive donors. We identified a dominant HLA-A*0201-restricted epitope in the VZV ribonucleotide reductase subunit 2 and used a tetramer to analyze the phenotype and function of epitope-specific CD8 T cells. Interestingly, CD8 T cells responding to this VZV epitope also recognized homologous epitopes, not only in the other α-herpesviruses, HSV-1 and HSV-2, but also the γ-herpesvirus, EBV. Responses against these epitopes did not depend on previous infection with the originating virus, thus indicating the cross-reactive nature of this T cell population. Between individuals, the cells demonstrated marked phenotypic heterogeneity. This was associated with differences in functional capacity related to increased inhibitory receptor expression (including PD-1) along with decreased expression of co-stimulatory molecules that potentially reflected their stimulation history. Vaccination with the live attenuated Zostavax vaccine did not efficiently stimulate a proliferative response in this epitope-specific population. Thus, we identified a human CD8 T cell epitope that is conserved in four clinically important herpesviruses but that was poorly boosted by the current adult VZV vaccine. We discuss the concept of a “pan-herpesvirus” vaccine that this discovery raises and the hurdles that may need to be overcome in order to achieve this.
| Human herpesviruses can cause a wide range of serious infections. They are extremely common and individuals remain latently infected lifelong, with reactivations often causing recurrent or severe disease. T-cells are important in controlling herpesvirus infections and preventing their reactivation, so vaccines that induce T-cells are likely to improve control. Here, we examined human T-cells against VZV that might allow focused vaccine development. We identified a dominant target against which the majority of subjects had mounted a CD8 T-cell response. We found that very similar targets also exist in three other important herpesviruses, HSV-1, HSV-2 and EBV. We showed that CD8 T-cells recognizing the VZV target could also recognize the others and we hypothesized that recurrent encounter with these viruses could boost this common response. In some individuals, immunization with a VZV vaccine did cause activation of these cells, but in most it did not. This reflects the variable efficacy of the currently available VZV vaccine. Our findings suggest that T-cell targets may be shared between herpesvirus species and may therefore contribute to a novel “pan-herpesvirus” vaccine. However, current VZV vaccines cannot reliably stimulate these T-cells and new strategies will be necessary to achieve this goal.
| The family Herpesviridae encompasses several highly prevalent human pathogens that cause a spectrum of diseases ranging from mildly symptomatic to severe life-threatening illness [1]. All herpesvirus subfamilies (α, β, and γ) share one important characteristic: the ability to evade the immune response while persisting as latent infections in a state of minimal gene transcription. In many individuals, latent herpesviruses cause no further disease. However, reactivations do occur that lead to considerable morbidity and mortality as well as promoting onward transmission. These events are most frequent in individuals with immunosuppression or immunosenescence [2]. However, asymptomatic reactivation can also occur in immunocompetent individuals, leading to recurrent stimulation of host immunity by herpesvirus antigens [3], [4]. T cells are essential both for recovery from primary herpesvirus infections and prevention of symptomatic reactivation [5]. VZV-specific T cells that secrete Th1 cytokines and exhibit cytolytic activity are detectable following chicken pox [6]. While virus exposure also induces antibodies, the absence of antibodies in children with agammaglobulinemia does not lead to more severe disease [7]. Conversely, the waning T cell immunity that occurs with older age is associated with greater frequency and severity of reactivations [8].
The only herpesvirus vaccines currently available are against VZV. This live attenuated vaccine prevents primary infection in children (i.e. chicken pox) and, when given at high dose, reduces the frequency and/or severity of shingles in elderly adults [9]. The vaccine induces both humoral and cell-mediated immunity [10]–[12], but vaccine-induced immunity can fail and effectiveness in the elderly is relatively poor [13]. The factors underlying this are poorly understood. Rational design of herpesvirus vaccines that elicit optimal protective T cell responses therefore remains an important goal.
However, in order to achieve this, further understanding of the role of human virus-specific T cells during herpesvirus infections is required. In this study, we aimed to comprehensively analyze the breadth of the CD8 T cell response to VZV in the context of the common HLA-A*0201 allele. The VZV genome is large, containing 69 unique open reading frames. This hampers the systematic identification of T cell epitopes and the generation of tools to study them. From VZV, only 7 class I-restricted epitopes from 3 proteins (gI, gB and IE62) have been reported thus far [14]. To address this, we used in silico prediction across the entire VZV proteome for epitope mapping. Screening of these candidate peptides in VZV seropositive individuals identified an immunodominant HLA-A*0201-restricted epitope that was conserved with three other herpesviruses. In this study, we characterized the phenotype and function of CD8 T cells that recognized this conserved epitope and also examined the responsiveness of these CD8 T cells to VZV vaccination in humans.
We recruited 21 HLA-A*0201 positive volunteers with a history of primary VZV infection, detectable VZV IgG but no previous VZV vaccination or clinical evidence of recent reactivation (Table 1). The median age was 63 years (range 25–77 years). Epitope predictions in the context of HLA-A*0201 were made using the Immune Epitope Database consensus prediction tool with the complete published VZV sequences (Table S1). The top 0.5% of 9- and 10-mers (367 peptides) were synthesized and peptide pools screened by IFN-γ ELISpot using PBMCs from each subject. In 15/21 subjects, the same peptide pool induced positive responses (Figure 1a). Deconvolution of this pool showed that two candidate peptides (ILIEGIFFV and MILIEGIFFV) from ribonucleotide reductase subunit 2 (RNR2) of VZV strongly induced IFN-γ production (Figure 1b). The predicted MHC-binding of the 9-mer (henceforth called ILI) had 11-fold higher affinity than the 10-mer (Table 2), so this was used to generate the A2-ILI tetramer used to label epitope-specific cells (Figure 1c). Tetramer+ cells were detected in 12/21 subjects with frequencies ranging from 0.01% to 1.8% of CD8 T cells (Figure 1d). ILI was thus identified as an immunodominant HLA-A*0201-restricted class I epitope.
To determine the phenotype of these ILI-specific CD8 T cells, we co-stained for memory subset markers; co-stimulatory molecules; and effector molecules. The majority of A2-ILI+ cells were CD45RA-/CCR7- indicative of an effector memory T cell phenotype (Figure 2a & 2b). However, between individuals, these cells displayed marked heterogeneity, which allowed further categorization into one of three phenotypic groups. Phenotype 1 (6/12 subjects) was the least effector-like, expressing high levels of the co-stimulatory receptors CD27 and CD28 with no expression of the cytotoxic molecules perforin or granzyme B; most A2-ILI+ cells of phenotype 2 (5/12 subjects) still expressed CD27 and CD28 and were still negative for perforin, but now expressed granzyme B; finally, phenotype 3 (1/12 subject) described the most effector-like cells with no CD27 and CD28 expression but high perforin and granzyme B (Figure 2a & 2b). We also investigated the expression of granzyme K, a serine protease that marks less differentiated CD8 T cells [15]. This also differed between groups and was inversely associated with granzyme B expression (in keeping with previous reports). In a subset of donors, we went on to examine the differentiation markers KLRG-1 and CD127 (Figure S1a). In all subjects tested, at least half the ILI-specific cells expressed KLRG-1 (a marker commonly expressed on terminally differentiated short-lived effectors), with a trend towards progressively higher expression in phenotypes 2 and 3. A variable proportion expressed CD127 (predominantly expressed on long-lived memory cells) but there was a trend towards more CD127+ cells in phenotype 1 and fewer in phenotype 3. These data were therefore consistent with those used to classify phenotypes 1, 2 and 3, suggesting that ILI-specific cells of phenotypes 2 and 3 were more likely to be terminally differentiated.
Primary VZV infection results in a multi-system disease that includes skin and neurotropic phases and antigen-specific cells may therefore need to localise to a variety of tissues. In most individuals, a major proportion of ILI-specific cells expressed the integrin CD62L, which allows homing to lymphoid organs (Figure S1b). In addition, a variable proportion expressed CCR5, indicating potential for homing to inflamed tissues. With neither marker was there a significant difference in frequency of expression between the 3 phenotypes. A2-ILI+ cells displayed no expression of cutaneous lymphocyte antigen (CLA) but a variable proportion did express the integrin α4β7. Again, there was no correlation with advancing phenotypic group.
One possible explanation for the heterogeneity of phenotype might have been recent or on-going activation, since the combination of markers characteristic of phenotype 3 would also be expected to occur in short-lived effector T cells. We therefore also analysed the expression of Ki-67 to determine whether any of these cells had undergone recent proliferation (Figure 2a & 2b). In a few subjects a minority of A2-ILI+ cells did express Ki-67. However, these never made up more than 5% of the population and there was no evidence that ILI-specific cells of phenotype 2 or 3 were more likely to have had recent proliferative activity. This was supported by analysis of the activation markers CD38 and HLA-DR, neither of which was up-regulated on ILI-specific CD8 T cells (Figure S1c).
In addition, the combinations of markers that segregated the phenotypic groupings did not change over time. Subjects were sampled at intervals between 1 and 6 months from baseline and the frequencies of ILI-specific cells expressing these combinations of markers remained stable (Figure 2c). These data therefore suggested that the ILI-specific memory T cell populations had been observed in a quiescent state and that, while they might express one of several stable phenotypes in any single subject, they were heterogeneous between individuals.
In view of their phenotypic differences, we proceeded to examine the functional capacity of ILI-specific CD8 T cells by measuring their ability to produce cytokines and undergo proliferation (Figure 3). Comparing the proportions of A2-ILI+ CD8 T cells capable of producing IFN-γ and IL-2, ILI-specific cells with phenotype 1 had the greatest cytokine producing capacity with a mean of 76% (range 37–100%) expressing IFN-γ (Figure 3b). As the phenotype changed from 1 to 2 and 3, the capacity of ILI-specific cells to produce IFN-γ fell. Furthermore, phenotype 1 ILI-specific cells were also the most polyfunctional, with a mean of 29% (range 22–51%) also staining for IL-2 (Figure 3c). Again, as the phenotype advanced, fewer ILI-specific cells produced this cytokine. In subjects in whom ILI-specific cells were at sufficiently high frequency for the assay, we then analyzed their in vitro proliferative capacity. This indicated that ILI-specific CD8 T cells from the individual displaying phenotype 3 were markedly impaired in their proliferation compared with those with phenotype 1 (Figure 3d). Thus the phenotypic and functional patterns displayed by the ILI-specific populations implied that they had been driven, to varying extents between individuals, towards more terminal differentiation with characteristics reminiscent of functional exhaustion.
To investigate the potential mechanism underlying this, we examined the association between differentiation phenotype and the expression of the inhibitory receptors PD-1 and 2B4, which are associated with exhaustion and up-regulated on virus-specific CD8 T cells during chronic antigen stimulation [16]. On ILI-specific cells, as the differentiation phenotype progressed, both PD-1 and 2B4 expression increased (Figure 4a). This was associated with a trend towards decreased capacity to produce cytokines such that as the frequency of cells expressing PD-1 and 2B4 increased, the frequency of IFN-γ producing cells fell (Figure 4b).
In chronic viral infections such as HIV, antigen-specific CD8 T cells are abundant, driven by continuous antigenic stimulation via the T cell receptor [17]. However, this increase in the frequency is balanced by increasing expression of inhibitory markers including PD-1, leading to functional exhaustion. Thus, although the frequency of memory T cells is increased, their functionality is restrained. Although herpesviruses do not continually produce antigenic proteins during latent infection, a strong correlation between the size of the population and the frequency of ILI-specific CD8 T cells that co-expressed both inhibitory receptors was seen (Figure 4c). These data imply that recurrent antigen exposure, for example via reactivation, may have driven the proliferation of these cells, thus increasing their frequency but also inducing the expression of inhibitory markers, which in turn affects their functional capacity.
RNR is one of a number of widely conserved proteins [18], [19]. We therefore hypothesized that the ILI epitope might be well conserved between the human herpesviruses. Indeed, we found that not only was the epitope present in all recorded VZV sequences but that there were also conserved homologues in the α-herpesviruses HSV-1 (ILIEGIFFA) and HSV-2 (ILIEGVFFA), and the γ-herpesvirus EBV (LLIEGIFFI)(Table 2). In contrast, poor homology was seen in the RNR2 of the γ-herpesvirus HHV-8, while the RNR2 gene is absent in the β-herpesviruses CMV, HHV-6 and HHV-7. Furthermore, the homologous epitope from the human RNR has little sequence identity with those of the herpesviruses and therefore unlikely to be responsible for any auto-reactive responses.
We tested the recognition of homologous peptides from VZV, HSV-1, HSV-2 and EBV by intracellular cytokine staining (Figure 5a & 5b). In all 11 donors tested, each peptide was capable of stimulating cytokine production. This occurred even when those individuals had no serological evidence of previous infection with the originating virus, with the response to the HSV-1 epitope equivalent to that against the one from VZV even in subjects who were HSV-1 seronegative (Figure 5b & Table 3). Conversely, the reduced response to the HSV-2 epitope occurred even in seropositive individuals. All subjects had been recruited on the basis of seropositivity to VZV and all but three volunteers also had evidence of previous EBV infection (Table 3). In contrast, only 11/21 subjects were positive for HSV-1 IgG and only 3 for HSV-2. Individuals with serological evidence of 3 or more herpesvirus infections were more likely to have detectable A2-ILI+ responses (9/12 subjects) while those with no detectable ILI-specific cells were more likely to have only EBV and/or VZV (6/9 subjects). These data therefore suggest that co-infection with more than two α- or γ-herpesviruses may increase the likelihood of generating a cross-reactive ILI-specific response.
In silico prediction suggested that residues at the anchor motifs of the VZV peptide (position 2 and the C-terminus) conferred optimal binding, while in vitro binding measurements showed that all four epitopes displayed extremely high affinities of <0.2 nM (Table 2). However, in vitro stimulation of CD8 T cells showed that the HSV-2 epitope was much less effective at stimulating cytokine production than those from VZV, HSV-1 and EBV (Figure 5a, 5b & 5c). Despite the calculated and measured binding affinities, peptide titrations showed that the epitopes from VZV, HSV-1 and EBV induced similar responses in any given individual while the peptide from HSV-2 only stimulated cytokines at its highest concentrations (Figure 5c). We therefore inferred that isoleucine at position 6, absent in the HSV-2 epitope, must be important for efficient TCR recognition.
These data support the hypothesis that herpesviruses are capable of inducing and boosting this epitope-specific response in a cross-reactive manner. Reactivations of one or several of these viruses may cause expansion of this population, increasing its size but also driving the cells towards an increasingly differentiated phenotype.
In order to examine the ability of the current VZV vaccine to boost ILI-specific responses, we immunized the study cohort with the live attenuated Zostavax vaccine and tracked the A2-ILI+ response. Following vaccination, the majority of subjects had no detectable change in the frequency of ILI-specific cells irrespective of their pre-vaccination frequency (Figure 6a). A greater than 2-fold increase in epitope-specific cells was seen in only one vaccinee (subject 105, phenotype 2). This individual had one of the lower starting frequencies at 0.05% and incremented to 0.17% at day 14 post-vaccination (Figure 6a & 6b). The increased ILI+ CD8 T cell frequency was associated with up-regulation of Ki-67 in 66% of epitope-specific cells, indicative of proliferation (Figure 6c & 6d). This occurred between 7 and 14 days post-vaccination, with Ki-67 completely down-regulated by day 28. In one additional subject (subject 126), there was evidence of minimal proliferation peaking at day 14 but no overall increase in the frequency of ILI-specific cells (Figure 6c). However, even in the best responder, the overall increase in the frequency of ILI-specific cells was modest despite the higher frequencies being maintained to day 28. Therefore live attenuated VZV vaccine only induced proliferation of ILI-specific cells in a single subject and even in the responding individual, the response was quantitatively poor.
Several T cell epitopes have previously been described that are conserved between HSV-1 and HSV-2 [20], including the epitopes described here [21]. However, it is interesting to find that these conserved epitopes exist more widely in viruses as divergent as the α- and γ-herpesviruses. Here, we have shown that this epitope is, in fact, broadly conserved between 4 different clinically important herpesvirus species despite their sequence divergence. Furthermore, all epitopes were capable of stimulating CD8 T cells even in individuals with no evidence of previous exposure to that virus. Earlier studies have indeed noted that HSV-specific T cells are detectable in a proportion of individuals seronegative for HSV. It was then proposed that these T cells might be induced by subclinical infection or exposure without infection, but our findings suggest that cross-reactivity of T cells may be an alternative explanation [22].
The frequency of ILI-specific CD8 T cells varied widely between individuals. We hypothesize that the number of herpesvirus infections and frequency of reactivations is responsible for these differences. Since herpesviruses are highly prevalent, multiple viruses invariably co-exist within a host and new herpesvirus infections may contribute to further stimulation of cross-reactive T cell populations. Furthermore, although herpesviruses downregulate their transcriptional machinery during latency, chronic expression of some viral genes still occurs [23], and subclinical reactivation has also been widely described. During stress, VZV can be detected in blood or saliva by PCR while HSV-2 often sheds in the absence of genital ulceration [4], [24]. Chronic or periodic antigen exposure may therefore boost T cell numbers over time.
However, in many experimental systems, increasing the number of antigen-specific T cells can also lead to immunopathology and there is increasing evidence in natural infections that this can be controlled by a number of feedback mechanisms. Under conditions of continuous antigen exposure in chronic infections such as HIV, antigen-specific T cells may be driven to proliferate to high frequencies and epitope-specific CD8 T cells identified by tetramer labelling are abundant [17]. However, constitutive expression of inhibitory receptors including PD-1 and 2B4 is also induced by chronic antigen stimulation, leading to reduction in further proliferative capacity and cytokine production [25]–[27]. In tandem with a decrease in co-stimulatory signalling via the down-regulation of CD27 and CD28, immunopathology is restrained despite the higher frequency of potential effector cells. Although continuous production of antigen does not occur in the same way during herpesvirus infections, recurrent antigen stimulation during reactivations may lead to a similar process.
In this context, our data suggest that the differentiation phenotype may act as a biomarker for frequency of reactivation. The phenotypic groups defined by stable expression of differentiation markers in the absence of recent proliferation (as evidenced by Ki-67 expression) may therefore be indicative of antigen exposure history. Furthermore, exhaustion may be one potential mechanism for the impaired T cell function that permits symptomatic reactivations in the elderly. If exhaustion could be reversed, there might be the possibility of enhancing antigen-specific immune responses in this population.
Although T cell immunity is believed to be essential for the control of herpesvirus infections, little is known about the role of T cells in the efficacy of the only currently available herpesvirus vaccines, Zostavax (which protects against shingles) and Varivax (which prevents chicken pox). Both are based on the live attenuated vOka strain of VZV, which has been passaged over 30 times in both human and animal cells. Despite the fact that Zostavax contains 14 times more virus than Varivax (which is administered to children), our data indicate that it does not efficiently stimulate a secondary ILI-specific response in seropositive adults. The explanation for this is likely to be multi-factorial. The numerous mutations that the vOka strain has acquired are likely to have contributed to poor replicative capacity and altered immunogenicity [28]. However, our data also suggest that VZV-specific CD8 T cells in many adults are intrinsically suboptimal, with a balance of co-stimulatory and inhibitory receptors that favors decreased responsiveness to antigen stimulation and impaired functionality. This may partially explain the incomplete protection provided against shingles and why VZV vaccine confers no cross-protection against other herpesviruses despite the presence of this cross-reactive CD8 T cell population in some individuals. Furthermore, VZV, HSV-1 and EBV have all been shown to evade host immunity by interfering with antigen presentation [29]–[31]. Therefore neither existing vaccines nor natural infection are therefore likely to induce cross-reactive T cell responses of sufficient magnitude to provide clinically relevant cross-protection. In addition, it is possible that conserved epitopes such as the ILI homolog are not equally processed and presented in all herpesvirus infections, even when the originating protein (e.g. RNR2) is expressed. The presentation of ILI homologs by HSV-1/2 and EBV in the absence of previous VZV infection was beyond the scope of our study and although CD8+ T cells specific for ILI homologs from HSV-1 and HSV-2 have been described, the VZV serostatus of donors in those studies was not assessed [21]. It therefore remains to be definitively shown whether the cross-reactive CD8+ T cell response demonstrated in vitro is effective in vivo.
However, in view of the evolutionary relatedness of the herpesvirus subfamilies, we hypothesize that there are still more conserved epitopes to be discovered. Vaccines that induce cross-reactive CD8 T cells might provide protection against clinical disease caused by multiple strains or even species of virus pathogens. The existence of such epitopes could open the way to the development of novel “pan-herpesvirus” vaccines if they can be made to induce responses of sufficient magnitude and functionality. Using the tools we have developed, further identification of the proteins from which similar conserved epitopes are derived may lead us to such a goal.
However, since even natural infection by virulent herpesviruses cannot adequately induce cross-protectivity, development of an effective pan-herpesvirus vaccine will require not only the identification of further cross-reactive epitopes across the major HLA supertypes but also completely new methods to specifically enhance the stimulation of cross-reactive CD8 T cells. In particular, we anticipate that novel adjuncts such as inhibitory receptor blockade will be required to tip the balance of signals that coordinate these responses in the direction of virus-specific T cell activation in order to overcome their relatively functionally impaired state. We anticipate that these strategies as well as advances in the rational design of improved immunogens will ultimately be necessary to achieve a vaccine that effectively protects against the wide array of diseases caused by herpesvirus infection.
All studies were approved by the Emory University Institutional Review Board (IRB #00050285). Study subjects provided written informed consent prior to participation. Clinical information is detailed in Table 1. Twenty-one healthy HLA-A*0201 +ve adults were recruited. Blood was obtained at baseline and multiple time points post-vaccination with the live attenuated Zostavax vaccine (Merck). Study subjects had a history of varicella zoster infection and serologic status against VZV was tested using the VZV IgG ELISAII (Wampole Laboratories, NJ, USA). HLA class I loci were genotyped using the sequence-base typing (SBT) method as recommended by the 13th International Histocompatibility Workshop (Tilanus et al. 2002).
The capacity of all VZV vOka (GI: 26665420, Acc. No. AB097932) derived 9- and 10-mer peptides to bind HLA A*02:01 was predicted using the command-line version of the consensus prediction tool available on the Immune Epitope Database (IEDB) web site (http://tools.immuneepitope.org/main/html/tcell_tools.html). Peptides were selected if they scored in the top 0.5% of predictions for each length. Additional peptides from non-vOka VZV strains (Table S1) were also included if they similarly scored in the top 0.5% of predictions. To assign gene names and locus tags to each peptide, genome sequences were run through an ORF finding algorithm (http://mobyle.pasteur.fr/cgi-bin/portal.py?form=getorf), and the corresponding data copied from the information available for orthologous Dumas proteins. MHC-epitope binding predictions were made with the Stabilized Matrix Method using the tool on the IEDB website. All peptides used in this study were synthesized by Mimotopes (Victoria, Australia) as crude material, and resuspended at 20 mg/ml in 100% DMSO (v/v).
Quantitative assays to measure the binding of peptides to HLA A*02:01 class I molecules are based on the inhibition of binding of a radiolabeled standard peptide (HBV core 18–27 analogue, FLPSDYFPSV). MHC molecules were purified by affinity chromatography from the EBV transformed homozygous cell line JY, and assays performed, as described previously. Peptides were tested at six different concentrations covering a 100,000-fold dose range in three or more independent assays, and the concentration of peptide yielding 50% inhibition of the binding of the radiolabeled probe peptide (IC50) was calculated. Under the conditions used, where [radiolabeled probe] < [MHC] and IC50 ≥ [MHC], the measured IC50 values are reasonable approximations of the true Kd values.
PBMCs were isolated using BD Vacutainer CPT tubes, washed, and resuspended in RPMI 1640 with 10% FCS (v/v) for immediate use or frozen in fetal calf serum with 10% dimethyl sulfoxide (v/v) for subsequent analysis. Plasma samples were saved at −80°C for subsequent analysis.
Gamma interferon (IFN-γ) enzyme-linked immunospot (ELISpot) assays were performed using 2×105 PBMC stimulated with peptide pools (10 µg/ml/peptide). Peptide pools yielding positive responses were deconvoluted, by testing individual peptides at 10 µg/ml. After 20 h of incubation at 37°C, plates were developed, and responses were calculated. Positive wells contained ≥20 spot-forming units (SFU)/106 cells and a P value of ≤0.05 using a Student's t test in at least 2 experiments.
MHC class I tetramer was prepared in-house. Surface staining of T cells was achieved by addition of tetramer to whole blood, incubation for 10 minutes at room temperature, followed by addition of antibody co-stains for 20 minutes. Whole blood was preferred to thawed PBMCs for tetramer labelling due to greater consistency and signal intensity. Following lysis of erythrocytes using BD FACS Lysing solution (BD Biosciences), cells were either fixed using 2% formaldehyde (v/v) or permeabilized using the BD Cytofix/Cytoperm kit for intra-cellular staining. The following antibodies were used for surface and intra-cellular staining: CD3-PerCP, CD8-Horizon V500, CD8-APCH7, Ki-67-FITC, Bcl-2-PE, HLA-DR-Horizon V450, HLA-DR-PerCP, Perforin-FITC, Granzyme B-Horizon V450, CCR7-PE, CD27-Horizon V450, CD27-FITC, CD28-PECy7, CD28-PE (all BD Biosciences) and CD38-PECy7 (eBioscience), Granzyme B-PE (Caltag), Granzyme K-PE (Santa Cruz), CD45RA-FITC (Beckman Coulter), PD-1-PE and 2B4-PerCPCy5.5 (Biolegend). Intra-cellular cytokine staining with IFN-γ-FITC, TNF-α-APC, and IL-2-PerCpCy5.5 (all BD Biosciences) was undertaken after in vitro stimulation of PBMCs using peptide for 6 hours. Flow cytometry analysis was performed BD LSRII and BD FACSCanto flow cytometers. Flow cytometry data were analyzed using FlowJo software.
PBMC were labeled for 7 minutes with 2.5 µM carboxyfluorescein succinimidyl ester (CFSE, Molecular Probes) in PBS at room temperature. Cold FCS was then added and cells were washed extensively with RPMI 1640 plus 10% FCS. CFSE-labeled cells were incubated with or without the ILIEGIFFV peptide (10 µg/ml) for 6 days. Responding CD8 T cells were subsequently identified by tetramer staining.
|
10.1371/journal.pntd.0005303 | High-Throughput Carbon Substrate Profiling of Mycobacterium ulcerans Suggests Potential Environmental Reservoirs | Mycobacterium ulcerans is a close derivative of Mycobacterium marinum and the agent of Buruli ulcer in some tropical countries. Epidemiological and environmental studies pointed towards stagnant water ecosystems as potential sources of M. ulcerans, yet the ultimate reservoirs remain elusive. We hypothesized that carbon substrate determination may help elucidating the spectrum of potential reservoirs.
In a first step, high-throughput phenotype microarray Biolog was used to profile carbon substrates in one M. marinum and five M. ulcerans strains. A total of 131/190 (69%) carbon substrates were metabolized by at least one M. ulcerans strain, including 28/190 (15%) carbon substrates metabolized by all five M. ulcerans strains of which 21 substrates were also metabolized by M. marinum. In a second step, 131 carbon substrates were investigated, through a bibliographical search, for their known environmental sources including plants, fruits and vegetables, bacteria, algae, fungi, nematodes, mollusks, mammals, insects and the inanimate environment. This analysis yielded significant association of M. ulcerans with bacteria (p = 0.000), fungi (p = 0.001), algae (p = 0.003) and mollusks (p = 0.007). In a third step, the Medline database was cross-searched for bacteria, fungi, mollusks and algae as potential sources of carbon substrates metabolized by all tested M. ulcerans; it indicated that 57% of M. ulcerans substrates were associated with bacteria, 18% with alga, 11% with mollusks and 7% with fungi.
This first report of high-throughput carbon substrate utilization by M. ulcerans would help designing media to isolate and grow this pathogen. Furthermore, the presented data suggest that potential M. ulcerans environmental reservoirs might be related to micro-habitats where bacteria, fungi, algae and mollusks are abundant. This should be followed by targeted investigations in Buruli ulcer endemic regions.
| Buruli ulcer is a neglected tropical disease which has been reported in over 33 countries, mainly located in tropical and subtropical regions. It is caused by Mycobacterium ulcerans, an environmental pathogen associated to slow-moving water. The sources and reservoirs of M. ulcerans remain elusive and are still to be discovered. In a first attempt to address this issue we used high-throughput carbon substrate profiling of M. ulcerans. The reported results show that some nutrients, naturally available in organisms present in M. ulcerans’ environment, are metabolized by this microorganism. This carbon substrate determination should help improve the culture of M. ulcerans as well as suggest potential environmental reservoirs in Buruli ulcer endemic regions.
| Mycobacterium ulcerans is the etiologic agent of Buruli ulcer, a disabling infection of the cutaneous and subcutaneous tissues [1–3]. M. ulcerans has been discovered in Bairnsdale, Australia, where Buruli ulcer was initially described [4,5]. Buruli ulcer is a World Health Organization notifiable infection and has been reported at least once by 33 countries located in the rural tropical regions of Africa and South America, in addition to Australia and Japan [6,7]. Over the past ten years, 83.6% (80.89–86.30) of cases were declared by eight West African countries [8]. In these highly endemic regions, the exact reservoirs of M. ulcerans remain elusive [6, 9–11]. However, epidemiological studies conducted in West African countries all indicated a significant association between the prevalence of Buruli ulcer and the contact of populations with stagnant water sources [12–17] through routine activities such as washing, swimming, fishing and farming [18,19]. A significant progress was recently made by narrowing the possible sources down to contacts with rice fields in Côte d’Ivoire which are sources of stagnant water [16,18,20,21]. Parallel environmental investigations of stagnant water [20,22], water insects [23–25], fishes [26,27] and aquatic mammals [12] showed the presence of PCR-amplified M. ulcerans insertion sequences (IS) IS2404, IS2606 and KR-B gene. Furthermore, M. ulcerans partial DNA coding sequences were also recovered from the soil in the vicinity of stagnant water [20,22,26,28,29]. This finding was strengthened by an experimental study confirming a four-month survival of M. ulcerans in soil [30]. M. ulcerans DNA has been also detected in water plants [28,31] and in Thryonhuomys swinderianus (agouti), a small mammal causing damages to rice fields and in close contacts with rural populations in West Africa [20].
Moreover, this compelling amount of information concerning the presence of M. ulcerans DNA-related sequences found in the environment has been strengthened by the isolation of five wild strains from those sources [3,32,33].
Here, we propose that a characterization of the metabolic profile of M. ulcerans may give clues to better define its natural environment including its environmental reservoirs. In this perspective, we used the Biolog Phenotype MicroArray (Biolog Inc., Hayward, CA) for high-throughput carbon substrate profiling of M. ulcerans. Indeed, Biolog Phenotype MicroArray was previously used to classify and characterize heterotrophic microbial communities from different natural habitats according to their sole-carbon-source utilization profiles [34]. Accordingly, this approach previously unraveled the phenotypic patterns of some Mycobacterium tuberculosis complex mycobacteria [35] and Mycobacterium avium subsp. paratuberculosis [36]. It is used here in the context of unique carbon metabolisms such as chitinase exhibited by M. ulcerans [37].
This experimental study investigated M. ulcerans strain CU001 (a gift from Pr V. Jarlier, Paris, France), a clinical isolate representative of the West African epidemic, M. ulcerans ATCC 19423 isolated in Australia, M. ulcerans ATCC 33728 isolated in Japan, M. ulcerans ATCC 25900 isolated in the USA and Mycobacterium buruli ATCC 25894 isolated in Uganda [38]. These strains were manipulated into a BLS3 laboratory and a clinical isolate of Mycobacterium marinum was isolated in our laboratory [39]. All strains were cultured at 30°C in Middlebrook 7H10 agar medium supplemented with 10% (v/v) oleic acid/albumin/dextrose/catalase (OADC) (Becton Dickinson, Sparks, MD, USA) and 0.5% (v/v) glycerol in a microaerophilic atmosphere for one week for M. marinum and four weeks for M. ulcerans.
The Biolog Phenotype MicroArray (Biolog Inc.), which consists of 96-well microtiter plates containing each a defined medium that incorporates a unique carbon source (plates PM1 and PM2A for 190 different carbon sources) plus a dye indicator of cell respiration was used, according to the previously reported standard Biolog Inc. protocol [40,41]. M. ulcerans and M. marinum colonies were removed from Middlebrook 7H10 medium using a cotton swab previously dipped in 0.1% Tween 80 (WGK Germany, Sigma Aldrich). Mycobacteria were taken with the wet swab off the agar plate culture by gently sweeping on the surface of the culture and then rubbed against the wall of a dry glass tube containing glass beads. The cells were then suspended in GN/GP-IF-0a (Biolog inoculating fluid n°133), the suspension was vigorously vortexed, passed three times through a 29-gauge needle in order to separate aggregates and adjusted to 81% transmittance using a turbidimeter (Biolog Inc). The PM-additive solutions for each plate were prepared according to Table 1. The inoculating fluid (Table 2) consisted of 20 mL of IF-0a GN/GP (1.2 x), 0.24 mL of dye mix G (100x) and 2.0 mL of PM additive (12x) added to the M. ulcerans or M. marinum suspension in IF-0a GN/GP (1.76 mL). Each PM plate was then inoculated in duplicate with 100 μL of inoculating fluid. The PM plates were incubated in the OmniLog PM System (Biolog Inc.) which measures the growth of mycobacteria every fifteen minutes for eight days at 30°C. In each well the substrate was reduced to a purple color which was directly proportional to the growth of the mycobacteria. The intensity of the purple color was recorded as dye reduction value, which was then plotted as area under the curve (AUC) by Biolog's parametric software. Negative control wells containing non-inoculated additive solutions in each PM1 and PM2 plates were run at the same time as a quality control element. The threshold separating the wells which exhibited a positive reaction from those with a negative reaction was set for each plate according to the value of the area under the curve (AUC) of the negative control Well (NCW). We defined moderately positive growing wells (MPW) and highly positive growing wells (HPW) as follows: MPW is when the AUC value of the well is equal to or lower than 1.25 times the AUC value of the negative control well, and HPW is when the AUC value of the well is equal to or higher than 1.50 times the AUC value of the negative control. PM plates were further examined visually at the end of each incubation period to ensure an independent verification of the results.
In order to find the potential environmental origin of the carbon substrates metabolized by M. ulcerans, we used the PubMed database to obtain information on the environmental sources for each of the 190 carbon substrates present in the PM1and PM2 plates. The environmental sources were organized in 10 categories (plants, fruits and vegetables, bacteria, algae, fungi, nematodes, mollusks, mammals, insects and the inanimate environment). The Chi-square test was used to compare the proportion of each category for substrates not metabolized by M. ulcerans versus substrates metabolized by all tested M. ulcerans strains; a P value < 0.05 was used as the criterion for statistical significance. We then used the PubMed database to match each substrate, used as a key-word, with all environmental sources significantly associated with substrates metabolized by all tested M. ulcerans strains, used as the second key-word (e.g., D-glucosamine and fungi). We calculated the number of hits obtained in this research and compared it to the number of hits obtained by searching only for the key word corresponding to the environmental sources (e.g., fungi).
The negative control wells remained negative in all the PMs plates, and results obtained with the five M. ulcerans strains and the M. marinum strain were duplicated. A total of 131/190 (69%) carbon substrates were metabolized by at least one of the five M. ulcerans strains, including 28/190 (15%) carbon substrates common to the five M. ulcerans strains and 16/190 (8%) carbon substrates metabolized by only one M. ulcerans strain (Table 3). A total of 21/28 (75%) substrates metabolized by all tested M. ulcerans strains were also metabolized by M. marinum (Table 3). In detail, 17/95 (18%) carbon sources in PM1 plates were metabolized by all M. ulcerans strains and comprised D-glucose-6-phosphate, D-ribose, L-asparagine, uridine, D-fructose-6-phosphate, adenosine, inosine, acetoacetic acid, methyl pyruvate, L-malic acid, D-psicose, L-lyxose, glucuronamide, pyruvic acid, L-galactonic acid-g-lactone, D-galacturonic acid and phenylethylamine. Six of these substrates exhibited a strong positive reaction (D-ribose, L-malic acid, L-lyxose, glucuronamide, pyruvic acid and D-galacturonic acid). Then, 11/95 (11.5%) carbon sources in PM2 plates metabolized by all M. ulcerans strains comprised D-raffinose, butyric acid, D-glucosamine, α-keto-valeric acid, 5-keto-D-gluconic acid, oxalomalic acid, sorbic acid, L-isoleucine, L-lysine, putrescine and dihydroxyacetone. Five of these substrates exhibited a strong positive reaction (D-glucosamine, 5-keto-D-gluconic acid, oxalomalic acid, sorbic acid and dihydroxyacetone). A total of 21/28 carbon substrates were also metabolized by M. marinum leaving D-galacturonic acid, uridine, methyl pyruvate, α-keto-valeric acid, L-isoleucine, L-lysine and putrescine as the only substrates specific to M. ulcerans (Table 3).
Comparing the potential environmental sources in search of substrates metabolized by all tested M. ulcerans strains versus non-metabolized substrates, we found a significant association between M. ulcerans metabolized substrates and bacteria (p = 0.000), fungi (p = 0.001), algae (p = 0.003) and mollusks (p = 0.007). The differences were not significant for plants (p = 0.535), fruits and vegetables (p = 0.870), mammals (p = 0.064), insects (p = 0.234) and the inanimate environment (p = 0.477). No carbon source was found to be associated with nematodes. Further MedLine research incorporating bacteria, fungi, algae and mollusks as keywords disclosed that 16/28 (57%) metabolized substrates were associated with bacteria, 5/28 (18%) were associated with alga, 3/28 (11%) were associated with mollusks and 2/28 with fungi. Discarding bacteria because of a potential bias since Biolog was designed for the study of bacterial metabolism, 15/28 (54%) metabolized substrates were associated with fungi whereas 6/28 (21%) were associated with the algae and 6/28 (21%) with mollusks (Table 4).
We determined that five different strains of M. ulcerans could use 28 different substrates as sources of carbon. These results were authenticated by the negativity of the negative controls introduced in every plate and the reproduction of data over two replicates. Moreover, stringent criteria were used to ensure the predictive value of the positive results. However, only seven of these 28 substrates were found to be specifically used by M. ulcerans and not by the phylogenetically closest species M. marinum. Three of these seven carbon sources indeed contain indispensable amino-acids.
The carbon sources here determined for M. ulcerans may by incorporated in culture media in the perspective of enhancing the isolation and culture of this pathogen. Indeed, M. ulcerans is a slow-growing mycobacterium and the availability of an improved method for its culture would improve the diagnosis of Buruli ulcer patients and the quest for environmental reservoirs [32]. As an example, it has been shown that the incorporation of chitin into the Middlebrook 7H9 broth enhances the growth of M. ulcerans [37]. Accordingly, our study points towards a possible association of M. ulcerans with fungi as a potential source of chitin, a polysaccharide possibly degraded by M. ulcerans’ genome-encoded chitinase [42]. Likewise, the other carbon sources here disclosed should be tested for their potential to increase the cultivation of M. ulcerans.
Moreover, our analyses suggested that M. ulcerans may have found some sources of carbon in microbial communities including alive and dead bacteria, fungi and algae. As for bacteria, it has been previously reported that M. ulcerans was isolated in environments where 17 other mycobacteria species were also isolated, including M. fortuitum as a constant co-inhabitant [3, 32, 33]. These results suggest cross-feeding between various bacterial complexes including mycobacteria, for the acquisition of carbon. Likewise, green algae extracts have been shown to halve the in vitro doubling time of M. ulcerans and promote the formation of biofilm [31]. We observed that M. ulcerans metabolizes D-galacturonic acid, the main component of pectin contained in the primary cell walls of terrestrial plants, and putrescine, a foul-smelling chemical derived from the decomposition of dead plants, which indicates that M. ulcerans may live in assemblages of dead aquatic plants. This finding is reinforced by the observation that M. ulcerans’s genome encodes five putative cutinases. Cutinases are mainly produced by phytopathogenic fungi to hydrolyze cutin (a main component of the cuticle which covers the aerial surfaces of plants) during plant colonization process [43].
Green algae are among the main food of freshwater mollusks pointed out in our study; mollusks are herbivores like other species of the freshwater snail family [44]. The principal genera of mollusks met in freshwater in West Africa are Bulinus, Planorbis, Pila, Lanistes, Melania, Bithynia, Lymnaea, Biomphalaria, Mutela, Aspatharia and Sphaerium [23,45]. Previous molecular investigations reported the detection of specific M. ulcerans DNA sequences in Bulinus spp. [23,46], in Planorbis spp. [23] and in mollusks of different Gastropoda order, Bivalvia order and Basommatophora order [26]. Furthermore, the experimental infection of Pomacea canaliculata (Ampullariidae) and Planorbis planorbis (Planorbidae) by plants contaminated by M. ulcerans- showed through optic microscopy digestive tract observation that snails remained infected by viable mycobacteria up to 25 days [23]. Small mollusks are also known to be a prey for water bugs which are involved in the transmission of M. ulcerans in Buruli ulcer endemic regions [3]. In West Africa, approximately 76% of the population lives next to rivers, lakes, and other water bodies contaminated with intermediate hosts such as snails [47].
In conclusion, our study is suggesting paths to improve culture media for the enhanced isolation of M. ulcerans by mimicking the natural ecosystem of M. ulcerans which is probably living in microbial communities with other bacteria, fungi and algae. These data support the recent hypothesis that mollusks could be part of a larger food chain including several hosts giving appropriate shelters to M. ulcerans, as recently reported [48]. Small mollusks should be further investigated using culture-based appropriate methods in the search for M. ulcerans.
|
10.1371/journal.pgen.1002473 | Heterochromatin Formation Promotes Longevity and Represses Ribosomal RNA Synthesis | Organismal aging is influenced by a multitude of intrinsic and extrinsic factors, and heterochromatin loss has been proposed to be one of the causes of aging. However, the role of heterochromatin in animal aging has been controversial. Here we show that heterochromatin formation prolongs lifespan and controls ribosomal RNA synthesis in Drosophila. Animals with decreased heterochromatin levels exhibit a dramatic shortening of lifespan, whereas increasing heterochromatin prolongs lifespan. The changes in lifespan are associated with changes in muscle integrity. Furthermore, we show that heterochromatin levels decrease with normal aging and that heterochromatin formation is essential for silencing rRNA transcription. Loss of epigenetic silencing and loss of stability of the rDNA locus have previously been implicated in aging of yeast. Taken together, these results suggest that epigenetic preservation of genome stability, especially at the rDNA locus, and repression of unnecessary rRNA synthesis, might be an evolutionarily conserved mechanism for prolonging lifespan.
| Aging is characterized by a progressive decline in vitality and tissue function, leading to the demise of the organism. Many models have been proposed to explain the aging phenomenon. Among the many competing and/or overlapping models is the heterochromatin loss model of aging, which posits that heterochromatin domains (which are set up early in embryogenesis) are gradually lost with aging, resulting in de-repression of silenced genes and aberrant gene expression patterns associated with old age. In this paper, we genetically tested the role of heterochromatin in Drosophila aging. We find that heterochromatin levels indeed affect animal lifespan and that heterochromatin represses, among other things, rRNA transcription. Loss of heterochromatin thus leads to an increase in rRNA transcription, a rate-limiting step in ribosome biogenesis and protein synthesis. We suggest that the biological functions of heterochromatin formation include controlling rRNA transcription, which might play an important role in general protein synthesis and animal longevity.
| Organismal aging is accompanied by the accumulation of damage to DNA and other macromolecules, and a progressive decline in vitality and tissue function. The underlying mechanisms remain unclear, and many models have been proposed to explain the aging phenomenon. Prominent among these models is the “free radical theory of aging”, which posits that the gradual and collective damage done to biological macromolecules (including DNA and proteins) by reactive oxygen species (ROS) from intrinsic (e.g., metabolism) or extrinsic sources (e.g., radiation), is the major cause of organismal aging [1], [2]. Other competing (although some are overlapping) models of aging include genetically programmed senescence [3], [4], heterochromatin loss [5], telomere shortening [6], genomic instability [7], nutritional intake and growth signaling [8]–[10], to name a few. In the heterochromatin loss model of aging, Villeponteau (1997) has proposed that heterochromatin domains, which are set up early in embryogenesis, are gradually lost with aging, resulting in derepression of silenced genes and aberrant gene expression patterns associated with old age [5].
Experimental tests of the role of heterochromatin formation in animal aging, however, have produced controversial results [11]. On the one hand, cellular senescence is associated with an increase in localized heterochromatin formation in the form of Senescence-Associated Heterochromatin Foci (SAHFs), which are a hallmark of replicative senescence of aged cells in culture, and have also been found in the skin cells of aged animals [12]–[14]. On the other hand, it has been shown that premature aging diseases in human and animal models correlate with global heterochromatin loss [15]–[17].
Heterochromatin is important for chromosomal packaging and segregation, and is thus important for genome stability [18], [19]. Indeed, it has been shown in Drosophila that heterochromatin is essential for maintaining the stability of repeated DNA sequences and of the rDNA locus in particular [20]. Loss of heterochromatin causes disruption of nucleolar morphology and formation of extrachromosomal circular (ECC) DNA, which results from an increase in illegitimate recombination at the rDNA locus [20]. Interestingly, disruption of heterochromatin and nucleolar structure, and the consequent increase in ECC DNA, have previously been shown to cause accelerated aging in yeast [21], [22]. These reports suggest a positive role for heterochromatin formation in promoting longevity.
To understand the role of heterochromatin in animal aging, and the underlying molecular mechanisms, we altered heterochromatin levels in Drosophila by genetically manipulating Heterochromatin Protein 1 (HP1) levels and JAK/STAT signaling, and assessed the effects on aging. Our results suggest that heterochromatin formation positively contributes to preventing premature aging and suppresses illegitimate recombination of the rDNA locus and unnecessary rRNA synthesis.
To investigate whether heterochromatin levels are important for longevity, we examined the life spans of flies with reduced or increased levels of HP1. These flies exhibit reduced or increased levels of heterochromatin, respectively, during development [23], as HP1 is an integral component of heterochromatin and controls heterochromatin levels [24], [25]. We found that reducing HP1 levels by half, as in Su(var)2055 heterozygotes, caused a dramatic shortening of life span compared to isogenic controls (p = 2.03−86) (Figure 1A). Similar results were found with a second allele, Su(var)2052 (Figure S1A). Conversely, a moderate overexpression of HP1, caused by basal activity of the hsp70 promoter, significantly extended life span, resulting in a 23% increase in median life span and a 12% increase in maximum life span (p = 6.31−24) (Figure 1A). Similarly, at non-heat shock conditions (25°C), a second (independent) line of hsp70-HP1 flies also lived significantly longer than their control flies (Figure S1B). At basal levels of transcription, hsp70-HP1/+ flies exhibited higher heterochromatin levels [26]. By quantitative real-time polymerase chain reaction (qPCR) measurements, we found that these flies had approximately 20% higher HP1 mRNA expression than control (Figure S2A). Over-expression of HP1 at higher levels, such as under heat-shock inducible conditions, however, caused developmental abnormality or lethality to the animal. These results suggest that heterochromatin levels significantly influence life span, and moderately higher levels of heterochromatin promote longevity.
Since both JAK overactivation and STAT loss reduce heterochromatin levels [26], [27], we investigated the effect of altering JAK/STAT signaling on aging. JAK/STAT signaling plays two roles: in the canonical pathway, JAK/STAT directly regulates target gene expression [28], [29], while in its non-canonical function, unphosphorylated STAT is essential for heterochromatin formation [26], [27] and genome stability [19]. In the canonical pathway, loss of STAT has effects equivalent to loss of JAK and opposite to JAK overactivation. However, in the non-canonical function, loss of STAT has the same effects as JAK overactivation, causing heterochromatin destabilization [26], [30] and genome instability [19]. We examined the life span of Stat92E+/− flies and those heterozygous for gain- or loss-of-function mutations of hop. We found that flies heterozygous for either the gain-of-function hopTum-l or the Stat92E mutation exhibited shortened lifespans compared with wild-type control flies (p = 8.87−23; 2.92−53, respectively), while flies heterozygous for a loss-of-function hop allele, hop3, had longer lifespans (p = 7.34−25) (Figure 1B). These results are consistent with the idea that heterochromatin levels influence lifespan.
A previous study has shown that Drosophila life span was only slightly reduced in Su(var)205 heterozygotes and was not affected by a chromosomal duplication that encompasses the Su(var)205 locus and many other genes [31]. Using chromosomal duplication, any effects associated with higher levels of Su(var)205 might be masked by higher levels of other neighboring genes. In studies with the loss-of-function heterozygotes, the authors in that study ensured isogenicity of their compared strains by extensively back-crossing the mutations into a common genetic background, and relied on the suppression of position-effect variegation (PEV) to determine the presence of the Su(var)205 mutation. PEV results from heterochromatin-mediated gene repression, commonly seen in loss of eye pigmentation [24]. The presence or absence of the Su(var)205 mutation was assumed to correlate 100% with the PEV phenotype. However, we have found that this is not the case. We examined the PEV phenotype of wm4 in the progeny of single pairs of wm4; Su(var)2055/CyO and w flies, and found that the PEV phenotype was not 100% correlated with the Su(var)205 mutation (Figure S3), rather, there was less PEV than would be expected, suggesting that, with regard to suppression of PEV, either the Su(var)205 mutation is not completely penetrant or that there is an incomplete epigenetic reprogramming at the wm4 locus, or both. On the other hand, it has been shown that many Su(var) mutations exhibit maternal-effect suppression of PEV [32], such that the PEV phenotype can be modified regardless of inheritance of the Su(var) mutation. It has also been shown that HP1 mutations disrupt epigenetic reprogramming, causing transgenerational inheritance of epigenetic information [33], [34]. Thus, in the aging study by Frankel and Rogina (2005) [31], the presence or absence of the Su(var)2052 mutation in the test flies may not have been accurately determined. In our current studies, we confirmed the presence of Su(var)205 mutations in the coisogenic strains by both suppression of PEV and homozygous lethality (see Materials and Methods). We found that heterochromatin levels are essential for longevity using both gain- and loss-of-function strategies.
To investigate the cause of altered life span in flies with different heterochromatin levels, we observed the behaviors of these flies by video-recording (see Methods). Video playbacks show that aged flies exhibited a gradual loss of mobility and eventually became immobile (Videos S1, S2, S3). By quantifying their mobility (see Methods), we found that, compared with wild-type controls, flies with reduced heterochromatin levels lost mobility much faster, and those with increased heterochromatin levels maintained their mobility for a longer period of time (Figure 1C; Videos S1, S2, S3).
It has been shown in C. elegans and Drosophila that old animals die of sarcopenia (muscle degeneration) [35] and impaired muscle function precedes aging [36], similar to the gradual loss of muscle function and frailty in aging humans. Since we found that heterochromatin levels influence Drosophila life span, and since the altered life span was associated with the animals' mobility, we investigated whether loss of heterochromatin is associated with muscle degeneration.
We used whole-mount fluorescent immunostaining to examine the integrity of the large intestinal wall muscle, which can be visualized readily in adult flies of different ages after minimal dissection. The fly large intestinal wall muscles consist of longitudinal (thick) and circular (thin) muscle fibers (Figure 1D, top left). We found that, wild-type flies exhibited progressive muscle degeneration as they aged (sarcopenia), such that the gut muscle fibers gradually showed breakage starting around day 20, and extensive breakage was seen in 40-day-old fly gut muscles. We found that heterochromatin levels affected the ability to maintain muscle integrity, with 20-day-old Su(var)205+/− flies showing extensive muscle fiber breakage (Figure 1D, top middle), whereas hsp70-HP1 flies maintained their muscle integrity beyond 40 days after eclosion (Figure 1D, top right). We quantified the breakages in longitudinal muscle fibers in a defined area of the midgut and calculated the muscle integrity index for each genotype and age (see Methods). We found that the muscle integrity indices correlate well with the mobility of flies of different genotype and age (Figure 1D, bottom). These results are consistent with the differences in fly motility that we directly observed. Thus, maintenance of heterochromatin levels is essential for the maintenance of muscle structure and function, which consequently affect animal mobility and lifespan.
If heterochromatin levels are important for longevity and tissue integrity, then normal aging should be accompanied by gradually decreasing heterochromatin levels. Indeed, it has been shown that normal aging in C. elegans, as well as the premature aging observed in human progeric syndromes, is correlated with changes in nuclear architecture and loss of heterochromatin [15]–[17], [37]. Since pericentromeric heterochromatin is readily observable in enterocytes, we examined HP1 foci in enterocytes of young and old adult flies. We found that, in contrast to young flies, whose enterocytes had prominent chromocenter enriched with HP1 (Figure 2A; top), old flies had much reduced levels of heterochromatin, with many nuclei in the gut epithelia lacking pronounced HP1 foci (Figure 2A; bottom). Since HP1 is recruited to heterochromatin by binding to histone H3 di- or tri-methylated at lys9 (H3K9m2 or H3K9m3), heterochromatin-specific chromatin modifications, we further investigated changes in the levels of H3K9m2 in flies of different ages. Interestingly, we found that total histone H3 levels decreased with age when compared with the non-histone nuclear protein HP1, which remained nearly constant relative to α-tubulin (Figure 2B). However, total levels of H3K9m2 showed a more dramatic decrease with age, and the decrease was obvious even relative to H3 levels (Figure 2B). Total H3K4m3 levels, on the other hand, showed a less dramatic decrease (Figure S4). Our results are consistent with previous reports that the levels of total histone H3 and heterochromatin marks decrease when animals age [15]–[17], [38], [39].
Taken together, the above observations suggest that, total histone H3 levels and their modifications by methylation, especially methylation of K9, exhibit gradual decline when animals age. The presence of excess HP1 throughout life might help preserving H3K9 methylation, thus delaying its decline. Although HP1 protein levels control heterochromatin levels during development, HP1 is not the sole factor determining heterochromatin formation post-development, especially in aged adult flies, where we have observed a decrease in the levels of H3K9m2, but not of HP1.
To further confirm that HP1 is not localized on heterochromatin sequences in old flies, we carried out chromatin immunoprecipitation (ChIP) experiments to determine HP1 occupancy on the transposable element 1360, which is highly enriched in constitutive heterochromatin [40]. Transposable element 1360 is present in >300 copies in the fly genome and has been used as a representative sequence for global heterochromatin [26], [40]. The ribosomal protein 49 (rp49) gene is a constitutively transcribed gene normally not associated with heterochromatin and can be used as a negative control. By performing ChIP experiments using anti-HP1 antibodies followed by PCR amplification, we assessed the levels of HP1 occupancy in these sequences. Indeed, we found that HP1 was found associated with 1360 in young but not old wild-type flies (Figure 2C), whereas in flies carrying the hsp70-HP1 transgene, HP1 was also found associated with 1360 even in old flies. These results are consistent with the idea that heterochromatin levels decrease with aging, and that over-expressing HP1 prevented heterochromatin decline. Taken together, these results suggest that there is a gradual loss of heterochromatin when flies age, as with C. elegans and humans, that the lack of HP1 localization to heterochromatin foci in old flies is likely due to the loss of H3K9 methylation, and that over-expression of HP1 throughout life can prevent or delay heterochromatin loss.
Heterochromatin loss could cause re-expression of genes that are normally repressed by heterochromatin. We thus examined the expression of a heterochromatinized lacZ transgene in young and old DX1 flies. These flies carry a tandem array of seven P[lac-w] transgenes, but lacZ (and white+) expression from these transgenes is normally repressed by DNA repeat-induced heterochromatin formation [41]. A reduction in heterochromatin can cause derepression of the lacZ gene contained in the P[lac-w] elements of DX1 flies [27], [41]. Indeed, we found that old, but not young, DX1 flies expressed lacZ (Figure 2D, 2E), consistent with the idea that heterochromatin levels decline with age. Thus, normal aging is accompanied by a gradual loss of heterochromatin in Drosophila as well.
We next investigated the possible mechanism(s) by which heterochromatin formation promotes life span extension. It has been shown previously that H3K9 methylation and RNA interference regulate nucleolar stability [20]. Loss of HP1 or Su(var)3–9 levels causes fragmentation of the nucleolus, as revealed by the nucleolar marker Fibrillarin [20]. When we examined the effects of heterochromatin levels on nucleolar morphology, which can be seen most easily in 3rd instar larval salivary gland giant nuclei, we found that conditions that decreased heterochromatin levels, such as JAK over-activation [27] or loss of STAT [26], were associated with nucleolar instability (Figure 3A). Conversely, conditions that increased heterochromatin formation, such as hop loss-of-function or HP1 over-expression [27], were associated with a stable nucleolus: the presence of a single, round nucleolus (Figure 3A). Moreover, HP1 over-expression suppressed the nucleolar fragmentation associated with hopTum-l (Figure 3A). These results are consistent with previous findings that JAK overactivation disrupts heterochromatin formation and that heterochromatin formation is important for nucleolar stability [20], [27].
Nucleolar fragmentation has been attributed to illegitimate recombination of repeated DNA sequences, resulting in instability of the highly repeated rDNA locus. Illegitimate recombination events can be assessed quantitatively by measuring the levels of extrachromosomal circular (ECC) DNA [20]. We isolated ECC DNA from flies of different genotypes and quantified ECC levels by calculating ECC index (see Methods). Indeed, we found increased ECC levels in mutants with decreased heterochromatin, such as hopTum-l, Stat92E, and Su(var)205, and decreased ECC levels in mutants with increased heterochromatin formation, such as hop+/− (Figure 3B).
Interestingly, increased ECC formation due to instability of the rDNA locus has previously been shown to cause accelerated aging in yeast [21], [22]. Taken together with our results from Drosophila, we suggest that the rDNA locus (or nucleolus) might be an important cellular target regulated by heterochromatin formation.
Finally, we investigated the functional consequences of nucleolar instability. The nucleolus is the site of rRNA biogenesis, where precursor rRNA molecules are transcribed and processed to give rise to 18S, 5.8S and 28S rRNAs (Figure 4A). In Drosophila, as well as in mammals, the rDNA locus consists of a few hundred rRNA transcriptional units in tandem repeats. The number of rDNA genes vastly exceeds what is needed for adequate rRNA transcription. Normally only 20 to 25 units (<10% of the total) are actively transcribed, while the majority of the rDNA locus is silenced presumably by unknown epigenetic mechanisms [42]. Since it has been shown that loss of heterochromatin, as in Su(var)205 transheterozygotes, leads to illegitimate recombination and thus instability of the rDNA locus [20], we investigated whether heterochromatin loss also leads to derepression of rDNA transcription.
The levels of rDNA transcription can be more sensitively detected by examining the transcription of a class of transposons (e.g., R2 elements) that are specifically inserted into, and are cotranscribed with, the 28S rDNA gene [42]. Normally the host “selects” a region of the rDNA locus free of R2 insertions for transcription and represses the R2-inserted rDNA units by unknown mechanisms [42]. We found that in Su(var)205 transheterozygous mutants, however, transcription of R2 elements was dramatically increased by >40 fold compared to their sibling heterozygous control flies (Figure 4B). This suggests that in preserving the structural integrity of the nucleolus, heterochromatin formation plays a crucial role in silencing the transcription of the majority of rDNA genes, and that loss of heterochromatin causes a dramatic increase in rRNA transcription, which could lead to an increased capacity for protein synthesis, conducive to growth and accelerated aging.
To investigate whether the moderately altered heterochromatin levels that have been shown to alter lifespan (Figure 1), affect the rate of rRNA synthesis, we measured pre-rRNA transcript levels in flies heterozygous for Su(var)205 or carrying an hsp70-HP1 transgene by quantitative real-time PCR (qRT-PCR). Indeed, we found that flies in which HP1 is moderately over-expressed (by basal activity of the hsp70 promoter without heat shock) contained >50% less pre-rRNA, whereas Su(var)205 heterozygous flies had levels of pre-rRNA transcripts that were >2 fold higher (Figure 4C).
To determine whether these altered rRNA transcription rates affect growth, and thus the body size of the adult fly, which has often been inversely associated with lifespan [43], we measured the body size and weight of larval and adult flies with altered heterochromatin levels as mentioned above. Indeed, we found that flies moderately over-expressing HP1 had a smaller body weight, and that Su(var)205−/− larvae had a larger body size (length) (Figure 4D), and that Su(var)205+/− flies had a larger body weight (Figure 4E). The differences in body weight were not as pronounced as those in rRNA transcription, suggesting that body size may not be solely regulated by the rRNA transcription rate. Nonetheless, these results are consistent with the idea that changes in the rate of rRNA transcription may impact global protein synthesis and thus the growth of the organism.
In summary, we have found that heterochromatin formation promotes longevity, genome stability, and suppresses rRNA transcription. The causal relationship between aging and rRNA transcription, however, awaits further investigation. It is interesting to note that factors that promote growth, such as insulin signaling and protein synthesis, usually accelerate aging, whereas inhibition of these pathways extends life span [9], [44]–[49]. Moreover, it has been shown in yeast that the Sir2 histone deacetylase counteracts aging by inducing heterochromatin formation at the rDNA locus [21], [22], thereby suppressing rRNA transcription and maintaining stability of the rDNA locus. In mammals, it has been shown that heterochromatin and Sirt1 epigenetically silence rDNA transcription in response to intracellular energy status [50]. Thus, loss of rDNA silencing due to heterochromatin loss could lead to instability and increased rRNA transcription, which promotes protein synthesis in general. We suggest that suppression of rDNA transcription might be an evolutionarily conserved mechanism essential for animal longevity.
All crosses were carried out at 25°C on standard cornmeal/agar medium unless otherwise specified. Fly stocks of hopTum-l, Stat92E06346, Su(var)20505, Su(var)20502, hop3, hsp70-Gal4, and UAS-eGFP were from the Bloomington Drosophila Stock Center (Bloomington, IN). Fly stocks of DX1 and 6-2 mini-white+ (J. Birchler), and hsp70-HP1 (G. Reuter; L. Wallrath) were generous gifts. All alleles used for life span analyses were extensively outcrossed before experiments (see below).
The following outcrossing schemes were used to minimize genetic background effects. hsp70-HP1 (line 1; p[hsp70-HP1-eGFP, ry+] carried on the 2nd chromosome; [51]) flies were outcrossed to a ry506 stock for ten generations, and the ry+ or CyO marker was followed to derive outcrossed hsp70-HP1/+ and CyO/+ flies, respectively. These flies were crossed to establish new “outcrossed” hsp70-HP1 (p[ry+])/CyO; ry506 flies. Su(var)20505/CyO flies were outcrossed to In(1)wm4 stock for ten generations, and the CyO marker or suppression of In(1)wm4 PEV was followed to derive outcrossed In(1)wm4; Su(var)20505/+ and In(1)wm4; CyO/+ flies, respectively. These flies were crossed in single pairs to derive new “outcrossed” In(1)wm4; Su(var)20505/CyO stocks (the presence of Su(var)20505 was confirmed by both suppression of PEV and homozygous lethality). For lifespan analysis, the “outcrossed” hsp70-HP1 (p[ry+])/CyO; ry506 flies were crossed to In(1)wm4 flies, and the “outcrossed” In(1)wm4; Su(var)20505/CyO flies were crossed to ry506 flies, and the F1 non-CyO flies were collected for lifespan analysis. “Wild-type” control flies were the F1 of In(1)wm4 and ry506 flies. An independent stock of hsp70-HP1 (line 2; an unmarked p[hsp70-HP1-lacI] inserted in the 2nd chromosome; [52]) was outcrossed to In(1)wm4 stock (with a CyO chromosome “floating”) for ten generations. An isogenic line of In(1)wm4; hsp70-HP1 (line 2)/CyO flies was established from a single male, and the presence of hsp70-HP1 was confirmed by its strong enhancement of PEV. This stock was used for lifespan studies, and a line that did not enhance PEV (which was considered not carrying the hsp70-HP1 transgene) was used as wild-type control.
To assess life span, 2-day-old females were separated from males and were transferred to fresh vials at 20 flies/vial, and were subsequently transferred to fresh vials every 2–3 days. Dead flies were counted upon each transfer. Only female heterozygotes were analyzed because hop is located on the X chromosome and the hemizygous mutants are not viable. Flies were cultured at 25°C and 70% humidity.
Mouse monoclonal anti-HP1 (C1A9; Developmental Hybridoma Bank, Iowa; 1∶200), rabbit anti-H3(di)mK9 (07-212; Upstate Biotechnology; 1∶200), rabbit anti-GFP (CloneTech; ), and rabbit anti-Fibrillarin (Abcam; 1∶500) were used as primary antibodies and fluorescent secondary antibodies (Molecular Probes) were used in whole-mount immunostaining. Tissues were fixed in 4% paraformaldehyde/PBS and 0.3% Triton-X/PBS. Stained tissues were photographed with a Leica confocal microscope. Images were cropped and minimally processed using Adobe Photoshop CS.
For chromatin immunoprecipitation (ChIP), adult flies of appropriate ages were snap frozen with liquid nitrogen, and then were cross-linked with 1.8% formaldehyde. The flies were homogenized in cell lysis buffer, and ChIP was performed as previously described [26].
Adult flies of appropriate genotypes were split into two groups, one for Hirt ECC DNA isolation as described [20], and the other for genomic DNA isolation by standard protocols to use as controls. Typically 10 adult males were ground in 500 µl Hirt lysus buffer (0.6% SDS, 10 mM EDTA, pH 8.0) for ECC isolation or in 200 µl Buffer A (0.5% SDS, 100 mM EDTA, 100 mM NaCl, 100 mM Tris-HCl, pH 7.5) for genomic DNA isolation. Genomic DNA was quantified by spectrometry, and 100 ng of genomic DNA and an equal volume of ECC DNA were used for PCR amplification. Primer sequences used for amplifying ECC and control DNA were as previously reported [20]. The PCR products from ECC sample and genomic control sample were loaded on the same agarose gel. PCR bands were revealed by ethidium bromide staining and photographed. The level of ECC was measured as the ratio of ECC band intensity to that from genomic control run on the same gel. Three independent experiments were done for each genotype. The ECC index was calculated as the sum of each ECC/genomic ratio divided by the total number of ECC bands:Where N denotes the total number of ECC species examined (N = 8 in this experiment).
Flies were outcrossed as described in “Life span analysis” to minimize genetic background effects. Virgin female flies of a particular genotype were grouped by 5 in a cornmeal food vial (without supplementing yeast) and were passed daily to a fresh vial. Flies of different ages were video recorded for 2 min, around 9:30 AM, with a mounted digital camera. Recording was started after flies fell to the bottom of the vial by knocking the vial on a bench top. Motility scores were assigned to each fly in the video playback according to the speed with which it moved upward in the vial.
Adult flies of desired ages were dissected and fixed with formaldehyde. The intestines were stained with phalloidin-fluorescein and observed with an epifluorescence microscope. At least 10 flies of each genotype and age were dissected. Each fluorescein-stained large intestine was assigned a Morphology Score (MS) of 0 to 10 based on the number of breakages in the longitudinal muscle fibers in a 3 gut-diameter long stretch of the midgut immediately adjacent to the hindgut. MS = 10−n, where n = the number of breakages in the defined region of the midgut. A score 10 represents prefect morphology: no muscle fiber breakage in the longitudinal muscles (n = 0). When n≥10, usually no continuous longitudinal muscle fibers can be identified and accurate counting breakages becomes difficult. In this case, a Morphology Score of 0 was assigned. So MS represents the worst morphology. The muscle Integrity Index is defined as the total MS of each genotype and age divided by the total number of intestines observed (N).
Total RNA was isolated from 10 adult male flies (2-day-old) of desired genotypes using the RNeasy Mini Kit (Qiagen) or trizol (Invitrogen) according to the manufacturer's instructions. One µg of total RNA and primers specific for pre-rRNA and rp49 (control) were used to make the first strand cDNA using Superscript III reverse transcriptase (Invitrogen) in 50 µl total reaction volume. The cDNA (at 1∶100 dilution) was used as template for qRT-PCR analysis using SYBR green based detection on a BioRad iCycler. Reactions were carried out in triplicate, and melting curves were examined to ensure the presence of single products. The levels of pre-rRNA were quantified relative to rp49 transcript levels (control) and were normalized to a wild-type control. The following primer pairs (forward and reverse) were used.
rp49: tcctaccagcttcaagatgac, cacgttgtgcaccaggaact
pre-rRNA (5′ETS): atcggccgtattcgaatggattta, ctactggcaggatcaaccaga
|
10.1371/journal.pntd.0004376 | Population Genetics of Plasmodium vivax in the Peruvian Amazon | Characterizing the parasite dynamics and population structure provides useful information to understand the dynamic of transmission and to better target control interventions. Despite considerable efforts for its control, vivax malaria remains a major health problem in Peru. In this study, we have explored the population genetics of Plasmodium vivax isolates from Iquitos, the main city in the Peruvian Amazon, and 25 neighbouring peri-urban as well as rural villages along the Iquitos-Nauta Road.
From April to December 2008, 292 P. vivax isolates were collected and successfully genotyped using 14 neutral microsatellites. Analysis of the molecular data revealed a similar proportion of monoclonal and polyclonal infections in urban areas, while in rural areas monoclonal infections were predominant (p = 0.002). Multiplicity of infection was higher in urban (MOI = 1.5–2) compared to rural areas (MOI = 1) (p = 0.003). The level of genetic diversity was similar in all areas (He = 0.66–0.76, p = 0.32) though genetic differentiation between areas was substantial (PHIPT = 0.17, p<0.0001). Principal coordinate analysis showed a marked differentiation between parasites from urban and rural areas. Linkage disequilibrium was detected in all the areas (IAs = 0.08–0.49, for all p<0.0001). Gene flow among the areas was stablished through Bayesian analysis of migration models. Recent bottleneck events were detected in 4 areas and a recent parasite expansion in one of the isolated areas. In total, 87 unique haplotypes grouped in 2 or 3 genetic clusters described a sub-structured parasite population.
Our study shows a sub-structured parasite population with clonal propagation, with most of its components recently affected by bottleneck events. Iquitos city is the main source of parasite spreading for all the peripheral study areas. The routes of transmission and gene flow and the reduction of the parasite population described are important from the public health perspective as well for the formulation of future control policies.
| We present the population genetics of malaria vivax parasites in a large area of the Peruvian Amazon. Our results showed that the parasite population had a predominant clonal propagation, reproducing themselves with identically or closely related parasites; therefore, the same genetic characteristics are maintained in the offspring. The clonal propagation may favour the higher levels of genetic differentiation among the parasites from isolated areas compared to areas where human migration is common. The patterns of gene flow have been established, finding Iquitos city as a reservoir of parasite genetic variability. Moreover, a recent reduction of the parasite population was observed in areas where recent control activities were performed. This research provides a picture of the nature and dynamics of the parasite population which have a significant impact in the malaria epidemiology; therefore, this knowledge is crucial for the development of efficient control policies.
| According to the World Health Organization (WHO), Plasmodium vivax caused about 14.2 million malaria cases outside sub-Saharan Africa in 2013 [1]. Despite considerable efforts, Asian and South American countries are still far from achieving malaria elimination [2]. In Peru, the vast majority of malaria cases (76% of 64,673) was reported in the Amazon basin area (Loreto region) for 2014 and about 83% of them are due to P. vivax [3]. Many P. vivax infections are asymptomatic and undetectable by microscopy, providing a potentially important reservoir sustaining local transmission [4–7]. In addition, multiple infections recur even after the administration of the WHO-recommended radical cure treatment against blood- and hepatic-parasite stages (chloroquine and primaquine) [5, 6, 8]. To understand the epidemiology, distribution and transmission dynamics of P. vivax and thus improve its control, it is necessary to unravel the parasite population genetics and dynamics [9, 10]. Such information would be extremely useful for the monitoring and evaluation of control activities, both in the short and long term [10–12].
The extreme genetic variations in the P. vivax populations has already been reported from several endemic areas [10, 13]. In the Peruvian Amazon, the few observations available on P. vivax population genetics were collected in small areas (dispersed villages and communities) and reported heterogeneous and clonal parasite populations [6, 14–16]. Hereby, we report the genetic diversity and population genetics of the P. vivax parasite population from the most important urban city in the Peruvian Amazon and 25 villages located around and along the Iquitos-Nauta road.
The clinical isolates were collected during an initial screening of the study sites in the Peruvian Amazon (April 2008) followed by active case detection of all fever cases (April to December 2008) during a longitudinal study assessing the efficacy of the recommended radical cure treatment for P. vivax malaria infection (chloroquine 25 mg/kg/day for 3 days and primaquine 0.5 mg//kg/day for 7 days) [17]. The study was conducted both within Iquitos city and in 25 neighbouring villages (Loreto region), some of them along the Iquitos-Nauta Road. Villages were geographically stratified in 5 study areas (A1- A5) (Fig 1A and S1 Table): A1 (3–7 km northwest from Iquitos city and only accessible by boat), A2 (Iquitos city and peripheral villages), A3 (villages situated along the Iquitos-Nauta Road, 11–13 km southwest from Iquitos city), A4 (villages situated 21 km southwest of Iquitos city and 2–9 km far from the Iquitos-Nauta Road) and A5 (villages situated 26–58 km southwest from Iquitos city, most of them along the Iquitos-Nauta road). Participants’ demographics are described in S1 Table.
In Fig 1B, we describe the main human mobility patterns in the study areas. Most human mobility occurs around A2 (Iquitos city, economic centre and big markets). Indeed, some people commute every day to A2 but live in A1 or A3, others travel from A5 to A2 during the weekends crossing the Nanay river or come from A4 a couple of times per year navigating along the Itaya river to sell their goods. Some sporadic movement between A1 and A5, crossing the Nanay river, is observed.
The study population consisted primarily of mestizos of low socioeconomic status. Villages located along Iquitos-Nauta road or next to the river had no electricity and in most cases drinking water was taken from the river or natural springs. Malaria transmission is perennial with peaks from November to May (rainy season)[18], and the majority of malaria cases are due to P. vivax. Recurrent sub-patent and asymptomatic infections are frequent [5, 6]. Anopheles darlingi is the main anthropophilic and exo/endophilic vector [19]. Malaria prevention and control activities were conducted before the sample collection in Loreto region including all our study areas but in A4 (S1 Table) as part of PAMAFRO (Malaria Control Program in Andean-country Border Regions) which started in 2005 [20].
Patients were examined daily during the treatment and followed up weekly with blood sampling according to WHO guidelines for P. vivax drug efficacy during the first 28 days [WHO, 2009]; and systematic monthly follow-up was carried out thereafter. For the purpose of the current study, all D0 (before-) and D1 (after treatment) isolates were analysed, including a blood sample for microscopy (thick and thick film) and a blood spot on filter paper (BSFP) (Whatman grade 3, Whatman, Springfield Mill, USA).
The study was approved by the Ethical Boards of Universidad Peruana Cayetano Heredia, Peru (Project PVIVAX-UPCH, SIDISI code: 053256), the institutional review board of the Institute of Tropical Medicine Antwerp and the University Hospital of Antwerp, Belgium. Adult participants provided informed written consent, and a parent or guardian of any child participant provided informed written consent on their behalf.
All slides were read by microscopy (thin and thick smear) to confirm P. vivax infection and estimate the parasite density (number of asexual parasites for 200 white blood cells (WBC) assuming 8000 WBC/μl) [6, 21]. Parasite and human DNA was extracted using the Chelex method [22]. P. vivax mono-infections were confirmed by species-specific PCR (ssPCR) [23] and genotyped using a panel of 14 well-described microsatellite markers (MS), namely MS1, MS2, MS3, MS4, MS5, MS6, MS7, MS8, MS9, MS10, MS12, MS15, MS16 and MS20 [24, 25]. Briefly, the DNA extracted from each sample was used to perform a separate PCR for each MS. The forward primer of each MS was labelled with a fluorophore to identify the size of the amplicon through capillary electrophoresis in a 3730 XL ABI sequencer (Applied Biosystems, Foster City, CA, USA) [6]. MS PCR was repeated for those isolates with no MS PCR amplification. Some isolates from A1 and A4 were previously analysed and already published [6, 14].
The allele fragment sizes recovered from the capillary electrophoresis were determined using Genemapper (Applied Biosystems, Foster City, CA, USA). Only fragments with ≥100 relative fluorescence units (RFU) were considered as ‘real’ alleles. In case of the presence of two or more alleles, only alleles with RFU ≥30% of the dominant allele RFU were considered for further analysis. The genetic and statistical analysis were performed mainly at study areas level but whenever feasible also at the village level to assess the influence of individual villages on the areas. SPSS for Windows v.20 (IBM Corp., NY) was used to perform non-genetic statistical analysis.
Overall we have analysed 292 P. vivax isolates (out of 302) with genotyping success on 56.2% (out of 292 isolates) with all 14 loci, 62.3% with 11–13 loci and 81.5% with 10 or less loci (efficiency of PCR amplification for each locus tabulated in S2 Table). Most of the isolates, 62.7% (183/292) were monoclonal infections and the polyclonal infections carried a minimum of 2 or 3 different haplotypes (34.6% and 2.7%, respectively). The proportion of mono/polyclonal infections significantly differed between areas (p = 0.002), i.e. A1 and A4 presented higher frequency of monoclonal infections while in A2 and A3 the proportions of mono/polyclonal infections were similar (Fig 2A). The median MOI was higher in A2 (MOI = 1.5) and A3 (MOI = 2) compared to the other areas (MOI = 1) (p = 0.003).
The overall genetic diversity estimates among the areas described a median He = 0.74 (range 0.66–0.76), median allelic richness = 4.6 alleles (range 3.6–4.9), and did not differ significantly between areas (He: p = 0.32); allelic richness: p = 0.17), Fig 2B). Private alleles were less in A3 and almost absent in A5 than in the other areas (Fig 2B). The level of polymorphism (He) of each locus is presented in S2 Table. The AMOVA revealed that most of the genetic variation of the parasite population relied within areas (83%) (Table 1), though genetic differentiation between areas was also observed (PHIPT = 0.17, p = 0.0001).
Pairwise calculations of the genetic differentiation and PCoA between areas and/or villages showed that parasites in the geographically isolated areas A1 and A4 were differentiated compared to those circulating in the other three areas with direct access to the Iquitos-Nauta road (Table 2 and Fig 3A). The first two coordinates of the PCoA explained 59.8% of the total variance pointing out clustering of parasites in villages from A1 (median PHIPT 0.02 within A1), presence of related parasites in A2, A3 and A5 and genetic differentiation of parasites from San Carlos village (A4) (Fig 3A).
When considering parasite populations per village, most of the genetic variation actually lied within villages (73% of the total genetic variation) with a high differentiation among the parasites within the villages (PHIPT = 0.27, p = 0.0001), while little differentiation was found between the villages (PHIPR = 0.06, p = 0.004) (Table 1). The PCoA at the individual haplotype level confirmed genetic differentiation among parasites within the same areas even within the same village: i.e. two different groups of haplotypes in San Carlos village (Fig 3B). The genetic distance of the parasite population was not correlated with geographic distance between villages (using PHIPT matrix Rxy = -0.41 p = 0.17) (scatterplot of the genetic and geographic distances in S1 Fig).
In all study areas except for A5, multilocus linkage disequilibrium (LD) was found when all isolates (MOI≥1) or only monoclonal (MOI = 1) isolates were analysed (IAs = 0.08–0.49, for all areas except A5: p<0.0001) (Table 3). LD remained in those areas when only unique haplotypes were considered for the analysis (IAs = 0.08–0.17, p<0.0001). The number of isolates decreased drastically when only monoclonal infections were considered (in A3, from 10 polyclonal isolates to 3 monoclonal isolates). Further LD analysis was performed at village-level only for those villages with more than four isolates. LD was found in villages within A1 (MN, SR and LP, IAs = 0.18–0.21 p<0.0001); A2 (VS, IAs = 0.38 p<2.0x10-04); A3 (VA, IAs = 0.39 p<1.0x10-05) and A4 (SC, IAs = 0.52 p<0.0002). In contrast to the multilocus LD analysis, the pairwise LD analysis was performed by including isolates with missing alleles and showed presence of LD in all areas. The pairwise LD was found mainly between loci located in different contigs therefore the LD found among loci within the same contigs did not alter the outcome (Fig 4).
87 unique haplotypes were found (52.9% of them in A1) in 128 isolates without allelic missing data (Fig 5A; list of haplotypes in S3 Table). The two most frequent haplotypes were found 18 and 11 times (both in A4), followed by 11 haplotypes found 2–4 times (10 from A1 and 1 from A3) and 74 haplotypes were present only once (Fig 5A). Haplotypes were not shared between areas and only shared within the villages of A1. The haplotypes were grouped into 9 to 13 genetic clusters when the eBURST criteria was set to 8 to 13 loci with identical alleles. When the criteria was set to 7 loci, eBURST assigned the parasite population into 3 genetic clusters (PHIPT among clusters = 0.25, p = 0.001) and 6 singletons, where one cluster accounted for 91% (116/128) of the isolates. The K-means clustering divided the population into 2 and 3 genetic clusters and significant genetic differentiation between clusters was found (Rho 0.35 and 0.23, respectively). STRUCTURE analysis was performed to infer cluster assignment including only isolates with up to 3 missing alleles and with known haplotypes. Using STRUCTURE results HARVESTER predicted the most likely number of clusters being K = 2, followed by K = 3, and K = 7 (Fig 6). Using a threshold of 85% for the assignment of group representatives to each cluster, 70–63% of the isolates were assigned to clusters for K = 2, K = 3 and K = 7 and the remaining isolates were assigned as admixed parasites (Fig 5B). The AMOVA analysis revealed high differentiation between these genetic clusters (PHIPT = 0.26–0.59) (Table 1) which were graphically displayed using PCoA (Fig 5C). When the number of clusters was set to K = 2, parasites from all areas belonged to cluster 2 (K2: He = 0.76; IAs = 0.06 p<0.0001) while 18 isolates from San Carlos village belonged to cluster 1 (K1: all the isolates shared the same haplotype) (Fig 5B). Hierarchical structure was subsequently found within cluster 2 where two sub-clusters were found (PHIPT = 0.20, p<0.0001), A1 contained one cluster and hybrid samples while the other cluster was present in all the other areas (S2 Fig). When K = 3, the former 18 isolates from San Carlos and 11 additional isolates from San Carlos were classified as admixed parasites where their genetic composition shared part of the cluster 1 or 2 with a third cluster. No parasites with >85% ancestry belonging to the third cluster were found within our study. Cluster 1 and 2 contained parasites from all areas (He = 0.77 and 0.35; IAs = 0.05 and 0.11, for both p<0.0001). The extent of genetic diversity and LD within the clusters decreased when K = 7 (He varied from no diversity to 0.74, median He = 0.35) and linkage disequilibrium was found for all the clusters (p<0.01).
The minimum spanning tree in Fig 5A displayed the phylogenetic relationship among the parasites coloured by its geographic origin. Besides most A1’s parasites were related to other A1’s parasite (probably due to its larger sample size), all the rest of A1 parasites shared a phylogenetic relationship with parasites from other areas. Among A2 and A3 haplotypes, most A2 haplotypes diverged from A3 haplotypes but almost none from A2 to A3. By the way, the two largest clusters of haplotypes from A4 have diverged from parasites similar to the ones found in A3 while all A5 parasites have diverged from parasites from other areas.
The gene flow was assessed evaluating parasite migration models which relied on the combined knowledge of the genetic structuring of the parasite population and the known human mobility patterns (Figs 1B and 7A). Thirteen migration models were evaluated through Bayesian analysis (marginal likelihoods and LBF of all the models tabulated in S4 Table). The high gene flow rates among all areas denoted the model XIII as the best model (log mL = -47467.4, prob>0.99), which describes a single panmictic population (random mating among the parasites from all the five areas) with an effective population size of 6,891 haplotypes (credibility interval 95% 4,144–9,640). Considering the genetic substructuring and asymmetric human mobilization observed for most of the areas which contrast with the panmictic model, the 2nd and 3rd best models which consider 3 and 5 populations with asymmetric migration were also explored in detail (Fig 7B). The 3-population model XI suggests A2, A3 and A5 as a panmictic unit with unequal migration among A1 to A2/3/5 and few parasite moving from A2/3/5 to A4 (Nm<1). The 5-population model III adds unidirectional gene flow from A1 to A5 which is in line with the phylogenetic results (Figs 5A and 7B). For Model III the highest rate of gene flow was found between the A2 and A3, areas which contain parasites with common genetic characteristics (PCoA and cluster analysis), have the highest multiplicity of infection rates compared to the other areas and is consistent with the current people’s mobilization patterns (Figs 1B, 2A, 3A and 7B). Other models were evaluated were A4 was treated as a fully isolated area or with bidirectional gene flow but these models had the lowest probabilities (Fig 7A and S4 Table). However, model V described unidirectional gene flow from A3 to A4 with a high number of migrants (Nm = 5.6) probably due to related parasite among A3 and A4.
BOTTLENECK analysis using 10 polymorphic markers showed a significant number of microsatellites had an excess of He in areas A1, A2 and A3 under IAM (p<0.002) and TPM (only for A2 and A3, p<0.04), indicating a recent bottleneck event and a deficiency of He for A4 under SMM (p = 0.002), possibly indicating a rapid expansion (S3 Fig). However, A1 also presented contrasting He deficiency under SMM (p = 0.002). A5 was not included in the analysis (n = 4). Similar results were obtained when the analysis was performed separately for perfect and imperfect microsatellites as shown in S3 Fig The presence of excess and deficiency of He in A1 was further investigated by grouping the isolates by village (whenever the sample size was sufficiently large): excess of He in all villages was found but in addition deficit was also found in Manacamiri village (MA: IAM for imperfect MS: excess p = 0.008 and deficit p = 0.04; SMM for perfect MS: excess p = 0.06 and deficit p = 0.03).
The present study provides a very comprehensive dataset on population genetics and gene flow analysis of P. vivax parasites circulating in the Peruvian Amazon. We determined and compared the genetic diversity and multiplicity of infection in five areas in and around Iquitos. We further unravelled the population structure by assessing the LD, genetic differentiation and determined the most likely number of genetic clusters and the genetic relationships among parasites. Multiple gene flow models were assessed determining the parasite migration patterns that may affect the genetic structuring. Moreover the occurrence of recent bottleneck events as result of recent malaria control programs were also explored. The findings support a sub-structured parasite population with a predominant clonal propagation and revealed that Iquitos city (A2) is the source of parasite spreading for all the other areas due to socio-economic patterns. Recent bottleneck events were found only in areas where intervention control programs were carried before we started the sample collection.
MOI, the average number of distinct parasites infecting an individual in a specific area, has been used as a proxy of malaria transmission [12], and it provides information about the configuration of the parasite population [10]. Overall, monoclonal P. vivax infections were the most frequent, confirming previous studies from the Peruvian Amazon [6, 14, 51]. Areas with some degree of isolation, such as areas A1 and A4, had mainly monoclonal infections, whereas the areas close to the Iquitos-Nauta road (A2 and A3) had polyclonal infections with 2 or 3 different clones and a similar proportion of monoclonal and polyclonal infections. This difference may be explained by a higher mobility rate of the people from A2 and A3 leading to a higher probability of being infected with distinct parasites. In areas with limited gene flow, A1 and A4, frequent mating between genetically identical or very related parasites may increase the odds that a person is re-infected with the same “clone”, which would result in a low MOI [52]. Moreover, the MOI will be affected by the patterns of hypnozoite activation where the probability of having a homologous activation may be higher in areas with one or few circulating clones compared to areas where unrelated clones are circulating.
Previously, varying levels of genetic diversity have been reported in endemic settings of the Peruvian Amazon revealing different transmission patterns [6, 14, 15, 51]. Multivariate analysis (AMOVA) indicated that the major source of the genetic variation was due to variation within villages instead of between areas. The greater genetic variation within the villages may be explained by the coexistence of different haplotypes within the villages as results of gene flow, genetic drift and/or a large hidden P. vivax reservoir but also by the high mutability of the microsatellites [53] or overestimation of the number of different haplotypes. The degree of polymorphism of the microsatellites or genotyping errors due to technical artefacts (false alleles) may influence the accuracy of defining the presence of one or more haplotypes within an infection [54].
In our study the levels of genetic diversity remained at intermediate levels with no substantial differences between areas. The high rates of gene flow found may have increased the levels of He. In the present study we reported coexistence of LD and extensive diversity also reported in other studies [55–57]. In the case of repeat-sequence in tandems like the microsatellites, the level of genetic polymorphism (‘genetic diversity’) may be maintained or increased due to the appearance of new alleles by mutational events during replication in the host cells without affecting the LD [55, 57]. Considering long-term P. vivax infections in the Peruvian Amazon the rate of allele mutations occurring within the host needs to be further explored. On the other hand, the extent of LD could be overestimated due to some of the recruited vivax patients were living in the same household [57]. The presence of LD and relatively high genetic diversity may be also favoured by a scenario where inbreeding of few sympatric divergent parasites is frequent [58].
Similarly to other P. vivax populations from South America [11, 53, 58–62], the presence of LD and low MOI indicated a clonal propagation type in our study population. The low malaria transmission and/or the long-term vivax infections in the Peruvian Amazon may also favour the predominant clonal propagation [6]. Passive clonality due to restrained diversity and low gene flow, a different scenario to what we found in the present study, may not be the only type of clonal propagation occurring in the study areas. Tibayrenc and Ayala (2014) coined the ‘in-built, active clonality’ whereas P. falciparum despite the possibility of recombination with sympatric unrelated clones P. falciparum would prefer self-fertilization in order to gain biological and evolutionary adaptations to its environment [52]. The high rate of gene flow and high variability within the study villages drive us to consider that P. vivax may have also in-built, active clonal behaviour in order to take advantage during its adaptation to the environment/hosts. Further study is needed to verify the in-built clonal behaviour and its impact on the epidemiology.
Genetic clustering approaches revealed the presence of at least two or three independent parasite clusters in our study population. We rerun STRUCTURE within one cluster of the cluster when K = 2 to: (1) look for hierarchical structure and; (2) to avoid misleading cluster assignment [37]. Sub-structured parasite population was confirmed in areas A1 and A4 and possibly misleading initial STRUCTURE assignment of the clusters (this is also graphically detected in Fig 5C). The cluster 1, which contained 18 isolates carrying the same haplotype, may influence the misleading assignment: STRUCTURE tends to group strongly related samples into one cluster and the rest of samples are assigned in a large cluster [37].
Little to moderate differentiation among the parasites was found in the urban areas A2, A3 and A5 (“urban cluster”) where continuous human mobilization and parasite gene flow on the Iquitos-Nauta Road occur. Conversely, different clusters of parasites were found circulating in the area A1 and A4 increasing the levels of genetic differentiation. In the San Carlos village (A4) was detected a group of parasites circulating only within this village (Fig 6). Major divergence of these parasites may have occurred due to the limited gene flow, genetic drift, bottleneck events, selection and/or recombination with imported parasites not being sampled in this study. The LD and the deficiency of He detected in A4 confirmed the rapid expansion of these clones with clustered transmission previously described after 2-year follow up [6].
Determining the transmission patterns is a priority for the implementation of control and elimination programs [63]. Our initial analysis revealed no genetic isolation of the study areas despite the geographic distance indicating that exist gene flow. Overall, few private alleles were found among areas where A3 and A5 had the lowest numbers supporting gene flow among the areas where A3 and A5 are beneficiated with more immigrants (parasites). However due to sample size bias on the calculation of private alleles we further used a Bayes approach where our sample size was not anymore an issue. We evaluated 13 migration models, using a Bayes approach based on the coalescence theory, which were proposed under the assumptions of current human mobilization and the genetic structuring data. The panmictic model was inferred as the best model, which is in line with the AMOVA results (less genetic differentiation among areas compared to within the areas). However, due to geographic constrains it is unlikely that people mobilizes to every area while parasites mate randomly. Possibly our genetic data was not strong enough to distinguish the best model: i.e. recent bottleneck events may affect the analysis [45]; however still we were able to recover two models that proposed a better explanation regarding the transmission dynamics and genetic structuring. Both models (XI and III) agreed that area A2 (for model XI: A2/3/5) is the source of parasites spreading for all the other areas: A2 comprises Iquitos city, place where people from A1 and A3 transmutes to every day, crossing the Nanay river by boat or going by car/bus/motorbike through the Iquitos-Nauta Road, for economic activities (Figs 1A and 1B and 7B). The influx of people from A1 to A2 is drives most of the parasite influx among these areas where people from A1 may be infected in A2 then importing parasites from A1 to A2 when they return to their households. Model XI described a panmixia between A2, A3 and A while the model III showed that most of migration occurs between A2 and A3, favoured by the proximity and the road (no geographic barrier) in line with the shared genetic parasite characteristics found previously. In line to the phylogenetic analysis, the model III described that A5 have influx of parasites from A2 which could occur during the visit of people from A5 to A1 on weekends to sell their products in the markets of Iquitos. Model III also described influx of parasites from A1 to A5 despite the Nanay river isolate both areas. High rates of parasite migration from A2 to A3 and A5 is in line with low (or none) number of private alleles recorded for A3 and A5. Regarding to area A4, is it known that occasionally people from A4 (especially from San Carlos) travels through the Itaya river to the A2 to sell their products and that may explain the importation of parasites from A2 to A4. As mentioned before, some parasites from A4 are related to parasites from A3, which in turn are highly related to A2 parasites. The relatedness of A4 and A3 parasites (Network analysis) and migration of parasites from A3 to A4 (referred only in model IV) may have occur by unknown human migration patterns or events where the vector mobilizes among these areas or genetic divergence of A3 parasites into the current A4 parasites. We have documented parasite transmission from A2 to the other areas but not significant immigration of parasites towards A2 which explains why A2 has the smallest population size (number of haplotypes) compared to the other areas (S5 Table) and why we still found significant pairwise genetic differentiation against A1 and A4. In addition to the low rate of immigrants to A2, recent bottleneck events may have also negatively affected the effective population size in A2.
By describing the parasite structure, genetic diversity and dynamics, population genetics can also contribute in assessing the impact of an intervention [10]. Before our sample collection in 2008, the PAMAFRO project which involved campaigns of malaria prevention and control program with active case detection and treatment as well as distribution of insecticide-treated mosquito nets was carried out in Loreto (including all our study areas except for area A4), resulting in a 49% drop of the incidence of clinical vivax malaria from 2005 to 2008 [20]. The expected impact of the intervention on the parasite population besides lowering the malaria incidence would be a reduction of the effective parasite population size, the so called bottleneck effect. Since no data prior to the intervention on the effective parasite population size were available, we performed a retrospective analysis looking for recent bottleneck effect. A parasite population having experienced a recent bottleneck shows a faster decline in the number of alleles compared to a He reduction because rare alleles will be lost with little influence on the He [47]. The predominant clonal propagation found in this study did not affect the Bottleneck analysis since moderate He and significant intra-area genetic variation were found. Only areas with n>15 isolates (S3 Fig) were used for the analysis to increase the resolution power as described by Luikart et al. [64]. In this study, bottleneck events were detected in all areas where control interventions were implemented. San Carlos village did not benefit of any control activity before the sample collection, and no bottleneck but rather a rapid clonal expansion was observed. This is the first report of bottleneck events for P. vivax population in the Peruvian Amazon following the implementation of prevention and control activities. Noteworthy that the reduction of malaria cases in Peru lasted only until 2011 coinciding with the finalization of the PAMAFRO project and since then there has been an steady increase of malaria cases: i.e. Loreto region reported 11,779 vivax malaria cases in 2011 and 60,566 in 2014 [3, 65]. Further and continuous monitoring of the population structure and dynamics of the parasite population is necessary to understand the factors that are involved in the evolution of malaria in the Peruvian Amazon. The detection of recent bottlenecks in the parasite population could be used as complementary tool to measure the efficacy and impact of malaria control programs.
In conclusion, we have elucidated the population genetics of Plasmodium vivax in a large geographical area in and around Iquitos, the main socio-economic capital city of the Peruvian Amazon. We have shown the use of a Bayes approach to infer the gene flow pattern among our study areas and the detection of the reduction of the population size as a result of a control program. The knowledge about the routes of malaria transmission (gene flow) and the effect of control policies on the parasite population is a priority from the public health perspective as well for the formulation of future control policies and assessment of current control/elimination strategies.
|
10.1371/journal.ppat.1004641 | CD200 Receptor Restriction of Myeloid Cell Responses Antagonizes Antiviral Immunity and Facilitates Cytomegalovirus Persistence within Mucosal Tissue | CD200 receptor (CD200R) negatively regulates peripheral and mucosal innate immune responses. Viruses, including herpesviruses, have acquired functional CD200 orthologs, implying that viral exploitation of this pathway is evolutionary advantageous. However, the role that CD200R signaling plays during herpesvirus infection in vivo requires clarification. Utilizing the murine cytomegalovirus (MCMV) model, we demonstrate that CD200R facilitates virus persistence within mucosal tissue. Specifically, MCMV infection of CD200R-deficient mice (CD200R-/-) elicited heightened mucosal virus-specific CD4 T cell responses that restricted virus persistence in the salivary glands. CD200R did not directly inhibit lymphocyte effector function. Instead, CD200R-/- mice exhibited enhanced APC accumulation that in the mucosa was a consequence of elevated cellular proliferation. Although MCMV does not encode an obvious CD200 homolog, productive replication in macrophages induced expression of cellular CD200. CD200 from hematopoietic and non-hematopoietic cells contributed independently to suppression of antiviral control in vivo. These results highlight the CD200-CD200R pathway as an important regulator of antiviral immunity during cytomegalovirus infection that is exploited by MCMV to establish chronicity within mucosal tissue.
| Immune inhibitory receptors, including CD200 receptor (CD200R), can limit immune responses in the mucosa to restrict reactivity to the plethora of harmless antigens that mucosal surfaces are continually exposed to. However, viruses may exploit these suppressive mechanisms to enable their persistence and spread. Many viruses, including herpesviruses, have acquired functional homologs of CD200, the ligand of CD200R, implying that viral exploitation of this pathway is evolutionary advantageous. We now show that the β-herpesvirus murine cytomegalovirus (MCMV) takes advantage of the CD200R inhibitory pathway to persist within a mucosal site of MCMV persistence, the salivary glands. Mice deficient in CD200R mounted elevated antiviral immune responses that were driven by the increased division and accumulation of myeloid cells that function to orchestrate the generation of antiviral effector immune responses. Interestingly, MCMV infection of myeloid cells up-regulated CD200 expression. Thus, MCMV exploits the CD200 pathway to persist within mucosal tissue.
| CD200R is an Immunoglobulin superfamily family member that is expressed by hematopoietic cells, with notably high expression on myeloid cells [1]. The ligand of CD200R, CD200 (OX2), is broadly expressed by cells of hematopoietic and non-hematopoietic origins [2]. The primary function of the CD200R pathway is to limit immune reactivity. CD200-CD200R interactions induce a unidirectional inhibitory signal within CD200R-bearing cells that is mediated by tyrosine motifs in the cytoplasmic domain of CD200R that recruit DOK2 and RasGAP, resulting in inhibition of the ERK pathway [3–6].
The CD200R pathway negatively regulates myeloid cell homeostasis in the periphery [7], and in the pulmonary [8] and, to a lesser extent, the intestinal [9] mucosa. CD200R signaling limits the rapid onset of experimental autoimmune encephalomyelitis [3, 7] and restrains bacterial-induced inflammation [10]. Importantly, CD200R also restricts viral-induced inflammation during respiratory influenza infection [9, 11] and herpes simplex virus (HSV) infection of the cornea [11]. However, CD200R also restricts IFN-dependent control of corona virus infection via regulation of TLR7 [12] and control of intracranial HSV infection [13], demonstrating that this inhibitory receptor can impinge on protective antiviral immunity.
During evolution, numerous herpesviruses have acquired proteins with the potential to induce immune inhibitory receptor signaling [14]. For example, human cytomegalovirus (HCMV) encodes a functional homolog of the inhibitory cytokine interleukin-10 (IL-10) [15]. Rhesus CMV lacking its IL-10 homolog induces increased virus-specific immune responses [16], and IL-10R signaling during murine cytomegalovirus (MCMV) infection antagonizes antiviral immunity and facilitates virus persistence [17–19]. Thus, these studies provide in vivo experimental evidence supporting a rationale for CMV exploitation of host immune regulatory pathways.
Intriguingly HCMV UL119–121 proteins display homology to human CD200 [20], although it is currently unknown whether they induce inhibitory signaling through CD200R. However, numerous herpesviruses are known to encode functional CD200 orthologs (vCD200s) implying that exploitation of this inhibitory pathway is potentially advantageous for herpesviruses. The most well-characterized vCD200 is the Kaposi’s sarcoma-associated herpesvirus (KSHV) protein K14, which suppresses the activation of neutrophils [21], basophils and NK cells [22], T cells [23] and macrophages [24] in vitro. Furthermore, the English isolate of rat cytomegalovirus (RCMV-E) encodes a CD200 homolog (e127) capable of binding CD200R [25, 26].
Despite the possible importance of the CD200-CD200R pathway in modulating anti-CMV immunity, how it influences antiviral immune responses and virus replication during infection in vivo requires clarification. To investigate this, we studied MCMV infection in wild type mice and mice lacking CD200R. Experiments revealed a pivotal role for CD200R regulation of myeloid cell responses in limiting antiviral CD4 T cell responses. We provide evidence that MCMV exploits the CD200-CD200R pathway to facilitate persistent infection within mucosal tissue.
MCMV replicates in numerous organs, including the spleen, liver and lungs, during acute infection, prior to dissemination to the salivary glands (SGs), in which MCMV replicates for 1–2 months [27, 28]. We hypothesized that CD200R signaling may facilitate MCMV replication in vivo. To test this, wild type C57BL/6 (wt) and CD200R-/- mice were infected with MCMV and virus load measured. Peak acute MCMV replication at day 4 post-infection (pi) in the spleen (Fig. 1A) and liver (Fig. 1B) was unaltered by CD200R deficiency. However, CD200R-/- mice exhibited a reduced burden of replicating virus (Fig. 1A) in the spleen 7 days pi.
We next investigated whether CD200R promoted MCMV persistence. In our model, replicating virus is first detectable in the SGs at day 7 pi. Virus load in wt and CD200R-/- mice day 7 pi was comparable (Fig. 1C), suggesting that improved antiviral control in spleens of CD200R-/- mice (Fig. 1A) did not influence dissemination to the SGs and associated brown fat in which MCMV replicates at this time-point [29]. Crucially, however, CD200R-/- mice restricted persistent MCMV replication in the SGs 14 days pi, and more CD200R-/- mice cleared MCMV by day 33 pi as compared to wt controls (Fig. 1C). Thus, intact CD200R during chronic infection promoted virus persistence within this mucosal organ.
Consistent with biological impact of CD200R within the SGs, we observed significant CD200R expression by CD11c+MHC II+ salivary gland (SG) APCs (referred to hereafter as SG-APCs, Fig. 1D&E), which are phenotypically indicative of tissue-resident macrophages [30], and NK cells (Fig. 1D) but not CD4 and CD8 T cells (Fig. 1D). CD200R expression by SG myeloid cells was notably higher than splenic counterparts (Fig. 1E), demonstrating enhanced expression of CD200R in mucosal versus non-mucosal sites of MCMV infection. CD200R expression by myeloid cells in both compartments was relatively stable during infection, with a slight reduction in the intensity of CD200R expression 4 days pi prior to recovery to steady-state levels by 14 days (Fig. 1E). Interleukin-10 (IL-10) is expressed in the SGs in response to MCMV infection and promotes virus persistence [18, 31]. Although IL-10 induces CD200R expression by macrophages in vitro [8], MCMV-infected IL-10-/- mice exhibited no alterations in CD200R expression by myeloid cells during infection (Fig. 1F). Thus, CD200R was expressed during infection but was not significantly upregulated in response to MCMV, by either an IL-10-dependent or independent mechanism.
Unlike certain herpesviruses [14, 24], MCMV does not encode an obvious vCD200 [32]. Within infected SGs, CD200+ cells were predominantly large CD31+ cells (Fig. 2A, isotype controls:S1A Fig.) that were EpCAM- (Fig. 2B), suggestive of endothelial cell origin, and not EpCAM+ acinar epithelial cells in which MCMV replicates during the persistent phase of infection [33]. CD200+ cells did not express alpha smooth muscle actin (S1B Fig.), also demonstrating these cells were not myoepithelial cells. Interestingly, CD200+ cells were often observed in ring-like structures around acinar epithelial cells (Fig. 2B), indicative of capillary networks that surround acini [34]. CD200+CD31+ cells were detectable in naïve SGs (S1C Fig.), and we observed no notable increase in the intensity of CD200 expression by CD31+ cells within infected tissue. In addition to abundant CD200+CD31+ cells, a more scarce population of CD200+CD45+ cells was also detectable within the SGs indicating the presence of CD200 on a hematopoietic cell type(s) (Fig. 2C). Analysis by flow cytometry revealed significant CD200 expression by APCs and T cells within the SGs and spleen (Fig. 2D&E) during MCMV infection.
Interestingly, we noted that CD200 expression by APC populations in the SGs and spleen was induced above baseline upon infection (Fig. 2D&E). We hypothesized that MCMV infection of myeloid cells may directly influence CD200 expression. We infected myeloid cell populations in vitro using a multiplicity of infection of 1 that leads to an infection efficiency of less than 60% (see Fig. 3A for example), enabling us to compare surface CD200 protein levels on uninfected and infected cells from the same well of a tissue culture plate, as identified by flow cytometric detection of the intracellular MCMV m06 protein. Infection of bone marrow-derived macrophages (BM-DM, Fig.3A–C) and splenic macrophages (Fig.3B) up-regulated CD200 (Fig.3A–C). Importantly, we observed a marked increase in CD200 expression by infected (m06+) as compared with uninfected (m06-) macrophages derived from the same wells (Fig. 3A–C), suggesting that CD200 is preferentially up-regulated by macrophages in which MCMV is actively replicating.
Infection of BM-DM with influenza did not trigger CD200 expression (Fig. 3D), demonstrating that CD200 up-regulation is not a generic macrophage response to viruses. However, CD200 expression is induced by ligation of TLRs, including TLR3 and TLR9 [10]; both of which are triggered by MCMV [35, 36]. In accordance, TLR3 ligation by PolyI:C induced substantial CD200 mRNA expression by macrophages in an IFNβ-dependent manner (Fig. 3E). To investigate whether MCMV induction of Cd200 transcription required productive replication, we compared expression following macrophage infection with IE3 knockout replication-deficient MCMV (ΔIE3)[37] and replication-sufficient wt MCMV (pSM3fr). Macrophage exposure to ΔIE3 MCMV induced a small, transient induction of CD200 mRNA in an IFNβ-dependent manner (Fig. 3F), consistent with TLR-mediated induction of CD200 triggered during incomplete MCMV replication, and the moderate CD200 protein expression by uninfected (m06-) macrophages derived from infected cell cultures (Fig. 3C). In contrast, replicating MCMV induced substantial and prolonged CD200 mRNA expression independently of IFNβ (Fig. 3G). Furthermore, inhibition of viral DNA polymerase with phosphonoacetic acid (PAA) antagonized MCMV-induced CD200 expression in BM-DMs (Fig. 3H), again demonstrating the requirement for productive virus replication in this process. Importantly, we observed that SG-APCs did not support MCMV replication in vitro in accordance with the absence of detectable infection in vivo [38], and MCMV infection of splenic DCs did not further induce CD200 expression (Fig. 3B). Thus, these data suggested that myeloid cells up-regulated CD200 during MCMV infection and, in the case of macrophages in secondary lymphoid tissues, MCMV induces CD200 expression independently of TLR stimulation during productive replication.
Given that CD200 expression by both hematopoietic and non-hematopoietic cells was observed in MCMV-infected mice, we sought to understand which cellular compartment was responsible for inhibiting antiviral immunity. We made bone marrow chimeras derived from wt mice or mice deficient of CD200, generating mice lacking CD200 within the hematopoietic and/or radiation-resistant (non-hematopoietic) compartment. We then studied virus load in SGs 14 days post-MCMV infection. Interestingly, deleting CD200 from either compartment reduced virus load as compared to wt>wt mice (Fig. 4), demonstrating that CD200 expressed by hematopoietic and non-hematopoietic cells both delivered immune suppressive signals that promoted MCMV persistence.
We assessed the impact of CD200R deficiency on virus-induced myeloid cell responses. Reduced MCMV persistence in CD200R-/- mice was accompanied by accumulation of splenic DCs 14 days pi (Fig. 5A). The inability of SG-APCs to cross-present antigen to CD8 T cells (in combination with MCMV down-regulation of MHC class I) has been shown to be responsible for the lack of CD8 T cell-mediated control of MCMV replication in the SGs, demonstrating that local antigen-presenting function of SG-APCs is a critical determinant of protective T cell immunity during MCMV persistence [38]. Interestingly, SG-APC accumulation in infected CD200R-/- mice was substantially increased 14 days pi (Fig. 5B&C). However, SG-APC numbers were comparable in wt and CD200R-/- mice following resolution of MCMV infection in our model (48 days pi, S2A Fig.). Thus, CD200R restricted mucosal myeloid cell accumulation during early time-points of SG infection rather than influencing myeloid cell turnover during the resolution phase of infection.
Tissue resident macrophages proliferate in response to inflammatory stimuli [39, 40]. Thus, we measured SG-APC proliferation before and after MCMV infection 7 days pi. Low levels of SG-APC homeostatic proliferation were measured in naïve wt and CD200R-/- mice (Fig. 5D). However, infection-induced SG-APC proliferation was further elevated in CD200R-/- mice as compared to wt mice day 7 pi (Fig. 5D&E), a time at which CD200R was expressed by these cells in wt mice (Fig. 1D&E). This suggested that increased SG-APC accumulation in CD200R-/- mice was a consequence of heightened proliferation. Importantly, visualization of MHC II+ cells within the SGs revealed that MHC II+ cells were consistently located adjacent to large CD200+ cells 7 days (Fig. 5F) and 14 days (S2B Fig.) pi. In accordance with CD200 expression by CD31+ cells (Fig. 2A), MHC II+ cells were also observed surrounding CD31+ vessels (Fig. 5G), suggesting that tissue-resident MHC II+ SG-APC interactions with CD200-bearing endothelial cells restricts infection-induced cellular proliferation. In support of this conclusion, chimeric mice lacking CD200 only in non-hematopoietic cells exhibited increased SG-APC accumulation (S2C Fig.). Furthermore, improved control of MCMV in these mice (Fig. 4) in addition to the absence of an impact of non-hematopoietic cell-derived CD200 on splenic DC responses (S2D Fig.) points towards a role for local SG-APC expansion in determining control of MCMV replication in the mucosa.
CD4 T cells are critical effector cells in the control of MCMV persistence that afford protection via expression of IFNγ [38, 41]. Despite the absence of measurable T cell expression of CD200R (Fig. 1D), SG-infiltrating CD4 T cells in CD200R-/- mice exhibited increased activation, indicated by CD69 and CD25 up-regulation 10 days pi (Fig. 6A&B). Enrichment of CD25hi CD4 T cells were not observed in either wt or CD200R-/- mice (Fig. 6A), consistent with the absence of regulatory T cell infiltration into the SGs in response to MCMV [31]. Importantly, IFNγ+ virus-specific CD4 T cell numbers were elevated in the SGs of CD200R-/- mice by 14 days pi (Fig. 6C). In addition to activated T cells, CD4+ tissue-resident memory T cells also express CD69 [42]. Interestingly, whereas CD69 expression by SG CD4 T cells was elevated in CD200R-/- mice 14 (Fig. 6D) and 30 (S3A Fig.) days pi, elevated prolonged expression of CD25 by CD200R-/- SG CD4 T cells was not observed (S3B Fig.), implying that CD200R may also restrict the accumulation and/or retention of CD4 T cells with a tissue-resident memory-like phenotype. Furthermore, MHC II+ SG-APCs that proliferate and accumulate to higher numbers in CD200R-/- mice (Fig. 5B–E) co-localized with CD4 T cells (Fig. 6E), suggesting that elevated myeloid cell responses within the SGs of CD200R-/- mice enhanced mucosal CD4 T cell responses. In addition, elevated splenic DC numbers 14 days pi in CD200R-/- mice (Fig. 5A) were accompanied by an increase in virus-specific CD4 T cells in this organ at this time (Fig. 6F). These data therefore suggested that CD200R restricted peripheral and mucosal CD4 T cell responsiveness during virus persistence through localized regulation of tissue APC accumulation.
We next investigated whether elevated CD4 T cell responses restricted MCMV persistence in CD200R-/- mice. Depletion of CD4 T cells abrogated the improved control of MCMV in CD200R-/- mice (Fig. 6G), which is consistent both with the established role for CD4 T cells in limiting MCMV persistence in the SGs [38, 41], and the conclusion that MCMV exploits CD200R to facilitate persistence predominantly by antagonizing proliferation and accumulation of MHC class II-bearing myeloid cells.
We demonstrate that MCMV exploits the CD200-CD200R pathway to restrict mucosal antiviral immunity in vivo to facilitate MCMV persistence in a secretory mucosal organ. Restriction of myeloid cell responses was central to the inhibitory action of CD200R. CD200R signaling limited accumulation of MHC class II-bearing APCs in both the periphery and mucosa thus restricting the ensuing virus-specific CD4 T cell response. CD200R restricted SG-APC responses by limiting virus-induced cellular proliferation. CD200R inhibition of this process has likely evolved to limit responses to harmless antigens that mucosal surfaces are continually exposed to. However our data demonstrate that MCMV benefits from this immune-regulatory pathway to persist within its mammalian host, and MCMV can actively induce CD200 expression during infection.
SG-APCs proliferated in response to MCMV infection, consistent with the ability of tissue-resident macrophages to undergo a proliferative burst following inflammation [39, 40, 43]. Our data suggest that the interaction of SG-APCs with the basal surface of CD200-bearing endothelial cells limits this process, and implies that the large vascular network within the SGs may function not only as a blood supply but also to deliver inhibitory signals that, in the context of homeostatic conditions, functions to limit immune responsiveness. The identification of ring-like structures surrounding acini implies a scenario during MCMV infection in which CD200R-expressing myeloid cells situated close to or migrating towards infected cells may receive inhibitory signals from CD200-expressing vascular structures.
Currently, there are no methodologies available to exclusively delete SG-APCs in vivo. Therefore we are unable to make definitive conclusions regarding the function of these cells in our experiments. However, our data supports a model in which CD200R-mediated restriction of SG-APC proliferation reduces CD4 T cell activation within the SGs, subsequently impairing CD4 T cell responsiveness and control of MCMV persistence. Our data also demonstrated that CD200R impaired the accumulation of virus-specific CD4 T cells in the periphery that was accompanied by reduced splenic DC accumulation. Thus, CD200R signaling impinges on antiviral protection from mucosal MCMV replication by restricting CD4 T cell activation and expansion both within the mucosa itself, but also in secondary lymphoid tissue.
Intriguingly, persistent MCMV infection of CD200R-/- mice led to the enrichment of CD69+ CD4 T cells not expressing the activation marker CD25. CD69 is expressed by tissue-resident memory CD4 T cells [42]. MCMV replication continued at the time-points at which CD69+ CD4 T cells were detected in our study, thus precluding definitive conclusions regarding bone fide tissue-resident memory cells. However, our data implies that CD200R may indirectly restrict the accumulation of CD4 T cells in the SGs that exhibit a phenotype indicative of tissue-resident memory CD4 T cells.
CD200R facilitates early viral replication in acute MHV [12] and HSV [13] infections in vivo. In contrast, we observed that early control of MCMV was unaffected by CD200R. This may reflect in part that CD200R deficiency did not influence MCMV replication in macrophages (S4A Fig.), unlike data reported in HSV infection [13]. Improved control of MHV infection in CD200-/- mice was associated with elevated type I IFN [12]. Type I IFN was not measured in our study and may not be altered in MCMV-infected CD200R-/- mice. Also, type I IFN exerts potent antiviral activity against MCMV in vivo in wt mice [44] and may therefore be produced at levels that exert maximal antiviral activity in our model irrespective of any impact of CD200R on cytokine expression.
Instead, we show for the first time that CD200R signaling influences persistent virus replication in vivo. Improved control of MCMV replication in the SGs in CD200R-/-mice was intriguing given that MCMV does not encode an obvious CD200 homolog. This may be explained in part by the existence of a structural CD200 ortholog encoded by MCMV that lacks sufficient sequence similarity to be detected, or by the existence of other viral ligands for CD200R. Importantly however, experiments utilizing CD200-/- mice highlighted a role for cellular CD200 in dampening antiviral immunity. Cellular CD200 restricts virus-induced immune responses in acute virus infections [8, 12, 45], and our data supports the conclusion that some viruses may exploit host CD200-CD200R interactions to establish persistence. Intriguingly, in vivo experiments investigating a functional role for CD200 orthologs expressed by RCMV [26] and Rhesus macaque rhadinovirus [46] failed to detect significant benefit of these vCD200s in promoting herpesvirus persistence in these experimental models. Our data suggest the benefit of herpesvirus exploitation of host CD200 expression, irrespective of whether the virus also encodes its own vCD200 protein.
Results obtained from bone marrow chimeras demonstrate the importance of non-hematopoietic cell-derived CD200 in facilitating MCMV persistence, thus supporting an important role for endothelial cells in indirectly restricting antiviral CD4 T cell responses via regulation of myeloid cells. However, a significant role for hematopoietic cells in promoting virus persistence was also revealed in these experiments. Peripheral and mucosal myeloid cells expressed CD200 during MCMV infection. Although SG-APCs did not support MCMV replication, splenic macrophages up-regulated CD200 following direct MCMV infection in vitro. MCMV infection of wt and CD200R-/- bone marrow-derived macrophages resulted in comparable expression of MHC II (S4B Fig.), suggesting that MCMV does not exploit macrophage expression of CD200 to deliver an autocrine inhibitory signal; a conclusion further supported by comparable MCMV replication in wt and CD200R-/- macrophages and consistent with the inability of CD200 to interact with CD200R in a cis-cellular fashion [47, 48].
Instead our data suggest that a CD200-bearing myeloid cell may restrict antiviral immunity and that, in the case of peripheral infection, MCMV influences this process.
CD200 may suppress myeloid cell activity and/or accumulation indirectly via an unknown CD200R-expressing cell subset, or by directly triggering CD200R signaling within a myeloid cell. T cells expressed CD200 during MCMV infection, implying that MCMV may also passively exploit a negative feedback loop by which CD200-bearing T cells deliver an inhibitory signal to CD200R-bearing myeloid cell with which they interact. Notably however, non-hematopoietic cell-derived CD200 restricted myeloid cell accumulation within the SGs, suggesting that T cells do not exert CD200-mediated inhibition of myeloid cell proliferation within this particular site of MCMV infection. Irrespective of the exact mechanism(s), our data suggest that CD200 expressed by hematopoietic cells impacts on the development of antiviral immunity that subsequently allows virus persistence within the SGs, and that MCMV actively exploits this process.
MCMV induced myeloid cell CD200 expression via two distinct mechanisms. Firstly, incomplete virus replication triggered TLR-induced IFNβ-dependent Cd200 gene expression. Importantly, replication-competent virus induced Cd200 expression in macrophages independently of this pathway, and CD200 induction was dependent upon viral DNA polymerase activity. The mechanism through which MCMV actively regulates CD200 is not clear. CMV infection induces profound alterations in host cell protein production and gene expression [49–52]. Concurrent analysis of Cd200 gene and surface protein expression highlighted that viral induction of CD200 occurred at the transcriptional level. The impact of PAA on virus-induced CD200 expression suggests the involvement of a gene product or products expressed during the latter stages of virus replication. However, this conclusion is guarded given that inhibition of viral DNA polymerase during HCMV infection also incompletely inhibits production of certain viral proteins expressed at early times during the virus life-cycle [53].
Whether a viral gene product(s) directly or indirectly induces CD200 expression and which viral protein is responsible remains to be elucidated. Influenza infection of macrophages did not trigger CD200 expression despite the CD200-CD200R pathway restricting influenza-induced T cell responses in vivo [8, 12]. Thus, CD200 induction is not a generic response mounted by macrophages in response to viruses. Instead, our experiments demonstrate that MCMV gene expression is essential for this process and implies that CD200 up-regulation represents a previously unappreciated mechanism exploited by CMV, and perhaps other viruses, to antagonize host antiviral immunity.
Collectively, our study highlights a central role for myeloid cells in modulating cytomegalovirus-specific T cell responses in mucosal tissue and the potential importance of regulation of tissue-resident macrophage proliferation in this process. Our study also points towards the manipulation of cellular CD200 expression as a mechanism through which herpesviruses evade host immunity, suggesting that MCMV exploits CD200R signaling to antagonize myeloid cell orchestration of antiviral immunity to promote persistence within and dissemination from the mucosa.
C57BL/6 experimental mice were obtained from Harlan UK. CD200R-/- mice were originally generated and provided by Reginald Gorczynski (University Health Network, Toronto), and David Copland (University of Bristol) provided the OX-2-/- mice, with kind permission from Jonathon Sedgwick (Eli Lilly, Indianapolis). IL-10-/- mice were purchased from Jackson Laboratories and maintained in-house.
MCMV Smith strain (ATCC) was prepared in BALB/c salivary glands and purified over a sorbital gradient. Mice were infected by the intra-peritoneal route (i.p) with 3 x 104 PFU MCMV. Some mice were injected i.p with 200µg αCD4 antibody (100µg clone YTS191, 100µg clone YTS3) on days 4 and 6 pi. To measure proliferation in vivo, mice were injected i.p with 1mg/mouse EdU (Life Technologies) at day 6 pi. To generate chimeric mice, recipients were irradiated at 2 x 550G, transfused intra-venous (i.v) with 1 x 106 bone marrow cells 24 hours later. Mice were then treated for 3 weeks with baytril-supplemented water. Mice were infected with MCMV 8 weeks after irradiation.
All experiments were conducted according to the UK Home Office guidelines at the designated facility at Heath Park, Cardiff University under UK Home Office-approved project licenses PPLs 30/2442 and 30/2969.
SGs and spleens were surgically excised from mice that were euthanized with carbon dioxide. SGs were cut into small pieces and incubated in RPMI 1640 medium (Invitrogen) supplemented with 5 mM CaCl2, 5% FCS (Invitrogen), 1 mg/ml collagenase D (Roche Diagnostics), and 10 mg/ml DNAse I (Sigma) at 37°C for 45 minutes, before passing through a cell strainer prior to red blood cell lysis. Leukocytes were then stained with Live/Dead (Invitrogen) prior to incubation with Fc block (eBioscience). Lymphocytes were then stained with a combination of αCD3e-PerCP (Clone 145.2C11, Biolegend), αF4/80-Pacific-Blue (Clone BM8; Biolegend), αIA/IE-PerCP-Cy5.5 (Clone M5/114.15.2, BioLegend), αCD11c-PeCy7 (Clone N418, Biolegend), αNK1.1-allophycocyanin (Clone PK136, BD Biosciences), αCD4-Pacific-blue (Clone RM4.5, BD Biosciences), αCD25-APC-Cy7 (Clone PC61, Biolegend) and αCD69-FITC (Clone H1.2F3, eBioscience).
To detect EdU incorporation, cells were stained as above, fixed with 4% PFA, permeabilized with Saponin buffer, and EdU was labeled with Alexa Fluor 647 using the Click-iT Plus EdU Alexa Fluor 647 Flow Cytometry Assay Kit (Life Technologies) as per manufacturer’s protocol. To detect MCMV-specific CD4 T cells, leukocytes were incubated with 3μg MCMV peptides (Genscript) listed in Figure legends for 6 hours, with BFA (Sigma) for the final 4 hours. CD4 T cells stained as above were permeabilized prior to staining with αIFNγ FITC (clone XMG1.2, eBioscience).
All data were acquired on a BD FACS Canto II. Electronic compensation was performed with antibody-capture beads (BD Biosciences). Data was analyzed using FlowJo software version 10.0.3 (TreeStar Inc, Ashland, OR). Total numbers of different cell populations were calculated by multiplying % positive viable cells detected by flow cytometry x the total number of viable leukocytes (assessed by trypan blue exclusion).
Femurs were surgically excised from wt and CD200R-/- mice, sterilized in 70% ethanol and washed in PBS. Bone marrow was isolated, cells centrifuged, washed in RPMI and passed through a 40µM cell strainer. Cells were incubated at 2 x 105 cells/well in D10 media supplemented with 20ng/ml of M-CSF (Peprotech) for 7 days, replenishing M-CSF after 3 days. Spleens and SGs were processed as previously described, with an additional Percoll (GE Healthcare) purification step for SGs after processing. Bone-marrow derived macrophages were infected with MCMV or influenza (PR8) at an MOI of 1. Some cells were also incubated with 300μg/ml phosphonoacetic acid (PAA, Sigma-Aldrich) for 1 hour prior to infection. Splenocytes (2 x 105 cells/well) and SG leukocytes (2 x 104 cells/well) were infected in 48-well plates and infected with MOI 0.5 MCMV. After 24hrs, all macrophages were gently scraped gently off the bottom of the wells, stained with Live/Dead® fixable aqua dead cell stain (Invitrogen) and Fc block (eBioscience), and surface stained with αCD200-PE (Clone OX-90, Biolegend), αCD80 Pacific blue (Clone 16–10A1, Biolegend), αCD86 FITC (Clone GL-1, BD Pharmingen), and αIA/IE PerCP/Cy5.5 (Clone M5/114.15.2, Biolegend) prior to permeabilization and staining with anti-m06 antibody (a kind gift from Stipan Jonjic, Rijeka) conjugated with APC (Innova Biosciences).
SGs were frozen in OCT and 5μm thick sections fixed in acetone. Sections were blocked with Avidin/Biotin Blocking Kit (Vectorlabs) and then with 2.5% Normal Horse Serum (Vectorlabs). Sections were incubated overnight at 4°C in the dark with CD31-Biotin (Clone MEC 13.3, BD Pharmingen) or CD200-Biotin (Clone OX-90, BioLegend), and MHC II-FITC (Clone M5/114.15.2, BioLegend) or EpCAM (Clone E144, AbCam). Alexa Fluor 488 anti-rabbit IgG (Invitrogen) and Streptavidin Alexa Fluor 555 conjugate (Invitrogen) were used as secondary stains for EpCAM, and CD200-Biotin and CD31-Biotin, respectively. Sections were counterstained with TOTO-3 (Invitrogen), then fixed with 1% PFA and treated with 0.3M glycine. To investigate CD200 colocalization with CD31 or CD45, and MHC II colocalization with CD4, sections were incubated with CD45 (Clone 30-F11, Biolegend), CD31-FITC (Clone 390, eBioscience) or CD4 (Clone RM4–5, Biolegend) overnight at 4°C in the dark. Alexa Fluor 488 anti-rat IgG (Life Technologies), FITC anti-rat IgG2b antibody (Biolegend) and Alexa Fluor 568 goat anti-rat (Life Technologies) were used as secondary stains for CD31-FITC, CD45 and CD4, respectively. Sections were fixed in 1% PFA and treated with 0.3M glycine and then incubated with anti-CD200-Biotin for 2 hours at room temperature, followed by Streptavidin Alexa Fluor 555 conjugate (Invitrogen), or MHC II-FITC (without secondary antibody). Sections were counterstained with TOTO-3 (Invitrogen) and fixed. The following isotype controls were used: Rat IgG2a-Biotin (BD Pharmingen) for CD200-Biotin and CD31-Biotin, Rat IgG2a-FITC (eBioscience) for CD31-FITC, Rat IgG2b-FITC (eBioscience) for MHC II-FITC, Rabbit IgG (AbCam) for EpCam, Rat IgG2a (eBioscience) for CD4, and Rat IgG2b (BD Pharmingen) for CD45. Images were collected with a Zeiss Axioskop 2 FS mot confocal microscope. Images were assembled using ImageJ software.
Wt and IFNβ1-/- bone marrow derived macrophages (BM-DM) were derived from C57/BL6 mice as previously described [54] and grown in 24 well plates. After 7 days of culture, BM-DM were infected with wt-MCMV, MCMVΔIE3 (MOI = 1) or mock infected [55]. Cells were then harvested at 2, 4, 6, 8, 10 and 24 hours post-infection for the isolation of RNA using an RNeasy Mini kit (Qiagen, UK) according to manufacturer’s instructions. After QC using an Agilent Bioanalyzer, total RNA was labeled and hybridized to Mouse Gene 1.0ST microarrays (Affymetrix, CA, USA) according to manufacturer’s instructions using a WT Expression kit (Ambion, UK). After data capture, quality control metrics were assessed using Affymetrix Expression Console software and then all arrays were imported into Partek Genomics Suite (Partek, USA) for downstream analysis. In brief, arrays were normalized using the gcRMA algorithm [56]. After normalization, to increase confidence in the genes taken forward to statistical analysis, data was filtered to include genes with at least 1 signal value of > = 150 across the time course.
For viral load analysis, statistical significance was determined using the Mann-Whitney U test for paired groups. To analyze viral load data from bone marrow chimeras, linear regression analysis was utilized. Data were first subject to square-root transformation to introduce stability. We then fitted a linear model for covariates (donor + recipient) with and without the interaction term. Subsequent ANOVA analysis of these models showed the interaction term not to be significant (p = 0.13). However a model without any interactions is strongly significant (p = 0.0015) and was therefore used. For paired analysis of flow cytometry data, the two-tailed Student’s t test was utilized. For bone marrow chimeras, linear regression of non-transformed data was used. *p<0.05, **p<0.01, ***p<0.001.
mCD200–17470; mCD200R- 57781l; IFNγ- 15978; mIL-10–16153; mCD69–12515; mCD31/PECAM-1–18613; mIFNB1–15977; mCD80–12519
|
10.1371/journal.pgen.1007884 | Diversification of DNA binding specificities enabled SREBP transcription regulators to expand the repertoire of cellular functions that they govern in fungi | The Sterol Regulatory Element Binding Proteins (SREBPs) are basic-helix-loop-helix transcription regulators that control the expression of sterol biosynthesis genes in higher eukaryotes and some fungi. Surprisingly, SREBPs do not regulate sterol biosynthesis in the ascomycete yeasts (Saccharomycotina) as this role was handed off to an unrelated transcription regulator in this clade. The SREBPs, nonetheless, expanded in fungi such as the ascomycete yeasts Candida spp., raising questions about their role and evolution in these organisms. Here we report that the fungal SREBPs diversified their DNA binding preferences concomitantly with an expansion in function. We establish that several branches of fungal SREBPs preferentially bind non-palindromic DNA sequences, in contrast to the palindromic DNA motifs recognized by most basic-helix-loop-helix proteins (including SREBPs) in higher eukaryotes. Reconstruction and biochemical characterization of the likely ancestor protein suggest that an intrinsic DNA binding promiscuity in the family was resolved by alternative mechanisms in different branches of fungal SREBPs. Furthermore, we show that two SREBPs in the human commensal yeast Candida albicans drive a transcriptional cascade that inhibits a morphological switch under anaerobic conditions. Preventing this morphological transition enhances C. albicans colonization of the mammalian intestine, the fungus’ natural niche. Thus, our results illustrate how diversification in DNA binding preferences enabled the functional expansion of a family of eukaryotic transcription regulators.
| Transcription regulation is the primary step by which most cells control the expression of their genes. At its core, this process is mediated by proteins (transcription regulators) that bind to short DNA regulatory elements in a sequence-specific manner. Recent research in multiple model organisms ranging from vertebrates to unicellular yeasts has revealed that evolutionary changes either in the DNA regulatory elements or in the transcription regulators themselves underlie the origin of many traits such as morphological innovations or the ability to colonize new environments. While the effects of mutations that abolish or create DNA regulatory elements are straightforward to rationalize, understanding what sort of modifications the transcription regulators undergo and how these changes impinge upon the regulatory circuitry of the organism remains a key challenge. Here we investigate the mechanisms whereby a family of conserved transcription regulators diversified the biological functions that they control. While in most eukaryotes this family of regulators governs lipid biosynthesis, three members of the family in the human pathogen Candida albicans have acquired different functions, some of which contribute to the ability of this yeast to reside in the human host and cause disease.
| Evolutionary changes in gene expression patterns constitute a major source of phenotypic diversity [1–4]. The primary step through which all cells regulate expression of their genes is the binding of transcription regulators to cis-regulatory sequences. Not surprisingly, gains and losses of cis-regulatory sequences have been found to underlie many cases of transcriptional rewiring [5–12]. Although changes in the transcription regulators themselves are also important sources of evolutionary rewiring [13–17], relatively few examples of how these proteins change are understood in molecular detail. In particular, little is known about how different DNA binding preferences arise within a family of transcription regulators and whether such variation results in the functional diversification of the family. We address this question here studying the SREBP (sterol regulatory element binding protein) family of transcription regulators (reviewed in [18–20]). While the SREBPs have been traditionally associated with the regulation of sterol biosynthesis genes, several members of this family appear to govern cellular processes unrelated to lipid synthesis.
SREBPs are basic-helix-loop-helix (bHLH) transcription regulators extensively distributed among eukaryotes. bHLH proteins contain a characteristic 60-to-100-residue DNA binding domain composed of two segments that form amphipathic α-helices separated by a loop region that varies in sequence and length [21, 22]. The SREBPs are unique among the bHLH proteins in that they have a tyrosine residue in a conserved position of the first helix of the DNA binding domain where bHLH proteins normally have an arginine [23, 24]. The tyrosine residue allows the human SREBP to bind, at least in vitro, to an additional DNA sequence besides the canonical, palindromic E-box (5’-CANNTG-3’) that is recognized by most bHLH transcription regulators [24]. The significance of this “dual” DNA binding ability remains unclear because chromatin immunoprecipitation (ChIP) experiments have shown that the human SREBP binds in vivo to the same canonical, palindromic E-box sequence [25] as other bHLH proteins.
In addition to higher eukaryotes, SREBP family members are also broadly distributed in fungi. While most fungal genomes encode one or two SREBPs, the family has expanded in some lineages such as the Candida clade of the ascomycete yeasts (Saccharomycotina). Strikingly, SREBPs do not regulate sterol biosynthesis genes in the ascomycete yeasts as this role was handed off to an unrelated transcription regulator in the common ancestor of all Saccharomycotina [26]. Yet the SREBPs appear to play critical and non-redundant roles in the biology of these fungi. In the human commensal and pathogenic yeast Candida albicans, deletion of any of its three SREBPs results in reduced ability to colonize and proliferate in the mammalian host [27–29]. One of the C. albicans SREBPs (TYE7) has been shown to regulate carbohydrate metabolism [27] but the function(s) of the other two regulators is (are) less clear.
Here we investigate the mechanisms that allowed the fungal SREBPs to expand their repertoire of regulatory targets beyond sterol biosynthesis genes. Phylogenetic reconstruction of the family in fungi indicates that the ascomycete yeasts’ SREBPs comprise three distinct branches. We establish that only one of the three branches binds the palindromic DNA E-box motif that SREBPs in higher eukaryotes are known to recognize. The second branch preferentially binds a non-palindromic DNA sequence, whereas the third branch appears to have reduced its DNA binding sequence to a single half-site. Each one of the C. albicans SREBPs belongs to a different branch of the family, explaining the non-redundant role(s) that each protein has in this organism. Ancestral protein reconstruction experiments indicate that the intrinsic DNA binding plasticity observed in the SREBPs—which is conferred by the characteristic tyrosine residue in the first helix of their DNA binding domain—has been resolved in fungi to give rise to extant proteins that exhibit different DNA binding preferences. Furthermore, we show that in C. albicans two of its SREBPs act in concert to inhibit a morphological switch under anaerobic conditions. Preventing this morphological transition enhances the fitness of C. albicans in the mammalian intestine, a natural niche where the fungus resides. Taken together, our results illustrate how generating variation in DNA binding preferences enabled the functional diversification of the SREBP transcription regulators in fungi.
The SREBP family of transcription regulators is widely represented in fungi. A distinctive feature of this family—which distinguishes them from other bHLH proteins—is the presence of a tyrosine residue instead of an arginine in the first helix of the DNA binding domain (Fig 1A). Using this hallmark as the main criterion for inclusion, we assembled a comprehensive phylogeny of the fungal SREBPs based on a manually curated alignment of the DNA binding domain of ~200 proteins (S1 Table; models and computational procedures used for phylogenetic reconstruction are described under Materials and Methods). Little, if any, sequence conservation was detected beyond the SREBPs’ DNA binding domain. Some of the most studied SREBPs (e.g. those in the model fungus Schizosaccharomyces pombe and in humans) harbor transmembrane domains which serve to localize the regulators to intracellular membranes. Upon protein cleavage, the DNA binding domain of these SREBPs is released from the membrane into the cytosol and can shuttle to the nucleus. However, other SREBPs (e.g. those in the ascomycete model yeast Saccharomyces cerevisiae) clearly lack transmembrane domains [30]. Thus, we also scanned the full length of each protein for putative transmembrane sequences to establish whether or not the presence of such domain was widespread across the fungal SREBPs.
The resulting phylogeny points to the existence of several sub-groups within the fungal SREBPs (Fig 1B and S1 Fig). Of particular interest to this report, the ascomycete yeasts’ SREBPs (i.e. the Saccharomycotina) partitioned in three different branches (labeled 1, 2 and 3 in Fig 1B). The separation in three clusters is also supported by other independent, large-scale reconstructions of fungal gene families such as Fungal Orthogroups [31]. As in most other organisms, the majority of species in the Saccharomycotina encode no more than one or two SREBPs. A few species in the Candida clade, however, encode three SREBPs (namely Cph2p, Hms1p and Tye7p) (Fig 1C). Remarkably, each one of these three proteins lies in a different branch of the phylogenetic tree (Fig 1B and 1C) indicating that the Candida proteins span a considerable distance in the phylogeny. Within the Saccharomycotina, only branch 1 (which includes the Candida Cph2 protein) contains putative transmembrane domains whereas the other two groups (branches 2 and 3 in Fig 1B) do not. A sub-cluster of SREBPs in Aspergillus spp. is the only other group outside the Saccharomycotina that appears to lack transmembrane domains. In this study, we focus on characterizing SREBPs that are representative of branches 1, 2 and 3.
bHLH proteins are known to recognize and bind to variants of a palindromic DNA sequence termed E-box (the core motif is 5’-CANNTG-3’) [21,25]. Chromatin immunoprecipitation (ChIP) experiments have shown that the archetype member of the SREBP family, the human SREBP1, indeed binds to an instance of this palindromic E-box sequence in vivo [25]. However, classic in vitro DNA binding assays initially demonstrated that the human SREBP1 can bind not only to the E-box but also to a non-palindromic sequence (5’-TCANNCCA-3’) [23]. Whether binding to this alternative DNA sequence happens only in vitro or also takes place in vivo remained unclear. Here we sought to evaluate the intrinsic DNA binding preferences of the three selected branches of fungal SREBPs.
As a first systematic approach to establish the DNA binding preferences of the three proteins (Cph2p, Hms1p and Tye7p, all from Candida albicans), we employed MITOMI [32], a large-scale microfluidic-based approach that enables the in vitro measurement of protein-DNA interactions at equilibrium between transcription regulators and a comprehensive library of oligonucleotides. In each experiment, we assessed binding to a set of 740 double-stranded 70-nt oligos designed so that all possible 8-mers were represented in the library. The binding was then quantified by measuring the ratio of fluorescence emitted from labeled DNA binding to surface-immobilized labeled transcription regulators [32]. Pair-wise comparisons of the oligos bound by the proteins indicate that the oligonucleotide binding patterns observed for Hms1p and Tye7p are largely orthogonal whereas the other two pairs (Cph2p - Hms1p and Cph2p - Tye7p) display some overlapping binding preferences (Fig 2A). Examining the top 10% of oligonucleotides bound by each protein shows that, to a significant extent, they bind different sets of DNA sequences (Fig 2B). A similar pattern emerges if the top 30% or even the top 50% of oligomers are considered (S2 Fig). Comparisons of the top oligonucleotides bound and shared by the proteins reveals that CaHms1p and CaTye7p are the most distant from each other whereas CaCph2p appears as an intermediate (i.e. it shares a similar number of bound oligomers with CaHms1p and with CaTye7p).
We next used MatrixREDUCE [33] to analyze the binding intensities from all oligonucleotides and find DNA motifs overrepresented in the MITOMI data. We run several iterations of the software varying the length of the motif to be searched and allowing or not the inclusion of a 2-nucleotide spacer. All the DNA motifs generated by MatrixREDUCE (at P < 1 × 10−10) were then compiled and ranked according to r2 and P-values (full list with scores can be found in S2 Table). Representatives of the top ranked motifs for each protein are shown in Fig 2C. To a large extent, the MITOMI motifs resembled either the palindromic E-box variant 5’-ATCANNTGA-3’ or the non-palindromic sequence 5’-ATCANNCCA-3’ (or their predicted half-sites). Consistent with the pattern of overlap in bound oligonucleotides, the CaHms1p- and CaTye7p-derived motifs were the least similar to each other. CaCph2p, on the other hand, appeared as an intermediate that could recognize both types of motifs.
As a complementary approach to determine the in vivo DNA binding preferences of the proteins, we analyzed genome-wide chromatin immunoprecipitation (ChIP) data. While such datasets have been generated for all three proteins in C. albicans, a clear DNA motif could be derived only for Tye7p and Hms1p [27, 29]. Thus, we performed our own ChIP-Seq experiment of the third SREBP in C. albicans, Cph2p. The putative DNA binding domain of Cph2p is located at the N-terminal portion of the protein and is followed by two transmembrane domains that anchor Cph2p to an intracellular membrane. An unidentified signal is thought to trigger the cleavage and release of the N-terminal portion of Cph2p from the membrane and its posterior shuttle to the nucleus [34]. To circumvent the need for an “activating” signal, we generated a C. albicans strain encoding a truncated version of the protein which ends immediately before the transmembrane domain and is Myc-tagged at this new C-terminus (S3A Fig). The ChIP-Seq experiment conducted with this strain identified 14 high-confidence binding regions located within intergenic sequences (Fig 2D and S3B Fig). A clear DNA motif could be derived from this in vivo Cph2p occupancy data set (Fig 2C). The derived motif represents a bona fide binding sequence because: First, the purified CaCph2 protein gel shifted a DNA fragment harboring an instance of the motif; and, second, point mutations introduced in the putative DNA binding site significantly impaired the shift (S3C Fig).
As illustrated in Fig 2C, the MITOMI- and ChIP-derived DNA motifs were, to a significant extent, congruent with each other and revealed distinct DNA binding preferences for each protein. CaTye7p bound to a singular variant of the palindromic E-box motif that consisted of an extended left half-site (5’-CATCA-3’) and a three-nucleotide right half-site (5’-TGA-3’). While the MITOMI analysis was unable to capture the full 10-nucleotide sequence in a single motif, the two separate Tye7p MITOMI motifs shown in Fig 2C could explain the full-length sequence when combined. CaHms1p bound to an alternative, non-palindromic sequence (5’-ATCANNCCA-3’). In this case, the Hms1p MITOMI motif that included a 2-nucleotide spacer was in very close agreement with the full Hms1p ChIP motif. In contrast to Hms1p and Tye7p, the Cph2p MITOMI motifs suggested that, at least in vitro, this protein may recognize both (5’-ATCANNTGA-3’) and (5’-ATCANNCCA-3’) sequences. The Cph2p ChIP motif, on the other hand, indicated that, in vivo, this protein might simply bind to the left portion of either sequence (5’-A/CATCA-3’).
Since manual, detailed examination of the DNA regions occupied by the Cph2 protein in vivo produced no evidence of a composite motif (i.e. a second half-site), we considered the possibility that co-factors could contribute to this protein’s binding in vivo. Indeed, DNA motif searches in our CaCph2 ChIP dataset revealed the co-occurrence of a DNA sequence that closely resembles the DNA motif recognized by the C. albicans regulator Efg1p (Fig 3A). Consistent with this result, we found that a significant proportion of these sites are occupied by Efg1p in vivo (P = 2.6 × 10−5) (Fig 3B; [35]). These observations suggested that the Cph2 and Efg1 proteins may interact in vivo (either by binding cooperatively or by competing with each other for binding) to regulate a subset of target promoters. Consistent with this hypothesis, we found that the expression of two direct targets of regulation (ORF19.921 and ORF19.4941; each containing Cph2p- and Efg1p-binding sites in their putative promoter regions as indicated in Fig 3B) is dependent, at least in part, on CPH2 and EFG1 (Fig 3C). Taken together, these results indicate that CaCph2p may by itself recognize a shorter DNA sequence (compared to the other two branches of SREBPs studied here) but it likely operates in concert with co-factors such as Efg1p.
MITOMI and ChIP data clearly indicate that the Hms1 protein from C. albicans binds a non-palindromic DNA sequence, which is unusual because most bHLH proteins bind a palindromic DNA motif. We wondered whether this unusual binding preference was exclusive to this protein in this species or extended to other SREBPs. To address this question, in addition to CaHms1p, we purified the putative DNA binding domains of the C. parapsilosis Hms1 (CPAR2_303750) and the A. fumigatus SrbA (Afu2g01260) proteins and carried out electrophoretic mobility gel shift assays (EMSAs). As shown in Fig 4A, all three proteins bound to a DNA fragment harboring an instance of the non-palindromic motif. This binding was specific to the analyzed DNA sequence because point mutations introduced in the putative binding site abolished or severely impaired binding (Fig 4A).
We next wanted to determine whether the proteins were able to discriminate between non-palindromic and canonical E-box sequences. For this, we carried out competition binding assays in which we incubated CaHms1p or AfSrbA with a 32P-labeled DNA fragment carrying the non-palindromic sequence. Upon binding, we competed the reactions with unlabeled DNA fragments harboring either the non-palindromic site or the canonical E-box sequence (Fig 4B). The former DNA fragment was a stronger competitor compared to the latter (Fig 4B and 4C and S4 Fig) indicating that the proteins exhibit a marked preference for the non-palindromic sequence. Taken together, these results demonstrate that the preferential binding to a non-palindromic DNA sequence is a property shared by multiple fungal SREBPs, both within and outside the ascomycete yeasts.
We showed above that the purified DNA binding domain of the CaHms1 protein exhibits a strong preference for its cognate DNA binding sequence (a non-palindromic DNA site) over the canonical E-box motif in competitive EMSAs (Fig 4B and 4C). Since CaHms1p’s ability to discriminate between DNA sequences is clearly an intrinsic property of the protein, we sought to determine what portion(s) of its DNA binding domain confer(s) this ability.
We constructed several chimeric proteins by exchanging one of three portions (first helix, loop region or second helix) of the DNA binding domains of CaHms1p and CaCph2p (the latter protein displayed little, if any, ability to discriminate in vitro between the two DNA sequences evaluated here (S5A and S5B Fig)). We then employed EMSAs to probe each chimeric protein for their ability to bind DNA fragments harboring either the cognate HMS1 binding site or the canonical E-box sequence. We found that a chimeric protein consisting of the first helix and the loop from CaHms1p and the second helix from CaCph2p recapitulated almost completely the ability to discriminate between the two DNA sequences as the native CaHms1p (Fig 5). Chimeric proteins containing only the first helix or only the loop from CaHms1p showed little if any discrimination. Therefore, from these experiments we conclude that residues within the first helix combined with residues in the loop region of CaHms1p confer the ability to bind specifically to the non-palindromic sequence.
A major difference between branches 2 and 3 of the fungal SREBPs (these branches are represented by CaHms1p and CaTye7p, respectively) is their intrinsic ability to discriminate between the canonical, palindromic E-box (core motif 5’-CANNTG-3’) and the non-palindromic DNA sequence 5’-ATCANNCCA-3’. While Hms1p and related proteins exhibited a strong preference for the latter (Fig 4B and 4C), Tye7p showed a preference, albeit slight, for the former (S5C and S5D Fig). The presence of a tyrosine residue in the DNA binding domain of the SREBPs (instead of a conserved arginine in other bHLH proteins) allows the promiscuous binding to either DNA sequence [24]. But, how did the preference to bind one or the other sequence come about? At least two scenarios could be envisioned. First, an ancestor that had a clear preference for one of the two sequences could have given rise to a lineage that reduced (and eventually flipped) its DNA binding preference. Alternatively, an ancestor that bound both sequences equally well could have given rise to one branch that tilted its DNA binding preference in one direction and another branch whose DNA binding preference tilted in the opposite direction. To empirically test these models, we used ancestral protein reconstruction [14, 36, 37].
We inferred the amino acid sequences of the putative ancestors at two selected nodes of the fungal SREBP phylogeny (Figs 1B and 6A and S6 Fig), then expressed and purified these ancestral proteins. We first assayed the ability of these proteins to recognize the canonical E-box and non-palindromic Hms1p binding sequence in gel shift assays. While a higher amount of ancestor protein was required to bind to the DNA sequences and produce a “shift,” the protein-DNA interactions were still sequence-specific (S6C Fig). We then carried out in vitro competition assays to determine the preference of the purified ancestral proteins for either the palindromic or the non-palindromic sequence. As shown in Fig 6B and 6C, the “oldest” ancestor (Anc5) displayed a slight preference for the non-palindromic sequence. Similarly, the more “recent” ancestor (Anc4) also showed preference for the same sequence (non-palindromic site over the canonical E-box). However, the preference exhibited by CaHms1p towards the non-palindromic sequence is still about an order of magnitude higher than what is exhibited by both ancestors (Fig 6C).
These findings support the notion that two extant branches of fungal SREBPs, which are represented by Hms1p and Tye7p, followed divergent paths after separating from their last common ancestor: The Hms1 lineage enhanced the ancestor’s initial preference for the non-palindromic sequence whereas the Tye7 lineage reduced, and eventually flipped, the DNA-binding preference of the ancestor.
The fact that the C. albicans genome harbors three SREBPs—whereas most other organisms have only one or two—raises the question of what functions they perform in this particular fungus. The most prevalent function associated with SREBPs in fungi is the regulation of ergosterol biosynthesis; however, this role was taken over by the unrelated protein Upc2p in the lineage leading to C. albicans [26]. Each one of the C. albicans SREBPs seems to play a critical role in the biology of the fungus in the mammalian host because strains deleted for any single SREBP gene have reduced fitness in murine models of Candida colonization [27–29]. While CaTye7p has been implicated in glycolysis and sugar metabolism [27], the function(s) of CaHms1p and CaCph2p remain(s) less clear.
The SREBPs in other species and the C. albicans SREBP Tye7 regulate cellular processes sensitive to oxygen [27, 38, 39]. We reasoned, then, that the other two C. albicans SREBPs might play a role when the fungus proliferates in a niche largely devoid of oxygen. To identify the repertoire of target genes regulated by HMS1 and CPH2, we performed transcriptome analyses (RNA sequencing) of the wild-type reference strain and isogenic cph2 or hms1 deletion mutants grown in an anaerobic chamber at 37°C (the temperature of the mammalian host) (Fig 7).
Overall, the RNA-Seq experiment revealed 202 and 192 protein-coding transcripts whose expression was dependent on CPH2 and HMS1, respectively (-log10 P > 10 and expression changes >2-fold; Fig 7A, S3 Table; 685 and 235 targets at P < 0.001 and expression changes >2-fold). There was a significant overlap in targets of regulation between CPH2 and HMS1 (P = 1.36 × 10−99) (Fig 7B and S7 Fig) implying that these two SREBPs form a regulatory cascade. Consistent with this idea, we found that HMS1 expression is dependent on CPH2 (but not vice versa) (Fig 7C). Cph2p binds in vivo to the intergenic region upstream of HMS1 (Fig 2D and [34]) further supporting a direct regulatory link between these two factors. Gene Ontology analysis of the differentially expressed genes revealed filamentous growth (P = 1.7 × 10−4), pathogenesis (P = 1.44 × 10−7) and biofilm formation (P = 1.82 × 10−13) as cellular processes or functions enriched in the dataset (it should be noted that over 50% of the genes in the dataset are annotated as having unknown functions). Indeed, the transcript levels of several well-established regulators of yeast-to-filament transition, e.g. CZF1, GRF10 and ACE2 [40–42], appeared to be under control of both HMS1 and CPH2. The direction of the change in expression in CZF1, GRF10 and ACE2 suggested that, under the environmental condition evaluated, both HMS1 and CPH2 would work by preventing filamentation. In other growth conditions, HMS1 and CPH2 have been associated with the opposite phenotype (i.e. promoting the yeast-to-filament transition) [43, 44].
To establish whether indeed HMS1 and CPH2 work as predicted by our RNA-Seq experiment, we examined the morphology of both deletion mutant strains under anaerobic conditions at 37°C. We found that the hms1 deletion mutant as well as the cph2 mutant strain formed filaments while the wild-type reference strain did not (Fig 7D). We have recently demonstrated that filamentation in C. albicans is detrimental for intestinal colonization [45]. Since the C. albicans hms1 or cph2 deletion mutant strains are impaired in their ability to persist in the murine gut [28, 29], our results suggest that HMS1 and CPH2 may promote gut colonization, at least in part, by preventing the yeast-to-filament morphology transition (Fig 7E).
In this study, we have explored the mechanisms driving the diversification of a eukaryotic transcription regulator family, the SREBPs. In the ascomycete yeasts, the genomes of several Candida species encode three SREBPs. Previous work has shown that in this group of fungi, transcription of ergosterol biosynthesis genes—the main function associated with the family in most organisms—is regulated by proteins unrelated to the SREBPs. These observations implied that the family diversified their function in the ascomycete yeasts, i.e. that the proteins adopted other genes and cellular processes as main targets of regulation. We report that concomitant with a diversification of the cellular functions governed by the SREBPs, these proteins underwent significant changes in their DNA binding specificities. Several lines of evidence support this statement. First, phylogenetic reconstruction of the SREBP family based on the DNA binding domain of the proteins revealed that each one of the three Candida SREBPs belongs to a different branch of the family tree (Fig 1), a pattern consistent with these three proteins being non-redundant. Second, the three Candida SREBPs displayed, to a significant extent, non-overlapping patterns of binding to a comprehensive library of DNA sequences (Fig 2). And, third, only one of the three SREBPs in Candida bound to the palindromic E-box motif which is recognized by most bHLH proteins (Fig 2); in contrast, the Candida Hms1p branch exhibited a strong preference for a non-palindromic DNA sequence whereas the third Candida SREBP, Cph2p, bound to a sequence consisting of only a half-site motif but likely in conjunction with a co-factor (Figs 2 and 3). The SREBPs played a key role in the regulation of a morphological switch (Fig 7) or in sugar metabolism in C. albicans [27]; therefore, the diversification in DNA binding specificities appears to be central to the SREBPs’ expansion in targets of regulation in the lineage leading to Candida.
The archetype and most studied member of the SREBP family, the human SREBP1, exhibits dual DNA binding specificity in in vitro DNA binding assays: It can bind the palindromic E-box (5’-CANNTG-3’) generally recognized by bHLH proteins as well as a non-palindromic sequence (5’-TCANNCCA-3’) [23]. The protein, however, appears to preferentially bind in vivo to the palindromic E-box as revealed by ChIP experiments [25]. Our results indicate that the branch of fungal SREBPs represented by the C. albicans Tye7 protein shares these same DNA binding features with the human SREBP1. That is, the C. albicans protein displayed the same dual DNA binding specificity in in vitro DNA binding assays (S5C and S5D Fig) and also bound in vivo preferentially to a palindromic E-box variant (Fig 2, S8 Fig and [27]). Their similarities in DNA binding profile are in stark contrast to the divergent cellular functions that they govern: While the human SREBP1 regulates the expression of sterol biosynthesis genes, the Tye7 protein controls the expression of sugar acquisition and sugar metabolism genes [27] (Fig 7E). Furthermore, the former harbors the transmembrane domains that are a feature of the family [19] whereas the Tye7 proteins have no traces of any transmembrane domain in their sequences (Fig 1). Thus, the Tye7 branch of fungal SREBPs shares the human SREBP1’s DNA binding features despite the distinct roles that each protein plays in their organisms.
The dual DNA binding ability of the human SREBP1 has been traced back to a tyrosine residue in the DNA binding domain of the protein [23]. Most bHLH proteins have a conserved arginine in this position (instead of the tyrosine) (Fig 1A). The arginine residue forms a stabilizing salt bridge with a conserved glutamate nearby; such structure underlies, at least in part, the protein-DNA contacts with the canonical E-box [24]. This salt bridge cannot be formed when the tyrosine is present, conferring conformational plasticity to accommodate protein-DNA contacts with the non-palindromic sequence (5’-TCANNCCA-3’) besides the palindromic E-box [24]. All the proteins included in our study (Fig 1) harbor the tyrosine residue characteristic of the SREBPs. Yet in contrast to the DNA binding patterns displayed by the C. albicans Tye7p and human SREBP1, the branch (or branches) of the fungal SREBPs represented by the C. albicans and C. parapsilosis Hms1p and the A. fumigatus SrbAp exhibited a marked preference, both in vitro and in vivo, for the alternative, non-palindromic sequence (5’-TCANNCCA-3’). These results suggest that the tyrosine residue that is the hallmark of SREBPs enables alternate binding specificity in addition to dual DNA binding (the latter is what has been reported in the human SREBP1).
Our DNA binding assays with purified CaHms1 and AfSrbA proteins demonstrate that their preference to bind the non-palindromic sequence (5’-TCANNCCA-3’) over the palindromic E-box is an intrinsic property of the proteins (Fig 4 and S8 Fig). Amino acid residues in the first helix and the loop region of the CaHms1p were necessary to confer the specificity towards the non-palindromic sequence (Fig 5). Based on crystal structures of various bHLH proteins, the amino acid residues making direct contact with DNA are located within the basic region and first helix of the DNA binding domain [21, 24, 46, 47]: A glutamine and an arginine residues in the first helix and a histidine in the basic region make direct contacts with the bases that comprise the E-box [47]. These three amino acids are fully conserved throughout the fungal SREBPs included in our study (S1 Table). Thus, the changes in DNA binding specificity that we identify in the SREBP family cannot be due to variation in any of these positions. In bHLH regulators such as the S. cerevisiae Pho4, residues at the boundaries of the loop region (i.e. towards the end of the first and beginning of the second helices) are known to interact to stabilize the overall structure [47]. We speculate that at least some of the amino acids underlying the change in DNA specificity in Hms1p may, similarly, be involved in “stabilizing” the structure rather than in making direct contacts with DNA. The fact that the first helix and the loop region were necessary imply that more than one single amino acid change is responsible for the switch in specificity. This finding is consistent with the observation that cumulative amino acid changes—which often must occur in a particular order—are usually responsible for the modifications in protein function that occur during evolution [48–50]. Furthermore, these results suggest that “disordered” regions of a protein’s DNA binding domain, such as the loop region in bHLH proteins, may also influence DNA binding specificity.
In contrast to CaHms1p and AfSrbAp, it is apparent that for other SREBPs, such as CaTye7p and human SREBP1, additional factors may contribute to their in vivo DNA binding specificity. We speculate that binding to target promoters with co-factors may be one such determinant. It has been shown, for example, that the human SREBP1 cooperates in vivo extensively with the co-factors NFY and SP1 [25]. The C. albicans Tye7 protein has also been shown to bind to many promoters together with another protein, Gal4 [27]; indeed, both regulators Tye7 and Gal4 are needed to control the expression of glycolysis genes in this species [27, 51].
In Yarrowia lipolytica, a species that lies at the very base of the Saccharomycotina, the SREBP YlSre1 has been shown to be required for switching from yeast to filamentous growth in hypoxia [26]. The data described in this report indicates that two of the three C. albicans SREBPs regulate the same morphological switch in anaerobic conditions although in the opposite direction: CaHms1 and CaCph2 were needed to prevent Candida from switching from yeast to filamentous form under these conditions (Fig 7). Thus, the connection of SREBPs to fungal morphology regulation appears to have been maintained throughout the Saccharomycotina evolution. CaHms1p and CaCph2p form a regulatory cascade through which the gene encoding the former protein is a direct target of regulation of the latter. Cph2 is the only SREBP in C. albicans that contains transmembrane domains, a distinctive feature of the family. It is plausible that by being inserted in intracellular membranes, the activity of Cph2p can be modulated by stimuli related to those that regulate the prototypical SREBPs. Hms1p, on the other hand, lacks the transmembrane domains; hence, the activity of this protein most likely responds to different intra- or extra-cellular signals. The Cph2-Hms1 regulatory cascade can, thus, expand the repertoire of stimuli that feed into the circuit to control yeast-to-filament transition [1, 52, 53].
In sum, we have shown that the fungal SREBPs comprise several branches that differ from one another in their DNA binding preferences and in the biological processes that they regulate. A key element in the diversification of the family appears to be the intrinsic structure of the DNA binding domain of the SREBPs which allows these proteins to adopt two distinct conformations and therefore recognize at least two different DNA sequences. Our findings suggest that this promiscuous state was resolved during evolution of the family: One branch tilted the preference towards one of the DNA motifs largely through amino acid changes in the same protein whereas another branch tilted the preference towards the second DNA motif. We posit that the diversification in their DNA binding preferences enabled the SREBPs to expand and regulate diverse cellular processes in fungi.
The standard and systematic names of the main SREBP genes included in this study are as follows: TYE7 (ORF19.4941 or C1_13140C_A in C. albicans; YOR344C in S. cerevisiae); CPH2 (ORF19.1187 or C6_00280W_A in C. albicans; YOR032C in S. cerevisiae); HMS1 (ORF19.921 or C5_00670C_A in C. albicans). Notice that although the standard name of the S. cerevisiae gene YOR032C in the Saccharomyces genome database is HMS1, our phylogenetic reconstruction places it closer to the C. albicans CPH2 branch (Fig 1B).
All C. albicans strains used in this study are listed in S4 Table and are derivatives of the clinical isolate SC5314 [54]. For the construction of the epitope-tagged strain Cph2-MYC, which was used in ChIP experiments, a DNA fragment encoding 13× MYC followed by the SAT1/flipper cassette was amplified by PCR from plasmid pADH34 [55] with oligos described in S5 Table and integrated in the CPH2 (ORF19.1187) locus. This construct effectively truncates one of the CPH2 alleles at codon 407 and inserts the MYC tag at this position. The SAT1 cassette was then removed as described [56]. DNA sequencing and Western blot analysis confirmed the correct insertion of the tag and the expression of the tagged protein at the expected size, respectively.
All plasmids used for recombinant protein expression are listed in S6 Table. The putative DNA binding domains of CaCph2 (amino acids 197–302), CaHms1 (amino acids 463–686), CaTye7 (amino acids 121–269), CpHms1 (amino acids 486–659) and AfSrbA (amino acids 145–266) were amplified from genomic DNA of each species and introduced into plasmids pLIC-H3 [57] and pbRZ75 [58] (both derivatives of pET28b). These plasmids were designed to produce recombinant N-terminal 6×His or 6×His-MBP (maltose binding protein) tagged proteins.
Chimeric proteins were constructed by (1) replacing residues 211–232 from CaCph2 by residues 489–510 from CaHms1 to generate chimeric helix 1 protein; (2) replacing residues 255–281 from CaCph2 by residues 625–651 from CaHms1 to generate chimeric helix 2 protein; (3) replacing residues 211–255 from CaCph2 by residues 489–629 from CaHms1 to generate chimeric helix1-loop protein; and (4) replacing the loop of CaCph2 (236–255) by the loop of CaHms1 (506–625) to generate the chimeric loop protein.
The DNA fragments encoding the reconstructed ancestral proteins were generated by gene synthesis (Invitrogen GeneArt Gene Synthesis). These fragments included restriction sites for cloning into pLIC-H3 [57].
E. coli BL21 was used as the host of the expression plasmids. For recombinant protein overexpression, bacterial cells were grown to an OD600 of approximately 0.8 and induced with 0.5 mM IPTG. Cultures were grown further for 3 hours, pelleted and stored at -80°C. Cells were lysed by sonication. His-tagged proteins were affinity purified from the lysate using Ni-NTA agarose beads (Qiagen). Amicon Ultra-15 centrifugal filters (Merck) (10 or 30K membranes depending on protein size) were used to exchange buffer and concentrate the proteins. Protein concentration was estimated in Rothi‐blue (Carl Roth, Germany) stained gels using known amounts of bovine serum albumin as standards.
EMSAs were carried out as described [58]. Competition assays with unlabeled DNAs were performed as reported [13]. Gel shift assays with fluorescently-labeled DNA sequences (shown in S4 Fig) were conducted in a similar fashion to those with radiolabeled DNA except that larger amounts of Cy5-labeled non-palindromic and Cy3-labeled palindromic DNA fragments, alone or together, were incubated with a fixed amount of protein.
MITOMI experiments were carried out essentially as described [13, 32]. Briefly, the DNA binding domains of C. albicans Hms1p (amino acids 463–685), Cph2p (amino acids 197–302) and Tye7p (amino acids 159–269) tagged with GFP were generated with an in vitro transcription-translation system (Promega) and added to a microfluidics device containing the Cy5-labeled DNA library. All experiments used a 740-oligonucleotide pseudorandom DNA library containing all possible 8-nucleotide sequences (S7 Table). The library was designed to minimize similarities between k-mers represented on a given strand and thereby reduce the chance of multiple binding sites. Protein-DNA interactions were trapped at equilibrium. After a series of washings where unbound DNA and proteins were washed out, the GFP/Cy5 intensity ratio was measured in every chamber of the device. Experiments were performed in duplicates. Cytoscape (v3.4) [59] was used to visualize the data. MatrixREDUCE [33] was used to search overrepresented DNA motifs. The model variants (topologies) X6, X7, X3N2X3 and X4N2X3 (in forward, reverse or both strands), as implemented in MatrixREDUCE, were employed to query the datasets. The DNA motifs with P < 1 × 10−10 were ranked according to their r2 and P-values, which were calculated by the same software.
Fungal protein sequences were retrieved from UniProt [60]. The maximum likelihood tree was constructed by aligning the basic region, first helix and second helix of the DNA binding domain of 198 fungal SREBPs in MEGA7 [61]. Due to its variability in sequence and length, only the last five residues of the loop were taken into account, the rest of the loop region was omitted from the alignment. Only proteins carrying the characteristic tyrosine residue in the first helix (a defining feature of the SREBP family) were included in the analysis. ProtTest [62] was used to find the best-fit model to infer the phylogenetic tree (LG+G). The presence of transmembrane domains was predicted with OCTOPUS [63].
Phylobot [37] was used to infer ancestral protein sequences at specific nodes of the phylogenetic tree. The model used to infer the tree was PROTGAMMALG (tested for best-fit model with ProtTest) for all cases. The reconstructed protein sequence for Anc4 exhibited low levels of uncertainty. For Anc5, due to higher levels of sequence uncertainty, we carried out two reconstructions: (1) Using Pho4 as the only outside sequence; and (2) including mouse and human SREBP sequences in addition to Pho4 as outside sequences. Two alignment models, MUSCLE and msaprobs, were considered. Eight versions of Anc5, which differed from one another at residues in the first helix, were synthesized and their overall ability to bind to DNA was evaluated by EMSAs. Due to the high variability in the loop segment, this particular portion of the DNA binding domain could be neither properly aligned nor reconstructed. Given its short size (55 amino acid residues), the corresponding amino acid sequence of CaCph2p was used to fill the loop segment in all ancestral proteins.
MYC-tagged and untagged C. albicans strains (the latter served as a negative control) were grown in YPD broth at 30°C until mid-log phase. ChIP was carried out as described [55] with the following modifications: Input and immunoprecipitated DNAs were directly used to generate libraries for sequencing with the NEBNext ChIP-Seq Library Prep Master Mix Set for Illumina (New England Biosciences). DNA sequencing was carried out by GATC Biotech (Konstanz, Germany) using standard procedures. The reads were aligned to the C. albicans genome using Bowtie2 [64] with default parameters. Between 3–10 million reads per sample were uniquely aligned to the genome. Peak calling and visualization were performed with MACS2 [65] (using default parameters) and MochiView [66], respectively. To ensure the generation of a high confidence dataset, in addition to the standard computational analyses we manually curated all the extracted peaks using the following criteria: (1) Peaks that appeared in both the untagged control and the Cph2-MYC tagged strain were removed; (2) peaks located within annotated ORFs were ignored; (3) peaks located around highly expressed genes (particularly ribosomal genes) were also discarded because based on previous experience (e.g. [29, 35, 67, 68]) these places tend to bind to almost all DNA binding proteins non-specifically; and (4) only peaks that appeared significant in the MACS2 analysis in at least two of three replicates were taken into account. Motif finding analysis was performed with MochiView by providing 500 nt DNA sequences surrounding the high-confidence peaks using the software’s default parameters.
C. albicans reference strain, hms1 and cph2 deletion mutants were grown in Todd-Hewitt broth in an anaerobic chamber at 37°C for 24 hours. The culture medium had been placed in the anaerobic chamber at least two days before inoculation to remove any oxygen traces. Two independent replicates were used for the analysis. Total RNA extraction and cDNA synthesis was performed as described [69]. Quality control, mapping and differential gene expression was carried out as reported [45]. We obtained 63–91 million reads per sample which were then aligned to the C. albicans genome using STAR v2.5.2b [70] with default parameters (>97% of reads of each sample were uniquely aligned to the C. albicans genome). Read counts were loaded into R (v3.3.2) and analyzed with the DESeq2 [71] package (v1.14.1). With our depth of sequencing, significant numbers of reads were detected for ~6,100 annotated ORFs. Cytoscape [59] (v3.4) was used to visualize the data and generate the network graphs.
Overnight C. albicans cultures (in Todd-Hewitt broth at 30°C) were diluted to OD600 ~0.1 in fresh Todd-Hewitt broth and incubated at 37°C under anaerobic conditions for 24 hours. The medium used to dilute the overnight culture had been pre-incubated in an anaerobic chamber for 48 hours to achieve complete anaerobiosis. After the 24-hour period of growth, cells were washed with sterile PBS and fixed in glass slides for morphology evaluation under the microscope.
Reference strain, cph2 and efg1 deletion mutants were grown under anaerobic conditions as described above. Total RNA purification and cDNA synthesis were performed as described [69]. Real time PCR was used to quantify specific transcripts (oligos listed in S5 Table). The experimentally validated TAF10 transcript [72] served as a reference control for the qPCR.
The significance of the overlap between the differentially expressed genes from our RNA-seq datasets was estimated using the hypergeometric test. The Gene Ontology Term Finder of the Candida Genome Database (www.candidagenome.org) was used to identify enriched processes in our RNA-seq dataset. The student t-test for unpaired samples was used to assess statistical differences between transcript levels.
The ChIP-Seq and RNA-Seq data reported in this article have been deposited in the NCBI Gene Expression Omnibus (GEO) database under accession numbers GSE118419, GSE118416 (ChIP-Seq) and GSE118414 (RNA-Seq).
|
10.1371/journal.ppat.1004434 | Sensing of Immature Particles Produced by Dengue Virus Infected Cells Induces an Antiviral Response by Plasmacytoid Dendritic Cells | Dengue virus (DENV) is the leading cause of mosquito-borne viral illness and death in humans. Like many viruses, DENV has evolved potent mechanisms that abolish the antiviral response within infected cells. Nevertheless, several in vivo studies have demonstrated a key role of the innate immune response in controlling DENV infection and disease progression. Here, we report that sensing of DENV infected cells by plasmacytoid dendritic cells (pDCs) triggers a robust TLR7-dependent production of IFNα, concomitant with additional antiviral responses, including inflammatory cytokine secretion and pDC maturation. We demonstrate that unlike the efficient cell-free transmission of viral infectivity, pDC activation depends on cell-to-cell contact, a feature observed for various cell types and primary cells infected by DENV, as well as West Nile virus, another member of the Flavivirus genus. We show that the sensing of DENV infected cells by pDCs requires viral envelope protein-dependent secretion and transmission of viral RNA. Consistently with the cell-to-cell sensing-dependent pDC activation, we found that DENV structural components are clustered at the interface between pDCs and infected cells. The actin cytoskeleton is pivotal for both this clustering at the contacts and pDC activation, suggesting that this structural network likely contributes to the transmission of viral components to the pDCs. Due to an evolutionarily conserved suboptimal cleavage of the precursor membrane protein (prM), DENV infected cells release uncleaved prM containing-immature particles, which are deficient for membrane fusion function. We demonstrate that cells releasing immature particles trigger pDC IFN response more potently than cells producing fusion-competent mature virus. Altogether, our results imply that immature particles, as a carrier to endolysosome-localized TLR7 sensor, may contribute to regulate the progression of dengue disease by eliciting a strong innate response.
| Viral recognition by the host often triggers an antiviral state, which suppresses viral spread and imparts adaptive immunity. Like many viruses, dengue virus (DENV) defeats the host-sensing pathway within infected cells. However, in vivo studies have demonstrated a key role of innate immunity in controlling DENV infection. Here we report that sensing of DENV-infected cells by non-permissive innate immune cells, the plasmacytoid dendritic cells (pDCs), triggers a cell-contact- and TLR7-dependent activation of a strong antiviral IFN response. This cell-to-cell sensing involves transmission of viral elements that are clustered at the interface between pDCs and infected cells and is regulated by the actin network. Importantly, we revealed that uncleaved prM surface protein-containing immature particles play a key function in stimulating the innate immune response. These non-infectious immature particles are released by infected cells as a consequence of a suboptimal cleavage site, which is an evolutionarily conserved viral feature that likely favors the export of infectious virus by prevention of premature membrane fusion in the secretory pathway. Therefore our results highlight a conceptually novel trade-off between efficient infectious virus release and the production of IFN-inducing particles. This concept may have broad importance for the many viruses that, like DENV, can disable the pathogen-sensing machinery within infected cells and can release uncleaved glycoprotein-containing non-infectious particles.
| The innate immune system acts as the first line of defense for the sensing of viral infection. This involves rapid recognition of pathogen-associated molecular patterns (PAMPs), including viral nucleic acids, by pattern recognition receptors (PRRs). This recognition results in an antiviral response characterized by the production of type I interferons (IFNs) and expression of IFN-stimulated genes (ISGs). This response suppresses viral spread by blocking the viral life cycle at multiple levels and also mediates immunomodulatory effects in surrounding tissues that impart the onset of the adaptive immune response [1]. The PRR can be cytoplasmic, e.g., retinoic inducible gene-I (RIG-I)-like receptors (RLRs) and NOD-like receptors (NLRs), or endosomal, e.g., Toll-like receptors (TLRs) [1]. Thus, depending on their intracellular localization, virus-induced innate immune signaling typically occurs within cells that are either productively infected or that have internalized viral particles [1], [2].
Recent studies illustrated the existence of alternative host sensing strategies by bystander plasmacytoid dendritic cells (pDCs), which recognize infected cells [3], [4], [5], [6], [7]. pDCs are immune cells known to function as sentinels of viral infection and are a major type I IFN-producing cell type in vivo [8], [9]. Using hepatitis C virus (HCV) as a model, we recently demonstrated that HCV infected cells can selectively package immunostimulatory viral RNA within exosomes that deliver their RNA cargo to pDCs, which, in turn, produce IFNα [3]. Exosomes also permit transfer to pDCs of distinct immunostimulatory viral RNAs, such as those of the negative strand lymphocytic choriomeningitis virus (LCMV) [4]. This sensing pathway is thought to assure recognition of infected cells and hence protects the host against viruses that defeat the pathogen-sensing machinery within the cells they infect.
Virtually all viruses have evolved strategies that preclude antiviral signaling in the cell they infect [10]. For example, dengue virus (DENV) has evolved several evasion strategies that prevent IFN and ISG expression within infected cells [11]. Notably, the DENV NS2B-3 protease complex, by cleavage and degradation of an adapter of the cytoplasmic sensor-mediated signaling (STING, also called MITA) and by preventing phosphorylation and nuclear translocation of the downstream transcriptional factor, IFN regulatory 3 (IRF3), inhibits type I IFN production in DENV infected cells [12], [13], [14], [15]. Despite these potent inhibitory mechanisms, expression of antiviral and inflammatory molecules is readily detected in DENV infected humans [16], [17]. Their levels play a pivotal role in DENV infection clearance and pathogenicity [16], [18], [19], thus highlighting the importance of elucidating the host sensing mechanisms leading to the IFN response during DENV infection.
Here, we showed that pDCs are robust IFNα producer cells in response to DENV infected cells. In addition, we demonstrated that cell-to-cell contact- and TLR7-dependent pDC responsiveness leads to an antiviral state, inflammatory cytokine production as well as expression of co-stimulatory molecules by pDCs. Newly formed particles of DENV, like many viruses, undergo maturation by cleavage of the virus envelope protein, premembrane (prM), in the secretory pathway that renders the virus infectious [20]. Yet, the prM cleavage site is suboptimal, leading to the secretion of about 30–40% immature, prM-bearing particles [21], [22], [23], [24], [25], [26]. This evolutionarily conserved suboptimal site may be critical for the export of the infectious viral particles and/or may also positively contribute to viral infection by usurpation of humoral immune response, because anti-prM antibodies facilitate efficient binding and cell entry of prM-containing immature particles into Fc-receptor-expressing cells, a process called antibody dependent enhancement (ADE) [21], [22], [23], [27], [28]. Here, we report a previously unsuspected function of immature particles in innate immunity. Although the immature particles are not infectious, they are fully competent to trigger a robust type I IFN response by contacting non-permissive pDCs. Our results highlight the trade-off between efficient secretion of infectious viral particles and the production of a large amount of IFN-inducing immature particles.
To investigate the mechanisms regulating the IFN response against DENV infection, primary human peripheral blood mononuclear cells (PBMCs) from healthy donors were exposed to supernatants containing DENV virions or DENV infected cells. We found that PBMCs specifically responded to co-cultivation with DENV infected cells but not to uninfected Huh7.5.1 cells, by a robust secretion of IFNα (Figure 1A). In sharp contrast, supernatants from the DENV infected cells failed to trigger IFNα secretion by PBMCs (Figure 1A). Plasmacytoid dendritic cells (pDCs), which represent a rare PBMC population, i.e. 0.41% of PBMCs (Figure 1B, upper panel), are known to produce IFNα [9]. Antibody-mediated pDC depletion from PBMCs (Figure 1B, middle panel) abolished IFNα secretion in response to co-culture with DENV infected cells (Figure 1A). Similar results were also obtained using DENV infected BHK-21 cells (Figures S1A and S1B). To rule out potential non-specific effects of the depletion procedure on innate cell responsiveness, we verified that IL-6 production triggered by lipopolysaccharide (LPS) exposure was maintained after pDC depletion (Figures 1C and S1C). Consistent with the depletion results, the isolated pDC population (Figure 1B, lower panel) potently produced IFNα in response to co-culture with DENV infected cells, but not in the presence of their supernatants (Figure 1A). A very limited number of pDCs (i.e., 2,000 pDCs) was sufficient to produce a robust secretion of IFNα (Figure 1A). Similar levels of IFNα production were detected after co-culture of infected cells with isolated pDCs as compared to total PBMCs, which contained a similar number of pDCs (Figure 1D), further suggesting that pDCs are the main IFNα producer cells among PBMCs. We showed that the cells productively infected with DENV did not produce IFNα themselves (Figure S2A). The pDC IFNα response increased as the duration of infection and, thus the replication levels, prior to co-culture increased (Figure S2A). Remarkably, similar levels of IFNα secretion were reproducibly obtained with pDCs isolated from the blood of a cohort of 20 healthy donors (Figure 1E). Together these results suggest that pDCs represent the main cell type in PBMC populations that produce IFN in response to co-cultivation with DENV infected cells and that this response was not induced by the addition of cell-free supernatants containing virus.
To exclude the possibility that pDCs respond transiently to supernatants containing DENV, we quantified IFNα secretion in time course experiments. IFNα secretion was already detectable as early as 4 hours after co-cultivation of pDCs and DENV infected cells (Figure 2A). IFNα levels concurrently increased over the time course of co-culture of DENV infected cells with either pDCs or PBMCs, and reached levels around 100 ng/mL after 16 hours of co-culture (Figure 2A). In contrast, cell-free supernatants containing DENV did not trigger detectable IFNα production by pDCs or by PBMCs at any of time points analyzed (Figure 2A). IFNα producer cells were markedly enriched in pDCs, characterized as a CD123-positive population, as compared to the CD123-negative population (Figure 2B). For example, 12 hours after co-culture of DENV infected cells with PBMCs, ≈0.05% and ≈25–30% of CD123-negative and –positive cells, respectively, were IFNα positive (Figure 2B). Consistently, the frequencies of IFNα producer cells (i.e., about 30%) among pDCs (i.e., CD123-positive populations) were comparable in co-cultures of DENV infected cells with PBMCs vs. isolated pDCs (Figure 2B). Together these results suggested that IFNα is robustly produced only by pDCs that are co-cultured with DENV infected cells.
Next, we showed that co-cultivation of DENV infected primary cells, i.e., monocyte-derived macrophages (mo-M) and monocyte-derived dendritic cells (mo-DC) with pDCs (isolated from the same donor), potently triggered pDC IFNα secretion (Figures 3A and 3B). This stood in stark contrast to the corresponding cell-free supernatants containing virus or the parental uninfected cells did not, or very weakly, induced pDC IFNα production (Figures 3A and 3B). Consistent with the previously reported inhibition of type I IFN production by the DENV NS2B-3 protease in infected cells [12], [13], [14], [15], DENV infected primary cells did not produced detectable levels of IFNα (Figures 3A and 3B).
Additionally, we determined if the production of IFNα by pDCs could be reproduced in response to co-culture with various cell types infected by DENV. Robust secretion of IFNα was triggered when pDCs were co-cultivated with DENV infected cell lines from different origins (i.e., human Huh7.5.1, Hela and 293T cells or non-human BHK-21 and Vero cells), but not by the corresponding supernatants containing virus or the uninfected cells (Figure 3C). DENV infected Vero cells were weaker IFNα inducers (Figure 3C), consistent with lower levels of intracellular DENV RNA (Figure 3D) and infectious viral particle (Figure 3E) produced by these cells, suggesting that pDC IFNα induction is proportional, to some degree, with the level of viral replication.
Remarkably, 293T cells infected by another member of the Flavivirus genus, West Nile virus (WNV), but not the corresponding cell-free supernatants containing virus, also triggered robust IFNα production when co-cultured with pDCs (Figure 3C). Similar to the results obtained using co-cultures with DENV infected cells, the pDC IFNα responses increased as the numbers of WNV infected cells increased (Figures S2B and S2C). Together, these results demonstrated that the production of IFNα by pDCs in response to co-culture with DENV infected cells is not cell type specific and that pDCs similarly respond to another member of the Flavivirus genus.
Cell-free supernatants containing virus from various infected cell types failed to trigger pDC IFNα production, even when added as crude non-filtered supernatants containing virus at concentrations as high as 20 infectious units per pDC (Figure S3), indicating that the transmission of the immunostimulatory signal to pDCs likely requires cell-to-cell contacts. To determine if contacts with DENV infected cells favors pDC sensing, we assessed IFNα production by pDCs cultured in transwell chambers with infected cells. Transwell cultures containing DENV infected monocyte-derived dendritic cells (mo-DCs) and pDCs separated by a 0.4 µm permeable membrane did not result in detectable levels of IFNα production by the pDCs (Figure 4A). Similar results were obtained using DENV infected Huh7.5.1, BHK-21, Hela and Vero cells as well as WNV infected cells (Figure 4B), confirming that this feature is not cell type specific or restricted to DENV. Similarly to IFNα, pDCs robustly produced IFNβ when in contact with DENV infected cells, but not when cells were physically separated by a transwell membrane (Figure 4D). Consistent with these results, IFNβ production by pDCs was not triggered by supernatants from DENV infected cells and DENV infected cells did not themselves release detectable levels of IFNβ (Figure 4D). In control experiments using identical transwell culture settings, an agonist of TLR7, a viral RNA immune sensor [9], triggered the production of both IFNα and IFNβ by the pDCs at levels similar to those obtained in the co-culture setting (Figure 4C and 4E), thus ruling out potential non-specific effects of the experimental setting on pDC responsiveness. In agreement with previous reports [29], [30], vesicular stomatitis virus (VSV) or Influenza virus (FluAV) containing supernatants robustly triggered IFNα production by pDCs (Figure 4F). Consistent with these results, VSV and FluAV infected cells in contact with pDCs (Figure 4F, cocult) or separated by a transwell membrane (Figure 4F, TW), triggered IFNα production at similar levels. This suggested that contact with virus infected cells is not a universally employed mechanism to promote pDC activation by RNA viruses. Next, viral transmission across the transwell-membrane was assessed by quantifying infectious DENV (Figure 4G) and WNV (Figure 4H) on both sides of the membrane that separated infected cells from recipient cells. To evaluate the possible interference of recipient cells on the extracellular infectivity detection, we compared two types of recipient cells, i.e., IFNα response-competent pDCs, which are non-permissive to infection (Figure S4) and permissive cells (Figure 4G and 4H, Naïve recipient cells). As expected from their size, infectious viral particles readily flowed across the 0.4 µm membrane (Figures 4G and 4H), thereby permitting viral transmission from infected cells to naïve cells in the absence of direct contact (Figures 4I and 4J). In sharp contrast, type I IFN production by the pDCs was induced exclusively under conditions where cell-to-cell contact was possible between infected cells and pDCs (Figures 4A, 4B and 4D).
Collectively, these results demonstrated that the exposure of pDCs to the DENV or WNV infected cell milieu either at defined time points (Figure 3A–C) or continuously (Figures 4A, 4B and 4D) failed to trigger a robust IFN response by pDCs, which were responsive to infected cells by cell-to-cell contact and/or in a short-range manner.
pDCs typically respond to viral infection via endolysosome-localized TLR7- or TLR9 sensors that recognize RNA or DNA viral genomes, respectively [9]. Accordingly, we examined the transmission of DENV RNAs to co-cultured pDCs. The presence of DENV RNA in infected cells and co-cultured pDCs (selectively labeled with DiI, a fluorescent membrane dye) was assessed using a highly sensitive DENV RNA-specific fluorescence in situ hybridization (FISH) assay (Figure 5A, upper panels). The analyses were performed after 5 hours of co-culture with DENV infected cells, at which time pDCs already produced IFNα (Figure 2A). DENV RNA (green) was detected as discrete dots inside pDCs (Figure 5A, lower panels). Inspection of consecutive Z-axis sections of co-cultures stained by combined DENV RNA FISH and anti-IFNα immuno-detections revealed that the frequency of DENV RNA-positive pDCs was elevated in both IFNα-positive (i.e., 85%) and IFNα-negative pDCs (i.e., 74.5%) (Figure 5A, summary table). The specificity of these examinations was validated by the absence of DENV RNA-positive pDC when co-cultured with uninfected cells and when the FISH procedure was performed in the absence of the DENV RNA specific probe (Figure 5A, summary table and Figure S5). The presence of DENV RNA in IFNα-negative pDCs may reflect the time required for DENV RNA to trigger pDCs to produce enough IFNα to be detectable, which may not have occurred by 5 hours of co-cultivation. Alternatively, differential DENV RNA localization in intracellular compartments may modulate their recognition by innate sensors, and/or potential subsets of pDCs may be differentially responsive to the DENV RNA stimulus, in accordance with the maximal detection of about 30% IFNα-positive pDCs at plateau (Figure 2B). Only a few DENV RNA dots were detected inside pDCs, suggesting that it is a rare event but sufficient to trigger pDC IFN production. Together, these results indicated that DENV RNA was readily transmitted from DENV infected cells to co-cultured pDCs, supporting the notion that DENV RNA might be recognized by pDC TLR7. Accordingly, a TLR7 antagonist significantly inhibited pDC IFNα production induced by DENV infected cells (Figure 5B). The specificity of this TLR7 antagonist was demonstrated by the inhibition of IFNα production induced by a TLR7 agonist (R848) but not by a TLR9 agonist (ODN2216) (Figure 5B). Collectively, these results suggested that DENV infected cells transfer viral RNA to co-cultured pDCs and trigger TLR7-dependent IFNα production.
Next, to further define the nature of the pDC-mediated antiviral state induced by contact with DENV infected cells, we examined the secretion of the inflammatory cytokines, IL-6 and tumor necrosis factor (TNF)-α, triggered by activation of the transcription factor NF-κβ, known to transduce antiviral signaling downstream of TLR7 [1]. TNF-α is known to play a pivotal role in the vascular leakage syndrome, a hallmark of dengue hemorrhagic fever [18]. Sensing of DENV infected cells, but not their supernatants, specifically triggered pDCs to produce IL-6 and TNF-α at levels comparable to those induced by treatment with a TLR7 agonist (Figure 5C). In addition, ISGs (i.e., MxA and ISG56) were specifically up-regulated in co-cultures of DENV infected cells with pDCs or PBMCs (Figure S6), thus indicating the establishment of an antiviral state.
Finally, we determined if DENV infected cells trigger pDC maturation as assessed by the up-regulation of the CD83 and CD86 markers at the cell surface. DENV infected cells, but not their supernatants, triggered a rapid increase in the surface expression of CD83 on co-cultivated pDCs (i.e., in CD123 marker-gated cells) (Figure 5D, left panel), accompanied by a slightly delayed augmentation of CD86 cell surface expression (Figure 5D, left panel) and by a concomitant increase in IFNα secretion (Figure 5D, right panel). Collectively, these results demonstrated that sensing of DENV infected cells by TLR7, a sensor of single stranded-RNA, triggers IFNα production by pDCs, along with the induction of the inflammatory response, an antiviral state and pDC maturation.
To define how pDCs sense DENV infected cells, we analyzed the ability of cells harboring recombinant DENV genomes containing mutations specifying phenotypes deficient in various viral functions to trigger IFNα production by co-cultured pDCs. First, we tested cells containing DENV genomes encoding lethal mutations in the methyltransferase domain of the viral NS5 polymerase (i.e., Rep−/−) [31]. As expected [31], the triple mutation significantly reduced the intracellular level of DENV RNA at 48 hours post-transfection as compared to the wild type (WT) genome (Figure 6A), reflecting a failure to amplify viral RNA (Figure S7A). Consistently, this mutant did not express detectable amounts of intracellular viral proteins (Figure S7B). Despite comparable intracellular viral RNA levels between the DENV WT and Rep−/− mutant genomes at the onset of co-culture i.e., 24 hours post transfection, likely reflecting the input transfected RNA (Figure 6A), Rep−/− DENV mutant genome harboring cells did not trigger IFNα production by co-cultured pDCs (Figure 6D). Similarly, cells harboring DENV genomes encoding a four amino acid deletion in the capsid (i.e., amino acids V51-to-L54), that significantly compromised both viral RNA replication (Figures 6A and S7A) and viral protein expression (Figure S7B), failed to induce IFNα production by co-cultured pDCs (Figure 6D). Together these results indicated that the pDC IFNα response requires active viral replication in neighboring DENV infected cells.
Next, to address the requirement of viral genome release for pDC activation, we tested the effects of co-culture with cells harboring DENV genomes encoding point mutations in the envelope (E) glycoprotein, i.e., the substitutions D215A, H244A or P217A, known to inhibit infectious viral production [32], [33]. Consistent with previous reports [32], [33], the E glycoprotein mutations did not impair intracellular levels of either viral RNAs or proteins (Figures 6A, S7A and S7B), but they all greatly compromised the production of infectious particles (Figure 6B). Both the D215A and H244A mutations abrogated the release of viral RNA and structural proteins and the pDC IFNα response (Figures 6C–D and S7C). Conversely, cells harboring DENV genomes encoding the P217A mutation triggered the IFNα response by pDCs (i.e., ≈36% relative to WT) (Figure 6D) at various inducer cell concentrations (Figure S7D) and in proportion to the release of extracellular DENV RNA (i.e., ≈60% and 26% relative to WT at 24 and 48 hours post-transfection, respectively) (Figure 6C) and viral structural proteins (Figure S7C). Remarkably, the production of infectious virus (Figure 6B) was severely and disproportionally inhibited by the P217A mutation (i.e., 40-to-1,000 fold-reduction at 24-to-48 hours post-transfection) as compared to the modest inhibition of the IFNα response by pDCs (i.e., ≈2.5 fold-reduction of IFNα response in the same time period) (Figure 6D and S7D). These results suggested that infectious virus production is not required and/or is not rate-limiting for pDC activation. Consistently, pDCs were not permissive to DENV infection (Figure S4A), this latter observation is in line with the previous demonstration of pDCs as refractory to infection by other viruses [30], [34], [35].
Altogether, these results suggested that glycoprotein-dependent release of non-infectious viral components by DENV infected cells might trigger the IFNα response by contacting pDCs.
To determine whether DENV surface proteins mediate the transmission of viral components to pDCs, we first assessed whether, similarly to DENV RNA (Figure 5A), the DENV envelope proteins are transmitted into the pDCs, by inspection of consecutive Z-axis sections of DiD-labeled pDCs in co-culture with cells harboring the WT and DENV genomes encoding E protein mutations (Figure S8). Similar to DENV genome, we observed the E glycoproteins (E GP) in dot-like structures inside the pDCs. The frequencies of E GP dot-positive pDCs were elevated when in the co-cultures with either cells harboring the WT genome (i.e., around 90%) or the P217A mutation (i.e., above 65%) (Figures 6E and S8), which was in proportion to the release of extracellular DENV RNA (i.e., 60% relative to WT particles at 24 hours post-transfection) (Figures 6C). In contrast, cells harboring the DENV genome encoding the H244A mutation in E, which do not release viral particles and fail to trigger the IFN response by pDCs (Figures 6B, 6C, 6D, S7C and S7D), demonstrated little to no transmission of the E GP into the pDCs (Figures 6E and S8). Because the intracellular levels of viral components (i.e., viral RNA, E and capsid proteins) were equivalent for cells harboring DENV genomes encoding the H244A point mutant, as compared to WT genome (Figures 6A and S7B), the results suggested that pDC IFNα production is activated by the glycoprotein-mediated transmission of viral components from DENV infected cells into contacting pDCs.
Next, we tested the impact of expressing the DENV surface proteins alone (Figure S9A) on pDC IFN induction. Expression of the envelope proteins alone is known to result, in absence of nucleocapsid, in the release of viral envelope containing-membrane vesicles, the sub-viral particles (SVPs) (Figure S9B) [36]. Although the glycoproteins were readily transmitted from cells expressing only the DENV surface proteins to the co-cultured pDCs (Figure S9D–F), IFNα production was not triggered (Figure S9C). These observations are in agreement with the transmission of DENV RNA to pDCs and activation by the TLR7 RNA sensor (Figure 5A and 5B).
To corroborate these results, we determined whether pDC activation by contact with DENV infected cells requires an internalization-dependent mechanism by testing inhibitors of dynamin (Dynasore) [37], of clathrin-mediated endocytosis (Chlorpromazine [38]) and of macropinocytosis (Gö6983-PKC inhibitor [39], [40]). Inhibitors of both dynamin and clathrin-mediated endocytosis, but not macropinocytosis, abrogated pDC IFNα production triggered by DENV infected cells (Figure S10A), without any effect on the ongoing DENV replication and viral production (Figures S10B–C). In addition, these inhibitors did not markedly impair pDC IFNα production induced by a TLR7 agonist (Figure S10A), a cell-permeable imidazoquinoline, which passively diffuses inside the pDCs [41], thus ruling out potential side-effect downstream of TLR7 recognition. These results, in agreement with the requirement of the endolysosome localized-sensor, TLR7 (Figure 5B), suggested that pDC IFNα production triggered by DENV infected cells requires glycoprotein-mediated secretion of non-infectious viral components, which are subsequently internalized by co-cultured pDCs. These results demonstrated that pDC activation triggered by DENV infected cells is distinct from that induced by cells infected by other viruses, such as HCV, LCMV and classical swine fever virus (CSFV), which does not require viral structural protein expression [4], [5], [7].
Next, we sought to study the regulation by cell contacts of DENV surface protein-dependent transfer and activation of pDCs. First, the cytoskeleton organization at the cell interface between pDCs and DENV infected cells was determined by confocal microscopy analysis. We observed an accumulation of the actin network at the cell contacts (Figure 7A–E), while the microtubule network was not markedly modified at this location (Figure S11A, left panel). In agreement with the importance of secreted structural components for pDC activation (Figure 6), specific immunostaining of non-permeabilized cells revealed that envelope proteins (i.e., E GP and prM) were both present as clusters at the interface between pDCs and infected cells (Figures 7F–Q and S12).
These observations prompted us to define the impact of the cytoskeleton network on cell contact-dependent pDC IFNα production. We showed that two inhibitors of the cytoskeleton network, Latrunculin B and Nocodazole (i.e., actin and microtubule depolymerizing drugs, respectively) disrupted the actin network in co-cultures of pDC/DENV infected cell (Figure 8A), consistent with previous reports [42], [43]. As expected, the microtubule network was only perturbed by Nocodazole treatment (Figure S11A) [44]. By imaging flow cytometry analysis of GFP expressing-DENV infected cells co-cultured with pDCs (stained by pDC marker CD123) (Figures S13), we showed that the frequency of conjugates between pDCs and DENV infected cells was greatly decreased by inhibitors of the cytoskeleton network (Figure 8B and S13). Both these inhibitors, in conjunction with the loss of actin accumulation at the contacts (Figure 8A), impaired E glycoprotein clustering (Figure 8C). Indeed, quantifications performed in a “double-blind” set-up revealed that, while E GP clustering was readily observed at the cell interface in untreated co-cultures (i.e., ≈60% of the pDCs at close proximity with DENV infected cells harboring E GP clustering), these frequencies were reduced to 15% for co-cultures treated with either inhibitor (Figure 8D). Importantly, similar treatments inhibited IFNα production by the pDCs (Figure 8E). Neither compound inhibited DENV RNA replication in the infected cells and infectious viral production (Figures 8F–G), nor did they prevent the internalization ability of pDC, as assessed by membrane dye uptake (Figure S11B). In addition, they did not inhibit pDC IFNα production triggered by a TLR7 agonist (Figure 8E), thus ruling out potential nonspecific effects of these compounds on pDC responsiveness.
Altogether, these results suggested that the cytoskeleton-dependent regulation of cell contacts and apposed GP clustering likely favors the subsequent activation of IFNα production by the pDCs.
The phenotypic analysis of a virus production defective mutant (i.e., P217A) (Figure 6) revealed that infectious virus production is not required and/or is not rate-limiting for pDC activation.
Like many viruses, DENV infected cells release immature non-infectious particles harboring uncleaved precursor membrane proteins (prM), that are generated by inefficient cleavage of prM by the resident trans-Golgi protease furin [21], [22], [23], [24], [25], [26]. To determine whether immature particles can serve as vehicles from DENV infected cells to contacting pDCs, we first determined the presence of prM protein dots inside co-cultured pDCs by using an antibody recognizing the pr peptide [45] and by examining consecutive Z-sections by confocal microscopy analysis. Dots of prM were observed inside pDCs co-cultured with DENV infected cells (Figures S14A and S14C) with very little background staining in pDCs co-cultured with uninfected control cells (Figures S14B and S14C), suggesting that prM (and/or pr peptide), along with the E GP (Figures 6E, S8 and S9) are readily transferred to the pDCs.
Next, to determine the ability of immature particles to convey immunostimulatory RNAs to pDCs, we tested the effects of DENV genomes encoding mutations in the furin cleavage site of the prM protein (i.e., the substitutions R88A, K90A and R91A), which, as expected from previous reports with single mutations [26], failed to produce infectious virus (Figure 9C). By contrast, RNA replication, intracellular viral protein expression (Figures 9A, S7A and S7B), release of viral components (Figures 9B and S7C), and transmission of viral components to the pDCs (Figures 9D and S8D) were maintained at levels comparable to the WT counterparts. Remarkably, the pDCs produced similar levels of IFNα when comparing contacting cells producing non-infectious immature virions vs WT DENV (Figure 9E). Similar results were obtained when using various concentrations of cells harboring WT/mutant DENV genome (Figure S7D). Therefore, these results suggested that cells producing immature particles potently trigger IFNα production by contacting pDCs.
Next, to define the specific function of uncleaved prM-containing particles in pDC activation by DENV infected cells, we designed experiments aiming at modulating the levels of prM maturation. Firstly, we assessed the impact of an inhibitor of furin. As expected, this inhibitor markedly decreased the maturation of DENV particles, as shown by an increased prM∶E ratio measured by ELISA (Figure 9F). The production of extracellular infectious DENV was also reduced in a dose-dependent manner upon furin treatment (Figure 9H), while the levels of intracellular DENV RNA were unchanged (Figure 9G). Remarkably, inhibition of prM cleavage enhanced IFNα productions by co-cultured pDCs in a dose-dependent manner (Figure 9I). Increased pDC activation was observed despite a reduction in the release of physical particles, as shown by extracellular DENV RNA measurement (Figure 9H). Altogether these results suggested that the activation of pDCs triggered by contacting infected cells inversely correlates with the levels of prM maturation.
To further confirm these results, we studied the impact of furin up-regulation. As expected, cells overexpressing furin produced viral particles containing reduced prM∶E ratios (i.e., ≈10-fold reduction) (Figure 9J). The specific infectivity of DENV particles was increased upon furin overexpression (i.e., ≈3-fold increase in the ratios of infectivity to extracellular DENV RNA, comparing furin-overexpressing cells to counterpart control cells). Thus, cells overexpressing furin were compared to counterpart cells that produced either similar levels of intracellular and/or extracellular DENV RNA, or alternatively, similar production of infectious virus, by using different MOIs (Figure 9K–L). Our results indicated that cells producing more mature particles were clearly impaired at triggering IFNα production by co-cultured pDCs (Figure 9M).
Altogether these results demonstrated that cells producing immature DENV particles are very potent at inducing IFNα production by pDCs, as compared to cells releasing mature virions.
DENV has rapidly emerged in recent years as the most significant arboviral disease of humans, with greater than half of the world population at risk of infection [46]. Despite many years of research, the virus–host interactions that determine dengue pathogenesis are still incompletely understood [47]. Nonetheless, the self-limiting febrile symptoms observed in most DENV-contracted cases and the short course of illness suggest a key role for innate immune defenses in controlling DENV infection at early stages [18]. Accordingly, in vivo studies have demonstrated a critical role for type I IFNs in the host defense against DENV [16], [18], [19]. Furthermore, the activation of pDCs strongly correlates with the disease outcome of DENV infected patients [48]. Importantly, a study of children with DENV infections across a broad range of illness severities suggested that a blunted blood pDC response to systemic infection was associated with higher viremia levels and was a key step in the pathogenic cascade toward severe disease [49]. Although the activation mechanism and exact function are still elusive, altogether, these findings highlight the critical roles played by pDCs and the IFN response on disease progression in DENV infected individuals. Here, we revealed that DENV infected cells potently trigger IFNα secretion by non-permissive pDCs, a host response that bypasses the evasion from the innate response within infected cells. Furthermore, we demonstrated that TLR7-dependent IFNα production by pDCs in response to infected cells is concurrent with other hallmarks of innate immunity, such as inflammatory cytokine secretion, ISG up-regulation and pDC maturation.
In agreement with our results, Rodriguez at al. showed that DENV-containing supernatants failed to trigger pDC IFNα production [50], while other reports suggested that they triggered pDC activation [48], [51]. This discordance may be explained by the preparation and concentration of supernatants and large number of pDCs that were used in the latter reports. Remarkably, the results of our study demonstrated that, despite continuous exposure to the infected cell milieu, physical separation from infected cells precludes the IFN response by pDCs. Consistently, strong pDC IFNα secretion was induced by co-cultured DENV infected cells (i.e., up to 0.5 µg/ml), indicating that cell-to-cell contact is a key feature of pDC activation. Interestingly, cell-to-cell transmission of immunostimulatory signals appears to be a common characteristic of pDC induction, as shown in this report for two members of the Flavivirus genus, DENV and WNV and as previously reported for other viruses, i.e. HCV, HIV, LCMV and CSFV [3], [4], [5], [6], [7]. Specifically, our previous results obtained in the context of HCV indicated that pDC stimulation occurs via viral RNA-containing exosomes. In this context, we suggested that the concentration of immunostimulatory exosomes in the supernatants was below an activating threshold for pDC stimulation, while this threshold might be reached in the intercellular space when cells are in contact [3].
Importantly, we showed here that viral structural components are detected in clusters at the interface between pDCs and infected cells. This finding suggests that cellular surface molecules and/or structures might concentrate the PAMP-carrier at the cell contacts, thereby enhancing transmission to pDCs. We further revealed that the actin network is pivotal for both this clustering of viral components at the pDC-infected cell interface, likely by regulating cell-to-cell contacts, and for pDC activation. Based on this observation, it is conceivable that the cytoskeleton structure serves as a platform contributing to the cell-to-cell transmission of viral components to the pDCs. Additional experiments will be required to test these hypotheses and to determine whether, for the various viruses that trigger the pDC IFN response in a cell-to-cell contact dependent manner, the mechanism of activation involves either common or distinct cellular factors and/or structures at the contacts.
The mechanism we have identified is distinct from the conventional induction of the innate response, which typically occurs by the recognition of viral nucleic acids within infected cells [1], [2]. Moreover, in contrast with the previously characterized induction of pDC IFNα production through contact with infected cells [3], [4], [5], [7], here we have defined a sensing pathway, which requires an E glycoprotein-dependent secretion of viral components, notably viral RNA, to trigger the pDCs. As such, it is different from the mechanism of induction by cells infected by other viruses, which does not require viral structural proteins [3], [4], [5], [7]. Indeed, our results illustrate the crucial role of DENV envelope proteins in the induction of the innate response by neighboring IFN producer pDCs that are not permissive to infection. Importantly, our results revealed that cells producing uncleaved prM-containing immature particles triggered IFNα by pDCs more potently than cells efficiently producing fully mature virions. These immature particles are known to be deficient for the membrane fusion step, which occurs in the endo-lysosomal compartment during cell entry [52], [53], [54]. Interestingly, recognition of viral RNA by TLR7 sensor also takes place in this cellular compartment [1], [55]. Therefore, based on these findings, we suggest a working model in which an extended retention within the endosomal compartment of fusogenic-deficient immature particles may favor the exposure of their viral genome for TLR7 recognition. In contrast, mature virions, which are fusion-competent, could escape from this compartment by membrane fusion. Additional experiments will be required to firmly validate and generalize this new concept.
Several reports have demonstrated that a large proportion of uncleaved prM-containing immature particles are released from DENV infected cells, i.e., 30-to-40% of viral particles [21], [22], [23], [24], [25], [26], on which prM content is variable on a per-particle basis [56], [57]. Consistently, we showed that furin overexpression reduced the levels of immature virus, otherwise, produced by DENV infected cells, and concomitantly with reduced pDC IFN response. Although direct proof is still required, current evidence supports the in vivo existence of uncleaved prM-containing virus. Previous studies have demonstrated that a proportion of B cells isolated from DENV infected individuals produces monoclonal antibodies against prM [58], [59]. In addition, the characterization of these anti-prM antibodies indicated that they are a major component of the serological response to DENV infection, leading to increased replication in Fc receptor-bearing cells via antibody-dependent enhancement (ADE) [56], [58], [60]. Importantly, our results illustrate a previously unsuspected function of these immature particles in innate immunity in mediating an IFN response by non-permissive bystander pDC. Indeed, the results of our study imply that the suboptimal furin-cleavage sequence, likely evolutionarily conserved to favor efficient export of infectious virus by preventing premature membrane fusion in the secretory pathway and cell entry of immature virus into Fc-receptor-expressing cells by ADE [21], [22], [23], [27], [28], might also, by producing an IFN-inducer, contribute to regulate dengue pathogenesis. It is possible that pDC activation by infected cells elicits a strong local innate response that may lead to viral replication suppression or, alternatively, to the possible subsequent recruitment of DENV permissive cells and systemic viral spread. It is also conceivable that the interplay between pDCs and other cells regulating the innate responses, in turn, modulates this newly identified innate sensing mechanism of infected cells and/or the homing of pDCs to the infection site.
Productive infection of cells with a wide range of enveloped viruses depends critically on the processing of the viral surface glycoproteins by cellular proteases [20]. Yet, depending on viral variants/strains, such cleavages might be limited by the differential requirement for certain host proteases, as their expression can be tissue-restricted. These selective requirements may contribute to their virulence, as proposed for influenza virus [61]. Additionally, suboptimal cleavage sites are evolutionarily maintained by sequence features, such as, e.g., the presence of acidic residues or glycosylation sites adjacent to the cleavage site [23], [62]. These events lead to the release of viral particles with uncleaved glycoproteins, as shown for viruses such as, e.g. measles virus [63], influenza virus [61], [64], DENV and WNV [53], [56]. Therefore our results, by uncovering a functional role of immature viral particles in innate immunity, may have broad implications for our understanding of the host-virus relationship.
Huh-7.5.1 [65], Vero E6 (ATCC CRL-1586), Hela (ATCC CCL-2) and HEK-293T (ATCC CRL-1573) cells were maintained in Dulbecco's modified Eagle medium (DMEM) (Life Technologies) supplemented with 10% FBS, 100 units (U)/ml penicillin, 100 mg/ml streptomycin, 2 mM L-glutamine and non-essential amino acids (Life Technologies) at 37°C/5% CO2. BHK-21 cells (ATCC CCL-10) were maintained in Eagle's MEM (Life Technologies) with the same supplements. pDCs were isolated from 450 ml of blood from healthy adult human volunteers which was obtained according to procedures approved by the “Etablissement Français du sang” (EFS) Committee. PBMCs were isolated using Ficoll-Hypaque density centrifugation. pDCs were positively selected from PBMCs using BDCA-4-magnetic beads (MACS Miltenyi Biotec) and cultured as previously described [3]. Monocytes were positively selected from pDC-depleted PBMCs using CD14-magnetic beads (MACS Miltenyi Biotec) according to the manufacturer's instructions, with a typical purity of 95% of CD11c-positive cells. CD14+ cells were then differentiated to monocyte-derived DCs (mo-DCs) by incubation for 6 days in RPMI 1640 medium supplemented with 10% FBS, 100 U/ml penicillin, 100 mg/ml streptomycin, 2 mM L-glutamine, non-essential amino acids, 1 mM sodium pyruvate and 0.05 mM βmercaptoethanol (Sigma-Aldrich) with 500 U/ml human granulocyte-macrophage colony-stimulated factor (GM-CSF) and 2,000 U/ml human interleukin 4 (IL-4) (MACS Miltenyi Biotec), as previously described [12]. To generate monocyte-derived macrophages, monocytes were cultured in the same medium as for the mo-DCs with 500 U/ml GM-CSF for 6 days.
The antibodies used for immunoblotting were mouse anti-E glycoprotein (4G2 and 3H5) kindly provided by P. Despres (Pasteur Institut, Paris, France); mouse anti-capsid (6F3) kindly provided by J. Aaskov (Queensland University of Technology, Brisbane, Australia); mouse anti-actin (AC74, Sigma Aldrich). The antibodies used for immunostaining were mouse PE-conjugated anti-CD123, mouse APC-conjugated anti-BDCA-2, mouse APC-conjugated anti-IFNα (MACS Miltenyi Biotec), mouse PerCP-conjugated anti-CD83 (eBioscience), and mouse APC-conjugated anti-CD86 (BD Bioscience), mouse anti-DENV prM (clone DM-1, Abcam), mouse anti-alpha tubulin (DM1A, Sigma Aldrich); Ficoll-Hypaque (GE Healthcare Life Sciences); LPS, TLR7 agonist (R848) and TLR9 agonist (ODN2216) (Invivogen); TLR7 antagonist, IRS661 (5′-TGCTTGCAAGCTTGCAAGCA-3′) synthesized on a phosphorothionate backbone (MWG Biotech); Fc Blocking solution (MACS Miltenyi Biotec); Golgi-Plug and permeabilization-wash solution (BD Bioscience); IFNα and IFNβ ELISA kit (PBL Interferon Source); IL-6 and TNFα ELISA kit (Affymetrix, eBioscience); Lipofectamine 2000 (Life Technologies); 96-well format transwell chambers (Corning); LabTek II Chamber Slide System, 96-Well Optical-Bottom Plates and Nunc UpCell 96F Microwell Plate (Thermo Fisher Scientific); CF488A-conjugated phalloidin (Biotium); Vibrant cell-labeling solution (CM-DiI, Life Technologies); Hoescht and Alexas-conjugated secondary antibodies (Life Technologies); iScript cDNA synthesis kit (Biorad), qPCR kit (Life Technologies). Latrunculin B, nocodazole, chlorpromazine, dynasore and Gö6983-PKC were purchased from Sigma-Aldrich.
Viral stocks of the prototypic DENV-2 strain New Guinea C (NGC) (AF038403) were produced using in vitro RNA transcripts prepared from DENV-2 infectious plasmid clone pDVWS601 plasmid [66] linearized with XbaI and using mMESSAGE mMACHINE T7 Kit (Ambion). In vitro transcribed RNA was introduced into BHK-21 cells by electroporation as previously described [31]. Briefly, 5 µg of in vitro transcribed RNA was used to transfect 4×106 cells by electroporation. Six hours post-transfection, the culture medium was refreshed. Virus containing supernatants collected at 3 days post-electroporation were clarified through a 0.45 µm filter (Corning).
Viral stocks of WNV (lineage II, strain 956 D117 3B) [67] were produced by transfection of 2×106 HEK-293T cells with 4 µg of the plasmid pWNII-GFP [67] using the Lipofectamine 2000 transfection reagent (Invitrogen) in optiMEM. Six hours post-transfection, the medium was refreshed with HEK-293T culture medium. Virus containing supernatants collected 48 h post-transfection were clarified through a 0.45 µm filter (Corning).
Viral stocks of vesicular stomatitis virus (VSV-GFP, infectious titer of ≈109 Tissue Culture Infectious Dose (TCID50)/ml) were produced as previously described [68] and kindly provided by Dr J. Perrault (Department of Biology, Center for Microbial Sciences, San Diego State University, CA, US). Viral stocks of Influenza A Virus (FluAV, A/WSN/33 strain, delta NS1, i.e., infectious titer of ≈106 plaque forming unit (PFU)/ml) were produced as previously described [69] and kindly provided by Dr R. Le Goffic (Unite de Virologie et Immunologie Moleculaires, Jouy-en-Josas, France). Sixteen hours prior to co-culture with pDCs, Huh7.5 cells, which are known to be deficient for the RIG-I signaling pathway [70], were used to rule out the confounding contribution of IFNα production by the infected cells themselves and infected at MOI of 0.1 and 0.5 for VSV-GFP and FluAV, respectively.
Introduction of mutations into the genomic length DENV-2 strain NGC cDNA clone pDVWS601, encoding amino acid substitutions into the E glycoprotein (i.e., H244A, D215A, P217A) and NS5 (i.e., Rep−/−, containing the multiple amino acid substitutions G81A, G83A and G85A) have been described previously [31], [32]. Mutations encoding amino acid substitutions in prM (amino acids R88A/K90A/R91A) and an in frame four amino acid deletion in the capsid (amino acids V51-to-L54) were first introduced into DENV-2 subgenomic cDNA fragments by overlap-PCR (OL-PCR) using mutagenic primers. The sequences of the primers are described in the Table S1. The OL-PCR fragments were purified and cleaved with BsrGI and SphI and then transferred into the pDVWS601 plasmid that had been cleaved with the corresponding restriction enzymes. The presence of the mutations and sequence of the PCR derived regions were confirmed by sequencing.
In vitro RNA transcripts were prepared from the parental and mutated pDVWS601 plasmids as described above and transfected into Huh7.5.1 cells using the Lipofectamine 2000 transfection reagent (Life Technologies), according to the manufacturer's instruction. One µg of RNA was used to transfect a 60% confluent cell monolayer contained in a single well of a 6-well plate following the manufacturer's protocol. Six hours post-transfection, the cells were either harvested for the quantification of viral RNA (6 hour time point) or washed 3 times with PBS and fresh culture medium added to the cells for additional incubation times. At 24 and 48 hours post-transfection, the cells were harvested for the determination of RNA and protein levels and their supernatants collected for the quantification of viral RNA and infectious titer or concentrated by ultracentrifugation for the determination of protein levels by Western blot. At 24 hours post-transfection, the cells were harvested and co-cultured with isolated pDCs for 18–20 hours.
Forty-eight hours prior to co-culture, cells were infected at a MOI of 3 using a viral stock of WNV or DENV. Unless otherwise indicated, 2×104 pDCs were co-cultured with 105 infected cells, transfected cells or uninfected parental cells, or treated with 100 µl of supernatant from the latter cells in a 200 µl final volume in 96-well round-bottom plates incubated at 37°C/5% CO2. Eighteen to twenty hours later, cell-culture supernatants were collected and the levels of IFNα, IFNβ, TNFα and IL-6 were measured using a commercially available ELISA kits specific for IFNα and IFNβ (PBL Interferon Source), TNFα and IL-6 (Affymetrix), following the manufacturer's instructions. When indicated, 105 infected cells or uninfected cells were co-cultured with 3×104 pDCs or with 105 naïve recipient cells, as indicated, in 96-well format transwell chambers separated by a 0.4 µm membrane (Corning).
At the indicated times, cells were harvested and resuspended using 0.48 mM EDTA-PBS solution (Life Technologies). After incubation with Fc receptor blocking reagent (MACS Miltenyi Biotec) for 10 minutes at 4°C, surface staining of pDC markers, CD123 and BDCA-2 and/or the cell differentiation markers CD83 and CD86 were detected by a 40 minute incubation at 4°C with 5 µg/mL of the indicated combinations of PE-conjugated mouse anti-CD123, APC-conjugated anti-BDCA-2, PerCP-conjugated anti-CD83, and APC conjugated anti-CD86, respectively, diluted in staining buffer (PBS without calcium and magnesium, with 2% FBS), followed by PBS washes. Cells were then fixed by incubation for 20 minutes at room temperature with 4% paraformaldehyde, followed by 20 minutes incubation with 0.1 M glycin-PBS at room temperature and two PBS washes. For intracellular-immunostaining of IFNα, cocultivated cells were treated with 1 µl/ml GolgiPlug solution (BD Bioscience) before collection. After fixation and CD123-staining steps, IFNα was detected by a 40 minute incubation with APC-conjugated mouse anti- IFNα (MACS Miltenyi Biotec) diluted at 1∶10 in permeabilization buffer (BD Bioscience). Cells were then washed twice with permeabilization buffer and resuspended in staining buffer. Flow cytometric analysis was performed using a Digital LSR II, and the data were analyzed with Flow Jo software (Tree Star). The corresponding control isotypes served to define the specific signal.
After isolation, 5×104 pDCs were stained by using 0.5 µM Vibrant cell-labeling solution (CM-DiI, Life Technologies) by successive incubations for 10 and 15 minutes at 37°C and 4°C respectively. Labeled pDCs were washed twice with PBS and then co-cultured with pre-plated DENV infected cells for 5 hours at 37°C in glass bottom 96 well-plate (Fisher Scientific), pretreated with poly-L-lysine at 8 µg/mL. After 4% PFA fixation at room temperature and PBS washing, DENV plus strand RNA was detected using a probe set that targets a region between nucleotide positions 8437-to-9685 in the DENV-2 NGC genome (Panomics/Affymetrix) according to the manufacturer's instructions. For IFNα immunostaining, the cells were permeabilized by incubation for 7 minutes in PBS containing 0.3% (v/v) Triton - and 3% (w/v) BSA, then incubated with mouse anti-IFNα antibody (MACS Miltenyi Biotec) at 2 µg/ml in PBS containing 3% BSA for 40 minutes at room temperature, followed by an incubation with Alexa 647-conjugated anti-mouse antibody (Life Technologies) and Hoechst dye for 40 minutes at room temperature. As controls, FISH detection of DENV RNA were performed in co-cultures of pDCs with non-infected cells and in co-cultures of pDCs with DENV infected cells by omitting DENV-specific probe and by following the same procedure of hybridization and immunostaining. Images were acquired with a Zeiss LSM 710 laser scanning confocal microscope and analyzed with Image J (http://rsb.info.nih.gov/ij) and IMARIS (Bitplane Inc.) software packages.
After immune-isolation, pDCs were stained with 0.5 µM Vibrant cell-labeling solution (CM-DiI) as above-described. 4-to-5×104 DiI-labeled pDCs (DiI-pDCs) were co-cultured with 4-to-5×104 DENV infected Huh7.5.1 cells for 8 hours at 37°C. For analysis of DENV E and prM transfer into pDCs and cell contacts, co-cultures were performed in LabTek II Chamber Slide System (Nunc). After 4% PFA fixation and three PBS washes, cells were permeabilized 7 min with 0.1% Triton in PBS prior immunostaining. For analysis of DENV surface protein clustering at the cell contacts, co-cultures were incubated in a 96-Well Optical-Bottom Plates. After 4% PFA fixation and three PBS washes, immunostainings were performed without permeabilization step, as previously described [71]. After blocking step (PBS 3% BSA) actin filaments were stained with CF488A-conjugated phalloidin (Biotium) at 1.25 U/mL, α-tubulin was stained with mouse anti-α tubulin (DM1A clone, from Sigma) at 1∶2000-dilution, DENV E glycoproteins were detected using anti-E antibody (3H5 clone) at 1∶500-dilution and anti-PrM antibody (DM-1 clone, Abcam) at 1∶50-dilution and IFNα was detected by a mouse anti-IFNα (Miltenyi) at 1∶20-dilution. Antibodies were diluted in 3% BSA-PBS and added to the cell for 1 hour incubation at room temperature. After three PBS-washes with PBS, cells were incubated with an Alexa 647-conjugated-anti-mouse antibody (for detection of anti-α-tubulin and anti-E antibodies) or Alexa 488-conjugated-anti-mouse (for detection of anti-PrM antibody) at 1∶1000-dilution in 3% BSA-PBS, added to the cells along with Hoechst diluted at 1∶500 (Molecular Probes) for 1 hour incubation at room temperature. After three washes with PBS, cells in 96 wells plate were directly observed and cells in Labtek were mounted with mowiol prior observation. Images were acquired with a Zeiss LSM 710 laser scanning confocal microscope and analyzed with Image J (http://rsb.info.nih.gov/ij) and IMARIS (Bitplane Inc.) software packages.
RNAs were isolated from cells or supernatants harvested in guanidinium thiocyanate citrate buffer (GTC) by phenol/chloroform extraction procedure as previously [3]. The efficiency of RNA extraction and reverse transcription-real-time quantitative PCR (RT-qPCR) was controlled by the addition of carrier RNAs encoding Xef1α (xenopus transcription factor 1α) in vitro transcripts in supernatants diluted in GTC buffer. DENV RNA and Xef1α and glyceraldehyde-3-phosphate dehydrogenase (GAPDH) mRNA levels were determined by RT-qPCR using an iScript RT kit (Biorad) and a One-Step PCR Master Mix kit for qPCR and analyzed using StepOnePlus Real-Time PCR system (Life Technologies). The sequences of the primers used for the RT-qPCR are described in Table S1. Extracellular and intracellular DENV RNA levels were normalized for Xef1α and GAPDH RNA levels, respectively.
Infectivity titers in supernatants were determined by end-point dilution using Huh 7.5.1 cells. Foci forming unit (ffu) were detected 72 hours after infection by GFP expression for WNV and anti-E glycoprotein specific immunofluorescence for DENV. Briefly, Huh 7.5.1 cells were fixed with 4% PFA and permeabilization by incubation for 7 minutes in PBS containing 0.1% Triton. Cells were then blocked in PBS containing 3% BSA for 15 minutes and incubated for 1 hour with mouse anti-E glycoprotein (clone 3H5) hybridoma supernatant diluted at 1∶200 in PBS containing 1% BSA. After 3 washes with PBS, cells were incubated 1 hour with secondary Alexa 555-conjugated anti-mouse antibody (1∶1'000-dilution) and Hoechst dye (1∶1'000-dilution) in PBS containing 1% BSA. Percentage of E-positive cells and GFP expressing cells was determined using a Zeiss Axiovert 135 microscope.
Viral supernatant were filtrated through a 0.45 µm filter (Corning) and concentrated prior to Western blot analysis by ultracentrifugation at 110,000× g for 2 hours at 4°C using a SW41 rotor. The pellets were re-suspended in PBS. Viral pellets and cell lysates were extracted using lysis buffer (150 mM NaCl 50 mM Tris HCl pH 8, 1% NP40, 0.5% Deoxycholate, 0.1% Sodium dodecyl sulfate) and analyzed by Western blotting using hybridoma supernatant-containing anti-E (4G2) and anti-capsid (6F3) at the dilution of 1∶500 and actin at 1 µg/ml followed by secondary horse radish peroxidase-coupled antibodies and chemiluminescence.
Huh 7.5.1 cells were transduced with retroviral based vector pseudotyped with VSV glycoprotein to stably express GFP, as previously described [72]. Forty-eight hours prior co-culture with pDCs, GFP-expressing Huh 7.5.1 cells were infected at a MOI of 3 using a viral stock of DENV. 105 GFP-expressing DENV infected cells were co-cultured with 3×104 pDCs in low-adherence micro-plate designed for cell harvesting by temperature reduction (Nunc UpCell 96F Microwell Plate from Thermo Scientific) for 5 hours at 37°C in presence, or not, of Latrunculin B and Nocodazole (1 µM), as indicated. After 4% PFA fixation, co-cultured cells were harvested by equivalent multi-pipetting at room temperature and washed three times with staining buffer (PBS without calcium and magnesium with 2% FBS). After incubation with Fc receptor blocking reagent (MACS Miltenyi Biotec) for 10 minutes at 4°C, surface staining of a pDC marker, CD123, was detected by a 40 minute incubation at 4°C with 5 µg/mL of APC-conjugated mouse anti-CD123, diluted in staining buffer, followed by washes with staining buffer. Co-cultured cells were analyzed by Image Stream X technology (Amnis) at magnification ×60 using IDEAS software. The cell population defined as pDC/DENV cell conjugates comprises conjugates of at least one CD123+ cell and at least one cell solely GFP+ cell among the total of APC+ cells, GFP+ cells and conjugates. The cell populations were sorted by using masks (IDEAS software) to eliminate i/the non-specific signals i.e., double positive single cells and ii/cells with background levels for APC signal. Post-cell sorting, the accuracy of the gated cell population in regards to the defined criteria was controlled by a visual inspection of the individual pictures in the gated cells population (i.e., assessment with 90 randomly picked pictures of the population defined as conjugates). The percentages of gated single cells or conjugates with an accurate phenotype according to the defined criteria among the total of examined pictures per category of cell population were: 97% for GFP+ gated population, 99% for APC-CD123+ gated population and 89% for conjugates.
293T cells, which stably express furin, were generated by transfection using Polyethylenimine and selected using hygromycin (at 5 µg/ml). The decRRVKR-CMK inhibitor (Calbiochem) was used to inhibit the Furin activity in Huh7.5.1 co-cultured with the pDCs, as the indicated concentrations. The levels of prM maturation were analyzed by detection of E and prM by ELISA, as previously described [58]. Briefly serial dilutions of viral supernatants were incubated on anti-E (4G2) antibody coated 96-well plate. Then, E and prM were detected by using a humanized version of 3H5 mAb (hu3H5) and anti-prM, respectively. The prM∶E ratios were calculated using the viral supernatant dilution with E detection in the linear range.
DENV-2 NGC prM and E genes were cloned under the control of CMV promoter, by amplification from pSVprME [73] using primers ADVprME_Fwd (GATCCCCGGGACCGCCACCATGGTGAA) and ADVprME_REV (GATCCCCGGGAGCTTGATATCAGGCCTGC) and cloned into the Sma I site of the adenovirus shuttle vector pDC104 under the control of the CMV promoter to produce pAdvprME. AdvprME was transfected into cells using the Xtreme-GENE HP DNA Transfection Reagent, follow the manufacturer's instructions. Six hours post-transfection, the cells were washed with PBS and fresh culture medium added to the cells for additional incubation times. At 48 hours post-transfection, the cells were harvested and co-cultured with isolated pDCs for 18–20 hours. Parallel determination of intracellular protein levels by Western blot in harvested cells and their supernatants concentrated by using vivaspin concentrator with centrifugation at 3000 g for 30 min (cut-off 100 KDa, Sartorius).
Paired Student's t-test was used to analyze data. Data considered significant demonstrated p-values less than 0.05. Data were also analyzed using a two ways non-parametrical analysis of variance (ANOVA), followed by comparison with Levene Test, analyzed with xlstat software. Triangles indicate the experimental conditions that belong to a separated group statistically different from the others.
|
10.1371/journal.pcbi.1005912 | LRSSLMDA: Laplacian Regularized Sparse Subspace Learning for MiRNA-Disease Association prediction | Predicting novel microRNA (miRNA)-disease associations is clinically significant due to miRNAs’ potential roles of diagnostic biomarkers and therapeutic targets for various human diseases. Previous studies have demonstrated the viability of utilizing different types of biological data to computationally infer new disease-related miRNAs. Yet researchers face the challenge of how to effectively integrate diverse datasets and make reliable predictions. In this study, we presented a computational model named Laplacian Regularized Sparse Subspace Learning for MiRNA-Disease Association prediction (LRSSLMDA), which projected miRNAs/diseases’ statistical feature profile and graph theoretical feature profile to a common subspace. It used Laplacian regularization to preserve the local structures of the training data and a L1-norm constraint to select important miRNA/disease features for prediction. The strength of dimensionality reduction enabled the model to be easily extended to much higher dimensional datasets than those exploited in this study. Experimental results showed that LRSSLMDA outperformed ten previous models: the AUC of 0.9178 in global leave-one-out cross validation (LOOCV) and the AUC of 0.8418 in local LOOCV indicated the model’s superior prediction accuracy; and the average AUC of 0.9181+/-0.0004 in 5-fold cross validation justified its accuracy and stability. In addition, three types of case studies further demonstrated its predictive power. Potential miRNAs related to Colon Neoplasms, Lymphoma, Kidney Neoplasms, Esophageal Neoplasms and Breast Neoplasms were predicted by LRSSLMDA. Respectively, 98%, 88%, 96%, 98% and 98% out of the top 50 predictions were validated by experimental evidences. Therefore, we conclude that LRSSLMDA would be a valuable computational tool for miRNA-disease association prediction.
| Discovering miRNA-disease associations promotes the understanding towards the molecular mechanisms of various human diseases at the miRNA level, and contributes to the development of diagnostic biomarkers and treatment tools for diseases. Computational models can make the discovery more efficient and experiments more productive. LRSSLMDA was proposed to computationally infer potential miRNA-disease associations via adopting sparse subspace learning with Laplacian regularization on the known miRNA-disease association network and the informative feature profiles extracted from the integrated miRNA/disease similarity networks. Experimental results in global and local leave-one-out cross validation and 5-fold cross validation showed a superior prediction performance of LRSSLMDA over previous models. Moreover, three types of case studies on five important human diseases were carried out to further demonstrate the model’s predictive power: respectively, 98%, 88%, 96%, 98% and 98% out of the top 50 predicted miRNAs were confirmed by experimental literatures. So, we believe that LRSSLMDA could make reliable predictions and might guide future experimental studies on miRNA-disease associations.
| MicroRNAs (miRNAs) are small (about 22 nucleotides) non-coding RNAs that regulate gene expression [1]. They normally cleave or translationally repress their target messenger RNAs (mRNAs) via base-pairing to the 3’ untranslated region (UTR) sites of the mRNAs [2–5], thereby influencing various biological processes including cell proliferation, development, differentiation, death, apoptosis, metabolism, aging, signal transduction and viral infection [3,6–11]. In addition, increasing studies have indicated a correlation between miRNAs and human diseases [12–19]. For example, the expression level of miR-195 is lowered in Alzheimer’s disease (AD) patients and the AD amyloid-β production could be downregulated by over-expressing this miRNA [20]. Another miRNA mir-26a contributes to the migration of Lung Neoplasms (LN) cells through modulating the expression of metastasis-related genes and suppressing phosphatase and tensin homolog (PTEN) to activate the Protein Kinase B (AKT) pathway [21]. In contrast, miR-145 is under-expressed in LN patients and its restoration inhibits the LN cell proliferation by targeting the EGFR and NUDT1 genes [22]. A further example of miRNA-disease association is miR-501 in Hepatitis B viruses (HBV). Knockdown of this miRNA in the HBV-producing cell line HepG2.2.15 could significantly reduce HBV replication [23]. These miRNAs and many other disease-associated ones may serve as biomarkers for disease diagnosis, progression, prognosis and treatment response [24–27]. Thus, identifying miRNA-disease associations promotes the understanding of complex human diseases and benefits disease treatment. Experimental methods such as microarray profiling and qRTPCR have been used to discover miRNA-disease associations [28]. But they suffer from false-positive microarray results [25,28–30] and are time-consuming and expensive, especially due to the high probe design cost [28]. Fortunately, the large amount of biological data enables researches to develop computational models for predicting disease-related miRNAs. The potential miRNAs are prioritized in terms of prediction scores and the most promising ones are selected for biological verification. This approach complements experimental methods, improving the accuracy of association identification and reducing time and cost.
Remarkable progresses have been achieved in developing prediction models for potential disease-miRNA associations in the past. Most models were based on the assumption that miRNAs with similar functions tend to be associated with phenotypically similar diseases [31–33]. Many previous models were based on network analysis algorithms. An early model for predicting disease-related miRNAs was devised by Jiang et al. [34] and it integrated the miRNA functional similarity network, the disease phenotype similarity network and the known disease-miRNA association network. The potential miRNA-disease associations were scored according to a discrete hypergeometric probability distribution. However, the model only considered each miRNA’s neighbor information rather than global similarity measures. Then, Chen et al. [35] proposed RWRMDA where novel miRNA-disease associations were predicted by implementing random walking with restart on the miRNA functional similarity network. Although the model achieved an improved prediction accuracy compared with previous models, it was unable to prioritize miRNAs for diseases without any known related miRNAs. Later, Xuan et al. [28] developed HDMP, a model that integrated the known miRNA-disease associations and the miRNA functional similarity calculated by incorporating the information content of disease terms and phenotype similarity between diseases. When scoring miRNA-disease pairs, the model included the information of each miRNA’s k most similar neighbors and assigned higher weights to miRNAs within the same cluster or family. However, HDMP faced the same problem of failing to predict potential miRNAs related to new diseases without any known associated miRNAs. Subsequently, Shi et al. [36] devised another random walk model with a focus on the functional link between miRNA targets and disease genes in a protein-protein interaction (PPI) network. In addition, miRNA-disease co-regulated modules were identified via a hierarchical clustering analysis of a bipartite miRNA-disease network. Nonetheless, involving known disease-gene associations and miRNA-target interactions in the computation impaired the model’s prediction accuracy, since 60% of human diseases have unknown molecular bases [37] and the miRNA-target interactions contain a high rate of false-positive and high false-negative results [35]. Mork et al. [38] used a protein-driven approach named miRPD to infer miRNA-protein-disease associations. The model provided not only the potential associations between miRNAs and diseases but also the protein links between them. To make the inference, known and predicted protein-miRNA interactions were coupled with protein-disease associations text-mined from experimental literatures. Then the inferred miRNA-protein-disease associations were ranked by confidence under two scoring schemes; and the ranking results were divided into a high-confidence subset holding the most probable associations and a medium-confidence subset including the less likely associations. Xuan et al. [39] further introduced a random walk model named MIDP that exploited the prior information of nodes and various ranges of topologies in a miRNA-disease bilayer network derived from the miRNA functional similarity network, the disease semantic similarity network, and the edges between the two networks. With an extended walk on the network, the model overcame the limitations of previous models and could make association predictions for diseases that has no known related miRNAs. Furthermore, the negative effect of noisy data was mitigated via adjusting the restart rate of the random walk. To improve the prediction accuracy, Chen et al. [40] released WBSMDA that calculated and combined the within and between scores from the views of miRNAs and diseases in a composite network, built from the known miRNA-disease associations, the miRNA functional similarity, the disease semantic similarity and the Gaussian interaction profile kernel similarity networks for diseases and miRNAs. Gu et al. [41] developed a non-parametric universal network-based model named NCPMDA. In this model, a miRNA similarity network was constructed by combining the miRNA functional similarity, the Jaccard miRNA similarity of the known miRNA-disease associations and the miRNA family information; and a disease similarity network was built by integrating the disease semantic similarity and the Jaccard disease similarity of the known associations. Then, network consistency projection was carried out on the miRNA similarity network to the adjacency matrix of miRNA-disease associations, and on the disease similarity network to the adjacency matrix, respectively. Lastly, the miRNA space projection scores and the disease space projection scores were combined and normalized to give the final prediction scores. Chen et al. [42] further presented HGIMDA in which a heterogeneous graph network was constructed using the same model inputs as WBSMDA. Then, an iterative process was carried out in the network until a stable association probability matrix was obtained. Following HGIMDA, MCMDA was published by Li et al. [43] utilizing a matrix completion algorithm on the low-rank miRNA-disease association matrix. The candidate miRNA-disease pairs in the matrix were iteratively updated with predictive association scores, yielding highly reliable outcomes. Yu et al. [44] proposed a combinatorial prioritization algorithm named MaxFlow. The model’s input included the miRNA functional similarity network, the disease semantic and phenotypic similarity network, and the heterogeneous miRNA-disease association network that integrated miRNA-disease associations, the miRNA family information and the miRNA cluster information. Subsequently, these three networks were further combined to form a directed miRNAome-phenome network graph, where the weight of each link was regarded as the flow capacity. For an investigated disease, a source node and a sink node were introduced to this graph; and the maximum information flow from the source over all links to the sink were calculated using the push-relabel maximum flow algorithm. The flow quantity leaving a miRNA node was used as the association score between the miRNA and the investigated disease. More recently, You et al. [45] devised path-based model named PBMDA, where a heterogeneous graph were built from the same input datasets as those in WBSMDA. In the graph, all paths between a miRNA-disease pair were traversed via the adoption of the depth-first search algorithm; and each path’s score was computed by multiplying all the edges’ weights along the path. For a longer path, the score would be penalized by a distance-decay function. The sum of scores for all the paths were used as the association score for the miRNA-disease pair.
In addition, other previous models were based on machine learning algorithms. Xu et al. [46] used a support vector machine classifier to separate positive and negative miRNA-disease associations in a heterogeneous miRNA-target dysregulated network (MTDN). Negative samples were required to train the model. However, finding negative miRNA-disease associations is a difficult or even impossible task [42], meaning that the prediction accuracy might be reduced because the model is learned from inappropriate training samples. To address this problem, Chen et al. [47] applied semi-supervised learning (RLSMDA) to the inference of miRNA-disease associations and only using positive samples would suffice the model-training. The ensuing model was RBMMMDA authored by Chen et al. [48]. Restricted Boltzmann machine was implemented to predict four different types of miRNA-disease associations from a two-layered (with visible and hidden units) undirected miRNA-disease graph. RBMMMDA was the first model not only prioritizing potential associations but also providing the corresponding association types. A more recent model developed by Chen et al. [49] was ranking-based k-nearest neighbors for miRNA-disease association prediction (RKNNMDA). It was a three-staged approach: initially running the k-nearest neighbors algorithm for miRNAs and diseases, then carrying out SVM Ranking to rank the neighbors and lastly weighted-voting for both miRNAs and diseases to reduce the prediction bias. Later, Pasquier et al. [50] introduced a vector space model named MiRAI that formed a large network via concatenating five association networks, namely, the miRNA-disease association network, the miRNA-neighbor association network with edges weighted by the genomic distance between two miRNA nodes, the miRNA-target association network, the miRNA-word association network with edges weighted by the term frequency–inverse document frequency (TF-IDF) information retrieval scheme on investigated miRNAs’ associated documents, and the miRNA-family association network. Then, the large combined network was decomposed by Singular Value Decomposition (SVD) into the form of UΣVT, where the columns of U were the left-singular vectors, Σ was the matrix of nonnegative real numbers on the diagonal, and the columns of V were the right-singular vectors. The association score for a miRNA-disease pair was calculated by the cosine similarity between the vector of the miRNA in the miRNA space (U) and the vector of the disease in the disease space (a part of V).
The above mentioned models had their own strengths and uniqueness, while several of them suffered from obvious weaknesses. More importantly, although most models exhibited a sound prediction accuracy, there still exist areas for a continued improvement. When informative feature profiles were extracted from the training data, the challenge would be how to achieve a single classifier that reasonably combine multiple profile spaces. Hence in this study we presented a model of Laplacian Regularized Sparse Subspace Learning for MiRNA-Disease Association prediction (LRSSLMDA) to meet the challenge. The Gaussian interaction profile kernel similarity for miRNA and diseases was computed and integrated with the miRNA functional similarity and the disease semantic similarity. Although the Gaussian interaction profile kernel similarity had been successfully used by Chen et al. [51] in the LRLSLDA model for lncRNA-disease association prediction, their data preparation process was different from that in our study. For LRLSLDA, data preparation involved the lncRNA expression similarity and the lncRNA-disease associations; and the disease semantic similarity was not used. The Gaussian interaction profile kernel similarity for diseases and lncRNAs were computed from the lncRNA-disease associations. Then, the disease similarity was calculated by performing logistic function transformation on the Gaussian interaction profile kernel similarity for diseases; and the integrated similarity for lncRNAs was built by combining the Gaussian interaction profile kernel similarity for lncRNAs and the lncRNA expression similarity. Moreover, a weight coefficient was used in the integrated similarity for lncRNAs. From this, it is apparent that our model and LRLSLDA had different data preparation processes. In addition, constructing the integrated similarity for diseases and miRNAs was only the first step of our model’s data preparation. As the ensuing and important step, feature extraction was performed on the integrated similarity to form the statistical profile and the graph theoretical profile, and these two informative feature profiles were a key to the success of LRSSLMDA. Subsequently, the model used sparse subspace learning to map high dimensional miRNA/disease spaces into a lower dimensional subspace; and it used Laplacian regularization to smooth the subspace and maintain the local structures of the high dimensional spaces. The combination of these two techniques has been successfully applied to web image categorization by Shi et al.’s [52] and drug-target interaction prediction by Liang et al. [53]. But different from Liang et al.’s model, our model made effective predictions with fewer input datasets, exploited informative disease-related feature profiles, and could be applied to diseases without known associations. LRSSLMDA achieved effective dimensionality reduction and could simultaneously analyze a large amount of unlabeled data and a small amount of labeled data. The model was evaluated in three cross validation schemes and three types of case studies on five diseases. In local leave-one-out cross validation (LOOCV), global LOOCV and 5-fold cross validation, LRSSLMDA outperformed ten previous models; and for each disease in case studies, our model predicted the top 50 potentially associated miRNAs and most of the predictions were confirmed by experimental literatures.
HMDD v2.0 is a human miRNA-disease association database that records 5430 experimentally supported associations between 495 miRNAs and 383 diseases (See S2 Table). We used nm to denote the number of miRNAs, nd for the number of diseases and MDA for the nm × nd adjacency matrix made up of the nm miRNAs and the nd diseases. If miRNA m(i) had a known association to disease d(j), the entity MDA(m(i), d(j)) would equal to 1, and otherwise 0.
MiRNA functional similarity scores used in our study were retrieved from http://www.cuilab.cn/files/images/cuilab/misim.zip and computed based on the hypothesis that miRNAs with a functional similarity are more likely to correlate with diseases with a phenotypical similarity [54]. A nm × nm miRNA functional similarity network FS was constructed with weighted edges. An entity FS(m(i), m(j)) denoted the functional similarity score between miRNA m(i) and m(j).
As illustrated in the literature [28], the semantic information of disease d(i) was explained by a Directed Acyclic Graph (DAG) where d(i) and its ancestor diseases were used as nodes. The DAGs were retrieved from the U.S. National Library of Medicine (MeSH) at https://www.nlm.nih.gov/mesh/. The relationship between a parent node and a child node was represented by a directed edge pointing from the former to the latter. For disease t in DAG(d(i)), its contribution to the semantic value of d(i) was computed by
Dd(i)(t)=−log(thenumberofDAGsincludingtthenumberofdiseases)
(1)
The rationale behind (1) was that a greater contribution should be made by a more specific disease t to the semantic value of d(i). Summing up all the contributions from d(i)’s ancestor diseases and itself gave its semantic value
DV(d(i))=∑t∈D(d(i))Dd(i)(t)
(2)
where D(d(i)) denoted the node set in DAG(d(i)). Subsequently, the semantic similarity between disease d(i) and d(j) was defined by:
SS(d(i),d(j))=∑t∈D(d(i))∩D(d(j))(Dd(i)(t)+Dd(j)(t))DV(d(i))+DV(d(j))
(3)
This equation implied that two diseases with a greater overlap of their DAGs would exhibit a higher semantic similarity score between them.
According to [55], the Gaussian kernel similarity between miRNA m(i) and miRNA m(j) was calculated as follows. Respectively, binary interaction profile vectors IP(m(i)) and IP(m(j)) were used to represent the ith column and the jth column of MDA and were then fed into the Gaussian interaction profile kernel similarity matrix for miRNAs, KM
KM(m(i),m(j))=exp(−γm∥IP(m(i))−IP(m(j))∥2)
(4)
where γm was the bandwidth parameter for the function. It was defined by another parameter γ′m and the average number of associated diseases for all miRNAs
γm=γ’m1nm∑i=1nm∥IP(m(i))∥2
(5)
Same to previous literatures [51,55], both the values of γm and γ′m were set to 1 for the simplicity of calculations.
Similar to miRNAs, the diseases’ Gaussian interaction profile kernel similarity matrix KD was calculated by
KD(d(i),d(j))=exp(-γd∥IP(d(i))-IP(d(j))∥2)
(6)
where binary interaction profile vectors IP(d(i)) and IP(d(j)) denoted the ith row and the jth row of MDA; and γd was the bandwidth parameter defined by another parameter γ′d and the average number of associated miRNAs for all diseases
γd=γ’d1nd∑i=1nd∥IP(d(i))2∥
(7)
Again, as with the literatures [51,55], in our study we set the values of γd and γ′d to 1 to make the calculations simple.
The miRNA functional similarity matrix FS and the Gaussian interaction profile kernel similarity matrix KM were integrated to form a more comprehensive similarity measure, which was the integrated similarity matrix for miRNAs SM
SM(m(i),m(j))={FS(m(i),m(j)),ifm(i)andm(j)havefunctionalsimilarityKM(m(i),m(j)),otherwise
(8)
This means that if miRNA m(i) and m(j) had a functional similarity, we chose their corresponding score in FS to be their integrated similarity score; otherwise, we chose instead their Gaussian kernel similarity score obtained from (4).
Similarly, the disease integrated similarity matrix SD was obtained from the disease semantic similarity matrix SS and the Gaussian interaction profile kernel similarity matrix KD
SD(d(i),d(j))={SS(d(i),d(j)),ifd(i)andd(j)havesemanticsimilarityKD(d(i),d(j)),otherwise
(9)
In this study, we developed LRSSLMDA to uncover potential miRNA-disease associations. The model inputs included the miRNA-disease association matrix MDA, the miRNA functional similarity matrix FS and the disease semantic similarity matrix SS. The procedure of implementing LRSSLMDA involved data preparation, model formulation and optimization, as depicted in Fig 1. In data preparation, the integrated similarity matrices SM/SD were constructed according to (8) and (9), respectively, before being used to form two types of feature profiles for miRNAs/diseases. The idea of performing feature extraction on similarity networks to obtain feature profiles originated from the literature [56]. In our study, the first type of profile summarized SM/SD from a statistical perspective, so it was known as the statistical profile. For miRNA m(i)/disease d(j), we calculated
The second type of profile described SM/SD using graph theories, hence was named graph theoretical profiles. We converted SM/SD into an unweighted graph version: miRNA m(i)/disease d(j) now became a node in the graph; and an edge would form between two nodes if their similarity score surpassed the mean value of all entities. For each node in the unweighted graph version of SM/SD, we calculated
Inspired by Liang et al.’s LRSSL model for drug-disease association prediction [53], we used the feature profiles for miRNAs and diseases separately to form and optimize two respective LRSSLMDA objective functions. Our model was an innovation to Liang et al.’s model in the following aspects. First, LRSSLMDA could make effective predictions with fewer input datasets than Liang et al.’s model. As aforementioned, the input to our model contained only three datasets, namely, the miRNA functional similarity, the disease semantic similarity and the known miRNA-disease associations. On the other hand, their model predicted associations between drugs and diseases by integrating five datasets: the drugs’ chemical substructure profile, the drugs’ target protein domain profile, the drugs’ gene ontology term profile, the disease semantic similarity and the known drug-diseases associations. Second, Liang et al.’s model was developed mainly based on the ready-made drug-related profiles and so was only able to work from the drug perspective. Liang et al.’s literature stated that a limitation of the method was not being able to exploit disease-related profiles. Without the involvement of disease-related profiles, the model could not achieve the best possible performance. To deal with this limitation, we made the most of the available disease information by constructing the integrated disease similarity, extracting the statistical profile and the graph theoretical profile for diseases from the integrated similarity, and building the objective function from the disease perspective. In this manner, our model could accurately infer miRNA-disease associations. Third, by intensively involving disease feature profiles, our model could be applied to diseases without known associated miRNAs, whereas Liang et al.’s model was not effective in uncovering drugs associated with a disease that had no known associated drugs.
Because the two objective functions from the miRNA and disease perspectives were constructed and optimized in a similar manner, the rest of this section elaborates the remaining data preparation step, the model formation step and the optimization step from the view of miRNAs, while briefly presenting these steps from the view of diseases.
For miRNAs, the two feature profiles were represented by Xp where p equaled 1, 2 to denote the first and second profiles; the dimension of Xp was dp × nm where dp was the number of features for the pth profile. For each profile, we further built a network graph Sp, whose elements were defined by
Sp(i,j)={1,ifXp(j)wasthek-nearestneighborofXp(i)0,otherwise
(10)
where Xp(i) and Xp(j) were respectively the ith and jth vectors of the pth feature profile. Their closeness was measured by the cosine similarity between them. Furthermore, for miRNAs with known related diseases, we constructed another network graph SMDA, whose elements were computed by
SMDA(i,j)={1,ifMDA(m(j))wasthek-nearestneighborofMDA(m(i))0,otherwise
(11)
where MDA(m(i)) and MDA(m(j)) were respectively the ith and jth row of MDA. Their closeness equaled the maximum integrated similarity score between m(i) and m(j)’s associated disease groups. The last part of the data preparation step was to construct graph Laplacian matrices Lp and LA
Lp=Dp-Sp
(12)
where Dp was the diagonal matrix of Sp in the form of
Dp(i,i)=∑jnSp(i,j)
(13)
Similarly,
LMDA=DMDA-SMDA
(14)
where DMDA was the diagonal matrix of SMDA and defined by
DMDA(i,i)=∑jnSMDA(i,j)
(15)
Lp and LMDA were used to form a Laplacian regularization term in our model and to smooth a subspace to which the miRNA profiles were projected. Lp reflected the trend that miRNAs with similar features should be related to similar diseases, while LMDA helped to maintain the similarity between different miRNAs’ related disease groups.
The subsequent step was model formation, where a common subspace for the miRNA profiles, a L1-norm constraint and Laplacian regularization terms were joint to construct the LRSSLMDA model. This formation was consistent with that presented in the literature [53] and conveyed as the objective function below. This function effectively projected the miRNA profiles to a common subspace and maintained both the local and global structure of the input data.
In (16), F was the predicted miRNA-disease association matrix. The first term ∥F−MDA∥F2 was to keep F aligned with MDA, and ||·||F was the Frobenius norm.
Tr(FTLF) was the Laplacian regularization term, where L=∑p=1mαpγL+αMDAγLMDA. Here, α controlled the contribution of different graph Laplacian matrices to the predictions and γ > 1 guaranteed that all graph Laplacian matrices made a contribution. m was the number of miRNA feature profiles and equaled 2 in this study.
μ∑p=1m∥XpTGp-F∥F2 was the subspace regression term, where XpTGp was a common subspace in the form of a linear transformation of Xp, and Gp was the projection matrix of the pth miRNA feature profile. The subspace was learnt by minimizing the regression errors and μ was the balancing parameter for the subspace learning.
λ∑p=1m∑j=1nd∥Gp(:,j)∥12 was the L1-norm constraint, used to impose sparsity on Gp and assign weights to miRNA features. Here, λ was the regularization parameter and Gp(:, j) was the jth column of Gp.
Finally, (16) was optimized in an iterative process where α1, α2 and αMDA were initialized to 1/3 and G1 and G2 began with random non-negative values from uniform distribution on the [0, 1] interval. According to [53], γ was set to 2; and since the algorithm was not that sensitive to the values of μ and λ, we have set both of them to 1 for the simplicity in calculation. All parameters could be optimized by further cross validation. Gp was interactively updated based on the auxiliary function approach [57]
Gp(i,j)←Gp(i,j)Ap-Gp(i,j)+(Bp+)(i,j)Ap+Gp(i,j)+(Bp-)(i,j
(17)
where
Ap=Xp(μI-μ2PT)XpT+λe1×dpTe1×dp
(18)
Bp=μXpPY+μ2∑q≠pmXpPTXqTGq
(19)
P=(L+(1+mμ)I)-1
(20)
and e1×dp was a 1×dp vector with all elements equal to 1. By fixing F and Gp, αp was renewed by the equation introduced in [52].
The derivation and convergence proof of the optimization algorithm were presented in [53]. The final Gp was multiplied by Xp and then was normalized by row sums, before further timed by the final αp. In this way, the predicted association scores for all miRNA-disease pairs from the view of miRNAs were obtained
MDA*fromthemiRNAperspective=∑p=1map(XpTGpnormalizedbyrowsums)
(22)
Similarly, for diseases in Data Preparation, the two feature profiles were denoted by Xp where p equaled 1, 2 to denote the first and second profiles; the dimension of Xp was nm × dp where dp was the number of features for the pth profile. The resulting network graphs for disease profiles were obtained in the same way as (10). For diseases with known related miRNAs, the network graph SMDA was given by
SMDA(i,j)={1,ifMDA(d(j))wasthek-nearestneighborofMDA(d(i))0,otherwise
(23)
Then graph Laplacian matrices Lp and LA were calculated according to (12) and (14). Again, we constructed the objective function based on (16) in Model Formation, and the Optimization step gave the predicted association scores for all miRNA-disease pairs from the view of diseases
MDA*fromthediseaseperspective=[∑p=1map(XpTGpnormalizedbyrowsums)]T
(24)
The final prediction scores for all miRNA-disease pairs were computed according to three scenarios. First, when predicting potential diseases associated with a miRNA that had no associated diseases, the final prediction scores were calculated according to (22) only, which was MDA* from the miRNA perspective. Second, when predicting potential miRNAs associated with a disease that had no associated miRNAs, the final prediction scores were calculated based on (24) only, which was MDA* from the disease perspective. Third, when making predictions for a miRNA/disease with some associated diseases/miRNAs, the final prediction scores were obtained by taking the average of (22) and (24). These three scenarios were depicted as in (25)
finalMDA*={MDA*fromthemiRNAperspective,investigatedmiRNAwithoutassociateddiseasesMDA*fromthediseaseperspective,investigateddiseasewithoutassociatedmiRNAsMDA*fromthemiRNAperspective+MDA*fromthediseaseperspective2,else
(25)
In this study, we implemented both global and local LOOCV validation methods based on 5430 known miRNA-disease associations between 383 diseases and 495 miRNAs from HMDD v2.0 to evaluate the prediction accuracy of LRSSLMDA. Global LOOCV focused on all potential miRNA-disease associations. Each known miRNA-disease association was left out in turn as the test sample (hence 5430 validation rounds in total), while all the other known associations were considered as the training samples. The remaining miRNA-disease pairs were regarded as candidates. A candidate means a miRNA-disease pair whose association was unconfirmed according to HMDD v2.0 and needed to be predicted by LRSSLMDA. In contrast, local LOOCV only considered miRNAs for a specific disease. Each known miRNA related to disease d(i) was left out in turn as the test sample. This time, we defined all other known disease-related miRNAs (including those related to diseases other than disease d(i)) to be the seeds, and the miRNAs under the unconfirmed association status with disease d(i) to be the candidates. For both global and local LOOCV, the test sample was ranked by LRSSLMDA against the candidates; a rank exceeding a predefined threshold would indicate a successful prediction made by the model and vice versa. Then we plotted a Receiver Operating Characteristics curve with the true positive rate (TPR, sensitivity) versus the false positive rate (FPR, 1-specificity) at various thresholds. Sensitivity meant the percentage of test samples ranked above the threshold and specificity represented the percentage of candidates ranked below the threshold. ROC was subsequently used to generate Area under the ROC curve (AUC), a statistic widely used for describing the prediction accuracy of computational model. An AUC of 1 indicates a perfect performance whereas an AUC of 0.5 implies a random performance.
As shown in Fig 2, in global LOOCV, LRSSLMDA achieved an AUC of 0.9178 and was superior to PBMDA (0.9169), MCMDA (0.8749), MaxFlow (0.8624), NCPMDA (0.9073), HGIMDA (0.8781), WBSMDA (0.8030), HDMP (0.8366) and RLSMDA (0.8426). RWRMDA was not compared in global LOOCV because the model was based on a local ranking approach and thus unable to simultaneously uncover potential miRNAs for all diseases. MiRAI was not implemented in global LOOCV, either. By analyzing association scores calculated by this model, we found that the scores were highly positively correlated with the seed count (i.e., the number of known associated miRNAs) of the investigated disease. We calculated the correlation coefficient between the mean/median score for a disease and the seed count of the disease:
correlation(meanassociationscore,seedcount)=0.4567
correlation(medianassociationscore,seedcount)=0.3979
From this, we can see that the more associated miRNAs a disease had, the higher the association scores for its candidate miRNAs would be; and vice versa. Thus, the association scores calculated by MiRAI for different diseases were not globally comparable and the model was a local method, not applicable to global LOOCV.
In local LOOCV, our model yielded an AUC of 0.8418 and outperformed PBMDA (0.8341), MaxFlow (0.7774), MCMDA (0.7718), HGIMDA (0.8077), MiRAI (0.6299), WBSMDA (0.8031), HDMP (0.7702), RLSMDA (0.6953) and RWRMDA (0.7891). Although our model underperformed NCPMDA (0.8584), the former was superior to the latter both in global LOOCV as mentioned above and in 5-fold cross validation to be subsequently discussed after local LOOCV. Furthermore, NCPMDA seemed sensitive to the percentage of known associations in the training data. In Gu et al.’s study [41], the model was evaluated by local LOOCV using the 1395 known associations between 271 miRNAs and 137 diseases in the HMDD v1.0 database; and the resulting AUC was 0.9173, much higher than the value of 0.8584 obtained in our study. This was due to a reduction of the ratio of known associations to all miRNA-disease pairs in the training data: in HMDD v1.0 there were 1395/(271×137) = 3.76% of miRNA-disease pairs known to be associated, whereas in HMDD v2.0 there were 5430/(495×383) = 2.86% of miRNA-disease pairs known to be associated. This reduction made NCPMDA not as performative as presented in Gu et al.’s study. In addition, it is worth noting that MiRAI had a low AUC of only 0.6299, worse than the AUC of 0.867 presented in Pasquier et al.’ literature [50], because the model was based on collaborative filtering that is known to have the data sparsity problem. The training dataset in our study was sparse, where the average number of miRNAs associated with a disease was 14, while the dataset in Pasquier et al.’ study included 83 diseases with at least 20 known associated miRNAs. Evaluated using a sparser dataset, MiRAI became less performative. We believe that using our dataset to assess models would be a more realistic evaluation than using Pasquier et al.’s dataset, because the relatedness between miRNAs and diseases remains mostly unknown—currently the biological datasets available to research have a just small amount of labeled data and a large amount of unlabeled data. Our method overcame the data sparsity problem and could be applied to this kind of datasets to make effective predictions.
To evaluate LRSSLMDA’s performance variance, we further carried out 5-fold cross validation on the same dataset as that in global and local LOOCV. Since 5-fold cross validation was a global evaluation, MiRAI and RWRMDA were not included in this comparison. The 5430 known miRNA-disease associations were randomly divided into five subsets with an equal size. Each subset was regarded as the test samples in turn and the rest four were used as the training samples. Again, the miRNA-disease pairs without known association evidences were considered as candidates and we recorded the rank of each test sample against them. Finally, an ROC was produced to calculate the AUC. We repeated this procedure for 100 times to achieve a sound estimate of the average prediction accuracy of LRSSLMDA and obtained an AUC of 0.9181+/-0.0004, surpassing that for PBMDA (0.9172+/-0.0007), MCMDA (0.8767+/-0.0011), MaxFlow (0.8579+/-0.0010), NCPMDA (0.8763+/-0.0008), WBSMDA (0.8185+/-0.0009), RLSMDA (0.8569+/-0.0020) and HDMP (0.8342+/-0.0010). Moreover, the AUC’s standard deviation of 0.0004 was one-fifth of that for RLSMDA (0.0020) and about one-third of that for MCMDA (0.0011), and was also noticeably less than that for the remaining five models. This means that, in addition to its superior prediction power, LRSSLMDA was also a stable model with a lower performance variance than others. Another observation was that the average AUC of 0.8763 for NCPMDA was noticeably lower than its AUC of 0.9073 in global LOOCV. In contrast, for all other models, the two values were very similar to each other. This observation again proved the sensitivity of NCPMDA to the percentage of known associations in the training dataset. In global LOOCV 5429/(495×383) = 2.86% of all miRNA-disease pairs in the training dataset were associated, while in 5-fold cross validation 4344/(495×383) = 2.29% of all miRNA-disease pairs in the training dataset were associated. Again, this reduction in percentage impaired NCPMDA’s prediction accuracy.
According to the above comparison, and to our knowledge, LRSSLMDA was by far the most performative machine learning-based model for miRNA-disease association prediction, whereas PBMDA and NCPMDA were the most state-of-the-art network analysis-based models, though there existed a high risk that NCPMDA was sensitive to the percentage of known miRNA-disease associations and would not perform as well with different datasets. Furthermore, it is worth mentioning that the dimensionality reduction technique used in LRSSLMDA facilitated its extendibility to high dimensional datasets. Therefore, the model’s superiority over other models would likely become even more significant in the future with the availability of more feature profiles for miRNAs/diseases as a result of continuous research.
Finally, to assess the predictability of the statistical feature profile and the graph theoretical profile in our study, we used each profile separately for prediction in the above-mentioned three cross validation schemes. Table 1 records the corresponding AUCs and the AUCs for LRSSLMDA with both profiles used. In global LOOCV, the graph theoretical profile achieved a slightly higher predictive accuracy (an AUC of 0.9174) than the statistical profile (with an AUC of 0.9171). This indicated that the former profile was more advantageous in simultaneously uncovering novel miRNA-disease associations for all diseases than the latter. But in local LOOCV, the statistical profile (with an AUC of 0.8405) became superior to the graph theoretical profile (with an AUC of 0.8375), implying that the former would outperform the latter when making predictions for a specific disease. In 5-fold cross validation, like global LOOCV, the graph theoretical profile (with an average AUC of 0.9177) performed better than the statistical profile (with an average AUC of 0.9174), although both of them had an equally low standard deviation of 0.0004. Overall, using either of the two profiles alone for prediction would yield a satisfactory performance; however, only by involving both profiles could our model achieve the best possible predictive performance, that is, an AUC of 0.9178 in global LOOCV, an AUC of 0.8418 in local LOOCV and an average AUC of 0.9181+/-0.0004 in 5-fold cross validation.
Three types of case studies on five important human diseases were carried out to demonstrate the predictive power of LRSSLMDA. The first type concerned with Colon Neoplasms, Lymphoma and Kidney Neoplasms. The known miRNA-disease associations in HMDD v2.0 were used as the training dataset for the model. For each investigated disease, candidate miRNAs were ranked in terms of their predicted association scores. Then, the top 50 candidates were validated by 1) two other prominent miRNA-disease association databases, namely, dbDEMC [58] and miR2Disease [59], and 2) more recent experimental literatures. As a result of inner joining the three databases, 232 of the 5430 known miRNA-disease associations in HMDD v2.0 also existed in miR2Disease, and 546 associations also existed in dbDEMC. Despite this, there was no overlap between the training samples and the prediction lists. This was because in case studies only candidate miRNAs for an investigated disease were ranked and confirmed by experimental evidences. As has been defined, a candidate miRNA was a miRNA unassociated with the investigated disease according to HMDD v2.0. Therefore, none of the top 50 predictions existed in HMDD v2.0 and validation of the predictions was completely independent of this training database. To facilitate further experimental validations, we used LRSSLMDA to produce a complete prediction list for all the 383 diseases in HMDD v2.0 (See S1 Table). In the second type of case study, we sought to demonstrate the model’s applicability to diseases with no known associated miRNAs and used Esophageal Neoplasms as an example. All the known miRNAs related to this cancer were removed from the training samples so that prioritizing candidate miRNAs would only depend on the information of other diseases’ known associated miRNAs and the similarity information of diseases and miRNAs. In this case study only, we built our model solely from the disease perspective, since the investigated disease was made to have no known associated miRNAs. In the third type of case study, the model was trained by 1395 known miRNA-disease associations between 271 miRNAs and 137 diseases from the old version of HMDD, that is, HMDD v1.0. Breast Neoplasms was the investigated disease and its predicted miRNAs were validated against databases including HMDD v2.0, dbDEMC and miR2Disease as well as more recent studies. We implemented this case study to illustrate the applicability of LRSSLMDA to different datasets other than that in HMDD v2.0. The results for the five cancers in the three types of case studies are listed as follows.
Colon Neoplasms (CN) is a cancer arising from the colon or rectum of humans and is more commonly found in developed countries than developing ones [60]. According to the most recent statistics [61], 135,430 newly diagnosed CN cases and 50,260 deaths caused by this disease are expected in the United States in 2017. Both the CN incidence and mortality rates experienced a continuous decline over the past several decades, partly because of the introduction and wide adoption of screening tests [62]. Nowadays, the screening technology could be improved by the utilization of miRNAs as new biomarkers [63,64]. Studies have shown that miR-126 and miR-145 suppress the CN cell growth via targeting the phosphatidylinositol 3-kinase signaling and the insulin receptor substrate-1, respectively [65,66]. We used LRSSLMDA to uncover more CN-related miRNAs and confirmed 43 out of the top 50 potential miRNAs based on dbDEMC and miR2Disease. Among the remaining seven predictions, six were validated by more recent studies: miR-92a was determined to directly target the anti-apoptosis molecule BCL-2-interacting mediator of cell death (BIM) in CN tissues and an anti-miR-92a antagomir led to the apoptosis of CN cell lines [67]; overexpressed miR-199a-3p (the 3p arm of the pre-miRNA for miR-199a) contributed to the late TNM stage in CN and transfecting miR-199a-3p inhibitor into CN SW480 cells could significantly limit the cell proliferation [68]; miR-142-3p (the 3p arm of the pre-miRNA for miR-142) functioned as a CN suppressor through targeting CD133, leucine-rich-repeat-containing G-protein-coupled receptor 5 (Lgr5) and ATP binding cassette (ABCG2) [69]; miR-146b enhanced the proliferation of CN by targeting the calcium-sensing receptor (CaSR) and impairing the anti-proliferative and pro-differentiating actions of calcium [70]; miR-150 was found to be a tumor suppressor in CN by targeting c-Myb [71]; overexpressed miR-122 and its concomitantly suppressed target gene, cationic amino acid transporter 1 (CAT1), would contribute to the development of CN liver metastasis [72]. Overall, combining the above experimental evidences gave a confirmation of 49 out of the top 50 potential miRNAs (See Table 2).
Lymphoma is the most common cancer in adolescents, accounting for 21% of all the cancer cases [61]. Across all age groups, 80,500 new lymphoma incidences and 20,140 mortalities due to the cancer are expected in the United States in 2017 [61]. There are many types of lymphomas but broadly they fall into Hodgkin Lymphoma (HL) or non-Hodgkin Lymphoma (NHL). Experiments have shown that miR-494, miR-1973 and miR-21 could not only be used as diagnostic biomarkers but also circulating cell-free treatment response biomarkers in HL [73]. An example of NHL-miRNA association is that the subtype of NHL, canine B-cell lymphoma, has been found to experience an upregulated expression of miR-19a in the normal lymph nodes [74]. We implemented LRSSLMDA to predict more lymphoma-related miRNAs. Out of the top 50 potential miRNAs, 41 were verified by dbDEMC and miR2Disease; and, among the rest nine predictions, three were confirmed by more recent literatures. MiR-125b-5p (the 5p arm of the pre-miRNA for miR-125b) could upregulate the growth of cutaneous T-cell lymphomas (CTCL) cells, shorten the median survival rate of CTCL patients and promote cellular resistance to proteasome inhibitors by modulating MAD4 proteins [75]. Overexpressed miR-142-5p (the 5p arm of the pre-miRNA for miR-142) was observed in gastric MALT lymphoma, playing a pivotal role in pathogenesis of this cancer [76]. Lastly, the overexpression of miR-146b-5p (the 5p arm of the pre-miRNA for miR-146b) impeded the diffuse large B-cell lymphoma (DLBCL) cell proliferation and this miRNA’s low expression level could predict ineffective treatment response of DLBCL to cyclophosphamide, doxorubicin, vincristine, and prednisone (CHOP) [77]. Consequently, 44 out of the top 50 potential lymphoma-associated miRNAs were proved by experiments (See Table 3).
Kidney Neoplasms (KN) constitutes about 3.8% of all new cancer cases [78] and so is a less common cancer compared with CN and lymphoma. It has been estimated that in 2017 the United States will witness 63,990 new KN cases and 14,400 deaths due to KN [61]. Renal cell carcinoma (RCC) accounts for nearly 80–85% of KN tumors [79] and its diagnosis was made easier by the application of imaging methods such as ultrasound and abdominal CT with or without pelvic CT [80,81]. MiRNAs hold the potential of being novel biological diagnostic targets for KN. For example, a systematic review [82] has reported the down-expression of miR-141 and miR-200 and the up-expression of miR-23b, miR-29b and miR-438-3p in RCCs. We used LRSSLMDA to discover more KN-related miRNAs. Out of the top 50 candidates, 41 were confirmed by dbDEMC and miR2Disease, while seven other candidates were verified by more recent studies as follows: a lately study [83] revealed that down-regulated miR-125b could inhibit the RCC cell migration and invasion, and result in cell apoptosis, though it had no observed impact on the RCC cell proliferation; miR-221 could promote clear cell RCC (ccRCC) proliferation, migration and invasion via directly inhibiting the tumor suppressor TIMP2 [84]; an inverse correlation between the Von Hippel-Lindau (VHL) gene expression and miR-92a was found in ccRCC patients in the study [85], suggesting this miRNA’s oncogenic role in the tumorigenesis of ccRCC; let-7b was considerably under-expressed in ccRCC tissues and its dysregulation was associated with the pathological grade of ccRCC [86]; a low expression of both miR-133a and miR-1 could up-regulate the oncogenic luciferase assay revealed transgelin-2 (TAGLN2), contributing to the development of RCC [87]; oncogene miR-142-3p (the 3p arm of the pre-miRNA for miR-142) was significantly more overexpressed in RCC tissues than adjacent normal tissues and down-regulated miRNA could induce the apoptosis in RCC 786-O and ACHN cells [88]; miR-30a-5p (the 5p arm of the pre-miRNA for miR-30a) experienced considerably downregulation in RCC tissues and cells [89]. As a result, 48 out of the top 50 potential KN-related miRNAs were confirmed by biological evidences (See Table 4).
Esophageal Neoplasms (EN) is a cancer developed from the esophagus and ranks sixth among all cancers in terms of mortality [90]. I the United States, for both sexes the total estimated new EN cases will be 16,940 in 2017, while the total projected death caused by EN will be 15,690 [61]. Population-based screening for EN was not viable due to the relatively low incidence, the absence of early symptoms and the rarity of a hereditary form of the cancer [90,91]. Fortunately, monitoring miRNA expression may be useful for detecting EN. Experiments have indicated that expression profiles of mir-203, mir-205 and mir-21 can determine esophageal tumor histology and discriminate normal tissues from tumorous ones [92]. We trained LRSSLMDA to uncover more EN-related miRNAs and illustrate our model’s applicability to diseases without known associated miRNAs. Out of the top 50 predictions, 49 were confirmed by dbDEMC and miR2Disease (See Table 5). The remaining candidate, mir-122, was found to assist Tanshinone IIA in inhibiting EN cell growth [93]. In addition, miRNA response elements (MREs) of miR-122 and mir-144 employed in EN patients would induce EN cell apoptosis while preserving normal cells [94]. However, whether a direct link exists between miR-122 and EN deserves further investigation.
Breast Neoplasms (BN) is a common cancer in developed countries. In the United States, for instance, one in eight of its population has acquired BN [95] and in 2017 there will be approximately 63,410 newly diagnosed cases [61]. The detection methods for BN mainly include clinical breast examination for earlier-stage cancers and mammography is recommended for women aged over 40 [96]. Curing BN is highly possible given an early stage diagnosis, which could be achieved by involving easily accessible and sensitive miRNAs [97]. MiRNA dysregulations exist in BN patients through polymorphisms in the sequence of the miRNA, its binding sites in target genes, or through epigenetic mechanisms [98]. An example is the elevated expression level of miR-195 which occurred exclusively in BN patients and could be used to differentiate BN from other Malignancies [99]. We trained LRSSLMDA by known miRNA-disease association data from HMDD v1.0. The HMDD v2.0, dbDEMC and miR2Disease databases confirmed 47 out of the top 50 potential BN-related miRNAs, while more recent experimental literatures verified two of the rest three ones. MiR-494 could suppress the progression of BN in vitro by targeting CXCR4 through the Wnt/β-catenin signaling pathway [100]; and the expression level of miR-30e was lowered in both plasma and breast cancer tissues of BN patients and plasma miR-30e expression was statistically related to the patients age and clinical stage of BN [101]. To conclude, experimental evidences from databases and other publications validated 49 out of the top 50 potential BN-associated miRNAs (See Table 6).
The clinical significance of uncovering disease-associated miRNAs lies in their potential roles of therapeutic targets and diagnostic biomarkers for diseases. We introduced a novel computational model for predicting disease-miRNA associations by Laplacian regularized sparse subspace learning (LRSSLMDA). It would effectively complement to existing experimental methods in a way that the candidate miRNAs would be initially prioritized based on available biological data, followed by experimental validations on the most promising candidates. LRSSLMDA was developed as follows. The first step was Data Preparation. The Gaussian interaction profile kernel similarity scores for miRNAs and diseases were calculated from known miRNA-disease associations. Then we constructed the integrated similarity for miRNAs and diseases. In addition, statistical features and graph theoretic features for miRNAs and diseases were extracted from the integrated similarity. The second step was Model Formation. From the respective miRNA/disease perspective, we built an objective function from the common miRNA/disease subspace for the miRNA/disease feature spaces, an L1-norm constraint and Laplacian regularization terms. This step resulted in two objective functions: one from the view of miRNAs and the other from the view of diseases. The third step was Optimization where we optimized the objective functions and lastly combined the optimization results to attain the final prediction outcomes. Albeit inspired by Liang et al.’s method, our model had a substantial innovation: less input data was needed for prediction without sacrificing the predictive performance; disease-related feature profiles were efficiently exploited; and the model could effectively prioritize candidate miRNAs for diseases without known associated miRNAs. Cross validations were carried out to assess the prediction performance of LRSSLMDA. Impressively, it outperformed ten previous models (MCMDA, HGIMDA, WBSMDA, HDMP, RLSMDA and RWRMDA) under the global and local LOOCV frameworks and its prediction stability was reflected by a low standard deviation in results of the 5-fold cross validation. To our knowledge, LRSSLMDA is one of the very few models that achieved an AUC greater than 0.9 in global LOOCV. In addition, three types of case studies on five diseases demonstrated LRSSLMDA’s prediction accuracy. For each disease, a majority of the top 50 potential related miRNAs were confirmed by experimental literatures.
The reliable performance of LRSSMDA stemmed from four factors. First, comprehensive statistical features and graph theoretic features were constructed from the integrated similarity matrices for miRNAs and diseases. The statistical profile included the mean, the sum, the quantiles and the histogram distributions of the similarity scores, while the graph theoretic profile recorded the neighbor count, the centrality measures and Page-Rank scores of the network graphs built from the integrated similarity matrices for miRNAs and diseases. Moreover, because these two feature profiles made full use of the miRNA similarity and the disease similarity, and because functionally similar miRNAs tend to be related to phenotypically similar diseases [31–33], our model could effectively uncover miRNAs associated with diseases that had no known associated miRNAs. This was demonstrated in the fourth case study on Esophageal Neoplasms, where 49 out of the top 50 predictions were confirmed by experimental literatures. Second, dimensionality reduction was implemented via projecting the profiles to a common subspace, which removed the multi-collinearity in them. LRSSLMDA sought to determine the most useful features for differently profiles simultaneously. Third, Laplacian regularization was used to keep the local structure of the feature spaces; it also captured the similarities between known miRNA-related diseases and between known disease-related miRNAs. This resonated with the assumption that functionally similar miRNAs tend to be related to semantically similar diseases. Fourth, the sparse feature selection facilitated by L1-norm assigned higher weights to the most useful features, further improving the performance of LRSSLMDA.
However, there is noticeable room for improvement in LRSSLMDA. The miRNA and disease similarity calculations presented in this study might not be the perfect methods and we expect more biological information to be incorporated into the calculations in the future to fine-tune the similarity measures. In addition, by far the known miRNA-disease associations have a large degree of sparsity (with only 2.86% of 189,585 miRNA-disease pairs being labeled). Accumulating experimental evidences will confirm more associations that would diminish the prediction bias of LRSSLMDA. As a final point, the increasing understanding towards miRNAs and diseases would eventually facilitate a miRNA-disease association prediction that not solely depends on miRNAs’ functional similarity and diseases’ semantic similarity, but also other possible miRNA and disease profiles. Adding new profiles into LRSSLMDA would lead to a more comprehensive analysis and hopefully an improved accuracy of miRNA-disease association prediction. Therefore, we believe that our model would perform even better in future research.
|
10.1371/journal.pbio.2004086 | Fundamental properties of the mammalian innate immune system revealed by multispecies comparison of type I interferon responses | The host innate immune response mediated by type I interferon (IFN) and the resulting up-regulation of hundreds of interferon-stimulated genes (ISGs) provide an immediate barrier to virus infection. Studies of the type I ‘interferome’ have mainly been carried out at a single species level, often lacking the power necessary to understand key evolutionary features of this pathway. Here, using a single experimental platform, we determined the properties of the interferomes of multiple vertebrate species and developed a webserver to mine the dataset. This approach revealed a conserved ‘core’ of 62 ISGs, including genes not previously associated with IFN, underscoring the ancestral functions associated with this antiviral host response. We show that gene expansion contributes to the evolution of the IFN system and that interferomes are shaped by lineage-specific pressures. Consequently, each mammal possesses a unique repertoire of ISGs, including genes common to all mammals and others unique to their specific species or phylogenetic lineages. An analysis of genes commonly down-regulated by IFN suggests that epigenetic regulation of transcription is a fundamental aspect of the IFN response. Our study provides a resource for the scientific community highlighting key paradigms of the type I IFN response.
| The type I interferon (IFN) response is triggered upon sensing of an incoming pathogen in an infected cell and results in the expression of hundreds of IFN-stimulated genes (ISGs, collectively referred to as ‘the interferome’). Studies on the interferome have been carried out mainly in human cells and therefore often lack the power to understand comparative evolutionary aspects of this critical pathway. In this study, we characterized the interferome in several animal species (including humans) using a single experimental framework. This approach allowed us to identify fundamental properties of the innate immune system. In particular, we revealed 62 ‘core’ ISGs, up-regulated in response to IFN in all vertebrates, highlighting the ancestral functions of the IFN system. In addition, we show that many genes repressed by the IFN response normally function as regulators of cell transcription. ISGs shared by multiple species have a higher propensity than other genes to exist as multiple copies in the genome. Importantly, we observed that genes have arisen as ISGs throughout evolution. Hence, every animal species possesses a unique repertoire of ISGs that includes core and lineage-specific genes. Collectively, our data provide a framework on which it will be possible to test the role of the IFN response in pathogen emergence and cross-species transmission.
| Most emerging human viruses have an animal origin [1]. The increase in the global human population, international travel, and ecological changes, in addition to changes in agricultural practices, has led to complicated interactions between wildlife, domestic species, and humans that has enhanced the opportunities for cross-species transmission of known, as well as newly discovered, viruses [1,2]. Physical and molecular components of the innate immune system represent early barriers to incoming viruses that must be overcome in order for an infection to prevail. In vertebrates, one of the key innate immune defences against virus infection is the interferon (IFN) system. Type I interferons (including IFN-β and IFN-α among others; here referred to simply as IFN), type II interferons (IFN-γ), and type III interferons (IFN-λ) are cytokines with antipathogen, immunomodulatory, and proinflammatory properties. The IFN system is usually stimulated by the detection of pathogen signatures, known as pathogen-associated molecular patterns (PAMPs), resulting in the secretion of IFN. In turn, IFN signalling results in the up-regulation of hundreds of interferon-stimulated genes (ISGs), collectively referred to as the type I ‘interferome’ (here simply the ‘interferome’) [3,4]. Unsurprisingly, given the importance of IFN in combatting pathogen invasion, there are numerous examples of coevolutionary arms races between ISGs and invading pathogens [5,6]. However, previous studies investigating ISG transcription have focused on the interferomes of single species [7,8]. Despite resources such as the Interferome database [9], variations in experimental and bioinformatic approaches make comparing interferomes derived from divergent species and collected from different studies a difficult prospect if significant technical caveats and confounding factors are to be avoided. Here, we used the same RNA sequencing (RNAseq) approach on 10 animal species to deliver a snapshot of the genes that are differentially expressed in cells (fibroblasts) in a type I IFN-induced antiviral state. This snapshot of the interferome from a single cell type at one point in time cannot capture the entire temporal and tissue-specific complexity of the interferome. Nonetheless, using this comparative approach, we have uncovered fundamental paradigms of the IFN system.
We first determined the individual interferomes of Homo sapiens (human), Rattus norvegicus (rat), Bos taurus (cow), Ovis aries (sheep), Sus scrofa (pig), Equus caballas (horse), Canis lupus familiaris (dog), Myotis lucifugus (little brown bat, microbat), Pteropus vampyrus (large flying fox, fruit bat), and Gallus gallus (chicken) cells. Interferomes were obtained from cells stimulated with type I IFN and experimentally confirmed as being in an antiviral state as described in the Materials and methods (S1 Fig). In our study, we defined an ISG as a gene up-regulated by IFN with a false discovery rate (FDR) of <0.05, regardless of the extent to which it was up-regulated. To facilitate mining of the data, we also developed an open access webserver (http://isg.data.cvr.ac.uk) capable of filtering the dataset based upon user-defined criteria.
The absolute number of ISGs differentially expressed in each species varied (S1 Table), but their pattern of differential expression in response to type I IFN was remarkably similar (Fig 1A). The presence of shared ISGs at specific nodes on a schematic phylogeny provided evidence that interferomes have been sculpted over time by lineage-specific pressures possibly exerted by different pathogens (Fig 1B). As expected, we observed that the most closely related species in our dataset, cows and sheep, showed the greatest similarity in the genes they up-regulate (Fig 1C). However, we also observed substantial levels of similarity in the interferomes of some species that are more distantly related phylogenetically, most notably pigs and humans (Fig 1C). Interestingly, this finding was reflected in a principal component analysis of the 35 one-to-one (i.e., single copy) ISG orthologs up-regulated by every mammalian species in our study (see below), whereby the patterns of differential expression were again similar between humans and pigs (Fig 1D).
Every species possessed unique ISGs that were not up-regulated by IFN in any of the other nine species. Furthermore, certain ISGs present in our dataset (despite being up-regulated by IFN >2 log2 fold change [log2FC]) had few (if any) orthologs in the other genomes in the Ensembl database. Examples include a gene (RGD1561157) that is annotated on chromosome 10 of the rat genome and two chicken genes (Ensembl IDs ENSGALG00000019325 and ENSGALG00000020899).
We identified a core set of 62 genes (“core vertebrate ISGs”, hereinafter corevert ISGs) that were up-regulated by IFN by all 10 species analysed in this study, with an additional 28 genes up-regulated specifically in the nine mammalian species (“core mammalian ISGs”, hereinafter coremamm ISGs) (Table 1). The corevert ISGs represent the ancestral functions of the IFN system and include genes encoding proteins broadly involved in (i) orchestrating antigen presentation, (ii) IFN induction and response, (iii) IFN suppression, (iv) ubiquitination and protein degradation, (v) cell signalling and apoptosis, and (vi) antiviral responses (Table 1, Fig 1E).
Nine of the 62 corevert ISGs (e.g., various HLA genes, TAP1, ERAP1, etc.) are involved in the generation, trimming, loading, and presentation of MHC-I–restricted antigens, thus providing a direct link between the IFN response and the adaptive immune response via the CD8 T-cell response (Table 1). Additionally, components of the immunoproteasome (e.g., PSMB8 and PSMB9) were specifically up-regulated in the mammalian core (Table 1), reflecting previous studies reporting the absence of the immunoproteasome in birds [10].
We found that various corevert ISGs are involved in IFN induction and response, including pattern recognition receptors, the key adaptor molecule MyD88, and transcription factors. The classical sensors for RNA PAMPs, including IFIH1/MDA5, DHX58/LGP2, and TLR3 (and RIG-I/DDX58 in the mammals) were among the corevert ISGs (Table 1, Fig 2). RIG-I is known to be absent in chickens (and other galliformes) but is present and active in other birds, including ducks and geese [11–13]. Furthermore, TRIM25, a ubiquitin E3 ligase responsible for ubiquitinating RIG-I, was also a corevert ISG [14]. DAI/ZBP1, which was originally classified as a DNA sensor but is now thought to be an RNA sensor [15], was also highly up-regulated by IFN in every mammalian species except humans (Fig 2). No DNA sensors were found among the corevert ISGs. However, cGAS was found to be an ISG in every mammalian species in this study (Fig 2). By contrast, the DNA sensor AIM2 was only up-regulated in human cells. Interestingly, whilst the basal expression (defined in terms of fragments per kilobase mapped values [FPKM] in the absence of IFN treatment) of RNA sensors was very low, many of the genes in the literature associated with DNA sensing are constitutively transcribed (Fig 2).
With the exception of MyD88, a key adaptor involved in both the RNA- and DNA-sensing pathways [16], we observed limited up-regulation among genes involved downstream of nucleic acid detection (Fig 2). On the other hand, core ISGs included key transcription factors involved in IFN induction and response (IRF1, IRF7, STAT1 and STAT2 are all corevert ISGs in addition to IRF9 among the coremamm ISGs) (Fig 2). Importantly, several ISGs that play a role in the suppression of the IFN system, including USP18, USP25, IFI35, and SOCS1 were up-regulated in all species under examination. The encoded proteins of these genes target different points in the IFN response. Thus, negative regulation of the IFN response is multifaceted and a fundamental, ancestral failsafe necessary to avoid excessive/perpetual up-regulation of IFN-induced pathways. Among the corevert ISGs, we found several genes relating to ubiquitination, such as the ring finger proteins RNF213 and RNF19B (Table 1, Fig 1E), highlighting protein modification as part of the IFN response. Interestingly, N4BP1, originally identified as a target of Nedd4-mediated ubiquitination, has not previously been directly linked to the IFN response.
The corevert ISGs contained 14 IFN-induced antiviral factors such as MX1, IFIT2, and viperin (Table 1). Interestingly, when we assembled a list of 40 genes that either create a cellular environment hostile to or act directly upon the virus lifecycle (based upon the scientific literature; S2 Table), we noticed that 75% of these antiviral genes were ISGs in at least eight of the 10 species analysed in this study (Fig 3A). In addition, antiviral ISGs were up-regulated to a significantly higher extent than randomly sampled ISGs (P < 0.01 for each species, Fig 3B). Furthermore, some well-studied antiviral ISGs were not up-regulated by IFN in certain species. For example, OAS2 was not up-regulated in the rat, SAMHD1 was not up-regulated in the horse, OASL was not up-regulated in either the cow or the sheep, and the IFITM genes were not up-regulated in the dog (Fig 3C). In general, genes encoding antiviral factors were transcribed to minimal levels in the absence of IFN. Indeed, the median FPKM level for antiviral genes was lower than that of the overall interferome for every antiviral factor except SAT1, SHISA5, and the IFITM genes. Of interest, we noticed particularly high FPKM values for IFITM1 and 3 in the rat and IFITM2 in the microbat (Fig 3C).
Two corevert genes, CD47 and IL15RA, encode proteins involved in signalling to components of the adaptive immune system. CD47 is involved in a variety of biological roles, including leukocyte and dendritic cell migration, the development of antigen presenting cells, and immune apoptosis, and it also provides a ‘don’t eat me’ signal [17]. The IL15–IL15Rα axis is well characterised as being important in the promotion of both natural killer cells and a variety of T cell populations, including activated CD8 T cells [18].
Among the corevert ISGs, we identified a number of genes with few or no reported associations with the type I IFN response (Table 1). Several of these genes have been studied extensively, but not always in the context of the IFN response. DNAJC13, for example, is reported to be involved in endosome trafficking, and it has been closely linked to Parkinson’s disease [19]. Zinc Finger CCHC-Type Containing 2 (ZCCHC2) has nucleic acid–binding properties and, interestingly, contains a single nucleotide polymorphism (SNP) associated with insect bite hypersensitivity in Exmoor ponies [20]. The fragile X mental retardation (FMR1) gene encodes an RNA-binding protein that plays a role in intracellular RNA transport and in the regulation of translation of target mRNAs. FMR1 was not previously linked to the IFN response, although it has recently been shown to be a proviral factor for influenza virus and previously was shown to induce mild restriction of HIV-1 [21,22]. Cap methyltransferase 1 (CMTR1), also known as ISG95, binds to RNA pol II and is a 2′-O-ribose methyltransferase that participates in the conversion of cap0 to cap1 type transcripts [23]. Interestingly CMTR1 is also known as an important component of IFIT-mediated antiviral activity [24]. We hypothesise that these genes, as corevert ISGs, likely play fundamental roles in host immunity that are underappreciated or have yet to be fully determined.
Our data also indicate a relatively underappreciated link between local synthesis of early components of the complement system and the type I IFN response [25,26]. C2 was among our coremamm ISGs. In addition, we found that C1r and C1s, essential components of the C1 complex, were up-regulated by IFN in cells from all species analysed in our study with the exception of the cow. Interestingly, this was also the case for a negative regulator of C1r and C1s (SERPING1/C1-INH).
Our data enabled an unprecedented opportunity to investigate the interferon-repressed genes (IRGs), which, to date, have received comparatively little attention with regards to their involvement in the innate immune system. Unlike ISGs, the extent of down-regulation of IRGs across all the 10 species used in this study was relatively modest (overall average −0.56 log2FC for IRGs as compared to 1.64 log2FC for ISGs). We found no IRGs shared by all species, although this may reflect the low fold change in expression and/or greater variability in the response of individual genes. This result could also imply that there has been less conservation of the down-regulated genes over time. The most consistently down-regulated genes were FAM117B and KDM5B, which were both significantly down-regulated in eight of the 10 species analysed in this study. Relatively little is known about the function of FAM117B with the exception that it is a risk factor for sarcoidosis [27]. On the other hand, it has been shown that suppression of the KDM5B gene product, a H3K4 demethylase causing transcriptional repression, results in increased expression of IFN-β and other inflammatory cytokines following infection with respiratory syncytial virus (RSV) [28]. We also noticed that, with the exception of the rat, every species analysed down-regulated at least one KDM gene in response to IFN.
It is already established that another form of epigenetic control, acetylation, is also required for robust ISG transcription [29]. ANP32A, a protein involved in acetyltransferase inhibition [30], was down-regulated in five species. Interestingly, ANP32A has been shown to be a host component necessary for influenza virus replication and influences the ability of the virus to replicate in a given animal species [31].
Our data have thus revealed that key epigenetic factors regulating ISG transcription are themselves frequently responsive to IFN treatment. Along these lines, we also investigated the presence of noncoding RNAs (ncRNAs), a class of RNA molecule increasingly recognised as being important in the antiviral response [32,33]. We found that in human cells 75 long intergenic noncoding RNAs (lincRNAs) were differentially expressed (38 up-regulated, 37 down-regulated) in response to IFN (S2 Fig), including NEAT1, a lincRNA that has previously been associated with viral infections [34–36]. As other genomes become increasingly well annotated, it will be possible to resolve a more in-depth understanding of the impact that ncRNAs play in the control of the innate immune system.
Virus–host coevolution has shaped the innate immune system, most frequently by placing antiviral genes under positive selection. We assessed the dN/dS ratios (as compared to the human) for one-to-one ISG orthologs. We observed that the overall distributions of dN/dS values of ISGs were significantly higher than those of randomly selected non-ISGs (Fig 4A). In addition, we assessed the copy number of each ISG across the different species studied here. Strikingly, for each species—with the exception of sheep and, to a lesser extent, cow—the proportion of ISGs with gene expansions was significantly above that of the genome as a whole (Fig 4B). Interestingly, sheep and cows are the only species with multiple copies of the IFN-β gene (generally a single copy gene in mammals).
The data described above suggest that, in general, expanded ISGs (compared to the rest of the genome) have an increased likelihood of conferring a selective advantage to the host species. Indeed, we observed that ISGs that are shared between multiple species have a higher likelihood of being expanded in the genome compared to other genes (P < 0.001, Fig 4C). Furthermore, among the corevert ISGs, one-to-one orthologs are induced by IFN to a significantly higher level than genes present as paralogs (two-way ANOVA, F = 2.284, P < 0.05).
We further analysed gene expansions and deletions among the coremamm ISGs in the genomes of 111 mammalian species using an in silico sequence-similarity screening approach [37]. Although the uneven quality of the genomes used in the analysis make this approach prone to artefacts, we were able to detect coremamm ISG deletions. For example, we observed that XAF1, which is a negative regulator of inhibitors of apoptosis, is deleted in cats (Felidae) (Fig 5A). In addition, we confirmed previously published deletions of IFIT3 among the Scandentia (tree shrews), Cetacea (whales and dolphins), and marsupials (Fig 5A) [38,39]. Furthermore, we observed that IFIT2 exists as a pseudogene in the Cetacea (Fig 5B).
In this study, we devised a systematic approach to unveil fundamental properties of the type I IFN system in vertebrates. We investigated the IFN response in several mammalian species and the chicken using a consistent experimental framework. By grouping ISGs and IRGs according to the number of species in which they were differentially expressed, we were able to reveal key facets related to the evolution and function of the IFN response.
We identified 62 corevert ISGs that were up-regulated both in the chicken and nine mammalian species. Similarities between the chicken and mammalian IFN systems likely reflect fundamental functions that were present in the common ancestor of birds and mammals that have remained conserved over the ensuing circa 300 million years. Orchestration of the adaptive immune response by IFN appears to be a fully conserved and prioritised function among vertebrates. Specifically, within the corevert ISGs, we found MHC class I components, along with genes involved at all levels of the antigen presentation process and genes involved in cell signalling to diverse cells of the adaptive immune system. Interestingly, we found that the type I IFN response may also facilitate local expression of factors of the complement system.
Only a limited set of genes relating to PAMP detection (MDA5, LGP2, TLR3) and downstream signalling (IRF1, IRF7, STAT1 and STAT2) were within the corevert ISGs. In general, we noticed a greater number of IFN-up–regulated RNA sensors (and a greater level of their up-regulation) compared to DNA sensors. These data imply that the type I IFN response biases the sensitivity of surveillance for RNA viruses and may also reflect the difficulty, and danger, of differentiating self from nonself RNA in the cytoplasm.
Known antiviral ISGs are, in general, shared by many species. It therefore appears that a large proportion of the known antiviral capability of the IFN system evolved at an early point, a finding consistent with the presence of antiviral activity among fish ISGs [40,41].
Characterisation of the corevert ISGs has also revealed several genes that hitherto have had little, if any, association in previous studies with the IFN response. Our data suggest that these genes play fundamental roles in the innate immune response of vertebrates that remain yet to be discovered. It is possible that genes such as these have been overlooked in previous studies simply as a result of their relatively modest fold up-regulation in response to IFN treatment in cells derived from individual animal species. This dataset therefore provides additional power with which to uncover novel genes central to the IFN system and an alternative approach by which to prioritise their relative biological significance and evolutionary conservation.
We observed that genes have arisen as ISGs throughout evolution, to the extent that certain genes are responsive to IFN only in particular phylogenetic groups. In addition, ISGs shared by multiple species have a higher propensity to be retained in genomes, yet another example of the pressure exerted by invading pathogens shaping vertebrate evolution. Hence, the result of these evolutionary processes is that every species possesses a unique repertoire of ISGs. These findings may help explain the differing sensitivities of certain animal species to specific viruses. For example, it has been widely hypothesised that the bat innate immune system has unique features that allow this species to withstand persistent, asymptomatic infection with viruses that are pathogenic in humans [42–44]. Our data reveal that the overall pattern of the bat interferome is relatively unremarkable: they up-regulate the core ISGs, have similar distributions of up-regulated and down-regulated genes, and up-regulate lineage (order Chiroptera), as well as species-specific, ISGs. However, we found that the basal transcription level of the type I interferome (including the known antiviral ISGs) to be higher in both the megachiropteran and microchiropteran cells compared to cells from other species (S5 Fig), a finding consistent with the previous observation that the interferon alpha (IFN-α) locus is constitutively active in bats [44]. Hence, bat cells might exist in a relatively constitutively active antiviral state.
The IRGs were differentially expressed to a conspicuously lower extent than ISGs, although the overall pattern was largely uniform across the species. The ability to assemble lists of shared IRGs allowed us to suggest that epigenetic control via down-regulation of genes involved in acetylation and methylation may be a relatively underappreciated function of the IFN response. For example, we found that ANP32A, a protein involved in acetyltransferase inhibition [30], was down-regulated by IFN in human, rat, sheep, cow, and pig cells. ANP32A was recently identified as a key cellular cofactor for avian influenza virus (AIV). Indeed, the avian influenza virus polymerase functions relatively poorly in mammalian cells, and this is due, at least in part, to the inability of AIV polymerase to bind efficiently to mammalian ANP32A [31]. It is intriguing that, in our data, chicken ANP32A is not significantly down-regulated by IFN.
Our study, like many of a similar nature, relies on the quality of the annotations of the genomes used. Indeed, many ISGs have been shown to have complex orthologies, and it is possible that some genes are misannotated or not yet annotated in the Ensembl Compara database. In order to decrease the impact of this factor in our data, we manually curated all ISGs that were initially identified in at least eight animal species. In addition, the use of primary cells for most species (in most cases derived from different individuals) reduced the possibility of artefacts deriving from cells that were passaged extensively in vitro. We also ensured that all RNAseq experiments were carried out in cells in which IFN stimulation resulted in an antiviral state (see Materials and methods). The system-level nature of RNAseq experiments and downstream bioinformatic analyses complicates direct comparisons between distinct studies. MORC3, for example, was previously suggested to be specifically up-regulated by the megabats, as it was not up-regulated in human A549 cells [45]. By contrast, in our study, we observed MORC3 among the corevert genes, albeit robustly up-regulated in the fruit bat cells (> 12-fold) and minimally in human cells (< 2-fold). Similarly, the scope of this study is limited to type I IFN and a single cell type per species. It will be interesting in future studies to observe the differences in interferomes generated by IFN-γ and IFN-λ and, additionally, the variation that exists between cell types.
It is notable that the list of species for which we generated interferomes includes wild (rat, microbat, and fruit bat), as well as domestic companion (dog and horse) and livestock (pigs, chicken, cow, and sheep), species. We observed clear species- and lineage-specific ISGs for every species examined, which, as more genomes become sequenced, can be explored for evidence of how, for example, the domestication process has influenced the evolution of ISGs. Overall, the dataset described here represents the most comprehensive, cross-species ‘snapshot’ of the IFN response published to date. Our data provide a framework with which it will be possible to test hypotheses pertaining to the role of host innate immunity on virus emergence, cross-species transmission and pathogenesis.
Ex vivo tissue samples were collected postmortem either at commercial slaughterhouses or at the University of Glasgow School of Veterinary Medicine. In all cases, animals had been euthanized according to protocols approved by the local ethical committee and in accordance with the Council of the European Communities Directive of 24 November 1986 (86/609/EEC).
Ex vivo skin samples were collected from chickens (n = 3), cows (n = 4), sheep (n = 3), horses (n = 3), a dog (n = 1), and pigs (n = 4) and primary fibroblasts isolated following an explant procedure. Briefly, the hair or feathers were removed from the skin prior to disinfection by soaking in 70% ethanol. After rinsing in PBS, the skin was removed from the underlying tissues, cut into circa 3-mm square explants, and added to cell culture dishes (without media, squamous surface uppermost) for one hour at 37°C before adding Dulbecco’s modified Eagle medium (DMEM) (Gibco) supplemented with 10% fetal bovine serum (FBS) (Gibco), 1% penicillin/streptomycin (p/s) (Sigma) and 100 U/ml nystatin (Sigma). Human primary dermal fibroblasts were purchased from PromoCell (catalogue number C-12302). Rat primary dermal fibroblasts were purchased from the European Collection of Authenticated Cell Cultures (ECACC) (catalogue number 06090769). M. lucifugus (little brown bat, representative of the microbats) primary dermal fibroblast cells, isolated from individuals caught in Oregon, United States of America, were kindly provided by William Kohler [46]. P. vampyrus (the large flying fox fruit bat, representative of the megabats) cells (PVK4) are an immortalised kidney cell line kindly provided by Megan Shaw [45]. The origin of each cell used in this study is summarized in S5 Table.
Human, rat, and dog cells were cultured in fibroblast growth medium 2 (PromoCell) supplemented with 10% FBS and p/s. 293T and PVK4 were all cultured in DMEM supplemented with 10% FBS and p/s. All cell cultures were incubated at 37°C with 5% CO2 in a humidified atmosphere.
IFN- or mock-treated cells were challenged with infectious VLPs of envelope minus vesicular stomatitis virus (VSV-ΔG-GFP) decorated with a VSV-G envelope (provided in trans during VLP production) in order to confirm the antiviral state of the cells at the time of RNA harvest essentially as already described [47,48]. Briefly, cells were harvested by trypsinization and fixed in 5% formaldehyde. The number of infected cells in IFN- and mock-treated cells was assessed by flow cytometry.
Parallel sets of cells were plated in multiwell plates 24 or 48 hours prior to IFN treatment and incubated at 37°C. Cells were treated with either 1,000 U/ml Universal interferon (UIFN, PBL InterferonSource), 200 ng/ml canine IFNα (Kingfisher), 1,000 U/ml porcine IFNα1 (Stratech), or 200 ng/ml chicken IFNα (AbD Serotec). Mock treatment was performed in parallel using DMEM lacking IFN. Cells were incubated for the indicated time period at 37°C, washed with PBS, and either lysed in Trizol (Thermo Fisher) for RNA extraction or challenged with VSV-ΔG-GFP to assess the antiviral state. Cells were only further processed for RNAseq analyses when they were in an antiviral state. In this study, cells were considered in an antiviral state when IFN stimulation induced at least 75% inhibition of VSV-ΔG-GFP infectivity (value chosen as average of three independent experiments) (S1 Fig). Pilot experiments were performed for each cell type in order to optimise conditions necessary for cells to reach an antiviral state. With the exception of dog cells, all cells reached an antiviral state after four hours of IFN treatment. The antiviral state in the primary canine cells isolated in these experiments required 24 hours of treatment with canine IFNα (S1 Fig). For comparative purposes, a complete set of experiments ranging from IFN stimulation to interferome analysis was performed in parallel in pig cells stimulated with either UIFN or porcine IFN-α (S3 Fig).
RNA was extracted using Trizol (Thermo Fisher) and RNeasy (Qiagen) protocols. Briefly, chloroform was added to the RNA-containing phase of the Trizol sample and centrifuged. The aqueous phase was then mixed with ethanol and purified using RNeasy columns, incorporating an on-column DNase step (Qiagen) to ensure the complete removal of genomic DNA. Total RNA samples were quantified using the Qubit (Thermo Fisher) and were assessed for integrity using the Bioanalyser pico eukaryotic II RNA chip (Agilent). Only samples with an RNA integrity number (RIN) value > 9 were taken forward for library preparation.
Libraries of mock- and IFN-treated cell RNA were assembled using equal masses of total RNA. The External RNA Controls Consortium (ERCC) spike control was added to the total RNA sample in order to assess library quality following sequencing. RNA samples were first enriched for mRNA by selecting poly(A) RNA using the Dynabeads mRNA DIRECT Micro purification kit (Thermo Fisher). The eluted RNA was used to generate RNAseq libraries using the Ion Total RNA-seq Kit v2 (Thermo Fisher) following the manufacturer’s instructions, with the exception that RNA samples were sheared for just 1.5 minutes. Amplified and purified libraries were checked for quality and quantity using the Agilent Tapestation (D1000 tape) and Qubit (hsDNA assay). Libraries were run on the Ion Proton (Thermo Fisher) according to the manufacturer’s instructions.
Raw data were trimmed and assessed for quality using FASTQC (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). We performed multidimensional scaling (MDS) of normalised counts per million data using EdgeR (Bioconductor) in order to assess the impact of IFN and assessed biological covariance as a further control for data quality.
In order to assess the presence of cell culture contaminants on the transcriptomic data, we used Kraken to assign taxonomic labels to the reads using the MiniKraken database, which contains all complete bacterial, archaeal, and viral genomes in RefSeq [49]. To assess mycoplasma levels in the cell cultures, we used Bowtie2 to map the data to six different strains of mycoplasma known as frequent contaminants of cell cultures (S4 Fig).
Reads were aligned to host genomes (S6 Table) using a two-step procedure. A first round of mapping used TopHat2, followed by a second round of mapping using Bowtie2 in an attempt to map the remaining unmapped reads [50]. HTSeq-Count [51] was used to count reads mapping to genes annotated in.gtf files. Genes with <1 read mapping in at least half of the samples were removed prior to differential gene expression analysis using the EdgeR package (Bioconductor) [52,53]. FDR values were calculated using the Benjamini–Hochberg method. MDS and statistical analyses of the data were performed in R.
We utilised the Ensembl Compara database [54], combined with our expression data, to generate a table of orthologs with the following associated data for each species: Ensembl ID, Gene name, log2FC, and FDR following IFN treatment. The Compara database provides a thorough account of gene orthology based upon whole genomes available in Ensembl and thus provided us with a standardised approach by which to define phylogenetically the clusters of orthologous genes relative to the chicken taken as an outgroup in the orthology assignment. However, certain gene families relating to innate immunity (for example, the IFITM genes) have undergone lineage-specific expansion, potentially resulting in genes not being annotated and clustered within the database [55]. To account for the misannotation and absence of genes in Compara, we improved the table by manually checking (and annotating, if necessary) genes initially found to be ‘missing’ in either one, two, or three of the 10 species analysed in this study. For this subset of genes, we searched for the presence of an as-yet-unannotated ortholog in the Ensembl genomes using blastn. In cases where a clear ortholog was detected, we included this gene within the appropriate orthogroup. In total, we identified an additional 18 genes that were added to the final table. In cases where a gene was not annotated in the Ensembl genome but a sequence (or predicted sequence) for the homologous species was available in NCBI, we mapped the RNAseq reads to the gene of the homologous species using Bowtie2. The number of reads mapping from each sample was then counted. In cases where an ortholog was present in NCBI from a closely related species, a relaxed Bowtie2 algorithm was first used to map reads to the sequence. The consensus sequence of the resulting contig was then used to count the reads in individual samples using Bowtie2 as above. Differential expression values were then determined using EdgeR having appended the results to the HTseq file. In total, 64 orthologs were added to the table as a result of the orthologous sequence mapping approach. Finally, we modified the .gtf file of sheep to reannotate STAT4 (to become STAT1 and STAT4) and SOCS1, and manually annotated the ZCCHC2 gene in the cow .gtf file.
Using the human genome as a gold standard for annotation, we next assessed the authenticity of each “species-specific” ISG (i.e., an ISG present only in one of the 10 species analysed in this study) based upon its current annotation within the genomes. We first generated a subset of species-specific genes by applying an arbitrary cutoff of their differential expression of ≥2 log2FC. This cutoff resulted in a total of 102 genes across eight species. A total of 33 genes (32%) were in this category as a result of misannotation. In particular, the current little brown bat and pig genomes appear to be currently less well annotated compared to other genomes, as a large proportion (73% and 79%, respectively) of seemingly species-unique genes were in reality misannotations.
We used the database-integrated genome screening (DIGS) tool [37] to systematically screen the genomes of 111 mammalian species for sequences disclosing homology to 79 of the 90 coremamm ISGs. The peptide sequences of the human copies of these 79 genes (obtained from BioMart [56]) were used as probes for tBLASTn-based searches of each species genome. Sequences disclosing above-threshold similarity to peptide probes were extracted and classified by BLASTx-based comparison to a reference library. This library contained, for each ISG, peptide sequences of human paralogs and orthologs from selected additional mammal species. The DIGS tool captures screening results in a relational database, wherein they could be interrogated using structured query language. The number of significant matches for each gene was determined using a gene-specific bitscore cutoff. These counts were normalised to the median value across the mammalian genomes screened to account for variation in exon numbers. Because DIGS is based upon sequence-similarity screening, high counts for a particular gene do not necessarily reflect bona fide gene expansion. Where no matches to a given gene were identified and no ortholog had been annotated in Ensembl, we attempted to confirm that deletion had occurred by viewing the corresponding genomic region in the UCSC and Genomicus genome browsers [57,58] and by comparing alignments of the orthologous genomic regions derived from species with and without the deleted gene. Deletions could not always be confirmed mainly due to low coverage or relatively poor quality assembly of the available genomes.
Clusters comprising one-to-one orthologs present in the interferome of each species were extracted and filtered to check for the presence of a gene in species X matching a human Ensembl ID. The human Ensembl ID was then used to query BioMart to extract the corresponding pairwise dN and dS values for species X against the human. Values for differential expression (log2FC) and FDR values for species X were merged with the dN/dS values and histograms of both significant (ISG) and nonsignificant (non-ISG) clusters plotted. dN and dS values were not available in Ensembl for the Large flying fox or the chicken. The overall distributions of dN/dS values for ISGs and non-ISGs were compared using the Kruskal–Wallis rank sum test and Wilcoxon rank sum test.
In order to allow mining of our data by the wider research community, we created a web-based public interactive data server, accessible at http://isg.data.cvr.ac.uk. The server hosts a database containing the orthologous ISG clusters studied in this paper. This web tool allows users to search for and download orthologous clusters and the associated experimental results. The search may be based on various search criteria:
It should be noted that genes can have different aliases to those in Ensembl, and these must be checked if a gene that is not initially present, e.g., MDA5, is present as IFIH1. These aliases are stated in Ensembl.
The webserver for querying the dataset is available at http://isg.data.cvr.ac.uk/.
The DIGS blast framework is available at http://giffordlabcvr.github.io/DIGS-tool/.
The raw fastq files generated during this project have been submitted to the European Bioinformatics Institute (EBI) under project accession number PRJEB21332.
|
10.1371/journal.ppat.1003722 | Hepatitis B Virus Disrupts Mitochondrial Dynamics: Induces Fission and Mitophagy to Attenuate Apoptosis | Human hepatitis B virus (HBV) causes chronic hepatitis and is associated with the development of hepatocellular carcinoma. HBV infection alters mitochondrial metabolism. The selective removal of damaged mitochondria is essential for the maintenance of mitochondrial and cellular homeostasis. Here, we report that HBV shifts the balance of mitochondrial dynamics toward fission and mitophagy to attenuate the virus-induced apoptosis. HBV induced perinuclear clustering of mitochondria and triggered mitochondrial translocation of the dynamin-related protein (Drp1) by stimulating its phosphorylation at Ser616, leading to mitochondrial fission. HBV also stimulated the gene expression of Parkin, PINK1, and LC3B and induced Parkin recruitment to the mitochondria. Upon translocation to mitochondria, Parkin, an E3 ubiquitin ligase, underwent self-ubiquitination and facilitated the ubiquitination and degradation of its substrate Mitofusin 2 (Mfn2), a mediator of mitochondrial fusion. In addition to conventional immunofluorescence, a sensitive dual fluorescence reporter expressing mito-mRFP-EGFP fused in-frame to a mitochondrial targeting sequence was employed to observe the completion of the mitophagic process by delivery of the engulfed mitochondria to lysosomes for degradation. Furthermore, we demonstrate that viral HBx protein plays a central role in promoting aberrant mitochondrial dynamics either when expressed alone or in the context of viral genome. Perturbing mitophagy by silencing Parkin led to enhanced apoptotic signaling, suggesting that HBV-induced mitochondrial fission and mitophagy promote cell survival and possibly viral persistence. Altered mitochondrial dynamics associated with HBV infection may contribute to mitochondrial injury and liver disease pathogenesis.
| Hepatitis B virus (HBV) chronic infections represent the common cause for the development of hepatocellular carcinoma. Mitochondrial liver injury has been long recognized as one of the consequences of HBV infection during chronic hepatitis. Mitochondria are dynamic organelles that undergo fission, fusion, and selective-autophagic removal (mitophagy), in their pursuit to maintain mitochondrial homeostasis and meet cellular energy requirements. The clearance of damaged mitochondria is essential for the maintenance of mitochondrial and cellular homeostasis. We observed that HBV and its encoded HBx protein promoted mitochondrial fragmentation (fission) and mitophagy. HBV/HBx induced the expression and Ser616 phosphorylation of dynamin-related protein 1 (Drp1) and its subsequent translocation to the mitochondria, resulting in enhanced mitochondrial fragmentation. HBV also promoted the mitochondrial translocation of Parkin, a cytosolic E3 ubiquitin ligase, and subsequent mitophagy. Perturbation of mitophagy in HBV-infected cells resulted in enhanced mitochondrial apoptotic signaling. This shift of the mitochondrial dynamics towards enhanced fission and mitophagy is essential for the clearance of damaged mitochondria and serves to prevent apoptotic cell death of HBV-infected cells to facilitate persistent infection.
| Hepatitis B virus (HBV) infection affects nearly 350 million people worldwide and leads to chronic liver disease, liver failure, and hepatocellular carcinoma (HCC) [1], [2]. HBV is an enveloped DNA virus that belongs to the Hepadnavirus family. HBV DNA genome encodes four overlapping open reading frames designated as pre-S/S (the hepatitis B surface antigen, HBsAg), C (core/e antigen, HBc/eAg), P (polymerase, reverse transcriptase), and X (HBx) [1], [2]. HBx is a regulatory protein with multiple functions involved in various cellular and physiological processes including a key role in the maintenance of viral replication [3]. It is predominantly localized to the cytoplasm and also associates with mitochondria via its interaction with voltage-dependent anion-selective channel 3 (VDAC3) [4]–[7]. This association leads to a decrease in mitochondrial transmembrane potential (ΔΨm) and depolarization of mitochondria [3], [4], [8]. HBx also participates in activating transcription of whole host of cellular genes via protein-protein interactions both in the nucleus and cytoplasm [3], [7], [9]–[11]. HBx is not directly oncogenic but participates substantially in the process of liver oncogenesis [9], [12]. HBx is a regulatory protein with pleiotropic activities and has been shown to promote endoplasmic reticulum (ER) stress, oxidative stress, deregulation of cellular calcium homeostasis, and mitochondrial dysfunction [3]. HBx also modulates the activation of several latent transcription factors such as nuclear factor-kappa B (NF-κB), signal transducer and activator of transcription 3 (STAT-3), with resultant activation of cytoprotective genes [8], [13]. The multiple effects of HBx protein may be a consequence of the trigger of the ER-mitochondria-nuclear nexus of signal transduction pathways.
Mitochondrial injury and oxidative stress are prominent features of chronic Hepatitis B and C [14], [15]. Histological manifestation of swollen mitochondria and mitochondria lacking cristae directly implicates mitochondrial injury in HBV-associated liver disease pathogenesis. HBV infection is associated with deregulated cellular Ca2+ signaling, mitochondrial depolarization and dysfunction and reactive oxygen species (ROS) generation [3],[8],[16]. HBV-induced elevated cellular ROS levels can also promote mitochondrial dysfunction [15]. Dysfunctional or damaged mitochondria trigger a vicious cycle of mitochondrial damage and ROS generation, which is detrimental for cell survival and can be confounded by rapid turnover of damaged mitochondria [17]. The removal of dysfunctional mitochondria is orchestrated by asymmetric mitochondrial fission to eliminate the damaged mitochondria by subsequent mitophagy (selective autophagy of mitochondria) [18]. Mitochondria subjected to physiological stress usually undergo perinuclear clustering, which precedes both mitochondrial fission and mitophagy [19].
HBV and in particular, HBx have been shown to induce bulk autophagy [20]–[24]. In this study, we investigated HBV-induced aberrant mitochondria dynamics, and mitophagy. Our data revealed that HBV shifts the balance of mitochondrial dynamics towards enhanced fission and promotes selective autophagic degradation of damaged mitochondria via mitophagy. HBV triggered mitochondrial fission by promoting mitochondrial translocation of Drp1 via upregulation of Drp1 Ser616 phosphorylation. HBV upregulated proteins that mediate mitophagy and induced the elimination of dysfunctional mitochondria via mitophagy. More specifically, mitochondrial translocation of Parkin a cytosolic E3 ubiquitin ligase was observed in HBV/HBx expressing cells resulting in its self-ubiquitination and of its substrate, mitofusin 2 (Mfn2). In addition to confocal microscopy, using a novel dual fluorescence reporter Mito-mRFP-EGFP, we demonstrated that HBV/HBx induces complete mitophagy evident by fusion of mitophagosome with lysosome. Our studies also showed that HBx protein alone or in the context of HBV full genome, is a critical activator of HBV-induced aberrant mitochondrial dynamics. Further, we demonstrated that inhibition of mitophagy by silencing Parkin results in enhanced mitochondrial apoptotic signaling in HBV-infected cells, suggesting that induction of mitochondrial fission and subsequent mitophagy subvert apoptosis impending due to accrued mitochondrial injury in HBV-infected cells. In summary, our results suggest that HBV-mediated modulation of mitochondrial dynamics may promote cell viability of infected cells. We envisage that the altered mitochondrial dynamics and induction of mitophagy possibly contribute to the persistence of HBV-infected hepatocytes. However a careful and rigorous examination of HBV-induced mitochondrial regulation and its relevance to persistent phenotype of infected hepatocytes is required to confirm this finding in in vivo conditions.
We investigated the HBV-induced morphological changes of mitochondria in the human hepatoma Huh7 cells transiently expressing wild-type 1.3mer HBV genome (hereafter referred to as HBV). As shown in Figure 1A, distinct fragmented mitochondrial morphology was observed in HBV-expressing cells, compared to the typical tubular mitochondria in untransfected cells. HBx-expressing cells also displayed similar mitochondrial fragmentation (Figure 1B). HBV/HBx-expressing cells displayed prominent mitochondrial clustering in the perinuclear regions (Figure 1A and 1B), consistent with a previous report [19]. We then determined whether HBV infection triggered Drp1-mediated mitochondrial fission. As shown in Figure 1C, HBV stimulated both the expression and phosphorylation (Ser616) of Drp1. Mitochondrial translocation of Drp1 is modulated by phosphorylation at Ser616 by cyclin B/cyclin-dependent kinase 1 (Cdk1) [25]. HBV gene expression has been shown to stimulate cyclin B/Cdk1 [26]–[30]. Similar results were obtained in HBx-expressing cells (Figure S1C). Next, we examined if HBV induces Drp1 translocation to mitochondria by confocal microscopy. HBV-expressing cells displayed enhanced mitochondrial translocation of Drp1, compared to uninfected cells. (Figure 1D, see merged yellow spots). Using Drp1 antibody that specifically recognizes phosphorylated Ser616 residue, we demonstrated that most of Drp1 recruited to the mitochondria in HBV-infected cells is phosphorylated at Ser616 residue (Figure 1E). Similar results were obtained in HBx-expressing cells (Figure 1F). In support of these results, we demonstrated the accumulation of phosphorylated Drp1 in purified mitochondrial fraction from HepAD38 cells that stably express whole HBV genome under tetracycline-repressible promoter (Figure 1C, bottom panel) [31]. Together, these results indicate that HBV/HBx promotes mitochondrial fission via Drp1 translocation.
When mitochondria are depolarized, PTEN-induced putative kinase 1 (PINK1), a mitochondrial Ser/Thr kinase, accumulates on the outer mitochondrial membrane and recruits Parkin to the mitochondria [18]. Parkin translocation to depolarized mitochondria is a hallmark of mitophagy [18], [32], [33]. Thus, we examined mitochondrial translocation of Parkin in HBV-expressing cells by confocal microscopy. Significant Parkin translocation to mitochondria was observed in HBV-expressing cells, compared to untransfected cells (Figure 2A, see merged yellow spots). Similar results were also observed in HBx-expressing cells (Figures S1A and B). Parkin is an E3 ubiquitin ligase which ubiquitinates itself and its mitochondrial substrates, Mfn2 and voltage-dependent anion-selective channel 1 (VDAC1) [18]. Mfn2 functions to promote mitochondrial fusion and its degradation by HBV will favor fission activities [34]. Using purified mitochondrial and cytosolic fractions of HBV-expressing cells, we further analyzed mitochondrial translocation of Parkin. As shown in Figure 2B, Parkin is accumulated and mostly ubiquitinated in the mitochondrial fraction. Parkin ubiquitination was further analyzed by immunoprecipitation with anti-Parkin antibody, followed by subsequent Western blotting with anti-ubiquitin antibody (Figure 2C). We also observed a moderate decrease in the expression levels of VDAC1 (Figure 2B, lanes 2 and 6), whereas Mfn2 is significantly degraded with a concomitant increase in its ubiquitinated form (Figure 2D, first and third panels, lane 3). Parkin-dependent ubiquitination of Mfn2 in HBV-expressing cells was verified by immunoprecipitation assay using HBV-expressing cells with Parkin knockdown (Figure 2D, upper panel, lane 4). These results indicate that HBV stimulates Parkin translocation to mitochondria and Parkin-dependent degradation of Mfn2 by ubiquitination.
We then investigated whether HBV stimulates the expression of mitophagy-related genes. As shown in Figures 2E and F, HBV gene expression modestly stimulated the expression of Parkin, PINK1, and microtubule-associated protein 1 light chain 3B (LC3B) at both the mRNA and protein levels in HepAD38 cells, but not in HepG2 cells, the parental cell line of HepAD38. An increase in LC3B-I and LC3B-II was also observed (Figure 2F). These results were further confirmed in Huh7 cells transiently expressing HBV genome (Figure 2G). Activating transcription factor 4 (ATF4) was also stimulated by HBV. ATF4 is a transcriptional factor that stimulates Parkin gene expression via unfolded protein response (UPR) [35] (Figure 2F). HBx-expressing cells showed a similar stimulation of Parkin, LC3B-I, and LC3B-II (Figure S1C). By coimmunoprecipitation assay using whole cell lysates extracted from Huh7 cells co-expressing Flag-HBx and mCherry-Parkin protein, physical interaction between HBx and Parkin is shown in Figure S1D. HBx-VDAC3 interaction has been previously reported [4]. Since Parkin binds VDAC [36], it is tempting to speculate that Parkin, VDAC, and HBx may form a ternary complex to expedite the process of Parkin recruitment to mitochondria. In summary, these results demonstrate that HBV/HBx stimulated the expression of mitophagy-related genes.
Next, we analyzed the formation of mitophagosome in Huh7 cells co-expressing HBV and GFP-LC3 protein by confocal microscopy, as we demonstrated the HBV-induced mitochondrial fission and stimulation of Parkin and LC3B expression in Huh7 cells (Figures 1 and 2G). As shown in Figure 3A, we observed that Parkin-containing mitochondria are associated with GFP-LC3 puncta in HBV-expressing cells (see white puncta in the zoomed images). These GFP-LC3 puncta were abrogated upon treatment of cells with autophagy inhibitor 3-methyladenine (3-MA), but not Bafilomycin A1 (BafA1) (Figure 3A). 3-MA inhibits autophagosome formation, whereas BafA1 inhibits fusion of autophagosomes with lysosomes [37], [38]. Quantitative analysis of these results is presented in Figures 3B, C, and D. HBx alone was also capable of forming Parkin-associated mitophagosome (Figure S2A). In contrast, cells expressing whole HBV genome defective in HBx expression (HBV-ΔX) failed to show any GFP-LC3 puncta containing mitochondria, (Figure S2). Quantitative analysis of these images is also presented in Figures S2C, D, and E. Together, these results indicate that HBV/HBx induces Parkin-dependent mitophagosome formation.
The final step of mitophagy is the fusion of mitophagosomes with lysosomes where the cargo is delivered, degraded and recycled [18]. To analyze the progression of mitophagy from mitophagosomes to lysosomes, we made use of a tandem-tagged RFP-EGFP chimeric plasmid pAT016 encoding a mitochondrial targeting signal sequences fused in-frame with RFP and EGFP genes (Figure 4A), which exploits the differential stabilities of RFP and GFP [39]. GFP signal is quenched in lower pH, while RFP can be visualized in both mitophagosomes and acidic mitophagolysosomes thus the prevalence of RFP fluorescence in the lysosomes indicates completion of mitophagic process. In contrast, under normal conditions, yellow structures indicating merged images of GFP and RFP that localize to mitochondria are observed (Figure 4A). Cells cotransfected with plasmid pAT016 (mito-mRFP-EGFP) and whole HBV genome or, HBV-ΔX, or HBx-flag expressing plasmid, respectively, were analyzed by confocal microscopy. In control cells not expressing HBV or expressing HBV-ΔX, the yellow merged fluorescence that indicates the presence of both EGFP and mRFP in mitochondria was observed (Figure 4B). In contrast, the cells expressing whole HBV genome or HBx displayed distinct red puncta, and fewer green puncta because mRFP is more stable in the lysosomes [39]. When EGFP and mRFP signals are merged, only red puncta were observed in mitophagolysosomes as EGFP signal is quenched in lower pH (Figure 4B). Quantitation of these puncta is presented in Figure 4C. We continued to examine HBV/HBx-induced mitophagolysosome formation by conventional confocal microscopy. As shown in Figures 4D and S3A, GFP-LC3 puncta-containing mitochondria associated with lysosomes were observed in both HBV- and HBx-expressing cells (see white puncta indicating merged images of GFP-LC3, TOM20, and LysoTracker). Treatment with 3-MA and BafA1 abrogated these puncta (Figures 4D and S3A). Quantitative analysis of these results is shown in Figures 4E and S3B, respectively. Collectively, these results demonstrate that HBV and HBx, either expressed alone or in the context of whole HBV genome, respectively, induce complete mitophagy.
As damaged mitochondria are eliminated via mitophagy, we observed a decline in mitochondrial number in HBV-expressing cells, but not in those treated with BafA1 or untransfected cells (Figures S3C and D). Parkin knockdown in HBV-expressing cells also abrogated the decline in mitochondrial number (Figures S3E and F). It should be emphasized that not all mitochondria in HBV/HBx-expressing cells engage in mitophagy. The fraction that undergoes mitophagy and fission is likely to impact the disease process.
Mitochondrial dynamics is integrally linked to apoptosis [40], [41]. We investigated if HBV enhances mitochondrial fission and mitophagy to modulate apoptotic cell death associated with mitochondrial injury. HepAD38 cells were transfected with Parkin-specific siRNA pool and analyzed for changes in mitochondrial apoptotic signaling pathway. Parkin silencing induced massive cytochrome C release from mitochondria to cytosol and promoted cleavage of poly(ADP-Ribose) polymerase (PARP) and caspase-3, activation of caspase-3/7, and prompted apoptosis as seen by TUNEL assay (Figures 5A–C). These results strongly suggest that HBV-induced mitochondrial dynamics protects virus-infected hepatocytes from apoptotic cell death to facilitate the persistent virus infection (Figure 5D).
Mitochondria are dynamic organelles that undergo fission (fragmented mitochondria), fusion (tubular mitochondrial network), and trafficking [18]. The rapid modulation of mitochondrial dynamics occurs in response to physiological stress, apoptotic stimuli, metabolic demands, and infections [42]. Perturbation in mitochondrial dynamics is involved in many diseases such as neurodegenerative disorders and cardiovascular diseases, underscoring the pivotal importance of this process in maintenance of mitochondrial and cellular homeostasis [42]. Clearance of damaged mitochondria is proposed is orchestrated by asymmetric mitochondrial fission and subsequent elimination of damaged mitochondrial pool by selective mitochondria autophagy (mitophagy) [18]. Fragmented mitochondria are better substrates for removal by mitophagy and elongated or fused mitochondria resist mitophagic degradation, suggesting that mitochondrial dynamics and mitophagy are two critical arms required for maintenance of mitochondrial homeostasis [18], [43]. HBV gene expression is associated with physiological aberrations such as perturbed calcium homeostasis and elevated ROS levels which promote mitochondrial dysfunction and damage [3], [6], [8]. We observed that HBV/HBx expression triggers sequential cascade of events with mitochondria, initiating with their perinuclear accumulation followed by mitochondrial fission and ultimately the clearance of damaged mitochondria by mitophagy. Clearance of damaged mitochondria via mitophagy is crucial for establishing mitochondrial homeostasis and cell survival [18], [43]. Impairment of this process by mutations in mitophagy-mediating genes such PINK1 or Parkin is linked to hereditary forms of Parkinson's disease, a neurodegenerative disorder, emphasizing the vital role of mitochondrial homeostasis in cell survival [44], [45].
Many DNA and RNA viruses have been shown to tightly modulate the autophagy process to prevent their clearance by autophagy, to inhibit host immune response, or to favor viral replication and maturation events [46]. Hepatitis C virus, a positive-sense single-stranded RNA virus, induces Parkin-mediated selective mitophagy which may benefit viral replication [32]. Recently, several reports have described that HBV promotes bulk autophagy to favor its own replication [20]–[24]. Although the molecular mechanisms responsible for induction of autophagy during HBV infection are unclear, it seems that autophagy can either enhance HBV DNA replication or favor HBV envelopment [20]–[24]. HBV may promote autophagy either directly via viral factors that trigger autophagy, or indirectly by mechanisms mediated by virus-induced physiological aberrations and stress. Here, our data suggests a functional crosstalk between virus and autophagy pathways providing evidence that selective, rather than bulk autophagy is probably involved in promoting the viability of HBV-infected hepatocytes from imminent cell death due to the virus-mediated mitochondrial injury. This is in line with the emerging concept that selective autophagy of organelles contributes to tight control of host-pathogen interactions [46].
Although the role of HBV and HBx in regulating apoptotic signaling has long been debated [47], [48], based on the in vivo studies, it is inferred that HBV-infected hepatocytes maintain a persistent phenotype and that HBV replicates within infected hepatocytes noncytopathically [49]. The noncytopathic viruses have evolved mechanisms to mitigate the adverse effects of the pathophysiological perturbations manifested in the host cells due to recurrent infection. Here, we demonstrate that HBV modulates mitochondrial dynamics to promote mitochondrial fission and subsequent clearance of damaged mitochondria by mitophagy. Damaged mitochondria are a major source of ROS and can trigger a continuous and vicious cycle of subsequent damage to healthy mitochondria followed by ROS generation, ultimately leading to cell death [17]. Hence, a rapid turnover and clearance of damaged mitochondria is needed to confound imminent cell death due to mitochondrial injury accrued during HBV infection and sustain the viability of the infected cells. In agreement with this assumption, we observed a surge in mitochondrial-apoptotic signaling and resultant death of the HBV-infected hepatocytes upon inhibition of the mitophagy pathway. Hepatocyte damage commonly observed during chronic hepatitis is widely believed to be immune-mediated [49] and the failure to mount an efficient immune response to eliminate the infected hepatocytes is considered the primary cause for persistence of chronic infections, including HBV. However the intracellular mechanism(s), which prevents cell death and support the viability of infected cells in conditions of virus-induced adverse intracellular physiology contributes to the lack of virus-induced cytopathology in non-lytic viruses and probably play a significant role in persistence of non-lytic chronic infections. However, further studies are required to confirm the significance of our findings in in vivo conditions. Currently, the most convenient way to test our hypothesis in vivo settings is the use of animal models of HBV infection such as Woodchucks, Ducks, or Tupaia to determine if abrogation of Parkin-mediated mitophagy pathway promotes specific death of HBV-infected hepatocytes and alleviates persistent HBV infection in these animals.
It is believed that HBx sensitizes and indirectly participates in the onset of liver oncogenesis [3], [47]. Its pleiotropic functions in activating host gene expression, interaction with numerous cellular targets, and its imminent role in altering mitochondrial physiology, promoting oxidative stress, and affecting the epigenetic changes in the chromatin collectively influence the early steps of liver neoplasia. Although not directly linked, HBx-mediated mitochondrial damage and aberrant mitochondrial dynamic may also contribute in pathogenesis of liver disease and HCC [50], [51].
In summary, this study provides a unique insight into the probable involvement of HBV-induced altered mitochondrial dynamics and mitophagy in possibly facilitating the persistence of infection and pathogenesis of liver disease associated with infection thus unraveling potential newer avenues for design of novel therapeutics against chronic HBV infection.
Human hepatoma cell line Huh7 was grown in high-glucose DMEM (Gibco) supplemented with 10% fetal bovine serum (Hyclone), 1% MEM non-essential amino acid (Gibco), 100 units/ml penicillin (Gibco), and 100 µg/ml streptomycin (Gibco). Human hepatoma HepG2 and HepAD38 cells harboring HBV full-length genome were maintained in RPMI 1640 (Gibco) supplemented with 20% fetal bovine serum, 1% MEM non-essential amino acid, 100 units/ml penicillin, and 100 µg/ml streptomycin. In addition, HepAD38 cells were grown in the presence of 0.5 mg/ml G418 (Invitrogen) and 1 µg/ml tetracycline. NT-KD (expressing non-target shRNA) and P-KD (expressing Parkin shRNA) cells used in this study were maintained in the presence of 2.5 µg/ml of puromycin, as described previously and knockdown level of Parkin gene was shown in the previous study [32].
The pHBV1.3mer and pHBV-ΔX plasmid DNAs encoding wild-type HBV genome and HBx-deficient HBV genome, respectively, were a kind gift from Dr. Jing-hsiung James Ou (University of Southern California). The pHBx-flag plasmid DNA was described previously [4]. The pEGFP-LC3 plasmid DNA was a kind gift from Dr. Tamotsu Yoshimori (National Institute of Genetics, Japan). The mCherry-Parkin plasmid DNA (plasmid #23956) was obtained from Addgene (a generous gift of Dr. Richard Youle, National Institute of Health, Bethesda, MD). HepAd38 cells were a kind gift of Dr. Christoph Seeger (Fox Chase Cancer Center, Philadelphia, PA). To create plasmid pAT016 (p-mito-mRFP-EGFP), plasmid ptfLC3 (Addgene plasmid #2174, a generous gift of Dr. Tamotsu Yoshimori) was double digested with Bgl2 and BamHI to remove LC3 coding sequences, resulting in plasmid p-mRFP-EGFP production and then, polymerase chain reaction product for mitochondrial targeting signal sequences of human cytochrome c oxidase subunit VIII amplified from pEYFP-mito (Clontech) was inserted N-terminally in frame into p-mRFP-EGFP.
To conduct laser scanning confocal microscopy, the cells grown on coverslips were transfected with the indicated plasmid DNAs followed by immunofluorescence assay, as described previously [32]. Images were visualized under a 60× or 100× oil objectives using an Olympus FluoView 1000 confocal microscope. Quantification of images (at least 10 cells per each sample) was conducted with ImageJ and MBF ImageJ softwares.
Chemical reagents used in this study were Bafilomycin A1 (Enzo Life Sciences) and 3-Methyladenine (Sigma). Primary antibodies used in this study include the following: rabbit monoclonal anti-Drp1 (Cell Signaling); rabbit monoclonal anti-phospho-Drp1 (S616) (Cell Signaling); rabbit monoclonal anti-ATF4 (Cell Signaling); rabbit polyclonal anti-Parkin (Abcam); rabbit monoclonal anti-LC3B (Cell Signaling); rabbit polyclonal anti-PINK1 (Abcam); rabbit polyclonal anti-VDAC1 (Cell Signaling); mouse monoclonal anti-Mfn2 (Abcam); goat polyclonal anti-VDAC3 (Santa Cruz); rabbit polyclonal anti-GAPDH (Santa Cruz); goat polyclonal anti-β-actin (Santa Cruz); mouse monoclonal anti-TOM20 (BD); rabbit polyclonal anti-TOM20 (Abcam); mouse monoclonal anti-Flag M2 (Sigma); rabbit polyclonal anti-DTKDDDDK-tag (GenScript); goat polyclonal anti-DDDDK (Abcam); mouse monoclonal anti-HBsAg (Thermo Scientific); mouse monoclonal anti-Ubiquitin (Cell Signaling); rabbit monoclonal anti-cleaved PARP (Cell Signaling); rabbit monoclonal anti-cleaved caspase-3 (Cell Signaling); rabbit polyclonal anti-cytochrome c (Cell Signaling); normal rabbit IgG (Cell Signaling); normal mouse IgG (Santa Cruz). The secondary antibodies used for immunofluorescence were Alexa Fluor 350, 488, 594, or 647 donkey anti-mouse, rabbit, or goat IgG (Molecular Probe). The secondary antibodies used for Western blot analysis were HRP-conjugated anti-mouse IgG (Cell Signaling), HRP-conjugated anti-rabbit IgG (Cell Signaling), and HRP-conjugated anti-goat IgG (Jackson Laboratories).
Small interfering RNA (siRNA) pools used in this study were siGENOME SMARTpool for Parkin (NM_004562) and non-targeting #1 control (NT) (Dharmacon). The cells were transfected with siRNA (50 nM) for the indicated times using DharmaFECT 4 transfection reagent according to the manufacturer's instructions (Dharmacon).
To analyze the expression levels of Parkin, PINK1, and LC3B genes, total cellular RNA and subsequent complementary DNAs were prepared, as described previously [32]. The RNA levels of Parkin, PINK1, and LC3B were quantified by real-time qRT-PCR using DyNAmo HS SYBR Green qPCR kit (Finnzymes). The following primer sets were used for RT-PCR: Parkin forward, 5′-TACGTGCACAGACGTCAGGAG; Parkin reverse, 5′-GACAGCCAGCCACACAAGGC; PINK1 forward, 5′-GGGGAGTATGGAGCAGTCAC; PINK1 reverse, 5′-CATCAGGGTAGTCGACCAGG; LC3B forward, 5′- GAGAAGACCTTCAAGCAGCG; LC3B reverse, 5′- AAGCTGCTTCTCACCCTTGT; GAPDH forward, 5′-GCCATCAATGACCCCTTCATT; and GAPDH reverse, 5′-TTGACGGTGCCATGGAATTT. Real-time qPCR was conducted by using an ABI PRISM 7000 Sequence Detection System (Applied Biosystems).
For Western blot analysis, whole cell lysates (WCL) were extracted from cells, subjected to SDS-PAGE, transferred to nitrocellulose membrane (Thermo Scientific), and Western blot analyzed with antibodies against the indicated proteins, as described previously [32]. For analysis of ubiquitinated Parkin and Mfn2 in WCL, immunoprecipitates were prepared followed by Western blotting with anti-ubiquitin antibody, as described previously [32]. For co-immunoprecipitation, Huh7 cells co-transfected with HBx-flag and mCherry-Parkin were suspended in 0.1 ml of RIPA buffer. The suspended cells were incubated for 20 min on ice and clarified by centrifugation at 15,000×g at 4°C for 20 min. The supernatant was mixed with 1.9 ml of RIPA buffer without SDS and immunoprecipitated with anti-flag antibody and protein-G Sepharose. The immunoprecipitates were Western blot analyzed with anti-Parkin, flag, and VDAC3 antibodies, respectively. The intensity of protein expression was quantified by ImageJ software.
To isolate pure cytosolic and mitochondrial fraction, HepAD38 cells were homogenized and isolated, as described previously [32], [52].
The activity of caspase-3/7 in HepG2 and HepAD38 cells transfected with siRNA were measured by using Caspase-Glo 3/7 assay kit according to the manufacturer's instructions (Promega).
Apoptotic cells death in HepG2 and HepAD38 cells transfected with siRNA were measured by using Click-iT TUNEL Alexa Fluor 488 imaging assay kit according to the manufacturer's instructions (Invitrogen). For quantitative analysis, at least 1,000 cells on immunofluorescence image were counted.
Statistical analyses using Student's t-test were performed by using Sigma Plot software (Systat Software Inc., San Jose, CA, USA).
|
10.1371/journal.ppat.1004526 | A Natural Genetic Variant of Granzyme B Confers Lethality to a Common Viral Infection | Many immune response genes are highly polymorphic, consistent with the selective pressure imposed by pathogens over evolutionary time, and the need to balance infection control with the risk of auto-immunity. Epidemiological and genomic studies have identified many genetic variants that confer susceptibility or resistance to pathogenic micro-organisms. While extensive polymorphism has been reported for the granzyme B (GzmB) gene, its relevance to pathogen immunity is unexplored. Here, we describe the biochemical and cytotoxic functions of a common allele of GzmB (GzmBW) common in wild mouse. While retaining ‘Asp-ase’ activity, GzmBW has substrate preferences that differ considerably from GzmBP, which is common to all inbred strains. In vitro, GzmBW preferentially cleaves recombinant Bid, whereas GzmBP activates pro-caspases directly. Recombinant GzmBW and GzmBP induced equivalent apoptosis of uninfected targets cells when delivered with perforin in vitro. Nonetheless, mice homozygous for GzmBW were unable to control murine cytomegalovirus (MCMV) infection, and succumbed as a result of excessive liver damage. Although similar numbers of anti-viral CD8 T cells were generated in both mouse strains, GzmBW-expressing CD8 T cells isolated from infected mice were unable to kill MCMV-infected targets in vitro. Our results suggest that known virally-encoded inhibitors of the intrinsic (mitochondrial) apoptotic pathway account for the increased susceptibility of GzmBW mice to MCMV. We conclude that different natural variants of GzmB have a profound impact on the immune response to a common and authentic viral pathogen.
| Granzymes (Gzm) are serine proteases expressed by cytotoxic T cells and natural killer cells, and are important for the destruction of virally infected cells. To date, the function of these molecules has been assessed exclusively in common laboratory mouse strains that express identical granzyme proteins. In wild mouse populations, variants of granzyme B have been identified, but how these function, especially in the context of infections, is unknown. We have generated a novel mouse strain expressing a granzyme B variant found in wild mice (GzmBW), and exposed these mice to viral infections. The substrates cleaved by GzmBW were found to differ significantly from those cleaved by the GzmBP protein, which is normally expressed by laboratory mice. Alterations in substrate specificity resulted in GzmBW mice being significantly more susceptible to infection with murine cytomegalovirus, a common mouse pathogen. Our findings demonstrate that polymorphisms in granzyme B can profoundly affect the outcome of infections with some viral pathogens.
| Cytotoxic lymphocytes, such as natural killer (NK) cells and CD8 T cells, are essential for the elimination of tumour cells or cells infected with intracellular pathogens. One mechanism cytotoxic lymphocytes utilize to initiate the destruction of target cells is the exocytosis of granules containing perforin (Pfp) and a family of serine proteases known as granzymes (Gzms) [1]. Pfp facilitates the entry of Gzms into the cytoplasm of target cells, where the Gzms cleave specific proteins triggering death of the target. Multiple Gzms have been identified in both humans and the mouse, with GzmA and GzmB being the most abundant and best characterized in both species. While non-cytotoxic functions of Gzms have been described, inducing target cell death appears to be a major function of GzmA and GzmB, and the increased sensitivity of mice lacking these proteins to infection with ectromelia virus (ECTV) and murine cytomegalovirus (MCMV) has been attributed to the role of the Gzms in the killing of infected cells [2]–[4]. Unlike GzmB, which is universally agreed to induce apoptosis [5], the mechanism employed by GzmA to induce cell death remains controversial [6]–[8]; however, it is agreed that this mechanism does not require activated caspases.
Human and mouse GzmB share extensive sequence homology and thus were predicted to kill cells by the same mechanism. However, amino acids that influence substrate binding differ between human and mouse GzmB, with the two proteins now recognized to have different substrate preferences [9]–[11]. A significant difference between the two proteins is that human, but not mouse GzmB, efficiently cleaves the BH3-only protein Bid [10], [12], [13]. Once cleaved, tBid is capable of inducing permeabilization of the mitochondrial outer membrane (MOMP) resulting in the release of pro-apoptotic mediators that ultimately activate a caspase cascade. The finding that cells lacking Bid or overexpressing Bcl-2 survive treatment with human GzmB is consistent with the theory that human GzmB indirectly activates caspases [12], [14], [15]. By contrast, mouse GzmB appears to mediate its effects by directly processing pro-caspases to their active form, and does not require MOMP in order to induce apoptosis [9], [10]. Thus, while both human and mouse GzmB efficiently induce the death of target cells, they achieve this by different mechanisms.
Many pathogens inhibit apoptotic pathways as a means of survival. The differences in mouse and human GzmB substrate specificity may therefore have arisen in response to pathogens targeting different apoptotic pathways in humans and mice. Alternatively, the need to directly target proteins produced by species-specific pathogens could have driven the divergence in GzmB substrate specificities. For example, GzmB inhibits the reactivation of HSV-1 by cleaving the virally encoded ICP4 protein [16]. Similarly, GzmH and GzmB cooperate to suppresses the spread of human adenovirus V by degrading viral proteins essential for replication [17]. Further evidence that selective pressure from pathogens has contributed to changes in GzmB has come from the finding that GzmB polymorphisms exist. In humans, a limited degree of GzmB polymorphism has been described [18], however, the significance of this finding is unclear as there is no difference in the proteolytic activities of the two common alleles and both have equivalent biochemical and cytotoxic functions, at least in vitro [19]. In the mouse, 13 common inbred laboratory mouse strains that were tested show no GzmB sequence variation, a finding that probably reflects the limited gene pool from which laboratory strains were derived [20]. By contrast, significant variation in the GzmB sequence was noted in wild mice or wild-derived inbred mouse strains, including in some of the residues that line the substrate cleft. Interestingly, the single allele found in all the inbred mouse strains was relatively uncommon in the wild, found in <20% of isolates [20].
Responses to pathogen challenges have been investigated almost exclusively using inbred mouse strains including knock-out mice rendered genetically deficient in GzmB, but the important question as to whether polymorphisms in GzmB influence the outcome of infections with common pathogens has not been addressed. Here, we have characterized a GzmB allele present in wild mice and found that although its substrate specificity differs from that of GzmB encoded by C57BL/6 (B6) mice, its in vitro cytotoxic potential is identical to that of the B6 allele. Nevertheless, substitution of the GzmB allele encoded by B6 mice with the GzmB allele from wild mice led to the inability to effectively control MCMV infection. These data provide novel insights about the relevance of GzmB polymorphisms and demonstrate that polymorphisms in GzmB significantly influence the response to infection with a common, natural viral pathogen.
We had previously shown that the mouse GzmB gene is highly polymorphic amongst outbred mouse populations and that some of the polymorphic residues are predicted to impinge on the substrate binding pocket and to potentially influence fine protease specificity [20]. To investigate this further, we selected a wild (w) mouse GzmB allele that is markedly divergent from the allele common to B6 mice as well as the 13 inbred mouse strains we previously typed. For clarity we will refer to the prototype B6 inbred allele as GzmBP, and the outbred wild allele as GzmBW. GzmBW encodes 13 differences in amino acid sequence from GzmBP or 94.7% amino acid identity over the entire 247 amino acid sequence (Fig. 1A).
We expressed and purified recombinant GzmBW and GzmBP in Pichia pastoris yeast cells, as previously described [21]. Both forms were indistinguishable in their ability to cleave the generic GzmB substrate AAD-SBzl (Fig. 1B), indicating that GzmBW possesses classic Asp-ase activity. Both forms also bound Serpinb9, the intracellular inhibitor of GzmB [22]–[24](Fig. 1C), suggesting that GzmBP and GzmBW are subject to similar regulation in vivo. Inhibition of any serine proteases by its cognate serpin can only occur if the protease can correctly recognize and cleave the relatively unstructured reactive site loop of the serpin. This further confirmed the proteolytic activity of GzmBW, and its cleavage after aspartate [22].
We have previously shown that GzmBP differs from human GzmB in that it cleaves peptide substrates based on Bid (e.g. Abz-IEPDSESQK-dnp) very poorly, and prefers substrates with an aromatic P2 residue and glycine at P2′ ([10] and Table 1). Strikingly, GzmBW cleaves peptide substrates based on Bid over 100 times more efficiently than GzmBP, and three-fold more efficiently than human GzmB (Table 1). Like GzmBP, GzmBW prefers substrates with glycine at P2′, but cleaves substrates with an aromatic P2 residue four- to five-fold less efficiently. To confirm these differences in relation to broader substrate specificity, both forms were used in a substrate phage display experiment on a library with Asp fixed at the P1 position [10]. The results indicated very similar requirements at the P4 position (I/L), as well as a requirement for glycine at P2′ (Figure S1). However, the strong preference of GzmBP for aromatic residues at P2 was not conserved in GzmBW, consistent with the peptide substrate results (Table 1).
We next examined whether the differences in turnover of Bid (IEPD) substrates is also reflected in a different capacity to cleave intact Bid or effector pro-caspases 3 and 7. We found that GzmBW is 50–100 fold more efficient than GzmBP at cleaving recombinant Bid, but far less efficient at activating pro-caspase-3 or pro-caspapse-7 (Fig. 2). To determine whether these changes in substrate preference translate into a variable capacity to kill target cells, we exposed P815 (mouse mastocytoma), EL-4 (mouse thymoma), HeLa (human cervical cancer) and Jurkat (human T lymphoma) cells to graded doses of each GzmB form and very low (‘sublytic’) quantities of purified recombinant Pfp (Fig. 3). For each cell line, we found no significant difference in susceptibility to apoptosis. Overall, we concluded that there is no significant difference in the intrinsic pro-apoptotic activity of GzmB expressed by the inbred laboratory mice and outbred wild mice.
To examine the role of the GzmBW allele in more physiological settings, we generated a congenic mouse strain carrying the w allele. This was achieved by backcrossing the w/w mouse for >20 generations with B6, at each generation selecting for the w allele. This strain was also crossed with B6.OT1 mice to create the GzmBW/W.OT1 strain, so that antigen restricted CTLs could be studied in both strains (see below). An alternative and probably less laborious approach to deriving the congenic line over so many generations would have been to develop a ‘knock in’ of the W allele on P strain. However, we and other groups have not been successful in this approach, due to the large number of highly homologous granzyme gene sequences closely linked in the GzmB locus.
Using whole exome DNA sequencing we confirmed that the GzmB locus on chromosome 14 including the GzmB gene, all 6 linked Gzm genes and the gene encoding Cathepsin G (expressed in neutrophils, but not cytotoxic lymphocytes) were derived from w. The genetic interval derived from w comprised approximately 18% of the chromosome 14 DNA. Overall, the genetic content of the w/w mouse was >99.1% derived from B6. Having derived the DNA sequence of the 0.9% of the genome that remained from the w strain, we were able to assess whether the high degree of polymorphism identified for GzmB was also the case for the other granzymes (GzmC-G and N). This was not the case: whereas >5.3% of the amino acids of GzmB differed between the two allotypes, the corresponding figure across the other 6 genes was <0.3% (a total of four polymorphic residues out of 1362 across the 6 coding regions; P<0.0001) (Table 2).
We wished to establish beyond doubt that our in vitro enzymatic findings, which were derived with purified recombinant proteases, would be replicated in bona fide antigen restricted CTLs, which, along with NK cells, are the authentic physiological context for granzyme expression. We therefore generated OVA257 specific activated T cells from the spleens of B6 and GzmBW/W congenic OT1 mice and tested the resultant cell lysates on peptide substrates. The generic Asp-ase substrate AAD-SBzl was cleaved with similar efficiency by B6 and w/w T cell lysates (Fig. 4A). By contrast, the tetrapeptide substrate IEPD was cleaved more efficiently by w/w lysate (with >3 times the maximum velocity of p/p, p<0.05), confirming differences in the fine specificity of the two alloforms of GzmB observed with in vitro-generated proteins (Fig. 4B). Structural predictions for the granzymes whose genes are closely linked to GzmB, indicate that these proteases should have chymotrypsin-like (‘chymase’) activity and cleave after hydrophobic P1 residues [25] as has been clearly demonstrated for the human orthologue GzmH [26]. As expected, the respective CTL lysates from the w and p mice had indistinguishable chymase activity (Figure 4C), further confirming that the minimal changes in amino acid sequence among the chymase granzymes had no impact on substrate preference. There was no turnover in AAD or IEPD in T cells derived from GzmAB-null mice, while B6 mice deficient in perforin (whose gene maps to Chr 10) showed similar activity to wild type B6 mice (Fig. 4A and C). Western blot of cell lysates also showed no quantitative difference in GzmB expression across the various strains tested (Fig. 4D).
Our in vitro analyses determined that the substrate specificity of GzmBW differs from that of the B6 allele. Since the Gzms have been shown to play a critical role in viral infections, we next examined whether the differences in substrate specificity observed in vitro and ex vivo can lead to functional differences after infection with bona fide mouse pathogens. MCMV, a natural pathogen of mice, is partly controlled by activities mediated by GzmB [4]. Thus, the effect of GzmBW on the ability of the host to control MCMV infection was investigated. In B6 mice, NK cells rapidly control MCMV infection via activation mediated by engagement of the Ly49H activating NK cell receptor. However, in the wild most (≥80%) MCMV variants encode m157 proteins that are unable to activate NK cells [27] and indeed, the frequency of Ly49H-resistance is rare in outbred wild mice [28]. These findings indicate that B6-like Ly49H-m157 interactions are not a feature of host–MCMV interactions in the wild. Thus, to examine the role of GzmBW/W in a setting that reflects the situation in wild mouse populations, we utilized a virus lacking the m157 viral protein (Δm157). In the absence of m157, MCMV replicates to high titers in the visceral organs of B6 mice, and is eventually controlled primarily by cytotoxic CD8 T cells. Unlike B6 mice (GzmBP/P), infection of the GzmBW/W mice with Δm157 MCMV resulted in rapid mortality (Fig. 5A). At day 7 post-infection, viral loads in the spleens and lungs of GzmBW/W mice were not significantly different from those observed in B6 mice (Fig. 5B). By contrast, the livers of GzmBW/W mice contained approximately 10 fold more virus than livers of B6 mice (Fig. 5B). Histological analysis of GzmBW/W livers harvested at day 6 post-infection revealed significant areas of focal necrosis and diffuse cellular infiltrates, while in B6 mice no significant damage was evident in liver sections (Fig. 5C). The advanced liver damage observed in GzmBW/W mice by histological analysis was also confirmed by measuring circulating liver transaminase levels. Serum levels of the liver enzymes alanine aminotransferase (ALT) and aspartate aminotransferase (AST) in GzmBW/W mice were significantly elevated at day 6 post-infection (Fig. 5D). These data indicate that GzmBW/W mice have an impaired response to Δm157 MCMV infection that manifests as significantly higher viral loads within the liver, and tissue damage to this organ, which markedly increases mortality.
GzmB, along with GzmA, produced by cytotoxic CD8 T cells is essential for host defense against the poxvirus ECTV [2], [29]. ECTV is a large DNA virus that is the causative agent of mousepox. GzmBW/W mice infected with 105 pfu of the Moscow strain of ETCV succumbed to infection at rate equivalent to that of B6 mice (Figure S2A), and viral load in the blood of GzmBW/W mice at day 8 post-infection was not significantly different to that of B6 mice (Figure S2B). Since mice that lack GzmB are 100-fold more susceptible to ECTV infection [2], these data indicate that GzmBW can substitute for the B6 allele of GzmB in the context of ECTV infection, and provide independent evidence that the GzmBW allele is functional not only in vitro (Table 1), but also in vivo.
The outcome of infection with the Δm157 MCMV virus was then compared in GzmBW/W mice and mice lacking GzmA, GzmB or both GzmA and B. Viral loads were measured in target organs (spleen, liver and lungs) at days 4 and 6 post-infection by plaque assay. Experiments were terminated at day 6 post-infection as GzmBW/W mice become highly sensitive to infection after this time. Viral loads in mice deficient for GzmA were equivalent to those of B6 mice in all organs tested, at both day 4 and 6 (Fig. 6), suggesting that, at least during the acute phase of infection, GzmA is not required for viral control. By contrast, viral loads in the livers of mice deficient in GzmB, either alone, or in combination with GzmA, were significantly higher than those observed in B6 mice at day 6 post-infection (Fig. 6B). Thus, GzmB is essential for effective control of MCMV Δm157 in the liver. Furthermore, replication of the Δm157 virus in the livers of GzmBW/W mice was equivalent to that of GzmB−/− and GzmA/B−/− mice, indicating that GzmBW/W cannot substitute for the B6 allele during the anti-viral response to MCMV.
The effect of the Δm157 virus on GzmBW/W mice is reminiscent of the effects of MCMV infection on pfp-deficient mice. Mice lacking pfp exhibit increased mortality after MCMV infection [30]. While MCMV replication in the liver of pfp−/− mice is significantly higher than that observed in B6 mice, uncontrolled viral replication is not the cause of mortality [4]. Rather, a fatal hemophagocytic lymphohistiocytsis (HLH)-like syndrome develops in pfp−/− mice due to the uncontrolled production of TNFα by accumulating activated macrophages [4]. To determine if GzmBW/W mice exhibited any signs of an HLH-like syndrome, we examined lymphocyte populations and cytokine production after MCMV infection. Following infection with the Δm157 MCMV mutant, the livers of GzmBW/W mice contained significantly more leukocytes at day 6 post-infection (Fig. 7A). Analysis of these leukocytes by FACS revealed that the number of inflammatory monocytes (CD11b+ Ly6C+ Ly6G−) in the liver of GzmBW/W mice were significantly increased (Fig. 7A). While a similar trend was observed for granulocytes (CD11b+ Ly6G+), this did not reach statistical significance (Fig. 7A). Given the increased numbers of inflammatory monocytes in the livers of GzmBW/W mice, we measured pro-inflammatory cytokine production to determine if this may be contributing to the observed mortality. The levels of TNFα and IFNγ in the livers of GzmBW/W mice were not significantly different from those observed in B6 mice (Fig. 7B). These findings indicate that GzmBW/W mice do not develop an HLH-like syndrome following MCMV infection.
In order to better characterize the effects of MCMV infection in GzmBW/W mice, immunohistochemistry (IHC) staining of liver sections was performed. IHC staining of liver sections with an antibody specific for the IE1 protein of MCMV revealed a stark difference between B6 and the GzmBW/W mice. In B6 mice inflammatory foci were evident at day 6 post-infection, consisting of small numbers of infected hepatocytes (brown stain), typically surrounded by a large number of lymphocytes (Fig. 7C). In GzmBW/W mice, the inflammatory foci were larger in size and contained significantly more infected hepatocytes (Fig. 7C). Furthermore, areas of necrosis and cell debris within the centre of the foci were apparent (Fig. 7C). Together the data indicate that the liver damage observed in GzmBW/W mice was the direct result of uncontrolled viral replication, rather than the outcome of immune-mediated pathology.
The inability of GzmBW/W mice to control MCMV in the liver suggests that these mice may not be generating an appropriate anti-viral CTL response. Serpinb9 is a potent inhibitor of GzmB that is expressed by CTL [24]. Expression of Serpinb9 is required to prevent the premature apoptosis of CTL generated in response to lymphocytic choriomenigitis virus (LCMV) or Listeria monoctogenes infection [31]. We found that the GzmBW protein effectively bound Serpinb9 in an in vitro assay (Fig. 1C), but this finding does not preclude the possibility that Serpinb9 is unable to efficiently inhibit GzmBW in CTL in vivo. We therefore infected Serpinb9−/− mice with Δm157 MCMV and quantified viral replication. Viral titers in the spleen, liver, and lungs of Serpinb9−/− mice were not significantly different from those of B6 mice (Figure S3). Thus, the effects observed in GzmBW/W mice are not the result of an inability of Serpinb9 to inhibit GzmBW.
Next, we assessed the generation and effectiveness of anti-viral T cell responses in GzmBW/W mice. The total numbers of CD8 and CD4 T cells localizing to the livers of GzmBW/W mice after MCMV infection were not significantly different from those observed in MCMV-infected B6 mice (Fig. 8A). A peptide derived from the M45 protein of MCMV is the immunodominant epitope recognized by CD8 T cells in B6 mice [32]. The percentage of CD8 T cells stained by an M45 tetramer in B6 mice following MCMV infection was similar to that observed in GzmBW/W mice (Fig. 8B), and there were no differences in the total number of M45-specific CD8 T cells generated (Fig. 8C). The capacity of M45-specific T cells to kill target cells was also assessed. Splenocytes isolated from B6 mice at day 6 post-infection efficiently lysed M45 pulsed target cells, while no significant lysis was apparent when splenocytes from uninfected mice were used (Fig. 8D). The capacity of GzmBW/W splenocytes to lyse M45 pulsed target cells was similar to that of B6 cells (Fig. 8D). Hence, GzmBW/W mice generate an effective CD8 T cell response following infection with Δm157 MCMV and these T cells are able to efficiently kill model target cells. In addition to peptide pulsed targets, we tested the capacity of GzmBW/W and GzmBP/P CD8 T cells to kill MCMV-infected cells. CD8 T cells were purified from the spleen of B6 mice or GzmBW/W mice 6 days after infection, co-cultured with MCMV-infected IC-21 macrophages and macrophage viability assessed 18 h later. GzmBP/P CD8 T efficiently lysed the MCMV-infected target cells in a dose dependent manner, whereas GzmBW/W CD8 T cells were almost completely ineffective (Fig. 8E). Collectively, the data suggested that CD8 T cells expressing the w allele of GzmB are elicited and activated in response to MCMV infection, but are unable to kill MCMV-infected targets, accounting of the susceptibility of these mice to the virus.
Murine GzmB is highly polymorphic in wild mice, but the physiological relevance of this finding has been unclear. Here, we have characterized the biochemical and physiological function of an allelic variant of GzmB, GzmBW, found in wild mice. We have found that GzmB polymorphism affects the substrates cleaved by the protease in vitro. GzmBW efficiently cleaved peptide substrates based on the Bid sequence, and Bid itself, and as such, has an activity that is distinct from the GzmBP allele expressed by inbred mouse strains. Despite having different substrate preferences, GzmBP and GzmBW induced apoptosis of uninfected target cells with similar efficiency in vitro, and GzmBW/W mice were as efficient as wild type B6 mice in controlling ECTV infection. These data clearly show that the GzmBW isoform is bioactive in vitro and, more importantly, in vivo, at least as far as the requirements for its critical role in inducing the death of “generic” targets cells and in controlling infection with ECTV.
A striking finding of the study is that B6 mice expressing GzmBW are extremely sensitive to MCMV infection demonstrating that polymorphism in GzmB can have a significant impact on the capacity of the host to control specific pathogens. We found that GzmB was essential for control of Δμ157 MCMV (the most common MCMV variant found in the wild), and that the GzmBW allele was unable to substitute for GzmBP in this setting. These results argue for a defect in cytotoxicity, however, several non-cytotoxic roles have been ascribed to various Gzms. For example, GzmA and GzmM have a role in the production/release of pro-inflammatory cytokines [33]–[35]. We found that production of the pro-inflammatory cytokines IFNγ and TNFα by GzmBW/W mice following MCMV infection was equivalent to that of B6 mice. Furthermore, GzmBW/W mice generated CD8 T cell effectors at the expected frequency with localisation of these cells to the liver equivalent to that observed in B6 mice. Thus, the inability of GzmBW/W mice to control MCMV infection was not the result of defects in the production of pro-inflammatory cytokines, nor was it due to defective CD8 T cell numbers.
The initiation of apoptosis by activating caspases is the best-characterized function of GzmB. Human GzmB initiates apoptosis by activating caspases via two distinct pathways [36]. A mitochondrial-dependent pathway is activated when human GzmB cleaves Bid resulting in MOMP and the release of pro-apoptotic mediators [13], [37]. Mitochondrial-independent pathways operate via direct pro-caspase activation when sufficiently high concentrations of GzmB are delivered to the target cell cytosol [36]. Given the substrate specificity of the w allele, apoptosis induced by this form of GzmB is likely to mirror that of human GzmB. In vitro, GzmBW was as effective as GzmBP at inducing apoptosis in uninfected target cells, and anti-viral specific CD8 T cells isolated from GzmBW/W mice killed peptide pulsed target cells efficiently, indicating that there is no intrinsic defect in the direct cytotoxic capacity of GzmBW. However, CD8 T cells isolated from GzmBW/W mice were unable to kill MCMV-infected target cells in vitro. MCMV encodes potent inhibitors of both Bax and Bak that together prevent MOMP in response to multiple stimuli [38]–[41]. Collectively, the data strongly suggest that as GzmBW preferentially cleaves Bid, rather than directly activating caspases, it is susceptible to inhibition by MCMV-encoded proteins that block the intrinsic cell death pathway.
Our previous work has studied some of the key residues that dictate the species-specific substrate preferences of human and mouse GzmB. We found that two residues whose side-chains impinge on the substrate cleft, 180Arg and 222Lys are important in this regard, as substitution of the corresponding human or rat residue conferred a greater capacity to cleave Bid [10]. Residue 222 was invariant in all of the mouse GzmB alleles we previously sequenced and is thus common to both the mouse w and p alleles. However at position 180, the w allele encodes His, rather than Arg (present in the p allele), or Tyr which is found in human GzmB. His is also present at the same position in M. casteneus and M. spretus, mouse subspecies commonly found in certain parts of Asia. Given that 222Lys is invariant, it is extremely likely that the altered substrate preferences that result in the w allele having a far greater ability to cleave mouse Bid than to activate pro-caspases directly must rely on 180His. We found that in contrast to GzmBP, we found that GzmBW is unable to effectively activate some pro-caspases directly. Thus, apoptosis induced by GzmBW is reliant on activating the intrinsic pathway. Overall, a combination of changes in GzmB's fine substrate specificity, together with the expression of MCMV-encoded inhibitors of MOMP accounts for the failure of GzmBW expressing CD8 T cells to kill virally infected cells, resulting in uncontrolled viral replication.
In summary, we have demonstrated that a GzmB polymorphism commonly found in the wild has a profound influence on the ability of mice to control a natural pathogen. Importantly, the devastating effects elicited in hosts carrying the GzmBW allele by what is a common viral infection suggest that this allele has been maintained in the population because it confers a survival advantage in a setting yet to be defined. Furthermore, the function of GzmB in the response to pathogens and tumours has been investigated almost exclusively using inbred mouse strains all of which express the same allele of GzmB. The results of this study suggest that the use of mouse strains expressing alternative alleles of GzmB will be important for gaining a full understating of the role played by GzmB during an immune response. These findings also raise the possibility that alleles of GzmB identified in humans could impact on the control of some human pathogens.
This study was performed in accordance with the recommendations in the Australian code of practice for the care and use of animals for scientific purposes and the Australian National Health and Medical Research Council Guidelines and Policies on Animal Ethics. Experiments were approved by the Animal Ethics and Experimentation Committee of the University of Western Australia (Protocol # RA3/100/1094), the Animal Ethics and Experimentation Committee of the Peter MacCallum Cancer Centre (Protocol #E381) and the Animal Ethics and Experimentation Committee of the Australian National University (Protocol # A2012/041).
Recombinant granzymes were produced as artificial zymogens in Pichia pastoris, activated using enterokinase following purification, and assessed for the ability to cleave the synthetic peptide thiobenzylester (Boc-Ala-Ala-Asp (AAD)-SBzl [10], [21]. A GzmBw cDNA with optimized mouse codon usage was synthesized in vitro (GenScript). Specific activity of purified mouse granzymes was assessed by SDS-stable binding to an enhanced form of Serpinb9 (Cys339Asp) produced as described [42]. Preparations were routinely >95% active.
35S-labeled mouse procaspase 3 or mouse Bid was produced from cDNAs in the expression vector pSVTf via in vitro transcription and translation [19].
Recombinant granzymes were used to probe a P1 Asp-anchored library, as described [43].
A wild mouse colony (B1–6) maintained at the Animal Resource Centre (Canning Vale, WA), which expressed the outbred GzmB allele (GzmBw/w) [20] was crossed with C57BL/6 mice. Mice homozygous for GzmBw/w were backcrossed to C57BL/6 for 22 generations. These F22 mice were also crossed with B6.OT1 (ovalbumin-specific, H-2b -restricted T-cell receptor transgenic) mice to generate a GzmBw/w.OT1 line.
Inbred C57BL/6J (B6) mice were obtained from the Animal Resources Centre (Perth, Western Australia, Australia), orWalter and Eliza Hall Institute (Melbourne). B6.granzymeA−/− (GzmA−/−), B6.granzymeB−/− (GzmB−/−), B6.granzymeAB−/− (GzmAB−/−), B6.perforin−/− (Pfp−/−) B6.OT1, B6.Gzm.AB−/−.OT1, and B6.pfp.OT1 mice were bred and maintained at the Peter MacCallum Cancer Centre, Melbourne. B6 mice carrying the Serpinb9tm1.1/Pib allele (Serpinb9−/− mice) were generated and maintained at Monash University [44]. Mice were used at 6–10 weeks of age.
OVA257 specific activated T cells were generated from the spleens of the various OT1 mice (B6, B6.Pfp−/−, B6.GzmAB−/−, B6.GzmBw/w) as previously described [45]. The cytotoxic activity of the CTL was verified in 51Cr release assays using H-2b-peptide pulsed target cells (EL-4), as previously described [46].
Whole cell lysates were prepared from CTL cultures and normalized for protein content. Granule enzyme activity was determined as previously described [47] using synthetic peptide thiobenzylester (Boc-Ala-Ala-Asp (AAD)-SBzl and Suc-Phe-Leu-Phe SBzl) and paranitroanilide (Acetyl-Ile-Glu-Pro-Asp-paranitroanilide, Ac-IEPD-pNA) substrates (SM Biochemicals, CA, USA).
Ten µg of whole cell lysate was separated on NuPAGE 4–12% Bis Tris gels (Life Technologies, CA, USA), transferred and probed for mouse GrB protein (rat anti-mouse GrB, clone 16G6, eBioscience, CA, USA), as previously described [12]. Equal protein loading was confirmed by re-probing the blot with an anti-mouse β-actin antibody (Sigma-Aldrich, USA).
MCMV viral titers in organs were determined by plaque assay using M210B4 cells as previously described [49].
ETCV genome copies in blood were measured by quantitative real time PCR (qRT-PCR) to amplify the target sequence of ECTV-Mos-156 gene, as described elsewhere [50]. Oligonucleotide primers used were, forward: CGCTACACCTTATCCTCAGACAC, and reverse: GGAATTGGGCTCCTTATACCA. Viral DNA was prepared using QiaAmp DNA Mini Kit (Qiagen Pty Ltd, Victoria, Australia) as per manufacturer's instructions. Serial dilutions of a plasmid encoding ECTV-Mos-156 were used as the standard. The qRT-PCR reaction was carried out in SYBR iQ Supermix (Bio-Rad Laboratories), in a total volume of 20 µl using the iQ5 cycler (Bio-Rad Laboratories Pty Ltd, New South Wales, Australia).
Single-cell suspensions were prepared by perfusing the liver via the portal vein with phosphate buffered saline (PBS). The liver was then digested with collagenase buffer (RPMI/2% FCS, 1% (w/v) collagenase IV (GIBCO) for 20 min before being passing through steel mesh. The resulting preparation was resuspended in in a 37.5% isotonic Percoll solution (Pharmacia) and centrifuged at 690 g for 12 min to separate lymphocytes from hepatocytes. Red blood cells were osmotically lysed using NH4Cl and cells washed in FACS buffer. Antibodies used Antibodies used for analysis (TCRβ, CD8, CD4, CD11b, CD11c, Ly6C, Ly6G) were purchased from BD Biosciences and the M45 tetramer was obtained from ImmunoID Tetramers, University of Melbourne, Australia. Dead cells were excluded from analysis using propidium iodide.
IFNγ and TNFα levels in the liver were measured by standard sandwich enzyme-linked immunosorbent assay (ELISA) with antibodies from BD Biosciences. Detection was achieved with poly-horseradish peroxidase (poly-HRP) conjugated to streptavidin (CBL, Amsterdam, Netherlands) and K-Blue (Elisa Systems, Brisbane, Australia).
Organs were removed from mice at the designated times and fixed in 10% buffered formal saline. Organs were then embedded in paraffin, tissue sections prepared and sections stained with haematoxylin and counter stained with eosin. IE1 protein was detected by staining slides with an anti-IE1 monoclonal antibody (Clone Chroma 101), and detected with goat anti mouse HRPO and metal enhanced DBA substrate (Thermo Fisher).
Cell death induced by perforin (0.135–1.3 nM) and the recombinant mouse inbred and the outbred wild GzmB (12.5–25 nM) was assessed by 51Cr release assays as previously described [51].
Cytotoxic activity of M45 specific T cells was assessed by preparing a single-cell suspension from the spleens. Splenocytes were diluted 2-fold on 96-well plates starting with 1×106 cells/well and 51Cr-labeled EL4 cells pulsed with M45 peptide (1×104 cells/well) were added. Each assay was performed in triplicate. Chromium release was measure after 4 h incubation. Data are presented as percentage of specific lysis, calculated by the following formula: percentage specific lysis = (experimental c.p.m.−spontaneous release c.p.m.)/(total c.p.m.−spontaneous release c.p.m.)×100.
The capacity of activated CD8 T cells to kill MCMV infected cells was assessed using IC-21 macrophages. B6 or GzmBW/W mice were infected with Δm157 virus and spleens removed at day 6 post-infection. CD8 T cells were purified from the spleen using a CD8a positive selection kit (Stem Cell Technologies) according to the manufacturers instructions. IC-21 macrophages were infected with MCMV 12 h prior to co-culture with the purified CD8 T cells. Cell viability was quantified after 16 h of co-culture by MTT assay.
In vitro assays were assessed using a one way ANOVA with Tukey's multiple comparison test. Statistical analysis of survival curves were performed using the Log Rank test. Differences in viral replication within organs were assessed using a two-tailed Mann-Whitney test. Statistical tests were performed using the statistical software package InStat (GraphPad Software, San Diego California USA).
|
10.1371/journal.pgen.1006919 | Adaptive introgression from distant Caribbean islands contributed to the diversification of a microendemic adaptive radiation of trophic specialist pupfishes | Rapid diversification often involves complex histories of gene flow that leave variable and conflicting signatures of evolutionary relatedness across the genome. Identifying the extent and source of variation in these evolutionary relationships can provide insight into the evolutionary mechanisms involved in rapid radiations. Here we compare the discordant evolutionary relationships associated with species phenotypes across 42 whole genomes from a sympatric adaptive radiation of Cyprinodon pupfishes endemic to San Salvador Island, Bahamas and several outgroup pupfish species in order to understand the rarity of these trophic specialists within the larger radiation of Cyprinodon. 82% of the genome depicts close evolutionary relationships among the San Salvador Island species reflecting their geographic proximity, but the vast majority of variants fixed between specialist species lie in regions with discordant topologies. Top candidate adaptive introgression regions include signatures of selective sweeps and adaptive introgression of genetic variation from a single population in the northwestern Bahamas into each of the specialist species. Hard selective sweeps of genetic variation on San Salvador Island contributed 5 times more to speciation of trophic specialists than adaptive introgression of Caribbean genetic variation; however, four of the 11 introgressed regions came from a single distant island and were associated with the primary axis of oral jaw divergence within the radiation. For example, standing variation in a proto-oncogene (ski) known to have effects on jaw size introgressed into one San Salvador Island specialist from an island 300 km away approximately 10 kya. The complex emerging picture of the origins of adaptive radiation on San Salvador Island indicates that multiple sources of genetic variation contributed to the adaptive phenotypes of novel trophic specialists on the island. Our findings suggest that a suite of factors, including rare adaptive introgression, may be necessary for adaptive radiation in addition to ecological opportunity.
| Groups of closely related species can rapidly evolve to occupy diverse ecological roles, but the ecological and genetic conditions that trigger this diversification are still highly debated. We examined patterns of molecular evolution across the genomes of a recent radiation of pupfishes that includes two trophic specialists. Despite apparently widespread ecological opportunities and gene flow across the Caribbean, this radiation is endemic to a single Bahamian island. Using the whole genomes of 42 pupfish we found evidence of extensive and previously unexpected variation in evolutionary relatedness among Caribbean pupfish. While adaptive introgression appears to be rare across the genomes of the San Salvador Island species, it may have introduced adaptive variants important in the evolution of the complex phenotypes of the specialists. Four of the 11 candidate adaptive introgression regions contain genes with known effects on jaw morphology in zebrafish or associated with pupfish jaw size, the primary axis of phenotypic divergence between species in this system. Our findings that multiple sources of genetic variation contribute to the San Salvador Island radiation suggests that a complex suite of factors, including hybridization with other species, may be necessary for adaptive radiation in addition to ecological opportunity.
| Adaptive radiations are central to our understanding of evolution because they generate a wealth of ecological, phenotypic, and species diversity in rapid bursts. However, the mechanisms that trigger rapid bursts of trait divergence, niche evolution, and diversification characteristic of classic adaptive radiations are still debated. The availability of resources in new environments with few competitors has long been seen as the major force driving adaptive radiations [1–3], but it is a longstanding question why only some lineages rapidly diversify in response to such ecological opportunities while others do not [4–9].
While gene flow can impede or reverse diversification among incipient species by reducing genetic differentiation and subsequent recombination can break down locally adapted haplotypes [10–13], it can also introduce adaptive genetic variants [14,15] and/or genetic incompatibilities [16–18] that initiate or contribute to the process of speciation. A growing number of studies have identified gene flow and genome-wide introgression across a range of adaptive radiations [19–26], contributing to the emerging view that gene flow is pervasive throughout the history of many young rapidly diversifying groups and may be necessary for adaptive radiation. Examples of adaptive radiations with histories of extensive hybridization include Heliconius butterflies [27–29], Darwin’s finches [21,30–32], Anopheles mosquitos [20,33], and cichlids [24,25,34–39]. The hybrid swarm hypothesis [40] proposes that hybridization among distinct lineages can introduce genetic diversity and novel allele combinations genome-wide that may trigger rapid diversification in the presence of abundant ecological opportunity. However, it is still unclear how often hybridization is necessary for rapid diversification, as opposed to simply being pervasive throughout the history of any young rapidly diversifying group [25,41]. One of the only examples with strong evidence of hybridization leading to ecological and species diversification is that of several hybrid species within a radiation of Helianthus sunflowers [42–47]. However, these may simply represent examples of multiple homoploid speciation events within an already radiating lineage rather than a hybrid swarm scenario. So while there is convincing evidence that hybridization can facilitate diversification among species pairs (but see [26,38] for a potential multispecies outcome of hybridization), it is still unclear whether gene flow is a major factor constraining adaptive radiation in some lineages or if ecological opportunity is the sole constraint.
The adaptive radiation of San Salvador Island pupfishes provides an outstanding system to compare the contributions of different sources of genetic variation to rapid diversification and the role of gene flow in the evolution of complex phenotypes. Pupfish species of the genus Cyprinodon inhabit saline lakes and coastal areas across the Caribbean and Atlantic and nearly all pupfishes are allopatric, dietary generalists consuming algae and small invertebrates [48]. In contrast, three Cyprinodon species live sympatrically in the hypersaline lakes of San Salvador Island and comprise a small radiation that has occurred within the past 10,000 years based on the most recent glacial maximum when these lakes were dry due to lowered sea levels [49–51]. This radiation is composed of the widespread generalist algae-eating species Cyprinodon variegatus and two endemic specialists that coexist with the generalist in all habitats in some lakes. These specialists have adapted to unique trophic niches using novel morphologies: the molluscivore Cyprinodon brontotheroides with a unique nasal protrusion and the scale-eater Cyprinodon desquamator with enlarged oral jaws and adductor mandibulae muscles [48,52]. Surveys of populations living on neighboring islands in the Bahamas and phylogenetic analyses with other Cyprinodon species indicate that these specialist species are endemic to the hypersaline lakes of San Salvador Island and that both specialists arose from a generalist common ancestor during this recent radiation [53].
The currently available ecological and genetic data on the group provides little indication as to why this radiation is localized to a single island. Variation in ecological opportunity among hypersaline lake environments in the Caribbean does not appear to explain the rarity of this radiation [53]. This finding suggests a potentially important role for sufficient genetic variation to respond to abundant, underutilized resources in these environments. However, a hybrid swarm hypothesis about the origins of the radiation does not appear to explain its rarity either: genetic diversity is comparable among islands and gene flow occurs among all Caribbean islands investigated, not only into San Salvador Island [53]. Novel traits and increased rates of diversification associated with them are well documented in this system [48,53,54], but understanding the rarity of this adaptive radiation requires a thorough investigation of the underlying genetic variation that accompanies these rare ecological transitions. A recent study investigating the genetic basis of trophic specialists in this radiation revealed very few regions underlying these phenotypes [55]. Only thousands of variants out of 12 million were fixed between the scale-eater and molluscivore species. Since genetic divergence is limited to particular regions, localized rather than genome-wide investigations of the genome will be important for understanding how genetic variation, possibly originating outside of San Salvador Island, has contributed to the exceptional phenotypic diversification restricted to this island. Here, we use a machine-learning approach to identify regions of the genome with different evolutionary relationships among 42 pupfish genomes sampled from the San Salvador Island radiation, two distant Caribbean islands, and 3 additional outgroups. We then scan the genome for evidence of localized introgression with pupfish populations outside of San Salvador Island and compare the relative contributions of adaptive introgression from two distant islands and hard selective sweeps to the divergence of each specialist species.
To identify localized patterns of population history across the genome, we used the machine-learning approach SAGUARO. SAGUARO combines a hidden Markov model with a self-organizing map to characterize local topologies across the genome among aligned individuals [56]. This method does not require any a priori hypotheses about the relationships among individuals, but rather infers them directly from the genome by finding regions of consecutive nucleotides with a similar pattern of genetic differentiation, building hypotheses about relationships among individuals from these genetic differences, and then assigning regions of the genome to these hypothesized local topologies. Since smaller segments with fewer informative SNPs are more likely to be incorrectly assigned to a hypothesized topology by chance (pers. comm. M.G. Grabherr), we tested various minimum SNP filters for reducing the amount of short, uninformative segments assigned to topologies by chance and found that increasingly stringent filters over 20 SNPs did not substantially reduce the number of uninformative segments. Using this approach and our 20 SNP filter, we partitioned the genome into a total of 15 unique topologies across 227,248 genomic segments that ranged from 101–324,088 base pairs in length (median: 852 bp) (S1 and S2 Figs; S1 Table). The 15th topology was uninformative about either species or population level relationships, so it was removed from downstream analyses.
The most prevalent history across 64% of the genome featured the expected species phylogeny for this group from previous genome-wide studies [48,53,57], in which all individuals from San Salvador Island grouped by species into a single clade with distant relationships to outgroup generalist pupfish populations from other islands in the Caribbean, Death Valley in California, and a second radiation in Mexico spanning the most divergent branch of the Cyprinodon tree (Fig 1). Unlike previous genome-wide phylogenies [53,57], and with the exception of a few individuals that grouped with molluscivores by lake, the generalists on San Salvador Island form a discrete clade from the molluscivores and scale-eaters.
Within this dominant topology, scale-eaters from six lakes on San Salvador Island fell into one of two separate clades: small-jawed individuals from Osprey Lake, Great Lake, and Oyster Pond and large-jawed individuals from Crescent Pond, Stout Lake, Osprey Lake, and Little Lake (Fig 1). Molluscivores did not form a single clade as individuals from some lakes (Crescent Pond and Moon Rock) were more closely related to generalists from the same lake than molluscivores from other lakes, similar to previous genome-wide phylogenies [57]. Another topology covering 10% of the genome was very similar to the dominant one, differing only in the relationships among San Salvador Island generalists (S1 Fig). Additional topologies spanning 7.6% of the genome featured a single San Salvador Island clade but also depicted a closer relationship between San Salvador Island and the outgroups as well as groupings of all three San Salvador Island species by lake in Crescent Pond and Moon Rock Pond. When combined with the dominant topology, only 82.6% of the genome supported the expected San Salvador Island clade (S1 Table).
In other regions of the genome, San Salvador Island did not form a single clade (Fig 2A–2C and S2 Fig, S1 Table). The most frequently observed alternative relationships depicted specialist individuals as a clade outside of the San Salvador Island group and sister to all the outgroup Cyprinodon species (Fig 2A and 2B). The ‘large-jawed scale-eater topology’ featured large-jawed scale-eaters outside of the San Salvador Island clade, sister to all other outgroups, and was assigned to 4,437 segments covering 3.77% of the genome (Fig 2A). Another topology, the ‘molluscivore topology’, showed a similar pattern in which the molluscivores formed a single clade outside of the San Salvador Island group and sister to all other outgroups (Fig 2B). This molluscivore topology was assigned to 3,916 segments and covered 3.12% of the genome. Another 2,029 segments covering 1.66% of the genome were assigned to a topology where both the large-jawed and small-jawed scale-eaters formed a combined clade outside of the San Salvador Island group, the ‘combined scale-eater topology’ (Fig 2C). Other topologies featuring one of the specialists separated from the rest of San Salvador Island covered 0.76%-2.48% of the genome (S1 Table).
Unexpectedly, all 14 informative topologies separated scale-eaters into groups of small- and large-jawed individuals and the relationships between these two groups and other species differed across different regions of the genome. In some regions, the small-jawed scale-eater individuals were sister to the large-jawed scale-eaters (Figs 1, 2B and 2C, S1 and S2 Figs). In other regions, the small-jawed scale-eaters were more closely related to the generalists and molluscivores (Fig 2A, S1 and S2 Figs). These small-jawed scale-eaters may be a product of ongoing hybridization between species on San Salvador Island or a new ‘occasional’ scale-eating ecomorph, perhaps representing an intermediate yet viable stage on the evolutionary path towards large-jawed scale-eaters, in which scales form the majority of their diet [54]. The presence of homozygous genotypes in all five individuals of small-jawed scale-eaters for variants fixed in both large-jawed scale-eaters and generalists is not consistent with first generation hybrids (S2 Table). They also do not fit the ancestry proportions expected in F2 hybrids (χ2 = 429.6, P = 5.16e-94). We might expect increased linkage disequilibrium (LD) in the small-jawed scale-eaters if they represent recent hybridization events between distinct populations. Consistent with this idea, LD decays more slowly in the small-jawed scale-eaters (after approximately 120 kb) than in the three San Salvador Island species (after approximately 50kb: S3 Fig). However, strong LD and long haplotype blocks may also result from other evolutionary phenomena like recent population bottlenecks (e.g. [58]). Demographic modeling with a larger sample will be needed to distinguish whether these small-jawed scale-eaters represent hybrids from ongoing or recent gene flow on San Salvador Island or a potential new ecomorph.
We examined signals of introgression from two distant pupfish generalist populations in the Caribbean: Lake Cunningham, New Providence Island in the Bahamas (described as the endemic species Cyprinodon laciniatus [59]) and Etang Saumatre / Lac Azuei in the Dominican Republic (described as the endemic species Cyprinodon bondi [60]). C. laciniatus exhibits morphological variation not observed in other generalist species, including laciniated scales and variation in oral jaw size [59], although not the extreme oral jaw morphologies observed in the specialists, and is an interesting candidate for looking at adaptive introgression of variants involved in oral jaw size morphology on San Salvador Island. C. bondi is a generalist species of the variegatus complex from the south-eastern end of the range of Greater Antillean pupfish and introgression with San Salvador Island populations would suggest that Caribbean-wide gene flow may have contributed to the adaptive radiation on San Salvador Island. We characterized the genomic landscape of introgression in the three San Salvador Island species using f4 statistics that were initially developed to test for introgression among human populations [61–63].
Genome-wide f4 tests provided evidence of introgression between Caribbean outgroups and San Salvador Island. f4 values significantly deviated from the null hypothesis of no introgression (f4 = 0) in the scale-eater/molluscivore (Z = 4.2, P = 2.67x10-5), and scale-eater/generalist combinations (Z = 4.67, P = 3.01x10-6), but were not significant in the molluscivore/generalist combination (Z = -1.63, P = 0.103).
When f4 was calculated in windows, we found that 181 10-kb regions out of 100,260 (0.18%) contained significant evidence of introgression between C. laciniatus or C. bondi and the San Salvador Island specialists (Fig 3A). Introgressed regions were scattered across the genome in 107 of the 9,259 scaffolds in our dataset. These regions were not typically concentrated in one section of the genome, with the largest cluster within a single scaffold containing 12% of the total (Fig 3A).
The genomic regions with significant evidence of introgression varied between the two specialists (Fig 3B and 3C): only 15 regions from the 176 and 112 regions with significant evidence of introgression were shared between generalist/scale-eater and generalist/molluscivore comparisons, respectively. This suggests that admixture with other Caribbean populations occurred multiple times and independently for each specialist or that different introgressed regions were used by the two specialists after a single admixture event (see S4–S6 Figs for full Manhattan plots).
We also tested for introgression with the small-jawed scale-eaters excluded to search for potential introgression with the large-jawed scale-eaters alone (S7 Fig). Introgressed regions were less variable between the two groups of scale-eaters, with 122 of 209 candidate introgressed regions shared. The 87 introgressed regions unique to the large-jawed scale-eaters suggest that some introgression may have occurred between populations on other Caribbean islands and the large-jawed scale-eater population independently from the small-jawed scale-eaters.
Regions of low diversity and low recombination may be biased when genome-wide tests of introgression, such as the f4 statistic, are applied to genomic windows [64]. To assess whether our introgressed regions were the result of this bias, we looked at π estimates across the detected regions of introgression in comparison to the genome-wide estimates (mean Dxy = 0.007; mean π scale-eater = 0.0048; mean π molluscivore = 0.0054) and variance in f4 statistic values. f4 statistics do appear slightly sensitive to the level of diversity in a region, with variance in f4 values having a weak negative correlation with mean scaffold π (Pearson’s r = -0.18; S8 Fig), and a weaker correlation between the value of f4 and π (Pearson’s r = -0.013; S9 Fig). However, in selecting our top candidate introgressed regions, we assessed π in all three San Salvador Island species and looked for other signals of introgression to complement the f4 test. This included pairwise estimates of Dxy between each San Salvador Island species and outgroups, TREEMIX analyses used to infer admixture events on population graphs [63], presence of alternative topologies in the regions, and maximum likelihood trees supporting close relationships between outgroups and either of the specialists. f4 outliers that appeared in extensive regions of low diversity in all three San Salvador Island species and did not have supporting evidence from other statistics or trees were excluded from the list of candidates as potential false positives in areas of low recombination (n = 2; S10 and S11 Figs).
The relationships observed in the three alternative topologies (Fig 2) underlie most of the divergence observed between the molluscivores and scale-eaters: 75% and 88% of the fixed SNPs between molluscivores and large-jawed scale-eaters and molluscivores and all scale-eaters, respectively, fall within these topologies that make up less than 5% of the genome in total. Many of these regions contained candidate genes previously associated with variation in Cyprinodon jaw size within the San Salvador Island radiation [55]: 18 of the 31 candidate jaw genes occurred in the combined scale-eater topology and 1 candidate region in the molluscivore topology.
We also assessed the relative contributions of different sources of genetic variation to the divergence between the two specialists (also see S12 Fig). Selective sweeps of introgressed variation from our two focal outgroups contributed 5 and 8 times less to species divergence between the scale-eaters and molluscivores, respectively, than sweeps of other sources of genetic variation (Fig 4). Adaptive introgression in regions of high divergence among the specialists appears to be rare, occurring in only 0.006 and 0.016% of the scale-eater and molluscivore genomes, respectively. The higher percentage in the molluscivore genome may be due to stronger bottlenecks in their past than in the scale-eaters, rather than more selective sweeps in this species. Within individual lakes, molluscivores have lower genetic diversity than both scale-eaters and generalists [57]. When segments are additionally separated based on topology assigned by SAGUARO, the alternative topologies contained a greater proportion of regions with introgressed genetic variation and selective sweeps than those regions assigned to the dominant topology. None of the fixed SNPs in adaptive introgression candidates occurred in a segment assigned to the dominant topology (S12 Fig).
In general, selective sweeps of introgressed genetic variation that contributed to species divergence between the specialists were rare. However, four of the 11 candidate adaptive introgression regions contained genes with known craniofacial effects in model organisms or have been strongly associated with oral jaw size variation in the specialists [55], the primary axis of diversification in this system (Table 1 and S3 Table). Only one of these, the proto-oncogene ski, has both known craniofacial effects and was associated with jaw size variation in the specialists. Ski encodes for a corepressor protein involved in the SMAD-dependent transcription growth factor B pathway [65–67]. Mutations in ski cause marked reductions in skeletal muscle mass, depressed nasal bridges, and shortened, thick lower jaw bones in mice [68,69] and malformed craniofacial cartilage and shortened lower jaws in zebrafish [70]. These phenotypic changes are remarkably similar to the novel craniofacial morphologies in San Salvador Island molluscivore pupfishes, including increased nasal/maxillary protrusion, shortened lower jaw, and thicker dentary and articular bones [52].
The candidate adaptive introgression region spans the start of ski and contains three fixed SNPs, one in the 3’ untranslated region, one in the 3rd codon position of an exon, and one in an intron. This region contains a signature of high absolute genetic divergence between the two specialists and a selective sweep in the molluscivore (Fig 5). This region also features low nucleotide diversity within scale-eaters and negative estimates of Tajima’s D, although this does not appear to be as strong as in the molluscivores. Several lines of evidence point towards the introgression of ski variants between molluscivores and C. laciniatus. Genetic differentiation is minimal between molluscivores and C. laciniatus (Dxy = 0.0011) (Fig 5) and higher in all other pairwise comparisons (Dxy > 0.013) between the two specialists and two outgroup Caribbean pupfish species (S4 Table), indicating gene flow between the molluscivores on San Salvador Island and the generalist C. laciniatus on New Providence Island. Taking a closer look at the genetic variation in this region, we observe that the ski SNPs fixed in the San Salvador Island molluscivores are homozygous in C. laciniatus and segregating in the generalists (Fig 6A), suggesting that they occur at an appreciable frequency in the generalists. The surrounding molluscivore genetic background of the fixed ski SNPs is very similar to C. laciniatus (Fig 6B). In this 10-kb region, only 62 SNPs differ between the molluscivores and C. laciniatus in our sample. Segments of this region were assigned to the combined scale-eater topology (Fig 2C) and a maximum likelihood tree of the SAGUARO segment containing these three fixed SNPs features C. laciniatus in a clade with molluscivores (S13 Fig).
In addition to ski, one other adaptively introgressed candidate region with known craniofacial effects in fish lies in the RNA-binding protein rbms3, a posttranscriptional regulator in the same SMAD-dependent transcription growth factor B pathway. Mutations in this gene cause cartilage and neural crest related abnormalities in zebrafish [71]. This region contains a non-coding SNP fixed in the San Salvador Island scale-eaters that is homozygous in C. laciniatus and segregating in the generalist population, a signature of high absolute genetic divergence between the two specialists, and a selective sweep in the scale-eater (Fig 7). Several lines of evidence point towards the introgression of rbms3 variants between scale-eater and C. laciniatus. First, genetic differentiation is minimal between scale-eaters and C. laciniatus (Dxy = 0.002) and higher in all other pairwise comparisons (Dxy > 0.0104) between the two specialists and two outgroup Caribbean pupfish species (S4 Table). Segments of this region were assigned to the combined scale-eater topology (Fig 2C) and a maximum likelihood tree of the segment containing the fixed SNP features C. laciniatus in a clade with scale-eaters (S14 Fig). Similar to the pattern we find in rbms3, another candidate region previously associated with oral jaw size variation on San Salvador Island spanning pard3 contained fixed scale-eater variants shared with C. laciniatus, strong genetic similarity in the surrounding region between the two and signs of a selective sweep in the scale-eaters (S15 and S16 Figs, Table 1).
In an unannotated candidate adaptive introgression region which has previously been associated with oral jaw size variation on San Salvador Island, we find a slightly different pattern than those mentioned above. The direction of introgression appears to be between C. laciniatus and the molluscivores, but is under a selective sweep in the scale-eaters (S17 and S18 Figs, Table 1 and S3 Table). We also see a similar pattern in nbea, where the direction of introgression appears to be between C. laciniatus and scale-eaters but is under a selective sweep in the molluscivores (S19 and S20 Figs, Table 1 and S3 Table). Nbea encodes for a scaffolding protein involved in neurotransmitter release and synaptic functioning and has been identified as a candidate gene for non-syndromic autism disorder [72–74]. In zebrafish, mutations disrupt electrical and chemical synapse formation and cause behavioral abnormalities such as decreased startle response [75]. Introgression in this regions is of interest because behavior is another axis of divergence between specialists in this system alongside craniofacial traits, as the species vary in mate choice [76,77], aggression, and prey capture behavior [54]. Both of these candidate regions feature nearly equivalent negative Tajima’s D statistics and low nucleotide diversity in the both of the specialists. The regions do not appear to be under strong selection in the generalist populations on San Salvador Island, so the signatures of selective sweeps in both specialists most likely stem from parallel molecular evolution in these regions rather than purifying selection in the ancestral population. Seven of 11 candidate regions show this pattern of equivalent low diversity and negative Tajima’s D statistics in both specialists (Table 1 and S3 Table).
The other 6 adaptive introgression candidates contained genes with a variety of functions including angiogenesis, calcium ion binding, embryonic eye morphogenesis, and RNA binding (Table 1) and had similar patterns to those mentioned above. Four of these regions feature low genetic diversity in both specialists. Two of these candidates lie in consecutive regions of the gene srbd1, which encodes for an RNA binding protein, and it appears that one has introgressed between the molluscivores and C. laciniatus and the other between scale-eaters and C. laciniatus. Both of these regions appear to be under a selective sweep in both of the specialists (Table 1 and S3 Table).
Overall, potential adaptive variants contributing to species divergence among the specialists appear to be coming from New Providence Island in the northern Caribbean, rather than the southern Caribbean (Table 1 and S3 Table). Since it is impossible to infer the directionality of gene flow directly from f4 values, we used TREEMIX [63] to visualize gene flow in adaptively introgressed regions. Across the candidate adaptive introgression regions, we found evidence of an admixture event directly from C. laciniatus into the molluscivores in ski and ltbp2 and C. laciniatus into scale-eaters in srbd1 (Fig 8 and S5 Table). This suggests that genetic variation found on New Providence Island introgressed into the San Salvador Island radiation. There is no direct evidence from the TREEMIX population graphs of admixture from C. bondi into a specialist in the candidate regions (S5 Table), and Dxy between C. bondi and the specialists in pairwise comparisons is greater than those found between C. laciniatus and specialists across these regions (S3 Table). Both lines of evidence suggest that the high f4 values in these regions stem from gene flow between C. laciniatus and the specialists rather than C. bondi.
Our investigation of genetic variation reveals that multiple sources of genetic variation were important for the assembly of the complex phenotypes associated with the novel ecological transitions seen only on San Salvador Island, Bahamas. While species divergence appears to mostly come from selective sweeps of variation from San Salvador Island (Fig 4), rare adaptive introgression has also played a role in the radiation (Table 1; Figs 5 and 7, S15, S17 and S19 Figs). The adaptive introgression we found in this study has come from large admixture events into San Salvador Island from a generalist pupfish population on another Bahamian island approximately 300 km away. In contrast, we found no evidence of introgression from a generalist population 700 km away in the Dominican Republic in our top candidate regions (Table 1 and S5 Table), although it is impossible to rule out that candidate adaptive variants may also exist in this population at lower frequencies. Importantly, our limited sampling of one individual from each of two distant islands suggests that long-distance adaptive introgression is common and arises from abundant genetic variation found in only some parts of the Caribbean. An intriguing implication of these findings is that adaptive variants within the San Salvador Island radiation may have been partly assembled from the overlap of different pools of standing variation distributed across different parts of the Caribbean.
We found introgressed variants in four genes associated with the primary axis of jaw size variation within the radiation, as well as one in a gene with known behavioral effects in zebrafish. Both specialists appear to have candidate introgressed adaptive variants implicated in jaw morphology. Our best candidate for molluscivores was a region containing three fixed variants previously associated with jaw size variation on San Salvador Island in the proto-oncogene ski, which introgressed from C. laciniatus, another pupfish species on an island 300 km away (Figs 5, 6 and 8A, Table 1). The best candidate for scale-eaters was a region containing a single fixed variant in the gene rbms3 (Fig 7, Table 1), which is also present in C. laciniatus. Other candidate regions contained genes with functions in behavior, angiogenesis, calcium ion binding, embryonic eye morphogenesis, and RNA binding (Table 1).
We rarely know the source of candidate variants involved in diversification or the contributions of multiple sources of genetic variation to rapid diversification. Genomic investigations of other adaptive radiations have also inferred roles for multiple genetic sources contributing to rapid diversification. For example, in the apple maggot fly, ancient gene flow from Mexican populations introduced an inversion affecting key diapause traits that aided the sympatric host shift to apples in the United States [78]. Hybridization within Darwin’s finches also appears to play a role in the origin of new lineages through adaptive introgression of functional loci contributing to beak shape differences between species [21]. In a Mimulus species complex, introgression of a locus affecting flower color appears to have been a driver of adaptation in the early stages of their diversification [79]. However, even in case studies demonstrating multiple sources of genetic variation, the relative contributions to the diverse ecological traits in these radiations still remain unknown in most cases (but see [80]).
Only 10% of all introgressed regions in either the molluscivore or scale-eater were shared between the two. This minimal overlap may reflect the complexity of different performance demands. Performance in the two specialists involves very different sets of functional traits (i.e. higher mechanical advantage and a novel nasal protrusion in the molluscivores vs. enlarged oral jaws and adductor muscles in the scale-eaters [54]) and divergent selective regimes (narrow and shallow vs. wide and deep fitness valleys [53,81,82]). The extensive variability in the genetic variation that introgressed between the two specialists may reflect multidimensional adaptation to two distinct trophic niches in this radiation, rather than variation along a linear axis (e.g. see [83–88]).
Although introgression is rare and localized across the genome, it was likely important for the assembly of the complex phenotypes observed on San Salvador Island (e.g. ski and rbms3). Our findings suggest two alternative possibilities. One intriguing possibility is that rare introgression of the necessary adaptive alleles into San Salvador Island may have been required to trigger the radiation in the presence of ecological opportunity. Indeed, a paradox in this system is why generalist populations in hypersaline lakes on neighboring islands with similar levels of ecological opportunity, lake areas, and overall genetic diversity have not radiated [53]. Alternatively, adaptive radiation on San Salvador Island may have initiated from standing and de novo variation and only later benefited from introgressed alleles to further refine species phenotypes. Of course, these scenarios are not mutually exclusive and may vary across loci. Based on our TREEMIX analysis, introgression from C. laciniatus into the molluscivores brought the ski variants (Fig 8A), but the candidate adaptive variants in this region are also segregating in the generalist population (Fig 6).
We can roughly estimate the timing of introgression for this ski region from the number of variants that have accumulated between the C. laciniatus and molluscivore haplotypes (n = 62 differences; Fig 6). Assuming neutrality, the observed genetic differences between the two lineages should equal 2μt, the time since their divergence in each lineage and μ, the mutation rate [89]. Using mutation rate estimates ranging from 5.37x10-7 (phylogeny-based estimate of Cyprinodon substitution rate [90]) to 1.32x10-7 mutations site-1 year-1 (estimated from a cichlid pedigree estimate of the per generation mutation rate [91] using a pupfish generation time of 6 months), introgression of the ski adaptive haplotype from C. laciniatus into the molluscivore specialist occurred between 5,700 to 23,500 years ago. The 10,000 year age estimate of the San Salvador Island radiation (based on estimates of dry lakes on the island [49–51]) falls within this window. This suggests the intriguing scenario in which widespread introgression during the last glacial maximum may have triggered adaptive radiation within pupfish populations isolated in the saline lakes of San Salvador Island during their initial formation. The 10-fold larger land mass of the Great Bahama Bank during this time could have created the opportunity for larger pupfish populations and greater genetic diversity. These pupfish populations would have been connected more extensively across the region than currently by the increased expanses of coastline habitats on the exposed bank. However, these are only exploratory inferences of the directionality of gene flow and timing of introgression. They should be confirmed with demographic analyses focused on testing different scenarios of admixture into San Salvador Island (e.g. [26,90,92–95]).
While there are rare and convincing examples of hybridization leading to homoploid speciation (reviewed in [47]), no study, including ours, has yet provided convincing evidence that hybridization was directly involved in triggering an adaptive radiation. For example, while there is strong evidence in Darwin’s finches that adaptive introgression of a loci controlling beak shape has contributed to phenotypic diversity of finches in the Galapagos, this hybridization occurred between members within the radiation [21]. Similarly, a recent study argued that hybridization between ancestral lineages of the Lake Victoria superflock cichlid radiations and distant riverine cichlid lineages fueled the radiations, based on evidence of equal admixture proportions across the genomes of the Victorian radiations from the riverine lineages and the presence of allelic variation in opsins in the riverine lineages which are also important in the Victoria radiation [38]. However, the timing of introgression and necessity of introgressed alleles for initiating adaptive radiations remains unclear in these systems, including our own. Admittedly, hybridization as the necessary and sufficient trigger of adaptive radiation is a difficult prediction to test.
Those examples with more direct evidence linking hybridization to adaptation and reproductive isolation within a radiation are often special cases where a single introgressed adaptive allele automatically results in increased reproductive isolation. Examples include introgressed adaptive loci controlling wing patterns in Heliconius butterflies involved in mimicry and mate selection [28,96], a locus controlling copper tolerance in Mimulus that is tightly associated with one causing hybrid lethality [16], and loci contributing to differing insecticide resistance in the M/S mosquito mating types [97–99]. While these cases provide convincing evidence that adaptive introgression can facilitate both ecological divergence and reproductive isolation, it is still unclear whether this introgression has actually triggered or simply contributed to the ongoing process of adaptive radiation.
Truly addressing the question of whether adaptive introgression triggered the radiation on San Salvador Island will require a better understanding of the timing of introgression and the necessity of introgressed variation for the speciation process. Although we have candidate alleles (e.g. in ski and rbms3) that we think play a role in the evolution of complex specialist phenotypes, it still remains unclear what minimal set of alleles is necessary for the major ecological transitions in this system. Knowledge of the age of variants important for these transitions, and whether these variants are present and adaptive in the other non-radiating lineages of Caribbean generalist populations is needed. Estimation of the age of introgressed variation relative to standing or de novo could also shed light on whether adaptive introgression simply contributed to an ongoing diversification process or triggered it on San Salvador Island.
We also found evidence of a distinct clade of small-jawed scale-eaters, separate from the large-jawed scale-eaters (Figs 1 and 2). The consistent clustering of this clade across the genome suggests that they may be a distinct, partially reproductively isolated population on San Salvador Island, rather than a product of hybridization between generalists and scale-eaters in the lakes where they exist sympatrically (Figs 1 and 2; S1 and S2 Figs). They have only been observed in six lakes connected to the Great Lake System on San Salvador Island (Great Lake, Mermaid’s Pond, Osprey Pond, Oyster Pond, Little Lake, and Stout’s Lake), but not in isolated lakes such as Crescent Pond. Consistent with this pattern of occurrence, F2 hybrid phenotypes resembling the scale-eaters have previously been shown to have extremely low survival and growth rates in these isolated lakes [81].
Small-jawed scale-eaters may represent a viable intermediate ecotype on the evolutionary path toward more specialized scale-eating. Small-jawed scale-eater diets appear to be consistent with intermediate levels of scale-eating. Preliminary gut content analyses revealed that scales were found in the stomachs of 33% of small-jawed scale-eaters (n = 33) compared to 91% of large-jawed scale-eaters (n = 53). The idea that specialization can open the door to further specialization has been seen in other systems, including pollinator syndromes for bees, hummingbirds, and hawkmoths in Mimulus [100–102], Darwin’s ground finch specializing on blood on two islands in their range [103], and transitions in mammals between omnivory, carnivory, and herbivory [104]. If small-jawed scale-eaters represent an ecotype stepping stone on the path toward more specialized scale-eating, we might expect regions of the genome to reflect a nested relationship between the large-jawed and small-jawed scale-eaters. We see this predicted pattern in the combined scale-eater topology that underlies most of the fixed variants between the two scale-eating species (Fig 2).
If the small-jawed scale-eaters were instead the result of recent or recurrent hybridization events, we would expect certain patterns of large-jawed scale-eater and generalist ancestry across their genomes. For example, if they represent F1 hybrids, they should have equal ancestry from the two parental species across their genomes. The lack of fit in all five small-jawed scale-eater individuals to the ancestry proportions excepted if they represent F1 hybrids, F2 hybrids, or backcrosses to parental species (S2 Table) suggests that the small-jawed scale-eaters are not the result of such recent hybridization events, although they might have resulted from more complicated scenarios of hybridization that do not follow these simple patterns of ancestry [105,106]. LD does appear to be stronger in the small-jawed scaler-eaters than in the three San Salvador Island species (S3 Fig), a pattern expected in recent hybrids of distinct populations. These small-jawed scale-eaters may indeed be the products of ongoing or recent gene flow on San Salvador Island. A reconstruction of the history of gene flow among San Salvador Island species from demographic modeling with a larger sample, along with estimates of selection and reproductive isolation in the small-jawed scale-eaters, will be needed to assess whether they represent the products of ongoing gene flow on San Salvador Island or a potential new ecomorph.
Here we demonstrate that the complex phenotypes associated with the novel ecological transitions within a nascent adaptive radiation of San Salvador Island pupfishes arose from multiple sources of genetic variation spread across the Caribbean. The variation important to this radiation is localized to small regions across the genome that are obscured by genome-wide summaries of the history of the radiation. Species divergence appears to mostly come from selective sweeps of standing or de novo genetic variation on San Salvador Island, but rare adaptive introgression events may also be necessary for the evolution of trophic specialists. This genomic landscape of introgression is variable between the specialists and has come from large admixture events from populations as far as 700 km across the Caribbean, although all top adaptive introgression candidates appear to have introgressed from a population 300 km away in the northwestern Bahamas. Our findings that multiple sources of genetic variation contribute to the San Salvador Island radiation suggests a complex suite of factors, including rare adaptive introgression, may be required to trigger adaptive radiation in the presence of ecological opportunity.
Individual pupfish were caught in hypersaline lakes on San Salvador Island in the Bahamas with either a hand or seine net in 2011, 2013, and 2015. Samples were collected from eight isolated lakes on this island (Crescent Pond, Great Lake, Little Lake, Mermaid Pond, Moon Rock Pond, Oyster Lake, Osprey Lake, and Stout’s Lake) and one estuary (Pigeon Creek). 13 Cyprinodon variegatus were sampled from all eight lakes on San Salvador Island; 10 C. brontotheroides were sampled from four lakes; and 14 C. desquamator were sampled from six lakes. The specialist species occur in sympatry with the generalists in only some of the lakes. Individual pupfish that were collected from other localities outside of San Salvador Island served as outgroups to the San Salvador Island radiation, including C. laciniatus from Lake Cunningham, New Providence Island in the Bahamas, C. bondi from Etang Saumatre lake in the Dominican Republic, C. diabolis from Devil’s Hole in California (collected as a dead specimen by National Park Staff in 2012), as well as captive-bred individuals of extinct-in-the-wild species C. simus and C. maya originating from Laguna Chichancanab, Quintana Roo, Mexico. Fish were euthanized by an overdose of buffered MS-222 (Finquel, Inc.) following approved protocols from University of California, Davis Institutional Animal Care and Use Committee (#17455) and University of California, Berkeley Animal Care and Use Committee (AUP-2015-01-7053) and stored in 95–100% ethanol. Only degraded tissue was available for C. diabolis, as described in [90]. Field research and export/collection permits were authorized by the BEST Commission in the Bahamas, the Ministry of Protected Areas and Biodiversity in the Dominican Republic, and the U.S. Fish & Wildlife Service and National Park Service.
DNA was extracted from muscle tissue using DNeasy Blood and Tissue kits (Qiagen, Inc.) and quantified on a Qubit 3.0 fluorometer (Thermofisher Scientific, Inc.). Genomic libraries were prepared using the automated Apollo 324 system (WaferGen Biosystems, Inc.) at the Vincent J. Coates Genomic Sequencing Center (QB3). Samples were fragmented using Covaris sonication, barcoded with Illumina indices, and quality checked using a Fragment Analyzer (Advanced Analytical Technologies, Inc.). Nine to ten samples were pooled in four different libraries for 150PE sequencing on four lanes of an Illumina Hiseq4000.
2.8 billion raw reads were mapped from 42 individuals to the Cyprinodon reference genome (NCBI, C. variegatus Annotation Release 100, total sequence length = 1,035,184,475; number of scaffold = 9,259, scaffold N50, = 835,301; contig N50 = 20,803) with the Burrows-Wheeler Alignment Tool [107] (v 0.7.12). Duplicate reads were identified using MarkDuplicates and BAM indices were created using BuildBamIndex in the Picard software package (http://picard.sourceforge.net(v.2.0.1)). We followed the best practices guide recommended in the Genome Analysis Toolkit [108](v 3.5) to call and refine our SNP variant dataset using the program HaplotypeCaller. We filtered SNPs based on the recommended hard filter criteria (i.e. QD < 2.0; FS < 60; MQRankSum < -12.5; ReadPosRankSum < -8) [97,108] because we lacked high-quality known variants for these non-model species. Our final dataset after filtering contained 16 million variants and a mean sequencing coverage of 7.2X per individual (range: 5.2–9.3X).
We used the machine learning program SAGUARO [56] to identify regions of the genome that contain different signals about the evolutionary relationships across San Salvador Island and outgroup Cyprinodon species. Saguaro combines a hidden Markov model with a self-organizing map (SOM) to characterize local phylogenetic relationships among individuals without requiring a priori hypotheses about the relationships. When diploid data is used, the SOM selects one allele at random for training. This method infers local relationships among individuals in the form of genetic distance matrices and assigns segments across the genomes to these topologies. These genetic distance matrices can then be transformed into neighborhood joining trees to visualize patterns of evolutionary relatedness across the genome. Three independent runs of SAGUARO were started using the program’s default settings and each was allowed to assign 15 different topologies across the genome. To determine how many topologies to estimate, analogous to a scree plot [109,110], we plotted the proportion of the genome explained by each hypothesized topology and looked for an inflection point (S21 Fig). We also looked at the neighborhood joining trees to assess whether additional topologies were informative about the evolutionary relationships among individuals (S21 Fig). The 15th topology and additional topologies that we investigated tended to be uninformative about the evolutionary relationships among individuals and represented less than 0.5% of the genome. We excluded the last topology (15th) from downstream analyses due to lack of genetic distinction at both the level of populations and species included in the proposed genetic distance matrix and the low percentage of the genome assigned to it. The 14 topologies included in downstream analyses and the total percentages of the genome assigned to them were robust across all three independent runs. These topologies also appeared to be fairly robust to the influences of poorly mapped regions in the genome. We generated a mask file to identify poorly mapped regions in our dataset using the program SNPable (http://bit.ly/snpable; k-mer length = 50, and ‘stringency’ = 0.5) and removed these segments from downstream analyses of the topologies. Rerunning the SAGUARO analysis on the masked dataset resulted in very similar trees across the 14 different topologies, with the exception of several generalist individuals grouping with molluscivores in the molluscivore topology (S22 Fig).
We calculated LD within each of the San Salvador Island species and compared it to estimates for the small-jawed scale-eaters to look for patterns of high linkage consistent with recent hybridization events. Pairwise LD across the largest scaffold in our dataset (4.2 Mb) was calculated for each species using the ‘r2 inter-chr’ function in PLINK v1.90 [111] for five individuals. These were chosen from a pool of individuals from Great Lake system populations (average genome-wide Fst < 0.05 across these lakes for each of the species) to balance the effects of small sample sizes and population structure on estimates of LD and more accurately compare LD decay between species. LD may be overestimated for each of the species due to the small number of individuals available to calculate it from in this study, and should be compared to estimates from other studies with caution.
We characterized the heterogeneity in introgression across the genome using f4 statistics that were initially developed to test for introgression among human populations [61–63]. The f4 statistic tests if branches among a four-taxon tree lack residual genotypic covariance (as expected in the presence of incomplete lineage sorting and no introgression) by comparing allele frequencies among the three possible unrooted trees. A previous study [53] provided evidence of potential admixture with the Caribbean outgroup species used in this study, preventing their use in a D-statistic framework which requires designation of an outgroup with no potential introgression.
To look for evidence of gene flow across the Caribbean, we focused on tests of introgression with the two outgroup clades from our sample that came from other Caribbean islands in the Bahamas and Dominican Republic. Based on the tree ((P1, P2),(C. laciniatus, C. bondi)), f4 statistics were calculated for all three possible combinations of P1,P2 among the pooled populations of generalists, scale-eaters, and molluscivores on San Salvador Island. These f4 statistics were calculated using the population allele frequencies of biallelic SNPs and summarized over windows of 10 kb with a minimum of 50 variant sites using a custom python script (modified from ABBABABA.py created by Simon H. Martin, available on https://github.com/simonhmartin/genomics_general; [64]; our modified version is provided in the supplemental material), allowing for up to 10% missing data within a population per site. All 10 molluscivore and 14 scale-eater individuals from San Salvador Island were used in the tests for comparison to the molluscivore and combined scale-eater topologies, respectively. In another calculation of f4 statistics across the genome, the 5 small-jawed scale-eater individuals were excluded for the comparison to the large-jawed scale-eater topology. Although only single individuals from New Providence Island, Bahamas and the Dominican Republic were used to represent C. laciniatus and C. bondi in the f4 tests, the individuals that were sequenced are a random sample from these populations and should be representative. This resulted in 100,276 f4 statistics (mean f4 = -2x10-4) calculated across the genome for the test that included all scale-eaters and 100,097 f4 statistics (mean f4 = -9x10-5) for the test excluding the small-jawed scale-eaters.
We conducted 1,000 permutations of the f4 test to evaluate the significance of f4 values in sliding windows across the genome. For each permutation, individuals from the four original populations were randomly assigned without replacement to one of the four populations based on the tree ((P1,P2),(P3,P4)) to assess how likely a given f4 value would be observed by chance within our empirical dataset. We calculated the 1% tails of this null distribution and used these thresholds for our candidate introgressed regions (i.e. significant at alpha = 0.02). The null distribution illustrating the 1st and 99th quantiles for all combinations of the sliding window f4 test are provided in the supplementary material (S23 Fig). Each candidate introgressed region was assigned a P-value by counting the number of permutations that had an f4 value greater than (or lesser than if the f4 value was negative) or equal to the observed value.
It is difficult to distinguish between genetic variation that is similar among taxa due to introgression from a hybridization event and that from ancestral population structure, so some of the regions with significant f4 values may represent the biased assortment of genetic variation into modern lineages from a structured ancestral population [51]. A recent simulation study [64] found that extending the use of genome-wide introgression statistics such as Patterson’s D statistic to small genomic regions can result in a bias of detecting statistical outliers mostly in genomic regions of reduced diversity. Although it hasn’t been formally tested, f4 statistics may be subject to the same biases, so we additionally considered the nucleotide diversity present in outlier f4 regions in downstream analyses by comparing π across the detected regions of introgression in comparison to scaffold- and genome-wide estimates among the three San Salvador Island species.
We then calculated several population genetic summary statistics in sliding windows across the genome to compare to the f4 patterns of introgression: Fst, between-population nucleotide divergence (Dxy), within-population nucleotide diversity (π) for pairwise species comparisons, and Tajima’s D estimates of selection in each species. Dxy between molluscivores and scale-eaters was calculated over the same 10-kb windows as the f4 tests using the python script popGenWindows.py created by Simon Martin (available on https://github.com/simonhmartin/genomics_general; [64]). Since our vcf file contained only variant sites and this script does not factor the missing sites into the calculation of Dxy by assuming they are invariant, we post-hoc incorporated the missing sites as invariant sites in the calculation of Dxy. Missing sites in our dataset may include poorly aligned regions with lots of variants, so by assuming the missing sites are all invariant, Dxy may be underestimated in this study and should be compared to diversity values from other organisms with caution.
The remaining statistics were calculated in non-overlapping sliding windows of 10 kb using ‘wier-fst-pop’, ‘window-pi’, and ‘TajimaD’ functions in VCFtools v.0.1.14 [112]. Negative values of Tajima’s D indicate a reduction in nucleotide variation across segregating sites [113], which may result from hard selective sweeps due to positive selection. To determine regions of the genome potentially under positive selection, we created a null distribution of Tajima’s D values expected for each of three species under neutral coalescent theory using ms-move [114], a program that adds more flexibility in incorporating introgression events into the coalescent simulator ms [115]. Based on the demographic history estimated for the three San Salvador Island species in a previous study [55], we incorporated a 100-fold decrease in population size approximately 10,000 years ago (-eN 0.8 0.01) and an introgression event from one population into another to mimic introgression between a San Salvador Island species and an outgroup population at the beginning of the radiation (ex. -ej 0.8 2 1 –ev 0.8 2 1 0.1). We estimated the null distribution of Tajima’s D for 100,000 loci for 10–14 individuals with a variable number of segregating sites (ranging from 50 to the maximum observed in a 10-kb window of the genomes of each species). We modeled the timing of introgression from approximately 6,000–23,000 years (based on the rough estimate of the timing of introgression of ski in this study) with 10% of population composed of migrants (although the distribution appeared robust to variations in this fraction). Tajima’s D values were calculated from the simulated loci using the ‘sample stats’ feature available in the ms package [114]. The simulated introgression event and bottleneck skewed the null distribution towards negative Tajima’s D values (S24 Fig). Windows from the observed genomes that had Tajima’s D values in the lower 2% tail of the null distribution were considered candidate regions for selective sweeps.
We also estimated regions under selective sweeps from the expected neutral folded site frequency spectrum calculated with SweeD [116]. In this calculation, we included the bottleneck of a 100-fold decrease around 10,000 years ago and the recommended grid size of 1 kb across scaffolds to calculate the composite likelihood ratio (CLR) of a sweep. The values of CLR from 1 kb windows were averaged across 10-kb to compare with the other statistics calculated in windows. Windows with an average CLR estimate above the 98th percentile across the background site frequency spectrum for their respective scaffold were considered candidate regions under a selective sweep.
We also used the function ‘wier-fst-pop’ to calculate Fst across individual SNPs to locate SNPs fixed between species and identify whether candidate adaptive introgression regions potentially contributed to species divergence. We assessed mean coverage across individuals at SNPs fixed between specialists and found that they ranged from 4.8–8.2x. The SNPs fixed in this study may be an overestimate of the variants potentially contributing to diversification in the specialists, as alleles may be missing from our individuals at these sites due to the low coverage. Average coverage and standard deviation across SNPs fixed in candidate regions are reported in the supplementary material (S3 Table). Only regions of overlap between significant f4 values, strongly negative Tajima’s D values, 98th percentile CLR estimates, and fixed SNPs between the two specialists were considered candidate adaptive introgression regions that have contributed to species divergence. For each of these regions, we looked for annotated genes and searched their gene ontology in the phenotype database ‘Phenoscape’ [117–120] and AmiGO2 [121] for pertinent functions, particularly skeletal system effects. Skeletal features, particularly craniofacial morphologies such as jaw length, have extremely high rates of diversification among the species on San Salvador Island [48,53] and likely play a key role in the diversification of this group.
While the sign of f4 hints at the directionality of introgression (e.g. for the tree (P1,P2),(P3,P4), a positive f4 value indicates gene flow either between P1 and P3 or P2 and P4), the lack of an explicit outgroup in the f4 statistics makes it difficult to determine the exact direction of gene flow among the included populations and limits our ability to determine if candidate introgressed regions came from admixture with C. laciniatus or C. bondi. We examined each candidate region for signs of directionality using several methods.
To visualize gene flow among the Caribbean populations included in this study, we used TREEMIX v1.12 [63] to estimate population graphs with 0–4 admixture events connecting populations. Population graphs were estimated for each region with a significant f4 value, each with a minimum of 50 SNPs. The number of admixture events was estimated by comparing the rate of change in log likelihood of each additional event, an approach similar to one used in Evanno et al. ([122]; also see [53]). However, this analysis should be viewed only as an exploratory tool as the reliability of TREEMIX to detect the number of admixture events has not been tested. This method was designed to be applied on genome-wide allele frequencies and estimates covariance in allele frequencies among populations in branch lengths using a model that assumes allele frequency differences between populations are solely caused by genetic drift [63]. The use of fewer SNPs (≥50) in our window-based approach also makes it harder to reliably distinguish between the different likelihoods for the number of migration events. The reliability of inference under these conditions has not been evaluated, however the migration events inferred in our TREEMIX results were consistent with our findings from our formal f4 test for gene flow.
We also compared pairwise nucleotide diversity between C. bondi, C. laciniatus, molluscivores, and scale-eaters to determine which pairs are most genetically similar in the candidate introgression regions. Since our genomic dataset only included single individuals from C. bondi and C.laciniatus and Fst estimates are a relative measure of divergence based on within population diversity, we calculated Dxy, an absolute measure of genetic divergence between-populations. Finally, we generated maximum likelihood phylogenetic trees for the SAGUAROsegment containing the fixed SNPs under a GTR+GAMMA model of sequence evolution using RaxML v.8.2.10 [123]. Support for nodes was assessed by bootstrapping, allowing the number of bootstraps determined by autoMRE function in RaxML, which ranged from 900–1,000 among regions.
|
10.1371/journal.ppat.0030118 | Nod1 Signaling Overcomes Resistance of S. pneumoniae to Opsonophagocytic Killing | Airway infection by the Gram-positive pathogen Streptococcus pneumoniae (Sp) leads to recruitment of neutrophils but limited bacterial killing by these cells. Co-colonization by Sp and a Gram-negative species, Haemophilus influenzae (Hi), provides sufficient stimulus to induce neutrophil and complement-mediated clearance of Sp from the mucosal surface in a murine model. Products from Hi, but not Sp, also promote killing of Sp by ex vivo neutrophil-enriched peritoneal exudate cells. Here we identify the stimulus from Hi as its peptidoglycan. Enhancement of opsonophagocytic killing was facilitated by signaling through nucleotide-binding oligomerization domain-1 (Nod1), which is involved in recognition of γ-D-glutamyl-meso-diaminopimelic acid (meso-DAP) contained in cell walls of Hi but not Sp. Neutrophils from mice treated with Hi or compounds containing meso-DAP, including synthetic peptidoglycan fragments, showed increased Sp killing in a Nod1-dependent manner. Moreover, Nod1−/− mice showed reduced Hi-induced clearance of Sp during co-colonization. These observations offer insight into mechanisms of microbial competition and demonstrate the importance of Nod1 in neutrophil-mediated clearance of bacteria in vivo.
| Pathogens are generally studied in the laboratory one species at a time. Most exist, however, in complex environments where they must adapt not only to their host but also to other members of the microbial flora. Using a mouse model of co-colonization, we have shown that one bacterial species (Haemophilus influenzae) can take advantage of the innate immune response of its host to outcompete and eliminate another species (Streptococcus pneumoniae) that resides in the same microenvironment of the upper respiratory tract. The molecular mechanism for this effect involves recognition of a cell wall fragment found on H. influenzae, but not on S. pneumoniae. The response to this immunostimulatory fragment requires Nod1, a host molecule that transmits inflammatory signals in response to specific peptides of the bacterial cell wall. This Nod1-mediated inflammatory stimulation triggers an increase in the ability of a type of white blood cell (neutrophil) to engulf and then kill S. pneumoniae, effectively removing it from its niche on the mucosal surface of the host airway. Our study, therefore, provides a demonstration of the importance of Nod1 in neutrophil-mediated clearance of bacterial infection. In addition, we have described a mechanism for interspecies competition between microbes that occurs through selective stimulation of host innate immune responses.
| Successful pathogens have mechanisms both to avoid triggering inflammatory responses and/or to evade the inflammatory response they induce in their host. In the case of the Gram-positive Streptococcus pneumoniae (Sp), a major pathogen of the human respiratory tract, infection involving normally sterile parts of the airway is characterized by acute inflammation with a marked and brisk recruitment of neutrophils [1]. This neutrophil influx, however, is often insufficient to clear the infection until type-specific antibody promotes opsonophagocytic killing. Before such antibody is generated, pneumococci are relatively resistant to neutrophil-mediated killing even when opsonized by complement [2]. The inability of phagocytes to eliminate pneumococci in this period may account for the rapid and often overwhelming progression of pneumococcal pneumonia, a disease responsible for more than a million deaths a year [3]. In fact, in experimental acute pneumonia, neutrophils enhance the likelihood of death without impacting bacterial clearance [4]. Likewise, in a murine model of carriage, intranasal inoculation of Sp induces recruitment of neutrophils into the nasal spaces, yet systemic depletion of neutrophils has little effect on the density of colonizing bacteria [5,6]. In contrast, when co-colonized with the Gram-negative respiratory tract bacteria Haemophilus influenzae (Hi), the neutrophil influx is sufficient to rapidly clear Sp from the mucosal surface [6]. Clearance during co-colonization is not seen if either neutrophils or complement are systemically depleted, indicating that killing occurs through neutrophil-mediated phagocytosis of Sp opsonized by complement. These in vivo observations demonstrate that one microbe can co-opt the innate immune response of the host to prevail over a competitor that resides within a similar niche. Enhanced killing of Sp can be modeled ex vivo using neutrophils derived from peritoneal exudates cells (PECs) treated in vivo with Hi or its products. Thus, components of Hi are sufficient to stimulate neutrophil activity that overcomes the resistance of complement-opsonized Sp to phagocytic killing.
The focus of this study is to define the mechanism leading to effective neutrophil-mediated killing of Sp that occurs in the absence of specific antibody. We observed that peptidoglycan fragments from Hi are sufficient to promote neutrophil-mediated phagocytosis of opsonized Sp. Pathways for the recognition of and response to peptidoglycan fragments leading to NF-κB-dependent transcriptional activation and pro-inflammatory responses have been partially characterized [7]. Peptidoglycan fragments containing the minimal structure γ-D-glutamyl-meso-diaminopimelic acid (meso-DAP) found in Gram-negative bacteria, including Hi, act through a cytoplasmic signaling molecule, nucleotide-binding oligomerization domain-1 (Nod1) [8–10]. In the peptidoglycan of most Gram-positive bacteria, including Sp, meso-DAP is replaced by lysine, a structural difference of a single carboxyl group that is sufficient to prevent effective signaling involving Nod1 [11]. In addition, another peptidoglycan fragment, muramyl dipeptide (MDP), common among most Gram-negative and Gram-positive bacteria, is the minimal structure needed for responses involving a separate cytoplasmic immune signaling molecule, Nod2 [12]. Our findings provide a demonstration of the contribution of Nod1-mediated signaling to the anti-bacterial activity of neutrophils and their ability to clear mucosal infection.
The increased ability of ex vivo PECs to kill Sp when elicited following intraperitoneal (i.p.) administration of heat-killed Hi (HKHi) allowed us to examine the mechanism whereby one species stimulates the killing of another. When HKHi-stimulated PECs were divided by density gradient centrifugation into mononuclear cell– and neutrophil-containing fractions, only the neutrophil-enriched fraction demonstrated killing of Sp (unpublished data). This result correlated with the absence of killing by HKHi-stimulated PECs when elicited from mice depleted of neutrophils by prior treatment with RB6-8C5, an antibody to murine Ly6.G [6,13]. Addition of HKHi correlated with increased neutrophil activation as confirmed by increased expression of the marker Mac-1(complement receptor 3 [CR3], CD11b/CD18) in cells co-expressing Ly6.G [6]. Moreover, increased killing of Sp following administration of HKHi was observed with neutrophil-enriched PECs derived from parental but not congenic Mac-1−/− mice (Figure 1). This finding pointed to the requirement of complement-mediated opsonization for neutrophil recognition. When heat-inactivated serum or serum from C3−/− mice was used as a complement source, no killing by HKHi-stimulated neutrophil enriched PECs was seen, confirming the requirement of active complement. Although C3 may be activated by either classical or alternative pathways, killing in the presence of serum from scid mice lacking antibody made it less likely that complement was being activated by the classical pathway [6]. The requirement for the alternative pathway was confirmed by showing a lack of Sp killing when serum from factor B–deficient mice was used as a complement source (Figure 1). Thus, results using PECs indicated that products of Hi stimulate neutrophil-mediated phagocytic killing of Sp opsonized primarily by activation of the alternative pathway of complement.
Products of Hi have previously been shown to signal pro-inflammatory responses through toll-like receptor (TLR) 2 and TLR4, through recognition of its lipopolysaccharide (LPS) and lipoproteins, respectively [14,15]. In addition, platelet-activating factor receptor (rPAF)-mediated signaling has been described for those Hi phase variants expressing the cell surface ligand phosphorylcholine [16]. Opsonophagocytic killing was assessed in neutrophil-enriched PECs derived from TLR2−/− mice. These showed increased killing in response to HKHi and were as active as cells derived from the TLR2-expressing mouse strain (Figure 2). Opsonophagocytic killing was also compared in neutrophil-enriched PECs derived from CH3/OuJ and C3H/HeJ mice, which express functional and non-functional TLR4, respectively. TLR4 did not contribute to Sp killing in response to HKHi stimulation. Moreover, HKHi derived from isogenic mutants expressing or not expressing phosphorylcholine stimulated similar levels of Sp killing by neutrophil-enriched PECs (unpublished data) [17]. Together, these results showed that the enhancement of opsonophagocytic killing occurs independently of non-redundant signaling involving known cell surface pattern recognition receptors for Hi, including TLR2, TLR4, and rPAF.
This unexpected finding led us to characterize the signal from Hi that enhances the opsonophagocytic killing of Sp. Neither lysis of HKHi by sonication, nor prior treatment with proteinase K, diminished stimulation of killing by neutrophil-enriched PECs, which indicates the involvement of a non-proteinaceous bacterial product. However, there was no stimulation of neutrophil-enriched PECs by purified LPS (in doses up to 50 μg/animal) extracted from Hi or Escherichia coli (Figure 3). These findings were also consistent with a signaling pathway other than recognition of Hi components by TLR2, TLR4, or rPAF.
In contrast, purified Hi peptidoglycan at a dose as low as 1 μg/animal was sufficient to stimulate increased killing of Sp by neutrophil-enriched PECs (activity equivalent to 107 HKHi). Purified peptidoglycan from Sp (or Staphylococcus aureus) was less active even when administered at a 10-fold higher dose (Figure 3 and unpublished data). The greater potency of Hi peptidoglycan correlated with the stimulation of killing by HKHi but not HKSp [6]. This observation indicated that structural differences between cell wall fragments of these species may be an important determinant of their peptidoglycan-mediated signaling. To confirm this hypothesis, FK-156, a synthetic muropeptide containing meso-DAP, was tested and showed a level of stimulatory activity equivalent to purified Hi peptidoglycan when administered at an equivalent concentration. Experiments with FK-156 also demonstrated that Hi peptidoglycan could provide a sufficient stimulus to neutrophil-enriched PECs that accounts for their enhanced killing of Sp and makes it unlikely that a contaminant in the peptidoglycan preparation could explain our findings. In contrast, MDP at the equivalent concentration was a relatively poor stimulus.
The potency of Hi peptidoglycan, as well as that of FK-156, suggested that stimulation of opsonophagocytic killing involved recognition of Hi components by Nod1. In order to examine this possibility, neutrophil-enriched PECs from Nod1−/− mice were analyzed for their response to HKHi and FK-156. As predicted, administration of FK-156 (10 μg/animal) stimulated Sp killing by cells in parental, but not in Nod1−/− mice (Figure 4A). Neutrophil-enriched PECs from Nod1−/− mice also showed a diminished response to HKHi, demonstrating that Nod1 accounts for a significant proportion of the signaling generated by innate recognition of this organism.
To further confirm this observation, a meso-DAP-containing peptide, murNAcTRIDAP, was synthesized using the Mur enzymes of Gram-negative bacteria [18]. As predicted, its ability to stimulate Sp killing by neutrophil-enriched PECs was equivalent to that of FK-156 and dependent on Nod1 (Figure 4A). In contrast, a synthetic form of the corresponding lysine-containing tripeptide found in Sp peptidoglycan, murNAcTRILYS, lacked stimulatory activity at the same concentration. In Figure 4B, the effect of these peptides on killing is compared in neutrophil-enriched PECs elicited with peptide in buffer alone without the addition of casein to show that administration of murNAcTRIDAP is sufficient and murNAcTRILYS is insufficient to enhance killing of Sp.
Our findings using neutrophil-enriched PECs stimulated in vivo and tested in killing assays ex vivo suggested that co-colonization with Hi in competition experiments with Sp should promote clearance of Sp in a Nod1-dependent manner. Indeed, a significant decrease in the density of Sp colonization was observed in Nod1+/+ but not in Nod1−/− mice co-infected with Hi (Figure 5). The reduced interspecies competition in the absence of Nod1 demonstrated an important role for peptidoglycan recognition in the innate response to Gram-negative bacteria on the mucosal surface.
Next, we explored the mechanism for increased opsonophagocytic killing stimulated through Nod1. Levels of the proinflammatory chemokine MIP-2, which functions as a murine neutrophil attractant and activator, were previously shown to correlate with neutrophil influx into the nasal spaces [6]. MIP-2 levels increased in response to co-colonization, but were not significantly different between co-colonized Nod1−/− and parental mice (Figure 6A). Analysis of tissue sections from co-colonized mice, both Nod1−/− and parental, showed an intimate association of both Sp (and Hi) with neutrophils in the lateral nasal spaces (Figure 6B). These results suggested that the recruitment of neutrophils and their migration to mucosal sites with bacteria were not affected by the expression of Nod1 in this model. Additional evidence that Nod1 did not impact neutrophil migration came from comparisons by flow cytometry of PECs elicited by HKHi or FK-156 (Figure 6C). No difference between Nod1−/− and parental mice was seen in the proportion of total cells expressing Ly6.G (neutrophils). For both Nod1−/− and parental mice, Ly6.G positive cells also expressed the markers CD18/CD11b (activated neutrophils).
We next considered whether Nod1 signaling affected uptake or killing of bacteria. Gentamicin sulfate was added at the end of killing assays to determine the proportion of Sp surviving within neutrophils as a measure of phagocytic activity and killing. Comparison of neutrophil-enriched PECs elicited by HKHi from Nod1+/+ or Nod1−/− mice showed no difference in the proportion of viable intracellular bacteria (Figure 6D). Preincubation of neutrophil-enriched PECs with cytochalasin D, to inhibit actin rearrangements and block phagocytosis, resulted in minimal survival after gentamicin treatment. These findings confirmed the role of phagocytosis in neutrophil-mediated killing and suggested that Nod1 signaling did not affect the uptake of Sp. Killing of Sp by neutrophil-enriched PECs elicited by HKHi from Nod1+/+ or Nod1−/− mice was also not affected by pretreatment with dibenziodolium chloride (DPI), a blocker of the oxidative burst. Together, these observations suggest that Nod1 signaling acts on events following phagocytosis on a non-oxidative pathway for killing Sp.
Although numerous studies have defined Nod1-mediated effects of bacteria or their cell wall products in vitro, our understanding of its contribution to innate immune responses to bacterial infection in vivo remains limited (reviewed in [7,19]). We demonstrate here that the Nod1 signaling pathway can respond to meso-DAP-containing compounds to increase clearance of Sp from the mucosal surface of the airway. Thus, Nod1 was shown to be important in dictating the outcome of competition between two pathogens that occupy a similar niche in their host [6]. Enhanced killing of Sp required products from another organism, since cell wall fragments from Sp, like most Gram-positive species, do not signal through Nod1. Our findings are relevant to polymicrobial infection and situations in which products from multiple types of organisms are present. This information adds to our previous report, which describes how combinations of microbes and microbial products synergize to enhance inflammatory responses [20]. Mucosal surfaces, in particular, are generally colonized simultaneously with multiple species. The paradigm of one species promoting an innate immune response that affects a competitor may be underappreciated, because most models of infection typically examine responses to individual microbial species.
While our model was useful in revealing a role for Nod1 in vivo, it also demonstrates that bacteria that succeed in such environments must have mechanisms to evade its clearance-promoting effects. The specificity for bacterial cell wall components that act through Nod1 suggests a mechanism whereby many Gram-positive pathogens that lack meso-DAP may avoid signaling events that lead to neutrophil-mediated killing. Likewise, the density of colonizing Hi during co-infection was not affected by Nod1 signaling, in contrast to clearance of Sp during co-colonization. This suggests that Hi may be resistant to the response induced by its meso-DAP-containing peptidoglycan, and also to the enhancement of opsonophagocytic killing by neutrophils seen against Sp. In addition to the synthesis of stem peptides without meso-DAP, there may be multiple mechanisms to evade peptidoglycan recognition and stimulation of immune signaling through Nod1 [21]. For example, it has recently been reported that modification of the α-carboxylic acid group of iso-glutamic acid, the residue proximal to meso-DAP, to an amide diminishes signaling through Nod1 and may be a mechanism for immune invasion by some pathogens [22].
Both Sp and Hi are considered extracellular pathogens, which are unable to effectively access intracellular pathways [23]. In the case of epithelial cells, pore-forming toxins or delivery via the type IV pilus have been shown to be necessary for peptidoglycan to gain access to the cytoplasm [24,25]. Moreover, Nod1-deficient mice were shown to be more susceptible to infection by Helicobacter pylori expressing the cag pathogenity island type IV secretion apparatus than were wild-type mice [25]. Our observation in this report that peptidoglycan fragments alone are sufficient to induce Nod1-dependent effects shows that access to these cytoplasmic pathways may not be similarly limited for professional phagocytes. In killing assays, however, bacteria or peptidoglycan fragments were delivered in vivo and their activity tested ex vivo. Thus, we cannot confirm whether the effect of injected compounds or bacterial products on neutrophil function was direct. Attempts to treat neutrophils in vitro with immunostimulatory fragments that are active when provided in vivo were not sufficient to elicit a direct effect in killing assays. It is unlikely that this is due to a lack of Nod1 expression in these cells, because in contrast to other members of the Nod protein family, Nod1 expression is ubiquitous [7]. It remains possible that Nod1-mediated signaling requires other cell types, such as epithelial cells, and that its effects on neutrophil function are indirect.
A further consideration is that neutrophils have cell wall–degrading enzymes, such as lysozyme, that may generate more biologically active peptidoglycan fragments. This could account for the effects of purified peptidoglycan in our study, which contrasts with prior reports where only synthetic products are active. Thus, both the processing of peptidoglycan and the ability of cell wall fragments to access the cytoplasm may be important factors for signaling events involving neutrophils. In this regard, it has been suggested that Sp and other Gram-positive pathogens synthesize modified peptidoglycan that is resistant to lysozyme [21,26,27]. Thus, a number of adaptations may contribute to minimizing Nod-mediated signaling by Gram-positive bacteria despite their greater quantity of peptidoglycan per cell.
Our study demonstrates that the resistance of Sp to killing by neutrophils (Figure 4B) can be overcome by a specific immune signaling pathway. Findings in this study with microbial products and synthetic meso-DAP-containing peptidoglycan fragments add to a prior report that systemic administration of FK-156 enhanced host resistance to various microbial infections [28]. Bacterial killing in our system required opsonization, which for Sp strain Sp1121 occurred through activation of the alternative pathway of complement, followed by phagocytosis by activated, Mac-1 (CR3, CD11b/CD18)-expressing neutrophils. One of the ligands of Mac-1, or CR3, is iC3b [29]. It remains unclear how Nod1-mediated signaling enhances Mac-1-dependent opsonophagoytic killing of complement-opsonized Sp. It has been suggested that Nod1 transduces signals that can stimulate chemokine production and neutrophil recruitment [30]. We did not observe, however, a Nod1-related effect on the increase in MIP-2 levels or influx of neutrophils into either the peritoneal cavity or the nasal spaces in response to bacteria. Likewise, no effect of Nod1 on the uptake of bacteria or generation of an oxidative burst was detected. Rather, killing of Sp in our model resulted from stimulation of non-oxidative activity of neutrophils. Reduced killing in the presence of inhibitors of actin polymerization and rearrangement, and the requirement for complement, argue against Nod1-mediated enhancement of previously described mechanisms for extracellular killing of Sp by neutrophils [31]. We are currently characterizing this oxidative burst–independent anti-pneumococcal effect of neutrophils and the contribution of Nod1 to stimulation of this biological activity.
Findings in this study also show a limited role of other signaling pathways in clearance of Sp from the mucosal surface of the murine airway. Sp has previously been shown to activate cellular NF-κB-dependent immune responses through Nod2. However, the effect of fragments acting through Nod2, including MDP, purified Hi or Sp peptidoglycan, and live or killed Hi or Sp, was minor in comparison to those acting through Nod1 [6,32]. Moreover, the Hi-induced increase in Sp killing by neutrophil-enriched PECs was not influenced by the pathogen-associated molecular pattern receptors, TLR2 or TLR4, in a non-redundant manner. Thus, our study provides an example where the predominant signaling response of the innate immune system to a bacterial challenge appears to be through Nod1.
Hi and Sp strains were grown as previously described [33]. Strains used in vivo were selected because of their ability to colonize efficiently the murine nasal mucosa and included Hi636 (a type b capsule-expressing, spontaneously streptomycin-resistant mutant of Hi strain Eagan), and Sp1121 (a type 23F capsule-expressing Sp isolate from the human nasopharynx [34]). Genetically modified Hi mutants of strain Eagan that constitutively express or lack phosphorylcholine on its LPS were previously described [17].
Six-week-old mice used in the study were housed in accordance with Institutional Animal Care and Use Committee protocols. Mouse strains included C57Bl/6J and congenic Nod1−/− (Millennium Pharmaceuticals, http://www.mlnm.com/), B6.129S4-Itgamtm1Myd/J (Jackson Laboratories, http://www.jax.org/), and TLR2−/− (provided by H. Shen, University of Pennsylvania). Mac-1 (CD11b/CD18)-deficient mice (Jackson Laboratories) have a targeted mutation in the gene for integrin alpha M or CR3 [35]. Neutrophils from these animals are deficient in phagocytosing complement-opsonized particles and in several Fc-mediated functions. The genotype of Nod1−/− (CARD4-deficient) mice was confirmed by PCR using primers CARD4-F2 (5′-CTTAGGCATGACTCCCTCCTGTCG-3′), CARD4-R1 (5′-GATCTTCAGCAGTTTAATGTGGGAGTGAC-3′), and CARD4-RB (5′-CCATTCAAGCTGCGCAACTGTTG-3′). Sequences and PCR protocols were supplied by Charles River Laboratories Genetic Testing Services (http://www.criver.com/), where the colony was derived. TLR2−/− mice have a targeted disruption of the gene encoding the C-terminus of the extracellular domain of TLR2 and display an increased susceptibility to bacterial infections [36]. Serum was also obtained from factor B–deficient and C3-deficient mice (provided by J. Lambris, University of Pennsylvania) [37,38]. TLR4-sufficient and -deficient mice were obtained from Jackson Laboratories. C3H/HeJ (TLR4-deficient) mice have a spontaneous mutation that occurred in wild-type C3H/HeOuJ (TLR4-sufficient) mice at an LPS response locus (mutation in TLR4 gene), making C3H/HeJ mice resistant to endotoxin [39].
Mice were used in a previously described model of nasal colonization with Sp and Hi [34]. Briefly, groups of at least ten mice per condition were inoculated intranasally with 10 μl containing 1 × 107 CFU of PBS-washed, mid-log phase Hi, Sp, or both applied separately to each naris. Then, 24 h post-inoculation, the animal was sacrificed, the trachea cannulated, and 200 μl of PBS instilled. Lavage fluid was collected from the nares for determination of viable counts of bacteria in serial dilutions plated on selective medium containing antibiotics to inhibit the growth of contaminants (100 μg/ml streptomycin to select for Hi636, and 20 μg/ml neomycin to select for Sp1121).
Neutrophil-enriched PECs were isolated as previously described [40]. Briefly, phagocytes were obtained by lavage of the peritoneal cavity (8 ml/animal with PBS containing 20 mM EDTA) of mice treated 24 h and again 2 h prior to cell harvest by i.p. administration of 10% casein in PBS (1 ml/dose). Administration of casein provided for a higher and more consistent yield of cells. Cells collected from the peritoneal cavity lavage (PECs) were enriched for neutrophils or monocytic cells using separation by a Ficoll density gradient centrifugation according to the manufacturer's protocol (MP Biomedicals, http://www.mpbio.com/). Neutrophil or monocytic cell-enriched fractions were collected and washed with 5 ml of Hank's buffer without Ca++ or Mg++ (GIBCO, http://www.invitrogen.com/) plus 0.1% gelatin. An aliquot of these cells was characterized using FACS for staining of granulocytes with anti-mouse Gr-1 mAb to Ly6.G (BD Biosciences, http://www.bdbiosciences.com/) and showed >90% positively stained cells following enrichment. Additional characterization involved staining for CD11b/CD18 (BD Biosciences). Where indicated, heat-inactivated Hi (Hi636), bacterial components, FK-156 (an analog of meso-DAP provided by Astellas Pharmaceuticals, http://www.us.astellas.com/), or synthetic peptidoglycan fragments were co-administered intraperitoneally with or without the casein solution as indicated. PBS-washed, mid-log phase bacteria (107 cells/animal) were heat-inactivated by treatment at 65 °C for 30 min and shown to be non-viable.
Neutrophil-enriched PECs were counted by trypan blue staining and adjusted to a density of 7 × 106 cells/ml. Killing during a 45-min incubation at 37 °C with rotation was assessed by combining 102 PBS-washed, mid-log phase bacteria (in 10 μl) with complement source (in 20 μl), 105 mouse phagocytes (in 40 μl), and Hank's buffer with Ca++ and Mg++ (GIBCO) plus 0.1% gelatin (130 μl). Earlier time points and fewer effector to target cells were shown in pilot experiments to result in less killing. The complement source consisted of fresh mouse serum from C57Bl/6 mice unless indicated otherwise. After stopping the reaction by incubation at 4 °C, viable counts were determined in serial dilutions. Percent killing was determined relative to the same experimental condition without i.p. administration of bacterial products or FK-156 (casein alone). For groups without co-administered casein, the percent killing was calculated by comparison to controls with inactivated complement (56 °C for 30 min) where there was no loss of bacterial viability. Additional controls consisting of heat-inactivated Hi636 administered without casein gave similar levels of killing, confirming that killing was stimulated by bacterial products rather than by casein.
Where indicated, neutrophils were preincubated with 10 μM DPI, an NADPH-ubiquinone oxidoreductase inhibitor, for 15 min at 37 °C. The respiratory burst of activated neutrophils and its inhibition by DPI was assessed by cytochrome C oxidation with activation by treatment with 25 nM phorbol 12-myristate 13-acetate (PMA; Sigma, http://www.sigmaaldrich.com/) as a control. To inhibit phagocytosis, neutrophils were pretreated with cytochalasin D (20 μM, Sigma) for 15 min at 37 °C. Intracellular pneumococci were quantified using viable counts following the addition of gentamicin sulfate (final concentration 300 μg/ml). After a 20-min incubation at 37 °C, the antibiotic was removed by serial washing prior to plating for viable counts.
Hi LPS was purified by hot-phenol extraction from strain Eagan as previously described [41]. E. coli LPS and Staphylcoccus aureus peptidoglycan were purchased from Sigma.
Preparation of peptidoglycan from Hi was modified from a previously described protocol [42]. Briefly, strain Hi636 was grown overnight in sBHI, pelleted at 6,000g at 4 °C, and washed with Tris-buffered saline (TBS). The pellet was resuspended in 5 ml of cold dH2O, and cells were lysed in boiling SDS (5%) for 30 min. Lysates were collected at 150,000g, resuspended in dH2O, and washed once with TBS. Glycogen and nucleic acids were removed by treatment with α-amylase (Fluka 10070, from Bacillus subtilis) and DNAse/RNAse A (Sigma) for 2 h at 37 °C, followed by overnight incubation at 37 °C with agitation and 100 μg/ml of trypsin (Worthington Biochemical, http://www.worthington-biochem.com/) in the presence of 10 mM CaCl2. To stop the reaction, 10 mM EGTA was added and the peptidoglycan preparation was boiled in 5% SDS for 30 min. After extensive washing, Hi peptidoglycan was lyophilized and resuspended at 5 mg/ml in endotoxin-free water.
For preparation of peptidoglycan from Sp, bacteria were grown in tryptic soy medium and treated as above, except cells were also treated in 0.5% Na-layrilsarcosin prior to boiling in SDS (5%) [43]. Sp peptidoglycan was additionally treated with hydrofluoric acid (49% for 48 h at 4 °C with agitation) to remove teichoic acid as described [44]. The pellet was washed extensively with dH2O, twice with acetone, and lyophilized.
N-acetlymuramyl-L-alanyl-γ-D-glutamyl-meso-2,6-diaminopimelic acid (murNAcTRIDAP), N-acetlymuramyl-L-alanyl-γ-D-glutamic acid (murNAcDI), and N-acetlymuramyl-L-alanyl-γ-D-glutamyl-L-lysine (murNAcTRILYS) were prepared as described previously [18,45,46]. Briefly, recombinant Pseudomonas aeruginosa (Pa) MurA, MurB, MurC, and MurD were used to synthesise uridine 5′diphosphoryl-N-acetlymuramyl-L-alanyl-γ-D-glutamic acid (UDP-murNAcDI); additionally, Pa MurE was used to synthesise uridine 5′diphosphoryl-N-acetlymuramyl-L-alanyl-γ-D-glutamyl-meso-2,6-diaminopimelic acid (UDP-murNAcTRIDAP), and Sp MurE was used to synthesise uridine 5′diphosphoryl-N-acetlymuramyl-L-alanyl-γ-D-glutamyl-L-lysine (UDP-murNAcTRILYS). Electrospray ionization mass spectrometry (negative ion) was used to confirm the molecular weight of synthesized compounds. Purity was assessed by analytical anion exchange chromatography using a GE Healthcare Mono Q HR5/5 column (http://www.gelifesciences.com/) and by continuous spectrophotometric enzyme assay with MurE for UDP-murNAcDI, and MurF for UDP-murNAcTRIDAP and UDP-murNAcTRILYS. N-acetlymuramyl-peptides were produced by mild acid hydrolysis (0.1 M HCl, 100 °C, 1 h) of the corresponding uridine 5′diphosphoryl-N-acetlymuramyl-peptides. Complete hydrolysis was confirmed by continuous spectrophotometric enzyme assay with MurE for murNAcDI (MDP), and MurF for murNAcTRIDAP and murNAcTRILYS. Peptides were analysed by electrospray ionisation mass spectrometry (positive ion). The concentration of other hydrolysis products (UDP, UMP, and Pi) was established by continuous spectrophotometric enzyme assay. Corresponding concentrations of UDP, UMP, and Pi were added to the casein-only control.
At 24 h post-inoculation, the animal was sacrificed and decapitated, and the head was fixed for 48 h in 4% paraformaldehyde in PBS. The head was then decalcified by serial incubations in 0.12 M EDTA (pH 7.0) at 4 °C over 1 mo before freezing in Tissue-Tek O.C.T. embedding medium (Miles, Elkhart, Indiana, United States) in a Tissue-Tek Cryomold. Then, 5-μm-thick sections were cut, air dried, and stored at −80 °C. Frozen-imbedded tissue sections were stained with hematoxylin and eosin (H&E) following a 10-min fixation step in 10% neutral buffered formalin (NBF). Sections were then dehydrated in alcohol, cleared in xylene, and mounted in cytoseal (Richard-Allan Scientific, http://www.rallansci.com/). Immunofluorescent staining on frozen tissue was performed and visualized as previously described [47]. Neutrophil-like cells were stained using rat anti-mouse Ly6G mAb (BD Biosciences) followed by anti-rat Ig secondary antibody [6]. To detect Sp1121, sections were incubated with antisera to Sp type 23F (Statens Serum Institut, http://www.ssi.dk/) followed by anti-rabbit Ig secondary antibody. To detect Hi636, sections were incubated with antisera to Hi type b (DIFCO Laboratories, http://www.bd.com/ds/) followed by anti-rabbit Ig secondary antibody.
Upper respiratory tract lavage fluid was assayed for the concentration of macrophage inhibitory protein (MIP-2) by ELISA in duplicate according to the manufacturer's instructions (Pharmingen, http://www.bdbiosciences.com/).
Statistical comparisons of colonization among groups were made by the Kruskal–Wallis test with Dunn's post-test (GraphPad Prism 4; GraphPad Software, http://www.graphpad.com/). In vitro killing assays were compared by ANOVA with Tukey post-tests as appropriate.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank/index.html) accession number for murine Nod1 is NM_172729. |
10.1371/journal.pntd.0002480 | Epilepsy and Neurocysticercosis in Latin America: A Systematic Review and Meta-analysis | The difference in epilepsy burden existing among populations in tropical regions has been attributed to many factors, including the distribution of infectious diseases with neurologic sequels. To define the burden of epilepsy in Latin American Countries (LAC) and to investigate the strength of association with neurocysticercosis (NCC), considered one of the leading causes of epilepsy, we performed a systematic review and meta-analysis of the literature.
Studies published until 2012 were selected applying predefined inclusion criteria. Lifetime epilepsy (LTE) prevalence, active epilepsy (AE) prevalence, incidence, mortality, treatment gap (TG) and NCC proportion among people with epilepsy (PWE) were extracted. Median values were obtained for each estimate using random effects meta-analysis. The impact of NCC prevalence on epilepsy estimates was determined using meta-regression models. To assess the association between NCC and epilepsy, a further meta-analysis was performed on case-control studies.
The median LTE prevalence was 15.8/1,000 (95% CI 13.5–18.3), the median AE prevalence was 10.7/1,000 (95% CI 8.4–13.2), the median incidence was 138.2/100,000 (95% CI 83.6–206.4), the overall standardized mortality ratio was 1.4 (95% CI 0.01–6.1) and the overall estimated TG was 60.6% (95% CI 45.3–74.9). The median NCC proportion among PWE was 32.3% (95% CI 26.0–39.0). Higher TG and NCC estimates were associated with higher epilepsy prevalence. The association between NCC and epilepsy was significant (p<0.001) with a common odds ratio of 2.8 (95% CI 1.9–4.0).
A high burden of epilepsy and of NCC in LAC and a consistent association between these two diseases were pointed out. Furthermore, NCC prevalence and TG were identified as important factors influencing epilepsy prevalence to be considered in prevention and intervention strategies.
| Epilepsy affects approximately 70 million people worldwide and at least five million people in Latin America. Many researchers have pointed out a different distribution of epilepsy in Latin American countries, with some regions presenting higher frequencies and others presenting lower frequencies. This difference in epilepsy distribution has been attributed to many factors, mainly related to the allocation of health resources and to the presence of environmental and infectious risk factors. Among the latter stands neurocysticercosis, a parasitic disease that has been recognized as the leading cause of acquired epilepsy in the developing world, with a particularly elevated distribution in rural settings. In this study, we performed a statistical analysis to investigate whether neurocysticercosis distribution affects epilepsy distribution among Latin American countries and the relationship between these two conditions. The combined results of the studies included indicated that neurocysticercosis influences epilepsy frequency in Latin America, as countries with higher epilepsy distribution presented also higher neurocysticercosis frequency. Moreover, another analysis pointed out an association between the two diseases. These results appear very important considering that parasitic infections are modifiable factors and that their reduction may contribute to decrease epilepsy burden worldwide.
| Epilepsy is one of the most prevalent non-communicable neurologic diseases [1], with an estimated aggregate burden of around 0.5% of the total disease burden [2]. It affects approximately 70 million people worldwide [3] and at least five million people in Latin American Countries (LAC) [4].
The epidemiological studies describing the burden of epilepsy across the world have frequently reported the presence of important differences in the estimate of prevalence and incidence [5]. The median lifetime epilepsy (LTE) prevalence, ranges from 5.8/1,000 (range 2.7–12.4) in developed countries to 15.4/1,000 (range 4.8–49.6) in rural areas of developing countries [3], and similar variations are also reported for active epilepsy (AE) prevalence and for incidence [6]. Considering LAC, the median LTE prevalence ranges from 6/1,000 to 43.2/1,000, while the median AE prevalence from 5.1/1,000 to 57/1,000 [7], showing a very large range of variability.
The wide prevalence difference existing among populations may be mainly attributed to country resources and development-related factors [8] to spatial clustering of etiologic and risk factors [9], and to methodological limitations of studies [10]. In a recent review on the global burden of epilepsy, a meta-regression analysis showed that location, study size and age of study participants explained 53% of the variance in LTE prevalence [3].
However among the factors probably responsible of the unexplained amount of variance, the distribution of epilepsy-related biological and environmental factors, such as infections of the central nervous system (CNS), may be important, especially in resource limited countries, but had never been taken into consideration in previous meta-analysis.
Among CNS infections, neurocysticercosis (NCC) is considered the leading cause of acquired epilepsy in the developing world [11], [12]. Although it has been declared eradicable by the International Task Force for Disease Eradication of World Health Organization (WHO) in 1993, NCC is still recognized as a “major neglected disease” due to the lack of information about its burden and transmission, the lack of diagnostic tools available in resource-poor areas, and the lack of intervention strategies for its control [13]. Recent data indicate that NCC represents a significant health problem in endemic areas, causing epilepsy in 0.6–1.8% of the population [13]. This indicates that between 450,000 and 1.35 million persons suffer from epilepsy due to NCC in LAC only [14].
Understanding the reasons that influence epilepsy distribution is crucial to improve and tailor intervention programs and prevention strategies. Thus, to better define the burden of epilepsy of NCC and their association in LAC, we conducted a systematic review and meta-analysis of epilepsy prevalence, incidence, mortality, treatment gap (TG) and of NCC prevalence among people with epilepsy (PWE) in LAC.
Two systematic searches, without language restriction, were conducted to identify all relevant articles concerning “burden of epilepsy” and “prevalence and association between cysticercosis (CC)/NCC and epilepsy”. The following electronic databases were independently examined by two authors (EB and JB) to identify articles published until the 1st July, 2012: MEDLINE, IMBIOMED, LILACS, EMBASE, SciELO, PAHO Library Online Catalog, PAHO Evidence Portal, WHOLIS, Cochrane Library. Additional searches were performed on bibliographies of pertinent original articles, reviews, abstracts and book chapters. Combined text words and Medical Subject Headings (MeSH) terminology were used. Searches were organized using the following search terms to develop a search strategy: “epilep*” and “mortality”, “ preval*”, “incidenc*”, “epidemiol*”, “ surve*”, “rate*”, “frequenc*”, “treatment gap”, “cysticerc*”, “neurocysticerc*”, “taenia*”, “Argentina”, “Bolivia”, “Brazil”, “Chile”, “Colombia”, “Costa Rica”, “Cuba”, “Dominican Republic”, “Ecuador”, “El Salvador”, “Guatemala”, “Guyana”, “Honduras”, “Mexico”, “Nicaragua”, “Panama”, “Paraguay”, “Peru”, “Puerto Rico”, “Suriname”, “Uruguay”, “Venezuela”, “Latin America”. The literature search was adapted for the different databases.
Two reviewers (EB and JB) independently assessed the titles and abstracts of all the studies identified. The full copies of papers requiring further consideration were obtained. Relevant studies were selected according to the criteria outlined above and data were independently extracted on a predefined collecting form.
Flowcharts of the literature searches are shown in figure S1 and S2. Of 48 retained articles on epilepsy burden (table S1), 41 reported prevalence, five incidence, 14 TG and five mortality (table S2). Most studies evaluated both adults and children and methods of ascertainment of epilepsy were mainly based on both questionnaire and neurological examination. The screening instruments more frequently adopted were the 1991 WHO questionnaire (WHO, 1991) and the questionnaires used by Placencia [25]–[31] and by Pradilla [32], [33]. The 1993 ILAE definition of AE was the most frequently used but some studies reported a narrower time frame, considering the previous 24 months [33], [34] or the previous 12 months [35].
Thirty-one studies described the proportion of CC/NCC among PWE in LAC (table S3 and S4). Proportion of patients with positive brain CT scan ranged from 8.8% [36] to 70% [37]. Serological diagnosis of CC/NCC with EITB was performed in 16 studies, and the proportion ranged among 0% [38] and 39.5% [30]. Thirteen studies associated to neuroimaging or EITB other ascertainment methods for the diagnosis of CC/NCC, such as serum AbELISA, serum AgELISA, CSF AbELISA, serum ELISA/serum immunofluorescence assay, or CT/CSF test/surgery [39], with proportions ranging from 0% [38] to 41.9% [40].
The association between CC/NCC and epilepsy was evaluated in ten of the 31 studies (table S4). In nine of them the association was significant, with a OR ranging between 2.92 [41] and 12.25 [42]. Only one study reported absence of CC/NCC cases among both PWE and controls [43].
Published data on epilepsy in LAC demonstrated a median LTE prevalence of 15.8/1,000, a median AE prevalence of 10.7/1,000 (both higher in rural areas), a median incidence of epilepsy of 138.2/100,000 and an enormous TG in rural areas. Furthermore a median NCC proportion among PWE of 32.3% (by CT scan) and a consistent association between NCC and epilepsy were found.
To minimize biases we performed a comprehensive systematic review, with particular attention to the main Latin American and Caribbean biomedical databases. Nevertheless, lack of epidemiological data on epilepsy from Nicaragua, Venezuela, Paraguay, Suriname, Guyana and a high information gap on incidence, mortality, TG and NCC prevalence in all LAC were pointed out, confirming epilepsy and NCC as major neglected diseases in this region.
The elevated burden of these diseases should be regarded as a primary public-health issue in LAC, especially in rural settings. Although slightly lower, our estimates were close to those reported in a previous meta-analysis (LTE prevalence 17.8/1,000; AE prevalence 12.4/1,000, [7]) and above the median values reported in a recent work considering both developed and developing countries [3]. Also epilepsy incidence in urban and rural setting was greater than that reported in another study analysing low- and middle-income countries [6]. The pooled NCC proportion among PWE was 32.3% (95% CI 26.0–39.0), higher than 29.0% (95% CI 22.9–35.5) reported in a meta-analysis including both rural and urban areas worldwide [12]. NCC proportion found among PWE in rural LAC (37.5%) appeared little lower than that reported in studies applying same criteria in rural Africa (Burkina Faso 46.9%) [46], rural India (40%) [47] and higher than that reported in rural Tanzania (17.9%) [48]. Considering urban areas, our estimate (29.4%) was similar to that reported in South Africa (28.0%) [49] and in urban India (28.4%) [50].
For the first time, to the best of our knowledge, this study demonstrated the influence of NCC prevalence on LTE and AE prevalence in LAC: countries with a NCC proportion (EITB assay) higher than 12.1%, among PWE living in urban areas, and than 23.7%, among PWE living in rural areas, presented higher LTE and AE estimates. We are aware that seropositivity at the EITB reflects an exposure to the parasite not necessary accompanied by a CNS involvement. However, this assay presents a high sensitivity and specificity and, together with the presence of compatible clinical manifestations (such as epilepsy) in people living in endemic areas, allows to formulate the diagnosis of probable NCC according the worldwide accepted diagnostic criteria for NCC [17]. We could then state that NCC prevalence seems to affect both LTE and AE prevalence, and that it could be considered a source of variability of prevalence estimates across LAC. The association between CC/NCC and epilepsy in LAC was evaluated in nine studies using “prevalent” cases. There was a consistent and significant association between epilepsy and CC/NCC, with ORs of 2.7 for studies using EITB serology and 5.6 for studies performing brain CT scan. Studies associating brain imaging to serology are likely more accurate as include those cases who present calcified or single parenchymal cyst which may be asymptomatic or seronegative [51]. Previous meta-analysis data from LAC did not exist, while in Africa a 3.4 to 3.8-fold increased risk for developing epilepsy was reported [52].
The meta-regression analysis has also pointed out an influence of TG on both AE and LTE estimates: since countries with higher TG are those presenting a larger burden of untreated epilepsy, the consequent increased number of active cases could lead to higher AE prevalence. Moreover, as AE prevalence is included in LTE measures (about the 38% for rural and 59% in urban population of developing countries; [3]) the role of TG could also be reflected on LTE estimates. On the other hand, in countries with lower TG, “not-active” cases are probably more frequent and more difficult to detect in prevalence surveys, leading to an underestimation of the LTE prevalence.
When considered together, TG and NCC prevalence together explained a very great amount of AE prevalence variability (up to 85.4%) suggesting that interventions on these modifiable factors could result in important reduction of AE burden.
Finally, our data showed a non significant increase in mortality in PWE in LAC (SMR 1.4; 95% CI 0.01–6.1). Studies in developed countries have reported mortality rates two to three times higher in PWE than in the general population [53]. This increase includes direct and indirect consequences of epilepsy, as well as underlying disorders responsible for secondary epilepsies [54].We found only five eligible papers reporting SMR in LAC (Argentina, Bolivia, Brazil, Chile and Ecuador) ranging from 0.76 in Brazil [44] to 6.3 in Ecuador [55]. These studies presented a prevalent cohort designs that might have underestimate short-term mortality, as patients with more serious disease die earlier and are not included in the observation period. Furthermore, SMR is highest in symptomatic epilepsy (ranging from 2.2–6.5, [53]) while, in idiopathic epilepsy, conflicting results have been reported, often showing a non significant increase, as the one found here. Only in the Bolivian study [56], SMR was estimated stratifying by symptomatic and idiopathic epilepsy, with increased mortality reported only among patients with symptomatic epilepsy (SMR = 3.0; 95% CI 1.2–6.3).
Concluding, this systematic review demonstrated a high burden of epilepsy and of NCC in LAC with marked detriment of rural areas, identified two important modifiable factors related to epilepsy prevalence and a consistent association between NCC and epilepsy in LAC. We are aware about the possible loss of power related to the dichotomization of continuous variables in the meta-regression. However, the paucity of studies found on NCC and on TG in particular, made our analysis not able to support a larger number of categories. Moreover, the reported narrow CIs suggest that the loss of statistical power was minimal. Additional data are needed to better understand the possible sources of heterogeneity among countries and determine the situation in those regions that are still under the shadows.
|
10.1371/journal.pcbi.1005253 | Competing Mechanistic Hypotheses of Acetaminophen-Induced Hepatotoxicity Challenged by Virtual Experiments | Acetaminophen-induced liver injury in mice is a model for drug-induced liver injury in humans. A precondition for improved strategies to disrupt and/or reverse the damage is a credible explanatory mechanism for how toxicity phenomena emerge and converge to cause hepatic necrosis. The Target Phenomenon in mice is that necrosis begins adjacent to the lobule’s central vein (CV) and progresses outward. An explanatory mechanism remains elusive. Evidence supports that location dependent differences in NAPQI (the reactive metabolite) formation within hepatic lobules (NAPQI zonation) are necessary and sufficient prerequisites to account for that phenomenon. We call that the NZ-mechanism hypothesis. Challenging that hypothesis in mice is infeasible because 1) influential variables cannot be controlled, and 2) it would require sequential intracellular measurements at different lobular locations within the same mouse. Virtual hepatocytes use independently configured periportal-to-CV gradients to exhibit lobule-location dependent behaviors. Employing NZ-mechanism achieved quantitative validation targets for acetaminophen clearance and metabolism but failed to achieve the Target Phenomenon. We posited that, in order to do so, at least one additional feature must exhibit zonation by decreasing in the CV direction. We instantiated and explored two alternatives: 1) a glutathione depletion threshold diminishes in the CV direction; and 2) ability to repair mitochondrial damage diminishes in the CV direction. Inclusion of one or the other feature into NZ-mechanism failed to achieve the Target Phenomenon. However, inclusion of both features enabled successfully achieving the Target Phenomenon. The merged mechanism provides a multilevel, multiscale causal explanation of key temporal features of acetaminophen hepatotoxicity in mice. We discovered that variants of the merged mechanism provide plausible quantitative explanations for the considerable variation in 24-hour necrosis scores among 37 genetically diverse mouse strains following a single toxic acetaminophen dose.
| Acetaminophen-induced liver injury in mice is a model for drug-induced liver injury in humans. Challenging an explanatory mechanism in mice is problematic because variables determining causes and effects cannot be controlled adequately. We circumvent that impediment by performing virtual experiments that challenge the prevailing scientific explanation for the characteristic spatiotemporal pattern of early acetaminophen-induced hepatic necrosis. Virtual mice utilize a biomimetic software liver. Results of virtual experiments provide compelling evidence that the prevailing explanation is insufficient. Without further studies in mice, we discovered a new, marginally more complex explanatory mechanism that met stringent tests of sufficiency. We argue that this virtual causal mechanism and the actual mechanism in mice are strongly analogous within and across multiple biological levels. Variants of the virtual mechanism provide possible explanations for the considerable variation in 24-hour necrosis scores among 37 genetically diverse mouse strains following a single toxic acetaminophen dose.
| Acetaminophen (APAP)-induced liver injury (AILI) in mice is the widely used model for drug-induced liver injury in humans. There has been dramatic progress in identifying involvement of a variety of molecular level events and pathways [1–3]. To use that knowledge effectively in predicting injury and developing new treatment strategies, we need more credible explanations of how, when, and where key injury features emerge within hepatic lobules, and why humans exhibit a wide diversity of responses. A characteristic early feature is that necrosis, proceeded by covalent adduct formation, begins perivenous, adjacent to the central vein (CV) of hepatic lobules, and then progresses (radially) outward [4, 5]. Given that the fraction of APAP converted to the reactive metabolite (NAPQI: N-acetyl-p-benzoquinone imine) within hepatocytes increases from lobule’s periportal (PP) space to CV (called zonation), the prevailing mechanistic explanations assume that zonation of NAPQI formation is a necessary and sufficient prerequisite to explain early perivenous necrosis. That NAPQI zonation hypothesis provides an explanatory foundation for current efforts to better understand downstream complexities leading to tissue regeneration or liver failure. However, challenging that explanation directly in mice is currently impracticable because it requires exercising precise control over zonation of many mechanism features along with methods to measure intralobular features sequentially within the same mouse. Those barriers have impeded progress.
Here, we successfully circumvent those impediments using multi-attribute software-based experiments. We challenge the NAPQI zonation explanation by experimenting on a strongly analogous software mechanism (as distinct from being described mathematically), called NZ-Mechanism, that is instantiated within a virtual mouse. This software analog is engineered to be quantitatively and qualitatively biomimetic across all anatomical, hepatic zonation, and cell biology features currently believed relevant to challenging the NAPQI zonation hypothesis. Results from virtual experiments provide strong quantitative evidence that zonation of NAPQI formation alone is insufficient to explain early perivenous necrosis. A parsimoniously more complex explanatory mechanism is needed, one involving additional feature zonation. After testing several, we focused on two equally plausible mechanism zonation features, but they too failed to meet stringent tests of sufficiency. However, when the two zonation features that had failed were combined into a unified mechanism, sufficiency was achieved, thus demonstrating the potential power of the virtual experiment approach.
Physiologically reasonable perturbations of the merged mechanism caused significant changes in both the numbers of hepatocytes killed and the intralobular patterns of dead hepatocytes for a fixed toxic APAP dose of 300 mg/kg. Injury diversity was similar to that reported for humans [6, 7] and to differences in 24-hour necrosis scores among 37 genetically different mouse strains [7]. For the latter, we hypothesized that variation of lobule location-dependent features within the combined mechanism, resulting from changing selected mechanism configurations, can be sufficient to account for such necrosis score diversity. We present results of virtual experiments supporting that operating hypothesis with the intent that these results will help formulate new wet-lab experiments to further improve explanatory mechanism insight (Fig 1).
Mouse Analog is the multiscale biomimetic software system (Fig 2) designed and built intentionally to be scientifically useful [8], especially, as in this work, for testing mechanism hypotheses of AILI, starting with the NAPQI zonation hypothesis: a configuration of NZ-Mechanism will be sufficient to explain early necrosis adjacent to CV. To first establish credibility that Mouse Analog can be used to challenge the NAPQI zonation hypothesis, we needed to achieve a variety of prespecified Target Attributes (performance requirements). A Target Attribute is a characteristic feature of APAP induced liver injury to which a prespecified Similarity Criterion is assigned. Each wet-lab measurement that we seek to mimic, such as hepatic extraction ratio, becomes a Target Attribute. The Similarity Criterion specifies the degree of similarity that must be achieved. An example is the mean virtual experiment measurement falling within ± 1 standard deviation of the mean wet-lab experiment measurement. The following are three of the Target Attributes achieved: measurements of APAP 1) hepatic extraction ratio, 2) intrinsic clearance, and 3) dose-dependent pharmacokinetic profiles (Supporting S1 Fig). By increasing the strength and variety of analogies between measurements made during experiments on Mouse Analog and corresponding measurements made during experiments on mice (Fig 3A), the credibility of explanatory inferences drawn from virtual experiments increases. Hereafter, Mouse Analog components and characteristics are capitalized to distinguish them from mouse counterparts; feature and parameter names are italicized.
For a particular Mechanism variant, Target Attributes are achieved by changing feature configuration values (Supporting S1 Table). Regardless of zonation, the probability of an event occurring within each Hepatocyte is fixed for the duration of the experiment. Probability of occurrence is the same within all Hepatocytes for events not subject to zonation. Note that to keep mechanisms parsimonious, some Fig 3B features, such as creation of Mitochondrial Damage (mitoD) objects, do not have direct mouse counterparts but instead subsume many.
For NZ-Mechanism, the zonation of APAP Metabolism events (Fig 4A) causes the largest amount of NAPQI to be generated adjacent to CV. As depicted in Fig 3B, at each time step, each bound APAP may be Metabolized (maps to rate of metabolism); the probability of a Metabolic event is Lobule-location dependent, increasing from 0.35 to 0.95 PP to CV, which maps to increasing expression of participating metabolic enzymes. The probability that the Metabolite is NAPQI increases from 0.33 to 0.9 PP to CV (otherwise, G&S Metabolites are produced). That increasing probability maps to increasing expression of enzymes responsible for NAPQI formation, primarily CYP2E [1–4].
Glutathione (GSH) Depletion Threshold and probability of a mitoD Mitigation event are independent of Lobule location for NZ-Mechanism. Changing one or both alters total Hepatocyte Death but does not alter their Lobular locations. However, making one or both location-dependent (Fig 4B) changes the causal cascade and consequently alters Hepatocyte Death locations.
Because hepatic lobules approximate regular polyhedra with CV located centrally, the percent of a given lobule’s hepatocytes within different regions varies dramatically. The same is true within Mouse Analog Lobules (Fig 4C). Note that Lobule does not have a direct mouse counterpart (Methods). Rather, it maps to a small random sample of all hepatic periportal to perivenous flow paths. Because of periportal interconnections among sinusoids, the number of hepatocytes encountered by a compound such as sucrose moving through a lobule is typically greater than the number of hepatocytes along straight-through sinusoids. Flow through a Lobule enables APAP and other mobile objects to mimic that phenomenon.
The temporal patterns for average location of Death events and Death trigger events differ only by the value of Death Delay intervals, which range from 1.2 to 12 hours (see Death Delay subsection in Methods). We explain in the Death Delay subsection that measuring Death events while keeping measurement of events separate from the Mouse Analog mechanism is computationally expensive. Thus, we elected to use measurements of Death trigger events for Fig 5 to challenge the NAPQI zonation hypothesis. Even though the highest probability of NAPQI formation occurs adjacent to CV, NZ-Mechanism is falsified because early average trigger events failed to fall within the Zone 3 target range close to CV (Fig 5). Falsification is discussed further in the Validation subsection under Methods. Prior to about 6 minutes, the location of early average trigger events shifts toward CV, but at about 10 minutes the trend shifts in the PP direction, away from CV. Such an unusual pattern suggests that NZ-Mechanism may not be biomimetic. Seeking evidence that challenges that prediction would be problematic in mice at such early times. Nevertheless, what is the explanation for this Mouse Analog phenomenon?
Because of event stochasticity, some Hepatocytes at similar distances from the PP space will experience trigger events before their neighbors. Consider two regions: RCV (nearer to CV) and RPP (near PP space). The fraction of Hepatocytes experiencing an early trigger event will be larger in RCV. However, the total number of Hepatocytes in RPP is larger (Fig 4C). Consequently, the number of early trigger events in RCV and RPP can be similar, sometimes even larger in RPP. Thus, the average location of all early trigger events (circled trend 1, Fig 5) can appear skewed toward RPP. Later, remaining RPP Hepatocytes will be less vulnerable as the amount of APAP decreases, which causes the average trigger event location to shift toward CV. Concurrently, the fraction of Hepatocytes that have already experienced trigger events increased faster in RCV than in RPP, explaining the direction change around 10 minutes.
The hypothesis that NZ-Mechanism alone is sufficient to explain early periportal necrosis is clearly falsified. We found no configurations that achieved that Target Phenomenon and shifted Hepatocyte Death patterns toward the CV (Fig 5). The results of these virtual experiments provide clear, strong evidence that an alternative, somewhat more complicated explanation is needed.
Having falsified NZ-Mechanism, we posited that at least one additional feature of the Mechanism must exhibit zonation. The following two (Fig 4B) seemed equally plausible. 1) GNZ-Mechanism specifies that GSH Depletion Threshold values decrease PP to CV and NAPQI formation values increase PP to CV. 2) MNZ-Mechanism specifies that each Hepatocyte’s ability to mitigate mitoD diminishes sigmoidally PP to CV and NAPQI formation increases PP to CV. Predecessors of MNZ-Mechanism employing a parsimonious linear mitoD Mitigation gradient failed to shift trigger events sufficiently in the CV direction. Using a sigmoidal distribution improved similarity to the Target Phenomenon, but not sufficiently. The results require that both MNZ- and GNZ-Mechanisms be falsified. We inferred that combining the two Mechanisms into MGNZ-Mechanism might be sufficient to achieve the Target Phenomenon, and identified values of MGNZ-Mechanism configurations that did so. Having achieved the Target Phenomenon, MGNZ-Mechanism stands as a plausible explanation of the target APAP-induced necrosis pattern.
Contents of each Hepatocyte were measured within three 5-grid-space-wide bands (Fig 6A) during execution of MGNZ-Mechanism. The APAP pharmacokinetic profile (Fig 6B) is sufficiently similar to a plasma profile in mice [16] (Supporting S1 Fig). To mimic hepatic blood flow, a particular Mouse Analog feature configuration determines the fraction of APAP in Mouse Body that is transferred to Liver each time step. G&S are transported out of Hepatocytes and allowed to accumulate in Mouse Body. NAPQI peaking first in CV region (Fig 6C) is a consequence of differences in APAP exposure, feature zonation, and the fact that Hepatocytes that have experienced a Death trigger event stop Metabolizing APAP. Mean amounts of G&S per Hepatocyte (Fig 6E) also reflect NAPQI patterns. G&S amounts in CV region prior to 20 minutes are greater than PP region amounts, which may seem counter-intuitive. Zonation of Metabolism is one explanatory factor, but zonal differences in Hepatocyte numbers (Fig 4C) are more important: 742 in Zone 3, 3,948 in Zone 2, and 9,310 in Zone 1.
Prior to 80 minutes, Death has been triggered and APAP metabolism has been halted in most Hepatocytes in CV region but not Midzonal region. Thus, the value of cumulative Midzonal GSH Depletion events continues to increase after 80 minutes (Fig 6H). Having early mitoD values in CV region that are much larger (10x) than those in Midzonal region (Fig 6I) resulted in Hepatocyte Death occurring first near CV. The probability of a mitoD Mitigation event is smallest adjacent to CV, yet that is where the number of early Mitigation events is largest (Fig 6J). Therefore, although the probability of any one Mitigation Event is small, the cumulative number of such events is large.
The space of MGNZ-Mechanism configurations contains both biomimetic and non-biomimetic variants. A variant that, for example, eliminates zonation of APAP metabolism (Fig 4A) is non- biomimetic. To test sensitivity and robustness of the MGNZ-Mechanism configuration, we conducted experiments using 64 plausibly biomimetic Mouse Analog variants of the MGNZ-Mechanism configuration in which values assigned to one-to-seven of ten influential configurations (shown in Supporting S2 Table) were changed. Each variant represents an arbitrary virtual mouse strain. Most changes included adjusting one or more zonation configuration. The Mechanism configurations included the five in Fig 4A and 4B plus the following: the probability that a NAPQI object will react, Death Delay value, the probability that a NAPQI reaction product will be mitoD or nonMD (non-mitochondrial damage products, discussed below), the probability of a nonMD Mitigation event, and/or Death Trigger Threshold value. Total Dead Hepatocytes 24 hours after dosing for all 64 variants were ordered in ascending order and plotted (Fig 7A). Comparing Deaths per zone rather than total Deaths (inserts, Fig 7A) provides concrete evidence of differences in how events within each Mechanism variant unfolded. Supporting S2 Fig provides additional examples.
We compared those results to data from Harrill et al. [7] (red bars, Fig 7B) in which mean necrosis scores of 37 different mouse strains were recorded 24 hour after dosing with 300 mg/kg of APAP administered intragastrically. We assumed a direct correlation between necrosis score and both numbers of necrotic cells (in mice) as well as total Dead Hepatocytes (Fig 7B, right axis). All Mouse Analog data in Fig 7A were scaled the same so that at least one (of the 64) Mouse Analog bar heights was similar to the smallest and largest values in Fig 7B. We arbitrarily matched the 24-hour necrosis score for strain FVB/J to total Dead Hepatocytes to Mechanism variant #37, and then selected the one variant that provided the closest match to the necrosis score for each of the other mouse strains (white bar).
Our intent is that parallel virtual and wet-lab experimentation will integrate into a faster and more effective scientific approach [17] (Fig 1). Most researchers seem to prefer including ample realistic detail in explanatory models. However, mechanisms like those employed in Mouse Analogs are scientifically preferable and more informative because they are only as detailed as necessary for explaining a particular phenomenon or challenging a particular hypothesis. A validated model Mechanism with limited detail subsumes many different yet equally plausible more detailed variants. When we falsify such a Mechanism (show that it is inadequate for its Target Attributes), we significantly shrink the space of equally plausible variants by eliminating whole classes of more detailed variants.
The results in Fig 5 document the first use of virtual experiments to challenge competing biomimetic Mechanism hypotheses. The fact that three were falsified demonstrates how virtual experiments can shrink the space of plausible explanatory Mechanisms and facilitate the more fundamental work of wet-lab experimentation.
When we started, we expected Mouse Analogs using NZ-Mechanism to show Hepatocyte Deaths occurring first near CV. We acquired new knowledge by examining why it, as well as MNZ- and GNZ-Mechanisms, failed to mimic that phenomenon. Explanatory insight was further improved by examining how MGNZ-Mechanism successfully achieved the Target Phenomenon. An example is the importance of zonated repair pathways in mitigating damage (Fig 6). We anticipate that other comparably parsimonious yet somewhat different Mouse Analog Mechanisms can do the same.
The Iterative Refinement Protocol (described under Materials and Methods) outlines how to identify and challenge comparably parsimonious explanations of target phenomena. Mouse Analog concreteness, coupled with the ability to observe and measure propagating causal events, makes it easy for experts and non-experts to form and discuss opinions about the credibility of similarities between virtual experiment results and those from real or envisioned wet-lab counterparts.
NZ-Mechanism was falsified because a small fraction of Hepatocytes in Zones 1 and 2 experienced an early Death trigger event (Fig 5) and consequently will experience an early Death event. The same phenomenon may occur in mice, but it has escaped detection because experiments were focused elsewhere. If deemed important, that phenomenon—that prediction—could be challenged directly using wet-lab experiments.
During execution, MGNZ-Mechanism is a concrete hypothesis for how key features of APAP hepatotoxicity in mice are generated (Fig 3A). Suppose that strong analogies do exist between virtual and real mice. It is clear from concurrent differences within the three regions (Fig 6C–6H) that explanatory inferences drawn from whole liver measurements or hepatic biopsies will at best be flawed and likely misleading.
Differences in zonal predictions can guide design of wet-lab experiments intended to challenge and possibly falsify the MGNZ-Mechanism hypothesis. For example, 1) at 30 min post-dose, GSH Depletion in the Midzonal region is significantly greater than in either CV or PP regions; and 2) at 10 min post-dose, counts of mitoD objects (maps to mitochondrial damage) in CV Hepatocytes are an order of magnitude greater than in PP Hepatocytes. Such experiments will be win-win: no matter the outcome, we will have useful new knowledge, and the new results are expected to motivate additional rounds of virtual experiments [17] (Fig 1). Note that, absent the data in Fig 6 there would be no basis to undertake such narrowly focused experiments.
Within individuals, expressions of hepatic proteins are known to change significantly in response to a variety of environmental and health factors. We simulated possible examples (Fig 7A). Genetic differences among mouse strains that influence hepatic zonation more than they influence particular pathways or molecular level events may account for the diversity of differences in necrosis scores (Fig 7B). However, considerable uncertainty remains. As illustrated by the green bars in Fig 7A, for each Mouse strain explanation, there will exist a sizable set of Mechanism variants that would be equally explanatory. Further, MGNZ-Mechanism is just one realization drawn from a plausible yet finite Mechanism space. Automated methods, as they become available, can be used to more systematically explore that space, prior to or in parallel with new wet-lab experiments to challenge MGNZ-Mechanism.
The data in Fig 7 demonstrate that, when targeting just one attribute, such as total Hepatocyte Death 24 hours post-dose, modest configuration changes for two or more Mouse Analog features can have counterbalancing consequences resulting in altered event details yet essentially no significant change in the measured feature. On the other hand, modest changes can have synergistic consequences resulting in a 3-fold change or more. Toxicity enhancing synergistic changes in cascade zonation, possibly initiated by changes in environmental and health factors, may be a contributor to idiosyncratic drug induced liver injury. In combination with the adaptation hypothesis [18], such change may help explain how and why only a small percentage of susceptible individuals develop overt liver injury.
Methods stem from four requirements [19], which are needed to implement the use cases illustrated in Fig 3B. Meeting them is outside the scope of systems biology models currently in use to predict APAP hepatotoxicity [20, 21]. We use the virtual experiment approach described by Kirschner et al. [17] along with the enhanced strategies detailed by Petersen et al. [22, 23]. Feature names (Supporting S1 Table) are italicized.
To achieve Requirement 2, Mouse Analogs are written in Java, utilizing the MASON multi-agent simulation toolkit [28]. In silico experiments are run using virtual machines [29] on Google Compute Engine, running 64-bit Debian 7. For longer simulations (e.g., Fig 6) Monte Carlo trials are run in parallel. The data presented herein along with Mouse Analog code are available [30].
We started with a previously validated version of Liver [9]. The goal of an Iterative Refinement Protocol cycle is to refine a formulated explanatory Mechanism by achieving Target Attributes, thus completing one right-side cycle in Fig 1. A concrete software mechanism can be falsified—shown inadequate for its Target Attributes—in two ways: 1) it cannot exhibit a Target Attribute and/or 2) we fail to discover a Mouse Analog configuration that achieves all Target Attributes.
Before starting, specify Target Attributes to be explained and rank-order them in terms of expected difficulty to achieve. For each Target Attribute, specify initial and possible subsequent Similarity Criteria.
Liver composition (Fig 2) is now stable and robust, having already achieved several Target Attributes [9–12, 31]. Bile Network does not influence the four Mechanisms. Because rat and mouse lobule structure and organization are similar, Liver can be used in simulations of phenomena measured in rats and mice. By altering analog-to-referent mapping functions [25], we can change experimental context. A Lobule maps to a small random sample of all lobular flow paths. The 3-zone network has 68 nodes, 45, 20, and 3 in Zones 1, 2, and 3, respectively. Nodes are connected using 99 edges: 55 from Zone 1-to-Zone 2; 10 within Zone 1; 24 from Zone 2-to-Zone 3; 10 within Zone 2; and 0 within Zone 3. To represent both uncertainty and variability, edges between Sinusoid Segments are randomly assigned at the start of each Monte Carlo execution. Having edges and Sinusoid Segment sizes assigned randomly at the start of each experiment simulates lobular variability within and between livers. See Validation subsection below for additional detail.
Hepatocyte objects, excluding Fig 3B capabilities, are identical to those described in Petersen et al. [13, 22]. Hepatocytes contain four physiomimetic modules [13]: InductionHandler, EliminationHandler, MetabolismHandler, and BindingHandler. The events and features illustrated in Fig 3B were added incrementally.
Liver sinusoidal endothelial cells have been implicated in AILI. They contain CYP2E1 and GSH, although the relative amounts are very small compared to hepatocytes. An unanswered question is whether lobular differences in CYP2E1 and GSH within liver sinusoidal endothelial cells contribute to zonation of injury [32]. Addressing that question is feasible using Mouse Analog, given Target Attributes specific to liver sinusoidal endothelial cells, but it is outside the scope of this work. Adhering to a strong parsimony guideline, and given that the vast majority APAP metabolism occurs within hepatocytes, we specified that all relevant APAP metabolism and GSH Depletion occur within Hepatocytes. Within Mouse Analog Endothelial Cells, we specified that only nonspecific APAP binding occur. It is straightforward to add modules to Endothelial Cells and configure them appropriately should evidence become available that falsifies those specifications.
The first performance goal was to discover Metabolism phase (Fig 3B) feature configurations that, upon achieving Target Attributes, would remain unchanged during experiments to challenge Mechanisms intended to explain necrosis patterns. There is quantitative variability (within and between experiments) in zonal measurements of CYP2E1 (primarily responsible for formation of NAPQI) [33], the relative proportion of the three primary metabolites (glucuronide, sulfate, and NAPQI) [34], hepatic clearance, and hepatic extraction ratio. However, there is qualitative agreement on their trends.
An APAP object maps to a tiny fraction of the 300 mg/kg dose. There is a direct mapping between the probability of an unbound APAP object being metabolized each time step (1 second) and amounts of metabolic enzymes [13, 22]. In mice, CYP2E1 levels per hepatocyte increase PP to CV by 2 to >10 fold. We specified this Target Attribute: probability of an APAP metabolic event generating NAPQI increases at least three-fold PP to CV. All other metabolites are lumped together and called G&S (maps mostly to the glucuronide and sulfate metabolites). In most reports, inactive metabolites are estimated to account for up to 85% of a dose, with NAPQI accounting for the balance. We specified that total NAPQI, as fraction of dose, be within 0.15–0.4; a large range was needed to ensure sufficient numbers of NAPQI are generated when dose is reduced. A Target Attribute is that toxicity be reduced by about 50% when the APAP dose is reduced by 50%.
This Target Attribute was the most demanding: APAP hepatic extraction ratio = 0.6 ± 0.06. To simplify achieving that Target Attribute, we used a virtual single pass Liver perfusion protocol with a constant rate of APAP input. When APAP outflow stabilized within 34–46% of input values, the Target Attribute was reached. For convenience, we began the initial Iterative Refinement Protocol cycle by specifying that APAP Hepatic Clearance be similar to that of prazosin, for which the Liver had already been validated [10–12]. We cycled through the Iterative Refinement Protocol seeking changes to configurations that would enable achieving additional Target Attributes specified below under Mouse Body. The resulting configuration values are illustrated in Fig 4A and 4B.
Lobules have a biomimetic PP to CV gradient [13] (Supporting S4 Fig), which Hepatocytes can use to create feature zonation. It maps to measures of one or more common blood attributes, such as pO2 [35]. However, because we needed the flexibility to explore multiple plausible, feature-specific PP to CV gradients for several different events, those in Fig 4A and 4B were implemented explicitly as functions of path length from analog Lobule PP entrance to the current Hepatocyte’s position. Path length (distance from PP entrance in grid spaces, dPP) is specified by this Hepatocyte’s distance from the inlet of its Sinusoid Segment (SS) plus the averages of the lengths of all Sinusoidal Segments in each of the previous zones. I.e., dPP(SSz) = ∑i=z1⟨|SSji−1|⟩j, where SSz is a given Sinusoid Segment in Zone z and |SS| is its length and j indicates each SS in zone i; j indicates each SS in zone i; and dPP(Hss) = xss + dPP(SSz), where Hss is a Hepatocyte in a given Sinusoid Segment, and xss < |SS| is a position along the length of a Sinusoid Segment. The angle brackets indicate the average of its contents. Each gradient has a shape (linear or sigmoidal) and its own start and end values, which are then scaled over the common path length computation. Gradients are turned on or off independently by specifying their start and end values. Once these feature-specific gradients have stabilized, they can be replaced by rules that use the common biomimetic gradient.
We started with a simple parsimonious operating hypothesis: when specific damage products accumulate within a Liver above some threshold, Death will be triggered irreversibly. The evidence shows that accumulation of APAP-induced damage products in mice occurs after available GSH has been sufficiently depleted. By analogy, we implemented a new feature: at each time step, with p = 0.5 (specified arbitrarily because we had no wet-lab data to guide doing otherwise), each NAPQI object may be destroyed, which maps to in vivo depletion of a fraction of hepatocyte’s available GSH. A NAPQI destruction event is also a GSH depletion event. A small GSH Depletion Threshold value maps to mice that are very sensitive to APAP hepatotoxicity; a larger GSH Depletion Threshold maps to increased resistance. The Threshold value needed to be small enough to allow sufficient accumulation of Damage products (below), but large enough to achieve this Target Attribute: toxicity is reduced by about 50% when the APAP dose is reduced by 50%. Setting GSH Depletion Threshold = 5 for NZ-Mechanisms proved adequate. In wet-lab studies, GSH is often measured in whole liver homogenates. To make NZ- and MNZ-Mechanisms more directly comparable to GNZ- and MGNZ-Mechanisms at the level of whole Lobule measurements, we specified that Depletion Threshold = 3.5 for NZ- and MNZ-Mechanisms, which is the average value for all Hepatocytes using GNZ- and MGNZ-Mechanisms. For some cases in Fig 7A, GSH Depletion Threshold values at the Lobule’s entrance of 8 or 3 are used (Supporting S2 Table).
Each time step after Depletion Threshold is reached, with probability to react = 0.5, each NAPQI object may be destroyed and replaced by a Damage product. Evidence implicates necrosis being triggered by mitochondrial damage. We parsimoniously specified that there be two classes of Damage product: “mitochondrial damage products,” called mitoD, and “non-mitochondrial damage products,” called nonMD. When a NAPQI object is destroyed, it is replaced with either a nonMD or mitoD object selected randomly (probability = 0.5); we had no wet-lab data to guide specifying differently.
To trigger Death, we implemented an analog counterpart to this simple yet widely accepted mechanism: upon accumulation of sufficient mitochondrial damage, a tipping point is reached and necrosis is triggered irreversibly. Triggering necrosis within one mouse hepatocyte requires reaction of possibly hundreds (or more) of NAPQI molecules, which is a very tiny fraction, f, of the administered APAP. One analog NAPQI object maps to a very small fraction, fA, of that same APAP dose. For mice resistant to APAP hepatotoxicity (case 1), it is possible that f > fA. When referent mice are much more sensitive to APAP hepatotoxicity (case 2), it is likely that fA > f. For case 2, replacing one NAPQI with one mitoD will be more than enough to trigger a Death event. Thus, the resolution [17] of one NAPQI → one mitoD is inadequate to simulate toxicity phase events. We solved that problem by specifying that each mitoD be amplified: one NAPQI → (1 + n) mitoD. Specifying that n be a random draw from either a uniform (1, 6) distribution enabled achieving Target Attributes. The mitoD amplification step is an example of making an analog Mechanism locally finer grain (increasing resolution) without changing granularity elsewhere.
As with Depletion Threshold, the Death Trigger Threshold value needed to be small enough to allow sufficient accumulation of mitoD without a trigger event, but also large enough so that toxicity is reduced by about 50% when APAP dose is reduced by 50%. Setting Death Trigger Threshold = 6 proved adequate. For some cases in Fig 7A, larger Death Trigger Threshold values are used.
Hepatocytes utilize multiple mechanisms to mitigate or reverse different types of damage, ranging from up-regulation of GSH synthesis to mitochondrial autophagy [36]. Consistent with our strong parsimony guideline, we started with a single mitigation Mechanism, which maps to a conflation of all actual mitigation/recovery mechanisms. Each time step, with p = 0.5 (specified arbitrarily), each nonMD and mitoD object may be destroyed. For NZ-Mechanism, lacking wet-lab data for guidance, we made probability-of-Mitigation event for each nonMD and mitoD at each time step independent of Lobule location.
Because hepatocytes closer to CV are known to experience increasing oxidative stress, we inferred that mitigation mechanisms would be more robust closer to CV. However, when we increased probability-of-mitoD-Mitigation event (simply p-mitoD-Mitigation below) PP to CV, the locations of Dead Hepatocytes were shifted further in the PP direction, relative to the NZ-Mechanism data in Fig 5. To shift Death trigger events in the CV direction, it was necessary to decrease p-mitoD-Mitigation PP to CV. Instantiating a linear 0.9-to-0 decrease in p-mitoD-Mitigation failed to produce a sufficient shift. We considered several options for moving mean trigger events toward CV. E.g., keep p-mitoD-Mitigation constant at 0.9 through Zone 2 and then decrease it linearly to 0 at CV. We rejected that feature because, absent any supporting observation, the risk seemed too great that it would be non-biomimetic. We considered adding a new biomimetic feature or event, but doing so conflicted with our strong parsimony guideline. Following several exploratory Iterative Refinement Protocol cycles, we opted for specifying that p-mitoD-Mitigation for MNZ- and MGNZ-Mechanisms decreases PP to CV following a reverse sigmoid with the inflection centered approximately midway in Zone 2. However, there is neither supportive nor refuting evidence for that configuration; therefore, this explanatory Mechanism hypothesis needs to be challenged. We could posit that the efficiency of one or more processes for mitigating normal mitochondrial dysfunction decreases PP to CV [36, 37] and that one or more of the processes that maintain necrosis-triggering pathways is preferentially sensitive to NAPQI damage. Consequently, their combined effect maps to p-mitoD-Mitigation decreasing sigmoidally PP to CV in Mouse Analog.
As with the GSH Depletion Threshold, to make NZ- and GNZ-Mechanisms directly comparable to MNZ- and MGNZ-Mechanisms at the whole Lobule level, we determined that the average p value for all Hepatocytes using is 0.67. Therefore, we specified that p-mitoD-Mitigation be 0.67 for NZ- and MNZ-Mechanisms in Fig 5 rather than 0.5.
Connecting Liver to Mouse Body produces Mouse Analog. Mouse Body contains a space that maps to all extrahepatic tissues including blood along with a space to contain dose, which enables simulating intravenous, intraperitoneal, and intragastric dosing. When additional details are required, new objects can be added without influencing preexisting Liver Mechanisms. To achieve APAP pharmacokinetic attributes, we adjusted values of two configuration features: one controls Absorption into Mouse Body and the other controls metering of APAP from Mouse Body to the Lobule’s entrance. The latter maps to hepatic blood flow. We specified that Absorption be first order and rapid.
We prespecified two Target Attributes: 1) mimic single dose APAP pharmacokinetics in mice [16] and 2) observe characteristic dose-dependent pharmacokinetics for APAP in Mouse Body. For (1), we prespecified that Mouse Body measurements be within one standard deviation of mean wet-lab values. Supporting S1A Fig shows that the first Target Attribute was achieved. For (2), we found no fine grain dose-dependent pharmacokinetic data in mice to use as direct validation targets. However, there is pharmacokinetic data in rats [38, 39] demonstrating characteristic dose-dependent pharmacokinetics. We reasoned that, by using a simple mouse-to-rat scaling, reported dose-dependent pharmacokinetic data in rats [38, 39] could serve as alternative targets, despite the fact that the metabolism and toxicity of APAP in mice and rats are quite different (we are targeting dose-dependent APAP clearance, not metabolite types). For comparable mg/kg doses, the plasma half-life of APAP in rats is approximately 3x-5x that in mice [40]. The dose of APAP objects used in Fig 6B maps to 300 mg/kg in mice. Using the median value of 4x, we determined that dose should scale to approximately 75 mg/kg in rats. Twice that dose of APAP objects should map to approximately 150 mg/kg, and half that dose should map to approximately 37.5 mg/kg in rats. Supporting S1B and S1C Fig show that characteristic dose-dependent pharmacokinetics is observed. Attaining Target Attributes from different species increases credibility and documents achieving the second capability under Requirement 2.
There are qualitative but not quantitative time-course data for relative amounts of necrosis in mice. Necrosis is not detectable during the first hour following a toxic but non-lethal APAP dose of 300 mg/kg, and there is no evidence of significant further induction of necrosis 12 hours after dosing [3]. Covalent adduct formation and histological evidence of hepatocyte damage always precedes measurable necrosis by tens of minutes to a few hours. That time interval maps to Death Delay. To specify the Death Delay feature, we needed a Target Attribute estimate for percent total necrosis as a function of time post-dose. Target range estimates (see Supporting S5 Fig) were provided by coauthor Kaplowitz for percent total necrosis at several times post-dose. The prespecified Similarity Criterion was that cumulative mean Death events fall within the estimated Target Attribute range. That Target Attribute was achieved in two different ways (Supporting S5 Fig): draw each Hepatocyte specific Death Delay value from a uniform (a pseudo-random draw from uniform [1.2, 12] hours) or a normal distribution. Although the second seemed more realistic, we used the first because it is more parsimonious and doing so enabled achieving the Target Attribute range.
Necrosis is the target wet-lab phenomenon. Our working hypothesis is that necrosis is a process that is triggered by early damage and that the wet-lab measurement is detecting a late stage in that process. Death Delay maps to the process of necrosis. The time of an Hepatocyte Death event = time of that Hepatocyte’s Death trigger event + a pseudo-random draw from [1.2, 12]. In Fig 5, we elected to focus on Death trigger events rather than time of an Hepatocyte Death events for three reasons. 1) Although there is no current wet-lab method to measure its short transient duration, the trigger event is a key causal event. 2) We know that final Lobular locations of trigger events and Dead Hepatocytes are the same. 3) To make the virtual experiment as much like a wet-lab counterpart as possible, we strive to measure the analog in the same way one measures the referent. Conventional models are not measured; rather, they simply output variables (e.g. concentrations of species) each time step. In contrast, Mouse Analog is measured. Doing so during executions requires keeping Death event measurements separate from Analog Mouse Mechanisms. Measuring Death events requires a high frequency polling effort, where during one poll the observer agent records a Hepatocyte in a particular location as not Dead, and the next poll records that Hepatocyte as Dead. Such polling increases considerably the time required to complete each simulation experiment.
Mouse Analogs are treated as a form of data, using both the implicit schema of Java, JavaScript, and R and the explicit schema of the configurations. Mouse Analogs and configuration data are maintained, archived, and released using the Subversion version control tool in two repositories, one private (Assembla) for rapid and prototyping development with project partners and another public for collaboration [40, 41]. Input-output (I/O) data is handled separately. Smaller data sets are stored in simple CSV. I/O data is tightly coupled to experiments and configuration details, requiring a common versioning system, aggregating all three data types. Versioned I/O data is archived as downloadable packages.
The entire toolchain, including the operating system, used for Mouse Analogs, configurations, and I/O handling is open-source, thereby ensuring repeatability. Similarly, all project generated and released data is available to be licensed as open data. Mouse Analogs are built for, maintained, and executed in a cloud environment (e.g., Google Compute Engine) to ensure platform and infrastructure repeatability across experiments, project team members, partners, and the wider community.
Regression and unit tests are special cases of canonical use cases (e.g., a single pass Liver perfusion experiment). While use cases are instances of the class of experiments to which Mouse Analogs are being applied (e.g., AILI), they also provide the measures by which the software and methods are maintained. Each toolchain iteration can execute the canonical use cases so that current results can be compared to prior results for which some degree of credibility has been documented. Significant variations are documented, investigated, and explained. While the majority of variations are the result of model iterations, the process does catch artifacts introduced by changes.
Unexpected variations in results during an Iterative Refinement Protocol cycle first trigger a software verification process designed to test the toolchain from the most general and widely used, up to the most specific tool: machine, OS, compiler, libraries, simulator, and model. Absent unexpected variations, that same process is applied at least yearly. So doing is rigorous and avoids wasteful iterations that result from assuming that all variations are caused by the most specific layer. Should anomalies occur at the Mouse Analog layer, they would trigger a source, data, configuration setting, and trace review in comparison with the design of experiments. Any unexpected variations that survive the verification process are then submitted to an Iterative Refinement Protocol based falsification battery designed to invalidate the Mouse Analog as compared to wet-lab observations. Results satisfying canonical use cases, but yielding unexpected variation as a result of changes to Mouse Analog or toolchain, provide material that can be used to formulate a useful hypothesis.
|
10.1371/journal.pcbi.1004052 | Correlations and Functional Connections in a Population of Grid Cells | We study the statistics of spike trains of simultaneously recorded grid cells in freely behaving rats. We evaluate pairwise correlations between these cells and, using a maximum entropy kinetic pairwise model (kinetic Ising model), study their functional connectivity. Even when we account for the covariations in firing rates due to overlapping fields, both the pairwise correlations and functional connections decay as a function of the shortest distance between the vertices of the spatial firing pattern of pairs of grid cells, i.e. their phase difference. They take positive values between cells with nearby phases and approach zero or negative values for larger phase differences. We find similar results also when, in addition to correlations due to overlapping fields, we account for correlations due to theta oscillations and head directional inputs. The inferred connections between neurons in the same module and those from different modules can be both negative and positive, with a mean close to zero, but with the strongest inferred connections found between cells of the same module. Taken together, our results suggest that grid cells in the same module do indeed form a local network of interconnected neurons with a functional connectivity that supports a role for attractor dynamics in the generation of grid pattern.
| The way mammals navigate in space is hypothesized to depend on neural structures in the temporal lobe including the hippocampus and medial entorhinal cortex (MEC). In particular, grid cells, neurons whose firing is mostly restricted to regions of space that form a hexagonal pattern, are believed to be an important part of this circuitry. Despite several years of work, not much is known about the correlated activity of neurons in the MEC and how grid cells are functionally coupled to each other. Here, we have taken a statistical approach to these questions and studied pairwise correlations and functional connections between simultaneously recorded grid cells. Through careful statistical analysis, we demonstrate that grid cells with nearby firing vertices tend to have positive effects on eliciting responses in each other, while those further apart tend to have inhibitory or no effects. Cells that respond similarly to manipulations of the environment are considered to belong to the same module. Cells belonging to a module have stronger interactions with each other than those in different modules. These results are consistent with and shed light on the population-based mechanisms suggested by models for the generation of grid cell firing.
| Grid cells are neurons in the medial entorhinal cortex (MEC), one synapse away from the hippocampus, that show a strikingly regular spatial selectivity [1]. Each grid cell has several firing fields that spread out in a hexagonal pattern, tessellating the environment in which the animal navigates. The locations of these firing fields are unaffected by the velocity of the animal, and they persist in the absence of external landmarks, suggesting that they make up an intrinsic metric for space [1–3]. These cells were first discovered in rodents [1, 2], but have recently also been reported in bats [4], monkeys [5], and humans [6], supporting the possibility that grid cells form a part of the neural circuitry underlying the brain’s internal representation of space in all mammals.
Two main properties of grid cells are their spacing (the shortest distance between two firing fields) and their orientation relative to an axis of the environment. Anatomically close grid cells tend to have the same orientation and spacing, with spacing increasing along the dorsoventral axis of MEC [1, 3]. This increase is stepwise rather than continuous, such that grid cells can be clustered with respect to spacing. These clusters also share other properties, such as orientation, and are therefore referred to as modules [7]. A third property of grid cells is their spatial phase, which is defined as the location of the grid pattern relative to a reference point in the environment. For cells with similar grid pattern, i.e. cells from the same module, one can also measure the difference in spatial phase by calculating the shortest distance between firing fields of two cells. No apparent relationship between the anatomical distance and the difference in spatial phase of pairs of neurons has been observed [1].
Since their discovery, grid cells have been under intense investigation, with studies ranging from experimental work to theoretical models, in hopes of revealing the underlying network mechanisms behind their coding; see [8, 9] for recent reviews. In particular, population-wise response properties [1, 7, 10] support the idea that the formation of grid cells is predominantly a network phenomenon, and that recurrent connectivity in MEC plays an important role. The main network model of grid cells, the continuous attractor model, would suggest that the hexagonal firing of grid cells emerges due to specific connectivity patterns between the neurons. In several of these models neurons are considered to be arranged in a two-dimensional network according to their phase. Cell pairs beyond a certain phase distance inhibit each other, while those closer to each other are coupled by excitation [11–13], or less inhibition [13, 14], as idealized by a ‘Mexican hat’ type of connectivity.
Although connectivity plays important roles in network models of grid cells and in shaping neuronal correlations, little has been done to study the correlation structure and functional connectivity in the MEC in vivo, as well as how they change with properties of grid cells, e.g. phase separation and theta modulations. In other words, statistical analyses of multi-neuronal spike trains of the type routinely performed on data recorded from other parts of the nervous system [15–17], is still lacking. Such analyses can shed light on how grid cells encode information at the population level and how they interact with each other, providing substance for understanding the network mechanisms behind the formation of grid cells.
In this paper we aimed at studying the statistical properties of grid cells’ multi-neuronal spike trains by analyzing recordings from two rats while they foraged freely in two-dimensional environments. We therefore first measured the correlations between these cells, beyond what is expected from space dependent rate variations, using the same approach as [18]: we averaged the Pearson correlation coefficients between firing rates of pairs of neurons during multiple passes through spatial bins covering the environment. With spatial bins small enough the effect of possible correlations due to rate covariations between two cells is removed. These correlations are referred to as noise correlations. We found that these correlations decay as the phase difference between cell pairs increases. This is consistent with previous analyses of pairs of grid cells recorded on a linear track [18]. Second, we fit a statistical model that assumes a pairwise maximum entropy distribution over the spikes generated in a time bin, given the spike pattern in the previous time bin and external covariates also referred to in the text as external fields. This model is known in the statistical physics community as the kinetic Ising model and belongs to the class of generalized linear models (GLMs) [19] with short time memory kernels. We considered an extensive list of external covariates known to modulate the firing of grid cells to explain the covariations in firing rates of neurons, ranging from spatially and temporally constant input, to spatial fields formed as the sum of Gaussian basis functions, as well as fields for speed, theta oscillations, and head and running directions. We evaluated the explanatory power of these models by comparing their likelihood values and found that speed, head direction and running direction had little power in explaining the data, while theta oscillation phase and pairwise couplings had more explanatory power. Although there were variations in terms of the relative strength of the couplings depending on the assumptions about the external fields, we consistently found that the inferred connections maintained a pattern that supports the attractor network hypothesis: cells with nearby phases tend to excite each other while those further apart inhibit each other. We also found that the strongest connections were among cells within the same module, that the connections were both negative and positive, and that none of our conclusions were sensitive to data limitations.
We analyzed two data sets with simultaneously recorded grid cells, one with a total of 65 cells, of which 27 were grid cells (referred to as data set 1), the other with 8 grid cells (data set 2). As mentioned, grid cells are known to cluster according to the spacing and orientation of their spatial fields, with cells with similar spacing making distinct functional modules that react in unison to external manipulations of the environment as quasi-independent populations [7]. In data set 1, all but 5 of the grid cells were easily identified into three distinct modules (see Material and Methods). In data set 2, all 8 cells belonged to the same module.
To calculate correlations between pairs of grid cells, beyond what is expected from spatial rate covariations, we binned the spike data into 1 ms intervals and smoothed the firing rates with a 20 ms Gaussian filter. The trajectory of the animal was then binned spatially by dividing the environment into a number of N × N square boxes, using different values of N = 2, 3, 4, 5, 10, 15, 20, 40, 75. Noise correlations, Cij, between cells i and j were then determined as the mean of the Pearson correlation coefficients, ρ, calculated over the trajectories through each spatial bin (see Material and Methods). As shown in Fig. 1, in the case of dividing the environment into 20 × 20 spatial bins, we found noise correlation values close to zero, or slightly negative, for cells with non-overlapping spatial fields. On the other hand, cell pairs close in phase distance showed positive noise correlation values that increased for cells closer to each other in phase; see Fig. 1A and B. The slope (β̂) and intercept (α̂) of a linear regression line (not shown) are β̂=−0.22 and α̂=0.09 for data set 1, and β̂=−0.25 and α̂=0.11 for data set 2, all significantly different from 0 (t-test, P < 0.001).
Since data set 1 included neurons from 3 separate modules, we also studied the dependence of the noise correlations on the phase difference between cells for each three modules separately. Except for the module with the largest field spacing (Fig. 1E), where the phase dependence was weak (intercept and slope of linear regression not significantly different from 0 (t-test, P>0.7)), the modules showed a significant pattern similar to that of all modules pooled together shown in Fig. 1A (intercept and slope of linear regression significantly different from 0 (t-test, P<0.001)). Similar results were found when other spatial bin sizes were used. This extends the results of [18] to two dimensions and also shows the variations in the phase dependence of the correlations to the module size.
Good empirical estimates of the noise correlations, as defined above, require that the rat makes enough passes through each spatial bin during the recording session. This means that the bins cannot be too small, otherwise there would be very few visits to most of the bins, and some of the bins may never be visited at all. On the other hand, if the bins were too big, the variations in rate from one pass through the bin to another would be be too large and, therefore, Cij would not exclude the rate covariations. We, therefore, looked at how consistent our estimates of the correlations were as a function of the spatial bin size by calculating the Pearson correlation coefficient between the correlations measured, using a random half of the visits to each spatial bin with those measured from the other half (see Material and Methods). The most stable estimate was with 20 bins per side of the box (or 7.5 cm), which is what we have used in Fig. 1. In this best case scenario, for data set 1, the Pearson correlation coefficient is 0.56 for the full data, with both halves of the data in all 20 sets of random halves still demonstrating the phase dependent pattern shown in Fig. 1. Cells with nearby grid patterns had stronger positive correlations, while those further apart in phase demonstrated a slightly negative, or no correlations (the slope and intercept of the linear regression lines were all significantly different from 0 (t-test, P < 0.03)). This was also the case for the 20 random halves of data set 2.
The pairwise correlation analysis done here is a good first step, however, it suffers from a number of shortcomings. First of all, it is really a pairwise measure, which excludes the interactions with other neurons, and thus a perceived correlation between two cells might really be explained by the presence of a third neuron or external covariates. Second, although we take into account spatial covariations in rate, there is no systematic way of evaluating how much other covariates, such as theta oscillations or head direction, contribute to the correlations between cells. Given the fact that grid cells are known to covary with these, it is important to evaluate their influence when analyzing correlations between grid cells. While pairwise correlation analysis suffers from these problems, they can be addressed, to a large extent, using statistical models of the GLM type. This is what we will do in the rest of the paper.
As a statistical model, we considered the simplest maximum entropy model to include both asymmetric couplings and time varying external input: the kinetic Ising model. The activity of the cells was binned in 10 ms bins, and a binary variable Si(t) was associated to each neuron in each bin, which would be equal to +1/-1 denoting the presence/absence of spikes emitted by neuron i within time bin t. Letting the state of each neuron at time t depend on the state of the population in the previous time step t − 1 and some covariates, independent of the state of the system, the maximum entropy distribution over the state Si(t) of neuron i at time t is [20]
P ( S i ( t ) | { S ( t − 1 ) } ) = exp [ S i ( t ) H i ( t − 1 ) ] 2 cosh [ H i ( t − 1 ) ] , (1)
H i ( t − 1 ) = h i ( t − 1 ) + ∑ j J i j S j ( t − 1 ) (2)
where Jij would be identified as the functional coupling from neuron j to neuron i, and hi(t) as the time varying covariate which in statistical physics terminology is called an external field. As mentioned in the introduction, Eq. 1 defines a GLM, where in each time bin, mostly only one or zero spikes per bin are observed and the interaction kernel extends one time step in the past. With binary states and only one time step kernels, this model represents the simplest possible model capable of capturing functional connectivity from neural data, which is convenient given the finite time in which the neural recordings were taken. This model should not be confused with the maximum entropy equilibrium models (equilibrium Ising model [21, 22]), which assume symmetric couplings and are not related to the GLMs.
Given Eq. 1, we asked what values of the parameters hi(t) and Jij are the most likely to generate the observed data. Both exact and fast approximate algorithms for solving the inverse kinetic Ising model have been developed [23] similar to other GLM models [15, 16, 19]. The exact solution is found by maximizing the log-likelihood function
L [ S , J , h ] = ∑ i t [ S i ( t + 1 ) H i ( t ) − log 2 cosh H i ( t ) ] (3)
with respect to hi(t) and Jij. The term ‘exact’ is used here in the sense that if data is generated by a kinetic Ising model, this learning algorithm would recover the parameters exactly in the limit of infinite data. The log-likelihood is the logarithm of the probability of observing the data at hand given that it was generated from the model, and thus measures how well the model explains the statistics in the observed data. In our analysis we have used the natural logarithm.
An important issue in dealing with a model of this type is choosing the external field. In the absence of couplings, the external field, hi(t), can explain the variations in the firing rate as the rat navigates in space. Ideally, the external fields can be inferred by binning the environment into small spatial bins, assuming that the external field in each bin takes a constant value for each neuron. If the rat passes through each bin many times, the external field in each bin can be reliably estimated. However, during a recording period, and as described above, the requirement of passing through small spatial bins many times is rarely satisfied.
Alternatively, the spatial input could arise as the sum of two-dimensional Gaussian basis functions with the basis set spanning the environment. By inferring the parameters of a linear combination of Gaussian basis functions (see Material and Methods for details), an accurate representation of the spatial field can be found, even with a reduced amount of data, as shown in the following.
Focusing on data set 1, which had the most cells, we first inferred couplings, assuming that each neuron receives an external field which is constant across time and space, hi(t) = hi. Next, we studied how the inferred couplings were affected by increasing the spatial resolution of the external fields, hi(t), to account for the spatial variation in firing rate by dividing the environment into spatial bins, considering the cases of bins of size 37.5 cm and then bins of size 7.5 cm, assigning one external field per box to each cell. We also considered external fields in the form of a sum of Gaussian basis functions. Fig. 2 shows the resulting couplings, plotted against couplings found in the model that assumed spatially and temporally constant external input, hi, for each neuron. As can be seen, increasing the resolution of the external fields made the couplings weaker but not inconsistent with the constant field case, even in the case of Gaussian fields, where the spatial rate maps were well captured by the model, as shown in Fig. 3. In this case, there was a significant weakening of the couplings (the estimated variance of the Gaussian field model couplings (S Gauss 2) was significantly smaller than that of the constant field model (S constant 2), (F-test for equal variances, P<0.001)). In each of the models, the total external fields were negative and often strong, as one would expect for data sets with low firing rates (mean firing rate 2.4 Hz).
Interestingly, no matter which of the various external fields we used, when neurons i and j both belong to one of the two smaller modules of data set 1, or the one module of data set 2, the inferred couplings, Jij, showed a consistent dependence on the spatial phase difference, with nearby phases showing positive Jij while those further away more negative values. This is shown in Fig. 4 for both data sets for the case of the Gaussian fields. The slopes and intercepts of linear regression lines were all significantly different from zero, both for the full data and the 20 sets of random halves (t-test, P<0.02) for all figures except for Fig. 4E, where the slope and intercept of linear regression were not significantly different from 0 (t-test, P>0.7). We remind that with the Gaussian fields, the correlations between two cells due to overlapping fields are explained away.
Since many cells in our data had some theta phase and head directional preferences, we also considered a model in which each cell was coupled to the head direction of the animal and the LFP theta oscillation through coupling constants that were inferred from the data; see Material and Methods. In general, there were only small differences between the couplings when theta and head direction were added. This can be seen in Fig. 5A, which shows the couplings in the model with Gaussian fields with and without theta included. In this case, we observed a small but selective change, depending on the phase preference of the neurons. The cells could be clustered into two groups according to their theta phase preference (see Material and Methods): one with connections between cells of similar theta phase preference, and the other with connections between cells with opposite preference.
Couplings between cells with similar theta phase preference were on average positive (average (μ) significantly different from 0 (t-test, P<0.001)), whereas couplings between cells of opposite theta preference were on average negative (μ < 0, P<0.001). As shown in Fig. 5B, including the time-varying phase of theta as an external covariate resulted in shifting the coupling strength towards less positive values for pairs of cells that prefer the same phase of theta (μno theta > μtheta, P<0.001), whereas the opposite was true for couplings between cells that showed preference to opposite phase of theta (μno theta < μtheta, P<0.001).
One would expect, based on the experimental indications of modules operating independently, that grid cells of the same module are more likely to participate in the same functional network than neurons from different modules. We found that the couplings within and between modules in data set 1 both had means close to zero (within modules (mean±std): −0.01±0.13, between modules: −0.01±0.09). However, the within module couplings had a greater variance (S within 2 > S between 2, P<0.001)), i.e. there was a higher proportion of couplings with high absolute values within modules than between, as can be seen in Fig. 6. This result was found to be stable with respect to data limitations, as shown in the next section.
In this section we consider a number of factors that could have influenced our estimations of the couplings, and show that our results were stable with respect to these factors.
It is known that some grid cells show phase precession. This could be an additional source of correlation, so we tried to address how phase precession can influence the couplings. We first investigated whether or not any of the cells in our data phase precess, focusing on data set 1. In general, quantifying phase precession in two-dimensions is a difficult task due to the changes in the animals movement direction within the field. To classify cells as phase precessing or not, we thus used a novel approach described in [24], correlating the distance to the field peak projected onto the current running direction with the phase of theta at the time of spikes. Our analysis revealed that 13 of the 27 grid cells showed significant phase precession (5 of 8 in module 1, 6 of 7 in module 2, and 2 of 7 in module 3). We then excluded the couplings between phase precessing cells from the analysis for the two smaller modules and found that this did not remove the trend reported in Fig. 4 between the spatial phase difference and the inferred couplings. As can be seen in Fig. 7A, there was still a significant negative relationship between coupling value and spatial phase distance for cell pairs in which at least one of the cells do not show significant phase precession (both the slope (β̂=−0.60) and intercept (α̂=0.21) of the linear regression line are significantly different from 0 (t-test, P<0.001)).
It has been suggested that correlations and thus inferred couplings from multi-electrode recordings can be biased due to problems with spike sorting [25,26]. Since the main part of our conclusion is on the phase dependence of the correlations and functional connections and not their actual value, and since the phase of grid cells appears to be not anatomically ordered, it is unlikely that a phase dependent bias would be introduced to the correlations due to mistakenly assigning spikes to wrong cells. In addition to this, the cells in the two data sets analyzed here were recorded using hyperdrives that consist of 14 independently movable tetrodes [7]. It has been suggested that a tetrode is unlikely to record signals from cells farther than 65μm away [27]. As the distance between tetrodes on the hyperdrive is approximately 250±50 μm, it was very unlikely that the same cell was recorded on two tetrodes, and in that way confound our results across tetrodes. We therefore examined the couplings versus spatial phase for cell pairs from different tetrodes, and found that this led to a qualitatively similar result, as shown in Fig. 7B (both the slope (β̂=−0.59) and intercept (α̂=0.20) of the linear regression were significantly different from 0 (t-test, P<0.001)).
In order to investigate the stability of the inferred couplings and the various covariates to data limitations we inferred the parameters of the models using only half of the data, and compared them with the ones from the other half. For this, we defined the spike data as being made up of consecutive time pairs, (S(t), S(t+1)) and created partitions by randomly selecting 50% of the pairs. In this way, we generated 20 random sets, and for each set inferred the couplings using constant fields without taking theta and head direction into account, and Gaussian fields with theta and head directional input included (the full model). In general, the inferred couplings from these random halves were correlated with each other. As shown in Fig. 7C and D, the within module couplings were more stable than the between-module ones, with an average Pearson correlation coefficient of 0.88 versus 0.73 for the constant field model, and 0.70 versus 0.51 for the full model. We noticed that the self-couplings are the ones that are most stable from one half to the other, showing a Pearson correlation coefficient of 0.94 between the couplings inferred from the two halves for the full model. We also found that the mean absolute values of the within and between module couplings maintained their relationship, with stronger couplings between cells within module than those between modules, for all 20 random partitions of the data (S within 2 > S between 2, P<0.005 for all 20 random partitions, in both constant field model and Gaussian field model).
The analyses reported here were produced using the data from two recordings of grid cells, the biggest of them consisting of 27 grid cells. This was the biggest data set we had access to, but still represents only a small fraction of the true local cell population. One might wonder how much the connections between these cells would be influenced if we had access to recordings from more cells. As described in Material and Methods, data set 1 included neurons which were not classified as grid cells. We found that using this entire data set (65 cells) did not affect the couplings between grid cells (see Fig. 7E).
In order to evaluate the strength of the statistical effect of the couplings and the external covariates on explaining the correlations in spike trains, we calculated the log-likelihood of half of the data using parameters inferred from the other half for various models for both data sets. The results are shown in Fig. 8A-D. To correct for the number of parameters, the total log-likelihood was penalized according to the Akaike correction, that is by subtracting the number of inferred parameters (covariates and couplings) used in each model (see Material and Methods) [28]. The negative log-likelihoods of the models without the couplings are also shown. In a likelihood ratio test, all covariates gave a significant increase (P < 0.001) compared to the constant field model. This was also the case where we included the couplings in each of the models compared to the same model without couplings. In general, adding head direction as a covariate had little effect on the likelihood. The effect was even weaker when including speed as a covariate, or using running direction instead of head direction (see methods), with the penalty from the Akaike correction larger than the increase in likelihood from the inclusion of the parameters. For the case of constant fields, adding couplings and then theta had the most significant effect. It is interesting to note that, when comparing the constant field model to the model with spatial fields, the impact on the likelihood from including the couplings is reduced, as would be expected by explaining away the spatial component of the correlations. Adding theta resulted in a consistent increase in the log-likelihood yielding 0.0025 for the model with constant fields and 0.0026 for spatial.
What is known about the connectivity in the grid cell network is primarily based on anatomical in vitro studies. Recent studies show that stellate cells in layer II are connected to each other primarily through inhibitory interactions [14,29], and that the inhibitory drive varies dorsoventrally as the size of the grid spacing changes [30]. As opposed to the connections between layer II stellate cells, within-layer recurrent excitation has been found between the main type of principal cells, namely pyramidal cells, in both layer III and V [31]. Although the picture drawn by these studies emphasizes the role of recurrent interactions in developing the properties of grid cells, it does not show how interactions between grid cells quantitatively depend on properties such as theta rhythmicity and spatial phase separation, properties that play a major role in computational models of grid cells. A previous work on in vivo recordings that studied phase dependence of the interactions between cells in MEC focused on pairwise correlation analysis by using recordings from one dimensional tracks [18], showing that cells with nearby phases have stronger correlations than those far apart in phase. Another recent in vivo study used strongly peaked cross-correlations as a signal for the presence of connections and has concluded that grid cells with a wide range of phases project to a given inhibitory neuron [32]. To analyse the multi-neuronal recordings in grid cells we took a different approach from previous studies: that of statistical inference. We used a kinetic Ising model and studied how functional connections depend on phase difference between grid cells, their level of theta modulation, speed modulation and head directionality, and the statistical role that these connections play in shaping multi-neuronal activity.
The kinetic Ising model that we used here for the inference is a model with minimal assumptions: (1) it is the maximum entropy distribution over the spikes of neurons at time t, given the spikes at time t−1 [20], and (2) it is pairwise (meaning it only takes into account the first-order non-trivial interactions). Being a generalized linear model, it is closely related to other GLMs used for analyzing population recordings from other parts of the brain [15–17], and it also employs the maximum entropy approach used by many in analyzing neural [21,22] or other biological data [33]. Our analysis showed that the correlations and the functional connections between grid cells demonstrate a spatial phase dependence, even when spatial variations in rate (as well as other possible sources of correlations, such as theta oscillations and head direction) are taken into account. Both correlations and functional connections were positive for small phase differences. Functional connections became negative, while the correlations approached zero, for larger spatial distances for cells in the one module in data set 2, and in the two smaller modules in data set 1. This connectivity provides support for a role played by attractor dynamics as suggested by several modelling efforts [11–14]. The trend in the phase dependence was, however, less clear in the third module in data set 1: the common inhibitory portion was represented, but we did not find any functional excitation between cells close in phase, possibly because of the lack of recorded cells with similar phase in this module. We also found that the absolute value of the couplings was bigger for pairs of cells that belonged to the same module than those belonging to different modules. This supports the idea that neurons in the same module form a more coherent population of neurons, bound together in a stronger manner than those in different modules.
In attractor models of grid cells, the phase dependent connectivity pattern allows the network to maintain a continuum of stable states such that, if the neurons of a single module could be aligned according to their phases, the activity on that neural sheet would itself show a regular pattern of activity. This local and relatively rigid relationship between within-module grid cells has been surprisingly well supported. First identified in [1], grid cells were found to locally share both orientation and spacing that were later observed to remap and deform coherently [7,10,34]. It has also been shown that the characteristics of the grid pattern of one cell were more stable relative to other grid cells than with respect to local features of the environment [10]. This was even more pronounced in novel environments where the individual fields were still changing significantly relative to the environment while remaining relatively stable between cells [10], further suggesting that the coding of the grid cells is more coherent within the grid cell population than it is with the actual space it is encoding. Even more convincing, a recent study looking at a large population of cells taken from single animals in the same environment showed that the cells clustered into a finite number of modules [7] suggesting there exists not only the large number of cells necessary for an attractor map but that there might be a finite number of these networks working together to better provide a metric of space. Our work complements these studies in that we show that there exists the functional connectivity of the type necessary to establish the patterned network activity that has been proposed to explain the above experimental observations.
As opposed to the attractor model [11–13], other grid cell model frameworks, the oscillatory interference [35] and the adaption model [36], were originally conceived as single cell models that suggest that the periodic firing comes from a combination of convergent input and cellular mechanisms within an individual neuron. As such, the role they have prescribed for the lateral connectivity has been mainly to align the grid patterns of the cells, without requiring any phase dependence in the couplings per se. However, it has recently been noted [9, 37] that in the adaptation model, interactions between grid cells can also be learned, resulting in a developmental model for the phase dependent connectivity which could later sustain a continuous attractor dynamics. In addition to aligning the grids, this connectivity will allow the adaptation model to code for novel environments much more rapidly while maintaining the stabilizing benefit of having convergent spatial input.
In our statistical inference, we considered various external covariates that comprise what is known about the single cell coding of these cells, including spatial, speed, theta oscillations, head direction and running direction inputs. Adding these additional covariates to the models with constant field or Gaussian fields had little effect on the connectivity, but there was a significant weakening of the couplings when we compared the couplings of the Gaussian model to those of the model with constant fields. This is not surprising, as a component of the correlations in the model with constant fields was likely due to overlapping fields which was better explained by the spatial component of the Gaussian model. One benefit of using a statistical model is the quantification of the relative contribution of the individual covariates to the overall likelihood of the data under the model, with the spatial component having the strongest impact followed by functional connectivity and theta preference. Speed, head direction and running direction, as covariates, had a small impact in all cases that we considered.
In all the statistical models, ranging from constant external field to Gaussian with and without theta and head direction, we found that the model without couplings was worse at explaining the statistics of the data than the same model with couplings, even when the Akaike corrections were taken into account. Further support for the significance of the couplings come from the stability of the connectivity when inferred from separate halves of the data.
Since the self-couplings appeared to be the most stable when one random partition of the data was compared to the other, we wondered how the rest of the couplings would react if we did not include the self-couplings. With the refractory period in mind, positive self-couplings might seem counter-intuitive. However, the refractory period lasts for only a few milliseconds, and we use 10 ms time bins. In addition, grid cells are primarily active only when the animal is in the cell’s spatial fields, and silent otherwise, i.e. the state of a grid cell in a time bin is likely to be equal to the state in the previous time bin, which a statistical model could interpret as a positive self-coupling. Removing self-couplings, however, had little effect on the couplings between cells (Pearson correlation coefficient > 0.98 for the constant model and the full model, for both data sets).
Stellate cells of MEC layer II, the main grid cell candidates, are known to functionally inhibit each other. In our analysis, the inferred connections were both inhibitory and excitatory. There are a few points to note regarding this apparent contradiction. First, considering the recording locations of the tetrodes in data set 1 (see Supplementary figure 4 (rat 14147) in [7]), and that a number of cells in this data set show head direction preferences, a property rarely observed in the layer II population [3], many of these cells are most likely recorded from deeper layers where, as mentioned, both intra- and interlayer excitatory connections between principal cells have been found. For data set 2, on the other hand, it seems probable that a bigger fraction of the cells is from layer II (see Supplementary figure 14 (rat 13855) in [7]). It is, however, not possible to confirm the exact location or principal cell type for the cells analyzed here. Second, the relationship between the inferred functional connections and the underlying anatomical connectivity is a nontrivial one which may involve other non-recorded neurons. It is also possible that the correlations driving the functional connectivity come from a common input that was not accounted for here. This input, however, should be non-spatial, non-directional and independent of theta phase, but still depend on the spatial phase difference between pairs of neurons and whether or not they belong to the same module. It would be interesting to see what such a signal could look like. The existence of such an input would, of course, leave the question open as to how the local network is connected, while opening a new possibility that the grid cell modules play a role in encoding currently unidentified features that are neither spatial or directional. Since it is possible in computer simulations to identify the presence or absence of a synapse based on the inference of functional connections [38, 39], it would be very interesting to see how the inferred functional couplings and correlations look like for a data set exclusively from layer II cells for which the actual functional connectivity between stellate cells is known. In addition, considering the fact that modules span layers [7], our results also make a case for taking a closer look at the between layer connectivity and how the different cell types and connectivity patterns might work together to develop the grid cell code.
With Gaussian fields, the model with only theta has a slightly higher likelihood than the one with only couplings, although the couplings still exhibit the phase dependence shown in Fig. 4. The relative improvement gained by pairwise connections in explaining the data is known to scale with the size of the recorded population [21,40,41], while other sources of higher order correlations will also scale up. It would therefore be interesting to see how the relative contribution of the various factors, in particular that of theta oscillations, will scale compared to that of the pairwise couplings. Future large-scale recordings of grid cells should allow us to perform such analyses.
Two recordings of the activity of cells in the MEC area of two Long Evans male rats (from [7]) were analyzed in this paper. One recording, referred to as data set 1, consisted of a total of 65 cells (rat 14147 in [7]), where 27 were classified as grid cells (mean firing rate: 2.4 Hz). These 27 cells distributed over 7 tetrodes, and 22 of them could be assigned to one of three modules (see [7] for methods). The number of cells in each module, along with mean spacing and orientation is given in Table 1. The other recording, data set 2, consisted of 8 grid cells (mean firing rate: 2.8 Hz) distributed over 3 tetrodes (rat 13855 in [7]). All 8 cells belonged to the same module. Mean spacing and orientation for this module is listed in Table 1. The movement of the rats is shown in Fig. 9.
The spikes were binned into 10 ms time bins, but using both 20 ms and 5 ms time bins led to similar results. Using the binned data, a spike matrix of −1’s and 1’s was constructed, where a ‘−1’ indicated that the cell did not fire in time bin t, and a ‘1’ indicated that the cell emitted one or more spikes in time bin t. More than one spike rarely happened (both data sets: average over cells = 0.1 (±0.1)% of the time bins).
Noise correlations were defined as
C i j = 〈 ρ ( r ¯ i a , r ¯ j a ) 〉 a
where r ¯ i a is a 1×k vector consisting of the average firing rate of neuron i in each of the k trajectories through spatial bin a, and ρ(⋅, ⋅) is the Pearson correlation coefficient (PCC), defined as:
ρ ( r ¯ i a , r ¯ j a ) = E [ ( r ¯ i a − 〈 r ¯ i a 〉 k ) ( r ¯ j a − 〈 r ¯ j a 〉 k ) ] σ [ r i a ] × σ [ r j a ]
with both the expectation (E) and the standard deviations (σ) over the k trajectories.
Random partitions: Each spatial bin has a given number of visits. To split the data into two random partitions, for all visited bins, a randomly chosen half of the visits to each bin was assigned to one partition, the other half to the other partition.
The cells could be divided in two clusters based on preferred phase of theta. The theta phase preference was defined as the peak in a circular kernel smoothed density estimate of the distribution of theta value at spike time. The number of clusters were defined as the number of local peaks in a kernel smoothed density estimate of the distribution of theta phase preference peaks for all cells. A circular k-means clustering algorithm were performed to assign cells to clusters. The clusters are shown in Fig. 10.
We used the kinetic Ising model to infer the functional network connectivity, i.e. we assumed that the observed spike train comes from the probability distribution in Eq. 1. We constructed different versions of the model by varying the form of the external field in several ways as described in the introduction and in more details below.
To allow the external field of the kinetic Ising model to account for the spatial variations in the firing of the grid cells, we started, for data set 1, by dividing the environment globally into K square boxes. We defined three models with increasing spatial resolution, with K = 4 × 4 (37.5 cm boxes) in the first model, and K = 20 × 20 (7.5 cm boxes) in the second. For each K, we defined external fields αik for each cell i and box k. The field resulting from this spatial discretization is then h i S ( t ) = Σ k α i k I k ( t ), where Ik(t) is a function indicating the presence (1) or absence (0) of the animal in box k at time t.
We further increased the resolution of the spatial fields using Gaussian basis functions centered on an evenly spaced M × M square lattice covering the recording environment. The spatial field for cell i at time t is then
h i S ( t ) = ∑ j k α i j k exp [ − ( ( x ( t ) − x j k ) 2 + ( y ( t ) − y j k ) 2 ) / r 2 ] + h i (4)
where (xjk,yjk) and r are the vertices of the regular lattice and the widths of the basis functions, respectively. To determine the optimal values of M and r (M = 15 and r = 8.5 cm), we maximized the likelihood for a range of values of M and r and chose the values of the parameters that gave the highest Akaike-adjusted likelihood value. To include the external theta phase preference, we computed the fast-Fourier transform of the local field potential (LFP) and set the theta rhythm to the maximum component between 4–12 Hz. From this, we constructed a theta input vector, where each element was the angular average ∈ (−π, π] of the theta phase in that time bin. The partial field for cell i at time t due to local field potential theta preference is then
h i LFP ( t ) = ∑ k α i k exp [ − d ( Θ ( t ) , Θ k ) 2 / ( π / 6 ) 2 ] + h i (5)
where d(Θ(t),Θk) is the minimum angular distance between Θ(t), the theta phase in time bin t, and Θk, the k’th component of a set of 10 equally spaced angular phases. The number of angles and width of Gaussian (π/6) was selected by maximizing the Akaike-adjusted likelihood of the model in the same way parameter values for M and r in the model with spatial fields were chosen, as described above.
The head and running direction components was also accounted for using sums of Gaussian basis functions
hiHD(t)=∑kαikexp[−d(φ(t),φk)2/(π/6)2]+hi (6)
where φ(t) is the head direction ∈ (−π, π] at time t, calculated from the projection of two LEDs onto the horizontal plane, and φk is the angular position of the kth basis function. The number of basis functions (10) and width of Gaussian (π/6) were selected by maximizing the Akaike-adjusted likelihood of the model, the same way it was done for parameter choice in the spatial and theta model. Speed was also incorporated into the model with a simple time-varying field, αi s(t), where s(t) is the average speed in the 100ms window around each time bin.
In all of the models, the parameters, Jij, hi and α’s, were found by maximizing the likelihood function given in (3) for the data under the different models by gradient ascent. When comparing the models, we first Akaike-corrected the log-likelihood. The Akaike information criterion (AIC) is a measure to compensate for overfitting by models with more parameters, where the preferred model is that with the minimum AIC value, defined as
A I C = − 2 ln ( L [ D | θ M L ] ) + 2 k (7)
where D is the observed data, and L[D∣θML] is the likelihood at the maximum likelihood (ML) estimates of the parameters θ (θML), and k is the number of parameters [28]. Equivalent to the method described above, we corrected the total log-likelihood as l n ( L A kaike ) = − A I C 2.
|
10.1371/journal.pbio.1001565 | Three-Dimensional Reconstruction of Bacteria with a Complex Endomembrane System | The division of cellular space into functionally distinct membrane-defined compartments has been one of the major transitions in the history of life. Such compartmentalization has been claimed to occur in members of the Planctomycetes, Verrucomicrobiae, and Chlamydiae bacterial superphylum. Here we have investigated the three-dimensional organization of the complex endomembrane system in the planctomycete bacteria Gemmata obscuriglobus. We reveal that the G. obscuriglobus cells are neither compartmentalized nor nucleated as none of the spaces created by the membrane invaginations are closed; instead, they are all interconnected. Thus, the membrane organization of G. obscuriglobus, and most likely all PVC members, is not different from, but an extension of, the “classical” Gram-negative bacterial membrane system. Our results have implications for our definition and understanding of bacterial cell organization, the genesis of complex structure, and the origin of the eukaryotic endomembrane system.
| The compartmentalization of cellular space has been an important evolutionary innovation, allowing for the functional specialization of cellular space. This compartmentalization is extensively developed in eukaryotes and although not as complex and developed, compartments with specialized function are known to occur in bacteria and can be surprisingly sophisticated. Nevertheless, members of the Planctomycetes, Verrucomicrobiae, and Chlamydiae (PVC) bacterial superphylum are exceptional in displaying diverse and extensive intracellular membranous organization. We investigated the three-dimensional organization of the complex endomembrane system in the planctomycete bacterium Gemmata obscuriglobus. We reveal that the G. obscuriglobus cells are neither compartmentalized nor nucleated, contrary to previous claims, as none of the spaces created by the membrane invaginations is topologically closed; instead, they are all interconnected. The organization of cellular space is similar to that of a classical Gram-negative bacterium modified by the presence of large invaginations of the inner membrane inside the cytoplasm. Thus, the membrane organization of G. obscuriglobus, and most likely all PVC members, is not fundamentally different from, but is rather an extension of, the “classical” Gram-negative bacterial membrane system.
| The compartmentalization of cellular space has been an important evolutionary innovation, allowing for the functional specialization of the membrane-bound organelles. This compartmentalization is extensively developed in eukaryotes, and although not as complex and developed, compartments with specialized function are known to occur in bacteria [1]. Some examples include protein-bound organelles, like carboxysomes, which increase the concentration of metabolite in a closed space [2] and gas vesicles, which are gas-filled protein-bound organelles that function to modulate the buoyancy of cells [3]. Other examples include the magnetosomes in magnetotactic bacteria, which are invaginations of the cytoplasmic membrane that enclose a magnetic mineral without achieving separation into individual vesicles [4]. Individual magnetosomes are arranged into one or more chains within the cell, where they act to orient the cell within a magnetic field. Photosynthetic prokaryotes including the purple bacteria, the cyanobacteria, and the green bacteria have photosynthetic membranes extending from their inner membrane (IM), also called cytoplasmic membrane, maximizing the size of the membrane surface exposed to light. These membranes can adopt diverse shapes, including the formation of membrane stacks continuous with the cell membrane, spherical invaginations of the inner membrane so that multiple membrane spheres are connected to one another or are folded in an accordion-like structure and adjacent to the cell membrane [5]. Lastly, the anammoxosome is a membrane-bound compartment found in the anammox bacteria, which are divergent planctomycetes. It houses the anaerobic ammonium oxidation reaction. Its membrane is enriched in unusual concatenated lipids, the ladderane lipids, which form an impermeable barrier preventing the diffusion of the toxic intermediates produced during the anammox reaction [6].
Bacterial cell organization can be surprisingly complex. Nevertheless, members of the Planctomycetes, Verrucomicrobiae, and Chlamydiae (PVC) bacterial superphylum are exceptional in displaying diverse and extensive intracellular membranous organization. For this reason they have been labeled the “compartmentalized bacteria” [7],[8]. The planctomycete Gemmata obscuriglobus is particularly interesting because a double membrane, formed from a folded single membrane, has been suggested to surround its genetic material. This double membrane is reminiscent of the eukaryotic nuclear envelope, leading to the name “nucleated bacterium” [7],[9]. Early ultrastructural analysis based on thin sections of cryo-substituted cells, freeze-fracture replicas, and computer-aided 3-D reconstructions has been used to argue that the DNA in G. obscuriglobus is enclosed within a compartment separated from the rest of the cytoplasm [8],[10]. However, the data are not entirely convincing. A three-dimensional (3D) reconstruction from serial sections and fluorescence microscopy of living cells was presented to support the claim of “the continuous nature of the membranous envelope surrounding the nuclear body and completely enclosing the nucleoid, apart from where gaps appear in the envelope” [8]. As stated by the authors, the “outer region of the nuclear body has a similar appearance to the cytoplasm,” and ribosomes are located in the same compartment as the DNA, arguing against the specific nature of this compartment. In addition, ribosomes line the walls of the internal membrane of the “nuclear compartment” [8], as observed along the inner membrane (IM) of classical bacteria.
This and other analyses have led to the suggestion that the PVC cell plan is different from “classical” Gram-negative bacteria, such as E. coli, because of the absence of a typical outer membrane (OM) [7],[8]. The outermost membrane closely juxtaposed to the cell wall was interpreted as the cytoplasmic membrane, while the remaining membrane was called the intracytoplasmic membrane (ICM), mainly based on the distinctive organization of the ICM supposedly surrounding the DNA. The claimed absence of an OM implied the absence of a periplasm, the volume located between IM and OM in Gram-negative bacteria. More recent evidence based on genomic information argue against this conclusion, including the presence of genes associated with the OM and the periplasm in Gram-negative bacteria [11],[12], and the presence of remnants of the division cluster and the peptidoglycan synthesis pathway (typically anchored in the OM) [13].
A more recent analysis of vitrified sections by cryo-electron tomography implied that the “internal membrane” system might be continuous with the ICM, but formed by membrane invaginations and that “the bacterial nucleoid is not completely sealed by the double-membrane system” [14]. It was observed that “the double-membrane network of G. obscuriglobus cells emanates from the intracytoplasmic membrane to form unsealed compartments.” In that study, the bacteria were preserved close to native state, sectioned, and imaged under cryogenic conditions to reduce preparation-induced artifacts. However, because of the difficulties involved in sectioning cells under liquid nitrogen temperatures and the technical challenges presented by the use of vitrified sections in obtaining serial sections of a whole cell, the analysis was based on tomographic reconstruction of only a fraction, up to 150 nm thick sections, of G. obscuriglobus cells, which are usually ∼2 µm in diameter.
We have recently contributed to this series of analyses and have described the cell organization in two types of G. obscuriglobus cells [15]. In the first type, the dividing form, the inner membrane protrudes deeply into the cytoplasm to form thin membrane sheet invaginations extending towards the inside of the cell. The second cell type is not budding, and has increased periplasmic volume populated by vesicle-like structures.
Till present, how the membranes are organized in 3D is not known for any of the PVC bacteria. We have thus investigated the 3D membrane organization in multiple cells of the species G. obscuriglobus. In order to capture the membrane organization of entire cells, we chose to use plastic embedding for this study. Here we present the reconstructed volume of one complete cell of the first, dividing type, where we followed the entire organization of internal membranes within the cell. We report for the first time the 3D reconstruction of a bacterium with a complex endomembrane system. Our 3D reconstruction reveals that G. obscuriglobus cells are neither compartmentalized nor nucleated. We show that the spaces created by the membrane invaginations are all interconnected and not closed. The organization of cellular space is similar to that of a classical Gram-negative bacterium modified by the presence of large invaginations of the IM inside the cytoplasm.
We acquired tomograms from 10 different bacterial cells (Table S1). We encountered difficulties with attempts to automatically track membranes and their interconnections. Currently, there is no software available that can accurately assign and follow the membranes in such a complex system and our attempts at automation did not achieve satisfactory results. We therefore manually assigned and traced membranes in more than 200 slices. In addition, reconstruction and modeling in 3D also required some manual intervention.
Three-dimensional reconstruction reveals that G. obscuriglobus cells have a cell plan that is not radically different from that of a typical Gram-negative bacterium (Figure 1; Figure S1). The organization is topologically compatible with an extension of the periplasmic space by invagination of the bacterial IM towards the cell's interior. This is supported by the fact that ribosomes line the IM and its invaginations in G. obscuriglobus cells, as they do along the IM of other Gram-negative bacteria. This similarity of topological organization is supported by genomic information [11]–[13]. The main difference is that the G. obscuriglobus IM invaginates extensively towards the interior of the cell to form a network of sheets within the cytoplasm (Figure 1; see Supplementary Movies 1 and 2, available at http://www.bork.embl.de/~devos/project/apache/htdocs/plancto/g3d/ [Text S1]). The space inside the invaginations is continuous with the periplasm and devoid of ribosomes, as in other bacteria (Figure 1; Figure S1).
We have observed ribosome-covered extended membrane sheets, as in the eukaryotic rough endoplasmic reticulum (ER) or the nuclear envelope, which have associated ribosomes, as opposed to membrane tubules associated to the eukaryotic smooth ER. The mean lumenal width of the internal membrane sheets is ∼20 nm (mean of 18.8). This is slightly smaller than the ∼30 nm and ∼50 nm reported, respectively, for yeast and mammalian ER sheets [16].
These membrane extensions have a significant impact on the cell organization, in particular on the ratio of OM versus IM. E. coli cells are about 1.5 µm long and 0.5–0.6 µm in diameter, with a cell volume of ∼0.65 µm3 [17]. Their periplasm comprises between 20% and 40% of the total cell volume [18]. With a diameter of ∼2 µm, the complete volume of the reconstructed G. obscuriglobus cell is 3.4 µm3, while the cytoplasm is 2.6 µm3. The periplasm, including the space created by the invaginations of the IM, has a volume of .82 µm3 (∼one third, 31.7%, of the cell's volume, similar to E. coli). The important difference is observed at the membrane surface. In E. coli, the IM/OM ratio is slightly below 1. In this particular G. obscuriglobus cell, the OM has a surface of 13.7×106 nm2, while the IM is almost exactly three times bigger, with a surface of 42.7×106 nm2 (Table S2). Based on our observations, this ratio likely varies from cell to cell.
Although extensively developed, the membrane does not create individualized compartments within the cytoplasm. All membranes are connected and isolated compartments defined by membranes within the cell volume do not exist. The only cellular volumes are the cytoplasm and the periplasm (Figure 1; see Supplementary Movies 1 and 2, available at http://www.bork.embl.de/~devos/project/apache/htdocs/plancto/g3d/ [Text S1]). G. obscuriglobus membrane invaginations and derived membrane morphologies appear to be dynamic and possibly cell-cycle-dependent [9],[15]. We have acquired partial volumes for six cells and complete volumes for four cells with various morphologies and believe it is highly unlikely that the membrane completely encloses or forms isolated compartments during any stage of the cell cycle. We have always observed connected pseudo-compartments that we could follow in 3D. The changes in membrane organization and connection of the pseudo-compartments, as well as the variation of periplasm organization, can be followed in consecutive slices from the tomograms (Figures S2, S3, S4, S5, S6, S7, S8).
We observed five isolated clusters of DNA in one completely reconstructed cell and similar results in other cells (Figure 1; Figures S2, S3, S4, S5, S6, S7, S8). Some regions appeared more condensed than others, possibly due to differences in the replication or transcriptional status of the genetic material, which is unknown since the cell is in a dividing state.
Importantly, the genetic material is not restricted to a closed compartment with communicating pores—that is, in a “nucleus-like” organization as previously concluded [8],[10]. Membrane invaginations are sometimes found close to the DNA, but never enclose it completely. It is, however, easy to see why this can lead to false interpretations when looking at 2D images of single sections (Figure 1). 3D reconstruction rules these out. We have obtained similar tomograms for nine additional G. obscuriglobus cells with distinct overall membrane organization, and reached the same conclusion in each case (Figures S2, S3, S4, S5, S6, S7, S8). This conclusion is consistent with the presence of ribosomes in the cytoplasm surrounding the nucleoid. The DNA appears to be floating freely within the cytoplasm and does not obviously interact with the membranes, as in other bacteria.
Almost all planctomycetes reproduce by budding [7], instead of fission, the most common form of bacterial division. During the early phases of the budding process, the bud is mostly devoid of membranes and DNA [9]. Consistently, we imaged the bud where only one membrane sheet is present and we do not detect any DNA. This membrane sheet is ∼20 nm thick, similar to those observed in the mother cell. Furthermore, the IM of the bud is continuous with the IM of the mother cell, as can be observed at the neck of the bud (Figure 2; see Supplementary Movie 3, available at http://www.bork.embl.de/~devos/project/apache/htdocs/plancto/g3d/ [Text S1]), implying continuity for all membranes between the mother and daughter cell. The cytoplasm of the mother and daughter cells are connected by a narrow channel through the neck of the bud. At its narrowest point, the channel is roughly 30 nm wide, explaining why it has been missed in previous studies. Moreover, electron dense material is observed inside the periplasm around the neck, possibly suggesting the periplasm as an alternative route for the transfer of material between the mother and the daughter cells. However, this dense material requires further study and confirmation. As the bud enlarges, the neck of the bud opens up, with dimensions ranging from 80 to 375 nm (Figure S9). The genetic material, being cytoplasmic, can pass freely into the bud without interference from a “nuclear membrane.” This structure must somehow close during completion of cell division. The membrane organization in the mother cell appears to become more complex in the proximity of the budding neck (Figure 3; see Supplementary Movie 3,, available at http://www.bork.embl.de/~devos/project/apache/htdocs/plancto/g3d/ [Text S1]), possibly due to the process of membrane transfer to the bud. However, also here, there are no defined compartments and all membranes are in continuity with the IM. These results have important implications for our understanding of planctomycete division.
Crateriform structures have previously been reported as homogeneously distributed in G. obscuriglobus as opposed to other planctomycetes [19]. These structures are associated with depressions of the OM as can be seen from the side view perpendicular to the membrane (Figure 4). They have an opening of ∼35 nm and are uniformly distributed around the cell periphery, except in the mother cell within ∼1 µm diameter around the neck of the bud, with a density of between 50 and 100 crateriform structures per µm2 (Figure 1).
We have observed the presence of intracellular electron dense granules that are not enclosed by a membrane (Figure 1, depicted in dark blue). These are visible in roughly 50% of the cells that have been observed, and there is generally one per cell. X-ray micro-analysis confirmed that those granules are mainly composed of poly-phosphate (PolyP; Figure S10). PolyP can perform different biological functions, such as serving as an energy source for ATP synthesis [20].
Previously, there have been two related claims of the uniqueness for the planctomycetes and verrucomicrobiae compared to other bacteria [8],[21]. The first claim was the distinctive status of its membranes with the lack of an OM and thus of periplasm, the presence of an outermost cytoplasmic membrane, and an ICM. The second claim was linked to the organization of this ICM, stating that it divides the cytoplasm into compartments. Most importantly, the presence of a membrane surrounding the DNA in a structure related to the eukaryotic nucleus was postulated [7],[8],[21]. The uniqueness of the PVC membrane organization has recently been challenged by genomic information. This includes the presence of remnants of the dcw cluster, typical of Gram-negative bacteria, including peptidoglycan synthesis and cell division genes—for example, the otherwise ubiquitous FtsZ [13]. In addition, proteins typical for the OM and periplasm of Gram-negative bacteria are present in the genomes of PVC species [11],[12]. Using electron microscopy we can show, with confidence, that the second claim is not justified—that is, that the G. obscuriglobus cell plan is not compartmentalized and does not contain a nucleus-like structure. Importantly, our findings show that chromosomal DNA is not enclosed by a single membrane in G. obscuriglobus. This has important implications for our definition of eukaryotes and bacteria. Combined with the genomic evidence [11]–[13], this strongly supports the suggestion that the membrane organization of the PVC superphylum is not different from that of a Gram-negative bacterium, but an extension of it based on numerous invaginations of the IM. PVC species, with the possible exception of the anammox, do not have unique compartments; rather, their periplasm is extended by IM invaginations containing the two classical cell volumes, the periplasm and the cytoplasm. In addition this suggests that there is no spatial separation of transcription and translation by a membrane, as supported by the presence of ribosomes in close proximity to the DNA (Figure S11) [8],[15].
Our conclusion that the internal membranes of G. obscuriglobus cells result from a “simple” expansion of the periplasm by IM invaginations are likely to be applicable to other planctomycetes and other members of the PVC superphylum such as the verrucomicrobiae [22]. The presence of this extensive membrane organization in most PVC members suggests that, despite important variations, the ancestor of the PVC superphylum already had the precursor of this feature—that is, some capacity to invaginate its membranes [23]. Thus, our results expand our understanding of the bacterial cell plan without challenging it.
Until now, it was believed that the genomic material of G. obscuriglobus is enclosed in a membrane. This created a problem in explaining genome segregation when the cells undergo division and a satisfying solution has proven difficult to find [9]. Here we show that the duplicated DNA is free to transfer to the daughter cell without membrane interference, the only restriction being the width of the neck of the bud. Division is directly linked to the fact that all planctomycetes have lost the otherwise ubiquitous cell division protein FtsZ, while it is still present in Lentisphaera and Verrucomicrobia [13]. How PVC cells lacking FtsZ divide is unknown. So far, the only clue available is the detection of a GTPase-related novel cell division ring gene in the anammox bacteria, which is unrelated to FtsZ [24]. However, the anammox bacteria are divergent planctomycetes, and homologues of this protein have not been found in other planctomycetes, making it unlikely to provide a global answer to this question.
A particularity of the planctomycete membranes is that, with the exception of the anammoxosome, they have no assigned function [25]. Here we calculated that IM invaginations triplicate the surface of membranes relatively to the cell volume. Similar membrane extensions in photosynthetic bacteria are linked to the synthesis of energy. However, photosynthesis is not known to take place in G. obscuriglobus. Similar membrane extensions in eukaryotic cells include the mitochondria (linked to energy) and the ER/Golgi (linked to protein synthesis and secretion). It is interesting to compare the G. obscuriglobus endomembrane system with the eukaryotic endomembrane system. The eukaryotic rough ER and the outer membrane of the nuclear envelope are formed by membrane sheets that are coated with ribosomes, while the smooth ER is formed largely by membrane tubules devoid of ribosomes. In G. obscuriglobus cells, we have not observed tubules, only membrane sheets. As in the eukaryotic rough ER, ribosomes coat the sheets of the G. obscuriglobus endomembrane (Figure S11) [8],[15]. The function of the rough ER includes protein translocation into and through the ER lumen, as well as modification of newly synthesized secretory and membrane proteins. Smooth ER might be involved in lipid metabolism or Ca2+ signaling [16] and is specialized in sterol synthesis, a function also described in G. obscuriglobus [26]. Sterol modifies lipid fluidity and is thus linked to membrane organization [27]. The capacity to synthesize sterol is a feature previously considered as mainly eukaryotic. As opposed to the other sterol-producing bacteria, sterol synthesis in G. obscuriglobus is unlikely to be the result of a lateral gene transfer event from eukaryotes [26],[28]. Instead, it has been suggested that G. obscuriglobus could retain the most ancient remnants of the sterol biosynthesis pathway [26]. It seems likely that sterol synthesis in G. obscuriglobus is directly linked to the diversity of its extensive membrane organization. It is interesting to consider that protein composition can influence membrane bending [29]. In this respect it would be particularly interesting to investigate the membrane-bound proteins in planctomycetes.
Eukaryogenesis has long been a question of major interest to biologists. Although it is increasingly accepted that eukaryotes and archaea share a common ancestor, the nature of this ancestor (if it was already an archaea per se or an intermediate organism) is still debated [30]. The eukaryotic cell is differentiated from bacterial and archaeal cells by many features whose origins are for the most part still unknown. These features include the actin- and tubulin-based cytoskeleton, the mitochondria, the nuclear pore, the spliceosome, the proteasome, and the ubiquitin signaling system [31]. Features reminiscent of these are increasingly detected in prokaryotes, including the members of the PVC bacterial superphylum [23],[32]. Because PVCs display some features related to eukaryotes or archaea, including sterol production [26] and ether-linked lipids [6], it has been proposed that the PVC ancestor might have shared a sisterhood relationship with the ancestor of the eukaryotes and archaea [23],[32]. Other scenarios involving a relationship between PVC and eukaryotes have also been proposed [21],[33]. However, whether the PVC features are homologous or analogous to their eukaryotic or archaeal counterparts is still under discussion [34]. If there is no evolutionary relationship between PVC and eukaryotes, the complex endomembrane system of those bacteria highlights that endomembrane systems have evolved more than once. The complex endomembrane system of G. obscuriglobus is in direct contact with proteins displaying structural similarities to eukaryotic membrane coat proteins like clathrin or sec31 that sustain the eukaryotic endomembrane system [15],[35]. In addition, G. obscuriglobus endomembranes are involved in the otherwise strictly eukaryotic process of endocytosis [36]. These data reinforce the possibility of an evolutionary relationship between the eukaryotic and PVC endomembrane systems and suggest that the latter could represent intermediary steps in the development of the former from a “classical” Gram-negative bacterium [23],. Deeper characterization of the PVC endomembrane system is therefore of great interest.
In conclusion, our analysis reveals that the membrane organization in G. obscuriglobus is not fundamentally different from that of “classical” bacteria, but a complex variant of it. The next step is to link those observations with the development of this endomembrane system in cells using live imaging methods.
G. obscuriglobus cells were grown as previously described [15]. The cells were frozen in an HPM010 (Abra Fluid, Switzerland) high-pressure freezing machine and freeze substituted with either 1% Osmium tetroxide, 0.1% uranyl acetate, and 5% H2O and embedded in Epon or with 0.5% uranyl acetate and embedded in Lowicryl HM20. Thin (60 nm) and thick sections (250 nm) were placed on formvar-coated grids and post-stained with uranyl acetate and lead citrate. Thin sections were imaged on a CM120 Phillips electron microscope. For tomography, acquisition was done on a Technai F30 300 kv (FEI Company) microscope with dual axis tilt series (first axis from −60° to +60° with 1° tilt increment, second axis from −60° to +60° with 1.5° increment). We acquired nine serial sections and reconstructed them using fiducial gold particles with the weighted back projection algorithm. We joined consecutive serial sections using the etomo graphical user interface from IMOD (Boulder Laboratory for 3-D Electron Microscopy of Cells). We fully acquired four cells (eight to nine sections), two cells that are ∼75% complete (six sections), two cells that are ∼50% complete (four sections), and two cells that are ∼40% complete (three sections) (Table S1). The budding cell was modeled with IMOD and we traced the contours on at least every fifth slice over a range of 1,130 slices.
Tomograms have been deposited in the EMDB (http://www.ebi.ac.uk/pdbe/emdb/) under the accession numbers EMDB-2362 and EMDB-2363.
|
10.1371/journal.pbio.1001613 | Anthranilate Fluorescence Marks a Calcium-Propagated Necrotic Wave That Promotes Organismal Death in C. elegans | For cells the passage from life to death can involve a regulated, programmed transition. In contrast to cell death, the mechanisms of systemic collapse underlying organismal death remain poorly understood. Here we present evidence of a cascade of cell death involving the calpain-cathepsin necrosis pathway that can drive organismal death in Caenorhabditis elegans. We report that organismal death is accompanied by a burst of intense blue fluorescence, generated within intestinal cells by the necrotic cell death pathway. Such death fluorescence marks an anterior to posterior wave of intestinal cell death that is accompanied by cytosolic acidosis. This wave is propagated via the innexin INX-16, likely by calcium influx. Notably, inhibition of systemic necrosis can delay stress-induced death. We also identify the source of the blue fluorescence, initially present in intestinal lysosome-related organelles (gut granules), as anthranilic acid glucosyl esters—not, as previously surmised, the damage product lipofuscin. Anthranilic acid is derived from tryptophan by action of the kynurenine pathway. These findings reveal a central mechanism of organismal death in C. elegans that is related to necrotic propagation in mammals—e.g., in excitotoxicity and ischemia-induced neurodegeneration. Endogenous anthranilate fluorescence renders visible the spatio-temporal dynamics of C. elegans organismal death.
| In the nematode Caenorhabditis elegans, intestinal lysosome-related organelles (or “gut granules”) contain a bright blue fluorescent substance of unknown identity. This has similar spectral properties to lipofuscin, a product of oxidative damage known to accumulate with age in postmitotic mammalian cells. Blue fluorescence seems to increase in aging worm populations, and lipofuscin has been proposed to be the source. To analyze this further, we measure fluorescence levels after exposure to oxidative stress and during aging in individually tracked worms. Surprisingly, neither of these conditions increases fluorescence levels; instead blue fluorescence increases in a striking and rapid burst at death. Such death fluorescence (DF) also appears in young worms when killed, irrespective of age or cause of death. We chemically identify DF as anthranilic acid glucosyl esters derived from tryptophan, and not lipofuscin. In addition, we show that DF generation in the intestine is dependent upon the necrotic cell death cascade, previously characterized as a driver of neurodegeneration. We find that necrosis spreads in a rapid wave along the intestine by calcium influx via innexin ion channels, accompanied by cytosolic acidosis. Inhibition of necrosis pathway components can delay stress-induced death, supporting its role as a driver of organismal death. This necrotic cascade provides a model system to study neurodegeneration and organismal death.
| While mechanisms of cell death such as apoptosis are well characterized [1], less is known about the mechanisms of organismal death, particularly in invertebrate model organisms. Here we investigate organismal death in the nematode C. elegans, using a newly discovered, endogenous fluorescent marker of death.
One possibility is that organismal death results from a cascade of cell death. As first defined by Kerr et al. in 1972 [1], cell death has been viewed as taking two forms: controlled (apoptotic) or uncontrolled (necrotic). However, more recent elucidation of the mechanisms underlying necrotic cell death reveals that it too can be a regulated process [2]–[5]. Biochemical hallmarks of necrosis include calcium-mediated initiation, lysosomal membrane permeabilization (LMP), and activation of noncaspase proteases (calpains and cathepsins) [5]–[7].
Necrosis as a regulated process has been characterized mainly in mammalian neuronal models. Excitotoxic neuronal cell death occurs in response to overstimulation with the excitatory neurotransmitter glutamate (e.g., under conditions of ischemia or stroke) [7]. Sustained activation of glutamate receptors causes a cytosolic influx of extracellular Ca2+ [8]. Increased Ca2+ levels lead to cell death, largely through activation of associated proteases [9]. Moreover, Ca2+ may spread between cells via connecting gap junctions, and gap junction inhibition reduces ischemia-induced neurodegeneration [10],[11].
Through the study of ischemia-induced death in mammalian CA1 hippocampal neurons, Yamashima and co-workers identified the calpain-cathepsin cascade as an effector of necrotic cell death. Ischemia increases intracellular Ca2+ levels, which activate Ca2+-dependent cysteine proteases (calpains) [12]. These calpains cause lysosomal lysis, leading to cytosolic acidosis and the destructive release of lysosomal cathepsin proteases [13].
Many components of the calpain-cathepsin cascade are present in C. elegans, where necrotic cell death can be induced in neurons by mutations such as mec-4(u231) [14]. For example, mec-4-induced neurodegeneration requires the calcium-dependent calpains TRA-3 and CLP-1 and the cathepsins ASP-3 and ASP-4 [15].
LMP is a central event in the necrotic cascade, and the degree of LMP can influence the cellular decision to live or to die via necrosis or apoptosis [3],[5],[16]. In C. elegans, lysosomes are required for osmotic stress-induced necrotic death [17] and interventions that increase lysosomal pH can ameliorate mec-4(d)-induced neurodegeneration [18].
C. elegans intestinal cells contain both lysosomes and gut granules, which are large, melanosome-like lysosome-related organelles [19]. Under ultraviolet light, gut granules emit blue fluorescence, with maximal intensity at λex/λem 340/430 nm (Figure 1A–B) [20]. This fluorescence has been attributed to lipofuscin [21],[22], a heterogeneous, cross-linked aggregate of oxidatively damaged lipids and proteins. Lipofuscin accumulates with age in postmitotic mammalian cells and so has frequently been used as a biomarker of aging [23]–[25]. Lipofuscin composition is highly variable but can be identified by virtue of its autofluorescence [24]. If excited by UV light in vitro it emits blue fluorescence, which may reflect formation of fluorescent Schiff bases between carbonyl and amino groups [26],[27]. However, UV excitation of lipofuscin in vivo results in peak fluorescence in the 540–640 nm (orange-yellow) range [28].
Several observations have led to the suggestion that the fluorescent material in the C. elegans intestine is lipofuscin. Its fluorescence peak at λex/λem 340/430 nm is similar to that of lipofuscin in vitro, it is localized to the lysosome-like gut granules, and its levels increase in aging populations [20]–[22],[29]. It is often used as a biomarker of aging—for example, to verify that treatments that shorten worm lifespan do so by accelerating aging. The presence of lipofuscin in C. elegans would support the view that aging is caused by accumulation of molecular damage. Yet it remains possible that the fluorescent substance in gut granules is not lipofuscin. For example, studies of flu mutations causing altered gut granule fluorescence suggest that it corresponds to fluorescent tryptophan metabolites [30].
In this study, we describe how a reassessment of blue fluorescence in C. elegans led to the discovery of the phenomenon of death fluorescence (DF), a burst of blue fluorescence that accompanies death in C. elegans. We establish that both DF and gut granule fluorescence originate not from lipofuscin, but from tryptophan-derived anthranilic acid glucosyl esters. We then show that DF is generated by the calpain-cathepsin necrotic cell death pathway, and requires calcium signaling for organismal propagation. Finally, we show that inhibition of this pathway can protect animals against stress-induced death, supporting a role of systemic necrotic cell death in organismal death.
Lipofuscin is formed through accumulation of oxidatively damaged proteins and lipids [24]. For example, raised oxygen level (40% O2) increases lipofuscin levels in human fibroblasts [25]. To probe whether the blue fluorescent material in C. elegans gut granules (Figure 1A–B) is lipofuscin, we exposed them to normobaric hyperoxia (90% O2), and elevated iron levels. Both treatments significantly increased protein oxidative damage but neither increased blue fluorescence levels (Figure 1C–F). Elevated expression of hsp-4::gfp is indicative of the unfolded protein response [31], symptomatic of protein damage. Heat shock increased hsp-4::gfp expression but not blue fluorescence (Figure S1). These results imply that C. elegans blue fluorescence is not generated by oxidative damage, suggesting that it is not lipofuscin.
Like lipofuscin in mammals, mean fluorescence levels rise gradually with age in C. elegans population cohorts [20],[29]. However, population mean data do not address heterogeneity in the fluorescence of individual worms. This concern was raised by a previous study [20], as follows. Aging worms can be classed according to their degree of motility: class A animals move normally, class B animals move more slowly, and class C animals do not move away when touched, and are near to death [32]. Notably, blue fluorescence levels did not differ significantly between class A and B, and only increased in class C worms [20]. This suggests that blue fluorescence levels in worms increase only as they approach death.
To test this directly, fluorescence levels of individually cultured, wild-type C. elegans in situ on nematode growth medium (NGM) agar plates were examined at intervals throughout life (DAPI filter; λex/λem 350/460 nm). As animals approached death (as indicated by reduced movement), time-lapse imaging was used to capture fluorescence changes during death. This revealed that fluorescence levels in individual animals change little until immediately prior to death. A striking and sudden ∼400% increase in fluorescence level then occurs, coinciding with cessation of movement (i.e., death) (Figure 2; Video S1). This rise begins at ∼2 h prior to death, and then fades by ∼6 h after death (Figure 2B–C).
Blue fluorescent bursts are not only associated with death from old age, but were also induced by killing—for example, by placing a heated worm pick on the agar adjacent to the worm (Figure 3A–B) or by freeze-thaw or low pH (Figure S2A,B). Hot pick-induced killing also caused fluorescent bursts in young adults of both sexes, and in larvae (Figures 3A–B, S2C; Video S2), and in the nematodes C. briggsae and Pristionchus pacificus (Figures 3A, S2D–E). In both aged and killed worms, fluorescence distribution also changed during death, from punctate to diffuse, and eventually spread from the intestine to other tissues (Figures 2B, S3). We named this phenomenon death fluorescence (DF).
Next we characterized the spatiotemporal dynamics of DF in C. elegans, as a potential marker of cellular and organismal death. DF typically originates in the anterior-most cells of the intestine (the int1 cells). It then spreads rapidly along the intestine in an anterior to posterior wave (Figures 3C–D, S3). In adults a second focus of fluorescence sometimes appears in the mid-body (Figure S3). When a hot pick was applied to the animals' tails (rather than near the head), DF initially only arose locally and did not spread from posterior to anterior (Figure 3D) but only, eventually, from anterior to posterior (unpublished data). This suggests that the anterior intestine represents an organismal weak point in C. elegans, where a local crisis in homeostasis can trigger a DF wave.
Several types of autofluorescence with different spectral properties have been described in C. elegans [29],[33],[34] (see Figure S4A for an overview of worm fluorescence). Using a more sensitive detection system [33], we examined the dynamics of blue, green, and red fluorescence over life and aging-induced death. Again, no significant age-increase in blue fluorescence was seen (Table S1). The much weaker green and red fluorescence did increase significantly with age (Figure S5; Table S1), and all three forms of fluorescence increased during death (Figure S5; Table S1). These results imply the presence of multiple fluorophores in C. elegans.
We then investigated the chemical identity of the blue fluorophore, using glo-1(zu437) (gut granule loss 1) mutants that lack gut granules [19]. glo-1 animals showed little blue fluorescence either during life or death due to aging or thermal injury (Figures S4B, S6A–B), implying that gut granule fluorescence and DF have a common origin. Blue fluorescence was present in aqueous extracts of N2 (wild type) worm homogenates. HPLC analysis revealed one major peak with fluorescence at λex/λem 340/430 nm in N2 but not glo-1 extracts (Figure S6C–D). We therefore used glo-1 mutants as a negative control for chemical identification of the blue fluorophore, using 2D NMR-based comparative metabolomics [35]. This approach allows identification of compounds whose production depends on a specific genetic background without extensive chromatographic fractionation.
Comparison of 2D NMR spectra acquired for the N2 and glo-1 extracts revealed several groups of signals that were much reduced or absent in the glo-1 spectra (Figure S7). The most differentially expressed compounds were the anthranilic acid glucosyl ester (angl#1 in Figure 4A) and N-glucosyl indole (iglu#1) and their corresponding 3′-phosphorylated compounds (angl#2 and iglu#2). N2 but not glo-1 extracts also contained smaller quantities of free anthranilic acid. These structural assignments were confirmed via high-resolution mass spectrometry and synthesis of authentic samples of anthranilic acid glucosyl ester and N-glucosylindole (Tables S2, S3, S4). Moreover, fluorescence spectra for angl#1 and worm blue fluorescence were alike (Figure S8).
Anthranilic acid (AA) is synthesized from L-tryptophan (Trp) by action of the kynurenine pathway (Figure 4B), and has previously been observed in C. elegans [36],[37], but neither angl#1 nor angl#2 have previously been identified in animals. AA derivatives show fluorescence at λex/λem 340/430 nm [38]. The HPLC retention times of these AA derivatives matches those detected in the initial HPLC analysis of the N2 extract (Table S4). The indole glucosides iglu#1 and iglu#2, also not previously reported in animals, did not emit blue fluorescence (unpublished data).
To verify the identity of the blue fluorophore, we next used a genetic approach. The first step in the conversion of Trp to AA is catalyzed by tryptophan 2,3-dioxygenase (TDO). The C. elegans gene tdo-2 (C28H8.11) encodes a putative TDO (Figure 4B) [36]. tdo-2(RNAi) suppressed both gut granule blue fluorescence and DF (freeze-thaw) (Figure 4C–D). RNAi or mutation of flu-2 (kynureninase) also reduced DF, while inactivating kynurenine 3-monooxygenase by kmo-1(RNAi) or flu-1 mutation increased DF, all as predicted (Figure S9). Both tdo-2(RNAi) and mutation of flu-2 greatly reduce AA levels [37]. We also tested whether exogenous AA is sufficient to cause blue gut granule fluorescence. tdo-2(RNAi) worms were incubated in a range of solutions of synthetic anthranilic acid (Sigma). Incubation with 5 mM AA restored gut granule fluorescence to a wild-type level (Figure S10A–B). We conclude that gut granule fluorescence and DF emanate from AA.
Why do fluorescence levels increase at death? One possibility is that AA levels increase at the point of death. To test this we compared DF in tdo-2(RNAi) worms with restored gut granule fluorescence and in L4440-treated control worms. Interestingly, the magnitude of DF was not reduced in the tdo-2(RNAi) worms (Figure S10C–D). This strongly implies that the DF burst is not the result of increased AA levels.
An alternative possibility is that concentration of AA within gut granules results in quenching of fluorescence. To probe this idea we employed the dye uranin, whose green fluorescence is partially quenched at low pH [39]. Treatment of wild-type worms with uranin led to punctate green fluorescence in the intestine that co-localized with blue gut granule fluorescence (Figure S11A). Thus, uranin accumulates within gut granules. Killing after uranin treatment caused a burst of green fluorescence in wild-type worms, but not glo-1(zu437) mutants without gut granules (Figure S11B–C). This supports the view that the burst of fluorescence at death is caused by dequenching of AA and uranin fluorescence due to increased pH upon release from the acidic milieu of the gut granules. It also shows that uranin is an excellent marker for loss of integrity of membranes bounding acidic compartments (e.g., lysosomes and lysosome-like organelles).
The presence of blue fluorescence within lysosome-related organelles (i.e., gut granules) and the central role of LMP in multiple instances of necrotic cell death suggested that DF might be generated by necrosis. If correct, then inhibition of necrosis might be expected to suppress DF. To test this we used freeze-thaw induced death, which is convenient for rapid and accurate quantitation of DF. The calpain-cathepsin necrotic cascade (Figure 5A) is involved in C. elegans neurodegeneration [15]. Neuronal necrosis requires Ca2+ release from ER stores. Mutations in the ryanodine and inositol-1,4,5-triphosophate receptors, and the ER Ca2+ binding protein calreticulin all suppress neuronal necrotic cell death [17],[40]. Each of these mutations, unc-68(e540), itr-1(sa73) and crt-1(bz29), respectively, also significantly reduced DF (Figure 5B). During cellular necrosis, increased intracellular Ca2+ can activate calpains (Ca2+-dependent cysteine proteases). The calpain TRA-3 is required for neuronal necrotic cell death in C. elegans [41]. tra-3(e1107) reduced DF (Figure 5B).
The necrosis cascade requires lysosomal lysis for cytosolic acidification and cathepsin release. In worms, the vacuolar proton-translocating ATPase (V-ATPase), which mediates lysosomal acidification, is required for necrosis [18]. We therefore tested two hypomorphic V-ATPase mutants, vha-12(ok821) and unc-32(e189), but these did not significantly reduce DF (Figure 5B). Finally, we asked whether cathepsins promote DF. cad-1(j1) mutants have 10%–20% of wild-type cathepsin D activity [42] and asp-4 encodes an aspartyl protease: both genes are required for necrosis [15],[18]. Again, both cad-1(j1) and asp-4(ok2693) reduced DF (Figure 5B).
ced-3, ced-4, and ced-9 are required for apoptosis in worms [43]. To test whether apoptotic cell death machinery contributes to DF, we examined DF in killed ced-3(n717), ced-4(n1162), and ced-9(n1950) mutants. ced-9 mutants showed no decrease in DF, ced-4 mutants only a slight decrease, and ced-3 actually showed an increase in DF (Figure S12A). In other negative controls (ftn-1, mdl-1, and rol-6) no effects on DF levels were seen either (Figure S12B). Thus, mutations that inhibit ER Ca2+ release, and calpain and cathepsin activity both inhibit necrosis and lower DF. We conclude that attenuation of elements of necrosis reduces DF. This suggests not only that cellular necrosis generates DF, but also that cellular necrosis occurs during organismal death. This in turn suggests the possibility that necrotic cell death contributes to organismal death.
The spread of DF through the intestine is reminiscent of Ca2+ wave transmission in the intestine during the defecation cycle [44]. Innexins (invertebrate gap junction proteins) are required for the defecation cycle as they create Ca2+ channels between adjacent intestinal cells, and the innexin INX-16 is required for Ca2+ transmission during defecation [44]. We asked whether Ca2+ signaling might play a role in DF wave propagation.
Upon being killed (by freeze-thaw), inx-16(ox144) mutants showed reduced DF levels, and (by a hot pick) a failure in DF wave propagation (Figures 6A–B, S13A; Videos S3, S4). Note that DF dynamics appear largely independent of the mode of killing (Figures 3, S2). Moreover, an intestinally expressed calcium reporter revealed increased Ca2+ levels during death (by oxidative stress; t-BOOH) (Figure 6C). This increase occurred first in the anterior and then in the posterior intestine (Figures 6C, S13B; Video S5), consistent with a wave of Ca2+ influx during death. The death-induced Ca2+ wave was blocked by inx-16(ox144), as observed for DF (Figure 6D). Thus, Ca2+ signaling is required for, and precedes, the spread of DF. This suggests that during death an anterior to posterior wave of Ca2+ influx drives a wave of necrosis that leads to DF.
Cytosolic acidosis and LMP also typically occur during necrotic cell death, and we therefore asked whether they accompany DF. To test for cytosolic acidosis, we used an intestinally expressed pH-sensitive reporter, pnhx-2::pHluorin [45],[46]. Upon killing (with t-BOOH), pH in the intestinal cytosol dropped from ∼pH 7.4 to 6.6 (Figure 6E, Video S6) and, again, cytosolic acidosis occurred first in the anterior intestine before spreading to the posterior (Figures 6E, S13C–D).
Next we used uranin (green) and the lysosomotropic dye lysotracker (red) to examine gut granule membrane integrity during organismal death. Killing (with t-BOOH) resulted in a burst of uranin fluorescence, and a loss of punctate green and red fluorescence that coincided with DF (Figure 7A). These changes were inhibited by inx-16(ox144) (Figure S14). The punctate red staining took slightly longer to decay than the green, likely reflecting residual staining of lysosomal membranes with lysotracker but not uranin. Thus, both cytosolic acidosis and LMP occur in intestinal cells at death. This provides further evidence that DF is generated by necrotic cell death.
If systemic necrosis contributes to organismal death, then its inhibition should prevent death. To test this we examined the effect of inhibiting necrosis on death induced by aging or stress. In most cases, inhibiting necrosis did not prevent death due to aging: necrosis mutants were either normal lived or short lived (Figure S15). The exception was inx-16, which was long lived; however, the slow growth and starved appearance of this strain suggests that its longevity may be caused by dietary restriction.
By contrast, inhibition of each point in the necrosis pathway (calcium release, calpains, lysosomal acidification, cathepsins, and innexins) was able to delay death induced by lethal osmotic stress (Figure 5C), although not all mutants showed resistance. Most necrosis mutants also showed resistance to lethal thermal stress (Figure S16). Moreover, it was previously reported that inhibition of necrosis can delay infection-induced death [47]. However, necrosis mutants showed little protection against death induced by oxidative stress (t-BOOH) (unpublished data). These findings suggest that some stressors cause death in C. elegans by triggering systemic necrosis.
One possibility is that the release of AA from gut granules stimulates intestinal necrosis. To test this we first examined the effect of removing AA by tdo-2(RNAi) on resistance to heat stress, and found that resistance was increased (Figure S17A–C). However, it was previously shown that tdo-2(RNAi) enhances proteostasis and lifespan by increasing Trp levels [37]. To test whether tdo-2(RNAi) protects against heat stress by reducing AA, we asked if replenishment in tdo-2(RNAi) worms would suppress their stress resistance, but it did not (Figure S17C). Moreover, AA supplementation of N2 worms did not reduce heat stress resistance. Also in a range of mutants with altered AA levels there was no correlation with resistance to either thermal or osmotic stress (Figure S17D–E). Thus, heat stress resistance resulting from tdo-2(RNAi) is not caused by reduced AA but may instead reflect increased Trp levels. These results imply that DF does not promote intestinal cell necrosis but, rather, is a bystander phenomenon (or epiphenomenon).
We also tested the effect of insulin/IGF-1 signaling on DF. Mutation of the daf-2 insulin/IGF-1 receptor increases lifespan, and this effect requires the daf-16 FoxO transcription factor [48]. In both freeze-thaw- and aging-induced death, the mutation daf-2(e1370) markedly reduced DF (Figure S18). This could imply that systemic necrosis is attenuated in daf-2 mutants. daf-16(mgDf50) modestly increased DF during death from old age only (Figure S18).
In this study, we have shown that conserved mechanisms underpinning neuronal necrosis can also contribute to organismal death. In so doing we provide the first insights into the last biological events in the life history of C. elegans: those leading to its final demise. Identification of an endogenous fluorescent marker of death led us to discover a calcium-generated wave of necrotic cell death that occurs during, and can contribute to, organismal death (Figure 7B). We have also chemically defined the source of the endogenous blue fluorescence that is a salient characteristic of C. elegans.
Evidence presented here implies that during death in C. elegans, the intestine, the largest somatic organ, undergoes a stereotyped process of self-destruction involving an intra- and intercellular cascade of cellular necrosis. The mechanisms involved are similar to those active in the propagation of cellular necrosis in mammals. In worms, necrotic propagation requires the innexin INX-16, while in mammals connexin (mammalian gap junction proteins) inactivation reduces ischemia-induced neurodegeneration [10]. Thus, the C. elegans intestine is a potential new model for understanding the propagation of necrotic cell death, and its prevention.
Previous studies of the cellular necrosis pathway have largely focused on neurodegeneration, in mammals and C. elegans. Our findings imply similar action of this pathway in the worm intestine. However, generation of DF appears to be restricted to the intestine, and is not detectable in necrotic mec-4(d) neurons (unpublished data).
Our results imply that intestinal self-destruction by systemic necrosis occurs during both stress- and aging-induced death. However, only in stress-induced death did inhibition of systemic necrosis prevent death. This suggests that while lethal stress causes death by inducing systemic necrosis, aging causes death by a number of processes acting in parallel, likely including systemic necrosis (given that it destroys a major organ). Here there are potential parallels in human aging: estimations of the likely upper limits of human longevity have calculated that removal of a major age-related disease (e.g., cardiovascular disease, cancer) would cause only small increases in lifespan [49]. This is because multiple pathologies act in parallel to increase age-related mortality.
A feature of intestinal necrosis is its origin in the anterior int1 cells. This suggests that the unusual vulnerability of these cells to necrotic death might represent a breaking point within organismal homeostasis; analogously, in humans localized failure (e.g., in the heart or kidneys) can cause rapid organismal death. The existence of an anterior to posterior (A-P) Ca2+ wave is unexpected, given that the defecation-associated Ca2+ wave previously characterized in the intestine flows in the opposite direction, from posterior to anterior [44]. How the A-P Ca2+ wave is specified is unknown. One possibility is that extracellular Ca2+ levels are elevated near the anterior intestine, creating vulnerability to necrosis [50].
In this study we have defined a new phenomenon, death fluorescence, which may be useful in future as a marker of death in lifespan assays. While DF means that blue fluorescence cannot be used as a biomarker of aging we confirm that red fluorescence does increase with age. Moreover, tdo-2(RNAi) can be used to abrogate blue intestinal fluorescence to aid the viewing of expression of intestinally expressed fluorescent reporters.
The anthranilic acid glucosyl ester angl#1 and its corresponding 3′-phosphorylated derivative angl#2 account for both death and gut granule-associated fluorescence. glo-1 mutants lack both forms of fluorescence and AA derivatives, and inhibition of the kynurenine pathway blocks both forms of fluorescence, establishing that this blue fluorescence is not lipofuscin. Whether lipofuscin accumulation occurs during aging in C. elegans remains an open question. However, our finding that blue fluorescence is not lipofuscin removes one reason for believing that aging in C. elegans is caused by accumulation of stochastic molecular damage. The kynurenine pathway that generates gut granule and DF is also involved in mammalian neurodegeneration, and has recently been shown to regulate protein folding homeostasis in C. elegans [37],[51].
During organismal death, AA fluorescence increases as a consequence of necrosis. That DF can occur in the absence of AA synthesis (Figure S10C–D) implies that the burst is not a consequence of synthesis of additional AA. One possibility is that AA fluorescence within gut granules is partially quenched, perhaps due to low pH and/or increased concentration. In this scenario, loss of gut granule membrane integrity causes rapid dequenching of AA fluorescence, leading to the burst. In a similar fashion, dequenching of uranin fluorescence upon gut granule permeabilization leads to a burst of green fluorescence (Figure S11). It remains to be investigated whether cellular necrosis in other organisms leads to increased AA fluorescence, but increases in blue fluorescence accompanying cell death have been reported—e.g., in budding yeast [52] and hepatocytes [53].
Despite their prominence, C. elegans gut granules are organelles whose function has yet to be established. The finding that they contain large quantities of AA further adds to the mystery. What is all this anthranilic acid for? Possibilities include protection against UV irradiation, or against pathogen invasion into the intestine. Notably, AA can be cytotoxic; for example, 3-hydroxyanthranilic acid can induce cell death in lymphocytes [54] and neurons [55], and AA can inhibit growth of bacterial pathogens (e.g., Legionella pneumophila) [56]. This might explain its presence in multiple species of soil nematodes.
The presence of a mechanism, systemic necrosis, that brings about organismal death in C. elegans raises questions about its evolutionary origin. Could such an organismal self-destruct mechanism serve as an adaptation? When food is limiting, gravid hermaphrodites typically die with multiple embryos in their uterus, which hatch internally and consume their mother's corpse (“bagging”). Potentially, this improves the mother's fitness by increasing survival of her genetically identical offspring [57]. One possibility, then, is that systemic necrosis enhances fitness by aiding efficient transfer of nutrients from mother to offspring during bagging. Alternatively, systemic necrosis may be the nonadaptive product of antagonist pleiotropy, or a quasi-program [58],[59]. By this view, elements of the necrosis cascade contribute to early life fitness, while systemic necrosis is an unselected, deleterious consequence of their action under lethal stress or as a result of aging.
Standard C. elegans strain maintenance and genetic manipulations were used [60]. All strains were grown at 20°C on NGM plates seeded with E. coli OP50 as food source unless otherwise specified. N2 (Bristol) was the wild-type.
The necrosis mutants are as follows: ZB1028 crt-1(bz29) V, CB540 unc-68(e540) V, JT73 itr-1(sa73) IV, CB4027 tra-3(e1107) eIs2137 IV, CB189 unc-32(e189) III, RB938 vha-12(ok821) X, RB2035 asp-4(ok2693) X, and PJ1 cad-1(j1) II. The apoptosis mutants are as follows: MT1522 ced-3(n717) IV, MT4770 ced-9(n1950) III, and MT2547 ced-4(n1162) III. The strains used for calcium measurements are as follows: KWN190 pha-1(e2123ts) III, him-5(e1490) V; rnyEx109 [pKT67 (Pnhx-2::D3cpv); pCL1 (pha-1(+)], KWN26 pha-1(e2123ts) III; him-5(e1490) V; rnyEx006 [pIA5nhx-2 (Pnhx-2::pHluorin); pCL1(pha-1(+)]. The strains used for pH measurements are as follows: KWN385 inx-16(ox144) I; pha-1(e2123ts) III; him-5(e1490) V rnyEx006 [pIA5nhx-2 (Pnhx-2::pHluorin); pCL1 (pha-1(+)]. Other: BA671 spe-9(hc88) I, CB1002 flu-1(e1002) V, CB1003 flu-2(e1003) X, GH10 glo-1(zu437) X, EG144 inx-16(ox144), GA91 ftn-1(ok3625) V; GA1200 mdl-1(tm311) X, GA200 wuEx41 [rol-6(su1006)], RB784 nkat-1(ok566) X, SJ4005 zcIs4 [hsp-4::GFP].
Worms were either imaged in situ on NGM plates or anaesthetized on agar pads on glass slides. Images were acquired using an Orca digital camera (Hamamatsu) and a Leica DMRXA2 microscope. Blue fluorescence was observed through a DAPI filter cube (λex/λem 300–400 nm/410–510 nm) (ET DAPI, set 49000, Chroma). Green fluorescent protein (GFP) fluorescence was observed through a GFP filter, (λex/λem 450–490 nm/500–550 nm) (Endow GFP Bandpass, 41017 Chroma). Images were acquired using the application Volocity Acquisition (Improvision, Perkin-Elmer). Fluorescence was quantified by manually tracing around worm peripheries using an Intuos graphics tablet (Wacom), and measuring mean pixel density using Volocity Quantitation. Worm fluorescence was estimated as the mean pixel density of the worm image area minus the pixel density of the image background.
Worms were killed in three ways, detailed below, all of which result in bursts of blue fluorescence of similar magnitude. Heat killing was used to observe DF dynamics in individual animals in situ on NGM plates. This approach allows lethal stress to be applied near the head or tail, and allows observation of spatial changes in fluorescence, but is relatively difficult to quantitate. Killing by oxidative stress was used for higher resolution microscopy for which it was necessary to view worms under cover slips. Freeze-thaw assays of worms in microtitre plates were used for accurate quantitation of DF to compare genotypes cohorts of worms, and for acquisition of whole spectrum excitation/emission scans.
Synchronous populations of L4 animals were transferred to NGM plates seeded with E. coli OP50 and containing 50 µM FUdR (24°C). One-day-old adults were then rinsed off the plates and washed using S buffer. For each strain, three replicate aliquots of 100 µl worm suspensions were loaded in black microtiter plates (Greiner). Aliquots resulted in ∼1 mg alkali-extracted protein, determined by standard BCA assay (ThermoScientific). The 10-nm step fluorescence emission spectra of living worm suspensions were measured upon excitation at 250–450 nm (10 nm intervals) using a microplate reader (Spectramax Gemini XS, Molecular Devices). Worms were then killed by freeze-thaw and fluorescence measurements repeated. All data shown are averages of three technical replicates, corrected by a blank measurement, and normalized by protein content of the worm suspensions.
A 55 mM stock of anthranilic acid (Sigma) was prepared by dissolving AA solid in PBS at 55°C and then diluted further in PBS. Worms were incubated in AA solution for 3 h at 20°C in a 96-well microtitre plate with constant shaking.
L4 or 1-d-old adult worms were incubated for 2 h in 475 µL M9 plus 10 µL Lysotracker Red DND-99 (Life Technologies, USA) plus 15 µL 20 mg/mL uranin. Worms were then washed 5× in 1 mL M9 and left to feed on OP50-seeded NGM plates for 30–60 min. They were then placed on a 2.5% agarose pad prepared on a glass slide between two cover slip spacers (to avoid squashing of adult worms between the pad and the cover slip), and mounted in 0.2% levamisole under a cover slip. Worms were then imaged under at 100× magnification on a DM RXA2 upright microscope (Leica, Germany) every 10 s for up to 1 h using Volocity software (PerkinElmer Inc., USA). We added 15 µL of Luperox TBH70X tert-butylhydroperoxide solution (Sigma Aldrich, Germany) using a pipette, assuring even dispersal of the liquid between the slide and the cover slip within the first minute of imaging. DF appeared within the first 15–30 min and the time-lapse acquisition was stopped once the blue fluorescence wave had propagated from head to tail.
The flu-2, kmo-1, nkat-1, and tdo-2 clones were acquired from the Ahringer library, and the inserts confirmed by DNA sequencing. E. coli HT115 were transformed with the clone and fed to animals as described [68].
Lifespans of synchronized population cohorts were measured as previously described [72] at 20°C with 15 µM FUdR topically applied.
Lifelong fluorescence measurements were acquired from individual animals. Data were normalized to time of each animal's death and average fluorescence level acquired at hours prior to death. Data measuring groups of animals used mean data, with Student's t tests performed to check for statistical significance. All mean data were repeated at least in triplicate. Lifespan and thermotolerance data were analyzed by log-rank for significance; osmotic stress by one-way ANOVA of means. All fluorescent images represent averages seen.
|
10.1371/journal.pbio.1001369 | Systematic Dissection of Roles for Chromatin Regulators in a Yeast Stress Response | Packaging of eukaryotic genomes into chromatin has wide-ranging effects on gene transcription. Curiously, it is commonly observed that deletion of a global chromatin regulator affects expression of only a limited subset of genes bound to or modified by the regulator in question. However, in many single-gene studies it has become clear that chromatin regulators often do not affect steady-state transcription, but instead are required for normal transcriptional reprogramming by environmental cues. We therefore have systematically investigated the effects of 83 histone mutants, and 119 gene deletion mutants, on induction/repression dynamics of 170 transcripts in response to diamide stress in yeast. Importantly, we find that chromatin regulators play far more pronounced roles during gene induction/repression than they do in steady-state expression. Furthermore, by jointly analyzing the substrates (histone mutants) and enzymes (chromatin modifier deletions) we identify specific interactions between histone modifications and their regulators. Combining these functional results with genome-wide mapping of several histone marks in the same time course, we systematically investigated the correspondence between histone modification occurrence and function. We followed up on one pathway, finding that Set1-dependent H3K4 methylation primarily acts as a gene repressor during multiple stresses, specifically at genes involved in ribosome biosynthesis. Set1-dependent repression of ribosomal genes occurs via distinct pathways for ribosomal protein genes and ribosomal biogenesis genes, which can be separated based on genetic requirements for repression and based on chromatin changes during gene repression. Together, our dynamic studies provide a rich resource for investigating chromatin regulation, and identify a significant role for the “activating” mark H3K4me3 in gene repression.
| Chromatin packaging of eukaryotic genomes has wideranging, yet poorly understood, effects on gene regulation. Curiously, many histone modifications occur on the majority of genes, yet their loss typically affects a small subset of those genes. Here, we examine gene expression defects in 200 chromatin-related mutants during a stress response, finding that chromatin regulators have far greater effects on the dynamics of gene expression than on the steady-state transcription. By grouping mutants according to their shared defects in the stress response, we systematically recover known chromatin-related complexes and pathways, and predict several novel pathways. Finally, by integrating genome-wide changes in the locations of five prominent histone modifications during the stress response with our functional data, we uncover a novel role for the “activating” histone modification H3K4me3 in gene repression. Surprisingly, H3K4 methylation appears to act in conjunction with H3S10 phosphorylation in the repression of ribosomal biosynthesis genes. Repression of ribosomal protein genes and ribosomal RNA maturation genes occur via distinct pathways. Our results show that steady-state studies miss a great deal of important chromatin biology, and identify a surprising role for H3K4 methylation in ribosomal gene repression in yeast.
| Packaging of eukaryotic genomes into chromatin has wide-ranging effects on gene transcription in eukaryotes [1]. There are two major ways in which cells modulate nucleosomal influences on gene expression. ATP-dependent chromatin remodeling machines utilize the energy of ATP hydrolysis to disrupt histone-DNA contacts, often resulting in nucleosome eviction and changed nucleosomal location or subunit composition [2]. In addition, the highly conserved histone proteins are subject to multiple types of covalent modification, including acetylation, methylation, phosphorylation, ubiquitination, SUMOylation, and ADP-ribosylation. These covalent histone modifications often occur during the process of transcription, and in turn have many effects on transcription. Moderately well-understood effects of histone modifications include epigenetic gene silencing, control of transcript structure via repression of “cryptic” internal promoters, control of splicing, and transcriptional activation [3]–[7]. Altogether, there are myriad interactions and feedback loops between chromatin state and transcription. At present, the effect of most modifications on transcription is unclear, even for reasonably well-characterized ones.
A large number of systematic genome-wide analyses have been carried out to characterize the complex interplay between chromatin regulation and gene transcription. Genome-wide mapping studies [8],[9] show that modification patterns are correlated with gene structure and gene activity levels. Genome-wide mRNA profiling has been used for over a decade to identify transcriptional defects in chromatin mutants [10]. A recent tour de force from the Holstege lab examined the effects on gene expression of deleting each of 174 different chromatin regulators [11]. Proteomic studies characterize many of the protein complexes that play a role in chromatin regulation [12],[13]. Systematic genetic interaction profiling (using growth rate as a phenotype) has been used to identify chromatin complexes, and to delineate interactions between chromatin pathways [14]–[16]. Importantly, most of these genomic screens have been carried out in steady-state conditions, typically in yeast actively growing in rich media.
In contrast, single gene studies suggest that chromatin regulators have important roles in dynamic processes that are masked at steady-state. For instance, deletions of the histone acetylase Gcn5 or the histone chaperone Asf1 have little effect on the eventual induction of PHO5 by phosphate starvation, but both of these deletions cause significant delays in PHO5 induction kinetics [17],[18]. Similarly, mutation of H3K56, whose acetylation plays a role in histone replacement, delays PHO5 induction by slowing nucleosome eviction upon gene activation [19]. Similar results hold for other classic model genes, such as the galactose-inducible GAL genes [20]. Because steady-state gene expression in mutants is subject to widespread compensatory or homeostatic mechanisms, we reasoned that analysis of mutant responses to a stressful stimulus would help reveal direct functions of transcriptional regulators. Thus, the dynamics of response to stimuli should uncover the transcriptional roles of histone-modifying enzymes and other chromatin regulators. We chose diamide stress in yeast as a model system, as it has been shown to involve a rapid, dramatic reorganization of the yeast transcriptome with 602 genes induced more than 2-fold and 593 genes repressed [21].
Here, we carried out a time course of diamide stress in 202 yeast mutants and characterized gene expression changes at 170 selected transcripts (Figure S1A–C). Importantly, analysis of thousands of genome-wide mRNA profiling studies shows that genes typically are co-regulated in coherent clusters [22]–[24], meaning that the behavior of the majority of co-regulated clusters can be captured by analyzing ∼100–200 transcripts. For example, analyzing mutant effects on six ribosomal protein genes suffices to capture the majority of mutant effects on all ∼250 of these genes.
We find that the majority of chromatin regulators have greater effects on gene induction/repression kinetics than they do on steady-state mRNA levels, confirming that dynamic studies can identify unanticipated functions for chromatin regulators. We show that grouping deletion mutants with similar gene expression defects identifies known complexes, and that joint analysis of histone mutants and deletion mutants associates many histone-modifying enzymes with their target sites. In addition to known relationships between chromatin regulators, we identify a number of novel connections, including a previously unknown connection between H3K4 and H3S10 modifications. We further carried out genome-wide mapping of five relevant histone modifications during the same stress time course (Figure S1D–E). By combining functional data with genome-wide mapping data, we identify a key role for Set1-dependent H3K4 methylation in repression of ribosomal biogenesis genes. H3K4 methylation and H3S10 phosphorylation are both required for full repression of ribosomal protein genes (RPG) and of genes involved in rRNA maturation (RiBi), but repression of RPGs and RiBi genes operate via two distinct pathways downstream of these histone marks. Thus, the classic “activating” mark H3K4me3 in fact serves primarily to facilitate repression in budding yeast under multiple stress conditions. Together, these data provide a rich multi-modal view on the role of chromatin regulators in gene induction and repression dynamics, and suggest that understanding the myriad roles of chromatin structure in gene regulation on a genome-wide scale will require extending mutant analyses to kinetic studies.
We used nCounter technology [25] to carry out genome-scale gene expression profiling. Briefly, this technology utilizes hybridization of labeled oligonucleotides in a flow cell to directly count individual RNA molecules, without any enzymatic steps, for several hundred RNAs in yeast extracts. For this experiment, we focused on gene expression during a stress response time course (using the sulfhydryl oxidizing agent diamide). We used whole genome mRNA abundance and Pol2 localization data from prior diamide exposure time courses [21],[26], along with a compendium of prior whole genome mRNA analyses and transcript structure analyses in various mutants [23],[24], to select 200 probes reporting on 170 transcripts (142 genes, of which 30 had two sense probes, as well as another 28 antisense transcription units) that capture the majority of the different patterns of gene expression behavior in this stress. Using this probeset, we measured transcript abundances over a 90-min time course of diamide exposure (Figure 1). Experimental replicates are highly reproducible (Table S1), and these data provide a detailed kinetic perspective on gene expression dynamics during the diamide stress response (Figure 1A–D).
We carried out identical time course experiments for 119 deletion strains for chromatin regulatory genes and for 83 mutants in histones H3 and H4 [27], covering the majority of individual K→R, K→Q, K→A, R→K, and S→A mutants, and several H3 and H4 N-terminal tail deletions. For most mutants, we analyzed mRNA abundance at four time points (t = 0, 15, 45, and 90 min) as these time points capture the major phases of the diamide stress response. Figure 1A–D show example data for wild-type yeast and three mutants in the HDA1/2/3 complex. The entire dataset, comprising ∼1,000 experiments carried out for 202 mutant strains, is shown in Figure 1F–G, with mutant time courses clustered according to the similarity between their effects on gene expression across all four time points (see also Table S1).
Close inspection of the cluster in Figure 1G (Table S1) revealed that many of the gene expression defects observed in these mutants were only observed during the stress response, but not before stress. This is apparent in Figure 1A and 1D, where many more genes exhibit different levels between wild-type and hda mutants at 15 and 45 min of the stress response than at t = 0 (midlog growth). These differences include both kinetic delays in gene induction/repression and defects in the extent of gene regulation (see below). To determine the generality of this phenomenon, we determined the distribution of mutant effects on RNA abundance at each of the four time points in the stress response. Many more significant gene expression changes relative to wild-type occur at 15 and 45 min (∼10% of probe/mutant pairwise interactions) after diamide addition than at t = 0 (∼3.5% of pairwise interactions, Figure 1E). As the yeast acclimate to the stress environment (e.g., at t = 90), the transcriptome reaches a new steady-state where we see fewer large mutant effects, although there are still more changes than at t = 0. Thus, consistent with observations from classical model genes such as PHO5, we find that chromatin mutants have much more extensive effects during changes in transcription than during steady-state conditions.
We sought to identify major classes of gene expression defect in various chromatin mutants, as a first step in eventually linking chromatin transitions to the genetic requirements for different chromatin regulators. Immediately apparent in Figure 1G (red boxes) are two large groups of mutants with opposing behaviors with respect to the stress response—mutants that appear to be transcriptionally “hyper-responsive” to diamide stress and “hypo-responsive” mutants that exhibit blunted stress responses. These two major classes of mutants are also captured by principal component analysis (PCA) of our dataset. Here, the first principal component, which explains 30% of the variance in the dataset, corresponds to hyper- and hypo-responsive mutants (Figure S2A–B). Interestingly, not all genes induced or repressed during diamide stress were affected by hyper- or hypo-responsive mutants. Genes whose induction was most affected by hyper-responsive mutants, for example, tended to be those with highly nucleosome-occupied promoters in YPD (Figure S2C) [28]–[30].
Hypo-responsive mutants to diamide stress included a number of expected mutants, including deletion mutants lacking the general stress transcription factors Msn2 and Msn4, or with compromised coactivator complexes such as Swi/Snf or SAGA. Hyper-responsive mutants, conversely, included a number of histone deacetylases such as Hda1/2/3. Beyond acetylation/deacetylation, hyper-responsive and hypo-responsive mutants included a variety of deletions known to affect histone turnover and/or occupancy. Several of these factors have previously been shown to affect bulk H3 turnover (Rtt109, Cac2/Rtt106, Htz1, Hat1, Rsc1, and Nhp10; [31]–[37]) or histone levels/occupancy (Rtt109, Yta7, Rtt106, Cac2, Spt21, H3K42Q; [38]–[40]). Interestingly, we noticed that among those histone mutants that decreased the stress response program, the subset of those mutations that are located in the globular domains of H3/H4 (as opposed to the N-terminal tails) are all situated at histone-DNA interfaces (Figure S2D), which we speculate could affect nucleosomal stability and/or replacement dynamics. Taken together, these results support a model in which many chromatin regulators have roles on global transcriptional responsiveness resulting from their overall effects on nucleosome stability.
Our RNA abundance measurements provide a population-averaged view of chromatin effects on gene expression, but hide a great deal of stochastic behavior that can be revealed by single-cell approaches. For example, RNA data on hyper-responsive mutants come from many thousands of cells, meaning the mechanistic basis for stress hyper-responsiveness is unknown. Do hyper-responsive mutants have a greater fraction of cells exhibiting diamide-driven gene induction (as might be observed if gene induction depends on cell cycle stage and mutants exhibit cell cycle delays), or do all individual cells exhibit greater amplitude responses?
We therefore extended our studies to include single cell analysis of protein expression using high throughput microscopy of GFP-tagged proteins in several key mutants. As protein stability significantly confounds measures of gene repression, we focused on four diamide-induced genes, and examined each reporter in wild type and in nine deletion mutants. We conducted time-lapse microscopy of yeast cells during the diamide response (Figure 2A, Methods). After detecting cells (average n = 120 for each of 40 strains, two biological replicates), we quantified the temporal profile of GFP intensity for each cell. Figure 2B shows the median intensity as a function of time for one reporter in wild-type and several mutants. Importantly, we found excellent agreement between defects in protein induction in various hypo- and hyper-responsive mutants and the corresponding nCounter RNA measurements (Figure 2C).
In general, we noted that GFP induction in individual cells followed a sigmoid-like curve consistent with a window of stress-increased protein production followed by a gradual return to baseline production levels. This behavior is consistent with a simple model in which there is a time window of diamide-induced gene transcription, followed by gradual mRNA decay. We implemented a simple mathematical model with cells transitioning from low expression to high expression and back, with a constant rate of mRNA production during the open window (Materials and Methods). This model is clearly oversimplified—each parameter covers multiple processes—but provides very good fit to the measured intensity profiles (Figure 2D). Fitting the model for each cell, we can estimate the transcriptional time windows for individual cells as well as the rate of protein production during this time and examine the variability in the timing and speed of transcriptional response in a genetically homogenous population of cells (Figure 2E).
We then used the extracted parameters for individual cells to determine whether hyper- or hypo-responsiveness corresponded to a change in the responsive fraction of cells, a population-wide change in promoter open time, and so forth. In general, we found that most mutants did not affect the fraction of cells responding to diamide. The fraction of cells exhibiting diamide induction of GFP was 87%±3% across all 40 strains, and no strain differed from wild-type by even 10% of cells responding. Notably, we found that different hyper-responsive mutants could act at different stages in gene expression. For example, deletion of YTA7, which is involved in histone gene transcription and affects nucleosome occupancy [40],[41], leads to accelerated promoter opening during diamide stress, whereas deletion of HDA2 predominantly affects GFP production rate rather than promoter opening (Figure 2F,G). Together, these results independently validate our RNA measurements, confirm that RNA changes are reflected in protein abundance, and show that, for the nine mutants analyzed, mutant effects on transcriptional response occur in the majority of cells rather than reflecting changes in the fraction of diamide-responsive cells.
Beyond the major groups of mutants that affect overall stress responsiveness and likely report on global histone occupancy/dynamics, we observed a wide variety of gene expression effects that were specific to smaller sets of mutants. For example, the white box in Figure 1G highlights the well-understood gene expression changes that occur in mutants related to the Sir heterochromatin complex—repression of mating-related genes secondary to the pseudodiploid state of these mutants [7]. To systematically group mutants according to their gene expression phenotypes, we calculated the correlations between the changes (relative to wild-type) in stress response in each mutant and clustered mutants according to these correlations (Figure 3A, Table S2, Materials and Methods). We kept histone mutants and deletion mutants separate to allow more intuitive interpretation of clusters.
Grouping deletion mutants by this method recovers a great deal of known chromatin biology, validating our approach. In general, mutants in different subunits of known chromatin complexes exhibit similar defects in gene expression, indicating shared function. Most white boxes in Figure 3A highlight a subset of clear examples, including the grouping of subunits of the Sir complex, the HDA1/2/3 complex, COMPASS, Cac2/Rtt106, Set3C, and the Ino80 complex. Furthermore, several pathways were recovered. The histone variant H2A.Z (encoded by HTZ1) was linked to components of the Swr1 complex responsible for H2A.Z incorporation [42]–[44], the H3K4 methylase Set1 was linked to the H2B ubiquitin ligase Bre1 whose activity is required for K4 methylation [45], and the H3K36 methylase Set2 was linked to Eaf3, the binding partner for H3K36me3 [16],[46],[47].
In addition to known chromatin regulatory complexes and pathways, our results also suggest a number of hypotheses for novel chromatin pathways. For example, we find strong correlations between gene expression defects in mutants lacking the H4K16 acetylase Sas2 and those lacking the proline cis/trans isomerase Cpr1. Similarly, our results link the H3K36 demethylase Rph1 with ATP-dependent remodeler Chd1, suggesting the possibility that H3K36 methylation regulates Chd1 in budding yeast, an idea that finds support in prior studies showing that H3K36 mutants and chd1 mutants have similar genetic interactions in vivo [48].
Analysis of histone mutations revealed similar structure. We observe two larger clusters that correspond to hyper- and hypo-responsive mutations (Figure 3A, yellow boxes), as well as many smaller groups. Many of these groups are comprised of several mutations in the same residue (e.g., all three mutations in H3K36 are tightly clustered together) or in the same tail (e.g., H3 tail delete and simultaneous K->Q/R mutations in H3 tail lysines 4, 9, 14, 18, and 23). Many other groups of histone mutants were unanticipated and may identify functionally relevant nucleosomal surfaces [27] or novel examples of histone crosstalk [49]. Below, we explore the relevance of one such novel connection, between H3K4A and H3S10A mutants.
Many of the connections between chromatin regulatory genes observed here also can be observed in systematic genetic interaction profiles, or in gene expression studies carried out in midlog growth conditions [11],[14]. A unique aspect of our study is the joint analysis of gene deletion mutants with histone point mutants. Many of the strongest correlations between deletion and histone mutants correspond to known enzyme-substrate and modification-binding partner relationships. For example, gene expression defects resulting from deletion of the H3K36 methylase Set2 were most strongly correlated with the defects in H3K36R and H3K36Q mutants, and with the H3K36me3-binding protein Eaf3 (Figure 3A).
Analysis of multiple different mutations of the same lysine residue can provide insight into the biochemical function of modifications at this residue. While both K→R and K→Q mutants disrupt modification-specific binding by proteins (e.g., bromo- and chromo domain proteins), they differ in their charge. Indeed, lysine mutants for which K→R and K→Q mutants exhibited similar gene expression defects tend to occur at lysines with well-characterized modification-specific binding partners (e.g., Eaf3, Sir3). In contrast, lysines for which K→R and K→Q mutants had opposing effects on gene expression often were known acetylation substrates, although we counterintuitively observe that for these lysines the K→R mutations were generally correlated with deletions in histone deacetylases (Figure S3).
To systematically identify relationships between chromatin factors, we identified significant correlations between mutants (Materials and Methods), recovering for example the Set2→H3K36→Eaf3 pathway (Figure S4A–B, Table S3). Data for all correlations above a threshold significance are visualized in a network view in Figure 3B to show not only connections within strongly connected pathways but also connections between pathways. Other known relationships recovered this way included the association between Set1 and H3K4, and the association between the Sir complex and H4K16 (Figure S4).
Furthermore, we found that Cac2, a CAF-1 subunit, and Rtt106, histone chaperones that were strongly correlated with one another, exhibited transcriptional effects most related to the H4K91R mutant (Figure 3A). H4K91 acetylation is a little-studied modification reported to occur on newly synthesized histones [50], and in systematic genetic interaction studies, H4K91R and mutations in the assembly-related lysine H3K56 exhibited similar genetic interactions [27]. We therefore hypothesize that H4K91 acetylation might affect chromatin assembly by CAF-1 or Rtt106. Other connections have no obvious literature precedent—the HMG protein Nhp6a, which plays a role in nucleosome positioning and dynamics at promoters [51],[52], was correlated with the H3R8K mutation (Table S2)—and thus represent potentially novel connections between histone residues and either modifying enzymes or binding partners. Below, we follow up specifically on one such observation, the surprising linkage between H3K4 methylation mutants and the H3S10A histone mutant.
Our data show that joint analysis of histone mutants with related gene deletion mutants can systematically link histone-modifying enzymes with their substrates, as well as modification-specific binding proteins to the relevant modified histone residue (Tables S2 and S3).
We next sought to understand why only particular genes were affected by mutants in various chromatin regulators. One of the central questions in chromatin regulation is why broadly localized histone marks appear to have extremely localized effects on gene expression? In other words, given that H3K4me3 occurs at nearly all +1 nucleosomes, why do set1Δ mutants exhibit relatively minor [11],[53] gene expression changes? Our functional results suggest that many transcriptional effects of chromatin mutants are masked at steady-state by feedback mechanisms, but can be uncovered during dynamic changes in gene expression. To address the relationship between histone mark occurrence and function in a dynamic context, we therefore extended our studies by carrying out genome-wide mapping of several histone modifications (Tables S4 and S5) during a six time point diamide stress time course (t = 0, 4, 8, 15, 30, and 60 min). We focused these experiments on two relatively well-characterized modifications: H3K36me3 and H3K4me3, and related marks H3K14ac, H3S10P, and H3R2me2a. Our mapping data for unstressed yeast are concordant with known aspects of modification localization patterns from either prior genome-wide mapping efforts [8],[9] or related studies (Figure 4A, Figure S5).
Given the surprising correlation between H3K4A and H3S10A mutants (Figure 3A, Figure S4), we focused on how the histone modifications H3K4me3 and H3S10P change genome-wide during diamide stress. As noted above, H3K4me3 occurs at the 5′ ends of transcribed genes, and genes induced during the stress response gained H3K4me3 over time, as expected (Figure 4C, Figure S5B). H3S10P, which had not been mapped genome-wide in yeast, is most strikingly localized to ∼20 kb surrounding yeast centromeres (Figure S5G), consistent with its pericentric localization by immunofluorescence in mammalian cells [54]. However, we also noted that H3S10P on chromosome arms was heterogeneous, and localized to coding regions with a pattern opposite to that of H3/H4 turnover [31],[35]. H3S10P is depleted from the 5′ ends of genes, and over coding regions anticorrelates with transcription rate (Figure 4B). Furthermore, during the stress response H3S10P levels increase over repressed coding regions, and decrease over induced genes, indicating that the anticorrelation between H3S10P and transcription is dynamic (Figure 4C).
Overall, many of the chromatin changes over stress-activated or repressed genes fit expectations. At stress-activated genes, promoter H3K4me3 levels increased while H3K36me3 increased over gene bodies. However, we also observed several unexpected dynamic behaviors (e.g., increasing H3K36me3 over the promoters of many stress-responsive genes). Furthermore, H3K14, whose acetylation scales with transcription rate during midlog growth [8],[9], was only deacetylated at a small subset of repressed genes during diamide stress, with most repressed genes exhibiting surprisingly minimal changes in H3K14ac (see below).
Most curiously, we found that H3K4me3 levels increase at the 5′ ends of a substantial number of diamide-repressed genes during their repression (Figure 4C, yellow box). Not only do these genes gain H3K4me3, they also gain H3S10P, and as noted above H3K4 mutants and H3S10 mutants exhibit similar gene expression defects (Figure 3A, Figure S4). Thus these marks are linked both functionally and in terms of dynamic localization changes. Curiously, the H3K4methylase Set1 and one of the H3S10 kinases, Ipl1, also share the nonhistone substrate Dam1 [55], indicating a more general connection between H3K4 and H3S10 based on shared nonhistone substrates for their modifying enzymes. It is unlikely that the gene expression defects observed here stem from nonhistone substrates of these enzymes as the gene expression changes are observed in histone point mutants as well as modifying enzyme deletions, but the connection is curious nonetheless.
Below, we attempt to connect the changes in H3K4me3 and H3S10P localization with the functional effects of relevant mutants. Are the genes that are misregulated in K4 and S10 mutants the same genes that exhibit dynamic changes in these marks during stress?
Set1 methylates H3K4 to create a gradient over coding regions from K4me3 at the 5′ end to K4me1 at the 3′ end, and this methylation pattern correlates with transcription rate during midlog growth ([8],[9], Figure S5B). The correlation between H3K4me3 and transcription rate leads to this mark being referred to as an “activating mark,” yet set1Δ mutants exhibit few gene expression defects in midlog growth, and in fact increasing evidence points to a primarily repressive role for K4 methylation in yeast. set1Δ mutants exhibit increased basal expression of repressed genes such as PHO5 [56],[57], and moreover exhibit widespread defects in repression of sense transcription by antisense transcripts [58]–[61].
We noted in our initial gene expression dataset that set1Δ and related mutants showed defects in repression of ribosomal protein (“RPG”) and ribosomal biogenesis (“Ribi”) genes (Table S1). We therefore extended these results to whole genome mRNA profiling, finding that the major gene expression defect in set1Δ mutants during diamide stress is a failure to adequately repress RPG and Ribi genes (Figure 5A). This result is interesting in light of prior observations that Set1 is required for full repression of the rRNA repeats [62],[63] during steady-state growth (when a subset of rDNA repeats are silenced), and shows that Set1 plays a general role in repression of all aspects of ribosomal biogenesis. Notably, although some snoRNA genes are found in RPG introns, we observed Set1 effects on the majority of ribosomal protein genes, most of which do not carry snoRNAs in their introns, indicating that the observed effect is not a consequence of Set1's known effects on termination at snoRNA genes [64].
Overall, deletion of SET1 resulted predominantly in diminished repression of ribosome-related genes, with very few large effects on diamide-activated genes (Figure 5B, Figure S6A–C). Importantly, loss of Set1 had a distinct effect on ribosomal gene repression from that observed in “hypo-responsive” mutants. Comparison of a given mutant's effects on overall gene repression to its effects on ribosomal gene repression identifies Set1-related and Sir2-related mutants as having specific defects in ribosomal gene repression (Figure S6D, see also below).
We next asked whether Set1's role in ribosomal repression was specific to diamide stress. We therefore assayed gene expression of our 200 probes in wild type and set1Δ yeast responding to another stress response, heat shock, or responding to nutrient deprivation signals induced by the small molecule rapamycin [65],[66]. Each of these stress responses exhibited different repression kinetics of the RPG genes, yet in all three stresses set1Δ strains suffered defects in RPG repression (Figure 5C). Thus, Set1 appears to act fairly generally as a repressor of ribosomal biogenesis under suboptimal growth conditions.
Comparing set1Δ effects on mRNA abundance with modification mapping data, we noted that many genes repressed in a Set1-dependent manner were often associated with stress-induced gains in H3K4me3 and H3S10P at their 5′ ends (Figure 6A, Tables S4 and S5). Focusing on the most highly Set1-dependent diamide-repressed genes revealed two clearly distinct clusters based on chromatin changes at the genes' 5′ ends (Figure 6B). Remarkably, we found that ribosomal protein genes (RPGs) were “paradoxically” associated with dramatic gains in H3K4me3 at their 5′ ends, as well as gains in H3S10P. The changes in H3K4me3 and H3S10P were strongest at the +1 nucleosome but occurred throughout the promoters (Figure S7A and analysis not shown). Conversely, non-RPG ribosomal biogenesis (Ribi) genes exhibited similar increases in H3S10P, but modest increases in 5′ H3K4me3. Instead, these genes were among the relatively few diamide-repressed genes associated with decreases in H3K14 acetylation. Importantly, these specific modification changes are quite specific for the gene classes in question. RPGs encompass the majority of genes gaining H3K4me3 during diamide repression, whereas Ribi genes provide the majority of cases with H3K14 deacetylation during repression (Figure S7B–C).
The distinct chromatin changes observed over RPG and Ribi genes during repression suggested that Set1-dependent repression of these genesets might operate via distinct pathways downstream of H3K4 methylation. We therefore sought to identify additional players in the pathways involved in repression of RPG and Ribi genesets. For each mutant assayed in our nCounter dataset, we compared the effects on diamide repression of RPGs to the effects on Ribi repression (Figure 7A). In general, mutants had similar effects on both gene classes, with globally hypo-responsive mutants such as H3K42Q failing to repress both RPGs and Ribi genes to similar extents. Intriguingly, we found a handful of mutants (several are shown in Figure 7B) with substantially different effects on RPG and Ribi repression: most notably, mutants in the RPD3L complex (e.g., sap30Δ, pho23Δ) exhibit defective repression of Ribi genes, yet have no effect on RPG gene expression during diamide stress. These results are consistent with prior genome-wide studies in yeast which found that repression of Ribi genes in response to heat shock, H2O2, or rapamycin was defective in the absence of RPD3L [66],[67]. Together, our results suggest that H3K4me3-dependent recruitment or activation of RPD3L (presumably via the PHD finger in Pho23; [57]) is required for Set1-driven repression of Ribi genes, whereas an alternative Set1-dependent pathway, potentially operating via Sir2 (see Discussion), represses RPGs.
Together, these results provide strong evidence for two distinct Set1-dependent gene repression pathways in yeast (Figure 7C–D). Both sets of genes require intact H3K4 and H3S10 for full repression. However, stress-dependent repression of ribosomal biogenesis genes not only requires H3K4 methylation but also is dependent on the RPD3L repressor complex (which likely is recruited to these genes via the PHD finger in Pho23), and these genes specifically are deacetylated during stress. In contrast, repression of ribosomal protein genes is delayed relative to Ribi repression, is largely unaffected by loss of the RPD3L complex, and furthermore these genes are associated with increased levels of the “active mark” H3K4me3 during repression.
Whereas mutants in our dataset that specifically affect Ribi gene repression suggested a clear mechanistic hypothesis regarding Set1's effects on these genes (H3K4me3-dependent recruitment of RPD3L), we observed relatively few mutants that disproportionately dampened RPG repression relative to Ribi repression. How does H3K4 methylation affect RPG expression? Our first hypothesis, that RPGs could be repressed via H3K4me2-dependent recruitment of the repressive Set3C [58], was ruled out by the observation that mutants in Set3C components do not affect RPG repression (Figure 7A–B).
An emerging concept in Set1 regulation of yeast genes is that Set1 is required for repression of transcription by trans-acting antisense RNAs [59],[60],[61]. Of the 28 antisense transcripts in our probeset, only a handful were significantly expressed above background during diamide stress. For example, in YPD we find that the BDH2 sense transcript is expressed at low levels, but its antisense is highly expressed (Figure 8A). Upon diamide treatment, the sense transcript is induced and the antisense is concomitantly repressed. We observed a widespread anticorrelation between mutant effects on sense versus antisense transcripts (Figure 8B). Notably, H3K4 methylation mutants expressed the antisense transcript at lower levels than wild-type in YPD, and conversely hyperinduced the sense transcript during diamide stress. Similar results were observed for the YTP1 sense/antisense pair (Table S1). In contrast, Set1 had little effect on the level of the antisense transcript at the ARO10 locus, but instead was required for full induction of the ARO10 sense transcript in diamide (Figure 8C). Thus, in both cases Set1 primarily affects one transcript in a sense/antisense pair, with the specific transcript being regulated in each case possibly reflecting the fact that the ARO10 sense does not overlap the TSS of its antisense [61], whereas for BDH2 the competing transcripts each overlap each other's TSS.
Many of the mutants that affect sense/antisense ratios also affected RPG expression, raising the question of whether expression of these classes of genes might be linked. However, using strand-specific q-RT-PCR we have been unable to find any evidence for antisense transcription of RPGs under our conditions (unpublished data). Instead, based on the curious observation that antisense-mediated repression of PHO84 in trans requires that the antisense RNA overlap with the PHO84 UAS [61], we wondered whether some aspect of RNA structure might affect Set1-dependent repression in yeast. Notably, 73% of ribosomal protein genes in yeast carry introns, and these introns are generally much longer than non-RPG introns [68]–[70]. Moreover, RPG introns tend to have more stable secondary structures, both in absolute predicted ΔG of folding and in ΔG per base pair (analysis not shown).
We therefore asked whether RPG introns might contribute to stress-dependent repression of these genes. Figure 8D shows change in expression of RPL16A, both for the native gene and for a chromosomally integrated intron-lacking version of RPL16A. Notably, diamide repression of this gene was far weaker in the absence of the native intron. We obtained similar results for three of four intronless strains tested, although one intronless gene exhibited hyperrepression in response to diamide stress (Figure 8E). We next asked whether the intronic contribution to RPL16A repression was in the same pathway at Set1-mediated H3K4 methylation. As expected from Figure 5, we confirmed that RPL16A repression was dramatically diminished in the absence of Set1. Notably, loss of the native intron had little additional effect on repression beyond that observed in the set1Δ mutant (Figure 8D), suggesting that Set1-dependent repression of RPGs is somehow connected to their long, potentially highly structured introns. Given the recent observation that RPG introns can affect RNA levels not only of their host genes but also of many paralogs [71], it will be interesting in future studies to determine if such trans-acting gene regulation by introns is Set1-dependent.
We report here a systematic functional genetic analysis of the roles for chromatin regulators and histone mutations in the dynamics of stress response in yeast. We analyzed the effects of 202 chromatin-related mutants on diamide-dependent transcriptional dynamics for 170 RNAs. Importantly, we generalize prior single-gene observations that many chromatin regulators have broader effects on gene induction/repression kinetics than during steady-state growth. Furthermore, we combined these data with whole-genome mapping data for five histone modifications. Together, this dataset provides a rich multidimensional resource for generating hypotheses regarding chromatin biology.
A major observation in this study is that chromatin mutants have far greater effects on gene expression during gene induction/repression than they do on steady-state gene expression in midlog growth. These results are consistent with observations using classic model genes such as PHO5 and GAL1-10, and suggest that a great deal of chromatin biology is obscured at steady-state due to homeostatic mechanisms that compensate for deleted chromatin regulators. These results also suggest that chromatin transitions may often be rate-limiting during transcriptional responses to the environment.
By grouping mutants according to their effects on gene expression, we were able to systematically construct chromatin regulation pathways. These analyses complement similar studies in which deletions are grouped by the similarity of their genetic interaction profiles [14], or according to their gene expression defects in YPD [11]. Importantly, by analysis of gene expression changes during a stress response, we uncover additional interactions that are not observed in YPD. For example, at t = 0 Rph1 and Rpd3 effects on gene expression are highly correlated (R2 = 0.51), but during diamide stress they exhibit opposite effects on gene expression (R2 = −0.38). These correlations may reflect stress-specific interactions between the factors in question, or they may reflect pathways that operate generally under all conditions but whose effects are only observed during dynamic reprogramming of transcription. Furthermore, by jointly analyzing histone point mutants and deletions of chromatin regulators, we correctly assign many histone-modifying enzymes to their known substrates. We uncover a small number of novel connections here (such as that between Nhp6a and H3R8), but did not find clear connections for predicted histone-modifying enzymes such as Set4. We believe the failure to identify a clear substrate for Set4 likely reflects the low levels of this protein in haploid yeast [72], although we cannot rule out that this enzyme primarily methylates nonhistone substrates, that it functions redundantly with another factor, or other possibilities.
As noted in the introduction, the disconnect between global localization of histone marks and their specific, local importance is a key mystery in chromatin at present [73]. Here, we carried out genome-wide histone modification mapping to enable comparisons between the functional effects of chromatin mutants with the locations of relevant marks in a dynamic context.
Overall, our modification mapping data were consistent with extensive prior knowledge about the modifications studied. However, we discovered a number of surprising aspects of histone modification changes during stress responses. For example, we found that H3K36 methylation, typically found over coding regions, was highly dynamic over promoters, suggesting a much more widespread role for this mark in regulation of open reading frames by cryptic transcription [74] than has been previously appreciated. We are currently following up on the role for promoter-localized H3K36me3 in gene regulation. Similarly, while H3K14ac is correlated with transcription rate of genes during steady-state growth (Figure 4, Figure S5; [8],[9]), we found that the majority of genes changing expression in response to diamide stress did not gain or lose H3K14ac in predictable ways. Among repressed genes, deacetylation occurred primarily at genes encoding ribosomal biogenesis factors (Figure S7).
Together, these results highlight the difficulty in understanding the function of specific histone modifications. Clearly, not every gene marked with H3K36me3 requires Set2 for expression. Understanding this phenomenon, often termed “context dependence” of histone modifications, is necessary for a deeper understanding of the biological roles for chromatin regulators [73],[75].
Our systematic analyses uncovered several surprising aspects of H3K4 methylation during diamide stress. As noted above, H3K4 methylation is associated with gene transcription at steady-state and thus is considered an “activating” mark, yet in budding yeast most evidence points towards H3K4 methylation as a repressive mark. Loss of Set1 results primarily in derepression of midsporulation and other repressed genes during midlog growth [11],[53],[56],[57]. Set1 appears to broadly play a role in control of sense/antisense ratios [59],[60],[76], as perhaps most clearly demonstrated in the case of the antisense transcript for PHO84 that is capable of repressing sense transcription in trans [61]. We extend these results, identifying additional sense/antisense pairs regulated by Set1 (Figure 8A–C). It remains unclear, however, what distinguishes sense/antisense pairs subject to Set1 regulation from those that are unaffected by Set1, although in general, transcripts that overlap the promoter of their opposing partner are more likely to regulate the other transcript [61].
Here, we dramatically extend the list of Set1 effects on transcription, finding that during diamide stress Set1 is required for full repression of genes involved in ribosomal biosynthesis. This effect is unlikely to result from nonhistone substrates of Set1, as it is recapitulated in H3K4A mutants. Together with prior observations demonstrating a role for Set1 in rDNA silencing [62],[63] during midlog growth in YPD, our results therefore identify Set1 as a general repressor of ribosomal biogenesis, with roles in repressing rRNA, ribosomal protein genes, and ribosomal biogenesis genes. Importantly, set1Δ mutants have no effect on RPG and Ribi gene transcription during active growth in YPD (Figure 5A and [11],[53]), when ribosomal genes are being extremely highly transcribed, meaning that the identification of Set1 as a broad repressor of ribosomal biogenesis could only be observed under stress conditions as in this study. Conversely, since a subset of rDNA repeats are repressed even during active growth, this enabled the discovery of this aspect of Set1 function in early midlog studies.
Based on chromatin mapping and on functional analysis of all 202 mutants, we find that distinct mechanisms operate in the repression of RPGs and the Ribi regulon. Ribi genes, but not RPGs, are not effectively repressed in mutants affecting the RPD3L complex. Moreover, Ribi genes are specifically associated with loss of H3K14ac during diamide stress, but exhibit little to no gain in H3K4me3. These results are consistent with a known pathway in which dephosphorylation of the transcriptional repressors Dot6 and Tod6 leads to RDP3L recruitment to Ribi promoters [67],[77], with binding of RPD3L component Pho23 to H3K4me3 contributing to either RPD3L recruitment or activity [57]. The molecular details underlying the presumptive “bivalent” recruitment/activation of RPD3L by Dot6/Tod6 and H3K4me3 remain to be elucidated. In striking contrast, we find no role for RPD3L in repression of RPGs (Figure 7A–B). Consistent with this, published gene expression profiles from rpd3Δ mutants in several stress conditions (diamide was not studied) reveal a far greater effect of Rpd3 loss on repression of Ribi genes than RPGs [67]. This raises the question of how Set1 contributes to RPG repression.
RPG repression was not accompanied by deacetylation of H3K14, and instead we observed that RPG promoters paradoxically gain H3K4me3 during diamide repression. It is not immediately apparent what aspect of RPGs makes them subject to Set1-regulated repression, but it is well known that RPGs represent roughly half (102 of 250) of all intron-containing genes in budding yeast. Given the emerging picture that Set1 affects gene regulation by antisense RNAs associated with promoters, we speculated that ribosomal introns and promoter-associated antisenses might share in common some unusual form of locally tethered RNA secondary structure. Ribosomal introns are longer than most other introns in yeast, and generally have much greater predicted RNA secondary structure than other introns. Consistent with the idea that RPG introns might contribute to Set1-dependent repression, we found that in several cases replacement of the native intron-containing RPG with its cDNA (in the native chromosomal context) abrogated repression of the RPG by diamide (Figure 8D–E), suggesting that either the intronic RNA or the corresponding DNA plays a role in Set1-dependent repression of some RPGs. As for the downstream repressor, we are currently investigating the hypothesis that RPG repression could be mediated by the Sir heterochromatin complex. Genome-wide mapping studies show that Sir3 binds to RPGs [78]–[80], and we show here that sir mutants and set1 mutants have similar effects on RPG repression (Figure 7B). Moreover, in vivo selection studies for RNA-based repressors in yeast found a surprisingly high fraction of tethered RNAs could repress a reporter gene in a Sir-dependent manner [81], suggesting that structured RNAs might recruit the Sir complex in a manner analogous to the role for lincRNAs in repressing metazoan genes by Polycomb recruitment [82],[83]. In this view, we hypothesize that ribosomal introns might serve in some sense as “domesticated” lincRNAs.
Alternative hypotheses include the possibility that the act of splicing per se could play a role in Set1-dependent repression of RPGs (whose splicing is mechanistically distinct from non-RPG splicing; [69],[70]) or that Set1 affects RPG expression by regulating Nrd1-dependent transcriptional termination [64]. Intriguingly, we observed in published NET-Seq data on Pol2 localization [84] that set1Δ mutants exhibit lower 5′ peaks of Pol2 over RPGs during midlog growth, with increased Pol2 levels downstream (Figure 8F). This decrease in 5′ Pol2 localization is consistent with the possibility that Set1 regulates RPGs via effects on transcriptional termination. It is also consistent with an alternative mechanism in which Set1 regulates Pol2 pausing at the 5′ ends of RPGs and that the delayed Pol2 in wild-type cells either allows intron folding or simply keeps 5′ RNA physically tethered near the promoter.
Future studies will be required to determine whether RPGs and antisense-regulated genes do in fact operate via a common mechanism, and to identify whether any specific aspects of RNA or RNA/DNA structures play a role in recruiting repressive complexes.
Taken together, these data show that chromatin regulators have far more effects on changes in gene expression than on steady-state transcription. Our approach allows systematic linking of chromatin regulators in complexes and of histone-modifying enzymes with their substrates. Finally, we show that joint analysis of functional gene expression data with localization data leads to novel insights even into extensively studied histone modifications such as H3K4me3.
Two collections of yeast mutants were used. Histone point mutants were described in [27], and were a kind gift from Jef Boeke. Diploid heterozygous deletion mutants with the SGA reporter developed by [85] were sporulated and selected to generate haploid Mata knockouts [86]. Yeast knockout mutants were grown on selective media (SC–Leu–His–Arg dropout mix+G418 200 mg/L+ L-Canavanine 6 mg/L) for two rounds to select for the deletion and for haploids, then used in the nCounter assays.
For Nanostring nCounter assays, each strain was grown in 80 mL YPD to mid-log phase (OD600 between 0.4 and 0.6) in a shaking 30°C waterbath. At “time zero” cells were treated with 1.5 mM diamide (D3648, Sigma), and 3 mL samples of culture each were taken at t = 0 (immediately prior to diamide addition), 4, 8, 15, 22.5, 30, 45, 60, and 90 min. Samples were immediately fixed with 4.5 mL cold (−45°C) methanol and kept in dry ice-ethanol bath throughout the time course. Cells in each sample were pelleted at 4,000 rpm for 2 min at 4°C, washed with 10 ml nuclease-free water, resuspended in 1 mL RNAlater solution (Ambion), and stored at −80°C.
For the histone modification mapping time course, six flasks each of 400 mL BY4741 cells were grown in YPD to mid-log phase shaking at 220 rpm at 30°C. Cells were treated with 1.5 mM diamide at time zero. At t = 0, 4, 8, 15, 30, and 60 min, cells were fixed by 1% formaldehyde, followed after 15 min by quenching with 125 mM glycine. Cells were then pelleted, washed with water, and subjected to MNase digestion as previously described [33],[87] and immunoprecipitation (see below).
Approximately 1×107 cells from individual samples were pelleted and resuspended in 600 µL Qiagen RLT buffer. After bead beating for 3 min, the supernatants were collected and 3–5 µL of the cell extracts were used for nCounter assays. The nCounter assays were performed as described [25] with customized probes corresponding to 200 S. cerevisiae RNAs.
The nCounter dataset reports on the measurement of 200 probes×202 mutants×4 time-points. We denote by Mi,j, the measurement of probe sample . To account for differences in hybridization, processing, binding efficiency, and other experimental variables, we used to following normalization procedure:
We computed a correlation matrix (Figure 3A) by first concatenating the measurements ( values) for all probes at the four time points to a single vector for each mutant, and then computing the Pearson correlation between the vectors for each pair of mutants. We clustered the correlation matrix using hierarchical clustering with Euclidian distance metric and unweighted average distance (UPGMA) linkage. Clustering was done using MATLAB 7.10 procedures “pdist,” “linkage,” and “dendrogram.”
To identify significant correlations between mutants we used a quantile-quantile plot. For a query mutant, we plotted the quantiles of its correlations vector with all other mutants versus theoretical quantiles from a normal distribution (function “qqplot,” MATLAB 7.10). Values that deviate from the line y = x were considered significant (Figure S4).
Principal component analysis was applied to the map of 200 probes versus 202 mutants using MATLAB 7.10 procedure “princomp.”
To evaluate transcriptional induction in individual cells in a population, we performed time-lapse microscopy of the induction of GFP-tagged protein. Nine deletion strains (hda2Δ, yta7Δ, spt8Δ, set1Δ, rph1Δ, snf1Δ, cac2Δ, and swc3Δ) were generated using KanMX in the BY4742 background. Four GFP-fusion reporters were selected (GCY1, GRE3, PGM2, and TSA2) from a library (Breker and Schuldiner, personal communication) based on the yeast GFP-tagging library [89] with an additional constitutive cytoplasmic mCherry (Genotype: xxx-GFP::HIS3, pTEF2-cherry::URA3, his3 leu2 met150 ura30 lyp1 can1::pMFA1-LEU2). Knockout strains were mated with GFP reporter strains and sporulated to generate haploid deletions carrying the GFP reporters.
Prior to assay, strains were grown in 96-well plates to mid-log (∼0.6 OD 600) in synthetic complete media (SC). We then transferred cells to glass bottom microwell plates (384 format, Matrical Biosciences) pre-treated with concavalin-A (incubation with solution at 0.25 mg/ml for 15 min). Cells were allowed to settle onto the glass surface for 30 min. We then removed the media and replaced with treatment media (SC with 1.5 mM diamide).
Following induction we placed the cell on an automated microscope (Scan∧R system, Olympus) and assayed with 40× objective at ∼10 min intervals, taking transmitted light, mCherry, and GFP images at each time point. Images were analyzed using custom-made software, written in python based on the OpenCV image analysis package (http://opencv.willowgarage.com/). Briefly, the procedure detects cells by thresholding the mCherry image and finding contours of bright objects. Contours that meet gating criteria for circularity and size were considered cells. The procedure matched detected cells in successive images based on a reciprocal closest hit procedure allowing a maximum of 5 pixel movement. Since cells are adhered to the glass surface, this procedure was effective in following a single cell. If there was a budding event, the closest-hit procedure returns an ambiguous result and the match is not made. Cells that were traced throughout the time course were used in the further analysis steps.
We represented each single cell time-course GFP measurements using a simple kinetic model. We assume that the transcription starts at a certain point following stimuli, termed ton, and stops at toff. Promoter behavior is represented by T(t):
Only during this time interval is mRNA being transcribed; to simplify the model we assume that transcription occurs with a constant rate of mRNA/min, and we also assume a constant exponential decay rate of mRNA denoted by . We present the mRNA levels as a function of time using the following differential equation:
Solving this equation we obtained a logistic equation describing mRNA level over time after stress induction at t = 0:
In the final step we assume that protein is the integration of the mRNA levels (assuming constant translation rate without degradation). We add a final parameter to account for the basal GFP level (prior to stress), solved the integral of , and obtained the following equation:
To estimate the parameters for each single cell track, we used MATLAB 7.10 function “fmincon” using the “active-set” optimization algorithm.
ChIP material was amplified using the DNA linear amplification method described previously [8],[90]. 3 ug of the amplified ChIP products was labeled via the amino-allyl methods as described on http://www.microarrays.org. Labeled probes (a mixture of Cy5-labeled input and Cy3-labeled ChIP-ed material) were hybridized onto an Agilent yeast tiled oligonucleotide microarray (G4495A) at 65°C for 16 h and washed as described on http://www.microarrays.org. The arrays were scanned at 5 µ resolution with an Agilent scanner. Image analysis and data normalization were performed using Agilent feature extraction.
|
10.1371/journal.pmed.1002159 | Impacts on Breastfeeding Practices of At-Scale Strategies That Combine Intensive Interpersonal Counseling, Mass Media, and Community Mobilization: Results of Cluster-Randomized Program Evaluations in Bangladesh and Viet Nam | Despite recommendations supporting optimal breastfeeding, the number of women practicing exclusive breastfeeding (EBF) remains low, and few interventions have demonstrated implementation and impact at scale. Alive & Thrive was implemented over a period of 6 y (2009–2014) and aimed to improve breastfeeding practices through intensified interpersonal counseling (IPC), mass media (MM), and community mobilization (CM) intervention components delivered at scale in the context of policy advocacy (PA) in Bangladesh and Viet Nam. In Bangladesh, IPC was delivered through a large non-governmental health program; in Viet Nam, it was integrated into government health facilities. This study evaluated the population-level impact of intensified IPC, MM, CM, and PA (intensive) compared to standard nutrition counseling and less intensive MM, CM, and PA (non-intensive) on breastfeeding practices in these two countries.
A cluster-randomized evaluation design was employed in each country. For the evaluation sample, 20 sub-districts in Bangladesh and 40 communes in Viet Nam were randomized to either the intensive or the non-intensive group. Cross-sectional surveys (n ~ 500 children 0–5.9 mo old per group per country) were implemented at baseline (June 7–August 29, 2010, in Viet Nam; April 28–June 26, 2010, in Bangladesh) and endline (June 16–August 30, 2014, in Viet Nam; April 20–June 23, 2014, in Bangladesh). Difference-in-differences estimates (DDEs) of impact were calculated, adjusting for clustering. In Bangladesh, improvements were significantly greater in the intensive compared to the non-intensive group for the proportion of women who reported practicing EBF in the previous 24 h (DDE 36.2 percentage points [pp], 95% CI 21.0–51.5, p < 0.001; prevalence in intensive group rose from 48.5% to 87.6%) and engaging in early initiation of breastfeeding (EIBF) (16.7 pp, 95% CI 2.8–30.6, p = 0.021; 63.7% to 94.2%). In Viet Nam, EBF increases were greater in the intensive group (27.9 pp, 95% CI 17.7–38.1, p < 0.001; 18.9% to 57.8%); EIBF declined (60.0% to 53.2%) in the intensive group, but less than in the non-intensive group (57.4% to 40.6%; DDE 10.0 pp, 95% CI −1.3 to 21.4, p = 0.072). Our impact estimates may underestimate the full potential of such a multipronged intervention because the evaluation lacked a “pure control” area with no MM or national/provincial PA.
At-scale interventions combining intensive IPC with MM, CM, and PA had greater positive impacts on breastfeeding practices in Bangladesh and Viet Nam than standard counseling with less intensive MM, CM, and PA. To our knowledge, this study is the first to document implementation and impacts of breastfeeding promotion at scale using rigorous evaluation designs. Strategies to design and deliver similar programs could improve breastfeeding practices in other contexts.
ClinicalTrials.gov NCT01678716 (Bangladesh) and NCT01676623 (Viet Nam)
| The benefits of exclusive breastfeeding are well documented, but only about a third of infants are exclusively breastfed during their first 6 months of life, well short of the 50% target endorsed by the World Health Assembly.
Past efforts to promote and support breastfeeding were rarely implemented at scale.
Previous studies on strategies to support breastfeeding have been limited to efficacy research; the few studies of programmatic approaches did not use a rigorous evaluation design.
We conducted two cluster-randomized impact evaluations of at-scale programs, in Bangladesh and Viet Nam (the Alive & Thrive initiative), to compare the population-level impacts on breastfeeding practices of two intervention packages: (1) an intensive package consisting of intensified interpersonal communication on breastfeeding, a mass media campaign, community mobilization, and policy advocacy to create a supportive environment for optimal breastfeeding practices and (2) a non-intensive package consisting of standard nutrition counseling on breastfeeding and a less intensive mass media campaign, community mobilization, and policy advocacy.
Surveys were conducted among households with children 0–5.9 mo of age at two points in time, before the interventions started (2010) and four years later (2014).
In both countries, we found significantly positive population-level impacts on breastfeeding practices, including higher rates of early initiation of breastfeeding and exclusive breastfeeding, and lower use of prelacteal feeding and bottle feeding, in areas that received the intensive package compared to areas that received the non-intensive package.
These findings provide much-needed evidence on what works to improve breastfeeding at scale.
Combining interpersonal counseling of mothers on optimal breastfeeding practices with a mass media campaign is more effective than a mass media campaign alone, suggesting that using multiple platforms and interventions to improve breastfeeding practices leads to greater improvements in practices than using one strategy alone.
The well-documented approach used by Alive & Thrive offers several implementation tools that can be used and adapted to inform the design of contextually relevant interventions to improve breastfeeding, a critical human development intervention.
| Globally, only 38% of infants are exclusively breastfed during their first 6 mo of life [1,2]. One of the six Global Nutrition Targets 2025 endorsed by the World Health Assembly is to “increase the rate of exclusive breastfeeding in the first 6 months up to at least 50%” [3]. Despite the well-documented benefits of exclusive breastfeeding (EBF) [4,5], and longstanding inclusion of interventions to improve EBF in child survival [6] and nutrition [4] strategies, tracking of global child survival goals shows slow progress in achieving adequate EBF levels in most developed and developing countries [2,7]. Past EBF promotion approaches, although often successful [8–10], have generally failed to achieve and maintain scale [11–13]. Achieving the ambitious World Health Assembly target for EBF will require a greater focus on effective delivery platforms [11] with multipronged approaches to implement successful programs at scale [10], including approaches that tightly link policy advocacy, investments in human resource development, data-driven implementation, research, and evaluation [12].
Suboptimal breastfeeding practices, including non-exclusive breastfeeding, contribute to 11.6% of mortality in children under 5 y of age [1]. Breastfeeding also prevents illnesses and provides essential nutrients for optimal child growth and development during the first 2 y [1,14,15], conveys several benefits for the mother [16], and may even have long-term effects on IQ and income [17]. Currently, the World Health Organization recommends initiation of breastfeeding within 1 h of birth, EBF until 6 mo of age, and continued breastfeeding until 2 y of age or beyond.
Systematic reviews have documented the impact on breastfeeding practices of various approaches, including those based on the use of individual- and group-based peer counseling and contact with lay counselors or trained personnel [8,9]. The studies reviewed, however, are generally implemented at small scale and under relatively controlled conditions. The reviews highlight the need to rigorously test the effectiveness of these and similar strategies when implemented at large scale. Although one study reports on the results of at-scale programs [18], the adequacy design used [19] does not provide irrevocable evidence that shifts in practices were attributable to the intervention strategies promoted by the program. A new review [10] provides evidence that combinations of interventions delivered through different platforms are more effective than single interventions, and recommends incorporation of combined strategies using complex adaptive systems for scaling up. However, to our knowledge, there are no studies, that combine at-scale programming with a rigorous evaluation using a randomized study design. Therefore, public health evidence on whether the magnitude of impacts on breastfeeding practices seen in efficacy studies and synthesized in systematic reviews can be achieved when programs operate at scale under real-life conditions is lacking [11,12].
This paper reports on findings from cluster-randomized impact evaluations of two at-scale programs, in Bangladesh and Viet Nam, that provided intensified interpersonal counseling (IPC) to pregnant women and mothers of children up to 2 y of age, combined with mass media (MM), community mobilization (CM), and policy advocacy (PA) intervention components, to improve breastfeeding practices [20]. The two program models that were evaluated reached large scale through distinct delivery platforms. In Bangladesh, IPC and CM were delivered by the community-based health platform of a large non-governmental organization (BRAC) through its network of field officers, community-based frontline workers (FLWs), and volunteers. In Viet Nam, IPC and CM were delivered through a social franchising approach integrated within the facility-based government health system [21]. In both countries, a nationwide MM campaign was implemented, as well as PA, to create a supportive environment for optimal infant feeding practices. Both implementation models used data-driven approaches that were focused on impact and delivery at scale and adapted to implementation conditions and context, thus fulfilling several elements recently identified as critical to scaling up impact on nutrition [22], and specific elements central to scaling up breastfeeding in particular [12].
Approval for the study was obtained from the institutional review board at the International Food Policy Research Institute (IRB #00007490), the Bangladesh Medical Research Council (IRB #BMRC/NREC/2007-2010/99), and the Vietnam Union of Science and Technology Associations (IRB #0904). All mothers of study children were provided with detailed information about the study verbally and in writing at recruitment. Verbal informed consent was obtained from mothers in Bangladesh; written informed consent was obtained in Viet Nam.
A cluster-randomized, non-blinded impact evaluation design was employed in each country to compare the impact of two Alive & Thrive (A&T) intervention packages, i.e., an intensive package consisting of intensified IPC, MM, CM, and PA compared to a non-intensive package consisting of standard nutrition counseling on breastfeeding along with less-intensive MM, CM, and PA on the primary and secondary outcomes of interest. Table 1 describes the interventions and the differences between the intensive and non-intensive packages. Cross-sectional household surveys were conducted at baseline (2010) and 4 y later (2014) in the same communities, among households with children under 6 mo of age. The questionnaires administered at each survey round were the same, with the exception of a detailed set of questions to measure program exposure included at endline. Within a large rollout of the intensive intervention package in both countries, 20 sub-districts (upazilas) in Bangladesh and 40 communes in Viet Nam were randomized to either the intensive or non-intensive intervention. The technical rationale for the impact and process evaluation approach used for A&T is available elsewhere [23,24]. The process evaluation used a variety of methods, tailored to implementation timelines, and data were collected in 2011, 2012, and 2013 in both countries. In both countries, the evaluation was restricted to a smaller geographic area than the full coverage area of the program.
In both countries, the evaluation was conducted in a subset of the total number of geographic areas chosen for programming. In Bangladesh, the program areas and the subset of evaluation areas were both rural. In Viet Nam, the program areas included both urban and rural areas, but the subset of evaluation areas were rural. At the outset of the evaluation, given the challenges of generalizability of a combined urban-rural evaluation model in Viet Nam, a decision was made to focus on rural areas only for the evaluation. By selecting rural areas only, we tried to ensure that the pool of areas included in the randomized allocation were somewhat homogenous.
The implementation teams were distinct from the evaluation teams. In Bangladesh, the implementation team was BRAC and FHI360, and the data collection team came from DATA, a survey research firm that had limited interactions with the implementation team. In Viet Nam, the implementation team was Save the Children and the government health system, and the data collection team came from the Institute of Social and Medical Studies. International Food Policy Research Institute researchers in charge of the evaluation connected and shared the information between the implementation and evaluation teams.
The study was registered at ClinicalTrials.gov in August 2012, about 2 y after the start of the program evaluation (April 2010). The sole reason for the delay in registering the trial was our understanding, at the time the study was initiated, that a program evaluation did not warrant registration, given the distinct differences between program evaluations and standard clinical trials. However, as discussions grew in the program evaluation community for preregistration of impact evaluations [25], a decision was made to register this program evaluation in a trial registry to ensure transparency and public availability of this information. No changes were made to the planned evaluation design or the planned analyses in the period between the design of the evaluation in 2010 and the registration of the evaluation in at ClinicalTrials.gov in 2012.
In Bangladesh, BRAC delivered intensified IPC and CM in 50 rural sub-districts in five of six divisions (later subdivided into seven) through its existing countrywide Essential Health Care program that already included standard counseling (S1 Fig) [26]. For standard nutrition counseling (available in both non-intensive and intensive areas), BRAC FLWs, called Shasthya Kormi, and health volunteers, called Shasthya Sebika, conducted routine home visits and provided standard information on infant and young child feeding (IYCF) practices, including breastfeeding. In intensive areas, IPC was based on multiple age-targeted IYCF-focused visits to households with pregnant women and mothers of children up to 2 y of age by the FLW and volunteer, as well as home visits by a nutrition-focused FLW, called Pushti Kormi, who was an additional human resource to provide more skilled support for breastfeeding and complementary feeding. In these areas, Shasthya Kormi and Pushti Kormi conducted monthly home visits and introduced age-appropriate IYCF practices, coached mothers as they tried out the practices, and engaged other family members to support the behaviors (see Table 1 for more details). These workers were supervised using observation checklists, and their workload was monitored. Cash incentives (US$6–US$8/mo) were given to the volunteers in the intensive areas for intervention delivery performance, which included ensuring high coverage, carrying out age-appropriate counseling at home visits, and collecting maternal reports of practicing the recommended behaviors. Households were unaware of this performance incentive. In intensive areas, CM included sensitization of community leaders to IYCF, and community theater shows focused on IYCF. The MM component consisted of the national broadcast of seven TV spots that targeted mothers, family members, health workers, and local doctors with messages on various aspects of IYCF; two of the spots focused on breastfeeding. Buys of media airtime were designed for multiple airings during the country’s most watched programs. In intensive areas that had limited electricity and TV access, supplemental activities were conducted to air the TV spots, and other IYCF films produced by the project, through local video screenings. PA included workshops to share data, engagement of journalists to broaden reporting on IYCF in the media, creation of an IYCF alliance and other such activities, which aimed at creating additional countrywide awareness of policies and programs to support breastfeeding.
In Viet Nam, standard nutrition counseling (available in both non-intensive and intensive areas) consisted of messages and information on IYCF delivered as part of routine child health care contacts at the government health facilities. Intensive areas received A&T program activities, which were implemented in 15 of 63 provinces (S2 Fig). Save the Children worked with the government of Viet Nam to establish a total of 781 government health facilities at the province, district, and commune levels that used a social franchising model, called Mat Troi Be Tho (MTBT), to deliver facility-based individual and group counseling. All facilities were required to meet minimum criteria including a standardized counseling room, trained staff, and availability of job aids and client materials. The program aimed to deliver 9 to 15 counseling contacts to each mother-child pair from the last trimester of pregnancy through the child’s first 2 y of life, including eight breastfeeding-focused contacts in the first 6 mo of life. Referrals, CM, promotional print materials, and TV advertising were used to generate demand for preventive IYCF counseling services, a concept new to most families. Training and supervision, incentives to the health facilities, and tools to collect data were applied to improve the quality of services. The MM component consisted of a nationally broadcast campaign using TV and the digital space (internet and mobile phone applications); three of four TV spots focused on breastfeeding. In intensive areas, the MM campaign also included additional out-of-home advertising through billboards and LCD screens. PA at the national and provincial levels targeted the extension of paid maternity leave to 6 mo, strengthening of the code of marketing of breast milk substitutes, and improving provincial planning for IYCF and nutrition actions.
The major components of the interventions started at the same time in all intervention areas in both Bangladesh and Viet Nam (Fig 1). In Bangladesh, intensified IPC began in 22 sub-districts (upazilas) by August 2010 and in another 28 sub-districts by August 2011. MM was launched in December 2010 and intensified to reach national coverage by February 2011. CM was operational in August 2010. In Viet Nam, IPC was initiated in March 2011, MM in November 2011, and CM in March 2011. PA activities commenced in August 2010 in both Bangladesh and Viet Nam. With the endline survey in 2014, the total duration of implementation of the full intensive package of interventions was about 3 y in both countries. All components continued uninterrupted throughout the implementation period once they had started. See Box 1 for information on A&T’s approach to operating at scale and Box 2 on the development of the MM campaign.
The primary outcome of the analysis presented in this paper is EBF in the previous 24 h, defined as the proportion of mothers of infants 0–5.9 mo of age who fed only breast milk (based on a previous-day recall of all foods and liquids) [30]. A key secondary outcome is early initiation of breastfeeding (EIBF), defined as the proportion of infants who were reported by mothers to have been put to the breast within 1 h of birth. Other related breastfeeding behaviors were also measured. Women with obstetric complications or cesarean delivery were included in the counts in determining EIBF. We categorized breastfeeding as exclusive, predominant (infant given water, water-based drinks, fruit juice, or ritual fluids in addition to breast milk), partial (infant given liquids and non-liquids such as milk, non-milk-based products, and other foods in addition to breast milk), and no breastfeeding [31], and also report EBF for the age bands 0–0.9 and 1–5.9 mo. We assessed prelacteal feeding based on maternal recall of foods consumed by the infant immediately after birth and during the first 3 d of life. Intervention exposure was measured by maternal recall of FLW home visits in the last 6 mo, recall of attendance at a CM event in the last 1 y, recall of ever having seen an A&T-promoted TV spot, and recall of specific breastfeeding messages contained within the TV spot. To help address the challenge of recall-based self-reported measures of breastfeeding, we collected information on infant diarrhea (defined as three or more loose stools in a 24-h period [32]).
Sample size calculations were carried out to determine the sample size needed to detect differences in the primary outcome, i.e., EBF, between the two intervention groups at endline, considering an alpha of 0.05, power of 0.80, intra-class correlation (ICC) of 0.01 (estimated from previous national or sub-national surveys), and estimated baseline prevalence of EBF of 43% in Bangladesh and 30% in Viet Nam. We hypothesized that the intensive intervention would increase the primary outcome indicator, and therefore a one-sided test was used. For cost and logistical reasons, 20 clusters (sub-districts) were selected to be randomized in Bangladesh, and 40 clusters (communes) in Viet Nam. In Bangladesh we estimated that a sample size of 49 infants aged 0–5.9 mo per cluster—for a total sample of 980 (490 per group)—was sufficient to detect a difference at endline in the practice of EBF of 10 percentage points (pp) or larger. In Viet Nam, we estimated that a sample size of 23 infants per cluster—for a total sample of 920 (460 per group)—was sufficient to detect a minimum difference of 9 pp. Prior to conducting the endline survey, we re-verified sample size estimates based on the original sample size, the observed baseline EBF prevalence in our sample, and the ICC for EBF from the baseline survey (ICC of 0.04 for Viet Nam, and 0.1 for Bangladesh). No changes in sample sizes were made at the endline survey, based on these parameters and anticipated effect sizes based on the baseline survey.
In Bangladesh, 100 sub-districts, across five divisions, were selected by BRAC as possible A&T intensive areas based on high poverty, stunting levels in excess of 30% among children < 5 y of age, and non-inclusion of the sub-district in the government National Nutrition Program. This list was narrowed to 78 based on geographic proximity (i.e., within the same agro-ecological and/or administrative zone, also called a division in Bangladesh), size, and other operational aspects to ensure homogeneity across the sample. The first phase of implementation took place in 50 of these sub-districts. Within each of five divisions, four sub-districts were randomly selected for inclusion in the evaluation sample using a computer program, for a total of 20 sub-districts. Sub-districts within each division were then randomly assigned, using a computer program, to either the intensive (10 sub-districts) or non-intensive (10 sub-districts) intervention. The randomization process was carried out in the presence of BRAC and A&T staff and the program evaluators in BRAC’s headquarters in Dhaka.
In Viet Nam, 15 provinces were selected for program implementation based on stunting levels, absence of other large organizations working in nutrition, population density, and representation of the different ecological regions covered by the initiative. Four rural provinces, representing four distinct ecological zones, were then selected for inclusion in the evaluation sample, and within these provinces, ten rural districts; two to six communes per district were selected for the evaluation sample based on the presence of a health center that met the eligibility criteria for the A&T franchise model; this ensured homogeneity across the sample. Communes were randomly assigned to either the intensive (20 communes) or non-intensive (20 communes) intervention. The randomization process was carried out using a simple public lottery system in the presence of local, district, and provincial health authorities as well as the program evaluators.
In both countries, households within the intensive and non-intensive areas were not explicitly made aware of the results of the randomization. Additionally, there was no blinding of the intervention at the level of service delivery.
In Bangladesh, within each sub-district, five unions and two villages within each union were randomly selected, to yield a total of 200 villages. Each village had an average size of 250 households. Within each village, a household census was conducted at baseline and endline to list mothers and infants and infants’ date of birth. A list of all households with infants less 6 mo of age was then drawn up. In Viet Nam, communes were sampled, and household listings were obtained from the health authorities, which maintain all birth records. In both countries, from these census lists that identified all households with children < 6 mo of age, households were selected for the survey using systematic sampling beginning with a random seed start point, to yield the desired sample size per cluster. Household visits for the survey were conducted only after sampling the desired number of households per village/commune. Desired sample sizes were 50 households per village in Bangladesh and 25 households per commune in Viet Nam. In Viet Nam, when there were insufficient infants within a commune due to small population sizes, we oversampled infants from another commune within the same district and intervention group.
A household questionnaire was administered to mothers of children under 6 mo of age to collect data on primary and secondary outcomes, along with data on several other household, maternal, and child characteristics (S1 Table). The interview was conducted in the local language, and almost all interviewees agreed to participate in the surveys. The refusal rate among mothers selected for inclusion through the random selection process and contacted was 0.9% for Bangladesh and 0.6% for Viet Nam. We excluded only mothers who had a clear mental disability, defined as inability to answer basic questions about their name or willingness to be interviewed that would prevent them from understanding and answering the questions.
Baseline differences between the two interventions were tested using ordinary least squares regression models (continuous variables) or logit regression models (categorical variables) with random effects, accounting for child age, sex, and geographic clustering [33]. We derived difference-in-differences estimates (DDEs) of impact using fixed-effects regression models that assessed differences between the two study groups over time [34]. The difference-in-differences method relies on comparing the difference in changes (i.e., differences) between baseline and follow-up in the intervention and control groups. Even though there are other methods of attributing the impact of interventions, the difference-in-differences method with randomization allows a best estimate of the impact attributable to the program by taking into account the changes in outcomes that may occur among the non-beneficiary population as a result of non-program factors such as secular trends, other types of programs, or climate-related or other types of shocks that may have independent positive or negative impacts on the study’s main outcomes. Difference-in-differences impact analyses have been widely used in health and nutrition impact evaluations [34].
We present intent-to-treat DDEs adjusting for geographic clustering, infant age, and gender. Using a subset of five items adapted from Reynolds [35] that were collected in the endline survey, we conducted a robustness analysis for EBF to test for social desirability bias, i.e., tendency of respondents to respond so as to be viewed favorably by others. Data analysis was performed using STATA 13; a statistical analysis plan was developed prior to endline data collection and discussed with the funder and program implementers. Although the prespecified plans assumed a one-sided test, we conducted analyses using two-sided tests to allow for the possibility of uncovering unexpected effects.
No evaluation clusters were lost to follow-up (Fig 2); none crossed from non-intensive to intensive during implementation. In Bangladesh, cluster size varied little across clusters or over time; in Viet Nam, there was greater variability, reflecting commune population size.
Desired sample sizes were achieved at baseline (June 7–August 29, 2010, in Viet Nam; April 28–June 26, 2010, in Bangladesh) and endline (June 16–August 30, 2014, in Viet Nam; April 20–June 23, 2014, in Bangladesh) in both countries. In Bangladesh, 977 households were surveyed at baseline, and 998 households at endline. In Viet Nam, achieved sample sizes were 948 at baseline and 1,002 at endline.
The two intervention groups were well-balanced for characteristics potentially associated with intervention effects in both countries (e.g., infant age and sex, household economic status, and maternal education and nutrition). In Bangladesh, there were small differences at baseline in maternal age, maternal employment, and household headship, and a 10.0-pp difference in ownership of agricultural land (Table 2). In Viet Nam, a higher mean level maternal mental stress and slightly higher mean level of mild food insecurity were seen in the non-intensive group at baseline.
DDEs (Fig 3) show that there were greater increases between baseline and endline in the proportion of women practicing EBF in the intensive group than in the non-intensive group, both in Bangladesh (36.2 pp, 95% CI 21.01–51.46, p < 0.001) and Viet Nam (27.9 pp, 95% CI 17.74–38.07, p < 0.001) (Fig 3). The proportion of women practicing EBF in the intensive group increased from 48.5% to 87.6% in Bangladesh, and from 18.9% to 57.8% in Viet Nam. There was no significant increase in the proportion of women practicing EBF in the non-intensive group over time in Bangladesh, but an increase of 10.0 pp in Viet Nam.
In the intensive group in Bangladesh, effects on EBF were primarily achieved through replacement of partial breastfeeding with EBF (Fig 4). In Viet Nam, effects on EBF were primarily achieved through replacement of predominant breastfeeding with EBF (Fig 5). In Bangladesh, DDEs were also statistically significant for EIBF (16.7 pp, 95% CI 2.78–30.57, p = 0.021) and feeding of prelacteals during the first 3 d after birth (−49.3 pp, 95% CI −65.60 to −32.92, p < 0.001); in Viet Nam, DDEs were statistically significant for feeding of prelacteals during the first 3 d after birth (−18.8 pp, 95% CI −30.86 to −6.15, p = 0.003), formula use during the first 3 d after birth (−22.4 pp, 95% CI −33.53 to −10.67, p < 0.001), and bottle feeding in the first 6 mo (−13.0 pp, 95% CI −20.26 to −5.54, p < 0.001) (Table 3).
EBF in children under 1 mo of age did not differentially improve by intervention group in either country. Impacts in both countries were due to the differential improvements in EBF in favor of the intensive group among children 1–5.9 mo (Table 3).
In the 6 mo preceding the survey, mothers in Bangladesh reported receiving 6.6 visits by the IYCF worker (Pushti Kormi) and health volunteer (Shasthya Kormi) trained to deliver nutrition-related IPC as part of A&T, while mothers in Viet Nam visited the health facilities for counseling an average of two times (Table 4). Fidelity in terms of contacts with FLWs and volunteers was high in Bangladesh (85%–98%, depending on the type of FLW or volunteer), whereas use of the health facility in Viet Nam was lower (<50% reported visiting the A&T franchise facility [MTBT] in intensive areas). In both countries, the breastfeeding-related knowledge of FLWs and volunteers was higher in intensive compared to non-intensive areas. Attendance at a CM session ranged from 12% to 22% in the intensive group in Bangladesh. Maternal exposure to the TV spots was 61%–64% in Bangladesh, and did not differ between groups. Similarly, exposure to TV spots was 70% in Viet Nam, with no statistically significant differences between groups. As expected, awareness and exposure to the social franchise brand was significantly higher among individuals in the intensive than in the non-intensive group.
We measured social desirability in both countries to assess, and account for, potential biases in our main impact estimates for reported breastfeeding practices (S1 Text). We found evidence of a social desirability bias for EBF in Viet Nam, but not in Bangladesh (S3 Fig). There was no evidence of this bias for other breastfeeding outcomes (S4 Fig). In Viet Nam, as social desirability scores for EBF increased, reported EBF increased in both groups, but more so in the intensive group. After adjusting for this differential increase, the impact estimate for EBF in Viet Nam remained strong and statistically significant (DDE 15.2 pp, p = 0.008). In addition, based on the premise that diarrhea could potentially be lower in the EBF group, we cross-checked EBF self-reports and infant diarrhea and found that the prevalence of diarrhea was indeed lower among EBF infants than among non-EBF infants (3.2% versus 5.0% in Bangladesh, p = 0.044, and 3.7% versus 9.7% in Viet Nam, p < 0.001).
An at-scale behavior change program that focused on integrating intensified IPC, MM, and CM (the A&T intensive intervention) had a greater impact on breastfeeding practices, especially EBF than MM with standard nutrition counseling (the A&T non-intensive intervention), within the context of national advocacy to create a supportive environment for optimal feeding practices. In Bangladesh and Viet Nam, the proportion of women who reported practicing EBF in the previous 24 h among children <6 mo was 36 pp and 28 pp higher, respectively, in the intensive compared to the non-intensive intervention areas. In Bangladesh, where intensified IPC was delivered through repeated home visits at critical ages combined with CM, the proportion of women practicing EBF at endline was 88% in the intensive group (from a baseline prevalence of 49%). In Viet Nam, the proportion of women practicing EBF reached 58% (from a low 19% at baseline) through a facility-based delivery model for IPC that was built on principles of social franchising and helped strengthen the health system’s capacity to deliver counseling. These impacts were seen despite positive changes in household economic conditions, maternal education, and occupation that in many settings in Asia are associated with lower rates of EBF [36].
The higher EBF prevalence in the intensive compared to the non-intensive group at endline in Bangladesh and Viet Nam is comparable to results from systematic reviews of breastfeeding promotion interventions [8,9]. By contrast with previous literature, the magnitude of the effect we report was achieved in a large at-scale program. In Bangladesh the intensive approach was implemented in 50 of 493 rural sub-districts, 20 of which were in the evaluation. Based on program monitoring, BRAC reported conducting 2.2 million home visits focused on the promotion of EBF to mothers of infants 0–5.9 mo old between December 2011 and February 2014. In Viet Nam, a total of 781 A&T franchises in 15 of 63 provinces (four of which were in the evaluation) were operational at endline, and there were 1.1 million counseling contacts with mothers of infants 0–5.9 mo of age at health facilities between January 2012 and December 2014. The Bangladesh results are generalizable to other program models that rely on incentivized community volunteers and/or skilled FLWs conducting home-based or community IPC and CM, and suggest that intensifying and strengthening contacts and linking them with MM could have significant benefits [37,38]. The results for Viet Nam are applicable to other countries where primary health care utilization is high, MM reach is almost universal, and facility-based platforms can be used to deliver preventive and curative health care. In Viet Nam, the four provinces selected for evaluation were representative of the A&T program areas in terms of geographic regions (north, south, and central). Compared to the rest of the country and parts of the program areas, however, they were more rural and had a higher level of stunting. Similarly, in Bangladesh, the program was operational primarily in rural areas. Therefore, the results are primarily generalizable to rural areas served by government health services (in countries like Viet Nam) and rural areas in countries like Bangladesh.
The A&T model for improving breastfeeding at scale was initiated in 2008, and developed in 2009–2010, based on prior experience [18]. Much of the evidence on the impacts of breastfeeding [1], and on elements necessary to assure scale up of breastfeeding programs using multiple platforms and data-driven coordinated efforts [12], was available only after the initial development of the A&T model. This rigorous evaluation of a complex intervention now deeply strengthens the programming evidence base for improving breastfeeding, and supports the data-driven and adaptive approach that has been recommended but not always put into practice. The A&T model, thus, provides guideposts for developing a practical and well-coordinated approach that links PA, strengthening of systems for implementation at scale, and the use of data from formative research, monitoring, and evaluation.
The plausibility of our findings are strengthened by the findings from a process evaluation [21,39,40] on service delivery and intervention exposure. Interventions were implemented largely as designed in both countries [21,40,41], but demand-side constraints to use of facilities were a challenge in Viet Nam [39]. In Bangladesh, intensified IPC was delivered well by FLWs, who were incentivized and monitored to cover all targeted households; achieved coverage was very high. In Viet Nam, high-quality counseling was established in the intensive group compared to the non-intensive group [21]; since reach was dependent on mothers visiting the health facilities, and subject to demand-side constraints, the model achieved lower coverage [39] than the outreach-based model in Bangladesh. MM reached a substantial proportion of the population in both settings, and we saw evidence of shifts in some breastfeeding practices in the non-intensive group as well. It is likely the MM played multiple roles and worked synergistically with the IPC, but our evaluation design does not tease these apart.
A challenge in assessing impact on feeding practices is the use of recall-based self-reported measures. Our study was subject to this challenge, particularly because the intervention included a MM component that provided regular reminders of recommended behaviors. Although we did not collect data on lactational amenorrhea, we did collect information on infant diarrhea (reported by the mothers). We compared differences in diarrhea prevalence between groups, finding, as expected, that there was lower prevalence of diarrhea among EBF infants compared to non-EBF infants. We also assessed the role of social desirability in relation to the main impact indicators [35]. In Bangladesh, the results were unaffected by respondents’ desire for social approval. In Viet Nam, it appears that respondent desire for social approval had some influence on reporting (S1 Text), which was differential between intervention groups, but strong and significant intervention effects were still demonstrated after accounting for this reporting bias.
In both countries, the evaluation was restricted to a significantly smaller geographic area than the true coverage of the program. In order to assess the possibility that there was a differential quality of service delivery due to the non-blinded nature of the intervention, we randomly assessed service delivery in a subset of intensive areas that were not included in the impact evaluation. We found no evidence of differential quality of service delivery in either country, further supporting the external validity of the evaluation findings to the at-scale program in both countries.
From the perspective of understanding the scaling up of a successful breastfeeding program, and the special efforts made by this initiative to ensure delivery of a comprehensive set of interventions at large scale, an analysis focused on the scaling up of this program identified the presence of several critical elements for scaling up [22]. Specifically, the program appears to have successfully included multiple elements deemed necessary for successful scale-up, including the following: (1) a vision for impact on breastfeeding; (2) finding the right combination of interventions and operational contexts in the two countries; (3) having access to adequate, stable, and flexible financing to stay adaptive and goal-focused; (4) actively engaging champions and alliances via PA; (5) using multiple pathways to scaling up—expanding and strengthening the capacities (increasing FLWs in Bangladesh and establishing counseling rooms within facilities in Viet Nam); and (6) including adequate learning through the use of data and learning. Much of what occurred in the context of this program also reflects elements of the complex adaptive health care systems framework [42,43]. Looking forward, assessments of the sustainability of these actions at scale will be needed to understand the extent to which these investments led to a sustained legacy focus on infant feeding in the context of the BRAC program in Bangladesh and the health system in Viet Nam.
Given that the A&T intervention was a large-scale intervention involving IPC, MM, and CM, there was some potential for between-group contamination. In Bangladesh, there was no between-group contamination because the sub-districts were administratively separated for implementation, and we verified the lack of contamination through routine interactions with the program implementation team. In Viet Nam, since the communes were situated within districts that were in turn located inside the provinces chosen for the overall A&T program, there was potential for some between-group contamination. We assessed this contamination informally in 2013 and ascertained that some content-related overlap had occurred in refresher training of health workers. Given that between-group contamination is likely to have dampened intervention effects, overall intervention effects may have been even higher than those seen.
This study was not designed to assess the impact of MM alone on breastfeeding practices, because the media campaign was implemented nationwide. Similarly, our design did not allow us to isolate the contributions of the nationwide PA activities, which aimed to create an enabling environment for implementation and scale-up of IYCF programs in the country. Our impact estimates may thus underestimate the full potential of such a multipronged intervention because the evaluation lacks a “pure control” area with no MM or national/provincial PA. Lastly, the EBF indicator itself (EBF in the previous 24 h), though currently recommended by WHO, has limitations. Most notably, it is reflective only of the previous day’s practice and does not inform about usual or continued breastfeeding practices over time. A study of the dynamics of EBF in Peru showed that mothers switched back and forth between exclusive and predominant breastfeeding during the first 6 mo of the child’s life [44]. The possibility of this practice is also supported by smaller scale qualitative research in our process evaluations [45,46]. Although there is some evidence that the EBF indicator is accurate [47], further research is needed to develop stronger and more reliable indicators of EBF, beyond the 24-h recall-based indicator that is currently used.
Evidence has been mounting on the positive impacts of breastfeeding on a range of maternal and child health outcomes in diverse settings [16,17], and yet progress on improving breastfeeding practices is held back globally [2], possibly because policies and interventions to date have not embraced full-scale efforts to deliver integrated, tailored, and well-coordinated strategies [12]. In this global health context, this study fills an important gap in the public health literature [11] on the impact of delivering a critically important child survival, nutrition, and human development intervention at scale. Using rigorous cluster-randomized evaluation designs, it shows that comprehensive behavior change strategies implemented at scale, under real-life conditions, and delivered through outreach-based (Bangladesh) and facility-based (Viet Nam) platforms have strong and significant impacts on breastfeeding practices. Strategies that combine intensive IPC with MM campaigns, CM, and advocacy are more effective than standard counseling with less-intensive accompanying strategies. Our study demonstrates the impacts of recommended complex adaptive systems-based approaches to programming for scaling up combined interventions to improve breastfeeding in low- and middle-income countries [10]. We conclude that investments in combined, coordinated, and data-driven strategies, especially those that intensify counseling and are supported by MM, CM, and PA, are recommended for replication in similar contexts and for sustained implementation in Bangladesh and Viet Nam.
|
10.1371/journal.pntd.0003748 | Rapid Emergence of Multidrug Resistant, H58-Lineage Salmonella Typhi in Blantyre, Malawi | Between 1998 and 2010, S. Typhi was an uncommon cause of bloodstream infection (BSI) in Blantyre, Malawi and it was usually susceptible to first-line antimicrobial therapy. In 2011 an increase in a multidrug resistant (MDR) strain was detected through routine bacteriological surveillance conducted at Queen Elizabeth Central Hospital (QECH).
Longitudinal trends in culture-confirmed Typhoid admissions at QECH were described between 1998–2014. A retrospective review of patient cases notes was conducted, focusing on clinical presentation, prevalence of HIV and case-fatality. Isolates of S. Typhi were sequenced and the phylogeny of Typhoid in Blantyre was reconstructed and placed in a global context.
Between 1998–2010, there were a mean of 14 microbiological diagnoses of Typhoid/year at QECH, of which 6.8% were MDR. This increased to 67 in 2011 and 782 in 2014 at which time 97% were MDR. The disease predominantly affected children and young adults (median age 11 [IQR 6-21] in 2014). The prevalence of HIV in adult patients was 16.7% [8/48], similar to that of the general population (17.8%). Overall, the case fatality rate was 2.5% (3/94). Complications included anaemia, myocarditis, pneumonia and intestinal perforation. 112 isolates were sequenced and the phylogeny demonstrated the introduction and clonal expansion of the H58 lineage of S. Typhi.
Since 2011, there has been a rapid increase in the incidence of multidrug resistant, H58-lineage Typhoid in Blantyre. This is one of a number of reports of the re-emergence of Typhoid in Southern and Eastern Africa. There is an urgent need to understand the reservoirs and transmission of disease and how to arrest this regional increase.
| Typhoid fever is a major cause of disease and death around the world, particularly in resource limited settings, although reports suggest that until recently it has been much less prominent in sub-Saharan Africa (SSA) than Asia. Estimates of the precise burden of this disease are, however, difficult, as diagnosis requires advanced laboratory diagnostics. This is a particular problem in much of SSA where long-term laboratory surveillance has been available in just a few settings. Queen Elizabeth Central Hospital (QECH), Blantyre, Malawi is one such setting; between 1998 and 2010, cases of Typhoid fever at QECH were both uncommon and responsive to all antibiotics. In 2011 a marked increase in highly antibiotic resistant Typhoid fever began, with 843 confirmed cases in 2013. A review of cases revealed that one in 40 patients died and one in five had complicated disease. A further study of the DNA of bacteria associated with the outbreak revealed a novel strain, common to Asia, has arrived in Malawi. This is one of a number of reports of the re-emergence of Typhoid fever in Southern and Eastern Africa. There is an urgent need to understand the reservoirs and transmission of disease and how to arrest this regional increase.
| Typhoid fever, caused by Salmonella enterica serovar Typhi remains one of the most important infectious diseases globally, responsible for an estimated 26.9 million infections and 269,000 deaths in 2010 [1]. The accuracy of this estimate in sub-Saharan Africa (SSA) is limited by the paucity of diagnostic microbiological facilities in this setting, and the non-specific nature of Typhoid fever, which typically presents with non-focal sepsis, making syndromic diagnosis unreliable [2]. It has been suggested that the burden of Typhoid fever in Africa before 2010 has been over-estimated [3].
Whilst Typhoid has remained a major public health problem in Asia, there have been numerous reports consistently showing that nontyphoidal serovars of Salmonella (NTS) are a more prominent cause of bloodstream infection (BSI) in sub-Saharan Africa (SSA) [4]. In Blantyre, Malawi, there have been two well-documented epidemics of BSI caused by multidrug resistant (MDR) NTS serovars since 1998, while S. Typhi, until recently, has represented only 1% of Salmonella BSI [5].
The sequential acquisition of antimicrobial resistance (AMR) genes has been a prominent feature of Typhoid fever in Asia. MDR to amoxicillin, chloramphenicol and cotrimoxazole has been common since the 1980s [6]. This phenotype was largely associated with an IncH1 plasmid [7] and the use of any of these drugs will act to maintain this plasmid. This phenotype has led to the widespread use of fluoroquinolones in the management of Typhoid fever in Asia, the emergence and spread of diminished ciprofloxacin susceptibility and the increasing use of 3rd-generation cephalosporins and azithromycin there [8]. AMR surveillance data from SSA are sparse.
The Malawi Liverpool Wellcome Trust Clinical Research Programme (MLW) has conducted longitudinal surveillance of BSI in adult and paediatric medical patients presenting to Queen Elizabeth Central Hospital (QECH) with clinically suspected severe bacterial infection since 1998. In 2011 we detected an increase in microbiologically confirmed S. Typhi and here we report the emergence of a rapid and sustained increase in MDR Typhoid fever.
QECH is the largest government hospital in Malawi with 1,300 beds and provides free healthcare to Blantyre district (population approximately 1.3 million) and tertiary care to the Southern region of Malawi. MLW has conducted routine BSI surveillance since 1998 [9], obtaining blood for culture from all adult patients (age≥16) admitted to the medical wards with an axillary temperature over 37.5°C or with clinical suspicion of sepsis. In addition, blood cultures were obtained from febrile children (age<16) that were malaria slide negative or positive and critically ill, or from afebrile children with clinical suspicion of sepsis. Aerobic blood cultures were taken under aseptic conditions and before the administration of antibiotics. Sampling criteria did not change between 1998–2014.
Automated blood culture has been undertaken at MLW using a standard aerobic bottle (BacT/Alert, bioMérieux Marcy-L'Etoile, France) since 2000, although prior to this, manual culture was undertaken [9]. The sample date, patient details (name, age and gender), blood culture result and antimicrobial susceptibility profile of any pathogen isolated were recorded. All isolates were identified using standard diagnostic techniques [10]. Salmonellae were identified by biochemical profile using API20E (bioMérieux Marcy-L'Etoile, France) and serotyped according to the White-Kauffmann-Le Minor scheme by antisera (Pro-Lab Diagnostics, UK) [11]. Coagulase-negative staphylococci, Bacillus spp., diptheroids and alpha-haemolytic streptococci other than S. pneumoniae (when there was no clinical suspicion of endocarditis) were considered as contaminants. MLW audits blood culture volumes and contamination rates [12].
Antimicrobial susceptibility testing was performed by disc diffusion using ampicillin, chloramphenicol, cotrimoxazole, cefpodoxime and ciprofloxacin discs according to British Society of Antimicrobial Chemotherapy methods and breakpoints, and isolates that were found to be resistant to ciprofloxacin or ceftriaxone by disc testing had e-tests (E-test macromethod, bioMérieux Marcy-L'Etoile, France) performed. Diminished ciprofloxacin susceptibility (DCS) was classified on the basis of a minimum inhibitory concentration (MIC) between 0.06–1 mg/L) [13]. Isolates were described as “fully susceptible” if susceptible to these five antimicrobials. MLW subscribes to the United Kingdom National External Quality Assessment Service (UK NEQAS) scheme. Isolates deemed clinically significant were frozen on beads at -80°C. Prior to October 2010, these data were entered into ledgers then double-entered into a validated database. Since then they have been directly entered into an electronic laboratory information management system (LIMS).
Once the rapid increase in S. Typhi had been identified, a retrospective case note review of adults (aged ≥ 16 years) identified by Salmonella disease surveillance (College of Medicine Research Ethics Committee P.07/09/808) in the first 2 years of the increase (June 2011- June 2013) was undertaken. Clinical presentation, including focus of infection, HIV status (when known) and case-fatality rate at discharge were recorded. Data on survival at discharge amongst paediatric patients during the first 12 months of the increase was collected from the paediatric admissions ledgers. HIV testing was performed according to the Malawi national HIV rapid antibody testing protocol, using Determine (Alere, USA) HIV-1/2 tests as the first test in a serial testing algorithm. All positive test results were confirmed by Uni-Gold (Trinity Biotech, USA). First line treatment for uncomplicated Typhoid fever at QECH was oral ciprofloxacin for 7 days, at a dose of 750mg twice daily (bd) in adults and 20mg/kg bd (maximum 750mg) bd in children.
Minimum incidence of Typhoid fever/100,000 patients/year was calculated using an estimated sensitivity of blood culture for the diagnosis of S. Typhi of 50%, therefore the number of positive blood cultures recorded in each year was doubled to provide a numerator. QECH serves the population of urban Blantyre, and the population size of Blantyre was estimated in the 1998 and 2008 census and population projections following the 2008 census have been published (National Statistics Office, Malawi: www.nsomalawi.mw/2008-population-and-housing-census.html). These numbers were used to provide a denominator.
In order to investigate whether this increase represented clonal expansion of a single lineage of S. Typhi, and to describe the full diversity of S. Typhi in Blantyre, blood culture isolates over the course of the surveillance period were selected for whole genome sequencing to represent the spectrum of antimicrobial susceptibility patterns in each year. These isolates were placed in a global context using previously sequenced S. Typhi isolates of known haplotype [14,15]. One S. Paratyphi (A270, Accession Number ERS223417) isolate was used as an out-group, to root the tree.
DNA extraction for whole genome sequencing was conducted on the Qiagen Universal Biorobot® (Limburg, Netherlands) using Qiagen All-for-one® extraction kits. Following DNA extraction, PCR libraries were prepared from 500ng of DNA as previously described [16]. Isolates were sequenced using Illumina HiSeq2500 machines (Illumina, San Diego, CA, USA) and 150 bp paired-end reads were generated.
Phylogeny was based on single nucleotide polymorphisms (SNPs) in conserved regions of the genome; WGS data for each of isolates was mapped to the reference S. Typhi CT18 [7] using SMALT (http://www.sanger.ac.uk/resources/software/smalt/: version 0.5.8). Phylogenetic modelling is based on the assumption of a single common ancestor, therefore variable regions, where horizontal genetic transfer occurs and repetitive regions, were excluded [15] [17]. A maximum likelihood phylogenetic tree was then built from the SNP alignments of the isolates using RAxML (version 7.0.4) [18]. The maximum-likelihood phylogeny was supported by 100 bootstrap pseudo-replicate analyses of the alignment data. The presence of plasmids was investigated using PlasmidFinder (Danish Technical University, Denmark: http://cge.cbs.dtu.dk/services/PlasmidFinder/).
Institutional approval for this study was obtained from the University of Malawi College of Medicine Research Ethics Committee (COMREC). All data analyzed were anonymised.
Between 1998–2010, there were 176 microbiologically confirmed cases of S. Typhi at QECH in Blantyre, a mean of 14/year (Table 1); 12/176 (6.8%) were multidrug resistant (MDR) to ampicillin, chloramphenicol and cotrimoxazole and 147/176 (83.5%) were susceptible to all these commonly used antibiotics. All 176 were susceptible to both ciprofloxacin and ceftriaxone. A rapid increase in isolation of S. Typhi began in Blantyre in 2011, with 67 cases that year, 186 in 2012, 843 in 2013 and 782 in 2014. Although the total number of blood cultures taken has varied, the proportion of blood cultures yielding S. Typhi has risen from a long-term trend of ≤0.3% before 2011 to 5.7% of all blood cultures in 2014. The minimum incidence of Typhoid fever for urban Blantyre was estimated at 9.1/100,000 in 1998 and was stable until 2010 (1.4–9.1), but rose to 23.4/100.000 in 2011 and was 184/100,000 in 2014. A seasonal pattern has become apparent, with peaks of cases towards the end of the wet season and early in the dry season, when the prevalence of malnutrition is also highest (Fig 1).
In 2014, 754/782 (97%) were MDR. One isolate with phenotypic DCS was detected in 2011(MIC 0.064mg/l), but none with outright ciprofloxacin resistance or ceftriaxone resistance have been detected as yet. There was greatest diversity of antimicrobial resistance patterns in 2010 and 2011 (Table 2); in 2010, the year before the increase began, 6/18 (33.3%) isolates were MDR and in 2011 42/67 (63.8%) were MDR (Table 3). To address the possibility that a greater diversity of strains was circulating in these years, isolates from these years were sampled most heavily for whole genome sequencing.
As is usual with Typhoid [6], the burden has fallen most heavily on children and young adults, with the majority of cases in children aged <16 years (1075/1693 [63%]). The age distribution since 2011 is depicted in Fig 2A. Of note, the median age (Fig 2B) has fallen from 14 (IQR 8–24) in 2012 to 12 (IQR 6–20) in 2014 (p = 0.01), as the increase stabilized.
A retrospective review of 77 adult admissions from the first 2 years of the increase was undertaken. 73/77 had typical presentations, consisting of fever without focus, or one or more of headache, severe sepsis or gastrointestinal symptoms. Although 5/77 (6.5%) were confused, none had any focal neurological syndromes similar to those reported in the outbreak in Neno, Malawi [19]. Only 1 patient (1.3%) had intestinal perforation. HIV status was known in 48 cases; 8/48 (16.7%) were HIV infected, which is comparable with the prevalence for the general population of Blantyre (17.8%), but is lower than for febrile adult admissions to QECH where 90% are HIV reactive [20]. 3/73 (4.2%) adult patients died.
A review of 330 paediatric admissions since the increase revealed 7 (2.1%) deaths. 212/248 (85.5%) children for whom fuller clinical details were available had a typical presentation with non-focal sepsis and a further 52/248 (21.0%) had disease complicated by one of the following; anaemia requiring transfusion (16 [6.5%]), meningism or decreased conscious level (13 [5.2%]), pneumonia (12/248 [4.8%]), intestinal perforation/peritonitis (9/248 [3.6%]), hepatitis (1/248 [0.4%]) and myocarditis (1/248 [0.4%]). All 7 deaths occurred in the group of patients presenting with complex disease including 4/9 intestinal perforation/peritonitis and the one patient with myocarditis. There were no cases of focal neurology or cranial nerve lesions [21]. HIV data were not available but the paediatric inpatient seroprevelence obtained through routine testing during this time was 13.6%.
Overall, the case fatality rate from culture confirmed Typhoid fever in adult and paediatric admissions to QECH following the increase in Typhoid fever was 2.5%.
112 S. Typhi isolates from 2004–2013 were selected for whole genome sequencing to give the maximum temporal and antimicrobial susceptibility diversity of S. Typhi isolates contained within the MLW strain collection (see S1 Table for accession numbers). The years 2010 (18 isolates) and 2011 (67 isolates) showed greatest diversity of antimicrobial resistance profiles and a rise in the proportion which displayed the MDR phenotype, suggesting that these were the years in which a novel clade or clades were most likely to have become prominent in Blantyre. Proportionately more isolates were therefore sequenced from these years.
Maximum likelihood phylogeny of the 112 isolates placed in the context of a collection of S. Typhi isolates previously characterised by haplotyping revealed that five clades of S. Typhi were isolated from patients presenting to QECH, but that the recent increase was dominated by one clade, previously described as the “H58-haplotype”(Fig 3)[14]. H58 was first observed in this collection of isolates in 2009 (Fig 4), although this isolate was on a different branch to the majority of the rest of the H58 isolates. It suggests that, prior to 2011, the sporadic Typhoid diagnosed in Blantyre was caused by a diversity of S. Typhi clades, but that the H58-haplotype rapidly expanded in 2011. In this study, the H58 haplotype was much more strongly associated with MDR (89.3%) than the other types of S. Typhi circulating (21.4%). Plasmid finder and in-silico PCR were used to investigate the accessory genome of the H58 S. Typhi isolates. Both suggested that the IncH1 plasmid was not present in any of the H58-isolates. Review of the chromosomes of these isolates revealed that the MDR region carried on an IncH1 plasmid by reference strain CT18 has integrated into the chromosome of these isolates on a Tn21-like element between genes cyaA and cyaY. This has been observed previously in a Zambian epidemic [22].
Seventeen years of longitudinal surveillance data demonstrate that S. Typhi was an uncommon cause of BSI in adults and children in Blantyre, Malawi prior to 2011, and that an increase of MDR Typhoid fever, initially involving a wide range of clades but subsequently dominated by the H58-lineage began in 2011. This increase has now stabilized, but shows no sign of abating. This finding is consistent with a number of recent reports from Eastern and Southern Africa of outbreaks of Typhoid [21,23,24], which poses the questions, why is Typhoid emerging in Southern and Eastern Africa and how can it be controlled?
The rapid emergence of the H58-lineage of S. Typhi is likely to be associated with its spectrum of antimicrobial resistance. Chloramphenicol and amoxicillin are widely available in the community in Blantyre district and cotrimoxazole prophylaxis therapy is used as part of the ARV-programme. This clade is associated with resistance to all three agents and the widespread and poorly regulated use of these antimicrobials is likely to maintain this phenotype. Although S. Typhi and NTS appear to cause BSI in different patient groups, all 3 outbreaks of invasive Salmonella disease in Blantyre have shared this resistance pattern [5]. In contrast, DCS is a common feature of H58-S. Typhi in South Asia, therefore it is notable that DCS has yet to be reported in Blantyre. It is also important to note that since the increase, Typhoid fever has been associated with significant mortality and with complicated disease causing morbidity, despite the availability of fluoroquinolones for treatment, which have been available for treatment at QECH since 2002.
There have been two previously documented discrete epidemics of iNTS disease in Blantyre, the first caused by Salmonella Enteritidis (1999–2002), and the second by S. Typhimurium (2003–2010). iNTS disease has been associated with malaria, malnutrition and HIV [25], whereas malaria does not appear to be a risk factor for Typhoid [26] and HIV may be protective against Typhoid [27]. In Blantyre, there has been a rapid and successful rollout of antiretroviral therapy, including an increasingly effective prevention of mother to child transmission programme [20]. Whilst this study does not provide evidence that HIV is protective against Typhoid fever, there is no suggestion that it is a predisposing factor as is the case for NTS. There has also been an expansion of malaria control interventions. Both of these interventions are predicted to reduce the proportion of the population that is susceptible to iNTS disease. It is also possible that the previous S. Enteritidis epidemic, which like S. Typhi is a member of Salmonella serogroup D, led to heterotypic immunity to Typhoid fever in the population of Blantyre, which 10 years after the end of the epidemic is now in decline and this possibility should be explored further. However, it is also possible that the arrival of the H58 MDR-associated haplotype that has been broadly reported in Asia and locally in Africa has in itself been a driving factor in this rapid increase in disease. Again, further work is required to determine this.
Typhoid control has previously been associated with improvements in drinking- water quality, sanitation and hygiene practices [6,28–30]. This outbreak has occurred following an increase in improved water source coverage in Malawi, however there are reports that many of these water sources may have become contaminated and river water pollution in the city has become a problem [31]. Use of inadequately treated sewage water and human faecal manure from ecological sanitation latrines for growing vegetables must also be considered as potential contributors to this increase [32]. A recent survey of the residents of informal settlements in Blantyre found that only 7% of respondents practiced hand hygiene after defecation [33]. These pressures have been compounded by a dramatic increase in population growth; the population of Malawi has been growing at an estimated 2.8%/year based on national census projections, and Malawi is also urbanizing at the rate of 3% p.a. [34]. These increases in urban population density without proportionate improvement in access to water and sanitary facilities may have played a role in facilitating this outbreak.
We report microbiological data from a single center in Blantyre, however QECH is the sole free-of-charge hospital in Blantyre, and this finding is likely to be generalizable to the rest of urban Malawi. The clinical data are retrospective, based on available case-notes. Although the DCS phenotype was not detected, the isolates were not tested against nalidixic acid, which is the most sensitive method of detecting emerging DCS by disc testing [35]. It was only possible to sequence a selection of isolates from the MLW archive and we have sampled fewer isolates from low-incidence years, so we cannot be sure precisely when H58-S. Typhi first arrived in Blantyre.
There has been rapid emergence of Typhoid in Blantyre not associated with HIV infection, initially involving a wide range of clades, but dominated by the MDR H58-lineage of S. Typhi. This is one of an increasing number of reports of outbreaks of Typhoid ever from the region. There is a critical need for a comprehensive description of the clinical and molecular epidemiology of this neglected tropical disease across Africa, in order to understand its true burden, to model its transmission dynamics and to inform vaccination trials.
|
10.1371/journal.ppat.1003049 | A Forward Genetic Screen Reveals that Calcium-dependent Protein Kinase 3 Regulates Egress in Toxoplasma | Egress from the host cell is a crucial and highly regulated step in the biology of the obligate intracellular parasite, Toxoplasma gondii. Active egress depends on calcium fluxes and appears to be a crucial step in escaping the attack from the immune system and, potentially, in enabling the parasites to shuttle into appropriate cells for entry into the brain of the host. Previous genetic screens have yielded mutants defective in both ionophore-induced egress and ionophore-induced death. Using whole genome sequencing of one mutant and subsequent analysis of all mutants from these screens, we find that, remarkably, four independent mutants harbor a mis-sense mutation in the same gene, TgCDPK3, encoding a calcium-dependent protein kinase. All four mutations are predicted to alter key regions of TgCDPK3 and this is confirmed by biochemical studies of recombinant forms of each. By complementation we confirm a crucial role for TgCDPK3 in the rapid induction of parasite egress and we establish that TgCDPK3 is critical for formation of latent stages in the brains of mice. Genetic knockout of TgCDPK3 confirms a crucial role for this kinase in parasite egress and a non-essential role for it in the lytic cycle.
| Toxoplasma gondii, an important human pathogen, is an obligate intracellular parasite, thus getting in and out of cells is key for its survival. The process by which Toxoplasma exits cells, known as egress, is controlled by calcium fluxes and can be triggered by ionophores. In vivo, rapid egress from the host cell has been identified as a means to escape attack by the innate immune system. At the molecular level, calcium dependent events in this parasite are regulated in part by plant like calcium dependent kinases, which share no homology to human kinases and are thus ideal drug targets. In this study we revisited 4 mutant parasite lines that were independently selected for an inability to undergo egress in response to ionophores. In all four mutants we have identified the Calcium Dependent Kinase 3 as the gene responsible for the defects. We have shown that two of these mutants, which are in a genetic background that allows virulence studies, also have a strong phenotype in vivo. That is, the parasites fail to form latent stages in mice. This work provides important information that a single kinase is responsible for the formation of latent stages that are important for transmission of the parasite.
| The obligate intracellular parasite Toxoplasma gondii chronically infects a third of the world's human population. While most infections are asymptomatic, in immunocompromised individuals such as those with AIDS, leukemia and lymphoma, new infections or rupture of pre-existing latent cysts can lead to toxoplasmic encephalitis [1]–[3]. Additionally, in congenital infections, the disease can lead to severe neurological problems or even death of the developing fetus [4].
Propagation within an infected host along with some of the dire consequences of an uncontrolled T. gondii infection are a direct result of its lytic cycle, which includes attachment to a host cell, invasion and egress [5]. The secretion of adhesins involved in attachment and the gliding motility of the parasite [6]–[8] and of perforins needed for egress [9], as well as the cytoskeletal rearrangements seen during T. gondii motility, invasion and egress [10] have been shown to be dependent on Ca2+ signaling [11]. The calcium critical for initiation of motility is released from intracellular compartments [12] and calcium regulated proteins such as calmodulin, centrins and calcium-dependent kinases (CDPKs) play important signaling roles [13]. In particular, TgCDPK1 has been shown to be upstream of a signaling pathway regulating microneme secretion during invasion and egress [14]. The relation between calcium fluxes and events involved in motility, invasion and egress is particularly evident in experiments with the calcium ionophore A23187. This ionophore induces intracellular parasites to become motile and exit their host cell in a process known as ionophore-induced egress (iiEgress) [15]. Similarly, when exposed to A23187, extracellular parasites activate the secretory and cytoskeletal events required for invasion [10], [16]. Prolonged ionophore exposure while extracellular causes T. gondii to irreversibly lose its ability to invade host cells, presumably due to the exhaustion of essential invasion factors [10]. This inhibition causes the parasite to die, and therefore is referred to as ionophore-induced death (iiDeath) [10]. Thus, both of these phenomena, iiEgress and iiDeath, can be used to dissect the processes that lead to the calcium-dependent initiation of motility, invasion and egress.
To study the signaling events involved in the parasite's response to calcium fluxes, we took a genetic approach by designing a screen for mutant parasites, generated with N-ethyl-N-nitrosourea (ENU), unable to exit the host cell after induction of egress with the ionophore A23187 [17]. Two independent mutant lines, MBE1.1 and MBE3.1, were established in this manner. While >95% of wild-type vacuoles are lysed after two minutes of exposure to A23187, less than 5% of mutant vacuoles were lysed at the same time point [17]. Importantly, unlike the egress phenotype in the TgCDPK1 knockdown strain, microneme secretion and gliding motility were not found to be impaired in these mutants when extracellular. Nonetheless, these mutants were shown to have a delay in parasite-dependent permeabilization of the host cell membrane [17] an event mediated by the micronemal perforin-like protein TgPLP1 [9]. A phenotypic difference between the MBE1.1 and the MBE3.1 strains was observed when testing for iiDeath: while MBE3.1 showed a normal iiDeath phenotype (i.e. parasites die after prolonged exposure to A23187), MBE1.1 parasites were significantly resistant [17]. To investigate the possibility that iiEgress and iiDeath are genetically related phenomena, a second selection was performed for parasites that survive extracellular A23187 treatment. Two mutant lines with iiDeath resistance were established in this manner, one of which also shows a severe delay in iiEgress (MBD1.1) and one that shows normal iiEgress (MBD2.1) (Table 1) [17]. More recently, we identified a mutant, 52F11, with both a delay in induced egress and extracellular resistance to the ionophore by screening through a signature-tagged library of mutants for those with an iiEgress phenotype [18]. Besides its ionophore-dependent phenotypes, 52F11 was less virulent than its parental strain in mice. Interestingly, a second independent clone from the same library, 91E4, originally selected for a reduction in virulence in mice, also exhibits a delay in iiEgress [18]. The fact that two independent strains show both a defect in iiEgress and in vivo cyst formation indicate that these two phenotypes are likely genotypically connected. Thus, in total we have a collection of six independent mutants that fall into three phenotypic categories: defective in both iiEgress and iiDeath (MBE1.1, MBD1.1 and 52F11, 91E4), defective in only iiEgress (MBE3.3) and defective in only iiDeath (MBD2.1) (Table 1).
The independent isolation of 4 mutants that have both iiEgress and iiDeath defects supports the idea that these two processes are genetically connected [19]. Whereas most data regarding ionophore-induced egress comes from studies performed in cell culture, there are strong indications that it might play an important role in vivo. In vivo and ex vivo it has been shown that Toxoplasma egress can be triggered during interaction with immune-effector cells in a perforin-, fas- and/or antigen-dependent manner [20], [21]. Intriguingly, rapid egress is often coupled with immediate invasion of the attacking immune cell. Immune cells have a prominent role in Toxoplasma infection and several studies indicate that A) the parasite can directly modulate the immune-system by injection of effector proteins and B) it uses immune-cells as Trojan horses to cross the blood brain barrier to reach the brain, where it encysts to reach chronic infection. The fact that mutants defective in iiEgress also have a deficiency in establishing a chronic infection fits the connection between rapid egress and in vivo propagation. Through the genomic analysis of several independent mutants with defects in iiDeath, iiEgress, and in some cases the capacity to establish a chronic infection in the brain, we show here that CDPK3 is key to calcium-dependent signaling and the virulence of the parasite in vivo.
To determine the mutation responsible for the phenotypes observed with mutant strain MBE1.1, we performed a low coverage screen of its genome for candidate loci using 454 pyrosequencing. As a control we also sequenced the genome of RHΔhpt, which is the strain used to generate the iiEgress mutants. Using the assembled genome of strain GT1 as a reference [22], we assembled both the RHΔhpt and MBE1.1 genome sequences. Single nucleotide polymorphisms (SNP) between the parental and mutant strains were given a score from 0 to 1 based on the consistency with which individual reads of the mutant strain were different from that of the parental strain. We found a total of 451 SNPs with a score of .8 or higher between the parental strain and MBE1.1. Of these SNPs, 56 were within the transcribed region of predicted protein-coding genes but only 2 affected exons. The large number of SNPs in non-coding sequences between MBE1.1 and RHΔhpt, may be due to the fact that the RHΔhpt strain sequenced was maintained in culture for a prolonged period of time after the generation of MBE1.1 and likely drifted from the original parent strain.
One of the mutations found in coding sequence (A428,372G in chromosome V, based on the ME49 genome) causes a silent mutation in gene TgGT1_015610. The other is a transition mutation of a C for T change in base 4,684,212 of chromosome IX, which causes a missense mutation in TgGT1_041610. This latter gene encodes a previously characterized calcium-dependent kinase, TgCDPK3 [23], [24]. Given that this was the only missense mutation we detected in our sequences, we confirmed the nucleotide change by sequencing a PCR fragment from genomic DNA, which included the potentially affected region. This experiment confirmed the C to T transition in MBE1.1 in relation to the parental strain (figure 1). This nucleotide mutation results in a change of threonine 239 for an isoleucine in TgCDPK3. This amino acid is within the catalytic domain of TgCDPK3 (see below, supplemental figure S1A) and it is conserved in 10 of the 11 CDPKs found in T. gondii (substituted for Ala in TgCDPK4A) and in all 7 CDPKs from Plasmodium falciparum (Supplemental figure S1B). Figure 1C shows a comparison of the region around Thr 239 with that of TgCDPK1 from T. gondii and of the closest homolog of TgCDPK3 in Plasmodium falciparum, PfCDPK1.
The mutation identified lies in the last base of exon 3 (supplemental figure S2A). Thus, the possibility of aberrant splicing as a result of the mutation was assessed by sequencing of cDNA and 3′ RACE (supplemental figure S2B), which both showed TgCDPK3 is expressed with the correct splicing. While we observe that the correctly spliced form is present in the mutant strain, it is possible that the level of TgCDPK3 transcript is affected by inefficient splicing. To explore this possibility, we performed the comparative Ct method of real-time PCR to assess TgCDPK3 RNA levels in the parental and MBE1.1 strains. Using primers directed against alpha-tubulin as an endogenous control, we determined that the relative quantity of TgCDPK3 RNA in MBE1.1 was 1.12±0.19 times that found in the parental strain, indicating that the TgCDPK3 transcript is present at approximately wild-type levels in MBE1.1 (supplemental figure S3). These results show that the phenotypes observed with T239I are not due to effects of the mutation on the processing or stability of the TgCDPK3 transcript.
To determine whether the mutation identified within TgCDPK3 is responsible for the delay in ionophore-induced egress, we introduced a C-terminally HA-tagged wild-type copy of the coding sequence under the control of the T. gondii SAG1 promoter (sagCDPK3::HA), into MBE1.1. When this complemented strain is exposed to A23187 while intracellular it exhibits nearly wild-type levels of ionophore-induced egress (figure 2A): while only 14.8±5% of vacuoles of the MBE1.1 strain are lysed by 2 minutes, 100% of MBE1.1+sagCDPK3::HA vacuoles have been ruptured by the same time point (figure 2A). The level of egress in the complemented strain at 2 minutes was not statistically different (based on paired T test) than what is observed for the parental strain (99±1%). We observed the same complementation effect when the wild-type copy of TgCDPK3 was under the control of a 2 kb region upstream of the TgCDPK3 start codon (i.e. its own promoter, supplemental figure S4). Complementation with a TgCDPK3 allele carrying the T239I mutation of MBE1.1 did not complement the phenotype (figure 2A). To determine if the T239I mutation affected the overall expression of the exogenous protein, western blot analysis was performed on the strains complemented with the wild-type or the T239I strain. The results (figure 2B) showed approximately the same level of TgCDPK3 expression using the wild-type and mutant alleles.
Besides the delay in iiEgress, mutant MBE1.1 is also resistant to iiDeath. To investigate whether the T239I mutation in TgCDPK3 is also responsible for this phenotype we tested the extracellular sensitivity to calcium ionophore of the complemented strain (figure 2C). When extracellular parasites of either the parental, the mutant MBE1.1 or the wild type-complemented strain are treated with 1 µM A23187 for 45 minutes we observe significantly higher levels of survival in the mutant strain as compared to either the parental or complemented parasites (33.7% vs. 10.3% and 2.1%, respectively, significance determined by Anova).
As part of their molecular characterization of TgCDPK3, Sugi et al. [24] investigated the localization of the protein using antibodies generated in mice against a GST-CDPK3 fusion. They reported that in intracellular parasites TgCDPK3 is localized to the cytosol and partially to the apical end of the parasite, while in extracellular parasites the protein is located solely at the apical end. Interestingly, staining MBE1.1+CDPK3::HA parasites with HA antibodies showed that the transgenic protein was clearly localized to the periphery of both intracellular and extracellular parasites (figure 3A). In neither intracellular nor extracellular parasite did we observe distinct, apical localization.
To eliminate the possibility that this localization is the result of possible over-expression of the transgenic protein, we tagged the endogenous locus of TgCDPK3 with an HA tag in a parasite line lacking the KU80 gene (which decreases the rates of non-homologous recombination [25], [26]) to generate the parasite line RHΔKU80Δhpt CDPK3::HA. Immunofluorescence assays of this strain confirmed that the endogenous protein is in the periphery of the parasite regardless of whether the parasites are inside or outside human cells (figure 3B). To investigate whether this localization changed during calcium fluxes, we stained parasites that were induced to undergo egress with A23187 for 2 minutes. In these parasites the majority of the TgCDPK3 signal appeared to still be associated with the periphery of the parasite (figure 3B). This was the case even for parasites that were in the process of either exiting their host cell or invading a neighboring one, which can be recognized by the constriction around the parasite's body (arrows in figure 3B). Moreover, we observe that during parasite division, TgCDPK3 remains in the periphery of the mother cell and it is not present in the inner membrane complex (IMC), which is associated with the nascent parasites (figure 3C) and we detected with antibodies against IMC3 [27].
To further validate the localization of TgCDPK3::HA, we mutated the two amino acids in the N terminus of the protein that are predicted to be modified post-translationally: the glycine at position 2 is predicted to be myristoylated (Myristoylator) [28] and the adjacent cysteine at position 3 is predicted to be palmitoylated (CSS-Palm 2.0) [29]. To determine whether localization to the membrane is dependent on these amino acids, we replaced them both for alanine in the rescue vector and introduced this into the MBE1.1 mutant and assessed both localization and function. In parasites expressing TgCDPK3(G2A, C3A) containing an HA tag, we detect the protein throughout the cytoplasm with little if any specific association with the parasites' plasma membrane (figure 4A). This suggests that modification at the N terminus of the protein is likely required for proper localization of TgCDPK3. Surprisingly, however, we observed that TgCDPK3(G2A,C3A)::HA was able to completely rescue the iiEgress phenotype of MBE1.1. While this might indicate that localization is not critical for the function of TgCDPK3 in induced egress, it is also possible that when over-expressed (as occurs with the strong SAG1 promoter, supplemental figure S4C), enrichment in the periphery is not necessary to complement the mutant phenotype. To address this possibility, we expressed TgCDPK3(G2A,C3A) under its own promoter in the MBE1.1 strain. While the wild-type TgCDPK3 expressed off its own promoter complemented the iiEgress phenotype of MBE1.1, this was not the case with the TgCDPK3(G2A,C3A) allele (figure 4B). This indicates that, under normal expression conditions, localization of TgCDPK3 to the membrane is needed for its function in iiEgress.
Besides MBE1.1, the original selection for mutants with a delay in iiEgress also resulted in the isolation of MBE3.3, which unlike MBE1.1 does not have a resistance to iiDeath (Table 1). Selecting directly for resistance to ionophore-induced death resulted in the isolation of MBD1.1, which also has a delay in iiEgress, and MBD2.1, which has normal induced egress (Table 1) [17]. We sequenced the TgCDPK3 locus in all of these strains and whereas MBE3.3 and MDE2.1 both harbor a WT CDPK3 sequence (data not shown), MBD1.1 carries a mutation in CDPK3, resulting in a Leu to Pro conversion in amino acid 184 (figures 5 A and D and supplemental figure S5), which lies within a helical region [30]. Introduction of a wild-type copy of TgCDPK3 under the SAG1 promoter complements the iiEgress phenotype of MBD1.1 (figure 5E). Thus, of the 4 mutants isolated by Black et al. only those that have both iiEgress and iiDeath phenotypes carry mutations in TgCDPK3, which strongly points towards this protein being essential in these processes.
As externally triggered egress has been suggested to be important for dissemination of the parasites, we previously screened a panel of chemically mutagenized, signature-tagged mutants that had decreased virulence in mice and identified two, 52F11 and 91E4, that also to have a defect in iiEgress and iiDeath [18]. We sequenced the TgCDPK3 locus of both and identified a mutation in TgCDPK3 in each (figures 5B and C and supplemental figure S5). In 52F11 we detected a G for A change, which results in a Gly for Asp change in amino acid 88 of TgCDPK3 (figures 5B, 5D and S1 and table 1). Sequencing of TgCDPK3 in the 91E4 mutant shows a C for G, which causes an Asn for Lys mutation in amino acid 204 of TgCDPK3 (figures 5C, 5D and S1 and table 1). Furthermore, the iiEgress phenotype in both of these mutants is complemented by the wild-type copy of TgCDPK3 (figure 5E). The fact that both 52F11 and 91E4 have mutations in TgCDPK3 is highly significant, as they both have a significant reduction in the number of cysts formed during mice infections [18]. Thus, TgCDPK3 appears to play an important role in vivo.
Analysis of the published structure of TgCDPK3 [30] revealed that all identified mutation sites (G88, L184, N204 and T239) lie within the kinase domain of CDPK3 and are predicted to inactivate kinase activity (figure 5A). Glycine 88 (mutated to Asp in 52F11) is part of the almost universally conserved Gly-X-Gly-X-X-Gly-X-Val domain, which anchors the non-transferable phosphates of ATP [31]. Leucine 184 (mutated in MBD1.1 to Pro) is within an alpha helix [30], [32], which would likely be disrupted by the introduction of a proline. Asparagine 204, which is affected in 91E4, is conserved in virtually all kinases and is part of the HRD catalytic domain of this class of protein kinases (HRDLKxxN) [31], [33]. Threonine 239, which is mutated in MBE1.1 is located in the kinase activation loop [31] which plays an important role in the recognition of peptide substrates [31]. The equivalent amino acid in PfCDPK1 is also a threonine and it has been shown to be auto-phosphorylated [34]. To test if all identified mutations inactivate TgCDPK3 we expressed wild-type TgCDPK3 and versions carrying the mutations found in MBE1.1, MBD1.1 and 52F11 and 94E1 (T239I, L184P, G88D and N204K, respectively) in bacteria. We compared kinase activity against the short peptide, syntide-2, which was previously used to monitor TgCDPK3 activity [32]. Whereas wild-type TgCDPK3 showed high levels of phosphorylation of the substrate, introduction of any of the mutations found in the iiEgress mutants reduced TgCDPK activity to at most 20% of wild-type levels (figure 6).
A screen for inhibitors of recombinant PfCDPK1, the closest homolog of TgCDPK3 in Plasmodium falciparum, identified the 2,6,9-trisubstituted purine purfalcamine, which was shown to have anti-plasmodial activity in culture experiments [35]. PfCDPK1 was confirmed as the target of purfalcamine through affinity chromatography of parasite lysate [35]. In addition, although at higher concentrations than what is needed to inhibit PfCDPK1, purfalcamine was shown to render T. gondii unable to invade [35]. Consequently, we tested whether purfalcamine could affect iiEgress, which we have now shown to be TgCDPK3 dependent. Accordingly, we pre-treated intracellular parasites with different concentrations of purfalcamine and then exposed them to 1 µM A23187 for 3 minutes. We noted that at both 25 µM and 50 µM purfalcamine largely blocked iiEgress 2 minutes after addition of the ionophore (figure 7A). These concentrations are identical to the ones found to have anti-Toxoplasma activity by Kato et al. [35]. To further analyze the inhibition of iiEgress by purfalcamine we treated intracellular parasites with 25 µM purfalcamine for 15 minutes before exposing them to 1 µM A23187 for 2, 5 and 10 minutes (figure 7B). While 100% of untreated parasites undergo egress by 2 minutes, we measured 10.5±0.5%, 57.8±9.7% and 98.7±1.4% egress at 2, 5 and 10 minutes of ionophore exposure respectively when parasites were pre-treated with purfalcamine. Thus, purfalcamine causes a delay in iiEgress that is very similar to what is seen when TgCDPK3 is mutated (figure 2B) suggesting that it could be a target of purfalcamine in T. gondii.
To determine whether purfalcamine can inhibit TgCDPK3 directly, we tested its effect on recombinant TgCDPK3. As shown in figure 7C, purfalcamine specifically inhibits CDPK3 activity against syntide-2 (IC50 = 800 nM); however, even at high concentrations (25 uM) it does not completely inactivate the kinase as the EDTA control shows (the EDTA control chelates the calcium in the reaction which is required for CDPK3 activity).
To assess the impact of a complete loss of TgCDPK3 we generated a knockout strain of TgCDPK3 and analyzed its phenotype. For this purpose we made a construct, pKOCDPK3, which consists of the selectable marker hypoxanthine-xanthine-guanine-phosphoribosyltransferase (HPT) flanked by fragments of TgCDPK3, such that when it integrates into the genome by double homologous recombination it replaces a 1543 bases fragment from the TgCDPK3 locus with HPT (figure 8A). This event would eliminate an entire exon and introduce a complete gene, including a polyA addition site, driven off the opposite strand, into its middle. This should result in no functional TgCDPK3 protein being produced. Following transfection of a parental RHΔhpt strain of T. gondii with pKOCDPK3 we established stable HPT-positive clones, which were tested by PCR to determine if they harbored a disruption of TgCDPK3. Unfortunately, six independent attempts failed at producing a knockout strain, which suggested that TgCDPK3 might be essential for parasite survival. The recent creation of a parasite line, RHΔhptΔku80 [25], [26], that is deficient in non-homologous end-joining, allowed us to revisit the disruption of TgCDPK3. Using this strain and the same approach described above, we were able to establish a parasite clone, RHΔhptΔku80Δcdpk3+HPT (Tgcdpk3 KO), that was confirmed to be disrupted at the TgCDPK3 locus by 3 independent PCR reactions each with different primer sets (figure 8B).
Using the knockout strain we first tested if it recapitulated the iiEgress and iiDeath phenotypes of the TgCDPK3 missense mutants. The knockout strain exhibited a marked delay in iiEgress defect akin to the one seen with the TgCDPK3 missense mutants: while the parental strain exhibited 98.0±1.3% egress at 2 minutes, at the same time point only 1.6±2.9% of knockout strain vacuoles had lysed (figure 8C). This phenotype was also seen with a second independent knockout clone (data not shown). Just as with the TgCDPK3 point mutants, the knockout strain was also resistant to the lethal effects of extracellular exposure to the calcium ionophore (figure 8D). Given previous reports that recombinant TgCDPK3 could phosphorylate components of the glideosome motility system in vitro [24] we tested whether the knockout strain for defects in this process. We allowed parasites to settle onto glass cover slips in a buffer mimicking intracellular environment (IC buffer), which keeps the parasites in a pseudo-intracellular, non-motile state. This buffer was changed to an extracellular buffer (EC buffer) to allow initiation of motility, and after 15 minutes the parasites were fixed and stained for the surface marker Sag1, which allows us to detect the parasites as well as the trails they leave behind during gliding. We did not notice any difference in the percentage of parasites that were associated with trails between the parental and the Tgcdpk3 knockout strains (figure 8E). Furthermore, we did not detect any obvious difference in either the shape or length of the trails (data not shown), indicating that parasite motility is not affected by the disruption of TgCDPK3.
The fact that we were not able to obtain a knockout strain of TgCDPK3 in a standard (KU80-expressing) parental strain could be due to the knockout strain being outgrown by others that incorporate the selectable marker without completely deleting the TgCDPK3 locus. Using the Δku80 strain allows for enrichment of parasites with insertion of the selectable marker by homologous recombination and reduces competition with wild-type parasites. Based on the observation that parasites of the knockout strain appear to lyse a fibroblast monolayer at a slower rate than the parental counterpart we suspected that it had a defect on one or more of the steps that influence propagation: invasion, division, and normal egress. When we inspected vacuoles of the knockout strain between 36 and 60 hours post invasion (period of time when egress starts occurring) we detected none with an abnormally high number of parasites, indicating that the knockout strain does not have a delay in normal egress (data not shown). To investigate if there is a delay in invasion, we allowed parasites to invade cells for between 15 and 60 minutes and observed no difference in the percentage of parasites that invade cells between the parental and knockout strain (figure 9A).
To determine whether the knockout strain divides at a different rate from that of the parental strain we allowed parasites from each strain to invade cells for 4 hours and then after a further 20 hours we counted the number of parasites per vacuole for at least 100 vacuoles per sample. The results showed a markedly different distribution of vacuole sizes between the two strains: while the majority of parental strain vacuoles, 79.3±7.3%, had 8 of more parasites, 82.4±25.4% of Tgcdpk3 knockout vacuoles had 4 or fewer parasites per vacuole (figure 9B). This indicates that the knockout strain divides at a slower rate, which likely accounts for the overall slower propagation observed.
Pathogenic intracellular parasites, including Plasmodium spp., the causative agent of malaria and T. gondii rely on calcium-signaling and kinases to elicit specific cellular responses and to interact with their hosts. They differ from their mammalian hosts, however, in having a family of calcium-dependent protein kinases (CDPKs). These unique kinases are abundant in plants (Arabidopsis thaliana has 42 CDPKs [36]), ciliates, and parasites of the phylum Apicomplexa, but are absent from animal cells. Consequently, parasite CDPKs have been suggested and studied as potential targets of drug therapies. With this in mind significant effort is being put forward to characterize the function of the many CDPKs found in T. gondii and P. falciparum [37]. Using a forward genetic approach, we have determined that TgCDPK3 is key to the parasite's response to calcium fluxes in T. gondii and shown that the virulence of strains harboring a mutation in this gene is substantially reduced in a mouse model [18].
The function of TgCDPK3 within the life cycle of T. gondii had not previously been elucidated although in vitro experiments using recombinant protein have shown that TgCDPK3 is capable of phosphorylating T. gondii Aldolase 1, an important component of the gliding motility machinery of this parasite [24]. While a role for TgCDPK3 in modifying components of the so-called glideosome is consistent with the fact that motility is Ca2+-dependent and our results that TgCDPK3 localizes to the periphery of the parasite, mutants in TgCDPK3 show no observable defect in motility [17]. Thus, the biological relevance of in vitro Aldolase phosphorylation, at least as a function of motility, is unclear.
Interestingly, the closest homolog of TgCDPK3 in Plasmodium falciparum, PfCDPK1, is also localized to the periphery of the parasite in its merozoite stage [38], [39] and recombinant protein can phosphorylate glideosome components in in vitro tests [38]. Specifically, recombinant PfCDPK1 can phosphorylate both the myosin light chain (MTIP) and the glideosome-associated protein GAP45 [38]. Both of these proteins are phosphorylated in the parasite [38], [40] and one of the GAP45 sites phosphorylated in vitro was shown to also be phosphorylated in vivo suggesting that PfCDPK1 might be responsible for this phosphorylation within the parasite. This modification of GAP45 could indicate a role in assembly of the glideosome since it has been shown that the phosphorylation state of GAP45 regulates assembly of the motor complex in T. gondii [41]. However, that fact that we do not observe a measurable phenotype in gliding motility in TgCDPK3 mutant parasites suggests that, at least in T. gondii, a role for TgCDPK3 in motility per se is unlikely.
While it is tempting to speculate that TgCDPK3 and PfCDPK1 directly regulate the function or assembly of the motility apparatus given their localization, other mechanisms by which these proteins might influence Ca2+-dependent motility should be considered. Motility in both T. gondii and Plasmodium depends on highly regulated fluxes of calcium and, prior to that, sensing of potassium levels, as only low levels of potassium allow motility, egress or invasion [42]–[44]. Anything that disturbs this ion homeostasis would likely affect events such as motility, invasion and egress. This is underscored by the observation that a knockout of a sodium/hydrogen exchanger, TgNHE1, results in a dysregulation of Ca2+ homeostasis and a delay in iiEgress [19]. Several plant CDPKs have been implicated in the regulation of ion homeostasis by phosphorylating proteins such as inward-rectifying K+ channels [45], [46], Ca2+ ATPases [47] and H+ pumps [48]. Thus, it is plausible that, upon activation by calcium, TgCDPK3 controls the amplitude and duration of the calcium fluxes by phosphorylating ion channels and transporters.
While a reduction of the calcium fluxes induced by the ionophore would delay egress, it could also allow extracellular parasites to resist the detrimental effects of exposure to the ionophore. Thus, effects on the amplitude or duration of calcium fluxes could be a factor connecting the iiEgress and iiDeath phenotypes that we observed. Defects in ionic homeostasis could also easily explain the delayed division of the knockout strain, as many steps of division and cytokinesis are calcium-dependent. Nonetheless, there are many other potential connections between all the phenotypes observed, including energetics, motility and secretion. Alternatively, TgCDPK3 could have multiple substrates and the phenotypes observed in the mutants could be due to parallel rather than overlapping pathways with only the kinase in common. Regardless, a better understanding of the mechanisms connecting the phenotypes requires the identification of the TgCDPK3 targets.
The main phenotype of our TgCDPK3 mutant strains when grown in culture is an inability to respond to artificially induced calcium fluxes. Careful phenotypic analysis of these mutants did not reveal any delay in normal egress parasite division, motility or micronemal secretion, in vitro, although a slight defect in invasion was detected [17]. This contrasts with results using purfalcamine, a potent and reportedly specific inhibitor of PfCDPK1, which completely inhibited the ability of extracellular T. gondii parasites to invade cells, leading to the conclusion that TgCDPK3 is likely to be essential for invasion [35]. Given our results, it seems likely that in T. gondii, purfalcamine has other targets besides TgCDPK3, as we do not observe a strong invasion phenotype in parasites lacking CDPK3 activity. It is interesting to note that we observed inhibition also of recombinant TgCDPK1 (data not shown), which has been associated with regulating invasion in Toxoplasma gondii. But whatever the other targets of purfalcamine are, we were able to show that purfalcamine inhibits TgCDPK3 and that iiEgress is affected in a manner similar to what is seen when TgCDPK3 is mutated.
The presence of two independent knockout lines proves the nonessential role that TgCDPK3 plays in the lytic cycle in cell culture. However, complete loss of TgCDPK3 function also results in a replication phenotype that was not observed in the mutant parasite lines. Whether this is a result of complete absence of TgCDPK3 activity (we cannot exclude minimal residual kinase activity in the mutants in vivo) or genetic manipulation of the locus remains unclear at the time. But whatever the cause for the growth phenotype might be, our data suggest that parasites lacking TgCDPK3 have an egress phenotype undistinguishable from the mutants.
A reason why in tissue culture our mutants only exhibit egress phenotypes during artificial induction is likely due to the fact the iiEgress phenotype represents a delay of only about 10 minutes which is not measurable in the normal 48–72 hour lytic cycle in vitro [49]. The function of TgCDPK3 will be critical in settings in which the parasite is required to rapidly exit a host cell. One such context is during the acute stages of an in vivo infection. Tomita et al. showed that in peritoneally infected mice, most T. gondii parasites divide at most 2 times before being induced by inflammatory cells to undergo egress in a calcium-dependent manner [21]. A possible mechanism for this in vivo rapid egress is suggested by the work of Persson et al. in which they observed that either death-receptor ligation or perforin exposure can trigger egress from infected T cells through the induction of intracellular calcium fluxes [20]. Consistent with this observation, it has been shown that after intraperitoneal inoculation of T. gondii into mice, parasites “jump” from infected dendritic cells (DC) to natural killer (NK) cells by a process that includes perforin-dependent killing of the infected cell and calcium-dependent egress from the dying cell [20]. Such cell jumping has also been observed in vivo [50]. Whether such a phenomenon is related to the decreased virulence of the 52F11 and 94E1 mutants [18] will require detailed in vivo characterization of these and complemented mutants. The results presented here, however, make clear that TgCDPK3 is a pivotal protein in the parasite's life in vitro and in vivo.
RH strain parasites lacking a functional hpt gene, RHΔhpt [51] and mutant parasites MBE1.1, MBE3.3, MBD1.1, MBD2.1, 52F11 and 91E4 [17] were maintained by passage through human foreskin fibroblasts (HFFs, culture cells obtained from ATCC) at 37°C and 5% CO2. Normal culture medium was Dubelcco's Modified Eagle Medium (DMEM) supplemented with 10% FBS, 2 mM L-glutamine and 100 units penicillin/100 µg streptomycin per ml. Ionophore assays were performed using Hanks Balanced Salts Solution (HBSS) supplemented with 1 mM MgCl2, 1 mM CaCl2, 10 mM NaHCO3, 20 mM Hepes, pH 7.2 (HBSSc). The calcium ionophore A23187 (Sigma) was dissolved in DMSO at 1 mM to make a stock solution.
Extracellular parasites from strains MBE1.1 and RHΔhpt (the parental strain) were purified through a 3 microns filter to eliminate human cell contamination. Genomic DNA from both strains was isolated using the DNeasy Blood and Tissue Kit (Qiagen). The DNA was then processed for emulsion PCR and sequenced on the Roche 454 GS FLX titanium platform. Sequencing resulted in a total of 674,003 reads (228,308,310 bases). Adapter sequences were removed from the raw reads using cross_match v1.08, a part of the Consed software package [52], and regions of poor quality (based on PHRED score) excluded using Lucy [53] with parameters: minimum good_sequence_length = 50, max_avg_error = 0.002, and max_error_at_ends = 0.002. Reads were further filtered using BLAT [54] to map the reads to the reference genome, and reads that matched over less than 70% of their length, or had 4 or more inserts were excluded [54] to map the reads to the reference genome, and reads that matched over less than 70% of their length, or had 4 or more inserts were excluded. After filtering, 147,903 sequence reads (51,249,990 bases) remained for the MBE strain, and 170,239 reads (60,381,739 bases) for the RH strain. This set of 318,142 filtered reads was then mapped to the reference genome using the Roche gsMapper software, where 316,992 (99.13%) of the reads were successfully mapped. The resulting differences file was parsed using a newly developed Java application, PyroSNP, to identify potential SNPs. The PyroSNP program parses the individual blocks (loci) in the 454AllDiffs file and identifies potential SNPs between the two mapped strains. Each potential SNP detected is assigned a “SNP score” ranging from 0 to 1 where a score of 0 indicates that a polymorphism occurs in the reads for each strain equally, and a score of 1 indicates that all the reads for each strain have a unique polymorphism. The specific algorithm is detailed in the PyroSNP documentation. Thus, the SNP score quantifies how consistently different samples are at this particular locus. Additionally, the PyroSNP application incorporates read quality information from the Roche454 output, allowing users to distinguish SNPs of low quality. The output from the PyroSNP parser application was then visualized using the SNPviewer application also developed to aid in SNP detection through complex sorting and the ability to view a potential SNP in its genetic context. The PyroSNP package is available at: (http://www.uidaho.edu/research/ibest/Tools).
Mutation in MBE1.1 was confirmed by sequencing a fragment of genomic DNA from both the parental and mutant strains obtained by PCR. The fragment was amplified using primers 5′ gcgcgttctcaggatgttcgt 3′ and 5′ cagtgtatctgcaacaaccaga 3′ and encompassed bases 4,683,601 to 4,684,889 of chromosome IX.
To obtain the sequence of the TgCDPK3 transcript, total RNA from RHΔhpt and MBE1.1 parasites was purified using RNeasy purification kit (Qiagen) and reverse transcribed using an oligo dT primer and the SuperScript III First Strand Synthesis System (Invitrogen). The resulting cDNA was used as a template for a PCR reaction using primers 5′-gaggaggcgagttgttgac-3′ and 5′-ggagcagaacttgacgatcc-3′. The resulting fragment was cloned into pCR2.1-TOPO TA vector (Invitrogen). Five different clones were sequenced using T7 and M13reverse.
For 3′ RACE, cDNA from either RHΔhpt or MBE1.1 parasites was used as a template in nested PCR reactions using the GeneRacer (Invitrogen) 3′ RACE primers in combination with 5′-gaggaggcgagttgttgac-3′ (for the primary reaction) and 5′-gctctctggcaccacttacc-3′ (for the nested reaction). The resulting fragment was cloned into pCR2.1-TOPO TA vector (Invitrogen). Five different clones were sequenced using T7 and M13reverse.
To sequence the TgCDPK3 locus in the different iiEgress mutants genomic DNA was obtained with DNeasy kit (Qiagen). PCR fragments were obtained using 9 different primer pairs as to span the entire genomic region (Supplemental figure S6) and sequenced directly.
Total RNA was isolated from intracellular parasites using the RNeasy Plus Mini Kit (Qiagen). The RNA was reverse transcribed using an oligo dT primer and the SuperScript III First Strand Synthesis System (Invitrogen). The cDNA was used for Comparative Ct real-time PCR. Briefly, the cDNA was used in conjunction with 2x SYBR Green Master Mix (Applied Biosystems) and primer pairs directed against CDPK3 (5′-CTTGTCATGGAGGTGTACCG-3′ and 5′-GTAAGTGGTGCCAGAGAGCA-3′) or alpha tubulin (5′-ACGCCTGCTGGGAGCTCT-3′ and 5′-TCGTCACCACCTCCAATGG-3′), which was used for normalization. Real time PCR was carried out using the StepOne Plus Real Time PCR system (Applied Biosystems).
The efficiency of egress after calcium ionophore exposure was determined using established protocols [18]. In brief, parasites were added to each well of a 24-well tissue culture plate containing confluent HFFs at a multiplicity of infection (MOI) of 1. After 30 hours of growth, the parasites were incubated at 37°C in HBSS containing 1 µM A23187 calcium ionophore for time periods ranging from 0 to 10 minutes, after which the cells were fixed in 100% methanol. To visualize intact and lysed vacuoles the cultures were stained using Diff-Quik (Dade-Behring) according to the manufacturer's instructions. Percent egress was determined by dividing the number of lysed vacuoles by the total number of vacuoles for a sample. For purfalcamine experiments intracellular parasites were treated with purfalcamine (at 0.25, 0.5, 2.5, 25, 50 µM) in HBSS for 15 minutes at 37°C before adding an equal volume of HBSS with purfalcamine and 2 µM A23187 (for a final concentration of 1 µM) for 2, 3, 5 or 10 minutes. Cultures were fixed, stained and analyzed as described above.
Freshly lysed parasites were incubated in serum free DMEM containing 1 µM A23187 or an equivalent amount of DMSO as a solvent control for 45 minutes at a concentration of 1×105 parasites/ml. Following the 45 minute incubation, 2,000 of the treated parasites for each of the time points were then added directly to triplicate wells of a 24-well tissue culture plate containing confluent HFFs in normal culture medium. Plates were incubated for 5 days at 37°C, after which the cells were fixed with 100% methanol and stained with crystal violet to visualize plaques. The number of plaques per well was counted and the percentage survival was determined by dividing the number of plaques formed at each time point in the presence of ionophore by the number of plaques formed in the solvent control well for the equivalent time point.
The TgCDPK3 coding sequence was amplified from RH parasites cDNA with the sense primer 5′-gcggggccatggggtgcgtccaaga-3′ and the antisense primer 5′-ttaattaatcacgcgtagtccgggacgtcgtacgggtagtgcttcactttgacgtcgca-3′. This primer set introduces an NcoI site at the 5′end (single line) and an HA epitope (double line) and a PacI (thick line) at the 3′end. This fragment was digested with NcoI and PacI and cloned into the equivalent restriction sites of plasmid pHEX2 [55] to generate pSagCDPK3::HA. This results in the TgCDPK3 coding sequences being between the T. gondii Sag1 promoter and 5′UTR and the tubulin 3′UTR.
To generate the different amino acid changes in the TgCDPK3 cDNA we used the QuikChange site-directed mutagenesis kit (Stratagene) to introduce the wanted mutations in pSagCDPK3::HA. The primers used were 5′-gaaggagcgccttggcaTagcctactacattgc-3′ and 5′-gcaatgtagtaggctAtgccaaggcgctccttc-3′ for the TgCDPK3 T239I mutant, and 5′-gagtatgcatgccatggCgGCcgtccactccaagaatc-3′ and 5′-gattcttggagtggacgGCcGccatggcatgcatactc-3′ for the TgCDPK3 G2A,C3A mutant. The bases in upper case are those that are mutated in relation to the wild-type sequence.
Extracellular parasites were resuspended in equal volumes of cytoskeletal protein extract buffer and 2× SDS-PAGE sample buffer (250 mM Tris, pH 6.8, 2% SDS, 20% glycerol, 0.05% bromophenol blue, 600 mM 2-mercaptoethanol) and boiled for 5 minutes. Proteins from approximately 1 million parasite equivalents were separated on a 4–20% SDS page gel and were transferred to a nitrocellulose membrane using standard methods. The membrane was probed with a mouse anti-HA tag monoclonal antibody (Cell Signaling Technology) followed by a peroxidase-conjugated goat anti-mouse antibody (Sigma). The membrane was incubated with SuperSignal West Femto Chemiluminescent Substrate (Pierce) and was used to expose X-Omat Blue film (Kodak).
Immunofluorescence assays (IFA) were performed as described previously [19] using rabbit, anti-HA monoclonal antibodies (Rockland Immunochemical Cell Signaling Technologies) in combination with Alexa fluor-594 or Alexa fluor-488 goat anti-rabbit secondary antibodies (Molecular Probes). Slides were viewed and images were captured with a 100× objective lens on a Nikon Eclipse E100080i microscope, and images. Images were captured using a Hamamatsu C4742-95 camera. The NIS Elements deconvolution software was used to generate images.
The endogenous CDPK3 locus was hemagglutinin (HA)-tagged using the method described by Huynh and Carruthers [26]. Briefly, a 1.4 kb genomic DNA fragment, which includes the region of CDPK3 immediately preceding the stop codon, was amplified by PCR using the primers, 5′- TACTTCCAATCCAATTTAgcgtcgggatcaagacacttcc-3′ and 5′- TCCTCCACTTCCAATTTTAGCgtgcttcactttgacgtcgc-3′. In addition to containing sequences for amplifying CDPK3, these primers also contain 5′ ligation-independent cloning (LIC) sequences (underlined), which allow for the subsequent ligation-independent cloning of the PCR product. The PCR insert was cloned into the p3XHA9-LIC-DHFR vector (a gift from Vern Carruthers) using previously described methods [26]. The resulting plasmid, pCDPK3-3XHA9-LIC-DHFR was linearized with MscI, which cuts within the CDPK3 genomic fragment, and was transformed into a T. gondii strain, RHΔku80Δhpt, which has a decreased rate of non-homologous recombination [25], [26]. Transformed parasites were treated with 1 µM pyrimethamine to select for the presence of the construct. Following selection, the parasites were cloned by limiting dilution, and the clones were tested for the presence of the HA-tag by immunofluorescence using a mouse anti-HA tag monoclonal antibody (Cell Signaling Technology) in conjunction with an Alexa Fluor 594-conjugated goat anti-mouse secondary antibody.
Recombinant TgCDPK3 was amplified with Phusion polymerase (New England Biolabs) from the previously published codon optimized expression construct [30], [32] and cloned into the pET28a vector (Novagen) using the Cold-Fusion enzyme kit (Biocat) with the primers 5′-agcggcctggtgccgcgcggcagcggctgcgtgcacagcaaaaatccgc-3′ and 5′-tggtggtggtggtgctcgagtcagtgtttgacttttacatcatcgcaaattt -3′ to generate a N-HIS-tagged TgCDPK3.
Point mutations were introduced by site -directed mutagenesis according to the Phusion protocol. The proteins were expressed in BL21-Rosetta (DE3)pLysS cells (EMD Chemicals) at 18°C and purified using Ni-NTA agarose according to the manufacturer's protocol.
Activity measurements of recombinant kinase were performed using phosphorylation of Syntide-2 and subsequent scintillation counting. 100 nM recombinant CDPK3 were pre-incubated at 37°C in a kinase reaction mix (New England Biolabs) containing 100 µM CaCl2 and 100 µM ATP to allow auto-phosphorylation in a final reaction volume of 25 µL at 30 degrees Celsius. After 5 minutes the reaction was supplemented to a final concentration of 1 mM Syntide-2 and 2.5 µCi 32[P]ATP and incubated at 30°C for 8 minutes. To test the effect of purfalcamine recombinant kinase was preincubated with CaCl2 and ATP in the presence of drug or DMSO for 5 minutes prior to addition of syntide-2 and 32[P]ATP. The reactions were stopped by the addition of phosphoric acid to 75 mM and half the reaction spotted onto P81 phosphocellulose squares (Millipore), air dried and subsequently washed 3 times for 2 minutes in 75 mM phosphoric acid. Syntide 2-phosphorylation was measured using a scintillation counter.
A knockout construct was generated using the pminiGFP.ht vector [19], which contains the T. gondii HPT gene flanked two multiple cloning sites. To design the knockout construct, first a PCR fragment amplified with primers 5′-ggggtaccagccgtatctcgtagcgaataaac-3′ and 5′-ggaagcttcgtgacttctacagatttgcgtct-3′ was digested with KpnI and HindIII and cloned into pminiGFP.ht digested with the same restriction enzymes. The resulting plasmid was digested with NotI and XbaI ligated to a PCR fragment amplified with primers 5′-aagggaaaagcggccgcgcaggttccctctggtggtgtactc-3′ and 5′-gagggtcagttcctccaagcctccc-3′ and also digested with NotI and XbaI.
The resulting vector, pKOCDPK3, was linearized with NotI and 30 µg of DNA were introduced into RHΔhptΔku80 strain [25], [26] by standard methods [56]. Parasites were then grown in T25 flasks containing HFFs in normal culture medium for 24 hours at which point the media was changed to culture medium containing 50 µg/ml MPA and 50 µg/ml xanthine. After parasites stably expressing HPT were established, the parasites were cloned by limiting dilution. 12 clones were tested by PCR for the desired disruption using genomic DNA isolated with the DNeasy Blood and Tissue Kit (QIAGEN).
The efficiency of invasion was determined by allowing 2×106 parasites to invade confluent HFFs grown on coverslips for 15, 30, 45 or 60 minutes. To differentiate those parasites that have entered cells from those that remained outside we utilized an immunofluorescence-based invasion assay. In summary, following fixation with 4% formaldehyde, external parasites were labeled with an antibody against the parasite surface generated in rabbits (Abcam Labs). Cells were then permeabilized with 0.2% Triton X-100 in PBS, and all parasites were labeled with a second SAG1 antibody that had been generated in mice (a gift from Dr. Peter Bradley). The two primary antibodies were visualized with an anti-rabbit IgG secondary antibody with a red fluorescent tag (Alexa Fluor 594 - Molecular Probes) and an anti-mouse IgG secondary antibody with a green fluorescent tag (Alexa Fluor 488 - Molecular Probes) respectively. Thus external parasites were co-labeled red and green, while intracellular parasites were labeled green only. In this manner we determined the number of intracellular parasites in 20 randomly chosen fields of view for each coverslip using a Zeiss Axiovert 40 CFL microscope (400× magnification). The data was expressed as the percentage of parasites detected in 20 randomly selected fields of view that are inside cells.
Cover slips were coated with 0.1% (w/v) poly-l-lysine in water (Sigma-Aldrich) for 10 minutes and washed with PBS, then dried out at 37°C for 1 hour. A total of 1×105 freshly lysed parasites were allowed to settle down on the poly L-lysine coated glass for 20 minutes in intracellular (IC) buffer (5.8 mM NaCl, 141.8 mM KCl, 1 mM CaCl2, 1 mM MgCl2, 5.6 mM glucose and 25 mM Hepes, pH 7.2) [57]. Media was then changed to extracellular (EC) buffer (141.8 mM NaCl, 5.8 mM KCl, 1 mM CaCl2, 1 mM MgCl2, 5.6 mM glucose, 25 mM N-2-hydroxyethylpiperazine-N′-2-ethanesulfonic acid (Hepes)-NAOH, pH 7.2 [43]) and parasites were incubated at 37°C for 15 minutes. After fixation with 3% formaldehyde, samples were stained with antibodies against SAG1 using standard IFA protocols [19]. The level of motility in each treatment was determined through fluorescence microscopy by counting the number of parasites on each cover slip that were associated with a Sag1 trail in 10 randomly selected fields.
To supplement EC buffer we added KCl (37 mM final concentration), MgCl2 (8 mM final concentration), MnCl2 (10 mM final concentration), CaCl2 (10 mM final concentration) or CsCl (30 mM final concentration).
To compare rate of division of the two strains, 1×105 parasites were allowed to invade confluent HFFs for 2 hours in a 24 well plate. Wells were then washed 2 times in PBS, and re-filled with normal culture medium. 24 hours after parasite invasion, the cells were fixed in 4% formaldehyde, and parasites stained using an antibody directed against SAG1 as previously described [58]. The number of parasites per vacuole for a minimum of 100 randomly chosen vacuoles was then counted for each strain using a Nikon Eclipse 2000-5 microscope at 1000× magnification.
|
10.1371/journal.pgen.1000350 | An ALS-Linked Mutant SOD1 Produces a Locomotor Defect Associated with Aggregation and Synaptic Dysfunction When Expressed in Neurons of Caenorhabditis elegans | The nature of toxic effects exerted on neurons by misfolded proteins, occurring in a number of neurodegenerative diseases, is poorly understood. One approach to this problem is to measure effects when such proteins are expressed in heterologous neurons. We report on effects of an ALS-associated, misfolding-prone mutant human SOD1, G85R, when expressed in the neurons of Caenorhabditis elegans. Stable mutant transgenic animals, but not wild-type human SOD1 transgenics, exhibited a strong locomotor defect associated with the presence, specifically in mutant animals, of both soluble oligomers and insoluble aggregates of G85R protein. A whole-genome RNAi screen identified chaperones and other components whose deficiency increased aggregation and further diminished locomotion. The nature of the locomotor defect was investigated. Mutant animals were resistant to paralysis by the cholinesterase inhibitor aldicarb, while exhibiting normal sensitivity to the cholinergic agonist levamisole and normal muscle morphology. When fluorescently labeled presynaptic components were examined in the dorsal nerve cord, decreased numbers of puncta corresponding to neuromuscular junctions were observed in mutant animals and brightness was also diminished. At the EM level, mutant animals exhibited a reduced number of synaptic vesicles. Neurotoxicity in this system thus appears to be mediated by misfolded SOD1 and is exerted on synaptic vesicle biogenesis and/or trafficking.
| A new animal model of the human neurodegenerative disease amyotrophic lateral sclerosis (ALS; Lou Gehrig's Disease) is presented. Two percent of ALS cases result from heritable mutations affecting the abundant enzyme superoxide dismutase (SOD1). Such mutations have been indicated to impair the folding and stability of the enzyme, leading it to misfold and aggregate in motor neurons, associated with the paralyzing disease. Here, when a mutant form of human SOD1 was produced in neurons of C. elegans worms, it led to a severe locomotor defect—the worms were essentially paralyzed. The protein formed aggregates in the neurons, including an intermediate form of aggregate, soluble oligomers, that has been linked to toxicity to cells. By contrast, worms expressing the normal version of human SOD1 protein exhibited normal movement and no aggregation. The movement defect was further analyzed using chemical inhibitors and found to result from defective function of synapses, the connections made between neurons, and between neurons and muscle. Finally, in a screen using RNA interference, we observed that the worms' aggregation and locomotor condition was worsened when a class of molecules called molecular chaperones, which assist protein folding in the cell, were impaired in function. This is consistent with the idea that misfolded SOD1 is directly involved with causing the neuronal dysfunction.
| A number of neurodegenerative diseases have been associated with protein misfolding and aggregation, with a specific protein in each case observed to aggregate in a particular population of neurons. For example, in the case of amyotrophic lateral sclerosis (Lou Gehrig's Disease), a dominantly inherited form of this condition, accounting for ∼2% of cases, is associated with mutant forms of the abundant cytosolic homodimeric enzyme superoxide dismutase (SOD1), which accumulate in insoluble aggregates in motor neurons [1]–[3]. Mutational studies of SOD1-linked ALS have uncovered single residue substitutions throughout the enzyme subunit [4], and studies in vitro indicate that the substitutions generally destabilize the protein, disposing to misfolding and aggregation [5]–[7]. It remains unknown, however, exactly how apparent misfolding and aggregation of SOD1 exerts toxic effects on motor neurons. Is there a central common effect shared by the various mutant alleles that comprises a common pathway of motor neuron injury?
Mice transgenic for a variety of mutant SOD1 alleles also develop motor neuron disease resembling that of affected humans [8],[9], enabling a variety of pathological and biochemical studies. A survey of pathology reported for various alleles implicates a variety of potential physical sites of toxicity, including mitochondria, endoplasmic reticulum, and axonal traffic. For example, abnormal-appearing mitochondria have been observed in animals transgenic for G93A and G37R SOD1 [9]–[11], and several mutant SOD1's have been coisolated with spinal cord mitochondria [12]–[15]. Concerning ER function, an unfolded protein response (UPR) was observed in spinal cord of G93A mice [16], and a recent report suggests that mutant SOD1 induces this response by binding to the ER membrane component Derlin-1, blocking retrograde traffic of ER proteins to the cytosol for proteasomal degradation (ERAD) at the level of ubiquitination [17]. Concerning axonal traffic, both anterograde and retrograde transport have been observed to be retarded in mice transgenic for mutant SOD1 [18]–[24]. Which, if any, of these effects is primary to mutant SOD1-induced motor neuron damage?
One approach to resolving this question is to produce mutant human SOD1 in neurons in an invertebrate system to inspect for effects on function, with the idea that this might reveal a minimal target of the toxic effect which could ultimately be further evaluated in the mammalian system. Such an approach has been taken, for example, with other neurodegenerative disease-associated proteins, expressing them in Drosophila or C. elegans, including polyglutamine repeat proteins [e.g. 25],[26] and α-synuclein [e.g. 27],[28]. Here we have taken such an approach with SOD1 using C. elegans, programming pan-neuronal expression of a mutant version of human SOD1, G85R, that obligatorily misfolds. We observe a locomotor defect in transgenic animals expressing the mutant SOD1, associated with aggregation of the mutant SOD1 protein and with synaptic dysfunction, involving deficient numbers and possibly deficient trafficking of pre-synaptic vesicles.
To examine the effects of an ALS-associated human mutant SOD1 on a collective of neurons in an optically accessible nervous system, we produced transgenic C. elegans expressing G85R mutant or wild-type human SOD1 (referred to hereafter as SOD). G85R SOD has been identified in human cases of ALS [1], and G85R SOD transgenic mice develop a similar disease [29]. In the latter setting, the protein has been shown to behave as a misfolded monomer [30]. That is, it fails to form the normal SOD homodimer, and it lacks the normal disulfide bond that is formed between Cys 57 and Cys 146 when the protein is properly folded. (This disulfide bond is normally formed despite localization of SOD to the relatively reducing cytosol). To express the wild-type and G85R SOD proteins in as many of the 302 neurons of a C. elegans hermaphrodite as possible, a pan-neuronal promoter, the promoter of the C. elegans synptobrevin gene (snb-1), was used to drive the respective wild-type and G85R human cDNAs encoding SOD. To allow direct observation of the expressed SOD proteins, two additional constructs were employed that join a YFP reporter sequence via a flexible peptide linker to the C-terminus of the wild-type or mutant SOD [31]. Multiple stable transgenic C. elegans lines of both unfused and fused constructs were produced.
We noticed immediately that G85R SOD transgenic animals exhibited minimal forward movement across the culture medium compared with normal movement of age and protein expression-matched wild-type SOD transgenic strains (Figure 1A and Video S1; see Figure S1 for immunoblot analysis and activity blot analysis). The wild-type SOD transgenic animals exhibited a forward movement speed similar to that of the parental nontransgenic C. elegans N2 strain (Figure 1A). In addition to severely reduced forward crawling of the mutant transgenic animals, their side-to-side thrashing movement in liquid was also severely affected (Figure S2 and Videos S2, S3). The rates of forward movement of the animals transgenic for the fusion proteins, WTSOD-YFP and G85R-YFP, likewise showed a large difference (Figure 1B), although the mutant now exhibited significant forward movement, and the wild-type fusion was somewhat slower than the nonfused wild-type transgenic animals (compare Figure 1B with 1A). These same effects were observed on thrashing (Figure S2). Thus the addition of the YFP moiety has effects on the mutant and wild-type SOD proteins, exerted in opposite directions. Overall, however, even with the YFP fusion protein, there was still markedly slower movement of the G85R animals as compared with the wild-type (Figure 1B).
To assess whether the biological behavior of the G85R-YFP fusion protein has any relevance to the mammalian context where expression is associated with an identifiable clinical neuronal disorder, G85R-YFP was produced in transgenic mice from a human SOD genomic clone with YFP fused to the last coding exon. This fusion produced an ALS phenotype in mice at 3–9 months of age, with the age of onset of the motor deficit depending on copy number and expression level of the transgene (JW, GF, KF, and ALH, unpublished). By contrast, expression-matched wild-type SOD-YFP transgenic mice produced in parallel did not develop disease even at ages of 2 years and beyond. Thus the SOD-YFP fusions tested here reflect the behavior of the nonfused counterparts in the context of production of ALS-like disease in the mammalian setting.
An additional SOD mutant transgenic C. elegans strain, H46R/H48Q-YFP, containing an SOD double mutant allele that blocks copper binding by SOD and produces ALS in transgenic mice [32], was also examined. Here it produced a movement defect less prominent than that seen in G85R-YFP (Figure 1B).
Fluorescence microscopy of transgenic C. elegans at both larval and adult stages revealed fluorescence in many neurons in the nerve ring (head region), the ventral nerve cord, lateral body wall, and tail ganglia (Figure S3A). Non-neuronal fluorescence of the spermatheca and two gonadal distal tip cells was also observed. The character of fluorescence of neuronal cell bodies of G85R-YFP transgenic animals differed from that of WTSOD-YFP transgenics as exemplified by cell bodies of motor neurons along the ventral nerve cord (Figure 2A, B). [Neuronal processes, by contrast, did not exhibit altered fluorescence (Figure S3B).] Cell bodies from the wild-type animals exhibited a more diffuse cytosolic fluorescence pattern, whereas those from the G85R-YFP mutant exhibited a well-demarcated pattern, suggestive of aggregate formation. Consistent with this, a FRAP experiment (fluorescent recovery after photobleaching) showed very slow recovery of fluorescence in mutant cell bodies as compared with wild-type (see Figure S4).
Aggregates were directly observed in the mutant by EM examination following chemical fixation (Figure 2C, D), which revealed well-delineated electron-dense inclusions in the cytosol of ventral nerve cord cell bodies of G85R-YFP animals (Figure 2C, Agg.). By contrast, there were no recognizable aggregates in WTSOD-YFP cell bodies (Figure 2D). The location of G85R-YFP aggregates in a perinuclear position is reminiscent of aggresomes [33] or of perinuclear structures distinct from the centrosome known as JUNQ, involved in juxtranuclear quality control [34], but aggregates formed in the unfused G85R animals exhibited a more diffuse, “fluffy” character (Figure S5A, asterisks). When high pressure freezing/fixation was employed on the G85R animals, these regions were now observed as amorphous inclusions (e.g. Figure 2E, Agg.), which at high magnification appeared to contain loosely stacked fibrillar material (Figure 2F).
To identify genetic modifiers of neuronal aggregation in the G85R SOD-YFP transgenic animals, we carried out an RNAi screen. Neurons of C. elegans are relatively resistant to RNA interference, so alleles previously identified to enhance interference activity, eri-1 (mg366) and lin-15B (n744), were introduced [39]. Strikingly, the introduction of these alleles led to a significant reduction in fluorescence from the G85R-YFP protein as compared with the parental strain (Figure S9A), associated with a decrease in the number of fluorescent inclusions observed in the ventral nerve cord (Figure S9B) and with somewhat improved movement. These results are consistent with a previous report that lin-15B could produce silencing of multicopy transgenes [43]. We observed that either allele alone could produce substantial silencing of the G85R-YFP transgene, although lin-15B exerted a greater effect. The combination of alleles proved to sufficiently reduce fluorescence and aggregate formation of G85R-YFP to a degree that could allow for a “dynamic” range of effects of RNAi screening, i.e. both reduction and enhancement of green fluorescence patterns reflecting aggregation would be detectable. A bacterial feeding library was employed for RNAi testing. Collectives of animals at all stages were transferred onto feeding plates and visually examined by multiple observers after 3–6 days for changes in the number and intensity of fluorescent aggregates in the nerve ring and ventral nerve cord as compared with the strain fed bacteria with an empty vector.
In the general case of observing decreased fluorescent aggregates when animals were placed on a particular interfering bacterial strain, the development of the animals was generally also slowed and viability in many cases reduced, reflecting likely effects on general health or on general gene expression, and these interfering RNAs were not studied further. There was one exception to this, involving an interfering RNA for a ubiquitin specific protease, where fluorescent aggregation was reduced and viability somewhat improved, and this gene is under further study.
In the general case of animals with increased fluorescent aggregates on a particular interfering bacterial strain, these animals generally exhibited larger numbers of fluorescent inclusions and diminished locomotion. There were 88 such hits (Table S1). 7 of these appeared likely to suppress eri-1; lin-15B action and were not studied further (see last entries in Table S1), while the remaining 81 hits were categorized into 10 groups (Table 1). The largest group of well-annotated hits comprised molecular chaperones and quality control components, amounting to about a quarter of the hits. These are discussed below. The other large group, of about the same size, comprised a collective of uncharacterized gene products.
To independently confirm RNAi hits in the parental genetic background, loss-of-function alleles were obtained for 12 hits of interest and were crossed with the parental G85R-YFP strain. 11 out of 12 of these enhanced the inclusion phenotype of G85R-YFP. For example, heat shock factor 1 (HSF1), which transcriptionally regulates a number of stress components [44], registered very strongly in the RNAi screen in increasing aggregate formation. Consistently, when the sy441 allele of hsf-1 was crossed into G85R-YFP, a strong increase in aggregate formation was observed (Figure S10A), and locomotion of these animals was substantially decreased as compared with either parental strain (Figure S10B).
Multiple interference hits of strong magnitude were observed in a number of pathways (Table 2), indicating that these pathways are likely playing a role in preventing aggregation of the misfolded G85R-YFP protein or facilitating its turnover. For example, hits of a collective of chaperone components registered strong increases of aggregation, including an Hsp110 (C30C11.4), a DnaJ (A2) (dnj-19), an Hsp70 (stc-1), and a neuron specific Hsp16 (F08H9.4). Three of these were validated as strongly increasing aggregation when mutant alleles were crossed with G85R-YFP. Two hits were identified in the ubiquitin-mediated turnover pathway at the level of E3 ligase SCF complexes, SEL-10, an F-box protein, and RBX-1, a RING finger protein, the latter confirmed with the ok782 knockout allele. Three hits were also observed in the sumoylation pathway by the RNAi screen, uba-2, encoding the E1 enzyme that activates SUMO, ubc-9, which encodes one subunit of the E2 SUMO-conjugating heterodimer, and an E3 SUMO-ligase component gei-17 (homologous to PIAS1). In validation of the interference effects, an allele of SUMO, smo-1 (OK359), increased aggregation when crossed into the parental G85R-YFP strain. The latter hits raised the question of whether G85R-YFP is itself sumoylated, and this is under study. In addition, three components involved with redox regulation were identified, PDI-2, an orthologue of a human thioredoxin domain-containing protein (C30H7.2), and BLI-3, a dual oxidase with both a peroxidase domain and a superoxide-generating oxidase domain.
Other pathways were also identified through strong effects of interference. For example, a hit in the TGFβ component DBL-1 was confirmed by the allele nk3. DBL-1 is expressed in neuronal cells, e.g. in the ventral cord, has recently been implicated in GABAergic synaptic transmission [45], and has been shown to affect both body size and male tail development [46]. Notably, in mammalian contexts, TGF-β family members have exhibited effects on axon outgrowth and protection from excitotoxicity [e.g. 47]. A hit of the dopamine transporter DAT-1 was confirmed with allele tm903. A link between neuronal activity and aggregation in the cases of GABAergic and dopaminergic transmission thus seems a consideration but requires further study. In the DNA replication/repair pathways, interference with top-1, the gene for topoisomerase 1, had a very strong effect, equivalent to the strong effect of hsf-1 interference. How such action could figure into aggregation behavior remains unclear. Similarly, interference of div-1, a subunit of the DNA polymerase α/primase complex, increased aggregation. Also a strong effect was observed with interference of pha-4, a FoxA transcription factor, implicated in calorie-restriction-mediated longevity, at least in part by regulating endogenous C. elegans SOD enzymes [48].
We have described here that pan-neuronal expression of a human ALS-associated mutant SOD in C. elegans produces substantial locomotor defects associated with macroscopic aggregation in neuronal cell bodies. By contrast, a wild-type human SOD produced neither locomotor defects nor aggregation. Further testing indicated that the human SOD mutant, G85R, unable to fold properly and producing both soluble oligomers and aggregates, appears to produce neuronal dysfunction in C. elegans at the presynaptic level. This was indicated by three types of observation. First, there was a reduced number and brightness of puncta of fluorescently-labeled synaptobrevin, RAB-3, and synapsin, as well as of the presynaptic active zone protein RIM1 in the dorsal nerve cord in transgenic G85R animals as compared with wild-type SOD transgenic animals. Second, mutant transgenic animals but not wild-type exhibited resistance to aldicarb paralysis, consistent with deficient acetylcholine release at cholinergic synapses in the mutant. Third, EM studies on a small number of animals revealed paucity of vesicles in the presynaptic region of nerve ring synapses of the mutant animals. In addition, EM showed reduced numbers of vesicles and mitochondria in ventral nerve cord processes (axons) of transgenic mutant animals. In contrast with the foregoing findings pointing to presynaptic dysfunction, there was a lack of postsynaptic effects. The cholinergic agonist levamisole produced the same efficient paralysis of both mutant G85R and wild-type SOD transgenic animals. In addition, EM inspection of muscle revealed normal morphology.
The putative presynaptic defect could result, in the first instance, from defective biogenesis, axonal transport, or recycling of synaptic vesicles. The diminution of vesicles in dorsal and ventral nerve cord processes and their sparsity in the presynaptic regions as demonstrated from the EM and fluorescence studies suggest that biogenesis or transport is affected. Because accumulations of vesicles were not detected in cell bodies, it seems less likely that transport is affected. Yet the failure of recovery of the GFP-synaptobrevin fluorescence following photobleaching could be consistent with an ongoing transport or recycling defect. Further studies will be required to resolve the steps affected.
Excitingly, effects of expression of human SOD1 on synaptic transmission in another invertebrate system have recently been reported. Watson et al [49] observed that expression of human SOD1 (wild-type or mutant) in motor neurons of Drosophila produced an age-progressive climbing defect that was associated in electrophysiological studies, stimulating muscle through the giant fiber circuit, with progressive loss of muscle response during high frequency stimulation.
Presynaptic effects have also been observed in several contexts in C. elegans expressing a number of other neurodegeneration-associated proteins. Animals transgenic for both wild-type and FTDP-associated mutant forms of tau presented with locomotor defects, associated with aldicarb resistance (and levamisole sensitivity), followed by appearance of macroscopic aggregates and then apparent neuronal loss [50]. Animals transgenic for α-synuclein exhibited locomotor defects when any of a number of synaptic proteins were knocked down, associated with aldicarb resistance and levamisole sensitivity [28]. In contrast with these studies, however, the studies of SOD1 presented here indicate a direct and specific effect of the mutant G85R SOD1 on presynaptic function.
So far it is not evident how or whether the presence of G85R-SOD aggregates in cell bodies of specific neurons, comprising a variety of neuronal cell types in the transgenic mutant animals, directly relates to the apparent presynaptic defect. We note, however, that the degree of locomotor defect correlates roughly with the degree of aggregation observed. For example, for the G85R-YFP transgene, out of multiple stable integrant lines, those with the highest level of fluorescence and greatest level of aggregation exhibited the strongest locomotor defect. Thus, either the aggregates themselves or perhaps the apparent precursors, soluble oligomers (Figure 3B), may be directly responsible for the presynaptic defect. The defect could come at the level of physical interactions of oligomers or aggregates either directly with synaptic vesicles themselves or with soluble or cytoskeletal components that are involved with vesicle biogenesis and traffic. A recent study producing pan-neuronal expression of dimeric versions of human wild-type and G85R SOD1 in C. elegans also observed locomotor defects associated with aggregation, but, in contrast to the direct correlation above, observed that a heterodimeric G85R-WTSOD-GFP molecule, while producing less aggregation than G85R-G85R-GFP, led to a greater paraquat-induced reduction of animal survival [51]. Thus in the context of a heterodimer, residual SOD enzymatic activity may contribute to toxicity.
Notably, despite evident protein aggregation and synaptic dysfunction, neurons in the mutant animals studied here were not subject to cell death, even during later adult life. This resembles the observation in the recent report of Watson et al [49] where human SOD1 expressed in Drosophila motor neurons produced both focal SOD protein accumulation and measurable electrical dysfunction but no observable cell death. Similarly, in earlier studies of C. elegans PLM neurons transgenic for polyglutamine expansion, touch sensitivity function was abolished but cell death did not occur [52]. This lack of cell death may either relate to the state of disease progression or be a function of the invertebrate neuronal systems themselves. For example, concerning stage of disease, in the mammalian context, SOD1-affected motor neurons appear likely to be functionally affected for a period of time before being subject to cell death. In C. elegans or Drosophila, by contrast, the trajectory may not extend sufficiently in time to produce cell death, albeit that added insults such as oxidative toxicity [50] may be capable of producing cell death. Alternatively, these systems may differ from that of mammals. C. elegans has, for example, a limited number of glia, and they may not function as in the mammalian context to hasten death of affected neurons [53]–[55]. Alternatively, the absence of cell death could be a function of a different neuronal response to chronic exposure to misfolded protein as compared with mammalian neurons.
The RNA interference screen conducted here, examining relative levels of G85R-YFP protein aggregation in relation to knockdown of various gene products, validated in many cases by crosses with corresponding mutant alleles (Table 2), provided evidence that a proteostatic network similar to that present in body wall muscles of C. elegans [56],[57] and elsewhere [44] is operative in C. elegans neurons. Whether it is induced in response to mutant human SOD1 remains to be determined, and whether a higher level of expression of some or all members of the network could be protective remains to be tested. Two transcriptional regulators that lie at the top of such a network were identified, HSF-1 and PHA-4/FoxA, which have a broad range of targets in, for example, chaperone pathways [44] and redox regulation [48], respectively. Components within the network were also identified, including chaperones, an E3 ligase, and redox components. Additional components had effects on aggregation but their mechanism of action remains unclear, including SUMO, the TGF-β homologue, DBL-1, expressed mainly in neurons, and two components of DNA maintenance, topoisomerase I and a subunit of polα/primase complex, implicating DNA integrity in the response. Notably absent from both the interference screen and from EM studies was evidence for involvement of the autophagy system. Consistent with this, administration of rapamycin was without effect on the extent of aggregation (data not shown).
Do the phenotypic properties observed here bear any relation to mammalian disease induced by expression of mutant SOD1? Could the transgenic C. elegans inform usefully about mammalian disease? As mentioned, the G85RSOD-YFP fusion protein expressed in mice indeed produces an ALS-like disease whereas WTSOD-YFP fusion produces no ill effect. Interestingly, the presynaptic effects observed here with the G85R transgenic C. elegans may have a parallel in mutant G85R SOD transgenic mice, as reported recently by Caroni and colleagues [58], who examined motor neuron axons and NMJs in hind limb muscle of mutant transgenic animals of varying age [see also 59]. In vulnerable motor neuron axons (FF and FR), they observed at early time (7 months of age; 2 months before end-stage) localized synaptic vesicle accumulation associated with diminished overall density of vesicles, reflecting apparent stalling of vesicle traffic, followed at a later stage by severe loss of synaptic vesicles associated with loss of presynaptic active zone markers. The findings at later time generally agree with the observations here in dorsal cord of our G85R C. elegans, where both fluorescent synaptic vesicle protein markers and an active zone marker (RIM1) appear to be reduced (Figure 6). The nature of the vesicle trafficking defect in either setting remains to be elucidated. Does it reflect failure of vesicles to be produced in the first instance, as suggested by the lack of organelles in ventral cord processes here (Figure 5), and/or is it a block of recycling, as might be suggested by the FRAP analysis of GFP-synaptobrevin (Figure 6C)? Is it a direct effect of mutant SOD, forming physical association with synaptic vesicles? Or is it a secondary effect, mediated e.g. via the motor/cytoskeletal trafficking system? Questions concerning both the basis to vesicular defects and the overall pathway of toxicity of mutant SOD protein remain to be resolved.
The C. elegans snb-1 promoter, a PCR-amplified genomic DNA segment extending from minus 3021 bp to just upstream of the SNB start codon, was inserted in place of the unc-54 enhancer/promoter in the plasmid pPD30_38 (Fire Lab Vector Kit, Addgene Inc., Cambridge, MA), and the various human SOD cDNA-containing segments were then adjoined. SOD mutations were generated by PCR, and the derived coding sequences confirmed by sequencing them in entirety. Fusion constructs joining human SOD with YFP via a linker segment (LQLQASAV) were kindly provided by Dr. R. Morimoto, Northwestern University [31].
The N2 Bristol strain of C. elegans was used as the wild-type strain. Standard culturing and genetic methods were used [60]. Animals were maintained at 20°C unless otherwise indicated. Mutant strains obtained from the Caenorhabditis Genetics Center (CGC), the National Bioresource Project in Japan, and the lab of Joshua Kaplan are listed in Suppl. Table 2. Germline transformation was performed by injecting DNA solution containing 20 ng/µl of an SOD construct and 5 ng/µl of myo2::GFP into hermaphrodite gonads [61]. Multiple extrachromosomal lines were established based on the fluorescent markers. They were further treated with trimethylpsoralen/UV to generate integrated lines that stably expressed the transgenes. At least three independent stable lines were produced for each variant, and each line was backcrossed with the N2 strain four times. The transgenic SOD and SOD-YFP lines used in these studies are designated in the legend to Figure 1. nuIs152 has the transgene Punc-129::GFP-snb-1 [39]. nuIs168 has the transgene Punc-129::YFP-Rab-3 [62]. nuIs163 and nuIs165 strains containing Punc-129::snn-1-YFP and Punc-129::unc-10-GFP were the kind gift of Joshua Kaplan [63].
For high resolution imaging, animals were immobilized with levamisole and examined by either differential interference contrast (DIC) or fluorescence with an Olympus IX81 microscope equipped with spinning disk confocal illumination. For fluorescence recovery after photobleaching (FRAP), a laser confocal microscope was used (Zeiss LSM510 META).
For transmission electron microscopy, animals were prepared by conventional two-step chemical immersion fixation or high-pressure-freezing [64],[65]. Serial thin sections were prepared and post-stained with heavy metals. At least four animals of each genotype were analyzed using a Tecnai 12 Biotwin at 80 kV.
A video-based assay was used to assess the locomotion speed of C. elegans. Animals were transferred to a plate with a fresh bacterial lawn on which movement tracks could be traced. Immediately upon release, worms exhibit a maximum movement response for a short duration. A 30 second movie was shot for each worm, and the ratio of the movement distance to the body length, measured by the NIH ImageJ software, was used as a movement index (see Video S1).
Mid L4 animals were transferred to freshly made NGM agar plates without bacterial food containing 1 mM aldicarb, and at different time points the animals were prodded on the nose to determine whether they had reached complete paralysis [42]. An identical experiment with 1 mM levamisole was also performed.
To assess solubility of SOD protein, animals were disrupted by sonication on ice in 0.5 ml extraction buffer (PBS, 1 mM EDTA, 1 mM EGTA, 1 mM TCEP, with half a tablet of Complete Mini protease inhibitor cocktail (Roche)) and left on ice for 10 min to allow large debris, including cuticle, to sediment. The supernatant fraction was then centrifuged in a Beckman TLA-100 rotor at 53,000 rpm (>120,000×g) for 15 min at 4°C. Pellets were washed once by resuspension in the extraction buffer and sedimentation. Supernatant and SDS-solubilized pellet fractions were analyzed in SDS-PAGE under reducing conditions. Supernatant fractions (1 mg total protein) were also subjected to gel filtration chromatography on a Superose 6 gel filtration column (GE Healthcare) and eluted with PBS supplemented to 0.1 mM TCEP at 0.5 ml/min. Individual fractions (0.5 ml) were examined by SDS-PAGE and Western blotting. For Western analysis, antibodies to human SOD1 (SOD-100, Stressgen, Canada), antibodies raised in rabbits against purified YFP (Cocalico), or antibodies to actin (C4, MP Biomedicals, Inc., Aurora, Ohio) were used.
An RNAi feeding library of 16,757 bacterial clones was employed for screening (GeneService, Cambridge, UK). Animals at mixed ages were screened in 96-well plates at 15°C as described [66]. “Hits” were identified by an increased number and intensity of fluorescent neuronal inclusions using a Leica fluorescence stereoscope with a 2.0× PLANAPO lens. All positives were subjected to secondary screening at both 15°C and 20°C in 6-well plates. The identities of all positive RNAi clones were confirmed by DNA sequencing of the plasmid insert. For selected hits, where loss-of-function alleles were available, the corresponding strains were crossed to the G85R-YFP parental strain (line 8). The genotypes of the product strains were verified by PCR or PCR/DNA sequencing and the phenotypes studied.
|
10.1371/journal.pbio.2005512 | GABAergic modulation of olfactomotor transmission in lampreys | Odor-guided behaviors, including homing, predator avoidance, or food and mate searching, are ubiquitous in animals. It is only recently that the neural substrate underlying olfactomotor behaviors in vertebrates was uncovered in lampreys. It consists of a neural pathway extending from the medial part of the olfactory bulb (medOB) to locomotor control centers in the brainstem via a single relay in the caudal diencephalon. This hardwired olfactomotor pathway is present throughout life and may be responsible for the olfactory-induced motor behaviors seen at all life stages. We investigated modulatory mechanisms acting on this pathway by conducting anatomical (tract tracing and immunohistochemistry) and physiological (intracellular recordings and calcium imaging) experiments on lamprey brain preparations. We show that the GABAergic circuitry of the olfactory bulb (OB) acts as a gatekeeper of this hardwired sensorimotor pathway. We also demonstrate the presence of a novel olfactomotor pathway that originates in the non-medOB and consists of a projection to the lateral pallium (LPal) that, in turn, projects to the caudal diencephalon and to the mesencephalic locomotor region (MLR). Our results indicate that olfactory inputs can induce behavioral responses by activating brain locomotor centers via two distinct pathways that are strongly modulated by GABA in the OB. The existence of segregated olfactory subsystems in lampreys suggests that the organization of the olfactory system in functional clusters may be a common ancestral trait of vertebrates.
| Olfactory-induced behaviors (homing, food or mate searching, etc.) are crucial for the survival and reproduction of most animals. A neural substrate underlying odor-induced behaviors in vertebrates was recently uncovered using a basal vertebrate model: the lamprey. It consists of a neural pathway extending from the medial olfactory bulb, a first-order relay of olfactory information in the brain, to locomotor regions. Here, we investigated modulatory mechanisms acting on this neural pathway. We show that an inhibitory circuitry that releases the neurotransmitter GABA in the olfactory bulb strongly modulates motor responses to olfactory stimulation. We also discovered and characterized a novel olfactomotor pathway that originates in the non-medial olfactory bulb and consists of a projection to the lamprey olfactory cortex that, in turn, projects to locomotor regions. This discovery of a novel pathway linking olfactory and motor centers in the brain indicates that olfactory inputs can activate locomotor centers via two distinct pathways. Both pathways are strongly modulated by the neurotransmitter GABA in the olfactory bulb. The existence of segregated olfactory subsystems in lampreys sheds light on the evolution of olfactory systems in vertebrates.
| Olfactory cues can trigger goal-directed locomotor behaviors, such as homing, predator avoidance, or food and mate searching [1–11]. It is only recently that the neural pathways and mechanisms involved in transforming olfactory inputs into locomotor behavior were characterized for the first time in a vertebrate species, the lamprey [12,13]. It consists of a specific neural pathway extending from a single glomerulus located in the medial part of the olfactory bulb (medOB) to the mesencephalic locomotor region (MLR), with a relay in the posterior tuberculum (PT) [12]. In all vertebrates, the MLR acts as a motor command center that controls locomotion via descending projections to brainstem reticulospinal (RS) neurons [14–22]. This olfactomotor pathway is present throughout the life cycle of lampreys, whether in larvae, newly transformed, parasitic, or spawning animals [12]. Yet, olfactory-induced motor behaviors can be life stage specific in lampreys. For instance, at the parasitic stage, lampreys feed on fish that they detect using olfactory cues [23]. Then, when sexually mature, the adults are attracted upstream by migratory pheromones released by larvae [24–26]. Once upstream, the females are attracted to males by sex pheromones [27,28].
The general organization of the lamprey olfactory system, from the periphery to the central nervous system (CNS), is very similar to that of other vertebrates. The peripheral olfactory organ is composed of a main olfactory epithelium and an accessory olfactory organ [29–31]. Axons from olfactory sensory neurons (OSNs) of the olfactory epithelium terminate in the olfactory bulb (OB). As in other vertebrates, the OB can be divided in two subregions, based on their inputs. The main olfactory bulb (MOB), which occupies the whole OB except its medial part (i.e., the medOB), receives inputs from the main olfactory epithelium. The medOB, on the other hand, receives inputs from OSNs located in the accessory olfactory organ [32–34]. The OB of vertebrates constitutes the primary olfactory center of the CNS and, as such, filters and actively shapes sensory inputs to secondary olfactory structures [35,36]. This processing of sensory inputs in the OB is driven by modulatory inputs coming from the numerous neurotransmitter systems present in the OB of vertebrates [37,38]. GABA is the main inhibitory neurotransmitter in the CNS, and numerous GABAergic processes are present in the OB of several vertebrate species [39–43]. GABAergic neurons of the OB are believed to play a critical role in olfactory processing by providing inhibition to the bulbar microcircuitry [44]. However, their effect on the outputs of the OB and ultimately on behavior is far less understood.
Here, we hypothesized that GABAergic neurons of the OB could play a significant role in modulating transmission in the olfactomotor pathway of lampreys. To address this, we used anatomical (tract tracing and immunohistochemistry) and physiological (intracellular recordings) techniques.
The present study showed abundant GABAergic cell bodies and processes in the OB (n = 10 adult animals, Fig 1), thus confirming the findings of Meléndez-Ferro and colleagues [43]. GABAergic neurons were mainly observed in the central region of the OB (internal cell layer [ICL], Fig 1A and 1D and S1 Fig), where the most common OB interneuron type, the granule cell, was described [45]. GABAergic processes were found all over the OB, including in and around the glomeruli of both the MOB (Fig 1A and 1C) and the medOB (Fig 1A and 1B and S2 Fig).
To investigate the physiological role of the GABAergic circuitry in the OB, local microinjections of the GABAA receptor antagonist, gabazine, were made into restricted areas of the OB, while stimulating the olfactory nerve (ON) and intracellularly recording from RS neurons on the same side of the brain. Gabazine injections (0.1 mM, 1.4 ± 1.6 nL) in the medOB (n = 60 synaptic responses; n = 6 neurons; n = 6 larval animals; Fig 2) were found to amplify synaptic responses of RS neurons to electrical stimulation of the ON (amplitude increase of 372.2 ± 277.5%; p < 0.05; no statistical differences between control and washout; Fig 2B and 2C). RS neurons from all four reticular nuclei (mesencephalic reticular nucleus and anterior, middle, and posterior rhombencephalic reticular nuclei) responded similarly as shown by calcium imaging experiments (n = 362 neurons; n = 6 adult animals, S3 Fig). Extracellular recordings of the OB further showed that responses of OB neurons to ON stimulation were greatly increased under gabazine (n = 60 responses; n = 6 animals, S4 Fig), thus corroborating our previous findings. In addition to increasing the responses of RS cells, stimulation of the ON after gabazine injection in the medOB even induced motor discharges in the ventral roots. The neural activity consisted of rhythmic discharges alternating on both sides, a hallmark of fictive swimming (in 58.1% of trials; n = 36 locomotor bouts out of 62 trials for gabazine versus 0 out of 78 for control; n = 9: three adult animals and six larval animals, Fig 3 and S1 Data).
Because the density of GABAergic processes seemed relatively similar in the medOB and the MOB, we hypothesized that the neural activity in the MOB could be modulated by GABA, as observed for the medOB. To test this hypothesis, the effect of gabazine injections in the MOB on RS cell responses was examined. As shown for the medOB, gabazine injections into the MOB (0.1 mM, 1.8 ± 2.0 nL) enhanced the RS neuron responses to ON stimulations (n = 60 synaptic responses; n = 6 neurons; n = 6 larval animals; amplitude increase of 174.4 ± 167.0%, p < 0.05; no statistical differences between control and washout; Fig 4A). However, stimulation of the ON does not activate MOB neurons specifically, as it also activates medOB neurons. To rule out any involvement of the medOB in the increased RS responses after MOB gabazine injections, the effect of an electrical stimulation of the MOB with a gabazine injection (0.1 mM, 2.9 ± 1.1 nL) in the MOB was tested. Under control conditions, MOB stimulation did not induce responses in RS neurons. However, after a gabazine injection in the MOB, electrical stimulation of the MOB elicited responses in RS cells (n = 70 synaptic responses; n = 7 neurons; n = 7 larval animals; amplitude increase of 286.3 ± 296.8%; p < 0.05; no statistical differences between control and washout; Fig 4B). As a further control, electrical stimulation of the MOB under gabazine elicited significant responses in RS cells, even when the medOB had been surgically resected (n = 5 larval animals, S5 Fig). Furthermore, recordings of the ventral roots of the spinal cord showed that electrical stimulation of the MOB after a gabazine injection in the MOB can induce fictive swimming (in 65.1% of trials; 54 locomotor bouts out of 83 trials for gabazine versus 0 out of 105 for control; n = 9 larval animals, Fig 5 and S1 Data). Taken together, these findings suggest the presence of a previously unknown pathway linking the MOB to RS cells that seems to be under a strong tonic GABAergic inhibitory control.
We investigated the spatial organization of projections from the MOB that would eventually reach the RS neurons. We injected the axonal tracer biocytin in the MOB (n = 13 adult animals, Fig 6A1) and found ipsilateral axonal projections to the lateral pallium (LPal), medial pallium, dorsal pallium, striatum, dorsomedial telencephalic neuropil, and habenula. Contralateral projections were found to the OB, dorsomedial telencephalic neuropil, striatum, and LPal. The MOB injections did not label any fibers in the PT. Similar olfactory projections from the OB have been reported in other species of lampreys [46,47], but the selective contribution from the medOB or the MOB was not investigated in these earlier studies. The LPal appears to be a major target of neurons in the MOB, judging by the numerous labeled fibers seen to enter this region. The fibers densely filled the outermost layer covering the entire rostro-caudal extent of the LPal (Fig 6A2). Many fibers were also seen in the more central layers of the LPal, where the neuronal cell bodies of that structure are located. Tracer injections in the LPal (n = 9 adult animals, Fig 6B1) retrogradely labeled many neurons in the MOB without ever labeling cell bodies in the medOB (Fig 6B2). The retrolabeled neurons in the MOB were found close to the glomeruli, but were almost never seen inside them.
Physiological experiments were then carried out to characterize the effect of the pharmacological inactivation of the LPal on the responses of RS cells to the electrical stimulation of the MOB. Based on the results reported in Fig 4B, these experiments were carried out after removing the local GABAergic inhibition with a gabazine microinjection into the MOB (0.1 mM, 0.9 ± 1.0 nL, just prior each stimulation). An injection of glutamate receptor antagonists (2-amino-5-phosphonopentanoic acid [AP5]: 0.5 mM, 6-cyano-7-nitroquinoxaline-2,3-dione [CNQX]: 1 mM, 5.2 ± 0.8 nL) in the LPal strongly decreased the RS neuron responses (amplitude decrease of 64.8 ± 21.3%; p < 0.05), thus confirming the role of the LPal in relaying glutamatergic outputs from the MOB to locomotor control centers (n = 50 synaptic responses; n = 5 neurons; n = 5 larval animals, Fig 6C).
Biocytin was injected in the LPal to examine its descending projections. Emphasis was placed on regions known to be involved in the medial olfactomotor pathway, such as the PT and the MLR (n = 5 adult animals, Fig 7). Numerous fibers terminated in the PT, predominantly on the ipsilateral side (Fig 7B), with fibers crossing locally to the contralateral side (arrows in Fig 7B2). At levels immediately caudal to the PT, in the rostral mesencephalon, the number of descending fibers decreased sharply. Only a few labeled fibers continued to the level of the MLR (Fig 7C), where many appeared to terminate (Fig 7C2). More caudal levels were not investigated in the present study, but it is not excluded that some fibers continued down more caudally [48].
Tracing experiments were carried out to further characterize the population of LPal neurons projecting to the PT and the MLR. The organization and anatomical boundaries of the LPal in lamprey are still debated [47,49–56]. In the present study, we followed the nomenclature of Northcutt and Puzdrowski [47] and Pombal and Puelles [54]. The part of the brain that was considered to be the LPal in the present study is illustrated in S6 Fig. In this series of experiments, a solution containing Texas Red-conjugated dextran amine (TRDA) was injected in the MOB to label olfactory projections from the OB and a solution containing biocytin was injected in the PT (n = 7 adult animals and 1 larval animal, Fig 8A and S7 Fig) or the MLR (n = 4 adult animals and 1 larval animal, Fig 9A and S7 Fig) to retrogradely label neurons projecting to the PT or MLR. Typical results are shown in Figs 8 and 9 and S7 Fig. Labeled cell bodies were distributed uniformly in all regions of the LPal, dorsal, ventral, rostral and caudal, when injections were made in the PT (Fig 8A and 8B) or the MLR (Fig 9A and 9B). The dendrites of cells often extended radially towards the outermost layer of the LPal, where secondary olfactory fibers, labeled from the MOB, are located (Fig 8B and Fig 9B). These results show that fibers originating in the MOB came in proximity with LPal neurons projecting to both the PT and the MLR, suggesting that the LPal is a relay for MOB inputs to the PT and MLR.
Electrophysiological experiments were then conducted to examine the effect of deactivating the PT and the MLR on the RS neuron responses to the electrical stimulation of the LPal (Fig 8C and Fig 9C, respectively). Glutamate antagonists were locally injected in either the PT (AP5: 0.5 mM, CNQX: 1 mM, 1.1 ± 1.2 nL, Fig 8C) or the MLR (AP5: 0.5 mM, CNQX: 1 mM, 3.6 ± 2.5 nL, Fig 9C), and the RS neuron responses were markedly decreased (PT: amplitude decrease of 48.7 ± 19.7%; p < 0.05; no statistical differences between control and washout; n = 50 synaptic responses; n = 5 neurons; n = 5 larval animals; MLR: amplitude decrease of 45.3 ± 21.7%; p < 0.05; n = 60 synaptic responses; n = 6 neurons; n = 6 larval animals). Taken together with our previous findings, these results show that glutamatergic olfactory outputs from the MOB are relayed via the LPal to the PT and to the MLR before reaching RS cells.
The relative importance of the projection from the LPal to the PT or to the MLR was examined by counting retrogradely labeled cells in the LPal after an injection of a fluorescent tracer in the PT or the MLR. Bilateral biocytin injections in the PT (n = 6 adult animals) followed by the analysis of 10 LPals revealed that, on average, 751 ± 283 LPal neurons (per LPal) projected to the PT (Fig 10A). Bilateral biocytin injections in the MLR (n = 5 adult animals) followed by the analysis of eight LPals revealed an average of 93 ± 62 neurons (per LPal) in these animals (Fig 10A). The size of LPal neurons projecting to the PT and MLR was measured along their long axis. LPal PT- and MLR-projecting neurons measured on average 16.3 ± 3.0 μm (n = 90 cells from a subset of three animals, Fig 10B) and 15.4 ± 2.4 μm (n = 90 cells from a subset of three animals, Fig 10B), respectively. Interestingly, a few medOB neurons were systematically labeled after an MLR tracer injection (Fig 9A), thus demonstrating a direct projection from the medOB to the MLR.
The lateral olfactomotor pathway (orange pathway in Fig 11) may contribute significantly to the motor responses of lampreys to olfactory cues in their environment, in parallel to the previously described medial olfactomotor pathway (green pathway in Fig 11).
The transformation of sensory inputs into motor outputs is critical for the survival and reproduction of animals. As such, it represents an important part of their behavioral repertoire. While phylogenetically old nervous systems are predominantly dedicated to sensorimotor behaviors, the underlying neural pathways have been characterized in only a few cases. Recent studies in invertebrates have shed light on the neural substrate underlying some sensorimotor behaviors [57–59]. In contrast, very limited information has been collected in relation to the mechanisms of the transformation of sensory inputs into a locomotor output in vertebrates. For instance, it is well known that most animals (including humans) display odor-guided behaviors. However, the neural substrate and mechanisms associated with these behaviors have remained poorly understood.
In this study, we show that a GABAergic circuitry in the OB modulates a hardwired sensorimotor pathway. This is, to the best of our knowledge, the first report of the role of the OB GABAergic circuitry in modulating a motor behavior in any vertebrate species. Moreover, we identified a lateral olfactomotor pathway originating in the MOB and reaching locomotor control centers via a relay in the LPal. Together with our previous discovery of the medial olfactomotor pathway [12], these findings support the existence of two olfactomotor pathways in one of the most basal extant vertebrate species, the sea lamprey. As basal vertebrates, lampreys share a common brain “bauplan” with jawed vertebrates, including modern-day mammals. Therefore, knowledge gained from lamprey circuits and mechanisms provides insight into fundamental principles of vertebrate brain organization and function.
Lampreys, like many other animal species, display sex- and life stage–specific olfactory-induced motor behaviors [60–63]. The neural mechanisms accounting for the behavioral variability associated with a specific neural pathway within a species are largely unknown. However, the long-standing hypothesis that it was due to fundamental differences in brain wiring is now being challenged (reviewed in [64]). Indeed, only very subtle sex-specific differences have been found in the structure and circuitry of the brain in mammals [65–68]. Likewise, we have shown in a previous study that a hardwired olfactomotor pathway is present in both sexes at all life stages in the sea lamprey [12]. For this reason, we hypothesized in the present study that modulatory mechanisms acting on this pathway could play a role in the variability of the behavioral responses of lampreys to olfactory cues.
The OB is the first relay of the olfactomotor pathway. As such, it interfaces sensory afferents with motor control centers and it is ideally located to modulate olfactory-induced motor responses in lampreys. It has been proposed that the main function of the OB in vertebrates is the filtering and transmission of olfactory inputs [69]. Studies in mammals and turtles have shown that the sensory inputs to the OB are modulated both at presynaptic and postsynaptic levels by two classes of local GABAergic interneurons: periglomerular and granule cells [69]. Periglomerular cells inhibit glutamate release from primary olfactory axon terminals via a GABAB-mediated mechanism [70–73]. On the other hand, granule cells inhibit projection neurons via a GABAA-mediated mechanism [74–78]. Despite a rather good understanding of the cellular mechanisms responsible for the modulation of olfactory inputs, little is known about their overall effect on the OB output and, ultimately, on the resulting behavior. Using an in vitro isolated preparation of lamprey CNS, we provide the first evidence linking cellular GABAergic modulatory mechanisms in the OB to the activation of a sensorimotor pathway producing locomotor behavior.
We showed that the lamprey OB anatomical organization is very similar to that of other vertebrates, regarding its GABAergic circuitry. Our material confirms previous work showing that the lamprey OB contains numerous GABAergic neurons of different morphological types [43]. The morphology and location of the GABAergic neurons suggest that they are mainly granule cells [43,45], but not excluding possible periglomerular cells [43,51]. We also showed that both medOB and MOB glomeruli are densely innervated with GABAergic processes. These GABAergic processes are in close proximity to both primary olfactory axon terminals and dendrites or somata of OB projection neurons; this suggests possible pre- or postsynaptic contacts (S2 Fig). We have not formally identified types (i.e., axons versus dendrites) and origin (i.e., intrinsic versus extrinsic) of the GABAergic processes. The abundant GABAergic cell bodies labeled in the OB suggest that they may be of intrinsic (OB) origin (i.e., granule cells or periglomerular cells), as seen in other vertebrate species [39–43]. The lamprey granule cells are axonless [45], as in other vertebrate species. These processes are thus likely to be dendrites of granule cells or dendrites and axons of periglomerular cells, but some of these processes could be axons originating from neurons located in other parts of the brain. In mammals, most neuromodulatory inputs to the OB originate from the locus coeruleus (noradrenergic inputs), the nucleus of the diagonal band of Broca (cholinergic inputs), and the midbrain raphe (serotoninergic) (reviewed in [69,79,80]). However, some cells located in the nucleus of the diagonal band of Broca are GABAergic and project to the OB [81,82].
We now show that injection of the GABAA receptor antagonist, gabazine, in the OB potentiates RS cell responses to ON or OB stimulation, thus suggesting an enhancement of the olfactomotor transmission. In the case of OB (MOB) stimulation, however, we cannot completely exclude that electrical stimulation of the MOB might recruit not only projection neurons but also local GABAergic interneurons, and that in such a case, an injection of gabazine might block the effect of their activation. Under gabazine, the electrical stimulation of the ON or OB can induce fictive swimming—the in vitro corollary of swimming behavior. Overall, these findings suggest that the GABAA antagonist gabazine increases the output of the OB. Indeed, downstream relays of the olfactomotor pathway (PT and MLR) control locomotion in a graded fashion [12,15,83]. Consequently, the increased RS cell responses observed under gabazine are likely to result from an increased drive from the OB to the PT and/or MLR. Studies in mammals have shown that GABA acts at several locations in the OB. OSN terminals express GABAB receptors [84–86], which inhibit transmission from OSN axons to mitral cell primary dendrites upon release of GABA by periglomerular cells [70,71,87,88]. Mitral cell dendrites express both GABAA and GABAB receptors [89–93]. Pharmacological blockade or genetic alteration of GABAA receptors in mitral cells alters the OB γ oscillations and leads to increased ON-induced mitral cell discharges [78,94]. The effect of GABAB receptor activation in these cells is less clear [93]. Both periglomerular and granule cells release GABA on mitral cells; periglomerular cells contact mitral cell primary dendrites, whereas granule cells contact mitral cell secondary dendrites [69,95]. Granule and periglomerular cells also express GABAA receptors [89,91,96,97]. Genetic alteration of the GABAA receptor subtype expressed in granule cells (i.e., expressing the β3 subunit) either globally or in a cell-specific manner increases the granule cell inhibition of mitral cells and results in increased OB γ oscillations [98,99]. To the best of our knowledge, the effect of periglomerular cell GABAA receptor activation on mitral cell activity has not been investigated, but an inhibition of periglomerular cells leading to the disinhibition of mitral cells could be expected. Finally, electrophysiological evidence suggests that granule cells also possess GABAB receptors whose activation modulates granule cell inhibition of mitral cells [100]. GABA can thus depress or potentiate mitral cell activity depending on its site of action (i.e., OSNs axons, OB interneurons, or mitral cell). However, as OB interneurons act on mitral cells via GABAA receptors, the net effect of the pharmacological blockade of GABAA receptors in all OB layers is likely to be a disinhibition of mitral cells. This is consistent with our results in lampreys and those found in other vertebrate species [77,94,101–104].
The presence of both tonic and phasic inhibition in the OB has been reported in fish, amphibians, and mammals [71,77,78,94,98,99,102,105,106]. It has been suggested that tonic inhibition may modulate the strength of sensory inputs to the OB [88] or the sensitivity of second-order olfactory neurons to sensory inputs [102]. Phasic inhibition has been shown to generate neuronal synchrony (i.e., oscillations) in projection neurons [98,99]. The role of these oscillations and thus of the phasic inhibition is still debated, but several studies in insects and mammals point toward a crucial role in coding olfactory information [107–109].
Our study shows that a strong GABAergic inhibition of the OB output is present in the lamprey, one of the most basal extant vertebrate, and thus may be a common ancestral feature of the vertebrate OB.
The GABAergic modulation of the olfactomotor pathways seen in lampreys could explain some of the life stage–specific behavioral responses to olfactory cues. For instance, migratory pheromones attract only pre-spawning adult lampreys [24–26]. This is surprising because these pheromones evoke strong responses in OSNs at other life stages [110]. Somehow, the activation of OSNs only leads to locomotor responses during the pre-spawning adult life stage. Meléndez-Ferro and colleagues [43,111] have stated that the density of OB GABAergic cells declines significantly between the newly transformed and pre-spawning life stages. Whether this apparent decrease in GABAergic cell density could account for some of the life stage differences is not known at present, but it could be one plausible mechanism worth investigating.
A series of recent studies have shown that a CO2-mediated water acidification significantly impairs several olfactory-driven behaviors in fish, including prey tracking, predator avoidance, alarm response, and homing [112–116]. The mechanism at play has not been fully characterized yet, but it involves an alteration of the normal functioning of GABAA receptors, as blocking these receptors with gabazine led to a behavioral recovery [115,117]. The authors of these studies proposed that a potentiation or a reversal of the GABAA receptor function (from inhibitory to excitatory) because of changes in anionic gradients over neuronal membranes could underlie these behavioral alterations. Taken together, these studies show that GABAergic mechanisms also play a crucial role in modulating olfactomotor behaviors in fish. Further studies are needed to establish whether the neural pathways and modulatory mechanisms characterized in lampreys are also present in fish and other vertebrates.
In the present study, we showed that stimulation of the MOB under gabazine led to excitatory responses in RS cells and to locomotion. This suggests the existence of a distinct pathway from the MOB to the RS cells and the presence of a strong tonic GABAergic inhibition in the MOB. We characterized the anatomy and physiology of this pathway. Anatomical data showed that the LPal receives a massive projection from the MOB and projects down to both the PT and MLR. The PT, in turn, projects to the MLR [12,118,119]. We also showed that the MLR receives a direct projection from the medOB, in addition to the already characterized projection via the PT [12]. The MLR then reaches the command cells for locomotion, the RS cells, via glutamatergic and cholinergic projections [120–123]. Physiological data confirmed that the LPal relays MOB olfactory inputs to the RS cells via the PT and MLR. This is consistent with the recent findings of Suryaranayana and colleagues [124] indicating that some LPal neurons receive monosynaptic inputs from the OB. Interestingly, Ocaña and colleagues [48] showed that a few fibers originating in the LPal could reach RS neurons directly and that some of these could be followed as far as the first spinal segments. This prompted the authors to conclude that the LPal possesses an efferent projection pattern similar to that of the amniote motor cortex [48]. It would be interesting to examine if these projections from the LPal to the RS neurons and spinal cord are also involved in olfactomotor responses. Although we cannot exclude that there may be other pathways linking olfactory centers to motor centers, our study demonstrates the existence of two distinct glutamatergic pathways linking the olfactory and motor systems in lampreys (Fig 11). Both these pathways share a common output via the PT/MLR–RS neurons system. However, they differ regarding their pathways from the OB to motor control centers (i.e., PT/MLR), as well as to their inputs from the periphery.
In lampreys, the main olfactory epithelium contains numerous tall, ciliated OSNs expressing the G-protein Golf, as in the main olfactory epithelium in other vertebrates [125–133]. The OSNs of the main olfactory epithelium project their axons to the MOB, which, in turn, projects mainly to the LPal, i.e., the putative homologue of the mammalian olfactory cortex in lampreys [134]. This pathway is strikingly similar to the main olfactory pathway of terrestrial vertebrates and thus further supports its evolutionary conservation. In addition to the main olfactory epithelium, lampreys possess an accessory olfactory organ [29–32,135–138]. The accessory olfactory organ contains short, broad, ciliated OSNs [32] that do not express the G-protein Golf and project only to the medOB [32,129]. Projection neurons of the medOB then project directly to the PT and MLR, bypassing the LPal. Taken together, these findings show that the lamprey accessory olfactory organ constitutes a discrete olfactory subsystem. It has even been suggested that the accessory olfactory organ represents a primordial vomeronasal system [29,31,138].
In other vertebrates, the presence of parallel olfactory pathways conveying the information from the periphery to high-order brain olfactory centers suggests that these systems subserve different behavioral functions [139–143]. For instance, in fish, segregated olfactory pathways, from the olfactory epithelium to the telencephalon, mediate feeding, reproductive, and alarm behaviors [139,142,144–151]. Similarly, the main and accessory (i.e., vomeronasal) systems of terrestrial vertebrates are segregated until at least the third-order neurons and their respective activation elicits different behaviors [152–155]. Physiological evidence in lampreys also supports this hypothesis, as OB local field recordings showed that the medOB and MOB have overlapping but different response profiles to feeding cues and pheromones [33,34]. Moreover, we show that the pathways from the OB to the motor control centers differ for the two olfactory subsystems. The medOB projects directly to motor control centers, whereas the MOB projects first to the LPal before reaching motor control centers. Not surprisingly, activation of both systems leads to locomotion. This could be attributed to the paucity of the behavioral repertoire of lampreys compared to mammals. However, it should be noted that reproductive, migratory, and feeding behaviors all require locomotion in lampreys. The distinction between these two subsystems thus lies in their inputs from the periphery (accessory olfactory organ versus main olfactory epithelium) as well as in the involvement of the LPal in the lateral pathway. It is tempting to propose that the medial pathway could mediate innate responses to chemical stimuli (for example, avoidance), whereas the lateral pathway could be involved in olfactomotor behaviors requiring further processing and perhaps learning (for example, olfactory navigation). A similar distinction between dual “olfactory” systems exists in invertebrates [156–158]. In mammals, it was shown that mitral cells of the MOB can develop differential responses to rewarded/unrewarded odors [159]. It has been suggested that the dichotomy between innate responses versus learned responses may be what distinguish the main and accessory systems of terrestrial vertebrates [154]. This hypothesis has, however, received little attention, and further studies are needed.
In conclusion, our study shows that olfactory inputs can activate the locomotor command system via two distinct glutamatergic pathways in lampreys. To the best of our knowledge, this is the first characterization of a dual olfactory pathway, from the periphery to the motor command system, in vertebrates. Both pathways are strongly modulated by the GABAergic circuitry of the OB that may account for some of the variability in behavioral responses to olfactory inputs in lampreys. The existence of two segregated olfactory subsystems in one of the most basal extant vertebrates sheds light on the evolution of the olfactory system and suggests that its organization in functional clusters could constitute a common ancestral trait of vertebrates.
For all procedures, the animals were deeply anesthetized with tricaine methanesulphonate (MS-222, 200 mg/L, Sigma-Aldrich, Oakville, ON) and then decapitated. All surgical and experimental procedures conformed to the guidelines of the Canadian Council on Animal Care and were approved by the animal care and use committee of the Université de Montréal (Protocol no. 18–018), the Université du Québec à Montréal, and the University of Windsor.
Experiments were performed on 57 larval and 61 adult sea lampreys (Petromyzon marinus) of both sexes. Some animals were used in more than one experiment. Larvae were collected from the Pike River stream (QC, Canada). Adults were collected from the Great Chazy River (NY, United States) and were kindly provided by agents of the U.S. Fish and Wildlife Service of Vermont. The permission to collect animals in the field was granted by the Quebec's Ministry of Natural Resources and Wildlife (permit no. 2017-03-30-2189-16-SP). All animals were kept in aerated fresh water maintained at 4–5 °C.
For all types of experiments, the animals were deeply anesthetized with tricaine methanesulphonate (MS-222, 200 mg/L, Sigma-Aldrich), decapitated caudal to the seventh branchiopore, and transferred into cold oxygenated Ringer's (8–10 °C) of the following composition (in mM): 130 NaCl, 2.1 KCl, 2.6 CaCl2, 1.8 MgCl2, 4.0 HEPES, 4.0 dextrose, and 1.0 NaHCO3, at pH 7.4. The branchial apparatus, myotomal musculature, and all soft tissues attached to the ventral side of the cranium were removed. The dorsal part of the vertebrae and cranium were removed to expose the brain and the rostral spinal cord. The peripheral olfactory organ was left intact with the ON still attached to the brain. All other nerves were cut and the choroid plexus covering the fourth and the mesencephalic ventricles was removed.
The preparation was pinned down to the bottom of a recording chamber lined with Sylgard (Dow Corning, Midland, MI) and continuously perfused with cold oxygenated Ringer's (about 4 mL/min). Intracellular recordings of RS neurons were performed under visual guidance through a M3C stereomicroscope (Wild-Heerbrugg, Heerbrugg, Switzerland) using sharp glass microelectrodes (60–130 MΩ) filled with 4 M potassium acetate. The signals were amplified with an Axoclamp 2A (20 kHz sampling rate, Axon Instruments, Foster City, CA) and acquired through a Digidata 1322A interface running on pClamp 9.2 software (Axon Instruments, Foster City, CA). Only RS neurons displaying a stable membrane potential lower than −70 mV for at least 15 min were considered in this study.
Electrical stimulation (1–3 pulses, 5–30 μA, 2-ms duration, and 20-ms pulse interval) was delivered using homemade glass-coated tungsten electrodes (0.8–2 MΩ, 10–50 μm tip exposure) connected to a Grass S88 stimulator via a Grass PSIU6 photoelectric isolation unit (Astro-Med, Longueuil, QC). A delay of 50 s was allowed between each stimulation. In the figures, the synaptic responses are illustrated as the mean of 10 responses obtained with the same stimulation parameters.
In some experiments, the left and right ventral roots from one spinal segment (usually around the 10th segment) were also recorded using extracellular glass electrodes (tip diameter about 5 μm) filled with the Ringer's solution. The signals were amplified (×10,000) and filtered (100 Hz–1 kHz band-pass) using an AM systems 1800 dual channel amplifier (AM systems, Sequim, WA) and monitored for the presence of neural activity. “Fictive locomotion” (originally defined by Perret and colleagues, 1972) [160] observed in the absence of muscles and movement was defined as a neural activation of the spinal ventral (motor) roots, with a similar pattern as myotomal contractions seen during swimming.
Drugs were purchased from Sigma-Aldrich (AP5), Tocris Bioscience (gabazine and CNQX), and Thermo Fisher Scientific (Fast Green). They were kept as frozen concentrated stock solutions and dissolved to their final concentrations in Ringer's solution prior to their use. Gabazine (SR-95531; 0.1–1 mM) and the CNQX/AP5 mixture (0.5 mM/1 mM) were pressure ejected (3–180 pulses; mean ± SD = 57 ± 46 pulses; about 4 psi, 20–40-ms pulse duration, injection volume: 0.1–6.6 nL; mean ± SD = 2.6 ± 2.1 nL) through glass micropipettes (10–20 μm tip diameter) in the brain tissue, using a Picospritzer (General Valve Corp, Fairfield, NJ). The inert dye Fast Green was added to the drug solution to monitor the extent of the injections. The spread did not exceed 300 μm in diameter for any microinjection.
RS cells were retrogradely labeled by placing crystals of the calcium-sensitive indicator dye Calcium-Green dextran (3000 MW, Invitrogen, Eugene, OR) on the rostral stump of the spinal cord, transected at the level of the first spinal segments. The preparation was then kept in cold, oxygenated Ringer's solution for 24–36 h of axonal transport. Labeled cells were observed under a Nikon epifluorescence microscope equipped with a 20× (0.75 NA) objective. A filter set appropriate for fluorescein isothiocyanate (FITC) was used to visualize the neurons. The emitted light was captured with an intensified CCD video camera (Photometrics CoolSNAP HQ, Roper Scientific, Tucson, AZ) and recorded at a rate of two images per second, using Metafluor imaging software (Molecular Devices, Sunnyvale, CA). Calcium responses are expressed as relative changes in fluorescence (ΔF/F).
Results are presented as mean ± SD. Statistical analyses were performed using Sigma Plot 11.0 (Systat, San Jose, CA), and statistical significance was set at p < 0.05. To test for differences in mean between groups, we performed a one-way analysis of variance for repeated measures, followed by a Holm-Sidak’s multiple-comparison post hoc test or Friedman analysis of variance for repeated measures on ranks followed by Tukey multiple-comparison post hoc test.
|
10.1371/journal.pbio.1000220 | Differentiation Driven Changes in the Dynamic Organization of Basal Transcription Initiation | Studies based on cell-free systems and on in vitro–cultured living cells support the concept that many cellular processes, such as transcription initiation, are highly dynamic: individual proteins stochastically bind to their substrates and disassemble after reaction completion. This dynamic nature allows quick adaptation of transcription to changing conditions. However, it is unknown to what extent this dynamic transcription organization holds for postmitotic cells embedded in mammalian tissue. To allow analysis of transcription initiation dynamics directly into living mammalian tissues, we created a knock-in mouse model expressing fluorescently tagged TFIIH. Surprisingly and in contrast to what has been observed in cultured and proliferating cells, postmitotic murine cells embedded in their tissue exhibit a strong and long-lasting transcription-dependent immobilization of TFIIH. This immobilization is both differentiation driven and development dependent. Furthermore, although very statically bound, TFIIH can be remobilized to respond to new transcriptional needs. This divergent spatiotemporal transcriptional organization in different cells of the soma revisits the generally accepted highly dynamic concept of the kinetic framework of transcription and shows how basic processes, such as transcription, can be organized in a fundamentally different fashion in intact organisms as previously deduced from in vitro studies.
| The accepted model of eukaryotic mRNA production is that transcription factors spend most of their time diffusing throughout the cell nucleus, encountering gene promoters (their substrate) in a random fashion and binding to them for a very short time. A similar modus operandi has been accepted as a paradigm for interactions within most of the chromatin-associated enzymatic processes (transcription, replication, DNA damage response). However, it is not known whether such behavior is indeed a common characteristic for all cells in the organism. To answer this question, we generated a knock-in mouse that expresses in all cells a fluorescently tagged transcription factor (TFIIH) that functions in both transcription initiation and DNA repair. This new tool, when combined with quantitative imaging techniques, allowed us to monitor the mobility of this transcription factor in virtually all living tissues. In this study, we show that, in contrast to the aforementioned paradigm, in highly differentiated postmitotic cells such as neurons, hepatocytes, and cardiac myocytes, TFIIH is effectively immobilized on the chromatin during transcription, whereas in proliferative cells, TFIIH has the same dynamic behavior as in cultured cells. Our study also points out that results obtained from in vitro or cultured cell systems cannot always be directly extrapolated to the whole organism. More importantly, this raises a question for researchers in the transcription field: why do some cells opt for a dynamic framework for transcription, whereas others exhibit a static one?
| Basal transcription/repair factor IIH (TFIIH) is a ten-subunit complex [1], essential for both RNA polymerase I and II (RNAP1 and 2) transcription initiation and nucleotide excision repair (NER) [2]. NER is an important DNA repair process, which is able to remove a broad spectrum of different DNA lesions. Inherited defects in NER cause severe cancer predisposition and/or premature aging, illustrating its biological significance [3]. In RNAP2 transcription and DNA repair, TFIIH acts as a DNA helix opener, required for transition of initiation to early elongation of RNAP2 and establishment of the preincision NER complex [4],[5]. Mutations in this complex are associated with a surprising phenotypic heterogeneity, ranging from the (skin) cancer-prone disorder xeroderma pigmentosum (XP) to the severe progeroid conditions Cockayne syndrome (CS) and trichothiodistrophy (TTD), the latter additionally characterized by brittle hair and nails [1],[6]–[8]. Since TFIIH is considered to be a general or basal transcription factor and essential NER component, it is surprising to note that TFIIH-associated syndromes present different pathologically affected tissues. For instance, within TTD, primarily differentiated cells appear to be affected. TTD-specific scaly skin and brittle hair features derive from defects in the latest stage of differentiating keratinocytes [9],[10]. In addition, reduced β-globin expression and subsequent anemia in TTD originates from a defective terminal differentiation of precursor erythrocytes [11]. Furthermore, both TTD and XP/CS patients and mice express neurological features caused by defects in final-stage differentiated postmitotic neuronal cells [9],[12]. Nevertheless, many other tissues/organs and cell types appear to be relatively unaffected. This observation can be partly explained by the hypothesis that some tissues are more susceptible than others to endogenous DNA damage [13],[14] and/or that TFIIH transcriptional function is differentially regulated in distinct cell types [15].
Most transcription factors, including basal factors and transcription activators, are only very transiently bound, on the order of a few seconds, to their substrate [16],[17]. This uniformly emerging concept of dynamic transient machineries, with the exception of components of the actual RNAP2 [18], is thought to have a number of advantages over previous proposed models based on stable preassembled large MW “holo” complexes. Live-cell protein mobility studies have culminated in unprecedented, novel insights into the spatial and dynamic organization of macromolecule machines within the context of the complex mammalian cell nucleus with a general, but not universal, modus operandi of dynamic exchange of reaction constituents (proteins). Exceptions to this general mechanism of action have been described for transcriptional activators, such as Gal4 [19] and hypoxia-inducible factor 1 (HIF) [20], and for molecular chaperones such as heat-shock protein 70 [21]. It was found that upon transcription activation these factors reside longer at the transcription sites. Interestingly, the basal transcription factor RNAP2 (GFP-RPB3) was also shown to be longer bound at transcribed regions when transcription was induced [22],[23]. However, it is currently not known whether the generally observed dynamic kinetic framework in in situ–cultured cell types for transcription initiation factors, such as TFIIH [24], can be extrapolated to other cell types in the organisms, under ground-state transcriptional conditions.
To study the consequences of different transcriptional programs on the kinetic behavior of TFIIH and further understand the complex phenotypic expression of mutated TFIIH, we created a knock-in mouse model (Xpby/y) that expresses homozygously a YFP (yellow fluorescent protein, a variant of the green fluorescent protein)-tagged TFIIH subunit under control of the endogenous transcriptional regulatory elements. Using this tool, we explored TFIIH binding kinetics directly in different cells and tissues of the organism.
In order to visualize and quantitatively determine dynamic interactions of transcription initiation factor TFIIH within different postmitotic and differentiated cells and living tissues, we created a mouse knock-in model that expresses the XPB protein (largest subunit and helicase of the ten-subunit TFIIH complex) tagged at its C-terminus with the yellow GFP variant (YFP). Gene-targeting constructs and strategy are schematically depicted in Figure 1A (see Materials and Methods and Figure S1A for details). Briefly, the targeting strategy was designed in such a way that interference of the genomic organization of the Xpb locus was kept to a minimum, keeping the integrity of the promoter, all the intron–exon boundaries, and the 3′ UTR (including the endogenous poly A signals) of the Xpb gene intact. The XPB-YFP fusion protein further contains additional convenient C-terminal His6- and HA-tags. Embryonic stem (ES) cells transfected with the targeting fusion construct were selected for proper targeting (by homologous recombination) by Southern blotting (Figure S1A). Immunoblotting of whole-cell extracts revealed that an intact XPB-YFP fusion protein was produced in recombinant ES cells (Figure S1B). Selected ES cells were introduced in C57Bl/6 blastocysts and transplanted into foster mothers. Chimeric offspring were further crossed for germ-line transmission of the targeted allele (Figure S1D and S1E).
To avoid any possible interference with expression of the fusion gene by the presence of the dominant selectable NeoMarker in the 3′ UTR, heterozygous Xpby−Neo/+ mice were crossed with a ubiquitous Cre-Recombinase–expressing mouse model [25] (Figure S1D). The subsequent “floxed” heterozygous offspring (Xpby/+) were intercrossed to generate homozygous knock-in mice (Xpby/y) (Figure S1E). Homozygous Xpby/y, heterozygous Xpby/+, and wild-type (Xpb+/+) progeny were obtained in a Mendelian ratio (32 homozygous knock-in mice out of 125 offspring), indicating that homozygosity for the knock-in fusion gene does not impair embryonic development. A small cohort of homozygous (males and females) and heterozygous (six of each) littermates was allowed to age, until natural death, which occurred around 2 years for both Xpby/y and Xpby/+ mice. No obvious features of premature aging or spontaneous carcinogenesis of the Xpby/y other than those occurring in Xpb+/+ mice were observed. Knock-in Xpby/y mice appeared healthy and fertile, indicating that the presence of the fluorescent tag does not significantly interfere with the vital functions (transcription initiation) of the Xpb gene.
Most viable, naturally occurring TFIIH mutations cause an overall reduction of the steady-state levels of TFIIH [26]–[28]. However, comparative immunofluorescence revealed that the intracellular concentration of p62 (Figure S2A) (another nontagged TFIIH subunit) and XPB were not altered by the presence of the tagged XPB subunit (Figure 1B) [26]–[28], suggesting that neither the expression nor the stability was affected by the presence of the YFP-His6_HA tag on the XPB protein. Unscheduled DNA repair synthesis capacity (UDS, a measure of NER activity) after UV damage (Figure S2B) and UV-survival of Xpby/y dermal fibroblasts were similar to wild-type (Xpb+/+) cells assayed in parallel (Figure 1C), indicating that the tagged XPB protein remains normally active in NER. Immunoprecipitation experiments also showed that the tagged XPB protein was incorporated into TFIIH (unpublished data), consistent with our previous observations that exogenously expressed GFP-tagged XPB is properly incorporated into TFIIH complexes [24].
In conclusion, the addition of the 27-kDa fluorescent tag to the strongly conserved XPB protein does not detectably affect the multiple functions of TFIIH in transcription and NER even at the critical level of an intact organism, whereas single amino acid substitutions in XPB patients give rise to severe skin cancer predisposition and dramatic premature aging [29],[30]. This demonstrates that the Xpby/y knock-in mouse model is a bona fide source to obtain relevant information on the spatial and dynamic organization of transcription and DNA repair in vivo in an intact organism. Fluorescence of TFIIH was detectable in all primary cultures of different cell types isolated from these mice, e.g., ES cells, dermal fibroblasts, and keratinocytes (Figure S2C). Note that the level and subnuclear distribution of fluorescence throughout the cell population is homogeneous (Figure S2C), in striking contrast to the heterogeneous expression characteristic of stably transfected cell cultures.
To study the spatiotemporal distribution of TFIIH and to determine its kinetic engagements in different cell types within intact tissues, we established organotypic cultures of several organs and tissues. Within organotypic slices of cerebral cortex, isolated and maintained according to established procedures [31], the different cortex layers and the neurons are easily recognizable (Figure S3A). We determined live-cell protein mobility of TFIIH by fluorescence recovery after photobleaching (FRAP) (see Figure 2A, top panel, and Materials and Methods). Using exogenously expressed XPB-GFP, we previously demonstrated that TFIIH in SV40-immortalized human fibroblasts is highly dynamic: the majority moves freely through the nucleus, and a fraction transiently interacts with promoters for only a few seconds (2–6 s) [24]. We found a similar high mobility in primary keratinocytes when monitored within the epidermis of skin explants from Xpby/y mice (Figure 2A, middle panel), where fluorescence fully recovered in the bleached strip within a few seconds. In sharp contrast, FRAP on neurons within cerebral cortex slices revealed a striking incomplete fluorescence recovery, even after 60 min postbleaching (Figure 2A, lower panel, and Figure S3B), indicating an unprecedented large (>80%) immobile pool of TFIIH (Figure 2B) stably bound to static nuclear structures, most likely chromatin. This unexpected static behavior of a transcription initiation factor, which can be compared to the static behavior of H2B in cultured cells [32], was also observed in Purkinje cells and cerebellar granular neurons in organotypic slices (Figure 2C and Figure S4) suggesting that this behavior is a common feature in various neurons, despite their different functions, chromatin compaction, and different TFIIH expression levels (Figure S5A). Moreover, computation of the (squared) Pearson product moment correlation coefficient between measured immobile fractions and single-cell TFIIH expression levels showed no significant linear relation between these parameters (r2 = 0.062). We also measured the mobility of nonfused GFP in neurons within organotypic cerebellar slices derived from a mouse that expressed GFP under the control of actin promoter [33]. In contrast to TFIIH mobility, GFP itself diffuses very rapidly in neurons (Figure S5B), as was previously found in cultured cells [34],[35]. These results indicate that this static behavior is specific for TFIIH and not a common phenomenon of nuclear protein mobility in neurons.
How does TFIIH immobilization relate to its multiple biological activities? Obviously, the engagement of TFIIH in transcription is most relevant in tissue sections not treated with DNA-damaging agents. However, to exclude that an eventual NER-dependent binding activity could account for the immobilized TFIIH in neurons, initiated by a possible high load of endogenously produced lesions, we measured TFIIH mobility in NER-deficient mice. For this purpose, we crossed Xpby/y mice with Xpc−/−, to generate Xpby/y•Xpc−/− mice. In the absence of XPC, TFIIH does not bind damaged DNA [36]. TFIIH mobility in neurons from Xpby/y•Xpc−/− mice appeared identical as in neurons derived from NER-proficient mice (Figure 3A and Figure S6), showing that the DNA repair function of TFIIH is not responsible for the protracted binding of TFIIH. To demonstrate that the transcription function is responsible for TFIIH immobilization, we inhibited transcription by treating organotypic brain slices with the RNAP2-specific transcription inhibitor α-amanitin [37]. α-Amanitin blocks the catalytic domain of RNAP2 [38], inhibiting both transcription initiation and elongation. Treatment of organotypic tissues with α-amanitin resulted in a release of immobilized TFIIH (Figure 3A and Figure S7), suggesting that the immobilization is due to the transcriptional function of TFIIH [17],[24]. In parallel, we verified that the condition used (incubation time and drug concentration) for transcription inhibition in tissues blocked transcription in cultured cells by measuring the BrU incorporation (Figure S8A and S8B). To further prove that the transcriptional engagement of TFIIH causes its high immobilization in neurons, we modulated transcription by inducing a cold shock (4°C or 27°C). As with the heat-shock response, cold shock generally induces a reduction in basal transcription and translation and a growth arrest [39]. This response is temporary, since after an adaptation period, cellular metabolism is resumed, although at a lower rate compared to growth at 37°C [40]. Moreover, after cold shock, a change in global expression of genes is observed [41] to allow adaptation to this environmental stress. As shown in Figure 3A and Figure S7, reducing the temperature of organotypic slices to extreme (4°C) and moderate hypothermia values (27°C) for 30 to 60 min fully remobilized TFIIH in neurons. To check whether this is a temporary stress response, we also measured the mobility of TFIIH in neurons embedded in organotypic slices kept at 27°C for 48 h. Under these conditions, TFIIH was found to bind as in untreated organotypic slices (37°C), demonstrating that the remobilization of TFIIH at low temperatures is a rapid (60 minutes at 4°C or 27°C) temporary response, likely reflecting a change in the transcriptional engagement (Figure 3A and Figure S7).
Additionally to RNAP2 transcription, TFIIH has been found to accumulate in nucleoli and participate in RNAP1 transcription [24],[42]. Although the clear TFIIH function in RNAP1 transcription has not been elucidated yet, dynamic studies showed that TFIIH residence time in RNAP2 transcription (2–10 s) is different from RNAP1 transcription (∼25 s), suggesting that the role played by TFIIH in these two cellular functions could be slightly different [24]. Cortex neurons and Purkinje cells show a strong localization of TFIIH in nucleoli (Figure S9A), allowing local FRAP analysis in this subnuclear compartment. Similarly to nucleoplasmic TFIIH immobilization (RNAP2 transcription), nucleolar TFIIH immobilization (RNAP1 transcription) was also very high (Figure 3A and Figure S10). In contrast, however, this immobilization is partly resistant to cold shock and resistant to α-amanitin treatment, in line with the expectation, as α-amanitin is known to exclusively inhibit RNAP2 transcription (Figure 3A and Figure S8). To determine that moderate cold shock would indeed inhibit RNAP1 transcription, we measured the amount of pre-rRNA 45S in cold shock–treated cells and tissues (organotypic brain cultures). Surprisingly, cold shock did not alter the amount of 45S, either in cultured cells or in organotypic cortex slices (Figure S9B). The absence of a reduction of the steady-state-levels of 45S rRNA by either a severe (4°C) or a mild (27°C) cold shock is likely explained by the fact that cold shock also interferes with pre-rRNA maturation or degradation. In fact, some proteins, induced by cold shock, have been showed to play a role in preventing the degradation of RNA molecules (reviewed in [43]). In contrast, RNAP1 transcription inhibition induced by actinomycin D (0.1 µg/ml) led to a clear reduction of the amount of pre-rRNA in cultured cells. In conclusion, our results suggest that TFIIH is highly immobilized in neurons in both RNAP2 and RNAP1 transcription.
As TFIIH is known to be involved in RNAP2 transcription initiation, its transcription-dependent immobilization in neurons predicts a favored binding of TFIIH at promoter sequences. To verify that indeed TFIIH in cortex neurons was bound to promoters of active genes, we performed chromatin immunoprecipitation (ChIP) on adult cortex tissues slices under normal conditions and after cold shock, and measured the proportion of active housekeeping genes (xpb, RnaPolI) promoter sequences versus adjacent untranscribed areas by semiquantitative PCR (Figure 3C and Figure S9D). As shown by the FRAP experiments (Figure 3A and Figure S7), during cold shock, TFIIH is released from chromatin (Figure 3B). Importantly, we found that TFIIH in cortex slices is more strongly bound to promoter sequences (40% of the input genomic DNA) than to untranscribed areas (19%) (see Materials and Methods for details, Figure 3C and Figure S9D). However, after cold shock, TFIIH is less bound to promoters, clearly showing that cold shock–induced transcription inhibition is associated with remobilization of TFIIH from housekeeping gene promoters.
In view of the notion that the majority of TFIIH is immobilized to chromatin, we wondered whether TFIIH would be available to act in NER after a sudden high dose of genotoxic stress. Since neurons located in the slice are inaccessible to UV-C light, we used multiphoton laser irradiation to locally induce DNA damage [44],[45] in cerebral neurons and compared the accumulation of TFIIH to that observed in skin keratinocytes damaged by the same procedure. Surprisingly, despite the large fraction of immobilized TFIIH, significant amounts of TFIIH were still recruited to damaged DNA (Figure 3D).
Since the observed high TFIIH immobilization is not found in all cell types (Figure 2C), we exploited the availability of an entire organism to analyze the dynamic distribution of TFIIH in different cell types within the context of living tissues (Figure 4A and Figure S11). Specifically in postmitotic and nonproliferative cells (neurons, myocytes, and hepatocytes), we identified a large pool of immobilized TFIIH, whereas in proliferating cells and/or cells that have the capacity to proliferate (intestine epithelium, epidermal keratinocytes, and dermal fibroblasts), TFIIH was found to be highly mobile (Figure 4A and Figure S11). This would suggest that the kinetic organization of this essential transcription factor would be determined by the proliferative capacity of cells. However cultured chondrocytes maintained in a confluent state under low serum were shown to become quiescent (as shown by the absence of the Ki67 marker, Figure S12A) and did not show a difference in TFIIH mobility when compared with proliferative chondrocytes (Figure S12B), suggesting that absence of proliferation is not a condition sufficient to cause a reduction in TFIIH mobility. In view of this result, we investigated whether TFIIH mobility is affected during the establishment of a differentiation-dependent specific transcriptional program. We measured TFIIH mobility during postnatal development of cerebral cortex and liver. Brain and liver were isolated from pups at different postnatal days (PNd), and TFIIH mobility was measured in cortex neurons and hepatocytes. Remarkably, we observed a progressive TFIIH immobilization during development (Figure 4B and Figure S13) (Figure 4C and Figure S14), which appeared time- and organ-specific. In cortex neurons, TFIIH bound fractions are gradually increasing from PNd 10 (Figure 4B and Figure S13). TFIIH mobility in neurons is very homogeneous throughout the tested population of cells at each different developmental stage, whereas in liver during development at, e.g., PNd 6, the kinetic pools differ over the population (see inset, Figure 4C) and becomes homogeneous at later stages. Our results demonstrate that the strong TFIIH binding is a physiological event that takes place during normal development of organs and suggests the establishment of a more fixed transcriptional program than in rapidly growing cells.
To further investigate differentiation-dependent mobility of TFIIH, we used an in vivo keratinocyte differentiation model. Within the hair shaft, highly differentiated nonproliferative keratinocytes, known as trichocytes [46], can be found. These cells produce the keratins and keratin-associated proteins that form the structure of hairs. Trichocytes are easily recognizable because of their position in the hair and of the melanin inclusions in their cytoplasm (Figure 5A, left panel). Indeed, in trichocytes (Figure 5B and Figure S15), TFIIH mobility is greatly reduced, almost to the same extent as in neurons and myocytes, whereas in other keratinocytes of the hair follicle (bulb) (Figure 5A, right panel), TFIIH is highly mobile (Figure 5B and Figure S15). To substantiate that indeed differentiation is an important determinant for the observed shift in TFIIH mobility, we measured TFIIH mobility during in vitro differentiation. ES cells were nonspecifically differentiated using the hanging drop technique [47]. Through the use of this method, several differentiated cell types can be obtained, organized in morphologically different areas of a developing clone. We measured different cells in several morphologically distinct regions of the differentiated ES clones. In these clones, three distinct TFIIH mobility groups were observed (Figure 5C and Figure S16). The vast majority of cells (∼70%) presented a high TFIIH mobility, a second group (∼20%) presented an intermediate (∼40%) bound fraction of TFIIH, and a small fraction of cells (∼10%) presented a high TFIIH binding capacity (> than 70%). Postlabeling showed that cells with high TFIIH binding were mainly osteocytes (Figure S17A and S17B). During ES hanging drop differentiation, cardiac myocytes were produced in culture. Morphologically indistinguishable from other cell types, however, cardiac myocytes have the property to beat in vitro, making them easily recognizable, but not easily measurable. Addition of Ca2+-free medium impedes cardiac myocytes beating, allowing measuring of TFIIH mobility in these differentiated cells. In these cells, we measured a TFIIH bound fraction of 43%, an intermediate fraction between the mobility observed in cardiac myocytes in situ and the undifferentiated ES cells. All together, our results show that during cellular differentiation of some cell types (neurons, hepatocytes, osteocytes, thricocytes, and myocytes), the dynamic organization of the basal transcription machinery is radically changed, whereas in other cell types (keratinocytes, fibroblasts, and chondrocytes), the dynamic framework of TFIIH activity is maintained.
Previous live-cell studies on complex multifactorial chromatin-associated processes that take place in mammalian cell nuclei, such as transcription, replication, and various DNA repair processes, have disclosed a general model in which highly mobile proteins (process factors) interact with sites of activity (e.g., promoters or DNA lesions) on the basis of stochastic collisions to form transient local machineries in an ordered but highly versatile manner [16],[48],[49]. These biologically relevant novel concepts, however, have been obtained mainly by exogenous expression of tagged factors within highly replicative cells in culture.
In an attempt to study the essential basal transcription initiation factor TFIIH within cells embedded in their natural environment (tissue), we designed a mouse model that allows quantitative determination of TFIIH dynamics. Using carefully designed targeted integration of the live-cell marker YFP at the Xpb locus (expressing the TFIIH subunit XPB), we obtained expression of functional YFP-tagged XPB protein under control of the endogenous promoter (guaranteeing physiological expression); we now find evidence for a fundamentally different scenario for the organization of basic transcription initiation in some cell types in the organism. Postmitotic cells (neurons, myocytes, and hepatocytes, etc.) appear to apply a largely static organization of transcription initiation with components being stably bound to chromatin that otherwise in other cell types (fibroblasts, chondrocytes, and keratinocytes, etc.) exchange constantly in a highly dynamic manner. In postmitotic cells, TFIIH is bound to promoters with a much longer residence time than in proliferative cells. A possible explanation for this static behavior is that in these postmitotic terminally differentiated cells, a large part of the transcription program is dedicated to a specific subset of genes defining cellular specialty and housekeeping functions, without the need to continuously switch to transcribed genes that are involved in proliferation (cell cycle, replication, and mitosis, etc.). It is possible that regular replication of the genome in proliferating (cultured) cells and tissues causes a continuous resetting of the transcription regulation machinery after each round of cell division and, in parallel, would involve a more open and accessible chromatin conformation.
It is generally accepted that differentiation requires and causes a resetting of the transcriptional program by activating and down-regulating specific genes in response to internal and external stimuli, thereby utilizing lineage-specific transcription activators and/or repressors. However, here, we have identified a novel concept of differentiation-dependent spatio-temporal organization of transcription initiation. This concept implies that transcription initiation factors, such as TFIIH, will be bound to promoters much longer in certain cell types than in others (Figure 6). Previously, differences in dynamic associations of lineage-specific transcriptional activators [50],[51] and elongation factors [21],[52] have been linked to activation of transcription. However, the herein described dynamic association of the basal initiation factor TFIIH in neurons, hepatocytes, and myocytes is likely not linked to a higher level of transcription than in, for example, keratinocytes, fibroblasts, or ES cells. We propose a model that the observed low mobility of TFIIH in highly differentiated postmitotic cells is derived form the establishment of a differentiation- and lineage-specific transcriptional program. However, slow mobility of TFIIH is not a general differentiation-dependent phenomenon because in, e.g., fibroblasts, chondrocytes, and keratinocytes, although differentiated, a much higher mobility of TFIIH is present. This observation excludes that this phenomenon is caused by a general change in mobility of histones during ES differentiation, as was previously described [53].
The fast TFIIH remobilization observed in neurons after a cold shock or the induction of local DNA damage demonstrates that, within neurons, TFIIH is still able to respond promptly to a “stress situation” that requires a rapid adaptation of the transcriptional program (in the case of the cold-shock response) or to be implicated in DNA repair. Thus remarkably, the static involvement of TFIIH in transcription initiation does not interfere with the flexibility of cells to change the nuclear organization in response to changing conditions. It is of interest to know how multifunctional factors such as TFIIH are still able to switch from one functional role to another and to relocate to other activity sites despite their virtually immobile nature. Further analysis is required to identify which subroute of NER (global genome NER, transcription-coupled NER, or differentiation-associated repair [DAR] [54]) is employed to repair genomic injuries in differentiated cells or whether lesions in permanently inactive sequences are repaired at all. The strategy outlined here has allowed us to address how transcription is organized in fully differentiated tissues or organs and during differentiation and development. Insights into these processes at the level of an intact organism are also relevant for a better understanding of the molecular basis of cancer and aging-related pathology. Importantly, our mouse model can be crossed into different genetic backgrounds, including existing TFIIH mutated mouse models [12],[55], mutated in another TFIIH subunit, i.e., XPD. These mice are associated with a puzzling clinical heterogeneity ranging from cancer predisposition to dramatically accelerated aging [1],[6]–[8],[12],. For instance, investigating TFIIH engagements directly in affected cells (neurons) in living tissues of XP/CS or TTD mice, which harbor a mutation in one of the other TFIIH components (XPD) [56], will help to elucidate the peculiar phenotype observed in these syndromes.
All animal work have been conducted according to Federation of European Laboratory Animal Science Associations (FELASA) ethical requirements and according to the respect of the 3R animal welfare rules.
The knock-in targeting vector (backbone pGEM5ZF) consisted of an approximately 7 Kb (NsiI/SalI fragment) mouse genomic DNA (isogenic to 129 OLA), which contains the 3′ part of the Xpb locus. The locus was modified by site-directed mutagenesis (SDM) to transform the stop codon into a coding amino acid (tryptophan) and two unique restriction sites were inserted for cloning purposes (i.e., SacII at the stop codon and NotI at a distance of 20 bp downstream of the SacII site). Between the SacII and NotI, a modified YFP was cloned, the YFP start was modified into a valine to avoid undesired translation of nonfused YFP. The YFP gene was further tagged at the C-terminus with a stretch of six histidines and an HA epitope sequence. A unique ClaI site was created 10 bp downstream of the stop codon of the modified YFP to introduce a neomycin gene-expression cassette, flanked by two LoxP sites [57], used as a dominant selectable marker. The dominant marker was inserted in the same transcriptional orientation as the Xpb gene.
ES cells (129 Ola, subclone IB10) were cultured in BRL-conditioned medium supplemented with 1,000 U/ml leukemia inhibitory factor. A total of 20 µg of the PmeI linearized targeting vector was electroporated into approximately 107 ES cells in 500 µl. Selection with 0.2 µg/ml G418 was started 24 h after electroporation. After 8–10 d, G418 resistant clones were isolated. Screening for homologous recombinants was performed using DNA blot analyses of NcoI-digested DNA with a 400-bp 5′ external probe (see Figure 1A and Figure S1A). Out of 128 G418 resistant clones, 12 ES clones had a correctly targeted Xpb allele. Two out of the 12 correctly targeted ES clones, checked for proper caryotype, were injected into blastocysts of C57Bl/6 mice and transplanted into B10/CBA foster mothers. Chimeric mice were further crossed, and germline transmission of the targeted allele to offspring was genotyped by PCR using primer sets (as described in Figure S1) and genotyping of offspring was done by PCR (see Figure S1). Primers sequences are available on request.
The knock-in Xpby (resulting fusion gene between Xpb and YFP, coding for XPB-YFP) allele was maintained in both FVB and C57BL/6 backgrounds. We thank P. Vassalli for pCAGGSCre plasmid used to generate the transgenic CAG-Cre recombinase–expressing mice. EGFP-expressing mice were kindly provided by Dr. Okabe [33] and Dr. R. Torensma (Nijmegen).
Murine dermal fibroblasts (MDF) were extracted from Xpb+/+, Xpby/y, and Xpby/y•Xpc−/− mice, following established procedures [58] and cultured in a 1∶1 mixture of Ham's F10 and DMEM (Gibco) supplemented with antibiotics and 10% fetal calf serum at 37°C, 3% O2, and 5% CO2. To induce in vitro differentiation of XPB-YFP–expressing ES cells, we applied the hanging drop method [47]. Briefly, 20 µl of ES cell suspension (2×105/ml in DMEM [Lonza] with 20% fetal calf serum [Lonza], 50 U penicillin/ml, and 50 µg streptomycin/ml [Lonza], 1% nonessential amino acids [Lonza], 0.1 mM B-mercaptoethanol [Sigma]) were placed on the lids of Petri dishes filled with PBS. After culturing for 3 d, the aggregates were transferred into bacteriological Petri dishes. Two days later, embryoid bodies were placed in Costar six-well plates with gelatin-coated coverslips for further development into different cell tissues. From day 3 on, retinoic acid (10−8 M) was added to induce skeletal muscle differentiation. Treatment with ultraviolet (UV) light at 254 nm (UV-C) was performed using a Philips germicidal lamp. For UV-survival experiments, cells were exposed to different UV-C doses, 2 d after plating. Survival was determined 3 d after UV irradiation by incubation at 37°C with 3H-thymidine, as described previously [59]. For unscheduled DNA synthesis (UDS), cells were exposed to 8 J/m2 of UV-C and processed as described previously [28].
Organotypic explants of cerebral cortex and cerebellum were produced as previously described [31],[60]. Tissues were analyzed on the same day of extraction in Neurobasal-A (GIBCO) medium, supplemented with antibiotics and B-27 at 37°C, 3% O2, and 5% CO2. FRAP analysis on cells within organotypic slices, maintained in culture for 1 wk, gave the same results as when performed on freshly extracted slices. Organotypic slices of heart, liver, and intestine were produced by cutting 300 µm of the organ with a Tissue Chopper (McIllwain). Slices were analyzed within 2 h following preparation, unless differently specified in the text. Skin tissues (epidermis and dermis) were prepared as described [10]. Before imaging, the two layers were mechanically separated and mounted on a coverslip for imaging and analysis.
Whole-cell extracts (WCE) of Xpb+/+ and Xpby/+ ES cells were prepared by isolating cells from a semiconfluent Petri dish (10 cm). Cells were washed with phosphate-buffered saline (PBS) and homogenized by sonication. Western blot analysis was performed as previously described [1].
A mixture of Xpb+/+ and Xpby/y cells were grown on glass coverslips and fixed with 2% paraformaldehyde at 37°C for 15 min. Immunofluorescence analysis was carried out as previously described [1]. Antibodies use were a rabbit polyclonal anti XPB (1∶500, S-19, Santa Cruz Biotechnology), a mouse monoclonal anti-p62 (1∶1,000, 3C9, kindly provided by Dr. J. M. Egly), and a rat monoclonal anti-HA (1∶1,000, 3F10, Roche).
FRAP experiments were performed as described before [24],[61] at high time resolution on a Zeiss LSM 510 meta confocal laser scanning microscope (Zeiss). Briefly, a narrow strip spanning the nucleus of a cell was monitored every 200 ms at 1% laser intensity (30 mW argon laser, current set at 6.5 A, 514-nm line) until the fluorescence signal reached a steady level (after circa 4 s). The same strip was then photobleached for 60 ms at the maximum laser intensity. Recovery of fluorescence in the strip was then monitored every 200 ms for about 30 s (1% laser intensity). All FRAP data was normalized to the average prebleached fluorescence after removal of the background signal. Every plotted FRAP curve is an average of at least ten measured cells.
To estimate the relative TFIIH bound fractions (BF) from the FRAP measurements, we used the first data point after photobleaching (Fmin) as an approximation of the baseline fluorescence recovery (BF = 1), i.e., the fluorescence level in the absence of recovery, when all proteins are considered immobile. We calculated the time-average fluorescence signal taken between 2 and 3.9 s prior to the photobleaching step to obtain the average prebleach fluorescence (Fpre), and then between 10 and 15 s to estimate the final fluorescence recovery level (Fmax). The bound fraction is then given by:
We corrected for the photobleached fraction, i.e., the incomplete recovery of fluorescence due to irreversible YFP bleaching during the FRAP procedure, as follows: whole nuclei of cultured keratinocytes were first imaged, subsequently strip-bleached (60-ms photobleach at maximum laser intensity), then imaged immediately after the bleach pulse and again 1 min later when no traces of the bleached strip were observed. The photobleached fraction (PBF) relative to the baseline fluorescence recovery (Fmin) was estimated as the average fluorescence intensity loss between the prebleached image (Fnucleus (pre)) and the last image of the nucleus (Fnucleus (last)):The corrected bound fraction is then given by:
Note: Measuring conditions where designed solely to measure immobile fractions, not diffusion or dissociation constants.
Laser-induced DNA damage was conducted as previously described [45]. Briefly, a Coherent Verdi pump laser with a Mira 900 mode locked Ti:Sapphire laser system (Coherent) was directly coupled to a LSM 510 NLO microscope (Zeiss) to obtain an 800-nm pulsed output (200-fs pulse width at 76 MHz, 10 mW output at the sample). Single nuclei targeted with the multiphoton laser received an approximately 250-ms exposure restricted to a 1-µm-wide strip.
The applied ChIP protocol was adapted from previously described methods [62],[63]. Briefly, brain fragments were fixed by adding 11% formaldehyde solution containing 50 mM Hepes (pH 8), 1 mM EDTA, 0.5 mM EGTA, and 0.1 M NaCl to a final concentration of 1% and incubated for 15 min at room temperature and 1 h at 4°C. Cross-linking was stopped by addition of glycine to a final concentration of 0.125 M. Fragments were washed twice with cold phosphate-buffered saline and treated with lysis buffer (50 mM Hepes [pH 8], 140 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 10% glycerol, 0.5% NP-40, 0.25% Triton X-100) containing 1 mM PMSF and a mixture of protease inhibitors. Fragments were homogenized (Ultra turrax, T25 basic), and ground on ice with an A-type and B-type glass pestle (20 strokes each) to allow nuclei release. Nuclear suspension was then sheared extensively by sonication on ice to obtain fragments of 200 to 600 bp (as revealed by ethidium bromide staining of aliquots run on agarose gels). For each ChIP reaction, 100 µg of cross-linked chromatin was immunoprecipitated with 0.5 µg of HA-antibody in RIPA buffer (10 mM Tris-HCl [pH 8], 1 mM EDTA, 0.5 mM EGTA, 140 mM NaCl, 1%Triton X-100, 0.1% Na-deoxycholate, 0.1% sodium dodecyl, 1 mM PMSF, and a mixture of protease inhibitors) overnight. The immunocomplexes were collected by adsorption (3 h) to precleared protein G sepharose beads (Upstate) precoated in RIPA containing 0.1 mg/ml sonicated salmon sperm DNA (ssDNA). The beads were then washed twice with 20 volumes of RIPA and once with RIPA containing ssDNA (Sigma), and twice with RIPA containing ssDNA and 0.3 M NaCl. Finally, the beads were washed with 20 volumes of LiCl buffer (10 mM Tris-HCl [pH 8], 1 mM EDTA, 0.5 mM EGTA, 0.25 M LiCl, 0.5% triton X-100, 0.5% Na-deoxycholate, 1 mM PMSF, and a mixture of protease inhibitors), resuspended in RIPA buffer, and divided into two equal parts to analyze coprecipitating proteins and DNA sequences.
For DNA analysis, the immunocomplexes were treated with RNAse (5 µg/µl) for 30 min at 37°C and by proteinase K (200 µg/µl) 3 h at 55°C in 50 mM Tris-HCl (pH 8), 1 mM EDTA, 100 mM NaCl, 0.5% SDS. Formaldehyde cross-links were reverted by heating the samples at 65°C for 6 h. The cross-linked DNA was extracted with phenol:chloroform and precipitated with ethanol in the presence of carrier glycogen. Pellets were resuspended in 30 µl of distillated water. Chromatin-immunoprecipitated DNA was subjected to PCR amplification. PCR was performed using 200 ng (for promoter XPB gene amplification) and 100 ng (for untranscribed sequences) of the chromatin immunoprecipitate and 400 nM concentration of both sense and antisense primers (primer sequence available upon request) in a final volume of 25 µl using the PureTaq Ready-to-go PCR beads (GE Healthcare). PCR products were analyzed on 2% agarose7 gels by SYBR Green reagent. Data were analyzed and quantified using Quantity One program (Bio-Rad). For adjustment of amplification efficiency of each primer set, PCR signal intensities from chromatin-immunoprecipitated DNA were normalized to those from the input genomic DNA and expressed as a percentage of the input (gDNA).
Brain slices and cultured cells were incubated at 37°C and 4°C. Additionally, cultured cells were incubated with 0.1 µg/ml of actinomycin D for 2 h. Samples were homogenized in Trizol using the Tissuelyser (Qiagen) for 90 s. RNA isolation was performed using the RNeasy Mini Kit (Qiagen). cDNA was produced using 2 µg of RNA with a Reverse Transcription Kit (Invitrogen), random primers (Invitrogen), 1 µg of cDNA, and 400 nM concentration of both sense and antisense primers (primer sequence and PCR cycle available upon request) in a final volume of 25 µl using the PureTaq Ready-to-go PCR beads (GE Healthcare). PCR products were analyzed on 2.5% agarose gels.
|
10.1371/journal.pntd.0002127 | Will Dengue Vaccines Be Used in the Public Sector and if so, How? Findings from an 8-country Survey of Policymakers and Opinion Leaders | A face-to-face survey of 158 policymakers and other influential professionals was conducted in eight dengue-endemic countries in Asia (India, Sri Lanka, Thailand, Vietnam) and Latin America (Brazil, Colombia, Mexico, Nicaragua) to provide an indication of the potential demand for dengue vaccination in endemic countries, and to anticipate their research and other requirements in order to make decisions about the introduction of dengue vaccines. The study took place in anticipation of the licensure of the first dengue vaccine in the next several years.
Semi-structured interviews were conducted on an individual or small group basis with government health officials, research scientists, medical association officers, vaccine producers, local-level health authorities, and others considered to have a role in influencing decisions about dengue control and vaccines. Most informants across countries considered dengue a priority disease and expressed interest in the public sector use of dengue vaccines, with a major driver being the political pressure from the public and the medical community to control the disease. There was interest in a vaccine that protects children as young as possible and that can fit into existing childhood immunization schedules. Dengue vaccination in most countries surveyed will likely be targeted to high-risk areas and begin with routine immunization of infants and young children, followed by catch-up campaigns for older age groups, as funding permits. Key data requirements for decision-making were additional local dengue surveillance data, vaccine cost-effectiveness estimates, post-marketing safety surveillance data and, in some countries vaccine safety and immunogenicity data in the local population.
The lookout for the public sector use of dengue vaccines in the eight countries appears quite favorable. Major determinants of whether and when countries will introduce dengue vaccines include whether WHO recommends the vaccines, their price, the availability of external financing for lower income countries, and whether they can be incorporated into countries' routine immunization schedules.
| Information gleaned from surveys of country-level policymakers and other opinion leaders can assist in planning the development, production and introduction of new or upcoming vaccines into public sector immunization programs. In the case of dengue vaccines, prevailing views among these leaders about the importance of the disease, their expressed level of interest in the government's use of the vaccine, and preferred strategies for vaccine introduction (e.g., geographically-targeted vs. nation-wide vaccination, specific age groups to target) can help to identify “early adopter” countries and indicate the level of demand for the vaccine. This information can be critical to current producers of the vaccine in planning their production capacity and to potential future producers in deciding whether to pursue development of the vaccine. This information also helps donors and international technical agencies, such as WHO and UNICEF, in setting their priorities and determining their level of technical and financial support to countries for the introduction of dengue vaccines. In addition, these surveys can provide crucial information to national governments and the above stakeholders about potential barriers to introducing dengue vaccines into national immunization programs, and what additional studies and data countries will require in order to make decisions about use of the vaccines in the public sector.
| Dengue, a mosquito-borne Flavivirus infection caused by four related viruses (DENV1 to 4), is a major public health problem in the tropics and subtropics. The greatest documented burden of dengue occurs in Asia and Latin America. Dengue's geographic range now places an estimated 3.97 billion people at risk and it continues to expand, causing epidemics that disrupt health care systems [1], [2]. The World Health Organization estimates that each year there may be up to 50 million dengue infections worldwide and 500,000 cases of dengue hemorrhagic fever (DHF) – a severe form of the disease [3]. According to the 2010 Global Disease Burden Study, dengue causes nearly 15,000 deaths per year, a 29% increase since 1990 [4].
Dengue vaccines have been under development since the 1940s, but the vaccine industry's interest in the vaccines languished throughout most of the 20th century [5]. However, dengue vaccine development has accelerated in recent years and several vaccine candidates are in or near to human clinical development. The most advanced is a recombinant live chimeric tetravalent vaccine (CYD TDV) developed by Sanofi Pasteur [6], consisting of four genetically engineered viruses in which several genes in a yellow fever DNA backbone have been replaced with comparable genes from the four dengue viruses. The vaccine is being evaluated in a three-dose regimen given over a one-year period (at six month intervals) in efficacy trials in multiple countries in Asia and Latin America. Although this and all other dengue vaccines are intended to be used with children and adults, initial licensure is likely to be limited to children 2–14 years of age, since the current clinical trials are evaluating the vaccine in this age group.
Two other live attenuated chimeric dengue vaccine candidates in are Phase II evaluations: one developed by the U.S. National Institutes of Health [7], and another being developed by Inviragen [8]. In addition, a recombinant subunit vaccine under development by Merck has completed Phase I trials [9].
This survey of eight dengue-endemic countries in Asia and the Americas was conducted by the Pediatric Dengue Vaccine Initiative (PDVI), managed by the International Vaccine Institute, to determine their possible interest in and perceived need for a dengue vaccine; what factors (e.g., vaccine characteristics, financing, political pressure, data needs) would drive or influence decisions; and what strategies countries would likely use to target, deliver and finance dengue vaccination. The main objectives of the survey were to: 1) provide an indication to donors, vaccine producers, and the international health community of the potential interest in and demand for dengue vaccination from endemic countries and what factors may affect this demand; and 2) anticipate the research, disease surveillance and other requirements that countries will have in order to make decisions about the introduction of dengue vaccines.
Surveys of policymakers and other stakeholders about new or under-utilized vaccines have been used in the past to inform research and advocacy activities of product development partnerships, including a multi-country study about rotavirus vaccines [10] and a seven-country survey concerning cholera, typhoid fever and shigellosis and vaccines against these diseases [11].
The survey was conducted in four Asian countries (India, Sri Lanka, Thailand and Vietnam) and four Latin American countries (Brazil, Colombia, Mexico and Nicaragua). This was a convenience sample of countries that are considered dengue-endemic where PDVI had contacts viewed as reliable authorities on dengue in their countries and who could facilitate arrangements for interviews with policymakers. Several of the countries were viewed by the investigators as potential early adopters of dengue vaccines based on reports of dengue outbreaks and expressed concern from health officials and the public about dengue in the country. No country that was requested to participate in the study refused to do so.
The study consisted of face-to-face interviews conducted on an individual or small group basis during country visits that took place between September 2008 and December 2010. All policymakers and other stakeholders interviewed consented to the interviews, which were voluntary, and participants were informed that their responses would be anonymous.
The interviews were semi-structured, using a question guide consisting entirely of open-ended questions (Appendix 1). The guide explored informants' views and perceptions about:
The semi-structured interview approach was felt to be the most appropriate for high-level informants, as opposed to a highly-structured questionnaire. This format facilitated the free expression of opinions and ideas among informants, allowed for probing and clarification of responses, and for the identification of new issues and topics as they arose. All informants were asked a core set of questions concerning their views about dengue, current prevent and control measures, and various aspects about dengue vaccines, while additional specific questions were asked to individuals, according to their expertise and position. For example, immunization program officials were asked about government plans and priorities for new vaccine introductions and vaccine producers were asked about the status and plans of vaccine development.
The interviews were conducted in English by one or more of the authors in each country. Local collaborators who arranged the interviews and meetings, also sat in on several of them and served as translators, as needed, in interviews conducted in the Latin American countries.
Persons to be interviewed in each country were identified based on a list developed by PDVI of the types of organizations and individuals to target. An effort was made to meet as many individuals and groups as possible who potentially have a role in making or influencing decisions about the future introduction of dengue vaccines into national immunization programs, including decision-makers within health ministries; chairs or members of national immunization technical advisory groups (NITAGs); immunization program heads; international aid agencies; leading scientists; national regulatory authorities; and local vaccine producers. Efforts were also made to meet with the highest level officials in each organization and category as possible. If people who were suggested to be interviewed were not available or did not respond, they or their office usually recommended a colleague within the organization or department to interview.
As shown in Table 1, those interviewed included senior officials from ministries of health (MOHs) (e.g., Directors General of Health Services, health secretaries and directors of communicable diseases) and other government health agencies (e.g., centers for disease control, national regulatory authorities, national health insurance agencies). They also included national immunization program managers (in five countries); and MOH officials and scientists from dengue control and vector control programs, disease surveillance and epidemiology, infectious disease control and planning departments. Other informants included officials and scientists from public and private research institutes and academic institutions; top officials of major hospitals; officers of medical associations; health authorities at the regional, state and/or municipal level (including state or provincial health ministers or the equivalent and chief medical officers); officials from international technical agencies (e.g., WHO, UNICEF); and representatives from local vaccine producers (in India, Vietnam, Brazil and Mexico) or multi-national pharmaceutical companies (in Colombia). Several of those interviewed also served on their country's NITAG.
Between 14 and 32 persons took part in the interviews in each country, for a total of 158 individuals (average of ≈20 per country).
The results were analyzed separately for each country. Extensive notes were taken for each interview and a complete set of notes were transcribed by topic area and then by person or group interviewed, including salient quotations. From these transcripts, the types and patterns of responses were analyzed by the type and level of informants. Responses were reported if at least two persons in a country gave a similar response. Individual country reports were then prepared that included sections on perceptions about the importance and priority of dengue in the country; perceptions about current dengue prevention and control; the perceived need for dengue vaccines; concerns, criteria and data needs regarding dengue vaccines; and possible vaccine introduction strategies and scenarios. Both the raw data (interview notes) and country reports were used in preparing this manuscript.
The majority of policymakers, hospital directors, and other informants whose work is not focused solely on dengue considered dengue an important disease and a priority in all eight countries included in the study. The disease was called a “major public health problem” by a senior health official in India, and “right at the top” [of health priorities] by a senior MOH official in Sri Lanka. Dengue was deemed “unappreciated by donors and the government” by preventive medicine officials and NITAG members in Vietnam. Similarly, senior health policymakers in Colombia described the disease as a “very high priority” and “very important” in terms of hospitalization and loss of productivity. In Brazil, the chair of the country's NITAG described dengue as one of only two diseases without effective control (the other being leishmaniasis) and considered it a high priority, along with tuberculosis, hepatitis and HIV/AIDS. In Nicaragua, dengue was described by a senior Ministry of Health official as not normally one of the top 20 priority diseases, but when there is an outbreak, it rises to the top and becomes the number one priority.
The disease is especially a high priority among those on the frontlines of dengue prevention and control. City health officials in Colombo, Sri Lanka ranked dengue their number one disease priority, and it is the only disease which the city government tracks through GIS mapping. City health officials also keep a map showing all cases of the disease on a large poster board (along with leptospirosis) that is updated daily.
Agreement among policymakers and other informants about the importance of dengue was not universal, however. A senior health official in India did not consider dengue a top priority like malaria, while some hospital officials in the country believe that the extensive media attention given to dengue outbreaks takes the focus away from other important endemic diseases, such as diarrhea, enteric fevers and hepatitis. And according to some hospital officials in Thailand, dengue is now less recognized by politicians as a major problem due to the country's success in reducing dengue shock syndrome (DSS) and deaths to a low level.
The main reasons why informants considered dengue a priority disease are the following:
This was a key factor mentioned by informants in all countries except Colombia and Nicaragua (where dengue is still largely considered an urban disease). According to interviewees in Sri Lanka, for instance, dengue was mainly confined to the city of Colombo in 1996, but by 2004 it had spread to 10 districts, and by 2007 to nearly all of the country's 26 districts. In Vietnam, where dengue has been endemic in the South for decades, outbreaks occurred in the North of the country in 2008 for the first time. The spread of the disease to peri-urban and even rural areas was a common perception and concern among informants in India, Sri Lanka, Vietnam, Thailand, Brazil, and Mexico, which they attribute to increased development of these areas, leading to building construction, proliferation of garbage, open water storage and other conditions favorable to the breeding of dengue-carrying mosquitoes. According to one informant in Brazil, 40% of dengue cases reported since 2007 have come from the outskirts of large cities. Some informants in Vietnam, Thailand and India believed that dengue is considerably more under-reported in rural areas than in cities and that the incidence could be as great in rural areas as in urban areas.
Informants in several countries, including Sri Lanka, Thailand and Nicaragua, report an increase in dengue incidence over the past several years. Those in the Asian countries of Sri Lanka, Vietnam, Thailand report that major outbreaks are now occurring every year, while as recently as five or six years earlier, they used to occur every three or four years, The disease was described in Sri Lanka as changing from an epidemic disease to a “hyper-endemic” disease that occurs throughout the year.
According to dengue researchers in Brazil, dengue overwhelms the health care system in urban areas during outbreaks more than any other disease, in terms of its numbers and severity. Informants in other countries report a similar phenomenon. During a major outbreak in Delhi, India in 2006, dengue wards were set up in tents outside of a major hospital, where 1,200 cases were treated daily. Dengue cases make up thirty to forty percent of all pediatric patients at a provincial hospital in Thailand during outbreaks. And at a national children's hospital in Managua, Nicaragua, the number of beds was increased 64% during an outbreak in 2009, mainly to accommodate dengue patients. The burden placed on hospitals during dengue outbreaks is increased even more by the panic that strikes the public, according to informants in Sri Lanka and Brazil. One expert in Sri Lanka describe “dengue phobia”, which causes parents to rush their child to a health facility at the first sign of a fever, fearing that it's dengue.
Because dengue occurs in outbreaks and urban populations are affected, it garners considerable media attention, placing pressure on politicians – especially at the local level – to find a way of controlling the disease. According to one informant in Thailand, whenever a child dies of dengue, it is reported in the media. Given this visibility, there can be high costs to political leaders viewed as failing to control outbreaks. In Brazil, mayors as well as officials responsible for vector control have been replaced because of dengue outbreaks, according to informants. Outbreaks in Colombia have prompted political leaders to declare a state of emergency, as occurred in Cali in 2010 after more than 1,200 cases and nine deaths were reported.
Informants in both Asian and Latin American countries (India, Thailand, Nicaragua) gave as a major reason for dengue being a priority the fear among physicians of an otherwise healthy, well-nourished child deteriorating in a matter of hours, leading to DSS, other complications, and even death. As one hospital director in Thailand stated, “Death due to dengue is particularly tragic, because children go from healthy to fatal very rapidly.” The speed at which complications developed during a major dengue outbreak in Managua, Nicaragua in 2008 took medical experts and hospital officials by surprise; even within the first day of the onset of symptoms, patients were developing severe symptoms, including circulatory shock. However, informants in Thailand and Vietnam pointed out that rates of DSS and dengue-associated deaths have remained the same or have even decreased in the past decade or so, due to early and more effective treatment.
Dengue has been viewed as mainly a children's disease in Asia and as an adult disease in Latin America. However, in some Latin American countries, notably Brazil, children are increasingly being affected by the disease, causing concern in the medical and public health community. A dramatic shift in DHF cases from adults to children occurred in Brazil in 2007 [12], which informants claim is responsible for the increased severity of the disease in the country.
Informants interviewed in all countries except India gave the economic impact of dengue as a reason for their concern about the disease. In Colombo, Sri Lanka, 20% of the city government's entire health budget – which covers the cost of health clinics, maternity homes, water quality and many other activities – is spent on dengue-related activities. Officials in one province in Nicaragua estimated that dengue control efforts during outbreaks typically consume 60% of the province's emergency budget. Hospital costs during dengue outbreaks can also put a severe strain on health budgets, as reported in Mexico.
Another reason given in some countries for the perceived growing importance of dengue is the great progress made in controlling or reducing mortality from other major infectious diseases at the same time as dengue incidence is increasing and/or expanding geographically. In Sri Lanka, the control of tuberculosis and malaria through intensive treatment, Japanese encephalitis (JE) through vaccination, and diarrheal disease deaths through oral rehydration therapy has reduced the relative importance of these diseases among policymakers, resulting in dengue rising to the top of infectious disease priorities. Similarly, in Nicaragua, informants pointed out that malaria and several other diseases (measles, rubella) have largely been brought under control in recent years, while little progress has been made on dengue.
A further reason for the sense of priority of dengue among informants, as mentioned in Thailand, Brazil and Nicaragua, is the fact that the disease strikes all sectors of society – rich and poor alike – and thus no one is immune from getting the disease.
Interest among policymakers and opinion leaders in the public sector use of dengue vaccine was on the whole quite high – though not universal – in the eight countries surveyed. A senior MOH official in Sri Lanka, where interest in dengue vaccines was universally high amongst those interviewed, believed that the government would introduce a dengue vaccine if it was affordable, even in the absence of donor funding. Similarly, a health policy expert in Mexico claimed that a dengue vaccine would be accorded a “high priority” by the government once one becomes available. And according to a high-level health ministry official in Nicaragua, the government would have a “genuine interest in making a [dengue] vaccine available to the population that needs it most.” A provincial health official in the country claimed that the introduction of a dengue vaccine would be “a priceless achievement in public health”, saving the country, the health system and people money.
Government officials in India were more hesitant to state an interest by the government in using dengue vaccines, in the absence of data about their safety and performance. A senior health official was skeptical of the need for the vaccine, claiming that vector control is preferable to a vaccine and is succeeding. However, expressed interest in dengue vaccines in India was high among non-government informants on the frontline of treating dengue cases, such as hospital officials and representatives of professional medical associations.
According to interviewees in several countries (Sri Lanka, India, Thailand, Brazil, Mexico), a key driver of government interest in dengue vaccines is the great pressure that the public and in some cases, the media will put on the government to introduce the vaccines, once available. The high expected public demand is due to the outbreak pattern of the disease, the media attention that it attracts, the public's high awareness of the disease in these countries, and the fear that it engenders in the population. In Thailand, interviewees believed that such pressure would result from the inequities created by having a dengue vaccine available in the private market but not through the national immunization program for free. Pressure would also come from the medical community in many countries, given the lack of specific treatment for dengue and the difficulty in predicting its course. According to several informants in Sri Lanka, such population demand, coupled with media pressure, is a stronger driver of new vaccine introductions than evidence of disease burden or cost-effectiveness, since politicians “are sensitive to population demand”. An official from an Indian vaccine producer, in fact, described vaccines against dengue and Japanese encephalitis (JE) as “political vaccines”, given the public's tendency to blame politicians for outbreaks of these vector-borne diseases in their community. The potential political benefits of introducing a dengue vaccine is described by a health policy expert in Mexico: “A courageous and early decision taken about dengue vaccines by the next government could be an early win for [them]. Launching a dengue vaccine to coincide with the political cycle could be most helpful.”
Despite the generally high level of interest in dengue vaccines expressed by most informants in the eight countries, several, including government officials in India, Thailand, Mexico and Nicaragua, were more cautious about embracing vaccines whose safety, performance and cost are not yet known. According to these informants, questions about a dengue vaccine's safety, the number of doses required, its effectiveness and duration of protection, as well as its affordability and financing, would have to be answered before their government would use it in the public sector. As one provincial government official from Nicaragua stated, “It is important not to raise unrealistic public expectations about dengue vaccines and their ability to stop infections soon after introduction.” The relatively low mortality of the disease could also be an obstacle to its rapid introduction into government immunization programs, according to some informants. As one informant in Mexico declared, “The government would need to be convinced of the importance of making large budget allocations for dengue vaccines, since there are only around 2,000 cases of DHF and 20 deaths per year.”
The top concerns that informants had about dengue vaccines were their safety, cost and whether they can be used in infancy or early childhood.
While all countries will require evidence of the safety and efficacy of any new vaccine before it can be introduced into the national immunization program, both Sri Lanka and Vietnam now require that safety and immunogenicity be demonstrated in the local population through a small Phase I/II study, even for vaccines pre-qualified by WHO. (An exception is made in Vietnam for vaccines supplied by the GAVI Alliance. Immunogenicity studies are also not required in Vietnam for vaccines that have been licensed in other countries for five years or more.) Health policymakers in Sri Lanka, Thailand and Mexico expressed strong interest in conducting Phase IV post-marketing surveillance (PMS) studies once the vaccine has been introduced into the national immunization program to further monitor its safety. A vaccine researcher in Thailand also recommended using PMS to track the effects of vaccine introduction on individuals previously exposed to dengue, while Indian health officials expressed interest in studying the vaccine's tolerability in HIV positive individuals.
Several countries would require additional evidence of disease burden from different parts of the country before making a decision to introduce a dengue vaccine. Additional surveillance data would help to strengthen the case made to policymakers for dengue vaccination, and in countries, such as India, Colombia and Mexico, where targeted vaccination is likely, it would help determine which areas to prioritize for vaccination. The need for additional disease burden data for decision-making was mentioned less often in countries confident in their surveillance and reporting systems or where nation-wide vaccine introduction is assumed (e.g., Sri Lanka and Thailand).
Economic data were mentioned as critical evidence needed for policy decisions in all countries. According to health officials in India, cost-effectiveness estimates are increasingly important to the government, especially for newer, more expensive vaccines and “niche” vaccines. In Vietnam, informants reported that the Finance Ministry will only approve financing for imported vaccines if they are determined to be cost-effective. Also viewed as critical in various countries – both as stand-alone evidence and as data needed for the cost-effectiveness analyses – were data on the cost of the disease in their country, including the cost of treatment and hospitalization, the cost of vector control, and the economic impact of the illness on the poor.
The views of policymakers and other stakeholders concerning dengue and dengue vaccines were first surveyed in 2002 in a study of four Southeast Asian countries (Cambodia, Indonesia, Philippines and Vietnam) [14]. This new survey of eight countries in both Asia and Latin America confirms the generally high level of importance accorded to dengue by government policymakers and other stakeholders across dengue-endemic countries that was found in the 2002 study, as well as strong interest in dengue vaccines for public health use. There were few clear distinctions found in the views about dengue and dengue vaccines between Asian stakeholders and those in the Americas. One of the only discernible differences was the greater interest in vaccinating adults in the Americas compared to Asia, due to the older age distribution of disease in the Americas.
As in the 2002 survey, the perceived importance of the disease and the perceived need to have more effective tools to control it are driven by the fact that dengue is increasing in incidence and spreading within countries. Adding to the sense of urgency since the 2002 study is the fact that major outbreaks are becoming an annual occurrence in several Asian countries and thus, the disease is transforming from an epidemic to an endemic or even hyper-endemic disease. While dengue appears to be worsening in magnitude and severity, substantial progress has been made in recent years in controlling such high priority diseases as tuberculosis, malaria and measles in many dengue-endemic countries, effectively increasing the relative importance of dengue.
Another key factor contributing to the sense of priority of dengue and interest in dengue vaccines is the highly visible nature of the disease. A large part of this visibility is due to the fact that it often occurs in epidemics – which attract media attention, stoke fear and even panic in the public, can overwhelm hospitals, and put a strain on municipal budgets. This is in contrast to non-epidemic diseases, such as diarrheal disease and pneumonia, which exact a higher toll than dengue in terms of morbidity and mortality in many countries, but which attract less public or media attention. The importance of an epidemic disease pattern in creating a demand for a vaccine is demonstrated by the high priority that governments in the meningitis belt of Africa placed on the development of an effective vaccine against meningococcal meningitis A – which is now being used in mass campaigns in several countries – despite the relatively low death toll from the disease [15], [16]. Dengue's visibility is further enhanced by its occurrence in cities – which are the centers of the media and political leadership – and the fact that it strikes all social classes and not just the poor.
These factors have created considerable political pressure on governments to control the disease, and according to informants, will create pressure to introduce a dengue vaccine once one is available. Recent studies into factors influencing government adoption of vaccines suggest that political pressure – often fueled by public fear or anxiety about a disease – has contributed to decisions to introduce certain new vaccines, even in the absence of solid disease burden or cost-effectiveness data [17]–[19]. Political considerations were paramount in the rapid introduction of HPV vaccines in seven developed countries, despite uncertainty about the vaccines' long-term effectiveness and the lack of country-specific cost-effectiveness data in four of the countries [18]. In some countries, decisions to fund the HPV vaccine bypassed the normal decision-making process and were even made in the context of current or upcoming elections. And in the Netherlands, the government decided to introduce meningococcal C vaccine without a favorable cost-effectiveness analysis, in part to assuage public anxiety about the disease [19]. While these examples are from developed countries, this survey strongly suggests that similar pressures will be applied to dengue vaccines in many dengue-endemic developing countries.
The findings of this and the 2002 survey also highlight the differences in how policymakers and opinion leaders in many endemic countries define disease burden as compared to global institutions and donors. While international organizations tend to define the burden of disease in terms of mortality and morbidity, and thus consider dengue a low-mortality disease of low priority, many of those dealing with the disease in endemic countries take a much more comprehensive view of the dengue disease burden. They also take into consideration the economic costs of outbreaks on health systems, the cost of vector control, as well as such immeasurable, more “political” variables as the panic that outbreaks can cause among the public; the fear among doctors of patients deteriorating rapidly due to the unpredictable nature of the disease; and the demand from parents, the media and society at large for the government to prevent and control the disease. Since country-level policymakers must consider all of these factors, they tend to accord a higher priority to dengue than do global institutions at present.
The findings of this survey provide a blueprint for research and other activities that are needed to accelerate dengue vaccine introduction in public sector immunization programs in endemic countries. Primary among the research needs are more systematic assessments of the local dengue disease burden (e.g., sentinel site surveillance). Many countries will also require locally-generated data on the cost of dengue, including cost-of-illness and vector control costs, and on the cost-effectiveness of vaccination. Studies to demonstrate the safety and immunogenicity of the vaccine in the local population will also need to be undertaken in countries with an explicit policy requiring such studies for new vaccines (e.g., Vietnam and Sri Lanka), and perhaps in other countries as well, given the unique safety concerns about dengue vaccines. Pilot introduction or vaccine demonstration projects may be the preferred route to making a decision about vaccine introduction, as mentioned by informants in several countries and as shown in the literature to be an important factor in the introduction of hepatitis B vaccine in some early adopter countries (e.g., Thailand, Taiwan, Indonesia) [17].
The development of international recommendations, such as by the WHO Strategic Advisory Group of Experts (SAGE) on immunization, could also accelerate dengue vaccine introduction. Not only was this mentioned as an important factor in several countries in the study, the history of the introduction of Haemophilus influenzae Type b (Hib) vaccine suggests that the development of stronger recommendations by WHO in 2006 calling for universal use of the vaccine was a contributing factor to the rapid adoption of the vaccine, along with intensive advocacy and GAVI support in the years following the new recommendations [20].
Three additional conditions or criteria could be critical to the introduction of dengue vaccines in developing countries. One is vaccine prices that countries consider “affordable”. Dengue vaccine introduction will likely be aided in the Americas if the vaccine can be purchased through the PAHO Revolving Fund, which obtains lower prices than countries can generally obtain on their own. Second, donor financing, especially through the GAVI Alliance, will be critical to avoid long delays in vaccine introduction in many low- and lower-middle income endemic countries. GAVI eligibility shortened the time for countries to decide to introduce Hib vaccine by 63% and made up for differences in income levels between countries [21]. Finally, according to informants in several countries, policymakers will be more inclined to introduce a dengue vaccine if it can be incorporated into the countries' childhood immunization schedules.
There are a number of limitations of this study that must be considered. First, the countries included in the study were a convenience sample and not necessarily representative of all dengue-endemic countries or of all potential early adopters of dengue vaccines. Since this study represents only a subset of potential early adopters of the vaccine, additional studies in other countries would be of value. Nonetheless, the combined population of these eight countries makes up a significant portion of the global population at risk for dengue, including the largest country in each region with a substantial dengue burden (India and Brazil). Secondly, the sample of informants may not have included all major stakeholders in each country or have been representative of all stakeholders, since the sample size per country was quite small and certain key sectors, such as finance ministries, were often not available for interviews. In addition, by the time dengue vaccines become available, many of the major decision-makers and opinion leaders may have changed. Nonetheless, those interviewed in each country included representatives of groups found in the literature to be highly influential in the adoption of new vaccines, including senior Ministry of Health officials, NITAG members, officials from medical professional societies, leading academicians, and local vaccine producers [17], [22], [23].
There is also the possibility that informants' responses were biased towards playing up the importance of dengue and interest in dengue vaccines, since they were aware that the survey was being conducted by a dengue vaccine project. However, a number of respondents expressed less concern about dengue or were hesitant to embrace dengue vaccines, suggesting an atmosphere of free expression. As with all qualitative studies with open-ended responses, there is also the possibility of misunderstanding or biased interpretation of informant's responses. The structure of the interviews, which allowed for probing and clarification of responses, was designed to minimize misinterpretation. Bias could also arise in the selection of responses to report in the paper, although efforts were made in the analysis to find and examine opposing views within a country. However, many responses, especially those concerning the importance of dengue and interest in a vaccine, occurred repeatedly across respondents and across countries and are thus likely to transcend these possible biases and reflect the prevailing views of stakeholders concerning dengue and dengue vaccines in these eight countries.
|
10.1371/journal.pbio.1002543 | Sam68 Is Required for DNA Damage Responses via Regulating Poly(ADP-ribosyl)ation | The rapid and robust synthesis of polymers of adenosine diphosphate (ADP)-ribose (PAR) chains, primarily catalyzed by poly(ADP-ribose) polymerase 1 (PARP1), is crucial for cellular responses to DNA damage. However, the precise mechanisms through which PARP1 is activated and PAR is robustly synthesized are not fully understood. Here, we identified Src-associated substrate during mitosis of 68 kDa (Sam68) as a novel signaling molecule in DNA damage responses (DDRs). In the absence of Sam68, DNA damage-triggered PAR production and PAR-dependent DNA repair signaling were dramatically diminished. With serial cellular and biochemical assays, we demonstrated that Sam68 is recruited to and significantly overlaps with PARP1 at DNA lesions and that the interaction between Sam68 and PARP1 is crucial for DNA damage-initiated and PARP1-conferred PAR production. Utilizing cell lines and knockout mice, we illustrated that Sam68-deleted cells and animals are hypersensitive to genotoxicity caused by DNA-damaging agents. Together, our findings suggest that Sam68 plays a crucial role in DDR via regulating DNA damage-initiated PAR production.
| Maintaining genome integrity is crucial for all organisms, and failure to do so can lead to fatal diseases such as cancer. Exposure to challenging environments can induce DNA strand breaks or other lesions; thus, rapid and appropriate DNA damage responses (DDRs) need to be in place to detect and repair the damage. Cellular networks use a variety of signaling molecules and post-translational modifications that are crucial for the signaling of DNA breaks to repair machineries. Poly(adenosine diphosphate [ADP]-ribosyl)ation (PARylation) and activation of the enzyme poly(ADP-ribose) polymerase 1 (PARP1) is a post-translational modification that occurs within seconds upon DNA damage detection and triggers downstream DDR signaling; however, it remains obscure whether other molecules, beyond DNA strand breaks, stimulate or control PARP1 activity. We report here that a novel DDR signaling molecule, Src-associated substrate during mitosis of 68 kDa (Sam68), has a crucial function in governing the DNA damage-initiated PARP1 activation and polymers of ADP-ribose (PAR) production. We show that Sam68 is recruited to and significantly overlaps with PARP1 at DNA lesions and that the Sam68-PARP1 interaction is critical for DNA damage-initiated PARP1 activation and PAR production both in vitro and in vivo. Sam68-deleted cells and animals have a diminished PAR-dependent DNA repair signaling and are hypersensitive to genotoxicity caused by DNA-damaging agents. Hence, our data reveal an unexpected function for Sam68 in DNA damage-initiated early signaling and provide a novel mechanism on the activation and regulation of PARP1 in DDR.
| DNA damage responses (DDRs) that occur promptly are essential for maintaining genome integrity, which is consistently challenged by internal and external insults [1–6]. Failure to do so can lead to loss of genomic integrity and also cause cancer, immune deficiency, premature aging, and other critical conditions [3,5]. Sophisticated cellular networks, consisting of a variety of molecules and post-translational modifications, are crucial for signaling the presence of DNA strand breaks to repair machineries [3]. In particular, poly(adenosine diphosphate [ADP]-ribosyl)ation (PARylation), catalyzed by the enzymes from the poly(ADP-ribose) polymerase/diphtheria toxin-like ADP-ribosyl transferase (PARP/ARTD) family of proteins [7,8], is one of the earliest events (within seconds) in DDR [9–11]. Previous studies have underscored an indispensable role of PARylation in DNA repair pathways including base excision repair (BER), single-strand break repair (SSBR), homologous recombination (HR), and nonhomologous end joining (NHEJ) [12–17]. Importantly, the elongated and branched structure enables polymers of ADP-ribose (PAR) to serve as a docking platform for the focal assembly of DNA repair complexes, thus orchestrating appropriate DDR signaling cascades [18–26]. For instance, following γ-irradiation, phosphorylation/activation of the proximal checkpoint kinase ataxia telangiectasia mutated (ATM) as well as ATM substrates checkpoint kinase 1 (Chk1) and Chk2 occurs in a PAR-dependent manner [14,27]. As the founding member of PARP/ARTD superfamily, PARP1 (also named ARTD1) is the major enzyme responsible for the rapid and vigorous PAR synthesis triggered by damaged DNA [1,10]. Binding of PARP1 to DNA strand breaks results in conformational changes in PARP1 and elevates its activity [4]. Upon activation, PARP1 vigorously synthesizes and adds ADP-ribosyl polymers to a variety of target proteins, including PARP1 itself [9]. Albeit these important advances in understanding of the critical function of PARylation in DDR, the precise mechanisms of stimulation and regulation of PARP1 catalytic activity during DDR are still obscure. In particular, recent studies showed that DNA strand breaks appear not to be the sole stimulatory factor for PARP1 activation [9,10,28–33], which suggests that a more complicated mechanism or mechanisms could be required to robustly activate and elegantly fine-tune PARP1 activity. Moreover, the inhibition of PARP1/PARylation has emerged as a promising therapeutic approach for treating human cancers and inflammatory diseases associated with impaired DNA repair activities [34,35]. Of note, the current classes of PARP1 inhibitors, either approved by FDA or undergoing clinical trials, are all based on a competitive binding strategy first observed with nicotinamide [36]. Therefore, elucidating the molecular mechanisms of activation and regulation of PARP1 activity in DDR may provide clinical relevance to aid the rational development of new PARP1 inhibitors.
Src-associated substrate during mitosis of 68 kDa (Sam68) is a versatile RNA-binding protein and plays a role in a wide range of cellular processes, including RNA stability, RNA splicing, RNA nuclear export, HIV-1 replication, adipogenesis, neuronal activity, and others [37–47]. We and others recently showed that Sam68 participates in the transcription of certain genes via its interactions with transcription factors [43,48–51], which is in line with the fact that Sam68 binds single- and double-stranded DNA, besides RNA. Moreover, emerging evidence suggests that RNA-binding proteins play critical functions in DNA damage signaling [52–54]. In particular, Sam68 was identified as a PAR-binding partner in cells with DNA damage and proposed as a putative substrate of ATM, ataxia telangiectasia and Rad3-related (ATR), and DNA-dependent protein kinase (DNA-PK) [55,56], which strongly indicates that Sam68 could execute an important function in DDR. Although Sam68, a ubiquitously expressed RNA-binding protein, has been long acknowledged as an almost strictly nuclear protein [38,41], its potential function in the nuclear-initiated signaling pathways, especially DDR, has not been the subject of intense investigation. In this manuscript, we report that Sam68, as a novel signaling molecule in DDR, plays a crucial function in governing the DNA strand break-triggered PARP1 activation and PAR production. Upon DNA damage, Sam68 is recruited to and significantly overlaps with PARP1 at DNA damage sites. Interaction between Sam68 and PARP1 via their N-termini is critical for DNA dependent-PARP1 activation and PAR production in vivo and in vitro. In line with the attenuated PAR-dependent repair signaling, DNA damage is poorly repaired in Sam68-deficient cells and animals in comparison to the Sam68-sufficient controls. As a consequence, Sam68 knockout mice are hypersensitive to genotoxicity caused by γ-irradiation and DNA alkylating agents. Hence, our data reveal an unexpected function for Sam68 in DNA damage-initiated early signaling and provide a novel mechanism on the activation and regulation of PARP1 in DDR.
Hypersensitivity to DNA-damaging agents is one of the hallmarks of defective DDR. To address the role of Sam68 in repair of DNA strand breaks, we first examined the effect of Sam68 deletion in clonogenic survival of mouse embryonic fibroblasts (MEFs) following exposure to genotoxic stresses. Sam68 deletion in MEFs led to an increased sensitivity to etoposide (a DNA-damaging agent that inhibits DNA topoisomerase II), γ-irradiation, and H2O2 compared with wild-type cells (Fig 1A and 1B and S1 Fig). To ascertain whether Sam68 is essential for the completion of DNA repair, we performed single-cell gel electrophoresis-based alkaline comet assays, a sensitive method for detecting DNA strand breaks [57]. No comet tails were observed in mock-irradiated Sam68-/- and Sam68+/- primary thymocytes, suggesting Sam68 deletion does not spontaneously cause DNA damage (Fig 1C and 1D). The vast majority of γ-irradiated thymocytes, in the presence or absence of Sam68, showed prominent comet tail moments, an indicator of DNA damage severity, at 15 min post γ-irradiation (Fig 1C and 1D). The comet tails lessened in a time-dependent manner in Sam68+/- thymocytes, and almost no comet tails were detected at 3 h post γ-irradiation. Strikingly, the comet tails remained prominent in Sam68-/- thymocytes during the same time period (Fig 1C and 1D), which strongly supports an indispensable role of Sam68 in repairing DNA breaks.
DNA double-strand breaks (DSBs) are the most severe form of damage to DNA, and homology-directed repair (HDR) and NHEJ have been proposed as the major mechanisms used to repair DSBs [58]. We sought to directly test whether Sam68 facilitates DNA repair through one or more such specific signaling pathways. To this end, we utilized U2OS cell lines that contain chromosomally integrated green fluorescent protein (GFP) reporters with recognition sites for the rare-cutting endonuclease I-SceI to assess the rates of HDR and NHEJ, as GFP positivity by flow cytometry analysis suggests that repair has occurred in these cells [58]. Upon down-regulation of Sam68 by small interference RNA (siRNA), we observed the repair efficiency of the NHEJ and HDR pathways was reduced to 42.3% and 12.2%, respectively, in comparison to the nonspecific control siRNA (Fig 1E and 1F). The efficiency of Sam68 silencing in the U2OS reporter cell lines was verified by IB (Fig 1F), and Sam68 down-regulation exhibited no substantial effect on transfection efficiency of these cells (Fig 1E). Moreover, ectopic expression of a Sam68 construct that is resistant to siRNA markedly rescued the repair efficiency of the NHEJ and HDR pathways in the Sam68 knockdown cells (S2 Fig). Together, these results suggest that Sam68 deletion causes defective DNA repair, thus resulting in persistent DSBs and hypersensitivity to DNA-damaging agents.
As Sam68 knockdown has a more profound impact on the efficiency of HDR (Fig 1E and 1F), which is the most important error-free pathway for repairing DSBs, we sought to examine how Sam68 deficiency affects the signaling cascade in response to DSBs caused by γ-irradiation. Phosphorylation of the histone variant H2AX (γH2AX), a DNA damage marker that promotes the recruitment of chromatin-modifying complexes and downstream repair factors [59–61], was robustly increased in Sam68-sufficient MEFs, primary thymocytes, and U2OS cells treated with γ-irradiation (Fig 2A–2C and S3 Fig). In contrast, such DNA damage-triggered γH2AX accumulation was severely attenuated in Sam68-deficient cells (Fig 2A–2C and S3 Fig). Moreover, the γ-irradiation-induced phosphorylation of ATM (which is the kinase of H2AX) and its substrates Chk1 and Chk2, all of which are essential DNA damage signaling transducers [1], was consistently more robust in Sam68-sufficient MEFs, thymocytes, and U2OS cells in comparison to Sam68-deficient cells (Fig 2A and 2D and S3 Fig). These observations further suggest that Sam68 deficiency leads to decreased DNA repair signaling. As DSB-induced ATM phosphorylation is known to rely on the activation of the DNA damage sensor PARP1 and subsequent PAR production [14,27], we examined PAR chain formation in γ-irradiated cells in the presence and absence of Sam68. As expected, PAR chains were rapidly and vigorously built up in Sam68-sufficient thymocytes and MEFs following γ-irradiation (Fig 2C, 2E and 2F). However, while Sam68 deficiency did not affect PARP1 levels (Fig 2E and 2F), γ-irradiation triggered PAR production was diminished in Sam68-deficient cells (Fig 2C, 2E and 2F), which demonstrates that Sam68 is crucial for DNA damage-initiated PAR chain formation. Moreover, supplementing with exogenous Sam68 markedly restored the DNA damage-initiated PAR synthesis and the PAR-dependent phosphorylation of ATM, Chk1, Chk2, and H2AX in Sam68 knockout (KO) MEFs following γ-irradiation (Fig 2A and 2E), highlighting the crucial role of Sam68 in the DSB-triggered PAR production and PAR-dependent signaling to DNA repair.
Given that Sam68 deletion almost abolished PAR production following γ-irradiation (Fig 2E and 2F), we first performed immunoprecipitation assays to examine whether Sam68 interacts with PARP family proteins in DDR, thus affecting the DNA damage-initiated PAR formation. Indeed, at 5 min post γ-irradiation, Sam68 substantially associated with PARP1 and ATM (Fig 3A and S4 Fig); in contrast, no detectable interaction between Sam68 and PARP2, PARP3, and PARP5a/b was observed under the same conditions (S4B Fig). Given the observed interaction and that PARP1 is the primary nuclear enzyme that transfers PAR chains to various target proteins in response to DNA damage [2], we further characterized the Sam68-PARP1 interaction following DNA damage. The γ-irradiation-triggered Sam68-PARP1 association was almost identical in the cells pretreated with the PARP inhibitor PJ-34 compared to control (Fig 3B), which suggests that PAR chain formation is dispensable for DNA damage-induced Sam68-PARP1 interaction. In contrast, the Sam68-PARP1 interaction in response to γ-irradiation was greatly reduced in the presence of the DNA-binding agent, ethidium bromide (Fig 3C), demonstrating that damaged DNA is critical for the DNA damage-induced Sam68-PARP1 interaction.
The fact that PARP1 is rapidly recruited to DNA damage sites [62] and damaged DNA enhances the Sam68-PARP1 interaction in DDR (Fig 3A and 3C) led us to assess whether Sam68 can be recruited to DNA lesions where it interacts with PARP1 and regulates PAR formation. As illustrated by chromatin fractionation assays, Sam68 was remarkably enriched in chromatin fractions in MEFs and U2OS cells following γ-irradiation (Fig 3D and S5 Fig). To ascertain whether the damaged chromatin-enriched Sam68 is recruited to DNA lesions rather than generally bound to chromatin, we performed chromatin immunoprecipitation (ChIP) assays to monitor the DNA damage-triggered Sam68 recruitment to the proximity of a DSB that is actively generated at a unique chromosomally integrated I-SceI site in HDR-GFP reporter U2OS cells (Fig 3E). While no association of Sam68 or of PARP1 was detected with chromatin near the I-SceI site in U2OS cells without I-SceI transfection, Sam68 and PARP1 were indeed enriched at the site near the generated DSB in the I-SceI-expressing U2OS reporter cells (Fig 3E). In contrast, we did not detect an enrichment of Sam68 or PARP1 at the site far away from the cut I-SceI site (Fig 3E), which suggests that Sam68 and PARP1 are specifically recruited to the DNA damage site. Moreover, we carried out laser microirradiation microscopy assays to visualize the recruitment of Sam68 to local DNA strand breaks, using MEFs expressing red fluorescent protein (RFP)-tagged Sam68 as well as GFP-tagged PARP1. Consistent with the previous studies [2], accumulation of ectopically expressed GFP-PARP1 at DNA lesions was manifested as cytological discernable foci in laser-irradiated Sam68 KO MEFs (Fig 3F and S6A Fig). Interestingly, RFP-Sam68 formed discrete, cytologically detectable foci, which significantly overlap with the damage foci formed by GFP-PARP1 and endogenous γH2AX, after laser microirradiation in parallel samples all collected at the same time post micro-irradiation (Fig 3F and S6A Fig). In contrast, the RFP control failed to form discrete foci, although GFP-PARP1 formed damage foci as expected, under the same condition, in the cells expressing RFP and GFP-PARP1 (Fig 3F). Of note, endogenous Sam68 and PARP1 also formed discrete DNA damage foci that significantly overlapped with those manifested by γH2AX, after laser microirradiation (S6B Fig), further supporting our assertion that Sam68 localizes at sites of damaged DNA significantly overlapping with PARP1 as well as γH2AX during the cellular responses to DNA damage. Moreover, the GFP-PARP1 fluorescence intensities after laser microirradiation were comparable in wild-type and Sam68-deleted cells (S6C Fig), indicating that Sam68 is not required for PARP1 localization to sites of damaged DNA. Similarly, PARP1 inhibition with Olaparib, compared to the vehicle control, did not affect the recruitment of GFP-PARP1 at sites of damaged DNA (S6D Fig). In contrast, Olaparib treatment substantially augmented, rather than attenuated, the localization of RFP-Sam68 at DNA damage foci in the MEFs expressing RFP-Sam68 and GFP-PARP1 (S6E Fig). These results suggest that Sam68 and PARP1 both localize to sites of damaged DNA independently. To assess the impact of Sam68 on the PAR formation at local DNA damage sites, we carried out immunofluorescence staining assays for PAR production at local DNA strand breaks generated by laser microirradiation in the nucleus in the presence and absence of Sam68. As expected, shortly (1 min) after laser microirradiation, vigorous PAR production was detected locally in the nuclear regions where laser beams were introduced in WT MEFs (Fig 3G). In striking contrast, PAR formation was barely observed in the laser microirradiated regions in Sam68 KO MEFs (Fig 3G), which suggests that Sam68 is crucial for local PAR synthesis at DNA lesions. Collectively, these results demonstrate that Sam68 is recruited to and significantly overlaps with PARP1 at DNA damage sites, where it regulates local PAR production in the cellular responses to DNA damage.
The findings that Sam68 deficiency attenuates DNA damage-induced global PAR production (Fig 2E and 2F) and focal PAR formation (Fig 3G) and that Sam68 and PARP1 interact and overlap at DNA strand breaks after DNA damage (Fig 3A and 3F) led us to hypothesize that Sam68 governs DNA damage-initiated PARylation via directly stimulating the catalytic activity of PARP1. To test this hypothesis, we performed in vitro PARylation assays using recombinant PARP1 and Sam68 proteins. As expected, damaged DNA-activated PARP1 automodified itself with the addition of PAR moieties from the supplemented nicotinamide adenine dinucleotide (NAD+), as indicated by the robust PAR chain formation (Fig 4A and S7A Fig). Strikingly, incubation of recombinant Sam68 protein, compared to the GST control, with PARP1 dramatically boosted PARP1 activation and PAR production in a dose-dependent manner (Fig 4A, compare lane 3 with lane 7, and S7B Fig) in the presence of damaged DNA and NAD+. Moreover, in the absence of PARP1, Sam68 recombinant protein was not able to produce a PAR chain with supplemented damaged DNA and NAD+ (S7C Fig), indicating that Sam68 per se does not possess the enzymatic activity to transfer ADP-ribosyl polymers. Hence, our results demonstrate that Sam68 simulates PARP1 activation and subsequent PAR production in the presence of damaged DNA and NAD+.
We sought to understand the key domain(s) in Sam68 critical for the interaction with PARP1 and the PARP1-stimulatory function using various Sam68 truncates. We detected the association of the full-length, ΔC, and ΔKH truncated Sam68 to endogenous PARP1, but not GFP vehicle (Fig 4B). In contrast, deletion of the N-terminal amino acids 1–102 (ΔN) of Sam68 almost abolished the association of Sam68 to PARP1 (Fig 4B), suggesting a key role of the N-terminus for the interaction between Sam68 and PARP1. Moreover, we carried out pulldown assays using GST-fused full-length and ΔN truncated Sam68 together with PARP1 recombinant proteins. While full-length Sam68 pulled down PARP1, the ΔN truncated Sam68 failed to do so (Fig 4C), which further confirms the critical role of the N-terminus of Sam68 for the Sam68-PARP1 interaction. To examine whether the Sam68 N-terminus-mediated Sam68-PARP1 interaction is important for the PARP1-stimulatory function, we conducted the in vitro PARylation assays, supplementing PARP1 with full-length or ΔN mutant Sam68 recombinant protein. Indeed, the stimulatory effect of Sam68 for PARP1 activation and PAR production was substantially reduced when Sam68 (ΔN) recombinant protein was utilized, in comparison to full-length Sam68 (Fig 4D, compare lane 11 with lane 15). This result underscores an important role of the Sam68 N-terminus in controlling PARP1 activity in vitro. Moreover, to further examine whether the Sam68-PARP1 interaction is functionally important for DNA damage-initiated PAR production in cells, we compared the γ-irradiation triggered PARylation in Sam68 KO MEFs transiently expressing full-length or ΔN truncated Sam68, both of which share strict nuclear localization (Fig 4E). The expression of full-length Sam68, but not GFP control, significantly restored the DNA damage-initiated PAR synthesis in Sam68 KO MEFs (Fig 4F), consistent with our previous observation (Fig 2E). However, ectopic expression of Sam68 (ΔN) mutant failed to restore the γ-irradiation triggered PAR production in Sam68 KO MEFs (Fig 4F). Thus, our results demonstrate that the N-terminus of Sam68, which is critical for the Sam68-PARP1 interaction, is functionally important for PARP1-catalyzed PAR production.
We then performed structural-functional analyses using full-length and truncated PARP1 proteins to explore the domain(s) on PARP1 essential for Sam68 interaction and Sam68-stimulated PARP1 activation. In contrast to the strong interaction between full-length PARP1 and Sam68, PARP1 (663–1,014) truncated protein, containing the catalytic domain, barely bound Sam68 (Fig 4G), which indicates that the stimulatory function of Sam68 on PARP1 activity may not be conferred via a direct interaction between Sam68 and PARP1 catalytic domain. Conversely, the PARP1 (1–662) truncated protein, which contains the DNA-binding domain and automodification domain, associated with Sam68 to a similar extent as full-length PARP1 (Fig 4G), indicative of a potentially important role of the PARP1 N-terminus for the stimulatory function of Sam68. Indeed, in striking contrast to the robust PAR chain formation by Sam68 and PARP1 in the presence of damaged DNA and NAD+, incubation of Sam68 with either PARP1 (1–662) or PARP1 (663–1,014) failed to form detectable PAR chains (Fig 4H, compare lanes 18, 20, and 22), suggesting that the Sam68-PARP1 interaction and PARP1 catalytic domain are both required for Sam68 to stimulate PARP1 activation. Moreover, in the absence of damaged DNA, incubation of Sam68 with PARP1 exhibited no detectable PARP1 activation (Fig 4I and S7D Fig), which demonstrates that Sam68 primarily stimulates the DNA-dependent PARP1 activation. Altogether, our results suggest that Sam68 stimulates the DNA-dependent PARP1 activation and subsequent PAR production through the interaction between their N-termini.
To recapitulate our observed phenotypes due to diminished PAR synthesis in the absence of Sam68, we performed parallel experiments in PARP1-deficient or -inhibited cells. Indeed, pretreatment of thymocytes with PARP1 inhibitors, Olaparib and PJ-34, significantly impeded γ-irradiation-triggered PAR production and the downstream phosphorylation of ATM and Chk1 (S8 Fig), mirroring the impaired PAR synthesis and DNA repair signaling in Sam68-deficient cells (Fig 2). As a result, PARP1 deletion in MEFs, in comparison to wild-type cells, resulted in increased sensitivity to genotoxicity of etoposide and γ-irradiation (S9A and S9B Fig), which is comparable to Sam68 KO in MEFs (Fig 1A and 1B). Moreover, alkaline comet assays revealed that DNA damage repair was substantially delayed in the PARP1-inhibited thymocytes compared to the control cells, as illustrated by the persistent comet tail moments within 3 h post γ-irradiation (S9C and S9D Fig), further supporting a similar effect of Sam68 deletion and PARP inhibition on DNA repair of γ-irradiated lesions in thymocytes. Furthermore, PARP1 down-regulation by siRNA significantly attenuated the repair efficiency of NHEJ and HDR in the U2OS reporter cell lines, in comparison to the nonspecific control siRNA (S9E Fig), which mirrors the impact of Sam68 silencing on the specific signaling pathways that repair DSBs (Fig 1F). These results suggest that deficiency in either PARP1 or Sam68 similarly causes defective DNA repair, thus resulting in persistent DSBs and hypersensitivity to DNA-damaging agents. To examine whether PARP1 loss further impacts the hypersensitivity of Sam68 KO MEFs to DNA-damaging agents, we down-regulated PARP1 expression by siRNA in Sam68 KO MEFs. Indeed, following exposure to DNA-damaging agents, etoposide and γ-irradiation, the clonogenic survival of Sam68 KO MEFs expressing PARP1-specific siRNA was almost identical to those expressing control siRNA (S10 Fig). These results further highlight the crucial role of Sam68 in controlling the PARP1-catalyzed PAR production in DDR.
In light of the crucial role of Sam68 in mediating DNA repair in cultured cells, we additionally examined the thymi, which is known to be hypersensitive to radiotoxicity [17,63,64], derived from mice at various periods post whole body γ-irradiation (WBIR) to assess the impact of Sam68 on DNA repair signaling in damaged organs in vivo. To assess the immediate effect of Sam68, we harvested the thymi from mock- and γ-irradiated mice at 20 min post WBIR. As visualized by immunohistological staining, the DNA repair signaling in mock-irradiated Sam68+/- and Sam68-/- mice was comparable (Fig 5A and 5C). Shortly after WBIR, vigorous PAR synthesis and phosphorylation of ATM, Chk1, Chk2, and H2AX were detected in the thymi derived from Sam68+/- mice (Fig 5A and 5C). In striking contrast, such a response was greatly diminished in the γ-irradiated Sam68-/- thymi (Fig 5A and 5C). Moreover, these results were further supported by IB of thymocyte lysates derived from γ-irradiated Sam68+/- and Sam68-/- mice (Fig 5B and 5D). Hence, Sam68 is essential for DNA damage-initiated PARylation and repair signaling in radiodamaged thymus.
To determine the importance of Sam68 in DDR in vivo, we assessed the impact of Sam68 deletion on γ-irradiation-caused genotoxicity in mice. Sam68+/- male mice subjected to a sublethal dose of WBIR initially showed a modest loss in body weight and consistently regained their weight by 18 d post WBIR, correlating with 100% survival from radiotoxicity (Fig 6A). In striking contrast, more severe weight loss and mortality were observed in Sam68-/- mice, with a survival rate of 17% over a period of 45 d post WBIR (Fig 6A). Likewise, Sam68-/- female mice exhibited hypersensitivity to radiotoxicity compared with their Sam68 sufficient littermates, reflected in a sharp reduction in both body weight and survival rate (Fig 6B). Moreover, we monitored mouse mortality following intraperitoneal administration of N-methyl-N-nitrosourea (MNU), a DNA alkylating agent that causes genotoxicity [17,65]. Over a period of 14 d, Sam68-/- mice had higher and accelerated mortality compared to Sam68+/- controls (Fig 6C). Together, these results demonstrate a crucial role of Sam68 in protecting mice from genotoxic challenges by γ-irradiation and alkylating chemicals.
We further compared the morphology of the thymus and small intestine in mock- and γ-irradiated animals by gross dissection and histological staining. Of note, the thymic size, structure, and lymphocyte subpopulations of mock-irradiated Sam68+/- and Sam68-/- mice were comparable, showing that Sam68 deficiency does not impair thymus development (Fig 6D and 6E and S11A Fig). Fourteen d post WBIR, the thymi in Sam68+/- mice were indistinguishable to those in mock-irradiated mice in size and morphology (Fig 6D), and all exhibited a well-delineated cortex and medulla zones by histological staining (Fig 6E), indicative of successful DNA repair in the thymi. In contrast, the thymi in γ-irradiated Sam68-/- mice were severely damaged, with reduced size and abolished cortex–medulla conjunctions, compared with those from γ-irradiated Sam68+/- or mock-irradiated mice (Fig 6D and 6E). Similarly, the small intestine develops normally regardless of Sam68 sufficiency in mice (Fig 6F and S11B Fig). However, Sam68-/- mice subjected to WBIR had more widespread damage to the shortened small intestine, particularly to the structure of the villi and crypts in the duodenum (Fig 6F and S11C Fig). The severe damage in the thymus and intestine in Sam68-/- mice strongly supports that Sam68 is essential for radioprotection in mice.
Herein, we report that Sam68 is a previously unappreciated early signaling molecule in the cellular response to DNA damage. Our data demonstrate that the deletion of Sam68 attenuates DNA damage-triggered PARP1 activation, PAR production, and PAR-dependent DNA repair signaling cascades, in line with our observations of an indispensable role for Sam68 in the HDR and NHEJ repair signaling pathways. Moreover, we show that Sam68 KO mice are hypersensitive to genotoxic challenges by DNA-damaging agents. Mechanistically, upon DNA damage, Sam68 is recruited to DNA damage sites, where Sam68 stimulates the catalytic activity of PARP1 at DNA lesions (Fig 7).
The evidence that Sam68 interacts with PARP1 and is recruited to and substantially overlaps with PARP1 at sites of damaged DNA and that Sam68 deletion diminishes the DSB-initiated PAR production suggests that Sam68 regulates PARP1 activity in DDR. The similarity in the phenotypes of Sam68- and PARP1-deficient/inhibited cells and animals in response to DNA damage further supports this notion. First, dramatically attenuated PAR production in response to DNA damage in Sam68-deficient cells (Fig 2E and 2F) mirrors the abolished DNA damage-triggered PAR synthesis caused by PARP inhibitors (S8 Fig). The role of Sam68 in controlling PARP1 activity is further supported by the evidence that recombinant Sam68 is sufficient to boost PARP1 activity in the presence of damaged DNA (Fig 4A). Secondly, the PAR-dependent DNA repair signaling cascade is dampened in PARP1 KO cells [14,27], PARP1-inhibited cells (S8 Fig), and Sam68-deficient cells (Fig 2A–2D and S3 Fig). Moreover, the retarded repair of DNA strand breaks, as illustrated by comet assays, is similarly observed in PARP1-deficient [66] or -inhibited cells (S9C and S9D Fig) and Sam68-deficient cells (Fig 1C and 1D), in line with the PARP1 malfunction-impeded DNA repair signaling and recruitment of the repair machinery. Hence, Sam68-deleted cells are hypersensitive to genotoxic stress (Fig 1A and 1B), similar to what has previously been reported in cells with repressed PARP1 activity through the use of chemical inhibitors or transdominant mutations [67] and in PARP1-deleted cells (S9A and S9B Fig). Furthermore, owing to its key function in DDR, PARP1 is known to be required for protecting mice from genotoxicity. PARP1 KO mice were reported previously to be hypersensitive to the DNA alkylating agent MNU and γ-irradiation [17,65], and γ-irradiation has been shown to cause severe acute damage to the small intestine in PARP1 KO mice [17]. Interestingly, Sam68 KO mice are also hypersensitive to MNU and γ-irradiation challenges and exhibit acute radiodamage in the small intestine and the thymus (Fig 6), thus exactly recapitulating the phenotypes in PARP1 KO mice.
One hallmark of DDR factors is that they are rapidly recruited to the proximity of sites of damaged DNA, and ideally, such enrichment can be visualized by laser microirradiation microscopy [1,68–70]. It has been widely acknowledged that activated PARP1 PARylates itself with such automodification stimulating more PARP1 to be recruited to DNA damage sites (Fig 3F) [2]. Notably, in our laser microirradiation microscopy assays, the accumulation of ectopically expressed and endogenous Sam68 also leads to the formation of discrete, cytologically detectable foci at sites of damaged DNA that significantly overlap with the damage foci formed by GFP-PARP1 or endogenous γH2AX (Fig 3F and S6A and S6B Fig). In addition, we demonstrate that Sam68 and PARP1 are specifically enriched on the chromatin close to the unique DSB at the I-SceI site (Fig 3E). Moreover, Sam68 and PARP1 are capable of physically interacting with each other (Fig 4C), and DNA damage-enhanced interactions between endogenous Sam68 and key initial components of DDR, i.e., PARP1 and ATM (Fig 3A and S4 Fig). Furthermore, recombinant Sam68 is sufficient to facilitate the DNA-dependent PARP1 activity in vitro (Fig 4A and S7D Fig), and the N-terminus of Sam68 is both critical for the Sam68-PARP1 interaction and functionally important for PARP1-catalyzed PAR production (Fig 4B–4D). Together, these results suggest that Sam68, functioning as a key initial signaling molecule, is crucial in regulating PARP1 activity and the PAR-dependent signaling cascade in DDR. It is noteworthy that our work here also elucidates a novel role for the less-characterized N-terminus of Sam68 in DDR signaling. Albeit not appearing to harbor any well-defined functional domains or motifs, the N-terminus of Sam68 does contain a cluster of several serine and threonine residues. This cluster could make Sam68 a potential target of serine/threonine phosphorylation, thus potentially serving as an important functional switch that allows Sam68 to interact with PARP1 and regulate PARP1-catalyzed PARylation in DDR, though further studies are needed to confirm this hypothesis.
Recent studies underscore the emerging role of RNA-binding proteins in DDR [52–54]. Heterogeneous nuclear ribonucleoprotein U-like proteins 1 and 2 were shown to promote recruitment of the Bloom Syndrome protein (BLM) helicase to DSBs by functioning downstream of Mre11-Rad50-Nbs1 (MRN) complex and carboxyl-terminal binding protein (CtBP)-interacting protein [53]. Moreover, RNA-binding proteins RNA binding motif protein, X-linked (RBMX) and non-POU domain containing, octamer-binding (NONO) were reported to be recruited to DNA damage sites in a PARP1/PAR-dependent manner, where RBMX regulates BRCA2 expression [54] and NONO stimulates NHEJ and represses HR repair signaling pathways [52], respectively. It is noteworthy that all these RNA-binding proteins have been proposed to function after PARP1 activation/PAR synthesis in DDR, as their enrichment on DNA lesions relies on the interactions with PAR chains [52]. In contrast, here we propose that the RNA-binding protein Sam68 is a signaling molecule that functions prior to PARP1 activation and in fact controls PARP1 activity in DDR (Fig 7). The enrichment of Sam68 on DNA lesions is apparently as rapid as PARP1, and it stimulates and controls rather than depends on PARP1 activation and PAR production. Thus, the identification of Sam68 as an early signaling factor proximal to PARP1 in DDR provides new insight into the sophisticated signaling mechanisms in DDR, in particular the stimulation and regulation of PARP1 activation and PAR production.
In light of their crucial roles in various DNA damage repair-signaling pathways, the inhibition of PARP1 as well as other PARP family proteins has emerged as a promising therapeutic approach for treating multiple human diseases associated with impaired DNA repair activities [34,35]. However, the current classes of PARP inhibitors undergoing clinical trials and the FDA-approved Olaparib are all based on a competitive binding strategy first observed with nicotinamide [36], and almost all PARP inhibitors are derivatives of the natural PARP inhibitor nicotinamide [10]. There is an urgent need for improving their specificity for each individual PARP family member and lowering their off-target effect and toxicity [2]. That said, the nature of PARP1’s rapid recruitment (within seconds) to DNA damage sites upon sensing lesions combined with the fact that activated PARP1 vigorously catalyzes the PARylation reaction makes it difficult to elucidate the mechanism of PARP1 activation in DDR. Nevertheless, PARP activation and PAR production need to be elegantly controlled during the cellular response to DNA damage—a sufficient amount of PAR at DNA damage sites is required to recruit downstream effectors to DNA lesions to fulfill their repair function. Meanwhile, excessive PAR chain buildup needs to be avoided, as hyperactivated PARP1 could exhaust intracellular pools of NAD+ and sabotage ATP production, thus resulting in cell death [71–74]. Although post-translational modifications and interactions with other proteins have been proposed to fine-tune PARP1 activity in DDR [32,75], the stimulatory mechanism of PARP1 activation at DNA damage sites has remained largely unknown. Our proposed model that Sam68, as a previously unrecognized stimulatory factor beyond DNA strand breaks, stimulates PARP1 activation and PAR production at DNA damage sites could provide a novel strategy to develop a new category of PARP1 inhibitors and therapeutics for human diseases involving DNA repair.
All animal experiments were performed according to protocol number MO13-H349, approved by the Johns Hopkins University’s Animal Care and Use Committee and in direct accordance with the NIH guidelines for housing and care of laboratory animals. Sam68-/− (Sam68 KO) mice and their gender-matched littermate Sam68+/− heterozygous mice (occasionally substituted with gender-matched littermate Sam68+/+ mice when Sam68+/− ones were lacking but referred to as Sam68+/− alone for simplicity) were produced using heterozygous breeding pairs and were genotyped for disrupted or wild-type Sam68 gene, as previously described [39]. Mice were maintained in a specific pathogen-free facility and fed autoclaved food and water ad libitum.
Wild-type, Sam68 KO, and PARP1 KO MEFs were kindly shared by Drs. Stephan Richard (McGill University, Canada) and Zhao-Qi Wang (Fritz Lipmann Institute, Germany). Wild-type U2OS cells and the DR-GFP and EJ5-GFP reporter U2OS cells [58] were kindly provided by Drs. Michael Matunis (Johns Hopkins University) and Jeremy Stark (Beckman Research Institute of City of Hope). MEF and U2OS cells were cultured in DMEM medium containing 10% fetal calf serum, 2 μM glutamine, and 100 U/ml each of penicillin and streptomycin, except for the addition of 10 mM HEPES (pH 7.2–7.5) for U2OS culture. U2OS reporter cells were cultured using similar medium as for U2OS cells, except without sodium pyruvate. Antibodies used were as follows: Sam68, GST, PARP2, PARP3, and PARP5a/b from Santa Cruz Biotechnology (Dallas, Texas); β-actin and myc from Sigma-Aldrich (St. Louis, Missouri); ATM, PARP1, Caspase-3, Chk1, Chk2, p-Chk1 (Ser345), p-Histone H3 (Ser10), and β-Catenin from Cell Signaling Technology (Danvers, Massachusetts); PAR from Trevigen (Gaithersburg, Maryland); γH2AX from Millipore (Billerica, Massachusetts); H2AX from Bethyl Laboratories (Montgomery, Texas); p-ATM from Rockland (Gilbertsville, Pennsylvania); p-Chk2 from Novus Biologicals (Littleton, Colorado); RFP from GenScript (Piscataway, New Jersey); Sam68 from GeneTex (Irvine, California); Histone3 from Abcam (Cambridge, Massachusetts); CD4 and CD8α from BioLegend (San Diego, California). MNU, etoposide (VP16), 4′, 6-diamidino-2-phenylindole (DAPI), and ethidium bromide (EtBr) were obtained from Sigma-Aldrich. 4-[(3-[(4-cyclopropylcarbonyl)piperazin-4-yl]carbonyl)-4-fluorophenyl]methyl(2H)phthalazin-1-one (Olaparib) and N-(6-oxo-5,6-dihydrophenanthridin-2-yl)-N, N-dimethylacetamide-HCl (PJ-34) were purchased from Fisher Scientific (Pittsburgh, Pennsylvania) and Enzo Life Sciences (Farmingdale, New York), respectively. The GFP-PARP1, RFP-Sam68, and I-SceI plasmids were kindly shared by Drs. Anthony Leung (Johns Hopkins University), Johnny He (University of North Texas Health Science Center), and Jeremy Stark (Beckman Research Institute of City of Hope), respectively. The GFP, GFP-Sam68, GFP-Sam68 (ΔC), GFP-Sam68 (ΔN), GFP-Sam68 (ΔKH), GST, GST-Sam68, and GST-Sam68 (ΔN) constructs were described previously [51].
The γ-irradiation on MEFs, U2OS cells, and primary thymocytes was performed using a 137Caesium source (dose rate 4 Gy/min). WBIR in mice was performed as previously described [17,65]. Briefly, 6–8-wk-old Sam68+/- and Sam68-/- mice were subjected to a single dose of sublethal γ-irradiation from an MSD Nordion Gammacell 40 Exactor, with a dual 137Caesium source (dose rate 1 Gy/min). The body weight, mortality, and survival of mice were monitored post irradiation, and in some circumstances, the γ-irradiated mice were sacrificed at indicated time points post WBIR for histological and immunohistological analyses.
The in vitro proliferation assays were performed as previously described [76]. Briefly, 1 × 103 MEF cells were γ-irradiated at the indicated dose and were seeded in the wells of a 6-well plate immediately after γ-irradiation. In certain cases, 1 × 103 MEF cells seeded in DMEM medium were treated with or without indicated concentrations of etoposide for 20 h, followed by extensive washes with phosphate-buffered saline (PBS). After incubation for an additional 96 h, the surviving cells were accounted using a Z1 Coulter Particle Counter (Beckman Coulter, Indianapolis, Indiana), and the survival fractions were calculated by comparing the live cell numbers in treated cultures to those in untreated controls.
Comet assays were conducted by using the Comet Assay Kit (Trevigen) following the manufacturer’s instructions. Briefly, isolated primary mouse thymocytes were mock-treated or γ-irradiated with indicated doses and then allowed to recover in normal DMEM culture medium containing 10 mM HEPES (pH 7.2–7.5) for indicated periods at 37°C. Cells were collected and washed once with PBS, and 3 × 105 cells were combined with 1% molten LMAgarose at 37°C at a ratio of 1:10 (v/v) and immediately pipetted onto slides. Slides were then immersed in prechilled lysis solution for 1 h on ice to lyse cells, followed by alkaline unwinding of chromatin. Alkaline electrophoresis of gelled slides was performed using an Econo-Sub Horizontal System (C.B.S. Scientific, Del Mar, California) at 24V (0.7 V/cm) at 4°C for 30 min. The DNA was visualized by SYBR Green staining, and images were taken under an Axio Observer fluorescence microscope (Zeiss, Oberkochen, Germany) and analyzed by CometScore software (TriTek, Sumerduck, Virginia).
The siRNAs targeting human Sam68 were described previously [51]. Human and mouse PARP1 siRNAs were purchased from Santa Cruz Biotechnology. Transient transfection of siRNA or plasmids cells was performed using Lipofectamine RNAiMax and Lipofectamine 2000 (Life Technologies, Frederick, Maryland), respectively, according to the manufacturer's instructions.
DSB repair assays were conducted as previously described [58]. In brief, 6 × 105 U2OS cells (DR-GFP/homologous-directed repair and EJ5-GFP/NHEJ) were transfected with nonspecific control or Sam68-specific siRNAs. Forty-four hours later, cells were transfected again with siRNAs together with I-SceI (to generate a DSB at the unique I-SceI site) or GFP (to indicate transfection efficiency) plasmids. Seventy-two hours later, cells were collected and subjected to flow cytometry analysis for GFP positive cells.
For flow cytometry, 0.5–2 × 106 cells were washed twice with PBS, resuspended in staining buffer (1% fetal bovine serum in PBS), and stained with appropriate antibodies for cell surface markers on ice for 30 min. Following staining and extensive washes with staining buffer, cells were analyzed on a FACSCalibur (BD Biosciences, San Jose, California). Events were collected and analyzed with the FlowJo software (Tree Star, Ashland, Oregon).
Freshly excised thymi from mice were gently teased with a syringe and forceps in DMEM containing 10 mM HEPES (pH 7.2–7.5). The mechanically disrupted cell clumps were poured through 70 μm nylon mesh cell strainers (BD Falcon, Bedford, Massachusetts) to remove connective tissue and prepare single cell suspensions. The cell suspensions were washed once with PBS, and the red blood cells were lysed with Ammonium-Chloride-Potassium (ACK) buffer. The remaining thymocytes were counted, washed twice in PBS, and recovered in DMEM containing 10% fetal calf serum, 2 μM glutamine, 100 U/ml each of penicillin and streptomycin, and 10 mM HEPES (pH 7.2–7.5) for 1 h, followed by further treatments as indicated.
Immunoprecipitation and immunoblot assays were conducted as previously described [51]. In brief, the cells were harvested and lysed on ice by 0.4 ml of lysis buffer (50 mM Tris-HCl [pH 8.0], 150 mM NaCl, 1% NP-40 and 0.5% sodium deoxycholate, 1 × complete protease inhibitor cocktail [Roche Applied Science, Indianapolis, Indiana]) for 30 min. The lysates were centrifuged at 10,000 × g at 4°C for 10 min. The protein-normalized lysates were subjected to immunoprecipitation by adding 10 mg/ml of the appropriate antibody, 30 μl of protein G-agarose (Roche Applied Science), and rotating for more than 2 h at 4°C. The precipitates were washed at least four times with cold lysis buffer followed by a separation by SDS-PAGE under reduced and denaturing conditions. The resolved protein bands were transferred onto nitrocellulose membranes, probed as described previously [77,78], developed by the Super Signaling system (Thermo Scientific, Waltham, Massachusetts) according to the manufacturer's instructions, and imaged using a FluorChem E System (Protein Simple, Santa Clara, California).
Immunofluorescence microscopy was performed as previously described [51,79]. Briefly, cells were fixed with 4% paraformaldehyde in PBS and then Cellspin mounted onto slides. After a 5-min permeabilization with 0.05% Triton X-100 in PBS and a 30-min blocking with 5% goat serum, the fixed cells were stained with the appropriate primary antibodies for 1 h and with fluorescence dye-conjugated second antibodies (Life Technologies) for 1 h together with 1 μg/ml of DAPI (Sigma-Aldrich) for 5 min at 25°C. The slides were then rinsed with PBS three times and cover mounted for fluorescence microscopy.
Cells were harvested at indicated time points after γ-irradiation, and cell pellets were resuspended in the NETN buffer (20mM Tris–HCl [pH 8.0], 100 mM NaCl, 1mM EDTA, and 0.5% NP-40) and incubated on ice for 20 min. Supernatant after 3,000 × g for 10 min was collected as soluble fraction. Pellets were recovered and resuspended in 0.2 M HCl on ice for 30 min and sonicated for 10 s to release chromatin-bound proteins; then, the soluble fractions were neutralized with 1 M Tris–HCl (pH 8.5) and collected as chromatin fraction, and the pellets were collected as insoluble fraction for further analysis, as described previously [33,76].
ChIP assays were performed as previously described [79]. Briefly, U2OS DR-GFP cells were transfected with a plasmid encoding I-Scel endonuclease to create a DSB. At 20 h post transfection, 4.5 × 106 cells were fixed by 1% formaldehyde at room temperature for 10 min, and the cross-linking reaction was stopped by 125 mM of glycine. Cells were homogenized in cell lysis buffer (5 mM PIPES, 85 mM KCl, 0.5% NP-40, 1 × protease inhibitor cocktail [Roche]), and the nuclei were resuspended in nuclei lysis buffer (50 mM Tris-HCl [pH 8.1], 10 mM EDTA, 1% SDS, 1 × protease inhibitor cocktail) on ice for 10 min and sonicated by Bioruptor UCD-200 (Life Technologies). The extract was then clarified by centrifugation and diluted 10-fold with dilution buffer (16.7 mM Tris [pH 8.1], 167 mM NaCl, 1.2 mM EDTA, 1.1% Triton X-100, 0.01% SDS) to yield the solubilized chromatin. For immunoprecipitation, anti-Sam68, anti-PARP1, or IgG control antibody, and Protein AG magnetic beads (Thermo Scientific) were added to the soluble chromatin, and the mixture was incubated at 4°C overnight. The beads were then washed sequentially with TSE-150 mM NaCl (20 mM Tris [pH 8.1], 2 mM EDTA, 1% Triton X-100, 0.1% SDS, and 150 mM NaCl), TSE-500 mM NaCl, buffer III (10 mM Tris [pH 8.1], 0.25 M LiCl, 1% NP-40, 1% deoxycholate, 1 mM EDTA) and three times with TE (10 mM Tris 8, 1 mM EDTA). The immunoprecipitants were eluted by elution buffer (1% SDS, 50 mM NaHCO3, 20 μg/ml glycogen), and DNA was extracted with phenol-chloroform and resuspended in TE. PCR (30–35 cycles) was performed using MyTaq Red Mix (Bioline, Taunton, Massachusetts) with following primers at or adjacent to the unique I-SceI site: P1F, 5′-GAGCAAGGGCGAGGAGCTGT-3′; P1R, 5′-CCGTAGGTCAGGGTGGTCAC-3′; P2F, 5′-TCTTCTTCAAGGACGACGGCAACT-3′; P2R, 5′-TGCCGTTCTTCTGCTTGTC-3′; P3F, 5′-CCGCGACGTCTGTCGAGAAG-3′; and P3R, 5′-GCCGATGCAAAGTGCCGATA-3′.
The GST pulldown assays were performed as previously described [79]. Briefly, 1 μg of GST/Ni2+ affinity-purified recombinant protein, as indicated, was applied to glutathione Sepharose 4B resins (Amersham Pharmacia) and incubated for 1 h at 4°C. After washing, bound proteins were eluted and subjected to SDS-PAGE, followed by immunoblotting.
Laser microirradiation microscopy assays were performed as previously described [62,80], with some modifications. Briefly, near infrared (NIR) excitation was provided by a Mai Tai HP Ti:Sapphire laser (Spectra Physics, Santa Clara, California) tuned to 800 nm, 125 fs pulses, and an 80 MHz repetition rate. The laser beam was introduced through a Leica DMi8 confocal microscope (Leica Microsystems, Mannheim, Germany). The pulsed NIR beam was focused to a diffraction-limited spot with a 63 × oil-immersion objective (1.4 NA). An electro-optic modulator (EOM) was used to create regions of interest (ROIs) targeted for microirradiation by position-dependent power adjustment of the NIR beam. Prior to every experiment, the NIR beam power at the objective was measured to ensure repeatable laser settings that generated a detectable DDR restricted to ROIs without noticeable cytotoxicity. A coverglass-bottomed dish (MatTek Corporation, Ashland, Massachusetts) plated with cells was placed into a temperature-controlled chamber (37°C, 5% CO2) (Life Imaging Services, Basel, Switzerland) enclosing the entire microscope stand. Immediately after microirradiation, the DDR was captured through time-lapse acquisition of the GFP- or RFP-fused protein recruitment using the FRAP application unit of LAS X acquisition software. All images were processed using the LAS AF Lite software (Leica Microsystems), and we quantified the pre- and post-microirradiation fluorescence intensity of the ROI and an adjacent nuclear area with a similar pre-microirradiation fluorescence intensity, to correct fluorescence intensity for transfection efficiency and photobleaching.
MNU injection-induced genotoxic stress in mice was carried out as previously described [17,65]. Briefly, 6–8-wk-old Sam68+/- and Sam68-/- mice were administrated a single dose of 165 mg/kg body weight of MNU in 200 μl of PBS by intraperitoneal injection. The body weight, mortality, and survival of mice were monitored until 14 d post injection.
After euthanizing mice, the thymus was removed, washed once with ice-cold PBS, immersed in optimal cutting temperature media (Tissue-Tek, Elkhart, Indiana), and frozen in dry ice to preserve the tissue. The duodenum and colon were removed, washed once with ice-cold PBS, fixed in 4% PFA at room temperature for 24 h, and embedded in paraffin. Five-micron sections were cut for all tissues and processed for hematoxylin and eosin (H&E) staining. Stained sections were microphotographed to perform histomorphometric analyses, as previously described [81].
Immunohistology was carried out as previously described [81]. In Brief, after euthanizing mice, the thymus was excised under aseptic conditions and frozen in optimal cutting temperature media (Tissue-Tek). Five-micron frozen sections were cut using a Microm HM 550 Cryostat (Thermo Scientific), collected on coated slides, fixed in 4% paraformaldehyde, washed with PBS, and blocked with appropriate sera in PBS. After incubating with appropriate antibodies, sections were washed and incubated with fluorescence dye-conjugated second antibodies and 1 μg/ml of DAPI (Sigma). Stained sections were washed and mounted under a coverslip using Fluoro-gel with Tris Buffer (Electron Microscopy Sciences, Hatfield, Pennsylvania) and examined using an Axio Observer fluorescence microscope (Zeiss).
For in vitro PARylation assays, PARP1, PARP1 (1–662), PARP1 (663–1,014), GST, GST-Sam68, and GST-Sam68 (ΔN) recombinant proteins were purified as previously described [82] using three chromatographic steps: GST/Ni2+ affinity, heparin-sepharose, and gel filtration. In vitro PARylation assays were performed as previously described [83]. Briefly, the indicated recombinant proteins were incubated for 20 min at 30°C in a standard assay buffer (100 mM Tris-HCl [pH 8.0], 10 mM MgCl2, 10% (v/v) glycerol, and 1.5 mM DTT) in the presence and absence of damaged DNA (sonicated) and NAD+. The reaction was terminated by the addition of SDS sample buffer (Life Technologies), and the boiled samples were subjected to SDS-PAGE. When indicated, the PARP inhibitor PJ-34 was added to the reaction mixture at a final concentration of 1 μM for 15 min prior to the reaction.
All statistical analysis was performed using GraphPad Prism version 6.0 (GraphPad Software, San Diego, California). The differences between treated and control groups were examined by unpaired Student’s t tests, except Gehan-Breslow-Wilcoxon tests were used for Kaplan-Meier survival curves. Standard errors of means (SEMS) were plotted in graphs. ns means nonsignificant difference, and significant differences were considered * at p < 0.05; ** at p < 0.01; *** at p < 0.001; and **** at p < 0.0001.
|
10.1371/journal.pgen.1008091 | Expression estimation and eQTL mapping for HLA genes with a personalized pipeline | The HLA (Human Leukocyte Antigens) genes are well-documented targets of balancing selection, and variation at these loci is associated with many disease phenotypes. Variation in expression levels also influences disease susceptibility and resistance, but little information exists about the regulation and population-level patterns of expression. This results from the difficulty in mapping short reads originated from these highly polymorphic loci, and in accounting for the existence of several paralogues. We developed a computational pipeline to accurately estimate expression for HLA genes based on RNA-seq, improving both locus-level and allele-level estimates. First, reads are aligned to all known HLA sequences in order to infer HLA genotypes, then quantification of expression is carried out using a personalized index. We use simulations to show that expression estimates obtained in this way are not biased due to divergence from the reference genome. We applied our pipeline to the GEUVADIS dataset, and compared the quantifications to those obtained with reference transcriptome. Although the personalized pipeline recovers more reads, we found that using the reference transcriptome produces estimates similar to the personalized pipeline (r ≥ 0.87) with the exception of HLA-DQA1. We describe the impact of the HLA-personalized approach on downstream analyses for nine classical HLA loci (HLA-A, HLA-C, HLA-B, HLA-DRA, HLA-DRB1, HLA-DQA1, HLA-DQB1, HLA-DPA1, HLA-DPB1). Although the influence of the HLA-personalized approach is modest for eQTL mapping, the p-values and the causality of the eQTLs obtained are better than when the reference transcriptome is used. We investigate how the eQTLs we identified explain variation in expression among lineages of HLA alleles. Finally, we discuss possible causes underlying differences between expression estimates obtained using RNA-seq, antibody-based approaches and qPCR.
| The level at which a gene is expressed can have important influence on the phenotype of an organism, including its predisposition to develop diseases. One way to estimate gene expression is by quantifying the abundance of RNA. RNA-seq has become the method of choice to provide such estimates at the genomewide scale. However, the application of RNA-seq to HLA genes —key players in the immune adaptive response— has remained a rarely explored approach. This is due to the problem of mapping bias, which causes deficient read alignment at genes which are very polymorphic and different from the reference genome. This has motivated approaches that replace the single reference genome with personalized sequences, comprised of the individual’s specific HLA genotype. Here we explore the use of computational frameworks to obtain reliable expression levels for HLA genes from RNA-seq datasets. We present a pipeline in which the quantification of HLA expression is carried out using methods which account for HLA diversity, avoiding the biases of standard approaches. We then evaluate the impact of this form of quantifying HLA expression on downstream analyses. The pipeline also allows us to integrate information on eQTLs with expression levels at the HLA allele-level, which can help disentangle different contributions to disease phenotypes and help understand the regulatory architecture at the HLA region.
| The HLA region is the most polymorphic in the genome, and also shows the greatest number of disease associations, which has made it very well characterized at the genomic, population and functional levels [1, 2]. Decades of research have also shown that the HLA genes are targets of natural selection, likely a consequence of their role in responding to pathogens [2, 3]. This combination of evolutionary and biomedical interest has resulted in an extensive catalogue of HLA variation in human populations, with the frequency of HLA alleles defined for various populations [4–7].
The most intensely studied genes in the region are the classical HLA loci. These include Class I genes (HLA-A, HLA-C, and HLA-B), whose proteins present endogenous antigens to CD8+ T cells, and also regulate innate immune responses by interacting with killer cell immunoglobulin-like receptors (KIR) expressed on natural killer (NK) cells, and Class II genes (HLA-DRA, HLA-DRB1, HLA-DQA1, HLA-DQB1, HLA-DPA1, and HLA-DPB1) whose proteins present exogenous antigens to CD4+ T cells [1, 8].
Variation within or near HLA genes has been convincingly linked to resistance and susceptibility to both autoimmune and infectious diseases [2, 9, 10]. Although in many cases the mechanistic basis of the associations remains poorly understood, there has been an effort to identify whether associations can be linked to features such as variation at the level of specific amino-acids, HLA alleles, HLA haplotypes, non-coding variants near HLA genes, or HLA expression levels (reviewed in [2, 10, 11]). In some instances, the association of one feature results from the fact that it tags another feature. Several studies have shown that HLA alleles may have a protective effect because they mark overall gene expression [12–15]. Other studies showed that a combination of two features may be important. For example, mortality of transplant recipients was associated with increased expression of HLA-C alleles which harbor specific amino-acids [16].
An understanding of how HLA expression varies among individuals, and the identification of genetic variants involved in the regulation of expression, will play a central role in understanding the contribution of HLA genes to normal and disease phenotypes. However, until recently, little information existed about the regulatory variation and population-level expression patterns of HLA genes, a result of the difficulty in quantifying expression for genes which show an unusually high polymorphism and are members of a multi-gene family [9, 17].
Efforts have been made to develop antibody-based methods to quantify HLA protein on the cell surface [12–14, 18–20], or hybridization-based approaches to quantify mRNA, such as qPCR [13, 21, 22] and microarray methods [23]. However, the design of PCR primers, microarray probes, or antibodies that span the diversity of possible variants represents a technically challenging and labor-intensive undertaking. In addition, qPCR technologies are not appropriate for comparison of expression levels among different loci, an important concern when seeking to understand how expression of HLA genes responds to environmental challenges.
The results obtained to date using qPCR [13, 21, 22], antibody-based approaches [12–14, 18–20] and customized arrays [23] have contributed to the understanding of HLA expression and its underlying genetic regulation. However, they do not take advantage of the large amount of RNA-seq data generated by studies of whole transcriptomes in large samples [24–26], often involving populations from various regions of the world, which represents an attractive resource to investigate HLA expression.
Although such whole-transcriptome RNA-seq studies do provide expression estimates for HLA genes, they bring new challenges. RNA-seq pipelines may provide biased expression estimates for two reasons: 1) many short reads originating from genes with extreme polymorphism fail to map to the reference genome, due to high degree of variation (which results in a large number of mismatches between the reference genome and that of most individuals being analyzed), and 2) the presence of paralogues makes it difficult to map a read uniquely to a specific gene, leading to the exclusion of many reads. This raises concerns about the reliability of RNA-seq approaches to quantify HLA expression, given that these loci represent both the extreme of polymorphism in the human genome and are part of a multi-gene family [27–29].
A strategy to overcome these challenges is the mapping of reads to an HLA-personalized reference (an index containing sequences of the individual being quantified), rather than to a single reference genome. For example, seq2HLA is a tool developed by Boegel et al. [30] to provide in-silico HLA types and expression estimates, and later applied to demonstrate that different tumor types are associated with different HLA expression levels [31], and also to provide a large catalog of HLA expression in 56 human tissues and cell types [32]. AltHapAlignR [33] is another software which infers the HLA references which are the closest to the individual’s HLA haplotypes, and maps reads to them. The authors reanalyzed a large RNA-seq dataset [24], and provided comparisons with conventional read mapping, showing an improvement in accuracy with the HLA-tailored pipeline.
In this article we present HLApers, a personalized pipeline which we have developed to reliably quantify HLA expression from RNA-seq data. We compare HLApers to conventional pipelines, and discuss for the first time the impact of accurate estimation of HLA expression on downstream analyses such as eQTL (expression Quantitative Trait Loci) mapping and allele-specific expression. We show that it is possible to adapt different computationally efficient methods to work under the personalized reference framework, providing reliable quantification of HLA expression from RNA-seq data. We find that implementations with either a conventional read mapper [34] or a pseudoaligner [35] show similar expression estimates.
We use simulations to assess accuracy, showing that the HLA-personalized pipeline is more accurate than conventional mapping, and apply the tool to reanalyze RNA-seq data from the GEUVADIS Consortium [24], which made available whole-transcriptome RNA-seq data for Lymphoblastoid Cell Lines (LCLs) from 462 European and African individuals. We evaluate the impact of more accurate expression estimates obtained with HLApers on downstream analyses by carrying out a detailed survey for allele-specific expression and eQTL mapping at the classical HLA loci.
Surprisingly, we find that conventional RNA-seq pipelines provide gene-level expression estimates and identify eQTLs which are highly correlated with those obtained under the HLA-personalized approach. However, we identify gains of using a pipeline tailored for HLA expression: higher accuracy, expression estimates at the HLA allele-level, and eQTLs with higher probabilities of being causal.
We developed the HLApers pipeline (for HLA expression with personalized genotypes) to measure HLA expression from whole-transcriptome RNA-seq data. The pipeline can use either (1) a suffix array-based read mapper (STAR [34]) followed by quantification with Salmon [36] (henceforth called STAR-Salmon), or (2) a pseudoaligner with built-in quantification protocol (kallisto [35]). The key feature of our implementation is the use of an index supplemented with a set of sequences covering the breadth of known HLA sequences (see Materials and methods: Index supplemented with the HLA diversity). We implemented a two-step quantification approach, where (1) we align reads to reference sequences corresponding to all known HLA alleles, and identify those which maximize the read counts at each locus to infer the genotype which is present (in-silico genotyping), and (2) we use this inferred HLA genotype to create a personalized index which we use to quantify expression (Fig 1).
A problem for the analysis of HLA is that of multimaps: reads originating from a particular allele which map to other alleles of the same locus, or to other loci. To deal with multimaps different approaches have been used. One possibility is to discard them (as in [33]), and another is to split them evenly among the compatible references (as in [30]). In HLApers, we use maximum likelihood estimates of expression obtained by an expectation-maximization (EM) algorithm (implemented within Salmon [36] and kallisto [35]), which infers the quantities of each HLA allele that maximize the probability of observing the set of sequenced reads.
Because the in-silico typing is an important step for accurate expression estimates, we assessed the concordance between our RNA-seq based HLA typing and the HLA allele calls experimentally determined for 5 HLA loci using Sanger sequencing [6]. The concordance was higher than 97% for all of the HLA genes compared (S1 Table). This is consistent with previous results showing that RNA-seq provides reliable HLA alleles calls [30, 37–40].
Our HLApers pipeline integrates widely used quantification tools such as Salmon and kallisto, and provides both allele and locus level estimates of HLA expression. The pipeline is available at https://github.com/genevol-usp/HLApers.
We next investigated how using the HLA-personalized index affects the quantifications of HLA expression. To this end we simulated an RNA-seq experiment using the Polyester package [41], with a dataset of 50 individuals, with the read length and counts matching those of the observed data (see Materials and methods: Simulation).
We analyzed the simulated dataset with three different methodologies: (1) the two-step approach in HLApers, where we first inferred the personalized HLA genotype and then aligned reads to it; (2) alignment to the reference transcriptome (Gencode release 25; primary assembly); and (3) alignment to the reference genome (GRCh38). In approaches (1) and (2) the alignment is followed by expression quantification using Maximum Likelihood (ML), which provides a statistical framework for dealing with multimap reads, whereas in approach (3) quantification is performed using only uniquely mapped reads.
For each HLA locus and methodology, we assessed the proportion of simulated reads which successfully aligned. Because previous studies for genome sequencing identified a correlation between mapping success and the number of mismatches between the HLA allele an individual carries and the reference genome [28], we analyzed how alignment success behaves as a function of the number of mismatches between each HLA allele and the reference genome (Fig 2).
The use of an HLA-personalized index results in the largest proportion of successfully aligned reads, no matter how different the allele carried by the individual is from the allele in the reference genome. This is expected, since the personalized HLA component guarantees that a sequence close or identical to that originating the read will be present.
When the alignment was performed using the reference transcriptome, there was a marked reduction in the proportion of successfully aligned reads for HLA-DRB1, HLA-DQA1, HLA-DQB1, driven by decreased alignment success for alleles with a greater proportion of mismatches with respect to the reference genome.
When using uniquely mapped reads there was a massive read loss for HLA-A, HLA-B and HLA-DPB1, regardless of the divergence to the reference genome, as well as a lower proportion of successfully aligned reads across other loci. This shows that both discarding multipmaps, as well as not including a personalized index, have a negative impact on mapping success.
Finally, for the least polymorphic HLA loci, mapping should not be sensitive to the specific reference used. This is precisely what we find, with all pipelines performing similarly for HLA-DRA and HLA-DPA1.
Having demonstrated that including an individual’s HLA alleles in the index improves the success of read alignment in the simulated data (Fig 2), we set out to address two questions with real data by applying HLApers to the GEUVADIS dataset [24]. First, we examined how expression varies among HLA loci, when the personalized index is used (Fig 3). Secondly, we compared expression estimates with and without the use of the personalized index, so as to evaluate the impact of its usage on real data (Fig 4).
By summing the estimates for the 2 alleles at each HLA locus, we obtain gene-level expression estimates (Fig 3). We observe that HLA-B is the highest expressed gene overall. Among the Class I genes, HLA-B is followed by HLA-A with similar levels, and by HLA-C which has about 50% of the expression levels of HLA-B. For Class II genes, HLA-DRA is the most highly expressed. Although we observe a general concordance with the original GEUVADIS quantifications [24], there are some notable differences: in the original quantifications, HLA-B is twice as expressed as HLA-A, and HLA-DPA1 is more expressed than HLA-DRB1 (S1 Fig).
We found that the correlation between results using the reference transcriptome or the personalized index was greater than 0.87 for every locus except for HLA-DQA1 (Fig 4). The loci with the lowest correlations between indices (HLA-DRB1, HLA-DQA1 and HLA-DQB1), are also those with the greatest read loss when divergence from the reference allele is high in the simulation (Fig 2).
We then investigated if the quantification tool used in HLApers (STAR-Salmon or kallisto) influences expression estimates. Correlations between STAR-Salmon and kallisto approaches were on average r > 0.99 for read counts, dropping to r = 0.8 for TPM estimates for Class I genes, likely due to different bias correction models (S2 Fig). These results show that the key features influencing alignment success are the use of a personalized index and the statistical treatment of multimaps (as opposed to discarding them), with the specific alignment tool being less influential.
Finally, we evaluated the reproducibility of the HLA expression estimates obtained with HLApers. To this end, we analyzed replicates for 97 European individuals in GEUVADIS. We observed an average correlation of expression estimates between replicates of 0.92 over the 9 classical HLA loci. This shows that HLA quantifications from RNA-seq with HLApers are highly reproducible (S3 Fig).
The key role played by HLA loci in the immune response and their strong and abundant associations with infectious and autoimmune diseases have motivated studies to uncover their regulatory architecture [15, 20, 21, 23, 42–45].
Here we use accurate expression estimates obtained with HLApers, together with genotype data from the 1000 Genomes Project [46], to identify SNPs which are associated with variation in expression levels (eQTLs).
Because multiple SNPs can affect expression, it is interesting to identify independent contributions made by distinct SNPs. This has previously been done by using a best eQTL (i.e., the one with the most extreme p-value) as a covariate in subsequent searches for an additional variant. Here we use a related approach implemented in QTLtools, a collection of tools to perform eQTL analysis [47]. The module QTLtools cis allows for the identification of groups (or “ranks”) of SNPs associated with independent signals. For each rank, we identified the site with the most extreme association (Fig 5).
Transcriptome studies can quantify expression for various biological features: individual SNPs, exons, isoforms, genes. In the case of HLA loci, a natural unit of interest is the HLA allele. HLApers provides expression estimates for individual alleles, since it is the allelic sequences which are included in the index. The immunogenetics literature has shown that many HLA alleles are associated with specific phenotypes of evolutionary and medical importance (reviewed in [2, 10, 11]). Gauging information about the expression levels of alleles can therefore provide an additional layer of information.
In order to explore allele-level estimates, we first grouped alleles in “lineages”, which comprise groups of alleles which are evolutionarily and functionally related, since the large number of alleles would make sample sizes per allele too sparse. Although the expression of individual allelic lineages is highly variable among individuals, there is an overall significant difference in expression among lineages (Welch’s ANOVA p-values ranging from 3.7 × 10−8 for HLA-DPB1 to 6 × 10−51 for HLA-DQA1).
Enhanced coordination of gene expression has been proposed as an advantage for gene clustering, as seen at the HLA region [1, 2]. We found a high correlation of expression both within the group of Class I and Class II genes, and lower levels between Class I and Class II genes, which are more than 1Mb apart (Fig 8A) (but see [22] for a result of no co-expression among Class I genes).
A possible cause for co-expression of genes which are physically close to each other is that regulatory activity is structured in domains (CRDs, for Cis Regulatory Domains) [52]. Such domains comprise contiguous regions along a chromosome, and their existence predicts co-expression along haplotypes for genes associated with the same CRD. In fact, previous studies did find evidence that expression of some genes in the HLA region was a feature associated with haplotype membership [23, 62].
We used our inferences of HLA haplotype structure to investigate if there is an haplotypic effect on coordination of expression among nearby HLA genes. Specifically, we tested the hypothesis that co-expression is stronger between alleles located within the same haplotype than between those on different haplotypes. We did not find a consistently higher correlation of expression for alleles on the same haplotype (correlation within haplotypes being higher in only 7 out of 18 locus pairs surveyed within Class I or Class II, Fig 8A).
This result suggests that correlation of expression among HLA loci in LCLs is a result of factors acting at the gene level, and is not driven predominantly by properties of the haplotype. For example, we observe correlation of expression between Class II genes and CIITA, the Class II Major Histocompatibility Complex Transactivator. This correlation is not driven by proximity (CIITA is on chromosome 16), but rather by a trans regulatory mechanism (Fig 8B).
The contribution of HLA variation to normal and disease phenotypes goes beyond peptide specificity, and includes other factors which influence the strength of immune responses, such as HLA expression levels. However, we are still only starting to understand the regulation of expression of these genes, a consequence of their extreme polymorphism which imposes challenges to the methodologies available to measure mRNA and protein levels.
There has been an increasing effort to use RNA-seq data for in-silico HLA typing ([63] and other methods reviewed in [64]) and to estimate expression levels for HLA genes [30, 33]. In this paper, we present HLApers, an HLA-personalized pipeline that quantifies HLA expression based on RNA-seq. We used these expression estimates to identify eQTLs for HLA genes, to estimate the degree of allelic imbalance in expression, and to investigate how the alleles at eQTLs explain variation in expression among HLA alleles.
When new technologies/methods are developed, it is a good practice to evaluate if the results are in agreement with the current knowledge, which in the case of HLA was mostly derived from qPCR or antibody-based studies. However, in our survey of this literature, we found it difficult to compare our newly obtained expression estimates based on RNA-seq with those from previous studies. There are several reasons underlying this difficulty, which we now discuss.
First, expression estimates based on qPCR and antibody-based approaches are in general not comparable among different HLA loci, since different primers/antibodies are often developed for each locus. Thus, the information for expression differences which we observe among HLA loci are usually not available from studies using these techniques. A few studies developed antibodies with comparable affinities for HLA-A/B/C, allowing comparison of expression between loci. Using such a method, Apps et al. [19] found an at least 12-fold reduction in HLA-C expression with respect to HLA-A and HLA-B as measured by flow cytometry, whereas the difference we documented based on RNA-seq was only twofold. However, the Apps et al. [19] study was restricted to a specific haplotype, and because our results show substantial within locus variation in expression levels, it is difficult to compare studies.
Second, many additional layers of biological factors make a direct comparison across studies challenging. For example, the abundance of total protein, of proteins on cell surface or of mRNA in the cell, represent different molecular phenotypes associated with gene expression, given the existence of various post-transcriptional and post-translational regulatory processes [18, 20, 42, 43, 65, 66]. As a consequence, there are good biological reasons to expect differences when expression is estimated at each of these levels. In addition, it is well known that gene expression varies among tissues and cell types [26, 32], posing an additional challenge in comparisons between our findings and those previously reported in the literature, which are often based on analyses of different cell types or tissues.
Third, regarding mRNA quantification, qPCR and RNA-seq each have their own sources of bias, and the finding that there are differences between studies using these approaches should not, on its own, provide definitive evidence that one or the other is inaccurate. Future studies will need to evaluate the degree to which these two technologies can be compared for quantification of HLA expression.
Nevertheless, given the availability of qPCR based quantifications for HLA expression, as well as the interest in obtaining allele-level estimates of expression, we compared the expression for individual HLA lineages obtained by RNA-seq with those from qPCR from previous studies. We found that RNA-seq and qPCR show high concordance for the relative ordering of lineage-level expression for HLA-B and HLA-C (66% and 100%, respectively), whereas for HLA-A concordance was low, at 28%. Therefore results suggest that for HLA-B and HLA-C, when we examine lineages that are sufficiently different in their expression relative to one another (see S1 Text), the two methods are in high agreement, but further research is required to identify the sources of discordance between RNA-seq and qPCR for HLA-A. It is important to recall that these analyses rely on lineage-level expression estimates obtained by qPCR which are indirect, since they estimate locus-level expression (whereas RNA-seq data analysis with HLApers distinguishes among alleles), adding another source of differences among methods.
Our perspective is that differences across studies call for further investigation, and highlight the potential biological variation that needs to be accounted for when comparing cell types, infection status, methodological approach, and molecule being quantified.
In this study we present HLApers, and use simulations and empirical data analyses to show that it produces estimates for HLA expression from RNA-seq with high precision and accuracy. HLApers incorporates a Maximum-Likelihood estimator to deal with instances of reads mapping to multiple alleles or loci, following a strategy that has been widely used in RNA-seq studies [26, 67]. Our pipeline can use different alignment strategies (e.g. STAR-Salmon [34, 36] or kallisto [35]), but we show that the accuracy of expression depends on the sequences contained in the index, and is less dependent on the program used.
The impact of using an HLA-personalized index on expression estimates varies markedly among loci. We found that for HLA-DQA1 there are large differences between gene-level expression estimates obtained using the reference transcriptome and those obtained using the HLA-personalized index. However, this difference is quite low for Class I genes, and intermediate for Class II loci other than HLA-DQA1 (Figs 2 and 4). However, even for the Class I genes, using the personalized index does result in changes in expression estimates with respect to the reference transcriptome. Therefore, we asked whether these changes have an impact on downstream analyses, such as eQTL mapping.
When we compare the eQTLs identified using either HLApers or the reference transcriptome, 6 out of 8 loci had non-overlapping eQTLs. However, in most cases the same biological signal was being captured (S2 Table), and the p-values (Fig 5) and the causality (S5 Fig) of the eQTLs obtained with HLApers are only modestly better than when the reference transcriptome is used. We show that most of the eQTLs we identified (14/17) share a signal with previously reported regulatory SNPs (either experimentally validated SNPs and eQTLs (S3 Table) or CRD-QTLs (S5 Table)). This indicates that, despite of the improvement in expression estimates that personalized pipelines can generate, larger sample sizes are necessary in order to identify eQTLs with greater probability of being causal, and to identify novel eQTLs with smaller effects. We note that the use of transformed cell lines (such as LCLs) may lead to expression profiles specific to this environment, and as a consequence some of the regulatory architecture underlying the expression of the HLA genes may not be shared with other tissues, cell types or treatments.
The HLA-personalized approach provides expression estimates at the HLA allele level, which is not a product of standard RNA-seq pipelines. We integrate allele-level information with the eQTLs mapped for the genes, showing that the HLA allele is a relevant layer of information to understand the regulation of gene expression, because in some instances the regulatory architecture is linked to specific HLA alleles. This joint mapping of regulatory variants and assessment of expression of HLA alleles can illuminate the understanding of the HLA regulation, and contribute to disentangle specific contributions to disease phenotypes.
All data were obtained from third party sources and no additional ethical approval was required.
In order to create the index, we downloaded 16,187 nucleotide sequences for 22 HLA loci (HLA-A, HLA-B, HLA-C, HLA-E, HLA-F, HLA-G, HLA-DMA, HLA-DMB, HLA-DOA, HLA-DOB, HLA-DPA1, HLA-DPB1, HLA-DQA1, HLA-DQB1, HLA-DRA, HLA-DRB1, HLA-DRB2, HLA-DRB3, HLA-DRB4, HLA-DRB5, HLA-DRB7, HLA-DRB8) from the International Immunogenetics/HLA database (IMGT) (release 3.31.0 available at https://github.com/ANHIG/IMGTHLA).
For many alleles, sequence data is not available for the entire coding region (e.g., only for exons 2 and 3 for class I, and exon 2 for class II genes, which are called ARS exons). Because the lack of sequences for much of the coding region would cause the exclusion of many reads, including those mapping to the boundaries of the available exons, for each allele with partial sequence we used the available sequence to find the closest allele which has the complete sequence, and attributed the sequence from this allele. This is expected to introduce little bias in either the genotyping step (because ARS exons are the most polymorphic and sufficient to distinguish specific alleles) or the expression estimation (because the non-ARS sequence attributed is likely very similar to the real one).
For the final index file, we replaced the HLA transcripts in the reference transcriptome (Gencode v25, primary assembly) with the HLA diversity described above. STAR’s module genomeGenerate, Salmon’s index, and kallisto’s index compile an index from these sequences.
In order to select the alleles to be used in the HLA-personalized index, we observed that a simple procedure of selecting the 2 alleles with the largest number of estimated read counts after applying a zygosity threshold is sufficient to produce calls with accuracy of ≥ 95%. However, in order to avoid false homozygotes and false heterozygotes, we implemented additional steps.
First, we selected the top 5 alleles and applied an intra-lineage threshold of 0.25, meaning that only alleles which had at least 25% of the total expression in their lineage were considered for further steps. For each individual, we compiled an index containing only these (up to) 5 alleles and estimated their expression. We then determined if the individual was heterozygote at the lineage level by applying a threshold of 0.15 on lineage expression levels. The lead allele from each lineage was selected to compose the genotype. A zygosity threshold of 0.15 was applied to decide whether the genotype was heterozygous at the allele-level. For each locus, the reads mapped to the lead allele were removed, and another step of alignment and quantification was performed in order to determine if the second allele was real, or just noise due to extensive similarity to the lead allele. If the second allele had at least 1% of the locus read counts, it was kept, otherwise the genotype was considered to be homozygous for the lead allele.
The thresholds described above were chosen because they maximized the concordance with the Sanger sequencing typings [6], while also minimizing the rate of false homozygotes and heterozygotes.
We implemented two versions of the HLApers pipeline: (1) one using STAR (v2.5.3a) [34] to map reads followed by Salmon (v0.8.2) [36] to quantify the expression, and (2) using kallisto (v0.43.1) [35], which performs pseudoalignment and quantification.
The quantification pipeline is structured in a two-stage process, first identifying the most expressed allele(s) at each HLA locus in order to infer the genotype which is present, and next quantifying expression for these inferred genotypes as well as for the rest of the transcriptome (Fig 1).
Reads were aligned directly to the transcriptome. STAR alignments were passed to Salmon for quantification (module quant under alignment mode), whereas kallisto directly produces quantifications with the quant module.
In both the HLA typing step (for which the index contains all HLA sequences in the IMGT database) and the quantification step (for which an HLA-personalized index is used), short reads can map to more than one locus, or more commonly to multiple alleles of the same locus (multimaps). The quantification methods we are using deal with multimaps by inferring maximum likelihood estimates optimized by an expectation-maximization algorithm to probabilistically assign reads to each reference in the index, and also include models to account for sequencing bias.
For the mapping with STAR, we tuned parameters in order to avoid discarding multimaps and to accommodate mismatches. For quantification, we used all bias correction options available (–seqBias and –gcBias in Salmon, and –bias in kallisto).
Polyester is an R package designed to simulate RNA-seq datasets [41]. We used the function simulate_experiment_countmat to simulate transcriptome data for 50 randomly chosen GEUVADIS individuals. Simulations were based on the read lengths and counts in the original data, with library sizes of 30 million reads, sampling without bias from a normal distribution of read start sites (with average fragment length of 250bp and sd of 25bp, and error rate of 0.005). The code for the generation of the simulated datasets is available at https://github.com/genevol-usp/hlaexpression/tree/master/simulation/data.
We processed the simulated reads with STAR-Salmon to perform the quantifications using different indices:
To investigate the relationship between quantifications and sequence divergence with respect to the reference, we used the R function adist to calculate the proportion of mismatches between the HLA alleles carried by the individuals and the alleles in the reference genome.
We quantified HLA expression based on RNA-seq data for 358 European individuals, in samples of LCLs (Lymphoblastoid Cell Lines), made available by the GEUVADIS Consortium [24] (we excluded samples from the original dataset which are not in the 1000 Genomes phase 3).
We performed expression quantification using the HLApers and reference transcriptome pipelines.
We evaluated the reproducibility of the HLA quantifications by analyzing a subset of 97 individuals for which a replicate was available (https://github.com/genevol-usp/hlaexpression/blob/master/geuvadis_reanalysis/replicates/data/write_sample_info.R).
For the eQTL analysis, we used only autosomal genes which are expressed in a large proportion of samples, exploring the thresholds of TPM > 0 in at least 25%, 50%, or 75% of samples. In order to correct the expression data for technical effects, we sequentially removed the effect of the first 0 to 100 PCs and ran an eQTL analysis for each condition (S6 Fig). The configuration of thresholds and number of PCs which maximized the eQTL discovery (at FDR = 5%) was considered. This resulted in the use of genes expressed in ≥ 50% of samples (19,613 genes), and 60 PCs.
The PCA analysis and data correction were performed with QTLtools v1.1 [47], using the modules pca and correct respectively. For the genetic variant data, we used the 1000 Genomes Phase 3 biallelic variants, lifted to GRCh38 coordinates, after filtering for MAF ≥ 0.05 in the individuals included in this study (6,837,505 variants in total).
In order to control for population structure in the eQTL analysis, we ran a PCA on the variant genotype data and assessed the PCs which captured the structure. We used QTLtools pca requiring that variants should be at least 6kb apart. After visual inspection of the plots in S7 Fig, PCs 1–3 were used as covariates in the eQTL analysis.
We used QTLtools cis to conduct the cis-eQTL analysis using the following model:
PCA - corrected and standard normal expression ∼ SNPs + covariates ( PCs for population stratification )
The permutation pass was performed with 1000 permutations and a cis-window of 1Mb. P-values were computed by beta approximation and significance was determined by running the script runFDR_cis.R provided by QTLtools with FDR of 5%.
Multiple eQTLs with independent effects on a particular gene were mapped with a conditional analysis based on step-wise linear regression (see Supplementary method 8 in [47]). The method automatically learns the number of independent signals per gene and provides sets of candidate eQTLs per signal.
In order to investigate the putative function of the eQTLs we mapped, we investigated whether these eQTLs were present in ENCODE [53] regulatory elements annotated for LCLs. We used three types of functional annotations: open chromatin regions given by DNAse footprinting, transcription factor binding sites (TFBS) assayed by ChIP-seq, and histone modifications.
We performed an RTC [48] analysis as described in [47] to investigate whether our eQTLs tagged the same causal variant as a GWAS variant or previously reported eQTL.
We downloaded the GWAS catalog data (v1.0.1) from https://www.ebi.ac.uk/gwas/api/search/downloads/alternative and selected associations with p-value < 10−8. We obtained the coordinates of recombination hotspots from http://jungle.unige.ch/QTLtools_examples/hotspots_b37_hg19.bed.
We applied the RTC module implemented in QTLtools (QTLtools rtc), selecting the HLA region only, using a D’ threshold of 0.5, and turning on the conditional flag (–conditional) to test all independent eQTLs for a gene.
Following the recommendation of Delaneau et al. (see Supplementary Note 7 in [47]), we considered that two SNPs tagged the same functional signal if the RTC score was >0.9.
To investigate whether there is a haplotypic coordination of expression at HLA, we used phased HLA genotype data to verify if there was more correlation of expression between alleles on the same haplotype than on different haplotypes (Fig 8). To also assess the phasing between HLA alleles and the eQTLs (Fig 6), we included the eQTLs mapped for each HLA gene in the phasing procedure, accounting for the fact that their phasing was already known from 1000 Genomes Project.
In order to be conservative in the phase estimation, we used only the haplotype calls which were concordant between two approaches for phasing. First, we used PHASE [68] to determine the haplotype of each allele in the genotype, providing HLA allele designations and phased eQTL genotypes as input. Second, we checked the compatibility of the individual HLA genotypes with the phased SNP haplotypes from 1000 Genomes. Given all possible haplotypes for each individual at HLA-G∼A∼E∼C∼B∼DRA∼DRB1∼DQA1∼DQB1∼DPA1∼DPB1, we checked which combination of 2 haplotypes minimized the number of differences to the 1000 Genomes data to infer the haplotypes present in the individual. This resulted in 417 haplotypes completely concordant between the two approaches.
The HLApers pipeline is available at https://github.com/genevol-usp/HLApers. The entire analysis, including simulations, index compilation, quantification of expression, eQTL mapping, etc is available at https://github.com/genevol-usp/hlaexpression.
|
10.1371/journal.ppat.1001069 | PKC Signaling Regulates Drug Resistance of the Fungal Pathogen Candida albicans via Circuitry Comprised of Mkc1, Calcineurin, and Hsp90 | Fungal pathogens exploit diverse mechanisms to survive exposure to antifungal drugs. This poses concern given the limited number of clinically useful antifungals and the growing population of immunocompromised individuals vulnerable to life-threatening fungal infection. To identify molecules that abrogate resistance to the most widely deployed class of antifungals, the azoles, we conducted a screen of 1,280 pharmacologically active compounds. Three out of seven hits that abolished azole resistance of a resistant mutant of the model yeast Saccharomyces cerevisiae and a clinical isolate of the leading human fungal pathogen Candida albicans were inhibitors of protein kinase C (PKC), which regulates cell wall integrity during growth, morphogenesis, and response to cell wall stress. Pharmacological or genetic impairment of Pkc1 conferred hypersensitivity to multiple drugs that target synthesis of the key cell membrane sterol ergosterol, including azoles, allylamines, and morpholines. Pkc1 enabled survival of cell membrane stress at least in part via the mitogen activated protein kinase (MAPK) cascade in both species, though through distinct downstream effectors. Strikingly, inhibition of Pkc1 phenocopied inhibition of the molecular chaperone Hsp90 or its client protein calcineurin. PKC signaling was required for calcineurin activation in response to drug exposure in S. cerevisiae. In contrast, Pkc1 and calcineurin independently regulate drug resistance via a common target in C. albicans. We identified an additional level of regulatory control in the C. albicans circuitry linking PKC signaling, Hsp90, and calcineurin as genetic reduction of Hsp90 led to depletion of the terminal MAPK, Mkc1. Deletion of C. albicans PKC1 rendered fungistatic ergosterol biosynthesis inhibitors fungicidal and attenuated virulence in a murine model of systemic candidiasis. This work establishes a new role for PKC signaling in drug resistance, novel circuitry through which Hsp90 regulates drug resistance, and that targeting stress response signaling provides a promising strategy for treating life-threatening fungal infections.
| Treating fungal infections is challenging due to the emergence of drug resistance and the limited number of clinically useful antifungal drugs. We screened a library of 1,280 pharmacologically active compounds to identify those that reverse resistance of the leading human fungal pathogen, Candida albicans, to the most widely used antifungals, the azoles. This revealed a new role for protein kinase C (PKC) signaling in resistance to drugs targeting the cell membrane, including azoles, allylamines, and morpholines. We dissected mechanisms through which PKC regulates resistance in C. albicans and the model yeast Saccharomyces cerevisiae. PKC enabled survival of cell membrane stress at least in part through the mitogen-activated protein kinase (MAPK) cascade in both species. In S. cerevisiae, inhibition of PKC signaling blocked activation of a key regulator of membrane stress responses, calcineurin. In C. albicans, Pkc1 and calcineurin independently regulate resistance via a common target. Deletion of C. albicans PKC1 rendered fungistatic drugs fungicidal and reduced virulence in a mouse model. The molecular chaperone Hsp90, which stabilizes client proteins including calcineurin, also stabilized the terminal C. albicans MAPK, Mkc1. We establish new circuitry connecting PKC with Hsp90 and calcineurin and suggest a promising strategy for treating life-threatening fungal infections.
| Microbial survival depends critically upon coordination of sensing environmental stimuli with control of the appropriate cellular responses. As a consequence, microbes have evolved elaborate mechanisms to sense and respond to diverse environmental stresses, including oxidative stress, osmotic stress, thermal stress, changes in pH, and nutrient limitation [1], [2]. Signal transduction cascades integrate recognition and response to these stresses as well as to challenges imposed by exposure to various small molecules that are a ubiquitous presence in the environment. Small molecules can have a dramatic effect on cellular signaling, mediate communication between microbes, or exert potentially lethal toxicity [3], [4], [5], [6], [7]. Many natural products are produced by microbes in competitive communities and can lead to selection for enhanced capacity to tolerate these agents. Since natural products and their derivatives are extensively used in medicine and agriculture [8], [9], the evolution of resistance to these agents can have profound consequences for human health.
The evolution of drug resistance in fungal pathogens poses considerable concern given that invasive fungal infections are a leading cause of human mortality worldwide, especially among immunocompromised individuals. The frequency of such infections is on the rise in concert with the growing population of patients with compromised immune systems due to chemotherapy, transplantation of organs or hematopoietic stem cells, or infection with HIV [10], [11]. The leading fungal pathogen of humans is Candida albicans, which ranks as the fourth most common cause of hospital acquired infectious disease and is associated with mortality rates approaching 50% [12], [13], [14]. There is a very limited repertoire of antifungal drugs with distinct targets for the treatment of fungal infections, in part due to the close evolutionary relationships between these eukaryotic pathogens and their hosts [15], [16]. Most of the antifungal drugs in clinical use target the biosynthesis or function of ergosterol, the main sterol of fungal membranes [2], [17], [18]. The therapeutic efficacy of most antifungal drugs is compromised by the emergence of drug resistant strains, superinfection with resistant strains, and by static rather than cidal activities that block fungal growth but do not eradicate the pathogen population. To improve clinical outcome it will be necessary to develop new antifungal drugs with different mechanisms of action and to discover drugs that improve the fungicidal activity of current antifungals.
The molecular basis of antifungal drug resistance is best characterized in the context of the azoles through studies with C. albicans and the model yeast Saccharomyces cerevisiae. The azoles have been the most widely deployed class of antifungal drugs for decades and inhibit lanosterol 14α-demethylase, encoded by ERG11, resulting in a block in ergosterol biosynthesis, the accumulation of a toxic sterol intermediate, and cell membrane stress [2], [17], [18]. The azoles are generally fungistatic against Candida species and many patients are on long-term therapy, creating favorable conditions for the emergence of resistance. Despite the evolutionary distance between C. albicans and S. cerevisiae, mechanisms of azole resistance are largely conserved [19]. Resistance can arise by mechanisms that minimize the impact of the drug on the fungus, such as the overexpression of multidrug transporters or alterations of the drug target that prevent the drug from inhibiting its target. Alternatively, resistance can arise by mechanisms that minimize drug toxicity, such as loss of function of the ergosterol biosynthetic enzyme Erg3, which blocks the production of the toxic sterol that would otherwise accumulate when the azoles inhibit Erg11. Recent studies have established that basal tolerance of wild-type strains and resistance due to mechanisms that mitigate drug toxicities without blocking the effect of the drug on the cell are often dependent upon stress responses that are critical for survival of azole-induced cell membrane stress [2], [18].
The key regulator of cellular stress responses implicated in both basal tolerance and resistance to azoles is Hsp90 [2], [18], [20]. Hsp90 is an essential molecular chaperone that regulates the stability and function of a diverse set of client proteins, many of which are regulators of cellular signaling [21], [22], [23]. In S. cerevisiae and C. albicans, inhibition of Hsp90 function blocks the rapid evolution of azole resistance and abrogates resistance that was acquired by diverse mutations [24], [25]. A central aspect of Hsp90's role in the emergence and maintenance of azole resistance is that it enables calcineurin-dependent stress responses that are required to survive the membrane stress exerted by azoles. In both yeast species, Hsp90 physically interacts with calcineurin keeping it in a stable conformation that is poised for activation [26], [27]. Inhibition of calcineurin function phenocopies inhibition of Hsp90 function, abrogating azole resistance of diverse mutants [24], [25]. This has led to the model that calcineurin is the key mediator of Hsp90-dependent azole resistance. Notably, in C. albicans both Hsp90 and calcineurin have recently been demonstrated to regulate resistance to the echinocandins, the only new class of antifungals to reach the clinic in decades; they inhibit the synthesis of (1,3)-β-D-glucan, a key component of the fungal cell wall [20], [27].
Another key cellular stress response pathway implicated in basal tolerance to antifungal drugs is the protein kinase C (PKC) cell wall integrity pathway, though it has only been implicated in tolerance to drugs targeting the cell wall. Central to the core of this signaling cascade is Pkc1, the sole PKC isoenzyme in S. cerevisiae that is essential under standard growth conditions and regulates maintenance of cell wall integrity during growth, morphogenesis, and response to cell wall stress [28], [29], [30], [31]. Signals are initiated by a family of cell surface sensors that are coupled to the small G-protein Rho1, which activates a set of effectors including Pkc1. Pkc1 signaling has been the focus of extensive study in S. cerevisiae where it is known to regulate multiple targets, most notably the mitogen-activated protein kinase (MAPK) cascade comprised of a linear series of protein kinases including the MAPKKK Bck1, the MAPKKs Mkk1/2, and the MAPK Slt2 that relays signals to the terminal transcription factors Rlm1 and Swi4/Swi6. While Pkc1 is not essential in C. albicans [32], the Pkc1-activated MAPK cascade is conserved in C. albicans with Bck1, Mkk2, and the Slt2 homolog Mkc1 [33]. In both species, components of the Pkc1 signaling cascade have been implicated in mediating tolerance to the stress exerted by the echinocandins that target the fungal cell wall [34], [35], [36], [37].
Here, we embarked on a drug screen of 1,280 pharmacologically active compounds to identify molecules that abrogate azole resistance of both an S. cerevisiae resistant mutant and a C. albicans clinical isolate. We identified a key role for PKC signaling in mediating crucial responses to azoles as well as to other drugs targeting the ergosterol biosynthesis pathway, including allylamines and morpholines. Pkc1 regulated responses to azoles at least in part via the MAPK cascade in both species via multiple downstream effectors. Strikingly, inhibition of Pkc1 function phenocopied inhibition of Hsp90 or calcineurin. In S. cerevisiae, compromise of PKC signaling blocked calcineurin activation in response to ergosterol biosynthesis inhibitors, providing a compelling mechanism for the impact on drug resistance. In C. albicans, we found that Pkc1 and calcineurin independently regulate resistance via a common target. The complexity of interactions linking PKC signaling, Hsp90, and calcineurin was further illuminated as genetic reduction of C. albicans Hsp90 resulted in destabilization of Mkc1 thereby blocking its activation. Deletion of C. albicans PKC1 rendered the fungistatic ergosterol biosynthesis inhibitors fungicidal and attenuated virulence in a murine model of systemic disease. Our findings establish an entirely new role for PKC signaling in basal tolerance and resistance to ergosterol biosynthesis inhibitors, a novel mechanism through which Hsp90 regulates drug resistance, and that targeting Pkc1 provides a promising therapeutic strategy for life-threatening fungal infections.
To identify compounds that enhance the efficacy of the azole fluconazole we screened the LOPAC1280 Navigator library. Our initial screen used an S. cerevisiae strain with azole resistance due to deletion of ERG3. This resistance phenotype is exquisitely sensitive to perturbation of stress response pathways [24], [25]. To enhance the activity of library compounds, this azole-resistant mutant also harbored deletion of PDR1 and PDR3, transcription factors that regulate the expression of numerous multidrug transporters which efflux structurally diverse compounds from the cell [38]. The library was initially screened at 25 µM in defined RPMI medium at 30°C in the presence of 8 µg/ml fluconazole, which reduces growth of this strain by less than 50% . The compounds that reduced growth by greater than or equal to 50% relative to the fluconazole-only controls were re-screened at 12.5 µM in the presence and absence of fluconazole to distinguish those that enhance the activity of fluconazole from those that are simply toxic on their own. This screen identified 185 compounds that enhanced the efficacy of fluconazole (data not shown). To prioritize compounds with synergistic activity with fluconazole against a clinical isolate of C. albicans, we then screened the 185 compounds at 12.5 µM for activity against an isolate from an HIV-infected patient undergoing fluconazole treatment, both in the presence and absence of fluconazole at 8 µg/ml. The capacity of this clinical isolate to grow in the presence of high concentrations of azole is critically dependent upon cellular stress responses [25], despite the fact that it has increased expression of the multidrug transporter Mdr1 relative to a drug-sensitive isolate recovered from the same patient at an earlier time point [39], [40], [41]. This secondary screen identified seven compounds that had little toxicity on their own but which enhanced the efficacy of fluconazole (Figure 1A). One hit from our screen, brefeldin A, was recently confirmed to exhibit potent synergy with antifungals against Candida and Aspergillus [42]. Strikingly, three of the seven hits were characterized as inhibitors of protein kinase C (PKC).
PKC governs the cell wall integrity signaling pathway so named for its role in regulating cell wall integrity during growth, morphogenesis, and exposure to stress in fungi [29], [30], [31]. In both S. cerevisiae and C. albicans, the PKC signaling cascade is known to regulate cellular responses crucial for survival of exposure to antifungal drugs targeting the cell wall, such as the echinocandins [34], [35], [36], [37]. Since the PKC inhibitors identified in our screen were characterized in mammalian cells [43], [44], we next turned to other pharmacological inhibitors of PKC whose mode of action had been validated in fungi. Cercosporamide was identified as a selective Pkc1 inhibitor through C. albicans Pkc1-based high-throughput screening and was shown to exhibit potent synergy with echinocandins [45]. We purified cercosporamide from the fungus Cercosporidium henningsii following standard protocols [46]. As a positive control, we tested the impact of a concentration gradient of cercosporamide on growth in the presence of a fixed concentration of the echinocandin micafungin that causes less than 50% inhibition of growth on its own and confirmed that cercosporamide had the expected synergistic activity with micafungin against the clinical C. albicans isolate (Figure 1B). Using a comparable assay, we determined that cercosporamide also enhanced the activity of fluconazole (Figure 1B), validating the results from our screen. We further confirmed our pharmacological findings with another PKC inhibitor characterized in fungi, staurosporine [47], [48]. Both cercosporamide and staurosporine enhanced the efficacy of antifungals targeting the cell wall, micafungin, and those targeting the cell membrane (Figure 1C), including fluconazole and the morpholine fenpropimorph, which inhibits Erg2 and Erg24 [49]. While staurosporine enhanced the efficacy of another ergosterol biosynthesis inhibitor that inhibits Erg1 [49], the allylamine terbinafine, cercosporamide did not (Figure 1C). The lack of effect of cercosporamide on terbinafine tolerance is likely an artifact of an inactivating drug-drug interaction given that mutants that are hypersensitive to terbinafine are rendered resistant by cercosporamide (data not shown).
In S. cerevisiae, PKC1 is essential [50], thus we used a strain harboring only a temperature-sensitive (ts) pkc1-3 allele [51] and assayed tolerance to three ergosterol biosynthesis inhibitors fluconazole, fenpropimorph, and terbinafine. Growth of the wild-type strain and the pkc1-3 ts mutant was assayed over a gradient of drug concentrations relative to a drug-free control at either the permissive temperature (30°C) or at a more restrictive temperature, but where the pkc1-3 ts mutant was still able to thrive in the absence of antifungals (35°C). At the permissive temperature, the wild type and the pkc1-3 ts mutant had comparable tolerance to all three drugs tested (Figure S1A). At the restrictive temperature, the pkc1-3 ts mutant was hypersensitive to all three drugs (Figure 2A). The same trend was observed when a dilution series of cells was spotted on solid medium with a fixed concentration of drug (Figure S1B). To determine if reduction of Pkc1 function rendered the fungistatic ergosterol biosynthesis inhibitors fungicidal we used tandem assays with an antifungal susceptibility test performed at the restrictive temperature followed by spotting onto rich medium without any inhibitors. The wild-type strain was able to grow on rich medium following exposure to all concentrations of drug tested (Figure 2B); compromise of Pkc1 function in the pkc1-3 ts mutant enhanced cidality of all three drugs with the most severe effect for fluconazole and fenpropimorph. Thus, reduction of Pkc1 activity increases sensitivity to drugs targeting the cell membrane and enhances cidality of these otherwise fungistatic agents.
Despite the simple linear schematic commonly used to illustrate the architecture of the Pkc1 cell wall integrity pathway (Figure 2C), there is evidence for additional Pkc1 targets [30] and multiple cases of cross talk with other stress response pathways [28]. We next sought to determine if the effects of Pkc1 on tolerance to ergosterol biosynthesis inhibitors are due to signaling via the downstream MAPK cascade. S. cerevisiae mutants lacking the MAPKKK Bck1 or the terminal MAPK Slt2 were hypersensitive to all three ergosterol biosynthesis inhibitors tested in both a liquid antifungal susceptibility assay measuring growth of a fixed concentration of cells across a gradient of drug concentrations (Figure 2A) and a spotting assay of a dilution of cells on solid medium with a fixed concentration of drug (Figure S1B). Deletion of the MAPK components also rendered these fungistatic drugs fungicidal. Thus, Pkc1 enables tolerance to ergosterol biosynthesis inhibitors via the MAPK cascade in S. cerevisiae.
In C. albicans, PKC1 is not essential though it does share a high degree of sequence conservation with S. cerevisiae PKC1 and has a conserved role in regulating cell wall integrity through a conserved MAPK cascade [32], [33]. To genetically validate the role of C. albicans PKC1 in tolerance to drugs affecting the cell membrane, we constructed a pkc1Δ/pkc1Δ mutant. Homozygous deletion of PKC1 rendered the strain hypersensitive to all three ergosterol biosynthesis inhibitors tested in liquid static susceptibility assays (Figure 3A) as well as on solid medium (Figure S2A). Comparable results were obtained in well-aerated shaking liquid cultures (data not shown). Restoring a wild-type PKC1 allele under the control of the native promoter to the native locus restored drug tolerance (Figure S2). To determine if deletion of C. albicans PKC1 renders the ergosterol biosynthesis inhibitors fungicidal, we used tandem assays with an antifungal susceptibility test followed by spotting onto rich medium without inhibitor. A strain with wild-type PKC1 levels was able to grow on rich medium following exposure to all drug concentrations tested (Figure 3B). Homozygous deletion of C. albicans PKC1 was cidal in combination with any dose of ergosterol biosynthesis inhibitor tested; no cells were able to grow on rich medium following exposure to the treatments. Thus, Pkc1 regulates crucial cellular responses for surviving the cell membrane stress exerted by antifungal drugs.
As an initial approach to assess whether the MAPK cascade was implicated in responses to drugs targeting the cell membrane, we monitored activation of the terminal MAPK in C. albicans. Mkc1 is known to be activated in response to distinct stress conditions including oxidative stress, changes in osmotic pressure, cell wall damage, and cell membrane perturbation [52]. To determine if Mkc1 is activated in response to ergosterol biosynthesis inhibitors we monitored Mkc1 phosphorylation using an antibody that detects dual phosphorylation on conserved threonine and tyrosine residues. Exposure to fluconazole, fenpropimorph, and terbinafine led to Mkc1 activation comparable to exposure to the cell wall damaging antifungal micafungin (Figure S3A). However, activation of signal transducers is not always coupled with functional consequences of their deletion. For example, Mkc1 is activated by exposure to hydrogen peroxide but is not required for survival of this stress [52].
To determine if the role of the MAPK cascade was conserved in C. albicans, we constructed homozygous deletion mutants lacking either the MAPKKK Bck1 or the terminal MAPK Mkc1 (homolog of S. cerevisiae Slt2). Homozygous deletion of either BCK1 or MKC1 rendered strains hypersensitive to fluconazole, fenpropimorph, and terbinafine (Figure 3A) but had negligible effect at elevated temperatures (Figure S3B). This stands in contrast to our results with S. cerevisiae that demonstrated an equivalent role of the MAPK cascade at all temperatures tested (Figure 2, Figure S1 and S4). While deletion of C. albicans PKC1 rendered the ergosterol biosynthesis inhibitors fungicidal, deletion of BCK1 or MKC1 did not (Figure 3B). These results not only implicate the MAPK cascade in C. albicans but also suggest that alternate effectors downstream of Pkc1 are more important at elevated temperature and enable survival in the presence of ergosterol biosynthesis inhibitors.
Effectors downstream of the terminal MAPK of the PKC signaling cascade have been well studied in S. cerevisiae and include both nuclear and cytoplasmic proteins. Slt2 is known to regulate activation of two transcription factors Rlm1 and SBF, which is comprised of Swi4 and Swi6 [30]. Rlm1 mediates the majority of the transcriptional output of cell wall integrity signaling, largely genes involved in cell wall biogenesis [53]. SBF drives cell cycle-specific transcription and is also regulated by Slt2 in response to cell wall stress (reviewed in [30]). Swi4 interacts directly with Slt2 and has additional roles in transcriptional regulation independent of the regulatory subunit Swi6 [54]. Slt2 translocates from the nucleus to the cytoplasm in response to cell wall stress [55]. Cytoplasmic Slt2 is required for activation of a high-affinity Ca2+ influx system in the plasma membrane that is comprised of two subunits, Cch1 and Mid1, in response to endoplasmic reticulum stress [56]. Activation of the Cch1-Mid1 channel leads to the accumulation of intracellular Ca2+ and activation of the protein phosphatase calcineurin [57].
To dissect the role of downstream effectors of Slt2 in ergosterol biosynthesis inhibitor tolerance, we tested the impact of their deletion individually and in combination on antifungal susceptibility. For reference, we included a strain lacking the regulatory subunit of calcineurin, CNB1, which is hypersensitive to ergosterol biosynthesis inhibitors [24]. For fluconazole, deletion of RLM1, CCH1, or MID1 had negligible impact on tolerance while deletion of SWI4 or SWI6 rendered strains almost as sensitive as the slt2Δ mutant (Figure 4). To determine if there was redundancy among the downstream effectors, we constructed strains harboring deletion of multiple effectors. Deletion of CCH1 phenocopies deletion of the entire channel and deletion of SWI4 abolishes SBF function as well as Swi4-dependent transcription independent of SBF. Thus, combined deletion of CCH1, SWI4, and RLM1 should eliminate the four known targets of Slt2 phosphorylation. No additional increase in sensitivity was observed in double or triple mutants. This suggests that the SBF transcription factor is of central importance for enabling responses to fluconazole. For fenpropimorph, deletion of RLM1, CCH1, or MID1 had no impact on tolerance individually while deletion of SWI4 or SWI6 caused a partial increase in sensitivity (Figure 4). Deletion of RLM1 in the context of the swi4Δ or swi6Δ mutants further increased fenpropimorph sensitivity. Deletion of CCH1 in the mutant backgrounds had little additional impact. This suggests that SBF is the major determinant of fenpropimorph tolerance with RLM1 enabling additional responses important in the absence of SBF. For tolerance to terbinafine, deletion of RLM1 had no impact while deletion of SWI4 caused a partial increase in sensitivity (Figure 4). Unlike tolerance to fluconazole and fenpropimorph, deletion of SWI6 had negligible impact on terbinafine tolerance while deletion of CCH1 or MID1 caused a partial increase in sensitivity. Deletion of both RLM1 and CCH1 in the swi4Δ mutant caused an incremental increase in sensitivity (Figure 4). These results suggest that Swi4 enables terbinafine tolerance independent of the SBF complex and that Rlm1 and Cch1 mediate responses that are important in the absence of Swi4. Thus, distinct downstream effectors are important for tolerance of S. cerevisiae to different ergosterol biosynthesis inhibitors.
Next, we tested a set of C. albicans mutants to determine if the role of the effectors downstream of the terminal MAPK of the PKC signaling cascade was conserved. As was the case with S. cerevisiae, deletion of RLM1 on its own had no impact on tolerance to the ergosterol biosynthesis inhibitors (Figure 5), consistent with recent findings [58]. Deletion of SWI4 rendered strains hypersensitive to all three ergosterol biosynthesis inhibitors tested (Figure 5). Deletion of SWI6 or combined deletion of both SWI4 and SWI6 conferred a comparable increase in sensitivity (data not shown; unpublished strains generously provided by Catherine Bachewich), implicating the SBF complex in responses to drug-induced membrane stress. Deletion of CCH1 or MID1 individually or in combination had a comparable effect to deletion of SWI4 rendering the strain hypersensitive to all three ergosterol biosynthesis inhibitors tested (Figure 5). Notably, C. albicans cch1Δ/cch1Δ and mid1Δ/mid1Δ mutants share some but not all phenotypes of a calcineurin mutant [59]. In terms of ergosterol biosynthesis inhibitor sensitivity, deletion of the gene encoding the catalytic subunit of calcineurin, CNA1, caused hypersensitivity akin to that of the cch1Δ/cch1Δ and mid1Δ/mid1Δ mutants for fluconazole and fenpropimorph but caused slightly greater sensitivity to terbinafine (Figure 5). Thus, in C. albicans both the SBF complex and the Cch1-Mid1 channel play critical roles in tolerance to drugs that target the cell membrane.
Given calcineurin's established role in mediating drug-induced membrane stress responses [24], [25], [57] and that Slt2 has been shown to enable calcineurin activation by phosphorylating Cch1 [56], we tested whether calcineurin was activated in response to ergosterol biosynthesis inhibitors and whether deletion of Slt2 blocked this activation. To monitor calcineurin activation, we used a well-established reporter system that exploits the downstream effector Crz1, a transcription factor that is dephosphorylated upon calcineurin activation [60], [61]. Dephosphorylated Crz1 translocates to the nucleus, driving expression of genes with calcineurin-dependent response elements (CDREs) in their promoters. We used a reporter containing four tandem copies of CDRE and a CYC1 minimal promoter driving lacZ [61]. We confirmed previous findings that fluconazole activates calcineurin ([27], [62] and Figure 6A). We also found that the other ergosterol biosynthesis inhibitors terbinafine and fenpropimorph activate calcineurin (P<0.001, ANOVA, Bonferroni's Multiple Comparison Test, Figure 6A). Deletion of SLT2 completely blocked calcineurin activation in response to ergosterol biosynthesis inhibitors as did deletion of the regulatory subunit of calcineurin required for its activation, encoded by CNB1 (P<0.001). Pharmacological inhibition of PKC signaling with staurosporine also blocked calcineurin activation (P<0.001, Figure 6B). The block in calcineurin activation was not an artifact of compromised viability as treatment conditions were optimized such that all cultures underwent comparable growth with equivalent protein yields. Given that the slt2Δ mutant is slightly more sensitive to ergosterol biosynthesis inhibitors than the mutant lacking calcineurin function, it is likely that Slt2 regulates responses to ergosterol biosynthesis inhibitors through additional targets. The swi4Δ mutant is less sensitive than the calcineurin mutant, suggesting that Slt2 regulates calcineurin function independently of Swi4 and that Swi4 regulates ergosterol biosynthesis inhibitor tolerance through additional targets (Figure 6C).
Given that deletion of CCH1 and MID1 had negligible effect on tolerance to fluconazole or fenpropimorph and only an intermediate effect on tolerance to terbinafine, it is likely that compromise of PKC signaling blocked calcineurin activation by a mechanism that is largely distinct from the Cch1-Mid1 channel. One possible mechanism is that the effects are transcriptional and mediated through a nuclear target of Slt2 such that inhibition of PKC signaling compromises the expression of calcineurin or CRZ1. However, deletion of SLT2 did not reduce the expression of genes encoding any of the calcineurin subunits (CNA1, CNA2 or CNB1) or CRZ1 as measured by quantitative RT-PCR in the presence or absence of ergosterol biosynthesis inhibitor (P>0.05, ANOVA, Bonferroni's Multiple Comparison Test, Figure S5). Thus, PKC signaling enables calcineurin activation in response to ergosterol biosynthesis inhibitors by a mechanism that is largely distinct from the Cch1-Mid1 channel or transcriptional control of calcineurin.
In contrast to the minor impact of deletion of the S. cerevisiae Cch1-Mid1 channel, deletion of the C. albicans Cch1-Mid1 channel had nearly as great an effect as deletion of the catalytic subunit of calcineurin in response to fluconazole and fenpropimorph; for terbinafine the effect was partial (Figure 5). To test if the ergosterol biosynthesis inhibitors activate calcineurin and if inhibition of PKC signaling blocks this activation, we monitored transcript levels of two calcineurin-dependent genes, PLC3 and UTR2 [63]. In a wild-type strain, fluconazole activated calcineurin as measured by an increase in PLC3 and UTR2 transcript levels (P<0.05, ANOVA, Bonferroni's Multiple Comparison Test, Figure 7A). As expected, deletion of the catalytic subunit of calcineurin, CNA1, blocked the induction of PLC3 and UTR2 transcripts (P<0.01). Deletion of PKC1 did not block induction of PLC3 or UTR2 indicating that impairment of PKC signaling does not block calcineurin activation (Figure 7A). At 35°C, conditions under which Pkc1 downstream effectors other than the MAPK cascade are more important in tolerance to ergosterol biosynthesis inhibitors, deletion of PKC1 increased the magnitude of induction of PLC3 and UTR2 (P<0.001, Figure 7A). At 30°C, conditions under which the MAPK cascade mediates tolerance to ergosterol biosynthesis inhibitors, deletion of PKC1 had no significant impact on calcineurin-dependent transcription (data not shown). Thus, drugs that inhibit ergosterol biosynthesis induce calcineurin activation in a manner that is independent of PKC signaling.
Next, we addressed alternative models that could explain the relationship between PKC signaling and calcineurin in C. albicans tolerance to ergosterol biosynthesis inhibitors. One possible model is that Pkc1 and calcineurin regulate tolerance through parallel but non-redundant pathways. This model leads to two predictions for ergosterol biosynthesis inhibitor tolerance: 1) there should be a synergistic effect of inhibiting both pathways simultaneously and 2) compromise of one pathway should confer increased sensitivity to inhibition of the other. To test the first prediction, we performed checkerboard assays in which a wild-type strain was exposed to a uniform concentration of fluconazole and a concentration gradient of both the calcineurin inhibitor cyclosporin A and the Pkc1 inhibitor staurosporine. There was no obvious synergy detected upon inhibition of both pathways in combination with fluconazole (Figure 7B). To assess this quantitatively we calculated the standard index of drug synergy, the fractional inhibitory concentration (FIC). The FIC value was 0.75 confirming that there was no synergy. To test the second prediction, we measured the impact of Pkc1 inhibition on fluconazole tolerance of a mutant lacking the catalytic subunit of calcineurin, CNA1. Growth of the cna1Δ/cna1Δ mutant was assessed in the absence or presence of the highest concentration of fluconazole that it could tolerate and with a gradient of the Pkc1 inhibitor cercosporamide. Fluconazole-sensitivity of the cna1Δ/cna1Δ mutant was not affected by cercosporamide (Figure 7C). The reciprocal was also true, such that the pkc1Δ/pkc1Δ mutant was not rendered hypersensitive to fluconazole by the calcineurin inhibitor cyclosporin A (data not shown). Thus, our results do not support either prediction of the model in which Pkc1 and calcineurin regulate ergosterol biosynthesis inhibitor tolerance through parallel pathways. We also ruled out the possibility that inhibition of calcineurin blocks PKC signaling as measured by levels of activated Mkc1 (data not shown). Taken together, these findings support a model in which Pkc1 and calcineurin independently regulate crucial responses to ergosterol biosynthesis inhibitors through a common target (Figure 7D). This target is not Crz1, the only well-characterized effector downstream of C. albicans calcineurin, given that transcription of the calcineurin-dependent genes PLC3 and UTR2 is mediated through the transcription factor Crz1 [63].
To determine if the role of PKC signaling in tolerance to drugs targeting the cell membrane was conserved in the context of bona fide drug resistance, we turned to C. albicans clinical isolates and resistant mutants (Figure 8). We tested the impact of two structurally unrelated Pkc1 inhibitors, cercosporamide and staurosporine, on azole susceptibility of a series of C. albicans isolates that evolved fluconazole resistance in a human host [40]. The isolates shown begin with the second isolate in the series, which is the first with elevated resistance and increased expression of the multidrug transporter Mdr1 [39], [40], [41]. Azole resistance of this series is known to have evolved from a state of dependence on calcineurin and Hsp90 to a state of independence and this change is associated with the accumulation of additional mutations [25]. The third isolate from the bottom (Figure 8A) has mutation (R467K) and increased expression of the azole target Erg11; the last two isolates, which show the least of effect of Hsp90 inhibition on resistance, also have increased expression of the multidrug transporter Cdr1 [39], [40], [41]. Inhibition of Pkc1 had a strikingly similar impact on azole resistance to inhibition of Hsp90 or calcineurin, reducing resistance of isolates recovered early during treatment to a greater extent than those recovered late during treatment (Figure 8A).
To further explore the relationship between stress response signaling and classic resistance mechanisms such as mutation of the drug target and overexpression of multidrug transporters, we characterized additional clinical isolates and laboratory-derived mutants. We tested an additional five sets of clinical isolates for which we had one isolate recovered early during azole treatment and one recovered later. In all cases, inhibition of Pkc1 phenocopied inhibition of Hsp90, with the least effect on azole resistance of isolates that overexpressed the multidrug transporter Cdr1 (Figure S6). Since the clinical isolates often harbor multiple mechanisms of resistance, we also tested specific laboratory-derived resistant mutants. Inhibition of Pkc1 abolished resistance of laboratory-derived C. albicans and S. cerevisiae erg3 loss-of-function mutants (Figure 8B), as does inhibition of Hsp90 or calcineurin [24], [25]. In contrast, inhibition of Pkc1 did not affect S. cerevisiae resistance due to an activating mutation in the transcription factor Pdr1 that causes overexpression of multidrug transporters including Pdr5 (Figure 8B), as was the case with inhibition of Hsp90 [25]. We previously confirmed that genetic compromise of Hsp90 function does not affect resistance due to overexpression of Pdr5, confirming that the stability of this resistance phenotype is not due to Hsp90 inhibitors being pumped out of the cell [25]. Given the equivalent impact on azole resistance of Pkc1 inhibitors and Hsp90 inhibitors with diverse mutants, this strongly suggests that the stability of resistance of the Pdr1 mutant cannot be attributed to Pkc1 inhibitors being pumped out of the cell. Thus, inhibition of PKC signaling phenocopies inhibition of Hsp90 or its client protein calcineurin, reducing resistance of clinical isolates and specific resistant mutants. These results are consistent with the circuitry connecting PKC signaling and calcineurin delineated above and may also suggest an additional functional connection between Hsp90 and PKC signaling in regulating responses to ergosterol biosynthesis inhibitors.
While our findings already establish a link between PKC signaling and calcineurin-mediated stress responses, we next explored the possibility of yet another functional connection between Hsp90 and PKC signaling. In S. cerevisiae, Hsp90 binds exclusively to the activated form of Slt2 and enables Slt2-mediated activation of the downstream target Rlm1 [64]. To determine if the connection between Hsp90 and the terminal MAPK is conserved in C. albicans, we tested the impact of genetic depletion of C. albicans HSP90 on Mkc1 levels and activation status. To deplete Hsp90, we used a strain with its only HSP90 allele under the control of a doxycycline repressible promoter [65]. To monitor total Mkc1 levels, this kinase was tagged at the C-terminus using a 6x-histidine and FLAG epitope tag. The Mkc1-6xHis-FLAG protein was functional and sufficient to confer wild-type tolerance to ergosterol biosynthesis inhibitors (Figure S7). To determine whether Hsp90 stabilized only the activated form of Mkc1, we used a strain lacking the upstream MAPKKK required for Mkc1 activation, Bck1. All strains were grown in the presence of terbinafine to induce Mkc1 activation. In the absence of doxycycline (Figure 9A, left panel), all strains had comparable levels of Hsp90 as measured relative to a histone H3 loading control. All strains also had comparable levels of activated dually-phosphorylated Mkc1, with the exception of the strain lacking Bck1 in which Mkc1 activation was blocked. Total Mkc1 levels monitored by a 6X-histidine antibody were comparable for the three strains harboring the tagged protein. In the presence of doxycycline (Figure 9A, right panel), Hsp90 levels were depleted only in the strains with the repressible promoter. Depletion of Hsp90 resulted in a corresponding depletion of total Mkc1 levels, even in the strain lacking Bck1 in which Mkc1 remains in the inactivate state. Depletion of Hsp90 did not affect MKC1 transcript levels as measured by quantitative RT-PCR (P>0.05, ANOVA, Bonferroni's Multiple Comparison Test, Figure S8), confirming that the chaperone influences Mkc1 stability at the protein level. Thus, Hsp90 stabilizes Mkc1 independent of its activation status and thereby regulates PKC signaling, providing a new mechanism through which Hsp90 regulates drug-induced membrane stress responses (Figure 9B).
Given that deletion of PKC1 enhances the efficacy of antifungal drugs, we next explored the therapeutic efficacy in a well-established murine model in which fungal inoculum is delivered by tail vein injection and progresses from the bloodstream to deep-seated infection of major organs, most notably the kidney [27], [65], [66]. We compared kidney fungal burden of mice infected with either a wild-type strain or a pkc1Δ/pkc1Δ mutant. The average kidney fungal burden in mice infected with 1×105 CFUs of the wild-type parental strain was 4.34+/−0.54 log CFU per gram of kidney (Figure 10A). In stark contrast, the kidneys of mice infected with 1×105 CFUs of the pkc1Δ/pkc1Δ mutant were sterile (Figure 10A). To determine if infection with higher inocula of the pkc1Δ/pkc1Δ mutant would lead to sufficient kidney fungal burden to enable assessment of antifungal efficacy in vivo, we tested the impact of infection with 10-fold and 100-fold higher inocula. Mice infected with 1×106 or 1×107 CFUs of the pkc1Δ/pkc1Δ mutant demonstrated significantly reduced fungal burden relative to those infected with only 1×105 CFUs of the wild-type strain (P<0.001, ANOVA, Bonferroni's Multiple Comparison Test). The average kidney fungal burden in mice infected with 1×106 or 1×107 cells of the pkc1Δ/pkc1Δ mutant was 0.19+/−0.66 and 0.23+/−0.67 log CFU per gram of kidney, respectively. Thus, while C. albicans PKC1 is dispensable for growth under standard conditions in vitro it is required for proliferation and infection in a murine model, providing evidence for a key role of this stress response regulator in virulence. While the attenuated virulence of the pkc1Δ/pkc1Δ mutant precluded straightforward studies to determine if compromising Pkc1 enhances the efficacy of antifungal drugs in vivo, it provides compelling support for therapeutic potential of compromising fungal Pkc1.
Given our findings that Pkc1 and calcineurin affect drug resistance via a common target in C. albicans (Figure 7), it is possible that Pkc1-mediated signaling may influence virulence by a target in common with calcineurin, which is known to be required for C. albicans virulence [67], [68], [69]. Calcineurin mutants are hypersensitive to calcium present in serum and are unable to survive transit through the bloodstream [68]. However, while a mutant lacking the catalytic subunit of calcineurin was unable to survive on medium containing 50% serum, the pkc1Δ/pkc1Δ mutant exhibited only an intermediate reduction in viability and the mkc1Δ/mkc1Δ and bck1Δ/bck1Δ mutants grew as well as the wild type (Figure 10B). Further, the pkc1Δ/pkc1Δ mutant grew as well as the wild-type strain in liquid serum while the calcineurin mutant was inviable (data not shown). These results suggest that Pkc1 exerts powerful control over C. albicans virulence by means of targets distinct from calcineurin.
Our results establish a new role for the PKC signal transduction cascade in resistance to drugs targeting the cell membrane in the model yeast S. cerevisiae and the fungal pathogen C. albicans. Three out of seven hits from our screen of 1,280 pharmacologically active compounds for those that abrogate azole resistance are classified as inhibitors of PKC, suggesting a central role for this cellular regulator in azole resistance (Figure 1). Pharmacological inhibition of Pkc1 with two additional structurally distinct PKC inhibitors whose mode of action has been validated in fungi or genetic compromise of Pkc1 function enhances sensitivity to azoles as well as other drugs targeting ergosterol biosynthesis, including allylamines and morpholines (Figures 1, 2 and 3). Pkc1 regulates responses to ergosterol biosynthesis inhibitors at least in part through the MAPK cascade in both species (Figures 2 and 3). In S. cerevisiae, signaling through the MAPK cascade is required for calcineurin activation suggesting that PKC signaling regulates crucial responses to ergosterol biosynthesis inhibitors through calcineurin in this species (Figure 6). In C. albicans, Pkc1 and calcineurin independently regulate responses to ergosterol biosynthesis inhibitors via a common target (Figure 7). Inhibition of Pkc1 phenocopies inhibition of calcineurin or Hsp90, reducing drug resistance of clinical isolates of C. albicans (Figure 8 and S6). We establish an additional level of regulatory complexity in the cellular circuitry linking PKC signaling, Hsp90, and calcineurin in that genetic reduction of C. albicans Hsp90 results in destabilization of the terminal MAPK, Mkc1, thereby blocking PKC signaling (Figure 9). This suggests that Hsp90 regulates basal tolerance and resistance to ergosterol biosynthesis inhibitors through Mkc1 in addition to the established connection with calcineurin. Our findings that compromising Pkc1 renders fungistatic drugs fungicidal (Figures 2 and 3) and attenuates virulence of C. albicans (Figure 10) suggest broad therapeutic potential.
The role of PKC signaling in basal tolerance and resistance to drugs targeting the cell membrane expands the repertoire of stress responses that depend upon this signal transduction cascade. In S. cerevisiae, it was previously appreciated that PKC signaling is required for basal tolerance to echinocandins, which target cell wall synthesis [35], [36]. This tolerance requires activation of the terminal MAPK Slt2 to drive Rlm1-dependent transcription of cell wall genes [36]. In C. albicans, the PKC pathway is activated by diverse stresses [52] and works in concert with calcineurin and the high osmolarity glycerol pathway to regulate chitin synthesis, which can enhance tolerance to echinocandins [37], [70]. As is the case with drugs compromising cell wall integrity, drugs targeting the cell membrane activate the terminal MAPK in the PKC cascade (Figure S3A). The role of PKC signaling in tolerance to drugs targeting the cell wall and the cell membrane raises the possibility that induction of cell membrane stress by ergosterol biosynthesis inhibitors could induce cell wall stress indirectly. This is consistent with the thought that the sensors involved in PKC cell wall integrity signaling are receptors that respond to changes in the structure of the cell membrane [71]. Despite the commonalities, the downstream regulation mediating responses to these different stresses diverge. Response to cell wall stress is largely dependent on the transcription factor Rlm1 [36], while regulation of cell membrane stress responses is largely independent of Rlm1. In S. cerevisiae, distinct downstream effectors contribute to tolerance to different ergosterol biosynthesis inhibitors (Figure 4). For fluconazole, the SBF transcription factor (Swi4/Swi6) is of central importance. For fenpropimorph, the SBF complex again is a major determinant with Rlm1 enabling responses important in the absence of SBF. For terbinafine, Swi4 enables tolerance largely independent of SBF and Rlm1 and Cch1-Mid1 mediate responses important in the absence of Swi4. These differences may be due to ergosterol depletion combined with the specific sterol that accumulates when ergosterol biosynthesis is inhibited at different points. In C. albicans, SBF and Cch1-Mid1 confer tolerance to all three ergosterol biosynthesis inhibitors tested suggesting that the point of inhibition of ergosterol biosynthesis has less impact than for S. cerevisiae.
The circuitry downstream of Pkc1 mediating membrane stress responses has been rewired considerably between S. cerevisiae and C. albicans. For S. cerevisiae, deletion of components of the MAPK cascade confers hypersensitivity to ergosterol biosynthesis inhibitors at all temperatures tested (Figure 2 and Figures S1 and S4). For C. albicans, deletion of components of the MAPK cascade confers hypersensitivity to ergosterol biosynthesis inhibitors at 30°C (Figure 3) but not at 35°C (Figure S3B), suggesting that the MAPK cascade is a key mediator of Pkc1-dependent cell membrane stress responses but that alternate downstream effectors play a dominant role in C. albicans at elevated temperature. The importance of alternate downstream effectors of Pkc1 in C. albicans is further emphasized as deletion of PKC1 renders fungistatic ergosterol biosynthesis inhibitors fungicidal, while deletion of MAPK components does not (Figure 3B). Our findings highlight another divergence between the two species. While inhibition of PKC signaling blocks calcineurin activation in response to ergosterol biosynthesis inhibitors in S. cerevisiae (Figure 6), this is not the case in C. albicans. Rather, our results suggest that Pkc1 and calcineurin independently regulate responses to ergosterol biosynthesis inhibitors via a common target in C. albicans (Figure 7). As with PKC signaling, calcineurin and Hsp90 regulate resistance to drugs targeting the cell membrane in both C. albicans and S. cerevisiae, however, they regulate responses to echinocandins in C. albicans but not S. cerevisiae [27], [66], suggesting both conservation and divergence in circuitry governing fungal drug resistance.
The cellular circuitry linking PKC signaling, Hsp90, and calcineurin is complex with multiple levels of regulatory control. On one level is the connection between PKC signaling and calcineurin, which is divergent between the two species. In S. cerevisiae, inhibition of Pkc1 blocks calcineurin activation. The terminal MAPK Slt2 has been found to activate the Cch1-Mid1 high-affinity Ca2+ channel in response to endoplasmic reticulum stress, thereby enabling calcineurin activation [56]. However, we found that deletion of this channel had little impact on drug tolerance (Figure 4), implicating calcineurin regulation via a distinct mechanism. Since inhibition of PKC signaling does not affect calcineurin expression (Figure S5), Slt2 likely regulates calcineurin activation by an alternative mechanism such as through a distinct calcium channel. In C. albicans, cch1Δ/cch1Δ and mid1Δ/mid1Δ mutants share some but not all phenotypes with a calcineurin mutant [59]. Consistent with this, we found that the cch1Δ/cch1Δ and mid1Δ/mid1Δ mutants are almost as sensitive to fluconazole and fenpropimorph as a calcineurin mutant but only show an intermediate sensitivity to terbinafine (Figure 5). In C. albicans, inhibition of PKC signaling did not block calcineurin function (Figure 7). Our findings support a model in which Pkc1 and calcineurin independently regulate responses to ergosterol biosynthesis inhibitors in C. albicans via a common target that remains to be identified. On another level is the connection between Hsp90 and the terminal MAPK. In S. cerevisiae, Hsp90 interacts with activated Slt2 and enables activation of Slt2 targets including Rlm1 [64]. In C. albicans, Hsp90 stabilizes Mkc1 independent of its activation status (Figure 9). Notably, in S. cerevisiae Hsp90 also chaperones Pkc1 [72], though this has yet to be investigated in C. albicans. In contrast to the extensive Hsp90 network in S. cerevisiae [73], [74], we identify Mkc1 as the second Hsp90 client protein in C. albicans. Our work suggests that Hsp90 regulates responses crucial for survival of drug-induced membrane stress through PKC signaling in addition to the established role through calcineurin [24], [25], [27]. These stress responses are less important for resistance due to overexpression of multidrug transporters but are critical for basal tolerance as well as resistance acquired by other diverse mutations. Future experiments will address the relative contribution of calcineurin and PKC signaling via the MAPK cascade in Hsp90-mediated resistance acquired by diverse mechanisms.
Our results highlight the central importance of fungal stress response pathways in enabling survival in the hostile host environment. We demonstrate that while deletion of PKC1 has little impact on growth in vitro, it drastically attenuates the capacity of C. albicans to proliferate in vivo and cause disease (Figure 10). While the attenuated virulence precludes studies to determine if compromising Pkc1 enhances the efficacy of antifungals in vivo, it provides compelling support for targeting fungal Pkc1 as a strategy to control fungal infections. The specific mechanism by which Pkc1 enables virulence has yet to be determined, however, it may operate in part via the downstream MAPK cascade given that C. albicans Mkc1 also contributes to virulence in a murine model [75]. While Mkc1 has little impact on susceptibility to oxidative-mediated killing by phagocytes [76], it is activated by physical contact and is required for invasive hyphal growth and normal biofilm development [77]. The mechanism by which Pkc1 influences virulence is distinct from calcineurin, which is required for C. albicans virulence and survival in the bloodstream [67], [68], [69]. While the calcineurin mutant is unable to survive in serum, the pkc1Δ/pkc1Δ mutant only exhibits an intermediate reduction in viability (Figure 10B), suggesting that Pkc1 regulates virulence via alternate targets. Notably, Pkc1 controls the expression of numerous virulence determinants in the fungal pathogen Cryptococcus neoformans [78], suggesting that Pkc1 governs virulence in phylogenetically diverse fungal species.
Our results suggest that targeting Pkc1 may provide a powerful strategy for the treatment of fungal infectious disease. In vitro, compromising PKC signaling renders laboratory strains and clinical isolates hypersensitive to drugs targeting ergosterol biosynthesis (Figures 1, 2, 3, and 8). These findings coupled with those established by others linking PKC signaling to tolerance of drugs targeting the cell wall [34], [35], [36], [37], suggest that compromising Pkc1 could have therapeutic benefits by enhancing the efficacy of the two most widely deployed classes of antifungals, the azoles and echinocandins. In a murine model of disseminated candidiasis, deletion of PKC1 attenuates C. albicans virulence (Figure 10A), suggesting therapeutic benefit of simply compromising fungal Pkc1 in addition to the benefits of combinatorial therapeutic strategies. Notably, in mammalian cells disruption of PKC signaling impairs tumor progression and drug resistance such that PKC inhibitors have entered clinical trials for the treatment of several human cancers as single or combination therapy agents [79], [80]. The complexity of functions and interactions of mammalian PKC isoforms poses a challenge for the development of anti-cancer therapeutics and current efforts focus on enhancing specificity of action to target specific isoforms. While C. albicans and other fungal pathogens only have one PKC isoform, the therapeutic challenge will lie in achieving fungal selectivity. The successful development of Hsp90 and calcineurin as therapeutic targets for fungal disease faces similar challenges due to complications of inhibiting the function of these key cellular regulators in the host [66], [81], [82]. As a complement to identifying fungal selective pharmacological agents, elucidating the architecture of cellular circuitry governing stress responses, drug resistance, and virulence is poised to reveal promising therapeutic targets as key points of regulatory control that diverged between pathogen and host.
All procedures were approved by the Institutional Animal Care and Use Committee (IACUC) at Duke University according to the guidelines of the Animal Welfare Act, The Institute of Laboratory Animal Resources Guide for the Care and Use of Laboratory Animals, and Public Health Service Policy.
Archives of C. albicans and S. cerevisiae strains were maintained at −80°C in 25% glycerol. Strains were grown in either yeast peptone dextrose (YPD, 1% yeast extract, 2% bactopeptone, 2% glucose) or in synthetic defined medium (SD, 0.67% yeast nitrogen base, 2% glucose) and supplemented with amino acids or in RPMI medium 1640 (Gibco SKU#318000-089, 3.5% MOPS, 2% glucose, pH 7.0) supplemented with amino acids. 2% agar was added for solid media. Strains were transformed following standard protocols. Strains used in this study are listed in Table S1. Strain construction is described in Text S1.
Recombinant DNA procedures were performed according to standard protocols. Plasmids used in this study are listed in Table S2. Plasmid construction is described in Text S1. Plasmids were sequenced to verify the absence of any nonsense mutations. Primers used in this study are listed in Table S3.
A seed culture of the fungus, Mycosphaerella (Cercosporidium) henningsii (IMI 176827) grown on potato dextrose agar (PDA) for two weeks was used for inoculation. Mycelia were scraped out and mixed with 20 mL sterile water and filtered through a 100 µm filter. Absorbance of the spore suspension was measured and adjusted to 0.4. A 2 L Erlenmeyer flask containing 1 L of M-1-D medium [83] was inoculated with 10 mL of the spore suspension and incubated at 160 rpm and 28°C for four weeks. Mycelia were then separated from the supernatant by filtration through Whatman No. 1 filter paper and the filtrate was extracted with EtOAc (6×500 mL). The combined EtOAc extracts were washed with H2O (3×500 mL), dried over anhydrous Na2SO4 and evaporated under reduced pressure to yield a dark brown semi-solid (51.2 mg). A portion (50.0 mg) of the EtOAc extract was separated on preparative TLC (Merck, TLC silica gel 60 F254 precoated Aluminum sheets) using MeOH/Et2O (3∶97) as eluant affording crude cercosporamide (8.1 mg, Rf 0.4). This was further purified by reversed-phase preparative TLC (Merck, TLC Silica gel 60 RP-18 F254 precoated Aluminium sheets) using H2O/CH3CN (3∶7) as eluant to give pure cercosporamide (4.5 mg, Rf 0.5).
Cercosporamide: red crystals; mp 187–188°C (lit. [46] 188–189°C); APCIMS (+)-ve mode, m/z 331 [M+1]+; 1H and 13C NMR spectroscopic data were consistent with those reported in the literature [46]. The structure of cercosporamide is shown in Figure S9.
Antifungal tolerance and resistance were determined in flat bottom, 96-well microtiter plates (Sarstedt) using a modified broth microdilution protocol as described [25], [27]. Dimethyl sulfoxide (DMSO, Sigma Aldrich Co.) was the solvent for fenpropimorph (FN, Sigma Aldrich Co) and terbinafine (TB, Sigma Aldrich Co.); fluconazole (FL, Sequoia Research Products) and micafungin (MF, generously provided by Julia R. Köhler) were dissolved in sterile ddH2O. Geldanamycin (GdA, A.G. Scientific, Inc.) was used to inhibit Hsp90 at the indicated concentrations. Cyclosporin A (CsA, Calbiochem) was used to inhibit calcineurin at the indicated concentrations. Cercosporamide and staurosporine (STS, A.G. Scientific, Inc.) were used to inhibit protein kinase C at the indicated concentrations. DMSO was the solvent for GdA, CsA, STS, and cercosporamide.
Minimum inhibitory concentration (MIC) tests were set up in a total volume of 0.2 ml/well with 2-fold dilutions of FL, FN, TB and cercosporamide. FL gradients were from 256 µg/ml down to 0 with the following concentration steps in µg/ml: 256, 128, 64, 32, 16, 8, 4, 2, 1, 0.5, 0.25. FN gradients were from 25 µg/ml down to 0 with the following concentration steps in µg/ml: 25, 12.5, 6.25, 3.125, 1.5625, 0.78125, 0.390625, 0.1953125, 0.09765625, 0.04882813, 0.02441406. TB gradients were from 250 µg/ml with the following concentration steps in µg/ml: 250, 125, 62.5, 31.25, 15.625, 7.8125, 3.90625, 1.953125, 0.9765625, 0.48828125, 0.24414063. Cercosporamide gradients were from 100 µg/ml with the following concentration steps in µg/ml: 100, 50, 25, 12.5, 6.25, 3.125, 1.5625, 0.78125, 0.390625, 0.1953125, 0.09765625. Cell densities of overnight cultures were determined and dilutions were prepared such that ∼103 cells were inoculated into each well. Plates were incubated in the dark at 30°C or 35°C for the period of time indicated in the figure legend, at which point plates were sealed with tape and re-suspended by agitation. Absorbance was determined at 600 nm using a spectrophotometer (Molecular Devices) and corrected for background from the corresponding medium. Each strain was tested in duplicate on at least 3 occasions. MIC data was quantitatively displayed with color using the program Java TreeView 1.1.1 (http://jtreeview.sourceforge.net).
Checkerboard assays were set up in a total volume of 0.2 ml/well with 2-fold dilutions of cyclosporin A across the x-axis of the plate and 2-fold dilutions of STS across the y-axis of the plate. STS gradients were from 0.5 µg/ml to 0 in the following concentrations steps in µg/ml: 0.5, 0.25, 0.125, 0.0625, 0.03125, 0.015625, 0.0078125. CsA gradients were from 48 µg/ml down to 0 in the following concentration steps in µM: 48, 24, 12, 6, 3, 1.5, 0.75, 0.375, 0.1875, 0.09375, 0.046875. Plates were inoculated and growth was measured as with MIC tests. To test for synergy, the fractional inhibitory concentration (FIC) was calculated as follows: [(MIC80 of drug A in combination)/(MIC80 of drug A alone)] + [(MIC80 of drug B in combination)/(MIC80 of drug B alone)]. Values of ≤0.5 indicate synergy, those of >0.5 but <2 indicate no interaction and those ≥2 indicate antagonism.
Strains were grown overnight to saturation in indicated media and cell concentrations were determined based on cell counts using a hemacytometer (Hausser Scientific). Five-fold serial dilutions of cell suspensions starting at indicated concentrations (105 or 107cells/ml) were performed in sterile ddH2O or sterile phosphate buffered saline. Cell suspensions were spotted onto indicated media using a spotter (Frogger, V&P Scientific, Inc). Plates were photographed after 3 days in the dark at indicated temperature.
For S. cerevisiae, MIC assays with two-fold dilutions of FL, FN, or TB were performed in SD as described above. For FL the gradients were from 256 µg/ml down to 0 with the following concentration steps in µg/ml: 256, 128, 64, 32, 16. FN gradients were from 100 µg/ml down to 0 with the following concentration steps in µg/ml: 100, 50, 25, 12.5, 6.25. TB gradients were from 250 µg/ml with the following concentration steps in µg/ml: 250, 125, 62.5, 31.25, 15.62. Plates were incubated for two days at 35°C. Cells from the MIC assay were spotted onto solid YPD medium and incubated at 30°C for two days before they were photographed.
For C. albicans, MIC assays with FL, FN, or TB were performed in YPD as described above with the following modification; four-fold dilutions of each drug were tested. For FL the gradients were from 256 µg/ml down to 0 with the following concentration steps in µg/ml: 256, 64, 16, 4, 1. FN gradients were from 25 µg/ml down to 0 with the following concentration steps in µg/ml: 25, 6.25, 1.5625, 0.390625, 0.09765625. TB gradients were from 250 µg/ml with the following concentration steps in µg/ml: 250, 62.5, 15.625, 3.90625, 0.9765625. Plates were incubated for two days at 35°C. Cells from the MIC assay were spotted onto solid YPD medium and incubated at 30°C for two days before they were photographed.
S. cerevisiae cultures were grown overnight at 25°C in SD medium supplemented for auxotrophies. Cells were diluted to OD600 of 0.05 and were either left untreated or were treated with FL (16 µg/ml), FN (1 µg/ml), or TB (25 µg/ml) for 24 hours at 25°C. When STS was used as an inhibitor in the assay, cultures were grown overnight in SD at 25°C and diluted to OD600 of 0.05 in SD with or without STS (2.5 µg/ml) for 24 hours at 25°C. Cells were then diluted to OD600 of 0.05 in SD with or without STS and with or without FL (32 µg/ml) for an additional 24 hours at 25°C. Cells were harvested, washed, protein was extracted, and protein concentrations were determined by Bradford analysis as described [27]. Protein samples were diluted to the same concentration and β-galactosidase activity was measured using the substrate ONPG (O-nitrophenyl-β-D-galactopyranosidase, Sigma Aldrich Co.) as described [27]. β-galactosidase activity is given in units of nanomoles ONPG converted per minute per milligram of protein. Statistical significance was evaluated using GraphPad Prism 4.0.
For the Mkc1 activation assay, yeast cultures were grown overnight in YPD at 30°C. In the morning, cells were diluted to OD600 of 0.2 in 50 mL YPD and were grown to mid-log (∼3 hours) at 30°C and then cultures were split into 5×10 mL cultures and were either left untreated or were treated with FL (8 µg/mL), FN (1 µg/mL), MF (30 ng/mL), or TB (25 µg/ml) for 2 hours at 30°C. Cells were harvested by centrifugation at 1308×g for 10 minutes at 4°C and were washed with sterile cold phosphate buffered saline (PBS). Cell pellets were resuspended in lysis buffer containing 50 mM HEPES pH 7.4, 150 mM NaCl, 5 mM EDTA, 1%Triton ×100, 50 mM NaF, 10 mM Na3VO4, 1 mM PMSF, and protease inhibitor cocktail (complete, EDTA-free tablet, Roche Diagnostics).
For the Mkc1 destabilization assay, cultures were grown overnight in YPD at 30°C. In the morning, cells were diluted to OD600 of 0.2 in 10 mL YPD with or without doxycycline (20 µg/mL; BD Biosciences) and left at 30°C for 24 hours. Cells were diluted once again to OD600 of 0.2 in the same treatment conditions as overnight and were grown at 30°C until mid-log phase (∼4 hours). Doxycycline reduces the growth rate of strains with the repressible promoter driving expression of the only HSP90 allele but does not affect stationary phase cell density [65]. Cells were then treated with 50 µg/mL TB for 3 hours at 30°C to elicit phosphorylation of Mkc1. Cells were harvested after TB treatment at 1308×g at 4°C and washed with ice-cold ddH2O. Cell pellets were flash frozen in liquid N2, resuspended in lysis buffer (50 mM HEPES pH 7.4, 150 mM NaCl, 5 mM EDTA, 1% Triton ×100, 100 mM NaF, 20 mM Na3VO4, 1 mM PMSF and protease inhibitor cocktail complete, EDTA-free tablet, Roche Diagnostics).
Cells suspended in lysis buffer were mechanically disrupted by adding acid-washed glass beads and bead beating for 1 minute for six cycles with 1 minute on ice between each cycle. Protein concentrations were determined by Bradford analysis. Protein samples were mixed with one-sixth volume of 6× sample buffer containing 0.35 M Tris-HCl, 10% (w/w) SDS, 36% glycerol, 5% β-mercaptoethanol, and 0.012% bromophenol blue for SDS-PAGE. Samples were boiled for 5 minutes and then separated by 10% SDS-PAGE. Protein was electrotransferred to PVDF membrane (Bio-Rad Laboratories, Inc.) and blocked with 5% skimmed milk in PBS with 0.1% tween or 5% bovine serum albumin in phosphate buffered saline with 0.1% tween. Blots were hybridized with antibodies against CaHsp90 (1∶10000 dilution, generously provided by Brian Larsen, [84]), histone H3 (1∶3000 dilution; Abcam ab1791), His6 (1∶10, P5A11, generously provided by Elizabeth Wayner) and phospho-p44/42 MAPK (Thr202/Tyr204) (1∶2000, Cell Signaling).
To monitor gene expression changes in response to FL treatment in S. cerevisiae, cells were grown overnight in SD supplemented for auxotrophies at 30°C. Cells were diluted to OD600 of 0.1 in SD and grown for 2 hours in duplicate at 25°C. After 2 hours of growth 16 µg/mL FL was added to one of the two duplicate cultures and left to grow for an additional 4 hours at 25°C. Cell pellets were frozen at −80°C immediately.
To monitor gene expression changes in response to FL treatment in C. albicans, cells were grown overnight in YPD at 30°C. Cells were diluted to OD600 of 0.1 in YPD and grown for 2 hours in duplicate at 35°C. After 2 hours of growth 16 µg/mL FL was added to one of the two duplicate cultures and left to grow for an additional 4 hours at 35°C. Cell pellets were frozen at −80°C immediately.
To monitor MKC1 transcript levels in response to decreased levels of Hsp90, cultures were grown overnight in YPD at 30°C. In the morning, cells were diluted to OD600 of 0.2 in 10 mL YPD with or without 20 µg/mL doxycycline (BD Biosciences) and left at 30°C for 24 hours. The next morning, cells were diluted once again to OD600 of 0.2 in the same treatment conditions and were grown at 30°C until mid-log phase (∼4 hours). Cell pellets were collected and immediately frozen at −80°C.
RNA was isolated using the QIAGEN RNeasy kit and RNAase-free DNase (QIAGEN), and cDNA synthesis was performed using the AffinityScript cDNA synthesis kit (Stratagene). PCR was performed using SYBR Green JumpStart Taq ReadyMix (Sigma-Aldrich Co.) with the following cycling conditions: 94°C for 2 minutes, 94°C for 15 seconds, 60°C for 1 minute, 72°C for 1 minute, for 30 or 40 cycles. All reactions were performed in triplicate, using primers for the following genes: CaGPD1 (oLC752/753), CaHSP90 (oLC754/755), ScACT1 (oLC1015/1016), ScCNA1 (oLC1286/1287), ScCNA2 (oLC1288/1289), ScCNB1(oLC1290/1291), CaCNB1 (oLC1292/1293), CaCNA1 (oLC1294/1295), ScCRZ1 (oLC1328/1329), CaCRZ1(oLC1330/1331), CaMKC1(oLC1332/1333), CaPLC3(oLC1432/1433), and CaUTR2(oLC1434/1435). Data were analyzed using iQ5 Optical System Software Version 2.0 (Bio-Rad Laboratories, Inc.). Statistical significance was evaluated using GraphPad Prism 4.0.
Inoculum was prepared as described [20], [27], [65]. Cultures were started from frozen stocks onto Sabouraud dextrose agar plates and incubated at 35°C for 48 hours. Colonies were suspended in sterile pH 7.4 PBS, centrifuged at 324×g for 5 minutes, washed with sterile PBS one time and diluted to the desired concentration as verified by counting on a Neubauer hematocytometer as well as by serial dilution and culture. Male CD1 mice (Charles River Laboratories, Wilmington, MA) age 8 weeks (weight 30–34 g) were infected via the tail vein with 100 µL of a 1×106 CFU/mL suspension of the wild type strain (CaLC239, 1×105 CFU per mouse, n = 9 mice), an inoculum previously determined to produce morbidity but not mortality when using C. albicans strain SC5314 at 4 days following tail vein injection (Zaas et al. unpublished data). We observed discordance between cell counts and CFU measurements for the pk1cΔ/pkc1Δ mutant, such that CFU values were ∼50% lower than expected; thus, inocula for the pk1cΔ/pkc1Δ mutant were prepared at higher concentrations based on cell counts and the effective concentrations in CFUs were confirmed by dilution plating. For infection with the pk1cΔ/pkc1Δ mutant, we used an inoculum equivalent to that for the wild type (1×105 CFU, n = 8 mice) as well a 10-fold and 100-fold increase in inoculum (1×106, n = 11 mice and 1×107 CFU, n = 8 mice). Mice were observed three times daily for signs of illness and weighed daily. At day 4 following injection, mice were sacrificed using CO2 asphyxiation and the left kidney was removed aseptically, placed in sterile PBS, homogenized using a FastPrep 120 (QBiogene) using 0.5 mm zirconium beads (Biospec, Inc.) for 1 minute and serial dilutions plated for determination of kidney fungal burden. The CFU values in kidneys were expressed as CFU/g of tissue and log-transformed. Statistical significance was evaluated using GraphPad Prism 4.0.
S. cerevisiae: PKC1 (852169); HSC82 (855224); HSP82 (855836); CNA1 (851153); CNA2 (854946); CNB1 (853644); ERG11 (856398); ERG3 (850745); RHO1 (856294); BCK1 (853350); MKK1 (854406); MKK2 (855963); SLT2 (856425); SWI4 (856847); SWI6 (850879); ERG2 (855242); ERG24 (855441); ERG1 (853086); RLM1 (856016); CCH1 (853131); MID1 (855425); CRZ1 (855704); PDR5 (854324); PDR1 (852871); PDR3 (852278); ACT1 (850504).
C. albicans PKC1 (3635298); HSP90 (3637507); CNA1 (3639406); CNB1 (3636463); MKC1 (3639710); ERG11 (3641571); ERG3 (3644776); RHO1 (3642564); BCK1 (3641434); MKK2 (3645580); MDR1 (3639260); ERG2 (3639416); ERG24 (3648198); ERG1 (3646509); CEK1 (3642789); CEK2 (3642459); RLM1 (3635703); SWI4 (3645507); SWI6 (3634957); CCH1 (3639950); MID1 (3647441); CRZ1 (3641722); PLC3 (3635941); UTR2 (3636747); CDR1 (3635385); GPD1 (3643986).
|
10.1371/journal.ppat.1006058 | Modulation of Re-initiation of Measles Virus Transcription at Intergenic Regions by PXD to NTAIL Binding Strength | Measles virus (MeV) and all Paramyxoviridae members rely on a complex polymerase machinery to ensure viral transcription and replication. Their polymerase associates the phosphoprotein (P) and the L protein that is endowed with all necessary enzymatic activities. To be processive, the polymerase uses as template a nucleocapsid made of genomic RNA entirely wrapped into a continuous oligomer of the nucleoprotein (N). The polymerase enters the nucleocapsid at the 3’end of the genome where are located the promoters for transcription and replication. Transcription of the six genes occurs sequentially. This implies ending and re-initiating mRNA synthesis at each intergenic region (IGR). We explored here to which extent the binding of the X domain of P (XD) to the C-terminal region of the N protein (NTAIL) is involved in maintaining the P/L complex anchored to the nucleocapsid template during the sequential transcription. Amino acid substitutions introduced in the XD-binding site on NTAIL resulted in a wide range of binding affinities as determined by combining protein complementation assays in E. coli and human cells and isothermal titration calorimetry. Molecular dynamics simulations revealed that XD binding to NTAIL involves a complex network of hydrogen bonds, the disruption of which by two individual amino acid substitutions markedly reduced the binding affinity. Using a newly designed, highly sensitive dual-luciferase reporter minigenome assay, the efficiency of re-initiation through the five measles virus IGRs was found to correlate with NTAIL/XD KD. Correlatively, P transcript accumulation rate and F/N transcript ratios from recombinant viruses expressing N variants were also found to correlate with the NTAIL to XD binding strength. Altogether, our data support a key role for XD binding to NTAIL in maintaining proper anchor of the P/L complex thereby ensuring transcription re-initiation at each intergenic region.
| Three proteins, the polymerase L, the phosphoprotein P and the nucleoprotein N, interplay to ensure transcription and replication of measles virus, a member of the Paramyxoviridae family. A regular array of nucleoprotein shields the viral genomic RNA. The resulting nucleocapsid constitutes the template of RNA synthesis used by the polymerase complex made of L and P, with the latter ensuring L anchoring onto the nucleocapsid. We herein report a correlation between the binding affinity of the C-terminal X domain of P (XD) and the intrinsically disordered C-terminal tail of N (NTAIL), the ability to reinitiate the transcription at the intergenic regions and the accumulation rate of viral transcripts from recombinant viruses. We therefore propose that the NTAIL/XD interaction contributes to maintaining the polymerase complex anchored onto the nucleocapsid while ending the upstream transcript and re-initiating the downstream transcript at every intergenic region. As such, the NTAIL/XD interaction strength must be controlled so as to keep the viral transcription gradient within an optimal efficiency window. The conservation of this mode of interaction between the viral P and N proteins in many members of the Paramyxoviridae family reflects one of the major evolution constraints to which their polymerase machinery is subjected.
| Measles virus (MeV), a member of the Morbillivirus genus, belongs to the Paramyxoviridae family of the Mononegavirales order [1]. These viruses possess a non-segmented RNA genome of negative polarity that is encapsidated by the nucleoprotein (N) to form a helical nucleocapsid. Not only does N protect viral RNA from degradation and/or formation of viral dsRNA, but it also renders the latter competent for transcription and replication. Indeed, the viral polymerase cannot processively transcribe nor replicate RNA unless the viral genome is encapsidated by the N protein within a helical nucleocapsid [2,3]. Transcription and replication are ensured by the RNA-dependent RNA polymerase complex made of the large protein (L) and the phosphoprotein (P), with P serving as an essential tethering factor between L and the nucleocapsid. The complex made of RNA and of the N, P and L proteins constitutes the replication machinery. In order to perform messenger RNA synthesis, the polymerase has not only to bind to the 3’ transcription promoter, but also to re-initiate the transcription of downstream genes upon crossing each intergenic region (IGR). Following polyadenylation, which serves as gene end (GE) signal, the polymerase proceeds over three nucleotides (3’-GAA-5’ or 3’-GCA-5’) without transcribing them and then restarts transcription upon recognition of a downstream gene start (GS) signal.
Within infected cells, N is found in a soluble, monomeric form (referred to as N0) and in a nucleocapsid assembled form [4]. Following synthesis, the N protein requires chaperoning by the P protein so as to be maintained in a soluble and monomeric form. The P N-terminal region (PNT) binds to the neosynthesized N protein thereby simultaneously preventing its illegitimate self-assembly and yielding a soluble N0P complex the structure of which have been characterised for MeV [5] as well as for four other members of the Mononegavirales order [6,7,8,9]. N0P is used as the substrate for the encapsidation of the nascent genomic RNA chain during replication [10], (see also [4,11,12,13] for reviews on transcription and replication). In its assembled homopolymeric form or nucleocapsid, N also makes complexes with either isolated P or P bound to L, with all these interactions being essential for RNA synthesis by the viral polymerase [14,15,16]. Throughout the Mononegavirales order, P and P+L binding to the nucleocapsid is mediated by interaction of the C-terminal region of P with either the C-terminal tail of N (Paramyxoviridae members), or to the N-terminal globular moiety (or core) of N (see [11,17] for review).
The MeV N protein consists of a structured N-terminal moiety (NCORE, aa 1–400), and a C-terminal domain (NTAIL, aa 401–525) [18,19] that is intrinsically disordered, i.e. it lacks highly populated secondary and tertiary structure under physiological conditions of pH and salinity in the absence of a partner (for a recent review on intrinsically disordered proteins see [20]). While NCORE contains all the regions necessary for self-assembly and RNA-binding [10,21,22] and a binding site for an α-MoRE located at the N terminus of the P protein, NTAIL is responsible for interaction with the C-terminal X domain (XD, aa 459–507) of P [11,18,21,23,24,25,26,27] (Fig 1A).
NTAIL binding to XD triggers α-helical folding within a molecular recognition element [28,29] of α-helical nature (α-MoRE, aa 486–502) located within one (Box2, aa 489–506) out of three conserved NTAIL regions [18,24,27,30,31,32,33,34,35,36,37]. XD-induced α-helical folding of NTAIL is not a feature unique to MeV, being also conserved within the Paramyxoviridae family [38,39,40,41,42,43]. XD consists of a triple α-helical bundle [27,34,44], and binding to the α-MoRE leads to a pseudo-four-helix arrangement that mainly relies on hydrophobic contacts [18,27,44]. The α-MoRE of NTAIL is partly preconfigured as an α-helix prior to binding to XD [31,32,35,37,45] and adopts an equilibrium between a fully unfolded form and four partly helical conformers [37]. In spite of this partial pre-configuration, NTAIL folds according to a “folding after binding mechanism” [45,46]. Previous mutational studies showed that Box2 is poorly evolvable in terms of its binding abilities towards XD, in that amino acid substitutions therein introduced lead to a dramatic drop in the binding strength, as judged from a protein complementation assay (PCA) based on split-GFP reassembly (gfp-PCA) [47]. In particular, substitutions within the N-terminal region of Box2 (aa 489–493) and at position 497 were found to lead to the most dramatic drops in the interaction strength [47].
In the context of the viral nucleocapsid, NTAIL points towards the interior of the latter and then extrafiltrates through the interstitial space between NCORE moieties, with the first 50 residues (aa 401–450) being conformationally restricted due to their location between successive turns of the nucleocapsid [37]. The NTAIL region spanning residues 451–525 and encompassing the α-MoRE is, by contrast, exposed at the surface of the viral nucleocapsid and thus accessible to the viral polymerase. Binding of XD to NTAIL has been proposed to ensure and/or contribute to the recruitment of the viral P/L polymerase complex onto the nucleocapsid template. However, its precise function has remained enigmatic so far with reports of apparent conflicting observations. From the analysis of four NTAIL variants it was concluded that the accumulation rate of primary transcripts is rather insensitive to a drop in the apparent XD to NTAIL affinity [26], while an XD variant showing a 1.7 times stronger interaction with NTAIL was associated with a 1.7-fold reduction in the accumulation rate of viral transcripts [48]. Furthermore, deletions studies of NTAIL have indicated that the interaction between XD and NTAIL may be dispensable for transcription and replication [49].
In the present work, we further investigate the molecular mechanisms by which substitutions in critical positions of NTAIL previously identified by a random approach [47] affect the viral polymerase activity. We did so by combining biochemical studies and molecular dynamics (MD) simulations on one hand with functional studies that made use of minigenomes and recombinant viruses on the other hand. Results identify positions 491 and 497 as the most critical in terms of both binding affinities and functional impact. In addition, thanks to the availability of a newly conceived minigenome made of two luciferase reporter genes, with the second one being conditionally expressed via RNA edition of its transcript by the viral polymerase, we could quantify the efficiency of transcription re-initiation after polymerase scanning through each of the five IGRs of MeV genome and on an elongated un-transcribed IGR (UTIGR). A low NTAIL-XD affinity was found to be associated to a reduced ability of N to support expression of luciferase from the second gene. Furthermore, in infected cells, the accumulation rate of primary transcripts and transcript ratios were found to correlate with the equilibrium dissociation constant (KD) of the NTAIL/XD pair. Altogether obtained data argue for a key role of the NTAIL/XD interaction in transcription re-initiation at each intergenic region.
In a previous random mutagenesis study that made use of a PCA based on split-GFP re-assembly (gfp-PCA) [47], we identified NTAIL variants that either decrease or increase the interaction strength towards XD [47]. Variant MX208, which bears the D437V, P485L and L524R substitutions that are all located outside the α-MoRE, is an example of the latter group. We previously reported the generation and assessment of binding properties by gfp-PCA of six single-site variants (R489Q, R490S, S491L, A492T, D493G and R497G) bearing each a unique substitution within the α-MoRE [47]. Here, we additionally designed and generated the MXSF variant, which bears D437V, R439S, P456S and P485L substitutions that are all found in variants displaying an increased fluorescence [47]. Gfp-PCA in E. coli showed that the binding strength of these variants towards XD is scattered over a wide range, with the S491L and R497G variants showing the lowest interaction and with variant MXSF displaying interaction strength only moderately higher than wt NTAIL (Fig 1B). Incidentally, this latter finding indicates that the effects of the substitutions are not cumulative.
We then sought at assessing to which extent results afforded by the split-GFP assay in E. coli cells reflect NTAIL/XD binding occurring in the natural host cells of MeV. To this end, the interaction between XD and NTAIL variants was measured using the split-luciferase reassembly assay [50]. This technique is based on the same principle as the split-GFP reassembly assay. The reporter (i.e. Gaussia princeps luciferase) and the measured parameter (luminescence) are however different, and the assay is performed in human cells. Moreover, contrary to the split-GFP reassembly assay where reporter reassembly is irreversible, in the split-luciferase assay (glu-PCA), association of the two luciferase fragments is reversible. As such, while the measured parameter in the former assay is dominated by the kon, the measured parameter in the latter assay does reflect the equilibrium between a kon and a koff and hence a true KD. A significant correlation was obtained between the two PCA methods (Fig 1C), a finding that provides additional support for the significance of the observed differences in binding strength among variants. Furthermore, a significant correlation was also observed when comparing binding strengths as obtained using monomeric constructs (i.e. NTAIL/XD) and binding strengths obtained using their natural multimeric counterparts, i.e. P multimerization domain (PMD)-XD (P303-507) and full-length mutated N protein constructs (Fig 1D). The rationale for using P multimeric constructs devoid of the N-terminal region (PNT) was to eliminate the binding site to NCORE located within PNT and involved in P chaperoning of N protein to form N0P complexes [5] (see Fig 1A for depicting scheme). Importantly, all N variants accumulated in cells in similar amounts (S1 Fig) indicating that variations in the level of reconstituted Gaussia luciferase likely reflects variations in NTAIL to XD binding strength.
In order to characterize Box2 variants (Fig 2A) at the biochemical level, we expressed and purified six α-MoRE variants of NTAIL as N-terminally hexahistidine tagged proteins. All NTAIL variants were purified to homogeneity from the soluble fraction of the bacterial lysate through immobilized metal affinity chromatography (IMAC) followed by size exclusion chromatography (SEC) (Fig 2B). The identity of all purified proteins were checked and confirmed by mass spectrometry. Even if their molecular mass is ~16 kDa, they all migrate on a denaturing gel with an apparent molecular mass of approximately 20 kDa (Fig 2B, inset). This aberrant electrophoretic migration has been systematically observed for all NTAIL variants reported so far [26,30,31,32,46] including wt NTAIL [19]. This anomalous migration is frequently observed in IDPs and is due to a high content in acidic residues [51] and/or a large extension in solution [43].
All NTAIL variants, including wt NTAIL, have the same SEC elution profile (Fig 2B). In particular, they are all eluted with an apparent molecular mass higher than expected and typical of a premolten globule (PMG) state [52], as already observed in the case of wt NTAIL [19]. Thus, the amino acid substitution(s) causes little (if any) effect on the hydrodynamic volume sampled by the protein.
Analysis of the secondary structure content of the NTAIL variants by far-UV circular dichroism (CD) shows they are all disordered, as judged from their markedly negative ellipticity at 200 nm (Fig 2C). In addition, they are all similarly able to gain α-helicity in the presence of 20% 2,2,2 trifluoroethanol (TFE) (Fig 2D), as already observed for wt NTAIL [19]. All variants have an estimated α-helical content similar (within the error bar) to that of wt NTAIL, with the only exception of variant R489Q that exhibits a lower α-helicity both in the absence and in the presence of TFE (Fig 2E). Thus, most of the amino acid substitutions cause little (if any) effect on the overall secondary structure content and folding abilities of NTAIL.
The binding abilities of the NTAIL variants, including wt NTAIL, were assessed using isothermal titration calorimetry (ITC). To this end, the purified NTAIL proteins were loaded into the sample cell of an ITC200 microcalorimeter and titrated with wt XD. For each variant, two independent experiments were carried out. Fig 3 shows, for each variant, one representative ITC curve along with the relevant binding parameters. The XD/NTAIL molar ratios achieved at the end of the titration were 2.0 (wt, R489Q, A492T, D493G), 2.5 (R490S) or 3.0 (R497G) (Fig 3). The data, following integration and correction for the heats of dilution, were fitted with a standard model allowing for a set of independent and equivalent binding sites.
Consistent with the unfavorable entropic contribution associated to the disorder-to-order transition that takes place upon NTAIL binding to XD, whenever binding parameters could be determined, they revealed a decrease in entropy, with a ΔS ranging from -13 to -29.5 cal mol-1 deg-1 (Fig 3). Binding reactions were all found to be enthalpy-driven, with ΔH values in the same order of magnitude and ranging from -10.9 to -14.5 kcal/mol (Fig 3). The estimates for the model parameters of the wt NTAIL/XD pair were found to be in very good agreement with those recently reported [46]. The estimates for binding parameters of variants R489Q, A492T and D493G yielded equilibrium dissociation constant (KD) very close to that observed for wt NTAIL, indicating that these substitutions poorly affect the interaction (Fig 3). On the other hand, the R490S substitution resulted in a 7-fold decrease in the binding affinity (KD of 20 μM). The decrease in affinity was even further pronounced (KD of 44 μM) in the case of the R497G variant, although the interaction remained measurable (Fig 3). In the case of the S491L variant the interaction strength was below the ITC detection limit and thus KD could not be estimated (Fig 3).
The n values for the A492T/XD and D493G/XD binding pairs were found to deviate from unit, a behaviour that is not unusual and that has been already observed with single-site tryptophan variants [46] and that may arise from relatively poorly defined baselines. In light of all the numerous previous studies [18,19,27,30,31,32,33,34,35,36,37] showing that NTAIL and XD form a 1:1 complex, these deviations were not taken to be significant.
We next focused on how binding affinities obtained by ITC correlate with binding strengths inferred from split-GFP and split-luciferase reassembly assays. In fact, although it has already been established that the higher the fluorescence the higher the interaction strength [53], no attempts were done at establishing which type of relationship exists between KD values and fluorescence or luminescence values. As shown in Fig 4, we found a significant correlation between fluorescence or luminescence values obtained by gfp-PCA [47] and glu-PCA and the ln of KD values (p = 0.02 in both cases). Although this finding needs to be confirmed on a larger set of data points, it lays the basis for the possibility of inferring KD values directly from fluorescence or luminescence values.
Notably, if the results obtained by gfp-PCA (see Fig 1B) pointed out similarly low interaction strengths for variants S491L and R497G, ITC studies yielded a different profile. Indeed, while no interaction could be effectively detected for the S491L/XD pair, the KD could be measured for variant R497G (see Fig 3). Using the empirically determined equation relating luminescence and KD values (Fig 4), the KD of the S491L/XD pair was estimated to be 85 ± 33 μM, a value consistent with our inability to detect the interaction by ITC. Indeed, an interaction characterized by a KD of approximately 100 μM could escape detection unless extremely high (and hardly achievable) protein concentrations are used (typically 1 mM NTAIL/10 mM XD) [54].
Altogether, obtained results confirmed that not all Box2 residues are equivalent in terms of their role in NTAIL/XD complex formation. In particular, while substitutions at positions 489, 490, 492 and 493 have a slight to moderate impact, substitutions at positions 497 and 491 drastically affect complex formation without having a strong impact on the overall α-helicity. The role of Box2 residues in complex formation follows the order 491>497>490, reflecting either the orientation of side chains towards the partner (residues 490 and 491) or involvement in stabilizing interactions with XD residue Tyr480 in spite of solvent exposure (residue 497), as already proposed (Fig 2A) [47].
In order to further investigate the mechanisms by which residues Ser491 and Arg497 stabilize the NTAIL/XD complex, we performed MD simulations in aqueous solvent using the CHARMM force field [55]. MD simulations were carried out starting from the X-ray structure of the XD/α-MoRE complex [44] or from the in silico generated XD/α-MoRE S491L and XD/α-MoRE R497G models. In the case of the S491L variant, the three most favorable orientations of the side chains were generated. We first assessed the dynamical stability of the complexes. For this purpose, we analyzed the root-mean square deviation (RMSD) of the Cα atoms with respect to the initial structure as a function of time for the three complexes (i.e. wt, S491L and R497G) (S1 Table). The RMSD values showed very little variations between the different constructs during the time course of the 50 ns simulations (S1 Table). The average RMSD for XD and for the α-MoRE were approximately 0.8 and 0.5 Å, respectively, indicating structural stability of each domain during the simulations (S1 Table). The relative orientation of the α-MoRE compared to XD was also assessed and revealed slightly higher average RMSD values for the two variants due to small local rearrangements of the structures to adapt to the substitutions. However, RMSD fluctuations were in the same order of magnitude. Secondary structure analyses of wt and mutated complexes confirmed that all α-helices are conserved during the whole trajectories. Overall, the different systems were stable during the whole simulation. Since the orientation of the side-chain of L491 showed no impact on the behavior of the complex during the MD simulation, only one conformer was selected in the rest of the study.
Although the association between XD and NTAIL is essentially driven by hydrophobic contacts, the two partners also interact through hydrogen bonds that are thus expected to play a role in the binding affinity. Two intramolecular hydrogen-bond interactions are present in the crystallographic structure of the complex (Table 1). These interactions involve the side-chain of NTAIL residue Ser491 and side-chain of Asp493 and main-chain of Lys489 from XD. These interactions are preserved in the simulations of the wt and R497G complex (Table 1 and S2 Fig). Due to the absence of the polar OH group in leucine, hydrogen bonds involving the OH group of Ser491 were lost in the simulations of the S491L complex. Three additional hydrogen bonds that are not present in the X-ray structure were observed in the MD trajectories of the wt complex (Table 1 and S2 Fig). Two of them involve the side-chain of Asp487 from XD and side-chains of either Arg490 or Arg497 of NTAIL. Only the former was also observed in the simulations of both variants (Table 1 and S2 Fig). The third one was formed between the side-chains of Lys489 from XD and Asp487 from NTAIL and was detected in the simulation of the three complexes with the two variants exhibiting even a higher frequency (Table 1 and S2 Fig). In addition, a water-mediated hydrogen bond could be identified between the side-chains of Tyr480 of XD and Arg497 of NTAIL, 55 and 41 percent of the time in the wt and S491L complex, respectively. This interaction was not maintained with the same water molecule throughout simulation. However, when a water molecule moved away from this site it was almost immediately replaced by another water molecule. This interaction could not occur in the R497G complex and was not compensated by another interaction. The presence of this water-mediated interaction correlates with the stabilization of the aromatic ring of Tyr480. The side-chain of Tyr480 was found in almost only one conformation corresponding to a χ2 angle (CA-CB-CG-CD) of approximately -130° in both wt and S491L complexes, whereas in the R497G complex, the ring oscillates between 2 conformations (50 and -130°) corresponding to a 180° rotation. Although the position of this water molecule in the crystal structure cannot be estimated with precision because the molecule is poorly defined in the electron density, the fact that a water molecule is systematically observed at this position during the simulation argues for its critical role in stabilizing the Arg497-Tyr480 interaction. That water molecules can play crucial roles in stabilizing protein-protein interactions has been widely documented [56].
To further investigate the importance of the effect of the substitutions on the binding affinity, additional MD simulations were carried out using the free energy perturbation (FEP) method (see details of the method in the Materials and Methods section). The calculations were based on the thermodynamic cycle shown in S3 Fig which allowed us to estimate the impact of an amino acid substitution on the binding energy by measuring the ΔΔG between the wt and mutated complexes at 300K. Replacement of Ser491 of NTAIL by Leu led to an average binding free energy change ranging from 3.22 to 3.91 kcal.mol-1. These ΔΔG values correspond to a 200-fold to 700-fold reduction in binding affinities for the S491L variant which is compatible, although a bit more pronounced, with the KD calculated for this variant using the empirically determined equation between luminescence and KD values (see above and Fig 4). In a similar manner, substitution of R497 with Gly led to ΔΔG values ranging from 1.29 to 1.85 kcal.mol-1. This corresponds to a 10 to 20-fold reduction of binding affinity which nicely correlates with the KD reduction-fold as measured by ITC.
The dissociation of the XD/NTAIL complex cannot be observed during the time course of free MD simulations. To obtain more insights into the dissociation process, we therefore performed simulations using adaptive biasing force (ABF), a method that allows overcoming barriers of the free-energy landscape [57]. The center of geometry between the two partners was selected as ordering parameter and both proteins were allowed to diffuse reversibly along this reaction coordinate during the different stages of the simulations (no average force was exerted along the ordering parameter). The free energy profiles of the wt and mutated complexes are shown Fig 5A. The global minimum corresponds to a distance around 11.3 Å, very close to the distance observed in the X-ray structure (11.03 Å). Analysis of the wt complex reveals that the dissociation between the two partners proceeds from the C-terminal part of NTAIL corresponding to the more hydrophobic residues (Fig 5B and S2 Fig). The final step of the dissociation corresponds to the disruption of hydrogen bonds between Ser491 of NTAIL and Lys489 and Asp493 of XD. The R497G complex exhibits an energy profile similar to that of the wt complex with slightly lower energy values indicating a lower resistance against disruption. In the case of the S49IL complex, the disruption can occur from either end of the α-helix of NTAIL depending on the trajectory. This behavior can be explained by the loss of hydrogen bonding with Lys489 and Asp493 of XD. As a consequence, the energy profile is profoundly affected and this variant shows less resistance toward disruption.
Altogether, these data provide a mechanistic basis illuminating the critical role played by NTAIL residues Ser491 and Arg497 in stabilizing the NTAIL-XD complex.
In order to investigate the functional consequences of attenuating the interaction between NTAIL and XD we tested the ability of each N variant to support the expression of a reporter gene from a minigenome rescued into a functional nucleocapsid by cotransfecting a plasmid coding for the minigenome under the T7 promoter together with P and L expression plasmid [58]. To take into account the transcription re-initiation at IGRs, we conceived and built new dual-luciferase minigenomes coding for Firefly and Oplophorus gracilirostris (NanoLuc) luciferase as first and second reporter gene respectively separated by each of the five IGRs of MeV genome (Fig 6A). To this end, the NanoLuc luciferase was chosen because it has a ~150-times higher specific activity compared to Firefly luciferase [59]. Like many other paramyxoviruses, MeV polymerase has the ability to edit P mRNA by adding one non-templated G when transcribing the specific sequence termed P editing site (3’-uguggguaauuuuuccc-5’) [12,60]. We introduced this editing site just downstream the 3’-UAC-5’ START codon of the NanoLuc gene so as to condition the creation of the NanoLuc ORF and the ensuing translation of NanoLuc to the co-transcriptional insertion of one non-templated G by MeV polymerase. If minigenome RNA transcripts made by the T7 RNA polymerase are basally translated in spite of the lack of both cap and polyA signals (S4A Fig), the T7 RNA polymerase does not recognize the P editing signal [61]. As a result, while the signal to noise ratio is ~24 for Firefly, it reaches ~521 for the edited NanoLuc, i.e. a 20-fold increase of the dynamic range (S4B Fig).
As a measure of the efficacy of each N variant to support the rescue of each minigenome, Firefly luciferase signals specifically driven by MeV polymerase from the first gene (as obtained after subtraction of background levels observed in the absence of a functional L, (see S5 Fig)) were compared. They were all found to be of similar magnitude irrespective of the MeV IGR within the minigenome and of the N variant, thus indicating comparable efficiencies of the rescuing step which relies on the random but ordinated encapsidation by the N protein of the naked RNA minigenome transcribed by the T7 polymerase ([62] see [63] for review) (S5A and S5B Fig).
We then verified that these newly built dual-luciferase minigenomes harboring individually one of the five IGR faithfully reproduce the expected re-initiation strength gradient. Indeed, when normalized to the NanoLuc/Firefly signal ratio observed with a minigenome carrying the N-P IGR, the ratios observed for the minigenomes harboring the downstream IGRs decrease with their remoteness from the genome 3’end with P-M being equivalent to N-P, M-F and F-H being significantly lower and H-L being the lowest of all (Fig 6B, wt N). These results are in agreement with the transcription gradient observed in MeV infected cells [12,64,65,66] and with the efficacy of Sendai virus re-initiation at each IGR as determined using recombinant viruses [67]. Interestingly, this trend was absolutely conserved for every NTAIL variant upon normalization to the ratio observed with N-P IGR minigenome (Fig 6B) indicating that the observed re-initiation strength gradient is an intrinsic property of each IGR region. When NanoLuc/Firefly ratios observed for each N variant were plotted without normalization as a function of NTAIL/XD binding strength for each of the five MeV IGR minigenome, the NanoLuc/Firefly signal ratio was found to decrease with decreasing binding strength, with the correlation being significant at p~0.05 or below for N-P, P-M, M-F and F-H IGR minigenomes (Fig 6C–6E), and the trend conserved for the minigenomes bearing the remotest H-L IGR (Fig 6F and 6G). Since in the natural situation MeV polymerase has to travel through every IGR, we estimated for each individual variant a mean re-initiation rate through all MeV IGRs by calculating the mean NanoLuc/Firefly ratio of the 5 IGR regions for each N variant. Remarkably this mean re-initiation rate correlates with the NTAIL/XD binding strength (Fig 6H, p = 0.021).
In few percent cases, the viral polymerase fails to recognize an intergenic region. This results in read-through transcripts. To investigate the possible impact of NTAIL/XD binding on read-through generation, a 3-gene minigenome was built as follows: the first gene code for the Firefly luciferase, the second gene codes for an irrelevant inactive protein (here the C-terminal half of the Gaussia luciferase (Glu2)) followed by a linker that remains in the same coding phase throughout the second downstream N-P IGR and the third gene which contains the NanoLuc luciferase coding sequence devoid of a start codon and out of frame by one missing nucleotide that can be restored by the editing signal. Consequently, among all possible viral transcripts, only the edited read-through mRNA over gene 2 and gene 3 can give rise to a NanoLuc luciferase activity (Fig 7A and 7B). Therefore, with the 3-gene minigenome, the NanoLuc/Firefly ratio is dependent on two IGR-related effects: the re-initiation of the transcription at the first IGR and the failure to recognize the second. As expected, the NanoLuc/Firefly signal ratios obtained with this 3-gene minigenome were found to be of a much lower level (i.e. few percent) than those observed with the 2-gene minigenome shown in Fig 6C. We normalized the NanoLuc/Firefly signal obtained with the 3-gene minigenome by the signal obtained with the 2-gene minigenome in order to cancel out the effect on the re-initiation at the first IGR and to focus on the generation of read-through transcripts at the second IGR. The resulting ratios are similar for all the variants, thus indicating they all roughly produce the same amount of read-though transcripts (Fig 7C). We conclude that the NTAIL/XD binding strength does not significantly impact the failure of the viral polymerase to recognize the N-P intergenic region.
Upon crossing an IGR, the polymerase from Mononegavirales having ceased RNA synthesis at the GE is able to scan forward and backward the genome template until it recognizes the transcription re-initiation site GS of the next downstream gene. This search for next GS had been initially observed as measurable temporal pause in transcription [68] (see viral transcription scheme in S6 Fig and [13,69] for reviews). Since the frequency of re-initiation decreases with the length of the un-transcribed IGR (UTIGR) [70,71] dual-luciferase Firefly/NanoLuc 2-gene minigenomes with elongated UTIGR based on MeV N-P IGR were also built (see scheme Fig 8A) according to previous work based on the related Sendai virus that has served as the reference study model for Paramyxoviridae [70]. The Firefly signals specifically driven by MeV polymerase (as obtained after subtraction of background levels observed in the presence of an inactive L protein) observed with each combination of minigenome of variable UTIGR length and N variant were of similar magnitude irrespective of the UTIGR length and of NTAIL variant (S7A Fig) and did not show any correlation with the NTAIL/XD binding strength (S7B Fig). These data confirmed that the rescue of the minigenome, is neither dependent on the sequence of the minigenome nor on the N variant. Incidentally, these experiments also allowed appreciating the reproducibility of our dual-luciferase minigenome-based experiments, as judged by comparing S5A and S7A Figs.
As observed in the previous set of experiments, the NanoLuc/Firefly signal ratios obtained with the N-P minigenome (i.e. UTIGR “+0”) nicely correlate with the NTAIL/XD binding strengths (Fig 8B, compare also with Fig 6C for data reproducibility). When N variants were tested with elongated UTIGR minigenomes, the NanoLuc/Firefly signal ratio exponentially declined with UTIGR elongation (Fig 8C, p<0.001 for every N variant). However the declining rate varied between N variants (compare the slopes in Fig 8C). This allowed us to calculate and compare the percentage of unpriming per UTIGR nt (%unpriming/UTIGRnt). The D493G variant exhibits a significantly lower %unpriming/UTIGRnt compared to wt N, whereas that of R490S, R497G and S491L variant was significantly higher (Fig 8D). Furthermore, the %unpriming/UTIGRnt of N variants tends to vary according to the log of the NTAIL/XD KD, (Fig 8E, p = 0.062). Remarkably, the %unpriming/UTIGRnt and mean re-initiation rate through the five MeV IGR regions significantly correlate to each other (Fig 8F, p = 0.0032). Overall these data reveal that lowering the NTAIL/XD binding strength significantly increases the unpriming rate of MeV polymerase during transcription re-initiation and its scanning over un-transcribed genomic sequences, i.e. over each UTIGR.
Since even NTAIL variants with the highest KD for XD were able to reconstitute functional dual-luciferase minigenomes, we sought at evaluating the impact of substitutions in the viral context by expressing N variants into two types of recombinant viruses, namely unigene and biG-biS viruses. Unigene viruses possess only one copy of the N gene and thus express solely the mutated N protein. By contrast, biG-biS viruses contain a duplicated viral gene, here the N gene, one encoding the wt N protein (wt Flag-N1) and one encoding the mutated N protein with a HA tag (HA-N2), the expression of which can be independently silenced thanks to the use of two cell lines expressing shRNA that selectively target one of the two N genes (S8 Fig) [48]. Unigene viruses harboring NTAIL variants were all rescued. The biG-biS viruses were also all rescued in cells allowing the selective expression of the wt Flag-N1 gene copy, although the too low virus production by the R489Q and R490S viruses prevented further analysis. Virus production by recombinant viruses at 3 d.p.i. were determined for unigene viruses in Vero cells, while that of biG-biS viruses was measured in three host cells allowing selective expression of either the wt Flag-N1 gene copy, the HA-N2 gene variant, or both of the N gene copies simultaneously (Fig 9A). Virus production was found to be very low (at least 2 log reduction with respect to the wt counterpart) in the case of unigene and biG-biS S491L viruses. Note that the possibility that the observed differences in virus production of unigene viruses could be ascribed to a defect in N variant expression (S1 and S9A Figs) or to a significant contamination by defective interfering (DI) mini-replicons was checked (S9B–S9D Fig) and ruled out.
When plotted against the NTAIL/XD binding strength as determined by glu-PCA, the virus production of unigene NTAIL variants does not significantly correlate with binding strength (Fig 9B). However, the virus titer of biG-biS viruses under the selective expression of HA-N2 variant and under the combined expression of both N copies were found to correlate with NTAIL/XD binding strength (p = 0.04 and p = 0.008, respectively) (Fig 9C and 9D), while no such a correlation was found upon selective expression of the wt Flag-N1 copy as expected (Fig 9E). We noticed that the coexpression of N wt with D493G variant appears deleterious for virus production (Fig 9D). However, in a minigenome assay such a mixture of N was as efficient as N wt alone (S10 Fig), thus ruling out the possibility that NTAIL heterogeneity could directly impact the polymerase activity. Overall these data indicate that the NTAIL/XD binding strength may control the virus production to some extent.
We then took advantage of unigene viruses expressing the single-site Box2 variants to determine which activity of the viral polymerase could be affected by a change in the NTAIL/XD binding affinity. Vero cells were infected with wt, R489G, R490S, A492T, D493G and R497G unigene viruses. Note that the S491L variant was not investigated since it could not be further amplified to reach a workable titer. RNA synthesis parameters reflecting primary transcription (i.e. mostly, if not solely, transcription, mediated by the active polymerases brought by infecting virions), secondary transcription and replication were determined by quantification of (+) and (-) RNA accumulation at different times post-infection as previously reported [48,65]. When RNA synthesis parameters were plotted along with NTAIL/XD KD, it appeared that both (+) RNA transcript accumulation rate and ratios between P (or F) and N transcripts could be roughly predicted from the interaction strength between the NTAIL variant and XD as measured by either method (Fig 10). The correlations were statistically significant between the accumulation rate of P (+) transcripts and NTAIL/XD KD (Fig 10A) and between the F/N transcript ratios measured at 24 h.p.i. and the KD (Fig 10B). In further support of the coherence of the results, a good correlation was found between the accumulated levels of N and P (+) RNAs during primary transcription and at 24 h.p.i. (S11A and S11B Fig), and between both N (+) and P (+) RNA transcripts and (-) genomic RNA (S11C and S11D Fig).
When the F/N mRNA ratios at 24 h.p.i. observed with unigene viruses were plotted against the calculated mean re-initiation rate of the 5 IGRs and the %unpriming/UTIGRnt a significant positive and a negative correlation were found, respectively (Fig 11). Altogether, these data support that the NTAIL/XD binding strength controls, at least in part, the steepness of the viral transcription gradient.
By combining in vitro biophysical and biochemical studies, in silico analyses (i.e. MD simulations) and in cellula polymerase functional investigations using recombinant viruses and dual-luciferase editing-dependent minigenome assays, we deciphered key molecular parameters that govern the NTAIL/XD interaction. Specifically, we uncovered a correlation between interaction strength and efficiency of transcription re-initiation at intergenic regions.
For most of the NTAIL variants the observed variations in binding affinities cannot be ascribed to differences in the extent of α-helical sampling of the free form of the α-MoRE, nor to differences in the ability of the latter to undergo induced α-helical folding. However, the R489Q substitution represents an exception in this respect: indeed, it has a reduced extent of α-helicity and a slightly increased KD towards XD. The reduced α-helical content of this variant is in line with secondary structure predictions, as obtained using the Psipred server (http://bioinf.cs.ucl.ac.uk/psipred/) [72], that predicts a slightly lower helical propensity. Whether the experimentally observed reduction in affinity towards XD arises from this lower helicity or from other attributes, including charge-related ones, remains to be established. This variant also displays a reduced accumulation rate of primary transcripts. The subtle molecular mechanisms underlying the peculiar behavior of this variant remain however to be elucidated.
The complex hydrogen bonding revealed by MD simulations of NTAIL/XD complexes allows the drops in binding affinities experimentally observed for the S491L, R497G and R490S variants to be rationalized. Interestingly, these substitutions, which have the most dramatic effects in terms of binding affinities, are also the ones that have the strongest effect on virus replication, with the S491L substitution being very poorly tolerated even in biG-biS viruses. The poor ability of the low-affinity S491L variant in mediating efficient virus replication is reminiscent of the comparable deleterious effect of the F497D XD substitution [48] and of the detrimental effect of the deletion of the NTAIL region encompassing the α-MoRE [49].
We provide here compelling evidence indicating that the strength of the NTAIL/XD interaction controls, at least in part, the ability of the P+L polymerase complex to re-initiate at IGRs: data obtained using our highly sensitive and reproducible dual-luciferase minigenome assay reveal a significant correlation between the NTAIL/XD binding strength and the efficiency of the transcription re-initiation. Since our minigenome assays rely on the edition of the second reporter gene, we cannot formally exclude that the editing may be also impacted by the NTAIL/XD binding strength. However, the calculated %unpriming/UTIGRnt only depends on the decrease of the NanoLuc/Firefly signals ratios with the length of the UTIGR. The observed effect is therefore independent of any potential effect on the edition (i.e. if N mutations only had an effect on editing, then this effect should be the same irrespective of the IGR under study and of its length, which is not the phenotype we observed). Moreover, the correlation in the viral context between the P/N and F/N mRNA ratio and the KD, supports a role for the XD/NTAIL interaction strength in the re-initiation at IGRs.
A N protein truncated of its last 86 C-terminal amino acids, i.e. truncated of most of NTAIL including the XD binding site, had been shown to be active in transcription and replication both in a minigenome assay and when introduced into a recombinant virus [49]. We confirmed that the N1-439 truncated protein is as good as, if not better than, the wt N in transcribing the Firefly gene from our N-P 2-gene minigenome construct (S12A Fig). However, its ability to support transcription re-initiation over the N-P junction was significantly reduced, with the extent of reduction being comparable with that observed with the low affinity R497G variant (S12B Fig, UTIGR 0 nt), thus confirming the role of NTAIL/XD interaction in transcription re-initiation. This low efficiency of transcription re-initiation may explain the extreme growth defect of the recombinant virus bearing the truncated N until reversion to a wt N [49].
Assuming a very slow degradation of viral mRNA [65,66,73], the transcripts accumulation rate in cells infected with unigene viruses reflects the RNA synthesis rate by the polymerase, the number of active polymerases (and their recruitment onto the nucleocapsid template), and the number of polymerases that are recruited per time unit on a given gene. For the same reason, the transcript ratios between the different genes are likely mostly governed by the efficiency with which the polymerase re-initiates the transcription at each IGR. Assuming this being a conserved feature for every N variant, we can reasonably interpret the inverse correlation we observed between multiple transcript ratios and KD as reflecting a direct control of the NTAIL/XD binding strength on the efficiency of the re-initiation at each IGR. A lower binding strength leads to lower levels of downstream transcripts. After completion of the polyadenylation of the messenger encoded by the upstream gene, the polymerase may remain firmly in contact with its genomic RNA template embedded into the nucleocapsid only if maintained by the anchoring of its P subunit via a dynamic binding of its X domain to the TAIL domain of N subunits located at the IGR (Fig 12). Therefore, a decrease in the XD/NTAIL affinity may favour the unpriming of the polymerase. Whether unprimed polymerases can detach from the nucleocapsid or stay on the template and move forward to the end of the nucleocapsid remain to be established. Hence, XD to NTAIL anchoring would tightly control the re-initiation level of the RNA synthesis by the polymerase in the transcription mode, thus determining the steepness of the transcription gradient (Fig 12, see also S6B Fig).
What could be the functional significance of the relationship between the accumulation rate of primary N and P transcripts and the XD/NTAIL binding strength? As speculated, the dynamics of XD/NTAIL binding and release may also affect the polymerase processivity on the nucleocapsid [48]. The XD/NTAIL interaction may act as a brake and slows down the polymerase: the weaker is the interaction, the weaker is the brake. Also, because of the efficient recycling of the polymerases on the promoter [65], if, in the absence of transcription re-initiation, the polymerase detaches from the RNA template, a steeper gradient would release more polymerases available for transcription of the first genes. With weaker NTAIL/XD interactions, the viral production by unigene viruses tends to be negatively affected although the correlation was not statistically significant likely because of the small number of available virus variants and the too high variability of the result due to the multiple intervening parameters (see S6A Fig and the complete scheme of virus replication dynamics in [65]). However with biG-biS viruses, we did observe a significant correlation between virus production and NTAIL/XD binding strength in conditions where the N variant was selectively expressed. This significance may reflect both the higher number of available virus variants and/or the higher impact of the modulation of the transcription re-initiation process in viruses possessing an additional transcription unit (i.e. where the polymerase has to go through one additional IGR). The similar correlation observed upon the co-expression of both wt Flag-N1 and variant HA-N2 copies may indicate similar impact on transcription re-initiation because of the tetrameric valence of the P anchoring on (contiguous?) heterogeneous NTAIL appendages. Alternatively, it is possible that the heterogeneity of NTAIL within a given nucleocapsid template may have a negative impact on other mechanisms such as nucleocapsids packaging into particles since NTAIL also recruits the M protein [74], a key virion assembly factor [75]. The discrepancy we observed between virus production from biG-biS viruses and minigenome data with mixed NTAILs argues for this later hypothesis.
Using minigenomes with elongated UTIGR, we were able to measure the unpriming rate of the polymerase in the “scanning mode” and we show that a decrease in the NTAIL/XD affinity induces an increase of the unpriming rate. In this situation, without the stabilization and the active motion of the polymerase due to the RNA synthesis, the role of the NTAIL/XD interaction in maintaining the polymerase on the nucleocapsid may overcome the “brake” effect. Alternatively, as suggested by Krumm et al [49], the NTAIL may need to be rearranged by P to allow an efficient RNA synthesis. In this case, a too low NTAIL/XD affinity may weaken the efficiency of P in rearranging NTAIL and would favor the unpriming of the polymerases. The fact that the N1-439 variant, that lacks most of NTAIL, has the lowest unpriming rate on UTIGR supports this second hypothesis (0.6 vs 0.81%unpriming/UTIGRnt for N1-439 and wt N respectively) (S12C Fig).
In conclusion, the XD/NTAIL interaction may play a critical role in the polymerase processivity, in maintaining the polymerase anchored to the nucleocapsid during its scanning upon crossing the intergenic regions, and/or in the transcription re-initiation at each intergenic region. Since both increasing [48] or decreasing (this study) the XD/NTAIL affinity negatively affect the viral growth, the wild type XD/NTAIL binding strength seems to have been selected to mediate an optimal equilibrium between polymerase recruitment, polymerase processivity and transcription re-initiation efficiency. A corollary of this is that substitutions that strongly affect affinity towards XD are poorly tolerated. Consistent with this, the substitutions with the most dramatic impact herein investigated (i.e. R490S, S491L and R497G) do not naturally occur in any of the 1,218 non-redundant MeV sequences, while those that have a less drastic impact (i.e. R489Q, A492T and D493G) are found in circulating measles strains [47]. Interestingly, in the case of Ebola virus (EBOV), an additional protein, i.e. VP30, serves as an anti-terminator transcription factor, and mutations that either decrease or increase the binding affinity between N and VP30, decrease RNA synthesis [76] thus arguing for a similarly tightly regulated interaction. According to our work, the NTAIL to XD binding strength tightly controls the transcription gradient. However, this does not rule out the possibility that other mechanisms may be at work in controlling the steepness of the gradient. Indeed, in the brain of three patients suffering from subacute sclerosis encephalitis (SSPE) or measles inclusion bodies encephalitis (MIBE) the transcription gradient was found to be steeper than the one measured in in vitro infected cells [77] although the amino acid sequences of NTAIL and XD were found to be unvaried [78]. Furthermore, in the absence of the C protein, a steeper transcription gradient is also observed [79]. These two lines of evidence advocate for a multi-parametric control of the transcription gradient.
The major role of the N binding site on the C-terminus of P has been postulated to mediate L anchoring to the nucleocapsid without understanding the implication of such anchoring on the polymerase and/or on the nucleocapsid dynamics. The need for an optimized interaction between the P and N proteins might be one of the major evolution constraints to which the polymerase machinery of MeV, and possibly of paramyxoviruses in general, is subjected. Our findings raise also the question as to whether binding of the C-terminus of P to the globular moiety of N, as observed in other Mononegavirales members, needs to be similarly controlled reflecting a similar functional role.
The bipartite nature of P to N binding (see scheme Fig 1A) is remarkably conserved throughout the Mononegavirales order [80]. An α-MoRE located at N-terminus of P binds to the C-terminal globular domain of the NCORE to form the so-called N0P complex that is used by the polymerase as the encapsidation substrate. Solved N0P structures from members of the Rhabdoviridae family (vesicular stomatitis virus, VSV) [8], Filoviridae family (VP35, of EBOV) [81,82], Pneumoviridae family (human metapneumovirus, HMPV) [9] and Paramyxoviridae family (Nipah virus, NiV a Henipavirus member) [7], MeV a Morbillivirus member [5], mumps virus, MuV a Rubulavirus member [83] revealed a common mechanism whereby the N terminus of P competes out with N arms that stabilize the oligomeric form of N and directly or indirectly prevents RNA binding. Structural and functional evidences indicate that, via its N-terminus, P can transiently uncover the genome at its 3’end from the first N subunits to give L access to its genomic RNA template (see [83] and [84] for review).
An additional N-binding site is located at the C-terminus of P (or VP35 for EBOV) (see scheme Fig 1A) [85] and allows binding to the assembled form of N. While this secondary binding site is required for the polymerase activity in minigenome experiments from several viruses [3,83,85,86,87], the structures of the reciprocal N-binding and P-binding site on P and N, respectively, look less conserved. In the case of NiV [42], Hendra virus [88], SeV [38,89] and MeV [27,30,44,89] the C-terminal domain of P (XD) is structurally conserved and consists of a bundle of 3 α-helices that are structurally analogous, and that dynamically binds to a α-MoRE located near the C-terminus of NTAIL ([90,91], see [92] for reviews). This NTAIL-XD interaction is commonly characterized by a rather low affinity (KD within the 3–50 μM range, [39,46,89] and this work). In Rubulavirus members, the C-terminal region of P spans in solution a structural continuum ranging from stable triple α-helical bundles to largely disordered, with crystal packing stabilizing the folded form [93,94]. In MuV, this triple α-helical bundle analogous to XD binds directly to the core of N subunits of the nucleocapsid [24] without excluding a complementary binding to the extremity of NTAIL [83]. By analogy with MuV XD, MeV XD might also bind to another binding site located on NCORE. This would explain how transcription and replication can still be observed in the presence of the N1-439 truncated where interaction of P relies only on NCORE ([49] and this paper). Indeed in other Mononegavirales members, the C-terminus of P binds to the core of N. In the case of RSV, the minimal nucleocapsid-binding region of P, which encompasses the last nine P residues, is disordered [95] and remains predominantly disordered even upon binding to the N-terminal lobe of NCORE [96].
The C-terminal domain of P from Rabies virus (RABV) [97], Mokola virus [98,99] and VSV [100] share a fold made of a bundle of α-helices that binds to the core of two adjacent N proteins of the nucleocapsid [101,102]. The N protein of Rhabdoviridae members, along with the N protein from RSV lacks the disordered NTAIL domain that characterizes N proteins from Paramyxoviridae members. In contrast to the XD-NTAIL interaction, the C-terminal domain of RABV P binds to the nucleocapsid with a high affinity (KD in the nanomolar range) [101]. In spite of the diversities of both structural features and binding modes within Mononegavirales members, does the binding of C-terminus of P to the assembled form of N fulfill common functions, namely ensuring the proper efficiency in polymerase scanning and re-initiation at intergenic regions? Further works will unveil to which extent our present findings are relevant for other members of the Paramyxoviridae or other families of the Mononegavirales order.
The pDEST17O/I vector [103], allowing the bacterial expression of N-terminally hexahistidine tagged recombinant proteins under the control of the T7 promoter, was used for the expression of all NTAIL variants. The pDEST17 derivatives encoding single-site NTAIL variants bearing substitutions within Box2 were obtained either by Gateway recombination cloning technology (variants R489Q, R490S, S491L and R497G) using the previously described pNGG derivatives [47] as the donor vectors, or by site-directed mutagenesis (variants A492T and D493G). In the latter case, we used a pair of complementary mutagenic primers (Operon) designed to introduce the desired mutation, Turbo-Pfu polymerase (Stratagene), and the pDEST17O/I construct encoding wt MeV (Edmonston B) NTAIL as template [47]. After digestion with DpnI to remove the methylated DNA template, CaCl2-competent E. coli TAM1 cells (Active Motif) were transformed with the amplified PCR product.
The pNGG derivative encoding the MXSF NTAIL variant N-terminally fused to the N-terminal fragment of GFP was obtained in four steps using pNGG/NTAIL as template [47,104] and site-directed mutagenesis PCR. In the first step, the pair of mutagenic primers was designed to introduce the first amino acid substitution. After PCR and DpnI digestion, CaCl2-competent E. coli TAM1 cells (Active Motif) were transformed with the amplified PCR product. After having sequenced the construct to ensure that the desired mutation had been introduced, a second PCR was carried out using another pair of mutagenic primers designed to introduce the second substitution. Repeating this procedure four times led to the final construct bearing the four desired substitutions (i.e. D437V, R439S, P456S and P485L).
The sequences of the coding regions of all constructs generated in this study were checked by sequencing (GATC Biotech) and found to conform to expectations.
The pDEST17/NTAIL construct encoding wt NTAIL has already been described [47], as is the pDEST14 construct encoding C-terminally hexahistidine tagged MeV XD [30].
The plasmid p(+)MVNSe previously described in [48] was used as the MeV genome backbone. MeV genomic plasmids were built by direct recombination of one or two PCR fragments according to the InFusion user manual (Clontech). To build biG-biS recombinant viruses, the N gene was duplicated in N1 and N2 in gene positions 1 and 2, respectively. N1 was tagged with an N-terminal Flag peptide and three copies of the GFP RNAi target sequence (GAACGGCATCAAGGTGAA) in the 3’UTR of its mRNA. N2 was tagged with an N-terminal hemagglutinin (HA) peptide and three copies of the P RNAi target sequence (GGACACCTCTCAAGCATCAT) in the 3’UTR. Mutations into the NTAIL domain of N, R489Q, R490S, S491L, A492T, D493G, R497G, MXSF (D437V/R439S/P456S/P485L) and MX208 (D437V/P485L/L524R), were introduced by subcloning PCR-amplified fragments from the pDEST17/NTAIL vectors.
Full length wt and mutated N, wt and mutated NTAIL, PPMD-XD and P376-507 fragments were subcloned downstream Gaussia glu1 and/or glu2 domains by InFusion recombination of PCR-amplified fragments as previously described [48]. All plasmids and viruses (N1, N2, P, M, and L gene) were verified by sequencing the subcloned PCR fragments or cDNA obtained by reverse transcription-PCR (RT-PCR) performed on virus stocks.
Plasmids encoding dual-luciferase editing-dependent 2-gene minigenomes were built by InFusion subcloning of PCR amplicons encompassing Firefly and NanoLuc coding sequences flanked by N UTR and L 3’UTR. The two luciferase coding sequences are separated by the N-P IGR either unmodified or exchanged with P-M, M-F, F-H and H-L IGRs (i.e. untranscribed 3’-GAA-5’ (or 3’-GCA-5’ for H-L) triplet flanked by canonical upstream and downstream gene end and gene start sequences arbitrarily fixed to 15 nt) or elongated by 12, 36, 108 or 324 nt (see sequences in S2 and S3 Tables) into the p107(+) MeV minigenome construct that drives the synthesis of (+) genomic strand under the control of the T7 promoter [62]. According to the rule of six that governs the strictly conserved hexameric length of measles virus genome [62,105], all minigenomes share identical phasing of the last U of the polyadenylation signal of the firefly gene (phase 6, i.e. the last nucleotide covered by the N subunit) and of the editing site with the C being in phase 6 as defined in [106]. A 3-gene minigenome coding for Firefly and chimeric Glu2-linker-NanoLuc luciferase as a results of read-through between gene 2 and gene 3 and RNA editing was built by modifying the N-P 2-gene minigenome. As a second gene, the ORF of the C-terminal domain of Glu (glu2) was inserted downstream to a START codon but without a STOP codon. This ORF is followed by a second N-P IGR and by the NanoLuc ORF without its own START codon, in frame “-1nt” to the upstream Glu2 ORF. Following addition of a G thanks to the presence of the P editing site, the NanoLuc ORF becomes in frame with the upstream Glu2 ORF. Consequently, the full-length chimeric Glu2-linker-NanoLuc can be uniquely translated from a read-through transcript over the second N-P IGR that is also edited (see sequence in S4 Table).
All plasmids will be deposited in the Addgene plasmid repository service except the glu1 and glu2 constructs that Addgene cannot accept. Those constructs are available upon request.
The E. coli strain Rosetta [DE3] pLysS (Novagen) was used for the expression of all recombinant proteins. Transformants were selected on ampicillin and chloramphenicol plates. 50 mL of Luria-Bertani (LB) medium supplemented with 100 μg/mL ampicilin and 34 μg/mL chloramphenicol were seeded with the selected colonies, and grown overnight to saturation. An aliquot of the overnight culture was diluted 1/25 in LB medium containing ampicillin and chloramphenicol and grown at 37°C. When the optical density at 600 nm (OD600) reached 0.6–0.8, isopropyl ß-D-thiogalactopyranoside (IPTG) was added to a final concentration of 0.2 mM, and the cells were grown at 37°C for 4 additional hours. The induced cells were harvested, washed and collected by centrifugation (5,000 g, 12 min). The resulting pellets were frozen at –80°C.
All the NTAIL and XD proteins were purified to homogeneity (> 95%) from the soluble fraction of bacterial lysates in two steps: Immobilized Metal Affinity Chromatography (IMAC), and size exclusion chromatography (SEC). Cellular pellets of bacteria transformed with the different expression plasmids were resuspended in 5 volumes (v/w) of buffer A (50 mM Tris/HCl pH 8, 300 mM NaCl, 20 mM imidazole, 1 mM phenyl-methyl-sulphonyl-fluoride (PMSF)) supplemented with lysozyme (0.1 mg/mL), DNAse I (10 μg/mL), 20 mM MgSO4 and protease inhibitor cocktail (Sigma). After a 30-min incubation with gentle agitation, the cells were disrupted by sonication. The lysate was clarified by centrifugation at 20,000 g for 30 min. The clarified supernatant, as obtained from a one-liter culture, was incubated for 1 h with 5 ml (50%) Chelating Sepharose Fast Flow Resin preloaded with Ni2+ ions (GE, Healthcare), previously equilibrated in buffer A. The resin was washed with buffer A supplemented with 1 M NaCl to remove contaminating DNA, and the proteins were eluted in buffer A containing 1 M NaCl and 250 mM imidazole. Eluents were analyzed by SDS-PAGE. Fractions containing the recombinant product were concentrated using centrifugal filtration (Centricon Plus-20, 5000 Da molecular cutoff, Millipore). The proteins were then loaded onto a Superdex 200 (NTAIL) or Superdex 75 (XD) 16/60 column (GE, Healthcare) and eluted in 10 mM Tris/HCl pH 8, 150 mM NaCl.
Protein concentrations were calculated using the theoretical absorption coefficients at 280 nm as obtained using the program ProtParam at the EXPASY server.
Mass analysis of the purified mutated NTAIL proteins was performed using an Autoflex II ToF/ToF (Bruker Daltonics). Spectra were acquired in a linear mode. 15 pmol of samples were mixed with an equal volume (0.7 μL) of sinapinic acid matrix solution, spotted on the target and dried at room temperature.
The identity of the purified NTAIL proteins was confirmed by mass spectral analysis of tryptic fragments obtained by digesting (0.25 μg trypsin) 1 μg of purified recombinant protein isolated onto SDS-PAGE. The tryptic peptides were analyzed as described above and peptide fingerprints were obtained and compared with in-silico protein digest (Biotools, Bruker Daltonics). The mass standards were either autolytic peptides or peptide standards (Bruker Daltonics).
The CD spectra of NTAIL proteins were recorded on a Jasco 810 dichrograph using 1-mm thick quartz cells in 10 mM sodium phosphate pH 7 at 20°C. CD spectra were measured between 190 and 260 nm, at 0.2 nm/min and are averages of three acquisitions. Mean ellipticity values per residue ([Θ]) were calculated as [Θ] = 3300 m ΔA/(l c n), where l (path length) = 0.1 cm, n = number of residues, m = molecular mass in daltons and c = protein concentration expressed in mg/mL. Number of residues (n) is 147, while m is 16 310 Da. Protein concentrations of 0.1 mg/mL were used when recording spectra. Structural variations of NTAIL proteins were measured as a function of changes in the initial far-UV CD spectrum following addition of 20% 2,2,2 trifluoroethanol (TFE) (Sigma-Aldrich).
The experimental data in the 190–260 nm range were analyzed using the DICHROWEB website which was supported by grants to the BBSRC Centre for Protein and Membrane Structure and Dynamics [107,108]. The CDSSTR deconvolution method was used to estimate the content in α-helical and disordered structure using the reference protein set 7.
ITC experiments were carried out on an ITC200 isothermal titration calorimeter (Microcal) at 20° C. Protein pairs used in the binding analyses were dialyzed against the same buffer (10 mM Tris/HCl pH 8, 150 mM NaCl) to minimize undesirable buffer-related effects. The dialysis buffer was used in all preliminary equilibration and washing steps.
The concentrations of purified wt and mutated NTAIL proteins in the microcalorimeter cell (0.2 mL) ranged from 25 μM to 180 μM. XD was added from a computer-controlled 40-μL microsyringe via a total of 19 injections of 2 μL each at intervals of 180 s. Its concentration in the microsyringe ranged from 300 μM to 960 μM.
A theoretical titration curve was fitted to the experimental data using the ORIGIN software (Microcal). This software uses the relationship between the heat generated by each injection and ΔH° (enthalpy change in kcal mole-1), KA (association binding constant in M-1), n (number of binding sites per monomer), total protein concentration and free and total ligand concentrations. The variation in the entropy (ΔS° in cal mol-1 deg-1) of each binding reaction was inferred from the variation in the free energy (ΔG°), where this latter was calculated from the following relation: ΔG° = -RT ln 1/KA.
All MD simulations were performed in explicit solvent with periodic conditions with CHARMM and NAMD software packages and CHARMM force field version 27 with CMAP corrections. The initial coordinates of the XD/α-MoRE complex were taken from the crystal structure (PDB code 1T6O) [44]. The two XD/α-MoRE mutated models bearing either the S491L or the R497G NTAIL substitution, were built with VMD plugin ‘mutator’ starting from the X-ray structure of the wt complex (PDB code 1T6O). In the case of the S491L variant, the three most favourable orientations of the leucine side chain were generated with Sybyl. Non-protein derivatives were discarded. Orientation of the side chains of Asn, Gln, and His residues were checked using in-house VMD plugin and the WHAT IF web interface (http://swift.cmbi.kun.nl/). Residue His498 of XD was assigned HSD type and all other titratable groups were assigned standard protonation state at pH 7.0. Coordinates of missing hydrogen atoms were added using the hbuild algorithm in CHARMM. To improve conformational sampling, three independent simulations were carried out using different initial velocities. The system was solvated with a pre-equilibrated solvation box (edge length around 60 Å) consisting of TIP3P water molecules. Crystallographic water molecules were included in the initial model. Chloride and sodium ions were added to achieve neutralization of the whole system. Periodic boundary conditions were applied. Unfavorable contacts were removed by a short energy minimization with conjugate gradient and ABNR. Electrostatic interactions were treated using the particle-mesh Ewald summation method, and we used the switch function for the van der Waals energy interactions with cuton, cutoff and cutnb values of 10, 12 and 14 Å respectively. Vibration of the bonds containing hydrogen atoms were constrained with the Shake algorithm and a 1-fs integration step was used. The system was heated gradually to 300K, followed by an equilibration step (500 ps). During these two early steps, harmonic constraints were applied to protein heavy atoms. The constraint harmonic constant (k) was equal to 1 and 0.1 kcal/mol/Å2 for the backbone and side chains, respectively, and was removed after 250 ps equilibration. The production phase of 50 ns was performed without any constraints. Snapshots of the coordinates were saved every 0.5 ps. Trajectories were analyzed using a combination of in-house and VMD scripts.
Overall <RMSD> variations were computed with VMD after superimposition of the Cα atoms of each conformation generated onto the initial structure (last structure of the equilibration step). Flexible N- and C-terminal residues were not included in the calculation. Three types of RMSD were computed as it follows. For each frame, the XD protein was superimposed onto the initial XD model and RMSD was computed over XD Cα atoms only. For each frame the α–MoRE was superimposed onto the corresponding region of the initial structure and RMSD was computed over NTAIL Cα atoms only. For each frame, the XD protein was superimposed onto the initial XD protein and RMSD was computed over NTAIL Cα atoms only.
Free-energy perturbation (FEP) module implemented in NAMD was used to perform alchemical transformation of Ser491 to Leu and Arg497 to Gly. Free energies differences resulting from the Ser to Leu or Arg to Gly substitution were computed using the thermodynamic cycle shown in S3 Fig. The free form of the α–MoRE in solution was taken from the XD/NTAIL complex.
The free energy profile for the dissociation of the XD/α-MoRE complex (wt and mutated forms) was computed using the adaptive biasing force (ABF) method, implemented in NAMD [109]. This method relies upon the integration of the average force acting on a selected reaction coordinate (here, the center of mass between the two partners). A biasing force is applied to the system in such a way that no average force acts along the reaction coordinate thus allowing overcoming free energy barriers. For a complete description of the method please refer to http://www.edam.uhp-nancy.fr/ABF/theory.html and references therein shown.
The distance separating the centers of mass of the two proteins was selected as reaction coordinate. The distance was calculated on the Cα atoms not taking into account the three atoms at each end (N-term and C-term end) of each protein partner due to their high flexibility. This distance is about 11 Å in the associated form and the partners are considered dissociated after a 10 Å increase in this distance. The reaction coordinate was subdivided into sections of 0.5 Å and each one was successively explored during 5 ns. Bin width was kept at 0.02 Å, the number of samples prior to force application was 500 and the wall force constant is 100 kcal.mol-1.Å2. Once a section is sampled, the conformation in which COM distance is the nearest to the upper boundary is selected as the starting point of the following 0.5 Å section. A post-processing step merges the sampling counts and the PMF of each part and generates the whole profile of PMF along the dissociation process. The trajectories were generated using the same protocol as described for free MD.
Cells were cultured in DMEM medium (Life Technologies) supplemented with 10% of heat-inactivated (30 min at 56°C) fetal bovine serum, 1% L-glutamine, gentamicin (10 μg/ml) at 37°C and 5% CO2. Medium of 293-3-46 helper cells was supplemented with G418 at 1.2 mg/ml. Vero (si2) and Vero-SLAM (si1) cells stably expressing shRNA targeting the P and GFP mRNAs, respectively, were previously described [48]. To rescue recombinant viruses, the helper cell line 293-3-46 stably expressing T7 polymerase, MeV N, and P was transfected by using the ProFection kit with two plasmids coding for the MeV genome and MeV-L protein (pEMC-La) [110]. Three days after transfection, the cells were overlaid on either Vero (single N gene virus) or Vero-si2 cells (bi-N virus). Upon appearance, isolated syncytia were picked and individually propagated on relevant Vero (from CelluloNet BioBank BB-0033-00072, SFR BioSciences, Lyon France) (single N virus) or Vero-si2 (bi-N virus) cells. Virus stock was produced after a second passage at a multiplicity of infection (MOI) of 0.03 in the relevant cell line. This stock was checked to rule out mycoplasma contamination, has its N1, N2, P, M, and L genes sequenced, and was titrated on the relevant host cell before use.
Gaussia princeps luciferase-based complementation assay and data analysis (normalized luminescent ratio, NLR) were performed according to [50]. Human 293T cells (from CelluloNet BioBank BB-0033-00072, SFR BioSciences, Lyon France) were cultured in Dulbecco's Medium Eagle’s Modified (DMEM) (Life Technologies) supplemented with 10% of heat inactivated (30 min at 56°C) fetal bovine serum, 1% L-Glutamine and 10 μg/ml gentamycin at 37°C and 5% CO2. Cells were transfected using the jetPRIME reagent (Polyplus transfection). NLR was calculated by dividing the luciferase value of the two chimeric partners by the sum of the luciferase value of every chimeric partner mixed with the other “empty” glu domain. Results were expressed as fold increase with respect to the reference NTAIL/XD, which was set to 1.
Parental Vero, si1 and si2 cells were infected at MOI 1 with recombinant viruses with or without addition of 10 μg/ml of fusion inhibitor peptide z-fFG to prevent syncytium formation. Virus production was measured after freeze-thaw cycles of infected cells using a 50% tissue culture infective dose (TCID50) titration assay. Contamination of virus stock with internal deletion and copyback defective interfering (DI) minigenomes were assessed according to the method of [111].
Detection of the expression of viral N, Flag-N1, HA-N2, P and cellular GAPDH proteins was performed by Western blotting. Infected cells were lysed in NP40 buffer (20 mM Tris/HCl pH 8, 150 mM NaCl, 0.6% NP-40, 2 mM EDTA, protease cocktail inhibitor Complete 1 x (Roche)) for 20 minutes on ice. The proteins were then separated from the cell debris by centrifugation at 15,000 g during 10 minutes. The proteins were denatured by the addition of Laemmli 1 x loading buffer before analysis by SDS-PAGE and immunoblotting using anti-N (cl25 antibody), anti-Flag (Sigma), anti-HA (Sigma), anti-P (49.21 antibody) and anti-GAPDH (Mab374, Chemicon) monoclonal antibodies. Western blotting was revealed by chemiluminescence as detailed previously [48].
Quantification of the MeV genome and mRNA contents of infected cells was performed by reverse transcription-quantitative PCR essentially as described previously [65], using the following primers. To quantify mRNA, sense N primer (5’-AAGAGATGGTAAGGAGGT-3’), antisense N primer (5’-ATGATACTTGGGCTTGTC-3’), sense P primer (5’-TGGACGGACCAGTTCCAGA-3’), antisense P primer (5’-GGCTCCTTTGATATCATCAAG-3’), sense F primer (5’-GCTCAGATAACAGCCGGCATT-3’), antisense F primer (5’-AGCTTCTGGCCGATTA-3’) were used. Negative-strand genome was reverse transcribed using sense 5’-tagged N primer (5’-gcagggcaatctcacaatcaggAAGAGATGGTAAGGAGGT-3’), and the cDNA was PCR quantified using sense tag primer (5’-gcagggcaatctcacaatcagg-3’) and antisense N primer. For the genome the results were expressed as copy number/μg RNA, and for transcripts the results were expressed either as the number of polymerized nucleotides/genome copy or as the viral transcript/μg RNA after normalization for the genome copy contents of each sample.
The assay was performed essentially as described in [62,112] with minor modifications. 2.104 BSRT7 cells that constitutively express the T7 phage DNA-dependent RNA polymerase [113] were seeded in 96-well plates and transfected the day after with 66 ng of pEMC-N (either wt or mutated) 44 ng of pEMC-(Flag/L+P) (a home-made T7-driven bicistronic construct) [58] and 90 ng of plasmid encoding for the different minigenomes mixed with the transfection reagent as indicated in the manufacturer protocol (jetPRIME Polyplus-transfection). Two days after transfection, Firefly and NanoLuc activity were measured using the Nano-Glo Dual-Luciferase Reporter Assay (Promega). The background luciferase activity from of both luciferases observed in the absence of active L protein was subtracted from the signal measured in the presence of L, and data obtained from three independent experiments were normalized to each other to level the mean signal observed for all combinations tested at the same time. The percentage of unpriming per nucleotide of the un-transcribed intergenic (UTIGR) region (%unpriming/UTIGRnt) was calculated as follow. The luciferase signals ratios were plotted in relation to the length of the UTIGR region and the equation of the exponential regression curve was calculated (y = b*ea). The %unpriming/UTIGRnt = 1-ea.
|
10.1371/journal.pntd.0002392 | The Outcome of Trachomatous Trichiasis Surgery in Ethiopia: Risk Factors for Recurrence | Over 1.2 million people are blind from trachomatous trichiasis (TT). Lid rotation surgery is the mainstay of treatment, but recurrence rates can be high. We investigated the outcomes (recurrence rates and other complications) of posterior lamellar tarsal rotation (PLTR) surgery, one of the two most widely practised TT procedures in endemic settings.
We conducted a two-year follow-up study of 1300 participants who had PLTR surgery, conducted by one of five TT nurse surgeons. None had previously undergone TT surgery. All participants received a detailed trachoma eye examination at baseline and 6, 12, 18 and 24 months post-operatively. The study investigated the recurrence rates, other complications and factors associated with recurrence. Recurrence occurred in 207/635 (32.6%) and 108/641 (16.9%) of participants with pre-operative major (>5 trichiatic lashes) and minor (<5 lashes) TT respectively. Of the 315 recurrences, 42/315 (3.3% overall) had >5 lashes (major recurrence). Recurrence was greatest in the first six months after surgery: 172 cases (55%) occurring in this period. Recurrence was associated with major TT pre-operatively (OR 2.39, 95% CI 1.83–3.11), pre-operative entropic lashes compared to misdirected/metaplastic lashes (OR 1.99, 95% CI 1.23–3.20), age over 40 years (OR 1.59, 95% CI 1.14–2.20) and specific surgeons (surgeon recurrence risk range: 18%–53%). Granuloma occurred in 69 (5.7%) and notching in 156 (13.0%).
Risk of recurrence is high despite high volume, highly trained surgeons. However, the vast majority are minor recurrences, which may not have significant corneal or visual consequences. Inter-surgeon variation in recurrence is concerning; surgical technique, training and immediate post-operative lid position require further investigation.
| Trachoma is the most common infectious cause of blindness worldwide. It causes trichiasis (inturning of the eyelashes to touch the eye), which can cause visual loss. Trachomatous trichiasis (TT) affects over eight million people, 1.2 of whom live in Ethiopia – the most affected country worldwide. Surgery is the mainstay of treatment for TT. However, results of surgery in the field are often very mixed. We investigated the surgical outcomes of one of the two most widely used surgical techniques (posterior lamellar rotation), in 1300 individuals in the Amhara Region of Ethiopia. We found that recurrence occurred frequently: 315/1276 (24.7%) participants. However, recurrence was rarely severe (greater than 5 lashes): 42 participants (3.3%). Recurrence occurred much more frequently in participants who had severe pre-operative disease and with specific surgeons. The high recurrence rates and inter-surgeon variation is concerning. Further research will be required to investigate factors such as surgical technique, surgeon training and immediate post-operative lid position, in order to improve surgical outcomes.
| Blindness from trachoma is the end result of progressive scarring of the conjunctiva driven by Chlamydia trachomatis. The major risk factor of blinding corneal opacification (CO) is trichiasis (TT), the in turning of the eyelashes. TT traumatises the delicate epithelium of the cornea, rendering it vulnerable to secondary infection. TT encompasses a range of eyelid and eyelash abnormalities from a few peripheral in turned lashes to the entire upper eyelid pulled inwards by scarring (entropion). TT can also occur without entropion, from metaplastic or misdirected lashes [1]. Recent global estimate suggested that in 2008 there were 8.2 million people living with trichiasis. Surgical treatment for TT is a key component of the SAFE strategy for trachoma control, directly reducing the risk of blindness [2].
Several different surgical procedures have been used to correct upper lid TT, some of which have been compared in randomized trials [3]–[6]. The technique of Bilamellar Tarsal Rotation (BLTR) has the lowest recurrence risk. However, the widely used Posterior Lamellar Tarsal Rotation (PLTR or ‘Trabut’ procedure) was not included in these comparisons. The World Health Organisation (WHO) advocates either BLTR or PLTR surgery for TT [7]. Both procedures involve a horizontal tarsotomy combined with everting sutures to rotate the inferior portion of the upper lid outwards [3].
TT surgery can prevent or reduce progression of corneal opacity, improve vision and relieve pain [6], [8], [9]. However, surgical outcomes are variable. Most studies of post-operative TT recurrence reports rates of 20%–40% at one-year, ranging from 7.4% at one year to 62% at three years [5], [6], [8], [10]–[22]. Recurrence is generally subdivided into early (before six months) and late (after six months). Early recurrence is probably attributable to a number of factors including the severity of the preoperative disease, the type and quality of the surgery, and post-operative wound healing events [8], [23], [24]. Substantial inter-surgeon variation of TT recurrence rates has been reported [8], [23]. After six months there is a steady accumulation of recurrence that probably results from progressive scarring disease [8], [22].
Serious surgical complications are rare in TT surgery. However, complications such as granuloma and lid contour abnormalities (notching) occur relatively frequently [23]. Granulomas are pedunculated masses of inflammatory tissue ranging in size from a few millimetres to over a centimetre. Larger granulomas can obscure the visual axis, and all except the smallest require surgical removal. The reported frequency of granulomas ranges from 0% to 14% [6], [10], [11], [16], [23], [25]–[28]. Lid notching, a focal overcorrection of the lid caused by irregular suture tension or an irregular tarsal incision is cosmetically unsatisfactory and may be associated with lagophthalmos, putting the cornea at risk [29].
During the course of two recently reported randomised controlled trials conducted in Ethiopia we recruited 1300 individuals with the full spectrum of TT type and severity, who received PLTR surgery with silk sutures, and were followed up for two year [30], [31]. This represents the largest data set to date on the results of PLTR surgery (recurrence risks, vision and other outcomes), and provides an opportunity to investigate outcomes in relation to the type and position of the trichiatic lashes.
The National Health Research Ethics Review Committee of the Ethiopian Ministry of Science and Technology, the London School of Hygiene & Tropical Medicine Ethics Committee and the Emory University Institutional Review Board approved this study. Informed consent was taken at the time of enrolment. The research adhered to the tenets of The Declaration of Helsinki. All participants gave written informed consent to take part in the study.
Two previously reported randomised controlled trials of the management of TT were conducted in Ethiopia from 2008 to 2010 [30], [31]. Each trial recruited 1300 individuals aged 18 years or older with previously unoperated TT: in each trial one arm comprised participants undergoing TT surgery with silk sutures. For the purpose of these studies, TT was defined as one or more lashes touching the eye or clear evidence of epilation (broken/re-growing lashes), without another obvious cause for the trichiasis, such as trauma, malignancy, involutional changes or severe blepharitis. Exclusion criteria were previous eyelid surgery and self-reported pregnancy.
Participants presented during TT surgical treatment campaigns in rural villages in the West Gojjam Zone, Amhara Regional State, which were advertised in local markets, churches and schools. Additionally, health extension workers from every sub-district (kebele) across West Gojjam were trained to recognize trichiasis. They visited each village in their kebele to identify potential participants.
In the first trial, individuals with major TT (>5 lashes touching the eye) were randomly allocated to PLTR surgery with either silk or polyglactan (vicryl) sutures. In the second trial individuals with minor TT (<6 lashes touching) were randomly allocated to either PLTR surgery with silk sutures or repeated epilation. In individuals with bilateral TT, one eye was randomly designated (sequentially selected from a blocked randomly generated list of right and left eyes) as the study eye and included in the analysis, although both eyes were treated. In both trials, participants were allocated to surgeons sequentially. Surgeons were not permitted to select specific participants and participants were not allocated according to severity.
The group described in this report is comprised of all the individuals who were randomly allocated to the PLTR with silk suture arms of the two studies. They represent the full spectrum of TT disease (both major and minor TT) and received exactly the same surgical intervention performed by the same group of surgeons.
Participants were examined immediately before surgery and again at 6, 12, 18 and 24 months. The methods used have been described in detail [30], [31]. Briefly, LogMAR visual acuity was measured in each eye. Participants' weight and height were measured in order to calculate the body mass index (BMI). Both eyes were examined for signs of trachoma using 2.5× magnification loupes and a bright torch. Baseline, one-year and two-year examinations were by a single ophthalmologist (SNR) and the six and 18-month examinations were by a single ophthalmic nurse (EH). The examiners were standardised to each other. Lashes touching the eye were counted (‘lash burden’) and sub-divided into the part of the eye contacted when looking straight ahead (corneal or peripheral (lateral or medial conjunctiva) and subdivided by the type of trichiatic lash (entropic, misdirected or metaplastic). Clinical evidence of epilation was defined as the presence of broken or newly growing lashes, or areas of absent lashes. In the absence of epilation, eyes with <6 trichiatic lashes were designated as minor TT and those with >5 lashes as major TT. In the presence of epilation, a clinical judgement was made of the number of epilated lashes, by assessing regrowing lash stubs that were pointing towards the globe; if the total trichiatic lashes+epilated lashes was <6, the lid was designated as having minor TT and >5 as major TT. Upper lid entropion was graded using a previously described system [1]. Corneal scarring was classified based on a modified WHO FPC grading system [32], [33]. Corneal opacity was graded in the field and with high-resolution digital photographs (Nikon D300, Nikon 105 mm macro lens). The eyelid was everted and the location of the muco-cutaneous junction (MCJ) graded [1]. Following surgery the presence/absence of notching and granuloma were noted. Notching was defined as central overcorrection of the lid such as to cause a clear deviation in contour of the lid margin. This would correspond to either moderate or severe lid contour abnormalities in a recently published grading system [34].
Surgery was performed under local anaesthesia using the technique previously described [30], [31]. Post-operatively, the operated eye was padded for a day and then tetracycline eye ointment was self-administered twice a day for two weeks. Five nurses, who had previously been trained in and were regularly performing PLTR surgery, performed the surgery. They were selected as the best surgeons from a larger group of 10, during a two-day standardisation workshop. The PLTR techniques of the five nurses were carefully observed and standardised to ensure that all performed the operation in the same way. The intra-operative quality of surgery was periodically reviewed during the course of the trials. Participants were seen at 7–10 days post-operatively, at which point silk sutures were removed. The presence of trichiasis and other complications were noted and treated as needed. Any individual who had five or more lashes at any follow-up examination was offered repeat surgery to be performed by a senior surgeon, within a few weeks of the follow-up assessment. Individuals in whom other ophthalmic pathology (e.g. cataract) was detected were referred to the regional ophthalmic services.
The primary outcome measure was trichiasis recurrence defined as either (1) one or more eyelashes touching the eye or (2) clinical evidence of epilation. Secondary outcome measures were surgical complications, entropion and conjunctivalisation.
Data were double entered into an Access (Microsoft) database and transferred to Stata 11 (StataCorp, College Station, TX). For participants who had bilateral surgery only the randomly designated ‘study eye’ was included in the analysis. The cumulative incidence of failure in each six-month block of follow-up was calculated using the Kaplan-Meier method. The associations of binary outcomes with exposures were assessed using logistic regression to estimate odds ratios (OR) and 95% confidence intervals (CI). Variables associated with the outcome on univariate analyses (p<0.2) were retained in the multivariable model. The p-values for the association between categorical variables and specific outcomes were calculated using likelihood ratios. For visual acuities of counting fingers or less, logMAR values were attributed: counting fingers: 2.0, hand movements: 2.5, perception of light: 3.0, no perception of light: 3.5.
We recruited 1300 individuals with previously unoperated trichiasis in at least one eye. No individuals refused participation. Recruitment took place between March and July 2008. The follow-up assessments were conducted at the following times: 6-month, August to December 2008; 12-month, February to July 2009; 18-months, August to December 2009; 24-months, January 2010 to May 2010. The final follow-up was conducted two months ahead of schedule because of the Ethiopian general election (May 2010), during which time follow-up would not have been possible. Almost all individuals (98.1%) were reviewed on at least one occasion.
Baseline characteristics are shown in Table 1; 650 eyes had major TT and 650 had minor TT. All participants were from the Amhara region of Ethiopia. The logMAR visual acuity was 0.3 (6/12 Snellen equivalent) or less in 853/1289 (66%) individuals it was possible to test. Of those with major TT, 359 had less than five lashes, but had evidence of epilation consistent with the eye having more than five trichiatic lashes. The majority of the participants were female (78.2% of those with major TT, 68.9% of those with minor TT: p = 0.005). Corneal opacity was present in 768 (59.9%) individuals of whom 419 (32% of all study eyes) had opacity within the central 4 mm of the cornea.
Overall, recurrence occurred in 315/1276 (24.7%) study eyes. Recurrence was more frequent in participants with pre-operative major TT (32.6%) compared to minor TT (16.9%): OR 2.39, 95%CI 1.83–3.11, p<0.005 (Figure 1 and Table 2). This association was found at each follow-up (Table 2). Within both TT groups, the risk of recurrence was much higher during the first six-month period compared to all subsequent periods (p<0.0001; Figure 1 and Table 2), with 58.0% and 48.1% of recurrences accruing during this initial period in major TT and minor TT participants, respectively. Thirty-eight participants had recurrence at the 6 months follow-up, but not at any subsequent timepoint, of whom ten had repeat TT surgery.
Overall, there was a significant reduction in mean lash burden between baseline (4.66 lashes) and 24 months (0.29 lashes; t-test p<0.0005) (Table 2). Amongst people with recurrence who were examined at 24 months, 14/198 (7.1%) of those with baseline major TT had more trichiasis at 24 months compared to baseline, compared with 2/107 (1.9%) of those with baseline minor TT (Table 2). There were similar risks of recurrence in right and left eyes in both groups (Table 2).
Individual surgeon's recurrence rates ranged from 17.7% to 52.6%. The pre-operative severity of cases operated by the different surgeons varied to a degree (X2: p = 0.035), as did their post-operative under-correction rates (X2 = 0.003) (Table 3). Their risk of recurrence was not affected by the variation in case mix (Table 4). There was no significant difference between the recurrence rate in the first 20 surgeries conducted by each surgeon (32/100, 32%) and the last 20 surgeries (29/100, 29%, p = 0.42).
Multivariable logistic regression modelling identified increased baseline TT severity, the presence of entropic lashes (compared with misdirected or metaplastic lashes without entropion), specific surgeons (No. 2 and No. 5) and older participant age as independent risk factors for recurrence (Table 4).
Early (noted by the 7–10 day suture removal follow-up) and later (seen at any subsequent follow-up) post-operative complications and their association with recurrent trichiasis are presented in Table 5. Over three-quarters (23/30, 77%) of individuals noted to have trichiasis at the 7–10 day follow-up had recurrent trichiasis at a subsequent follow-up (OR: 10.3, 95% C.I.: 4.33–24.23, p<0.001) (Table 6). Post-operative granuloma (OR: 0.39, 95% C.I.: 0.19–0.83, p = 0.014) and notching (OR: 0.44, 95% C.I.: 0.28–0.72, p = 0.001) were both significantly associated with lower recurrence rates (Table 6).
Univariate and multivariable associations for developing a granuloma and lid notching are shown in Tables 7. Surgeons (No. 1 and No. 4) who had the lowest recurrence rates also had significantly higher rates of granuloma and notching. There was no association between granuloma formation and visible suture fragments being left in the lid (X2: p = 0.495), gender (X2: p = 0.239) or younger (<41 years) age (X2: p = 0.41)
Surgery successfully corrected entropion: 1148 (93.5%) participants at 12 months and 1126 (92.1%) at two years had no entropion, compared to 327 (25.2%) at baseline (Table 8). Surgery reduced the entropion grade in 886/918 (97%) (Paired t-test: p<0.0001). Entropion grade worsened in three participants. In the 1213 participants with conjunctivalization of the lid margin at baseline, an improvement was seen in 699 (58%) (paired t-test: p<0.0001) and worsening in 29 individuals (Fig. 1a and b).
Trichiasis recurred in a quarter of study eyes by two years. Recurrence severity was very variable, ranging from a single peripheral metaplastic lash to complete entropion. However, only 13% of recurrences had more than five lashes touching they eye and there was a substantial reduction in lash burden. Therefore, although any recurrence is unsatisfactory, the likely severity of the recurrence should be considered in a balanced assessment of the risks and benefits of surgery.
Over half of all recurrences occurred by six months and the risk decreased significantly for each subsequent six-month period. Higher rates of early recurrence have been reported previously. In a study from The Gambia the recurrence at four years was 41%, over three quarters of which occurred within the first six months after surgery [22]. In a study from Southern Ethiopia, the recurrence rate at 6 weeks was 2.3% and at one year was 7.6% [17]. Taken together, these observations indicate the importance of understanding and addressing the determinants of early (<6 months) recurrence. A combination of risk factors are likely to be important: baseline disease severity, choice of procedure, the surgeon's ability, and early wound healing responses are likely to be dominant. Overall, PLTR was effective at correcting entropion, with only 8% of participants having residual entropion at the end of the follow-up. Conjunctivalization of the lid margin reversed; the epithelium with conjunctival appearance recedes and meibomion gland openings become surrounded by macroscopically normal looking skin, presumably in response to an altered epithelial environment following entropion correction.
Recurrence was twice as frequent in individuals with major TT pre-operatively. Furthermore, pre-operative entropic trichiasis (rather than misdirected or metaplastic) was an independent risk factor for recurrence. More severe pre-operative trichiasis is consistently a major risk factor for recurrent TT [8], [11], [15], [17]–[20], [22], [35], [36]. Such individuals generally have more conjunctival scarring and may have horizontal or vertical lid shortening. The lid surgery is technically more challenging as the anterior and posterior lamellae are more difficult to dissect and post-operatively there may be strong contractile forces pulling the lid back to an entropic position. These cases, who are at higher risk of sight threatening disease, should be treated by more experienced surgeons and have enhanced follow-up to detect recurrence. Interestingly, metaplastic lashes, even in the absence of entropion, appear to be ‘cured’ by surgery. It is unclear whether they cease to grow, or whether they are simply rotated far enough away from the globe.
In our study only the PLTR procedure was used and gave recurrence rates in the middle of the reported range for this procedure: 12% to 55%, with reported follow-up periods of between 3 months and four years [8], [10], [20]–[22], [26], [37], [38]. One randomised trial has compared the PLTR and BLTR procedures and found similar outcomes [10]. However, ophthalmologists performed all the surgery, follow-up was only three months and sample size insufficient to address the question. The range of outcomes in these different studies suggests that a comparative trial of PLTR and BLTR is required under more representative operational conditions to determine if one procedure is superior, particularly for more severe cases.
The surgeons in this study were selected for their surgical ability and given additional training. Their technique was intermittently observed. Nevertheless, two surgeons had significantly higher recurrence rates than the best performing surgeon. Surgeon No. 5, who had the highest recurrence rate, did operate on a higher proportion of major TT cases than the other surgeons, but remained an independent risk factor for recurrence after adjusting for baseline TT severity.
Inter-surgeon variability has previously been highlighted as a concern in trachoma surgery with one study finding recurrence rates ranging from 0–83% between surgeons [8]. Several factors may contribute to variable outcomes. Firstly, surgical training varies in quality and number of cases performed [39]. Secondly, supervision and refresher training is often sporadic and of variable quality and content, with many surgeons operating entirely independently [39], [40]. Thirdly, surgical volume may be low which may lead to loss of surgical skills. In cataract surgery, for example, higher volume is associated with better outcomes [41]. The WHO advises that a minimum of 10 TT procedures per month should be conducted [7]. Studies from Ethiopia and Tanzania found few high volume surgeons, with the vast majority of TT surgeons perform few cases [39], [42]. In our study surgeon 5 did perform less procedures than the other surgeons as she was dismissed for disciplinary matters mid-way through the trial. However, she still conducted over 150 procedures during the trials from which this study emanates, so low surgical volume is unlikely to explain the variation. Finally, despite attempts to standardise, subtle residual variation in technical ability and technique probably remain. For example, short incisions have been associated with increased recurrence following BLTR surgery (crude OR: 3.58, 95% C.I.: 1.39–9.23) [23]. The immediate post-operative lid position warrants further investigation: if this is predictive of outcome, immediate revision could be undertaken. In programmatic settings, if individual surgeons are underperforming this needs to be addressed. Ideally, they would receive refresher training and be reassessed. Unfortunately, TT surgical audit is rarely conducted, so poor performance is probably frequently missed.
Notching is focal external rotation or irregularity of part of the lid usually caused by excessive suture tension. Some authors include notching within a broader category of ‘lid contour abnormalities’ [11]. Large notches may cause lagophthalmos and disruption of the tear film, leading to corneal exposure. Notching can be cosmetically unsightly, in contrast to general overcorrection which is less noticeable. Other studies have reported notches in 6–30% for PLTR surgery and 0–14% in BLTR surgery [11], [21], [23], [27], [37]. The association between notching and reduced recurrence is not surprising, as notching usually reflects a degree of overcorrection. Notching occurred more frequently in older and less well nourished people (lower BMI), which may reflect age and nutrition-related reduction in tarsal plate rigidity, leading to a more pliable eyelid.
Granulomas usually develop at the incision site within weeks of surgery. They require excision when they are large. In ophthalmic surgery they have been described following tarsal rotation and chalazion surgery and found to be associated with residual suture fragments, male gender and younger age [23], [43]. Here we report an association between granulomas and a lower recurrence rate and increased baseline papillary inflammation. Granulomas do not usually develop following surgery that tightly closes the incision site. In tarsal rotation surgery, the everting sutures hold the lid in an out-turned position, which may slightly part the edges of the incision from where granulomas develop. With greater external rotation, the posterior incision is less well opposed, leading to more granulation tissue formation. Granulomas may therefore be an inevitable consequence of tarsal rotation surgery with a good degree of eversion.
This study has a number of limitations that potentially constrain the generalisation of the conclusions. It is possible that the results are better than those achieved under routine operational conditions. The five surgeons were selected for their technical skill, received additional training and supervision and performed relatively large volume surgery. They are, therefore, not truly representative of many ‘field’ TT surgeons, who typically perform few cases, have limited training and supervision [39], [42]. Participants were not randomly assigned to a surgeon, however, the risk of selection bias was low, as participants were allocated on a “first-come-first-served” as surgeons became available. Finally, only one operation type, the PLTR, was used for all cases.
Set against these limitations, this study has a number of strengths. Firstly, we report the results of a large number of operations performed in a standardised manner. Secondly, follow-up rates are high despite the inaccessibility of many participants; reducing follow-up bias. Finally, participants were representative of the spectrum of TT disease in the wider population of TT patients in Ethiopia, which remains the country with the highest prevalence of TT in the world.
Recurrence rates were comparable to previous studies. Baseline disease severity and inter-surgeon variation are major determinants of recurrent disease. However, PLTR surgery successfully corrected most entropion and much of the recurrence was minor, which may not represent a significant risk for most patients. The inter-surgeon variation in recurrence rates is concerning. Further research is needed to ascertain whether recurrence can be predicted immediately after surgery, and whether it can be ameliorated.
|
10.1371/journal.pgen.1006008 | Identification of a Functional Risk Variant for Pemphigus Vulgaris in the ST18 Gene | Pemphigus vulgaris (PV) is a life-threatening autoimmune mucocutaneous blistering disease caused by disruption of intercellular adhesion due to auto-antibodies directed against epithelial components. Treatment is limited to immunosuppressive agents, which are associated with serious adverse effects. The propensity to develop the disease is in part genetically determined. We therefore reasoned that the delineation of PV genetic basis may point to novel therapeutic strategies. Using a genome-wide association approach, we recently found that genetic variants in the vicinity of the ST18 gene confer a significant risk for the disease. Here, using targeted deep sequencing, we identified a PV-associated variant residing within the ST18 promoter region (p<0.0002; odds ratio = 2.03). This variant was found to drive increased gene transcription in a p53/p63-dependent manner, which may explain the fact that ST18 is up-regulated in the skin of PV patients. We then discovered that when overexpressed, ST18 stimulates PV serum-induced secretion of key inflammatory molecules and contributes to PV serum-induced disruption of keratinocyte cell-cell adhesion, two processes previously implicated in the pathogenesis of PV. Thus, the present findings indicate that ST18 may play a direct role in PV and consequently represents a potential target for the treatment of this disease.
| Pemphigus vulgaris is a life-threatening autoimmune skin blistering disease. A large body of evidence indicates that the propensity to develop this condition is in part genetically determined. Using a genome wide association approach, we recently identified pemphigus vulgaris-associated genetic variations in the vicinity of the ST18 gene. In the present study, we identify a risk variant residing within the ST18 promoter region which drives ST18 gene promoter activity in a p53/p63-dependent manner, which is in line with the fact that ST18 is up-regulated in the skin of PV patients. Using functional assays, we show that ST18 overexpression increases PV serum-induced expression of pro-inflammatory mediators, as well as augments PV serum-induced disruption of keratinocyte cell-cell adhesion, which are hallmarks of pemphigus pathogenesis. Our findings therefore support a direct role for ST18 in the pathogenesis of pemphigus vulgaris, and position ST18 as a new molecular target of potential interest for the treatment of disease. From a broader perspective, these observations underscore the importance of genetic variations affecting the susceptibility of target tissues to autoimmunity.
| Pemphigus refers to a group of autoimmune blistering disorders which affect mucocutaneous tissues [1,2]. Pemphigus vulgaris, the most common subtype of the disease, is estimated to have a worldwide annual incidence of 0.76–6.7 new cases per million [1] and is between 4- to 10-fold more common among Jews as compared with other populations [3]. The disease is characterized by the development of flaccid blisters over the skin and mucosal surfaces, which rupture easily to form large painful erosions with little tendency to heal and which, if left untreated, increase the probability of life-threatening complications [1]. Since the advent of corticosteroid treatment, mortality has dropped to 10%, though morbidity is still considerable [1].
PV is traditionally considered to result from abnormal desmosome function caused by circulating auto-antibodies (auto-Abs) directed against desmosomal antigens, mainly desmoglein (Dsg) 3 and Dsg1 [2], which in turn leads to loss of adhesion (acantholysis) between keratinocytes. More recently, additional pathogenetic mechanisms, not directly involving desmosome destabilization, have also been suggested to be operative in PV IgG-induced blister formation, such as activation of apoptosis; increased pro-inflammatory cytokine secretion; aberrant cell-cell signaling; and activation of muscarinic receptors uniquely expressed by basal keratinocytes [4]. The propensity to develop the disease is believed to be to a large extent genetically determined as attested by familial occurrence of PV, the presence of circulating PV IgG Abs in healthy first-degree relatives of PV patients and ethnic clustering [3,5–7]. This in turn offers the possibility to identify elements of importance to PV etiology through a genetic approach. We recently conducted a genome wide association study (GWAS) in a genetically homogenous population of Jewish extraction, followed by replication in two cohorts, and identified an association between PV and the ST18 gene locus [8]. Although ST18 encodes a transcription factor possibly regulating apoptosis and inflammation [9], two processes of potential relevance to PV [10,11], it is not clear whether this genetic association reflects causal involvement of ST18 in PV pathogenesis. In the present study, we examined the possibility that ST18 plays a direct role in PV pathogenesis.
Given the fact that a previous GWAS demonstrated an association between PV and variants in the vicinity of the ST18 gene locus [8], we aimed at characterizing this risk region and therefore performed targeted deep sequencing of the ST18 locus in 16 Jewish PV patients, initially comparing the sequencing results to the 1000 Genomes Project (1000GP) data (http://www.1000genomes.org). Since PV is a complex disease, our primary targets were common single nucleotide polymorphisms (SNPs) that were enriched in the sequenced cohort as compared with controls. We therefore examined all SNPs with non-zero frequencies in the public databases that were not in repetitive regions, totaling 789 SNPs (Fig 1). A case-control association analysis for each SNP, using chi-square and permutation test, led to the identification of a genomic haplotype block strongly associated with the propensity to develop PV (p<0.001) (Fig 1 and S1 Fig). We discovered that this risk haplotype harbors two variants previously found to be most significantly associated with PV in the original GWAS, rs4074067 and rs2304365 [8]. Collectively, the SNPs found within the risk haplotype block had a frequency of 50% in the sequenced cohort and of only 8% in the total 1000GP population.
Within the risk haplotype block, we identified a genetic variant, rs17315309 (Fig 1), which displayed significant association to PV based on the deep sequencing data. We then replicated this association in an independent set of 185 Jewish PV patients compared with 183 population-matched healthy controls (p<0.001; S1 Table).
A number of bioinformatic analyses suggested that this variant may be of functional importance and possibly up-regulate ST18 promoter activity. First, using HMR Conserved Transcription Factor Binding Site database, implemented in the UCSC Genome Browser (https://genome.ucsc.edu/), this variant was found to reside within a p53 transcription factor binding site consensus sequence (Fig 2a). The p53 binding motif consists of two very similar, closely located, half-sites, each 10 bp long [12,13], although it has been shown that one decamer is sufficient for p53 or p63 binding and activity [14–19]. Second, the binding motif harboring rs17315309 is located in an intron of ST18, upstream to the gene coding sequence and inside a 170 bp long DNAse hypersensitivity cluster (chr8:53207581–53207750, ENCODE; http://genome.ucsc.edu/ENCODE/), lending further support to the possibility that this region plays a regulatory role. Of note, p53 is known to recognize consensus binding motifs located proximal to the transcription start site of target genes, either within the promoter region or a gene intron [20]. Third, using the 100 Vertebrae Conservation by PhastCons, implemented in the UCSC Genome Browser (http://compgen.cshl.edu/phast/), we found that both rs17315309 and the p53/p63 binding motif are highly conserved with a maximum conservation score of 1 (range 0 to 1). Similarly, using Biobase Transfac Matrix Database (v7.0), implemented in the UCSC Genome Browser (http://www.gene-regulation.com/pub/databases.html), the 10 bp long binding motif containing rs17315309 was found to be remarkably conserved (computed score: 992; maximal score: 1000), supporting the possibility that it represents a biologically functional binding site. Lastly, as rs17315309 results in a T to C substitution at position 6 of the p53/p63 binding motif, we wished to examine the conservation of this nucleotide in the binding site consensus sequence. Both MotifMap (http://motifmap.ics.uci.edu) and Jasper (http://jaspar.genereg.net) databases indicated that position 6 in the half-binding site consensus sequence of both p53 and p63 is highly conserved and consists usually of a T nucleotide with a minimum to no abundance of a C nucleotide (Fig 2a). Taken collectively, these data suggested that modification of the wild type rs17315309 allele within the p53/p63 binding site may be of biological significance.
To examine this possibility, we transfected normal human keratinocytes (NHEKs) with a PGL4.17 luciferase reporter construct under the regulation of a 282 bp fragment spanning the p53/p63 binding motif containing either the wild type (T) or the risk (C) rs17315309 allele. The C allele was found to induce a more than 5-fold increase in luciferase activity (p<0.001) (Fig 2b). Moreover, when the construct was co-transfected with p53- or p63-specific siRNAs (S2a and S2b Fig), this effect was markedly attenuated (Fig 2b), indicating that the PV-associated rs17315309 risk allele increases ST18 promoter activity in a p53-/p63-dependent manner. These results are in line with previous data showing that a single nucleotide change in a canonical p53/p63 binding sequence is enough to affect p53 or p63 binding [21].
Given these data, the physiological roles of ST18 and the fact that ST18 is markedly overexpressed in the non-lesional epidermis of PV patients [8], we sought to ascertain the consequences of ST18 overexpression on pathophysiological hallmarks of the disease.
We first examined the effect of ST18 on keratinocyte secretion of pro-inflammatory cytokines which are believed to contribute to PV disease phenotype [1,4,11]. Overexpression of ST18 (S2c Fig) in the presence of normal serum or control IgG did not affect the secretion of either TNFα, IL-1α or IL-6 (Fig 3). In contrast, when overexpressed in the presence of PV serum, ST18 was found to drive the secretion of all three cytokines (Fig 3a–3c), indicating that ST18 functions by promoting PV-induced keratinocyte secretion of pro-inflammatory cytokines. This effect was seen early with TNFα and IL-1α but late with IL-6 (Fig 3a–3c). We then repeated the same experiments, comparing the effect of control and PV sera to the effect of control IgG and PV IgG. Overexpression of ST18 was found to increase PV serum-induced and PV IgG-induced secretion of TNFα, IL-1α and IL-6, to the same extent, while not affecting the secretion of these cytokines in the presence of control serum or control IgG (Fig 3d–3f).
As disruption of epidermal cell-cell adhesion is a pathogenic hallmark of PV [1], we investigated ST18 effect on PV serum-induced cell-cell disadhesion. For this purpose, we used the dispase-based dissociation assay. In this system, PV serum destabilizes intercellular bonds, compromising epidermal sheet resilience to mechanical stress [22] (Fig 4a). NHEKs overexpressing ST18 and exposed to PV serum exhibited a more than 2-fold decrease in cell-cell adhesion, as compared to cells exposed to PV serum and transfected with an empty vector (p<0.05) (Fig 4b) or with a vector overexpressing a non-relevant gene (S3 Fig). The deleterious effect of ST18 overexpression on cell-cell adhesion was similar when cells were exposed to pooled PV sera or to PV IgG (Fig 4b). Taken together, these results demonstrate that ST18 may also contribute to PV pathogenesis by potentiating PV IgG-induced acantholysis.
Despite a large body of epidemiological evidence supporting a role for genetic elements in determining the propensity to develop the disease [3,5–7], little is currently known about the genetic basis of PV. Using a GWAS approach, we previously identified an association between PV and a genomic segment on chromosome 8q11 spanning the ST18 gene [8]. ST18 encodes the suppression of tumorigenicity 18 (ST18), a 115kD member of the myelin transcription factor 1 (MyT1) family of transcription factors containing several zinc-finger DNA-binding domains [23]. ST18 is constitutively expressed in the brain, and less so in the heart, liver, kidney, skeletal muscle, pancreas, testis, ovary and prostate [24]. ST18 is undetectable in normal skin, but is significantly expressed in the epidermis of PV patients [8]. Two recent studies [9,25] demonstrated the role of ST18 in apoptosis and inflammation, two processes of direct relevance to the pathogenesis of PV [10,11]. ST18 was shown to mediate TNFα-induced transcription of pro-apoptotic and pro-inflammatory genes in fibroblasts, including TNFα, IL-1α and IL-6 [9].
In the present study, using targeted deep sequencing of the ST18 locus, we identified within the ST18 promoter region a PV-associated genetic variant, rs17315309, which was shown to drive gene transcription in a p53/p63-dependent fashion. Of interest PV serum was previously found to induce p53 expression [26] and p63 is overexpressed in the skin of pemphigus foliaceus patients [27].
Together with the fact that ST18 expression is up-regulated in the skin of PV patients [8], these data suggested that ST18 overexpression may be directly contributing to the disease pathogenesis. And indeed, ST18 up-regulation was found to induce the secretion of TNFα, IL-1α and IL-6 in the presence of PV serum as well as in the presence of PV IgG antibodies. The level of all three cytokines has been previously reported to be increased in the serum as well as in the lesional skin and blister fluid of PV patients [11,28–34]. In addition, serum levels of TNFα and IL-6 were shown to negatively influence PV outcome [28,35]. Finally, both TNFα and IL-1α have been reported to be up-regulated by PV IgG and to contribute to PV IgG-induced acantholysis and apoptosis in keratinocytes [32,36].
Most importantly, ST18 was found to potentiate PV IgG-induced cell-cell disadhesion. The mechanism of action of ST18 in inducing cell-cell disadhesion remains to be fully elucidated but may involve some of the cytokines whose secretion was found to be increased in the presence of ST18 overexpression [32,36]. Clearly, other pro-inflammatory factors may also contribute to keratinocyte dissociation. The fact that PV IgG cause secretion of cytokines and loss of cell adhesion to a comparable extent as PV serum demonstrates that ST18 promotes the effect of autoantibodies rather than serum factors on the secretion of cytokines and on loss of cell adhesion.
Taken collectively, our data indicate that a PV-associated risk allele at the ST18 gene locus may drive ST18 up-regulation which in turn could contribute to PV pathogenesis by stimulating keratinocyte-derived cytokine release and by compromising epidermal cell-cell adhesion. Thus, ST18 is likely to contribute to PV pathogenesis by increasing keratinocytes susceptibility to the deleterious effects of PV-associated autoantibodies rather than by affecting the production of these antibodies. Supporting this possibility, we did not detect any effect of ST18 genotype on anti-Dsg3 ELISA status in a series of PV patients (S4 Fig). The present results therefore underscore the importance of genetic variations affecting target tissues in the pathogenesis of inflammatory diseases as previously shown for other skin disorders [37,38].
The study was conducted according to a protocol approved by our institutional review board and the National Committee for Genetic Studies of the Israeli Ministry of Health in accordance with the Declaration of Helsinki Principles (102-2006/TLV-0537-15). All family members provided written informed consent to participate in this study
All family members provided written informed consent to participate in this study. Blood samples were obtained from all participants according to a protocol approved by our institutional review board and the National Committee for Genetic Studies of the Israeli Ministry of Health in accordance with the Declaration of Helsinki Principles. The diagnosis of PV was posed based upon clinical features, suprabasal separation on histology, positive direct and indirect immunofluorescence microscopy, and/or ELISA detection of anti-Dsg Abs. Genomic DNA was extracted from peripheral blood leukocytes using the 5 Prime ArchivePure DNA Blood kit (5 Prime Inc., Gaithersburg, MD, USA).
We used two different mixes of pooled sera from newly diagnosed PV patients (n = 3 and 4) with active disease and an anti-Dsg3 titer above122 relative units/ml, as measured by the anti-desmoglein 3 ELISA (IgG) test kit (Euroimmune AG, Luebeck, Germany). The sera were obtained prior to the initiation of any systemic immunosupressive treatment. PV IgGs were purified as previously described [39] and used at a final concentration of 65 μg/ml.
DNA enrichment was performed using HaloPlex kit (Agilent Technologies, Santa Clara, CA, USA) and sequencing was conducted on a MiSeq system sequencer (Illumina, San Diego, CA, USA) with 150 bp paired-end reads. A total of 463407 bp were included in the capture design, covering the entire ST18 gene (chr8: 53,023,399–53,373,519, GRCh37/hg19 assembly) as well as 10 kb downstream and 50 kb upstream to the gene and an additional 2 Mb located upstream and downstream to the gene and predicted to harbor putative regulatory regions (https://www.encodeproject.org).
The sequencing data were processed using MiSeq Reporter 2.0.26 and Casava softwares (Illumina, San Diego, CA, USA) and analyzed for quality control using FastQC software (http://www.bioinformatics.babraham.ac.uk/projects/fastqc). Reads were aligned to the Genome Reference Consortium Human Build 37 (GRCh37/hg19) using Burrows-Wheeler Aligner [40] and variant detection was achieved using The Genome Analysis Toolkit [41]. Variants were annotated by ANNOVAR [42] and the frequency of each variant was determined using data from dbSNP138, the 1000 Genome Project and an in-house database. Case-control association test for variants was performed with chi-square, and permutation test, using the Caucasian population from the 1000 Genome Project data (http://www.1000genomes.org) as a control.
To screen for the rs17315309 allele, we PCR-amplified a 317 bp fragment, with ReddyMix PCR Master Mix (Thermo scientific, NH, USA) and the following primers 5`- TGCTTGCCGTTTGTAAGATG-3`and 5`-AGCCTGGTTCAAGAGCCTTC-3`. Cycling conditions were as follows: 94°C, 4min; 94°C, 30 sec; 61°C, 30 sec; 72°C 30 sec, for 2 cycles, 94°C, 30 sec; 59°C, 30 sec; 72°C 30 sec, for 2 cycles, 94°C, 30 sec; 57°C, 30 sec; 72°C 30 sec, for 38 cycles, 72°C for 10 min. The T allele is associated with the presence of a recognition site for endonuclease NspI (New England Biolabs, Hitchin, UK). After incubation at 37°C for 16 hours followed by 20 min of inactivation at 65°C, the digested PCR products were electrophoresed in ethidium bromide-stained 3% agarose gels.
NHEKs were extracted from skin discarded during plastic surgery procedures, after written informed consent had been obtained as previously described [43]. The keratinocytes were seeded on feeder plates containing 3T3-J2 fibroblasts and were grown in medium Green containing 42.5% DMEM (Biological Industries, Beit-Haemek, Israel), 42.5% DMEM/F12 (Biological Industries, Beit-Haemek, Israel), 10% FCII serum, 1% penicillin-streptomycin, 1mM L-glutamine, 1mM sodium pyruvate, 5 μg/mL Insulin, 0.2 mM adenine, 0.5 μg/mL hydrocortisone, 2nM Triiodothyronine, 0.1 nM cholera toxin and 10 ng/mL EGF, and were frozen upon 90% confluence. Cell were then thawed and cultured in KGM media (Lonza, Basel, Switzerland).
For the dispase-based dissociation assay, NHEKs were extracted from foreskin using the same conditions and were thawed and cultured in M154 media (Life Technologies, Carlsbad, CA).
A 8.0 kb-clone containing the ST18 open reading frame in a pCMV6-Entry vector (4.9kb) was purchased from Origene Technologies Company (Rockville, MD, USA). The empty pCMV6-Entry was used as a control. Additionally, a 6.6 kb-clone containing the CNBD2 open reading frame in a pCMV6-Entry vector (4.9kb) was purchased from Origene Technologies Company (Rockville, MD, USA) and used a negative control for protein overexpression. NHEKs were cultured to 80% confluence and subjected to a transient transfection using lipofectamine2000 (Life Technologies, Carlsbad, CA).
A 282 bp ST18 gene fragment spanning rs17315309 was PCR-amplified using ReddyMix PCR Master Mix (Thermo scientific, NH, USA), primers 5’-AAAATTAGGTACCGCGTTCAAGCACTCTATTACCT-3’ and 5’-AAAAGGACTCGAGGCTTGCCGTTTGTAAGATGA-3’, and DNA extracted from two patients homozygous for rs17315309 wild-type allele T, and for rs17315309 minor allele C, respectively. Cycling conditions were as follows: 94°C, 4 min; 94°C, 30 sec; 61°C, 30 sec; 72°C 30 sec, for 2 cycles, 94°C, 30 sec; 59°C, 30 sec; 72°C 30 sec, for 2 cycles, 94°C, 30 sec; 57°C, 30 sec; 72°C 30 sec, for 38 cycles, 72°C for 10 min. The resulting amplicons were cloned into pGL4.17 vector (Promega, Madison, WI, USA).
NHEKs were co-transfected with the various pGL4.17 vectors and Renilla expression vector and control siRNAs (Life Technologies, Carlsbad, CA) or p53 specific siRNA (Santa Cruz Biotechnology, Santa Cruz, CA) or p63 specific siRNA (Dharmacon, Inc., Lafayette, CO) with lipofectamine2000 and Opti-MEM medium (Life Technologies, Carlsbad, CA). Efficiency of gene knock down was assessed by qRT-PCR (S2 Fig). Cells were grown in KGM medium (Biological Industries, Beit-Haemek, Israel). Twenty-four hours post transfection, cells were harvested and luciferase expression was evaluated using the Dual-Luciferase Reporter Assay System (Promega, Madison, WI, USA) and Tecan Infinite M200 device (Tecan Group Ltd, Männedorf, Switzerland)
Supernatant collected from NHEKs was evaluated using Elisa assays specific for IL-1α (Human IL-1 alpha/IL-1F1 DuoSet, R&D systems, Minneapolis, MN, USA), TNF-α (Human TNF-alpha Quantikine HS ELISA, R&D systems, Minneapolis, MN, USA) and IL-6 (Human IL-6 DuoSet, R&D systems, Minneapolis, MN, USA). All ELISA assays were read and quantified using Tecan Infinite M200 device (Tecan Group Ltd, Männedorf, Switzerland). For IL-6, a Human Cytokine Array / Chemokine Array 41-Plex (Eve Technologies Corporation, Calgary, Alberta, Canada) was additionally used using a Millpore MILLIPLEX kit (Merck KGaA, Darmstadt, Germany) read by BioPlex 200 (Bio-Rad, Hercules, CA, USA)
NHEKs were grown to confluence in triplicates on 6-well plates and exposed to PV or normal serum in a 1:10 dilution or to PV or control IgG antibodies in a final concentration of 65 μg/ml. After 24h the cells were washed twice with PBS, incubated in 2 ml of dispase II (2.4 units/ml, Roche Diagnostics, Basel, Switzerland) at 37°C for 40 minutes and detached from the plate as monolayers. Cell sheets were carefully transferred to 15 ml tube containing 5 ml PBS and subjected to mechanical stress using 5 inversions. The number of fragments was counted by two independent evaluators.
All pairwise comparisons were performed using the 2-tailed Student’s t test, unless otherwise indicated. Differences were considered significant if the P value was less than 0.05.
|
10.1371/journal.pntd.0006609 | Vector competence of biting midges and mosquitoes for Shuni virus | Shuni virus (SHUV) is an orthobunyavirus that belongs to the Simbu serogroup. SHUV was isolated from diverse species of domesticated animals and wildlife, and is associated with neurological disease, abortions, and congenital malformations. Recently, SHUV caused outbreaks among ruminants in Israel, representing the first incursions outside the African continent. The isolation of SHUV from a febrile child in Nigeria and seroprevalence among veterinarians in South Africa suggests that the virus may have zoonotic potential as well. The high pathogenicity, extremely broad tropism, potential transmission via both biting midges and mosquitoes, and zoonotic features of SHUV require further investigation. This is important to accurately determine the risk for animal and human health, and to facilitate preparations for potential epidemics. To gain first insight into the potential involvement of biting midges and mosquitoes in SHUV transmission we have investigated the ability of SHUV to infect two species of laboratory-colonised biting midges and two species of mosquitoes.
Culicoides nubeculosus, C. sonorensis, Culex pipiens pipiens, and Aedes aegypti were orally exposed to SHUV by providing an infectious blood meal. Biting midges showed high infection rates of approximately 40%-60%, whereas infection rates of mosquitoes were only 0–2%. Moreover, successful dissemination in both species of biting midges and no evidence for transmission by orally exposed mosquitoes was found.
The results of this study suggest that different species of Culicoides midges are efficient in SHUV transmission, while the involvement of mosquitoes has not been supported.
| Arthropod-borne (arbo)viruses are notorious for causing unpredictable and large-scale epidemics and epizootics. Apart from viruses such as West Nile virus and Rift Valley fever virus that are well-known to cause a significant impact on human and animal health, many arboviruses remain neglected. Shuni virus (SHUV) is a neglected virus with zoonotic characteristics that was recently associated with severe disease in livestock and wildlife. Isolations from field-collected biting midges and mosquitoes suggests that SHUV may be transmitted by these insects. In this study, four main vectors that transmit other arboviruses were selected to test their susceptibility to SHUV. Laboratory-reared biting midge species (Culicoides nubeculosus and C. sonorensis) and mosquito species (Culex pipiens pipiens and Aedes aegypti) were exposed to SHUV via an infectious blood meal. SHUV was able to successfully disseminate in both biting midge species, whereas no evidence of transmission by both mosquito species was found. Our results suggest that SHUV can be transmitted efficiently by diverse Culicoides species, and thereby that these insects could play a major role in the disease transmission cycle.
| Arthropod-borne (arbo)viruses continue to pose a threat to human and animal health [1, 2]. In particular the order Bunyavirales comprises emerging pathogens such as Crimean-Congo haemorrhagic fever virus (CCHFV) and Rift Valley fever virus (RVFV) [3, 4]. The World Health Organization (WHO) has included both CCHFV and RVFV to the “Blueprint” list of ten prioritized viruses likely to cause future epidemics and for which insufficient countermeasures are available [5]. In the veterinary field, prioritized viral diseases of animals, including RVFV, are notifiable to the World Organization for Animal Health (Office International des Epizooties, OIE). Apart from pathogens that are recognised as major threats by WHO and OIE, many have remained largely neglected. Before the turn of the century, West Nile virus, chikungunya virus, and Zika virus were among these neglected viruses until they reminded us how fast arboviruses can spread in immunologically naïve populations [2]. Although these outbreaks came as a surprise, in hindsight, smaller outbreaks in previously unaffected areas could have been recognised as early warnings.
Shuni virus (SHUV; family Peribunyaviridae, genus Orthobunyavirus, Simbu serogroup) is a possible arbovirus that recently emerged in two very distant areas of the world [6]. SHUV was isolated for the first time from a slaughtered cow in the 1960s in Nigeria [7]. During subsequent years, the virus was isolated on several occasions from domestic animals including cattle, sheep, goats, and horses [7–10], from wild animals including crocodiles and rhinoceros [10], and from field-collected Culicoides biting midges and mosquitoes [8, 11, 12]. More recently, SHUV was associated with malformed ruminants in Israel [13, 14]. Emergence of SHUV in areas outside Sub-Saharan Africa shows the potential of this virus to spread to new areas, and increases the risk for SHUV outbreaks in bordering territories such as Europe. Isolation of SHUV from a febrile child and detection of antibodies in 3.9% of serum samples from veterinarians in South Africa shows that SHUV can infect humans as well, although its ability to cause human disease is still uncertain [7, 15, 16].
Proper risk assessments rely on accurate knowledge of disease transmission cycles. Arbovirus transmission cycles can only become established when competent vectors and susceptible hosts encounter under suitable climatic conditions. Although SHUV has been isolated from pools of field-collected Culicoides biting midges and mosquitoes [7, 11, 12], the role of both insect groups as actual vectors remains to be confirmed. Detection of virus in field-collected insects is not sufficient to prove their ability to transmit the virus. Arboviruses need to overcome several barriers (i.e. midgut and salivary gland barriers) inside their vector, before they can be transmitted [17, 18]. In addition to virus isolation from field-collected vectors, laboratory studies are therefore needed to experimentally test the ability of vectors to become infected with, maintain, and successfully transmit arboviruses (i.e., vector competence) [19]. To gain insights into the potential of Culicoides biting midges and mosquitoes to function as vectors of SHUV, we studied the susceptibility of four main arbovirus vector species (Culicoides nubeculosus and C. sonorensis biting midges, and Culex pipiens pipiens and Aedes aegypti mosquitoes) for SHUV.
African green monkey kidney cells (Vero E6; ATCC CRL-1586) were cultured in Eagle’s minimum essential medium (Gibco, Carlsbad, CA, United States) supplemented with 5% fetal bovine serum (FBS; Gibco), 1% non-essential amino acids (Gibco), 1% L-glutamine (Gibco), and 1% antibiotic/antimycotic (Gibco). Cells were cultured as monolayers and maintained at 37°C with 5% CO2.
Vero E6 cells that were used in biting midge and mosquito infection experiments in the biosafety level 3 (BSL3) facility were cultured in Dulbecco's modified Eagle medium (Gibco) supplemented with 10% FBS, penicillin (100 U/ml; Sigma-Aldrich, Saint Louis, MO, United States), and streptomycin (100 μg/ml; Sigma-Aldrich). Prior to infections in the BSL3 facility, Vero E6 cells were seeded in 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid-buffered DMEM medium (HEPES-DMEM; Gibco) supplemented with 10% FBS, penicillin (100 U/ml), and streptomycin (100 μg/ml), fungizone (50 μg/ml; Invitrogen, Carlsbad, United States), and gentamycin (50 μg/ml; Gibco).
C6/36 cells (ATCC CRL-1660), derived from Aedes albopictus mosquitoes, were cultured in L15 medium (Sigma-Aldrich) supplemented with 10% FBS, 2% Tryptose Phosphate Broth (Gibco), 1% nonessential amino acids solution, and 1% antibiotic/antimycotic. Cells were cultured as monolayers and incubated at 28°C in absence of CO2.
KC cells [20], derived from embryos of colonized C. sonorensis biting midges, were cultured as monolayers in modified Schneider’s Drosophila medium (Lonza, Basel, Switzerland) with 15% FBS and 1% antibiotic/antimycotic at 28°C in absence of CO2.
SHUV (strain An10107, P2 Vero, 1980) was kindly provided by the World Reference Center for Emerging Viruses and Arboviruses (WRCEVA). The virus was originally isolated from the blood of a slaughtered cow in 1966 in Nigeria by inoculation of neonatal mice [21]. The P3 cell culture stock was generated by inoculation of Vero E6 cells at a multiplicity of infection (MOI) of 0.001. The supernatant was harvested at 6 days post inoculation, centrifuged, and stored in aliquots at -80°C. The P4 stock was generated by inoculating Vero E6 cells at MOI 0.01 using the P3 stock. At this MOI, full cytopathic effect (CPE) was present at 3 days post infection. Virus titers were determined using endpoint dilution assays (EPDA) on Vero E6 cells [22]. Titers were calculated using the Spearman-Kärber algorithm and expressed as 50% tissue culture infective dose (TCID50) [23, 24].
Cells were seeded in T25 cell culture flasks at densities of 7.5 × 105 (Vero E6), 1.5 × 106 (C6/36) or 2.5 × 106 (KC cells) per flask in 10 ml complete medium. After overnight incubation, the flasks were inoculated with SHUV at an MOI of 0.01 (P4 stock). The MOI calculation for each cell line was based on the virus titer that was determined on Vero E6 cells. One hour after inoculation, the medium was removed and replaced with fresh medium. At time points 0 (sample taken directly after medium replacement), 24, 48 and 72 h post infection, 200 μl samples were taken and stored at -80°C for later analysis. Virus titers were determined by EPDA using Vero E6 cells [22].
Culicoides nubeculosus were kindly provided by the Institute for Animal Health (IAH), Pirbright Laboratory, United Kingdom, in 2012 [25], and were maintained at 23°C with 16:8 light:dark cycle and 60% relative humidity. Culicoides sonorensis were kindly provided by the Arthropod-Borne Animal Diseases Research Laboratory, USDA-ARS (courtesy of Dr. Barbara Drolet) in 2017, and were maintained at 25°C with 16:8 light:dark cycle and 70% relative humidity. Similar rearing protocols were used for both biting midge species. Eggs were transferred to square larval holding trays (C. nubeculosus: 25 x 25 x 8 cm, Kartell, Noviglio, Italy; C. sonorensis: 19 x 19 x 20 cm, Jokey, Wipperfürth, Germany) with filter wool (Europet Bernina International, Gemert-Bakel, The Netherlands) attached with double-sided tape to the bottom. Trays were filled with tap water, a few millilitres of rearing water in which larvae had completed their life cycle, and two drops of Liquifry No.1 (Interpet, Dorking, United Kingdom). Larvae were fed with a 1:1:1 mixture of bovine liver powder (MP biomedicals, Irvine, CA, US), ground rabbit food (Pets Place, Ede, The Netherlands), and ground koi food (Tetra, Melle, Germany). Culicoides nubeculosus larvae were additionally fed with nutrient broth No. 2 (Oxoid, Hampshire, UK). Pupae were transferred to buckets (diameter: 12.2 cm, height: 12.2 cm; Jokey), and provided with 6% glucose solution ad libitum. Cow blood (Carus, Wageningen, The Netherlands) was provided through a Parafilm M membrane using the Hemotek PS5 feeding system (Discovery Workshops, Lancashire, United Kingdom) for egg production.
The Culex pipiens pipiens colony was established in the laboratory from egg rafts collected in the field in The Netherlands during August 2016. Egg rafts were individually hatched in tubes. Pools of approximately 10 first instar larvae were identified to the biotype level using real-time PCR [26]. The colony was started by grouping larvae from 93 egg rafts identified as the pipiens biotype. Mosquitoes were maintained at 23°C with 16:8 light:dark cycle and 60% relative humidity [27, 28]. Adult mosquitoes were kept in Bugdorm-1 rearing cages and maintained on 6% glucose solution ad libitum. Cow blood or chicken blood (Kemperkip, Uden, The Netherlands) was provided through a Parafilm M membrane using the Hemotek PS5 feeding system for egg production. Egg rafts were transferred to square larval holding trays (25 x 25 x 8 cm, Kartell) filled with tap water and two drops of Liquifry No. 1. Hatched larvae were fed with a 1:1:1 mixture of bovine liver powder, ground rabbit food, and ground koi food. Pupae were collected every 2 days and placed in Bugdorm-1 insect rearing cages.
Aedes aegypti mosquitoes from the Rockefeller strain (Bayer AG, Monheim, Germany) were used in all experiments. The mosquito colony was maintained as described before [29]. In short, mosquitoes were maintained at 27°C with 12:12 light:dark cycle and 70% relative humidity. Adult mosquitoes were kept in Bugdorm-1 rearing cages and maintained on 6% glucose solution ad libitum. Human blood (Sanquin Blood Supply Foundation, Nijmegen, The Netherlands) was provided through a Parafilm M membrane using the Hemotek PS5 feeding system for egg production. Eggs were transferred to transparent square larval holding trays (19 x 19 x 20 cm, Jokey), filled for approximately one-third with tap water and three drops of Liquifry No. 1. Hatched larvae were fed with Tetramin Baby fish food (Tetra). Larval trays were closed with fine-meshed netting, to allow adult mosquitoes to emerge inside larval trays. Twice a week, adults were aspirated from larval trays and collected in Bugdorm-1 insect rearing cages.
Groups of adult C. nubeculosus (1–6 days old), C. sonorensis (1–11 days old), Cx. p. pipiens (4–20 days old), and Ae. aegypti (4–11 days old) were transferred to plastic buckets (diameter: 12.2 cm, height: 12.2 cm; Jokey) closed with netting before being taken to the BSL3 facility. Culex p. pipiens mosquitoes were kept on water for 3 days, whereas the other species were maintained on 6% glucose solution until being offered an infectious blood meal. SHUV P3 stock with a mean titer of 3.0 x 106 TCID50/ml was mixed 1:1 with cow blood. The used cow blood was tested negative for Schmallenberg virus (SBV) antibodies, to prevent cross-neutralisation with SHUV. The infectious blood meal was provided through Parafilm M membrane using the Hemotek PS5 feeding system, under dark conditions at 24°C and 70% relative humidity. After 1 h, insects were anesthetized with 100% CO2 and kept on a CO2-pad to select fully engorged females. For each species, five fully engorged females were directly stored at -80°C for each replicate. These samples were used to determine the ingested amounts of SHUV for each species. All remaining and fully engorged females were placed back into buckets with a maximum group size of 110 individuals per species per bucket. All insects were provided with 6% glucose solution ad libitum. Culicoides sonorensis and Ae. aegypti were kept at 28°C for 10 days, whereas C. nubeculosus and Cx. p. pipiens were kept at 25°C for 10 days. These temperatures were selected for optimal replication of the virus, and to reflect differences in the natural environmental temperature for each species. Three replicates of C. nubeculosus (total N = 243), C. sonorensis (total N = 48), and Cx. p. pipiens (total N = 211) were carried out, and two replicates of Ae. aegypti (total N = 149). During each replicate, biting midges and mosquitoes were fed in parallel with the same infectious blood meal.
Adult female Cx. p. pipiens (3–9 days old) and Ae. aegypti (4–6 days old) mosquitoes were injected with SHUV into the thorax to investigate the role of mosquito barriers on dissemination of SHUV. Mosquitoes were anesthetized with 100% CO2 and positioned on the CO2-pad. Female mosquitoes were intrathoracically injected with 69 nl of SHUV (P3 stock with a titer of 3.0 x 106 TCID50/ml) using a Drummond Nanoject II Auto-Nanoliter injector (Drummond Scientific, Broomall, Unites States). Injected Cx. p. pipiens were maintained at 25°C and injected Ae. aegypti were maintained at 28°C. Mosquitoes were incubated for 10 days at the respective temperatures, and had access to 6% glucose solution ad libitum. Injections were done during a single replicate for Cx. p. pipiens (N = 50) and Ae. aegypti (N = 50).
After 10 days of incubation at the respective incubation temperatures, samples from surviving biting midges and mosquitoes were collected. Biting midges were anesthetized with 100% CO2 and transferred individually to 1.5 ml Safe-Seal micro tubes (Sarstedt, Nümbrecht, Germany) containing 0.5 mm zirconium beads (Next Advance, Averill Park, NY, United States). For a selection of C. nubeculosus (N = 77) and C. sonorensis (N = 30), heads were removed from bodies and separately stored in tubes. All samples were stored at -80°C until further processing.
Mosquitoes were anesthetized with 100% CO2 to remove legs and wings. Mosquito saliva was then collected by inserting the proboscis into a 200 μl yellow pipet tip (Greiner Bio-One) containing 5 μl of a 1:1 solution of 50% glucose solution and FBS. The saliva sample was transferred to a 1.5 ml micro tube containing 55 μl of fully supplemented HEPES-DMEM medium. Mosquito bodies were individually stored in 1.5 ml Safe-Seal micro tubes containing 0.5 mm zirconium beads.
Frozen biting midge and mosquito tissues were homogenized for 2 min at maximum speed in the Bullet Blender Storm (Next advance), centrifuged for 30 seconds at 14,500 rpm in the Eppendorf minispin plus (Eppendorf, Hamburg, Germany), and suspended in 100 μl of fully supplemented HEPES-DMEM medium. Samples were blended again for 2 min at maximum speed, and centrifuged for 2 min at 14,500 rpm. Mosquito saliva samples were thawed at RT and vortexed before further use. In total 30 μl of each body or saliva sample was inoculated on a monolayer of Vero E6 cells in a 96 wells plate. After 2–3 h the inoculum was removed and replaced by 100 μl of fully supplemented HEPES-DMEM medium. Wells were scored for virus induced CPE at 3 and 7 days post inoculation. Virus titers of infected biting midge bodies and heads, as well as mosquito bodies and saliva were determined with EPDA on Vero E6 cells [29]. Virus titers were determined using the Reed & Muench algorithm [30].
Infection, dissemination, and transmission rates were calculated, respectively, by dividing the number of females with virus-containing body (infection), virus-containing head (dissemination), or virus-containing saliva (transmission) by the total number of females tested in the respective treatment, and multiplied by 100. Dissemination and transmission success was calculated by dividing the number of virus-positive head or saliva samples, respectively, by the number of virus-positive bodies, and multiplied by 100. Two biting midge samples of which only the head was virus-positive, but not the body, were excluded from further analysis.
Mammalian, mosquito, and midge cells were inoculated with SHUV to gain insight into the replicative fitness of this virus and strain in different host cell types. The results show that SHUV is capable to produce progeny in all three cell types (Fig 1). Of note, a strong CPE was observed in the Vero E6 cells upon infection whereas no CPE was observed in the insect cell lines.
To evaluate the susceptibility of two species of biting midges (C. nubeculosus and C. sonorensis) for SHUV, groups of individuals of both species were orally exposed to an infectious blood meal with a mean SHUV titer of 3.0 x 106 TCID50/ml. SHUV titers of ingested blood were determined for a selection of 10 fully engorged females for each species, that were directly stored at -80°C after feeding. Both species ingested low titers of SHUV which were below the detection limit of the endpoint dilution assay, indicating that the estimated number of ingested infectious viral particles was below 103 TCID50/ml.
Infection rates were also determined after 10 days of incubation at temperatures of 25°C (C. nubeculosus and Cx. p. pipiens) or 28°C (C. sonorensis and Ae. aegypti; Fig 2). Both biting midge species showed high infection rates of 44.4% for C. nubeculosus (N = 243), and 60.4% for C. sonorensis (N = 48; Fig 2A). SHUV replicated to mean titers of 9.2 x 103 TCID50/ml in body samples of C. nubeculosus and 3.3 x 104 TCID50/ml in body samples of C. sonorensis (Fig 2C). For one replicate experiment, heads were separated from the bodies and tested for presence of SHUV to assess whether the virus successfully passed from the midgut to the haemocoel, indicative of dissemination throughout the body. Dissemination rates were 18.2% (14/77) for C. nubeculosus and 10.0% (3/30) for C. sonorensis. Dissemination success, defined as the percentage of virus-positive heads out of the total number of virus-positive body samples, was 29.8% (14/47) for C. nubeculosus and 13.6% (3/22) for C. sonorensis. In all virus-positive heads that induced CPE, SHUV titers were all below the detection limit of 103 TCID50/ml. Because only very low amounts of SHUV were detected in biting midge heads, the actual percentage of disseminated infections might be higher. Considering the relatively high infection rates observed in this study and the absence of a salivary glands barrier in biting midges as shown in previous studies [17, 31], both C. nubeculosus and C. sonorensis can be considered highly competent vectors for SHUV.
SHUV was previously isolated from field-collected mosquitoes [8]. Therefore we determined vector competence for two mosquito species (Cx. p. pipiens and Ae. aegypti) which are important vectors for several arthropod-borne viruses [22, 27, 29]. Similar to the biting midges, SHUV titers of ingested blood were determined for a selection of 10 fully engorged female mosquitoes that were directly stored at -80°C after feeding on an infectious blood meal with a SHUV titer of 3.0 x 106 TCID50/ml. Similar to results obtained with the biting midges, both mosquito species ingested low amounts of SHUV that were below the detection limit of 103 TCID50/ml of the endpoint dilution assay.
No SHUV infection was observed in the Cx. p. pipiens mosquitoes (N = 211) following oral exposure, whereas infection rates of 2% were found for orally exposed Ae. aegypti mosquitoes (N = 149; Fig 2B). SHUV replicated to mean titers of 8.5 x 103 TCID50/ml in body samples of Ae. aegypti (Fig 2D), which was comparable to titers found in biting midges. No SHUV was detected in any of the saliva samples taken from either Cx. p. pipiens or Ae. aegypti. Thus, SHUV was able to successfully infect a small proportion of Ae. aegypti mosquitoes but not Cx. p. pipiens, and no evidence was found for transmission of SHUV by mosquitoes.
The very low infection rates of mosquitoes triggered further investigation into potential mosquito barriers against SHUV infection. To this end, Cx. p. pipiens and Ae. aegypti mosquitoes were intrathoracically injected with SHUV, to bypass the potential midgut barrier. Direct injection of SHUV into the thorax resulted in high infection rates of 68% for Cx. p. pipiens (N = 50), and 100% for Ae. aegypti (N = 50; Fig 3A). Transmission rates of 32% (16/50) were found for Cx. p. pipiens and 8% (4/50) for Ae.aegypti. This corresponds to transmission success of 47.1% (16/34) for Cx. p. pipiens and 8% (4/50) for Ae. aegypti. Interestingly, although infection rates of Cx. p. pipiens were below 100%, the transmission success was relatively high. This indicates a relatively weaker salivary gland barrier in Cx. p. pipiens compared to Ae. aegypti mosquitoes which had 100% infection rate, but relatively low transmission success.
To gain more insight in replication of SHUV, virus titers were determined for virus-infected mosquito body and saliva samples. Titers of virus-infected Cx. p. pipiens body samples were almost all below the detection limit of 103 TCID50/ml of the endpoint dilution assay (Fig 3C). This indicates that even when SHUV is injected into the thorax, there is no productive virus replication. In contrast, we found mean titers of 1.1 x 105 TCID50/ml for virus-infected Ae. aegypti body samples. This shows that SHUV is able to successfully replicate in Ae. aegypti when the midgut barrier is bypassed. In the majority of mosquito saliva samples, SHUV titers were below the detection limit of 103 TCID50/ml of the endpoint dilution assay (Fig 3D). Taken together, SHUV is able to disseminate in mosquitoes, but both the midgut and salivary glands form a barrier for SHUV.
SHUV was previously isolated from field-collected pools of Culicoides biting midges and from mosquitoes, but their relative importance in SHUV transmission remained to be confirmed. Here, we show for the first time that SHUV is able to infect and replicate in biting midges as well as in mosquitoes, but only the biting midge species evaluated in the present study can be considered competent vectors.
Both C. nubeculosus and C. sonorensis showed high infection rates of 44.4% and 60.4% when incubated for 10 days at 25°C and 28°C, respectively. The absence of a salivary gland barrier in biting midges [17, 31], and evidence of successful dissemination of SHUV to the heads indicates that the biting midge species evaluated in the present study are competent vectors of SHUV. Importantly, the finding that two different biting midge species from European and American origin are highly competent vectors suggests that various species of Culicoides may function as vectors of SHUV.
SHUV infection and replication in biting midges seems more efficient compared to other biting midge-borne viruses such as SBV and Bluetongue virus (BTV), which generally result in infection rates up to 30% [31–35]. Both SBV and BTV have caused sudden and large-scale epizootics in Europe, with devastating consequences for the livestock sector [36, 37]. The relatively high SHUV transmission potential by biting midges and ongoing emergence of SHUV to areas outside Sub-Saharan Africa [13], should therefore be interpreted as a warning for its epizootic potential.
In contrast to the high infection rates in biting midges, only few orally exposed Ae. aegypti mosquitoes became infected with SHUV during 10 days of incubation at 28°C. In addition, no evidence of successful dissemination to the salivary glands was found. SHUV replication and dissemination (8%) was observed when the virus was directly injected into the thorax of Ae. aegypti mosquitoes. This indicates that both the midgut infection barrier and the salivary gland barrier prevent infection and subsequent transmission of SHUV by Ae. aegypti mosquitoes. None of the Cx. p. pipiens mosquitoes that were orally exposed to SHUV became infected during 10 days of incubation at 25°C. Moreover, replication of SHUV was low in Cx. p. pipiens, because the virus was not able to replicate to high titers when it was directly injected into the thorax. However, a relatively high percentage of mosquito saliva samples contained SHUV. We therefore conclude that the midgut barrier is the main barrier that prevents infection of Cx. p. pipiens with SHUV. Considering our results obtained with both a tropical and temperate mosquito species, it seems unlikely that mosquito species play an important role in the SHUV transmission cycle. Our findings are in line with an earlier study on the closely-related SBV, which showed no evidence for involvement of mosquitoes in transmission, although SBV was able to infect Cx. pipiens mosquitoes [38].
Recent outbreaks of SBV and BTV showed the tremendous impact of midge-borne viruses on animal health [36, 37]. Our study demonstrates highly efficient infection and dissemination of SHUV in two biting midge species (C. nubeculosus and C. sonorensis), which illustrates its potential for emergence. SHUV should therefore be considered as an important arbovirus which may emerge further internationally in the near future. Future studies should test vector competence of field-collected Culicoides species for SHUV, to more accurately predict the efficiency of SHUV transmission following a first introduction into currently free areas. In addition, we recommend the development of diagnostic assays and a vaccine. These actions are essential to be prepared for newly emerging arboviruses with zoonotic potential such as SHUV.
|
10.1371/journal.pntd.0007574 | Estimated incidence and Prevalence of noma in north central Nigeria, 2010–2018: A retrospective study | Noma is a spreading and fulminant disease believed to be native to Sub-Saharan Africa over the last decade and associated with low socioeconomic status of citizens of the region. Within this noma belt, most epidemiological reports regarding the disease have emanated from the north western region of Nigeria. However, our indigenous surgical mission encountered a substantial number of cases of noma and post-noma defects noteworthy of epidemiological representation across north central Nigeria.
All noma cases encountered within the 8-year study period were included and divided based on clinical signs into acute and sequelae groups. Incidence estimation was based on acute/recently active cases and was calculated using the statistical method proposed by the WHO Oral Health Unit (1994). Period prevalence of noma was calculated considering the population at risk in the zone.
A total of 78 subjects were included in the study with age ranging from 2–75 years. Twelve subjects (15.4%) presented with acute disease while 66 (84.6%) had various forms of post-noma defects. The estimated incidence of noma in the north central zone was 8.3 per 100000 with a range of 4.1–17.9 per 100000 across various states. Period prevalence of noma which incorporated all cases seen within the study period was 1.6 per 100000 population at risk.
Although noma may be more prevalent in the north western region of Nigeria, substantial number of cases occurs within the north central zone which calls for deliberate public awareness campaign on disease risk factors and prevention, and education of primary health-care providers.
| Noma, a devouring disease of the orofacial complex, is commonly associated with poverty and impoverished regions of the world especially Sub-Saharan Africa termed the noma belt region of the world. With more reports advocating for full inclusion of noma in the WHO Neglected Tropical Diseases (NTDs) program, the apparent neglect of the disease in north central Nigeria compared to other sub-regions is worrisome as the disease burden in the sub-region has not been reported till date. In this light, a retrospective, cross-sectional survey was conducted to provide epidemiological representation to the cases encountered within an eight-year period by the Cleft and Facial Deformity Foundation (CFDF), an indigenous surgical mission. The incidence of noma was estimated using methods recommended by WHO while the period prevalence was calculated considering the population living below poverty line in the sub-region. This study extrapolates an incidence of 8.3 cases per 100000 and a period prevalence of 1.6 per 100000 persons at risk. Notable is the finding that most individuals encountered were above thirty years of age and suffered varying degree of facial disfigurement consequent to acute noma disease experienced in their childhood/early adolescence. Therefore, we advocate public awareness on the disease risk factors and prevention within the sub-region as well as training of primary health personnel on disease identification, primary care and nearest referral centres. We also identify the need to bolster the efforts of existing health facilities and indigenous surgical missions in the management and rehabilitation of noma cases and survivors.
| Noma is a disease of the orofacial region that has been unanimously described as devastating, mutilating, destructive and debilitating due to its appearance and the nature of spread of the acute necrotizing lesion which runs fulminating courses. Alternatively known as Stomatitis gangrenosa or Cancrum oris, the aetiology of noma is infectious, yet unclear as regards the exact causative microorganism(s) [1,2]. Initially, a fuso-spirochetal microbial complex was implicated due to the higher level of these organisms in individuals with noma; however, this notion has been dispelled due to nonreproducibility in animal models following inoculation with these microorganisms under similar noma predisposing conditions [3]. Nonetheless, more recent breakthrough into the inquiry of noma microbiology has revealed a polymicrobial interaction between intraoral commensal organisms and extraoral opportunistic microbes as the most likely cause of the disease [2]. Although noma is almost exclusive to young children within ages two to six years; it has been shown to affect individuals across all age groups, progressing through unique clinical stages beginning at the reversible and seemingly inconsequential necrotizing gingivitis/oedema stages, to the grotesque gangrenous stage associated with extensive soft and hard tissue necrosis and a high mortality rate of 90% in untreated individuals [4–5]. In the presence of appropriate medical intervention at the latter stage of acute disease, scarring occurs–leaving sufferers with various forms of socially incapacitating facial defects which defines the chronic phase of the disease [5]. Despite being a disease first described over four centuries ago as affecting several world regions, noma is currently regarded as being exclusive to the tropics (notably sub-Saharan Africa) [6–7], which is accredited to the preponderance of noma predisposing factors in the region. These factors include socioeconomic factors such as low standards of living, extreme poverty, poor sanitary conditions and close proximity of residence to livestock. Oral conditions such as poor oral hygiene and presence of simple gingivitis; systemic conditions like severe malnutrition, measles, malaria, tuberculosis, HIV infection, leukaemia, Non-Hodgkin’s lymphoma and cyclic neutropenia; and miscellaneous factors including low birth weight, improper weaning, birth position within the family and absence of mother as primary care giver [4,6,8–9].
Epidemiological research targeted at determining noma incidence and prevalence has been highlighted as a main feature of public health action programs against the disease. As determination of actual epidemiological parameters of noma is difficult due to high mortality associated with untreated disease, regional health data record inadequacies, remoteness of affected areas in addition to sufferers’ lack of access to primary health centres; current epidemiological data estimates a global incidence of 30,000–40,000 cases annually with seventy-five percent of these occurring in sub-Saharan Africa (the noma belt) [9–10]. Furthermore, in Nigeria (particularly the north west and south west sub-regions), incidence rates between 0.8–6.4 per 1000 children have been reported in the last two decades mostly according to data provided by foreign non-governmental organizations or surgical missions [10–11].
Since 2010, our indigenous surgical mission (Cleft and Facial Deformity Foundation [CFDF]) has embarked on organizing free intervention programs for individuals with orofacial conditions and deformities requiring urgent or elective surgical intervention in north central Nigeria–a region challenged with the dearth of craniofacial surgical expertise in secondary and tertiary health institutions. Although it is widely presupposed that the noma scourge is exclusive to northwest Nigeria as evidenced by the number of reports that have emanated from the region and the establishment of an health institution–Noma Children Hospital, solely concerned with treatment of acute stages of the disease and rehabilitation of survivors in the sub-region; cases of noma have also been encountered and successfully managed by our surgical mission across north central Nigeria within eight years. As epidemiological data is important for planning and prioritisation of service delivery as well as formulation of disease preventive strategies, we aim to provide an epidemiological report on noma disease in north central Nigeria by determining the incidence, prevalence, trend and risk factors for noma in the sub-region based on the health data records of noma cases encountered by our foundation over an eight years period spanning from 2010 to 2018.
This is a retrospective cross-sectional, epidemiologic study involving the health and treatment records of all noma cases encountered at our various surgical outreach locations in north central Nigeria between 2010 and 2018. Cleft and Facial Deformity Foundation (CFDF) is an indigenous surgical nongovernmental organization that has its focus on providing free surgical care for individuals with orofacial diseases. The organization is based in the north central geo-political zone of Nigeria, with its Head office in the more central Federal Capital Territory, from where surgical missions are being embarked upon to different locations within the zone.
Comprehensive information was obtained from stored records of patients encountered in all surgical outreach programmes organized in north central Nigeria from June 2010 till September 2018. All cases diagnosed as noma were included in this study, and this comprised both individuals with the acute disease or its sequelae. Since noma and orofacial cleft may share some similarities in clinical presentation, cases of the latter were excluded based on their congenital nature of occurrence and absence of significant morbidity associated with the deformity. Other head and neck or orofacial disease conditions were also excluded.
The information obtained from the records included participants’ bio-data, year of encounter and facial location of defects at presentation. Distance between patients’ location of residence and the health institutions where the surgical outreaches were conducted was estimated for each participant. Other information included the number of siblings in the family (<18 years), proximity of residence to livestock (cattle, pigs, horses etc), primary caregiver around the time of disease onset, and history of visits to referral centres. Cases were also categorized into one of the five stages of the disease proposed by World Health Organization (WHO)–necrotizing gingivitis/beginner, oedema, gangrene, scar and sequelae; with the latter two stages indicative of long-standing or resolving disease [5].
In the custom of the Cleft and Facial Deformity Foundation Data Management Team (CFDF-DMT), all health records are scrutinized at the end of every outreach program and variables perceived to be missing from a patient’s record are identified and eventually obtained from them at recall visits (usually organized about two months following the surgical outreach program). For the purpose of this study, participants whose missing information could not be updated at the follow-up visits were excluded. This was necessary to ensure validity and reduce uncertainty of the research outcome.
Data obtained from the study was analyzed using Statistical Package for Social Sciences (SPSS) version 23.0 (IBM Corp Armonk, NY, USA). Descriptive statistics such as frequencies, mean and standard deviation were explored for quantitative and categorical variables as appropriate. The normality of the distribution was ascertained using the Shapiro-Wilk’s test. Difference between quantitative variables was determined using the Mann-Whitney U test while relationships between categorical variables were determined using the Pearson’s Chi-square test. The significance value of all statistical tests used were set to 5% (p<0.05).
The prevalence of noma in the region was calculated by utilizing the total number of noma cases (both acute and chronic phase) seen within the study period as numerator and the population at risk as denominator. The population at risk considered only 45.7% of individuals residing below poverty line in the north central zone of Nigeria [12].
Incidence estimation analysis was done in line with the 1994 consultation report of the WHO Oral Health Unit using the Delphi method [13]. Only confirmed cases of acute noma (≤10years; beginner, oedema and gangrene) or older sequelae cases who marked their sixth birthday within the study period (giving due consideration to their year of encounter) were included in the analysis. According to the method, estimating the total incidence (I) involves a two-step process, beginning with the determination of the total surviving cases. The number of surviving cases (S) is expressed as a function of the number of referred cases reaching our outreach or resident centres (R) and an approximation of the percentage of the total surviving cases that were referred (χ) which was approximated as 15%, considering the notion that about ‘one out of every five’ noma cases presenting to referral centres [13] and the fact that our programs were carried out quarterly.
Thereafter, the total incidence (I) was extrapolated based on ‘S’ and the case survival rate of noma (y; 10%).
Ethical approval for the study was obtained from the Research Ethics review board of the International Craniofacial Academy. Prior notification of all outreach participants regarding the use of their health records and/or photographs for research purposes was done at each outreach event, with consents obtained. Records or images of participants/beneficiaries who refuse consent were never selected included or illustrated.
Within the eight-year study period, our indigenous surgical mission encountered 78 noma cases in twelve secondary health centres across Kogi, Nasarawa, Niger and the Federal Capital Territory of north central Nigeria; although, some centres in Kogi, Nasarawa and Niger states were visited at least twice between 2010 and 2018. A total of 12 subjects had acute noma while 66 participants presented with defects consistent with noma sequelae.
Age and sex variables of the noma cases seen were not normally distributed (Shapiro-Wilk’s test, p<0.05). Participants encountered were within ages 2–75 years with a majority of 43.6% (n = 34) being above 30 years (Table 1). The average age of participants in this study was 29.6±18.84 years. Most individuals presenting with acute noma were between ages 2–10 years (n = 10; 83.3%), with two subjects being adults aged 30 and 35 years; while approximately half of the sequelae cases which accounted for the most of the cases seen were above 30 years of age (n = 33, 50%) [p = 0.001]. Analysis of their sex distribution revealed that males (n = 42, 53.8%) were slightly more than females (n = 36, 46.2%), with a similar pattern of distribution obtained for both acute and sequelae cases [p = 0.735]. Most noma cases were observed in centres within Niger and Nasarawa states (n = 48; 61.1%) with only nine cases (11.5%) recorded in Kogi state (Table 1). Four acute noma cases were encountered in Niger and Nasarawa states respectively (denoting most of the cases), while subjects with noma sequelae defects were mostly encountered in Niger state (n = 21; 31.8%). Since acute noma participants were mostly pre-schoolers or middle-age children, records based on occupation was only made to reflect if they were properly enrolled in school or not, which all participants in this category had no formal education at the time of encounter (Table 1). Comparatively in terms of economic status, most participants with noma sequelae (n = 28, 42.4%) were not productively engaged, while 18.2% had careers centred on agriculture (farming, fishing or cattle rearing).
Of the twelve (12) acute noma cases recorded, 91.7% exhibited features of gangrenous stage of the disease (n = 11), while only one subject was noted to have presented with facial swelling and necrotic ulcerations of the mucosal lining of the upper lip and cheek which are consistent with the oedema stage of noma. Thirteen of the 66 participants with post-noma defects (19.7%) had nascent scarring indicative of recent active disease, with 53 (80.3%) showing clinical signs that are indicative of stage five (full blown sequelae) of the noma disease spectrum. Regarding the facial locations of the noma defect, 50% of all cases (n = 39) had deformities involving the nose while 28.2% (n = 22) and 32.1% (n = 25) had lesions that affected their right and left cheek respectively. Of both lips, the upper lip was mostly affected (n = 30, 38.5% > n = 13, 16.7%), and the occurrence of trismus among noma sequelae sufferers was 13.6% (n = 9).
Within the study period, acute noma cases were first encountered in 2012, with a steady rate of occurrence from 2013 to 2016, and increase in the number of cases encountered in 2017 (n = 5, 41.7%). However, no cases of acute noma were seen within the study period of 2018 (Fig 1). Participants with post-noma defects were seen annually throughout the study period, with cases increasing from two (3.0%) to fourteen (21.2%) from 2010 to 2011. Alternating decline and increase of noma sequelae cases were thereafter observed from 2011 to 2018 (Fig 1). The highest number of noma cases seen in a single year was 14 (17.9%), which occurred in 2011 and 2017. While all the participants encountered in 2011 had post-noma defects, five out of the fourteen cases in 2017 had acute noma lesions (35.7%). Table 2 shows the assortment of all noma participants (acute and chronic) presenting to the outreach referral centres within different parts of the zone.
Pertinent information collected regarding risk factors associated with noma included the number of siblings in the family, being raised by extended family members (especially grandparents), proximity of household residence to livestock and distance between residence and location of health outreach facility. The number of siblings of participants (< 18 years) ranged from 3 to 18 in total with an average of 8.6 ± 5.06. Furthermore, 85.9% (n = 67) answered positively that they lived in close proximity to livestock or even reared them while only 19.2% (n = 15) admitted to residing with extended relatives around the time of onset of noma disease.
The mean distance between patients’ residence and location of the health facility used for the surgical outreach was 124.8 ± 96.713km, with 41(52.6%) participants residing at locations 30 to 100 km from the referral centres. Twenty-eight (35.9%) individuals live between 101 to 300km away from the host secondary health facilities while 5 (6.4%) and 4(5.1%) participants dwell in areas <30km and >300 km from the centres respectively. Participants’ records further revealed that 60 (76.9%) subjects had never visited a health referral centre in the past and cited our indigenous surgical mission as the first centre of presentation since the disease onset.
The total estimated incidence of noma in the north central region of Nigeria between 2010 and 2018 is 8.3 per 100,000 population, with approximately 109 new cases diagnosed annually (17–42 cases across the various captured states, within the geopolitical zone). Noma incidence was highest in Nasarawa state with a rate of 17.9 cases per 100,000 population and lowest in Kogi state with an incidence of 4.1 cases per 100,000 population. The incidence extrapolated for Niger state and the Federal capital territory (FCT) was 5.1 and 14.2 per 100,000 respectively.
The period prevalence of noma in this study is 1.6 cases per 100,000 population at risk, with calculated sex occurrence rates being 1.7 per 100000 for males and 1.5 per 100000 for females. Noma prevalence was highest in the Federal capital territory (3.3 per 100,000 population at risk), with proportions ranging between 0.6–1.4 per 100,000 obtained in Kogi (0.6), Nasarawa (1.3) and Niger (1.4) states.
As a means of overcoming the challenges posed by the paucity of noma epidemiological data, nongovernmental organizations were one of relevant stakeholders saddled with the responsibility of reporting cases of noma sufferers and survivors encountered, in a bit to raise awareness on disease occurrence across their various stations or referral centres within the noma belt region [14]. Over the study period, our volunteer-based surgical mission discovered a substantial number of noma cases noteworthy of epidemiologic representation in north central Nigeria, which would allow for adequate characterization of the disease burden in this region–an area resident to the third-highest number of citizens living below poverty line in Nigeria [12]. In addition, our study shifts major attention from the recent norm of conducting noma epidemiological surveys in the north western region of Nigeria (Sokoto state in particular) over the last decade to the north central region comprising six member states (Benue, Kogi, Kwara, Nasarawa, Niger, Plateau) and the Federal Capital Territory.
Attempts at estimating noma incidence commenced towards the end of the 20th century. Barmes et al [15] first reported case incidence extrapolations of noma from Niger, Nigeria and Senegal, which followed in 1998 with the world health report by Bourgeois and Leclerq–initially estimating an incidence of 140,000 cases worldwide from interviews with health workers in noma prevalent areas [16]. Fieger et al [10] from the most recent report on noma incidence in north west Nigeria, estimated an incidence of 25,600 cases in developing countries bordering the Sahara Desert (the noma belt of the world), and a global incidence of 30,000–40,000 cases. In like manner, our study estimates a noma incidence of 8.3 per 100000 in north central Nigeria from 2010–2018, with a range of 4.1–17.9 per 100000 observed across different states within the geopolitical zone. This estimate is approximately eighty folds less than the calculated incidence of noma reported by Fieger et al [10] from 378 noma patients encountered between 1996–2001 in Sokoto, Nigeria. In the latter study (based on a multiple logistic regression model of deducing unknown noma incidence from available incidence data of orofacial cleft within the region), the estimated incidence was 6.4 per 1000 with values varying between 4.4 and 8.5 per 1000 in individuals aged 10–30 years. Our lower incidence estimate may be clearly attributed to the wide variation in poverty indices of the north western and north central region of the country over several years, with the north western region serially recording the highest number of individuals living below poverty line in the country (>80.0%) [12]. By inference therefore, the north central zone may have less residents with severe malnutrition, unsafe drinking water, poorer sanitation practices and limited access to proper healthcare when compared to the north western sub-region. A supporting reason for the wide variation observed as compared to the reports of Fieger et al [10] may be the inclusion of noma cases above 10 years of age and possible noma sequelae cases in the sample utilized for the incidence estimation in the latter which implies probable over-estimation in the incidence values extrapolated for the north west region of Nigeria. Our calculated incidence was also lower than the values obtained by Denloye et al [17] in Ibadan, south west Nigeria, where an incidence rate of 7.0 per 1000 cases was reported in individuals within ages 1 to 12 years from 1986–2000. This finding may be attributed to the disparate methodology of incidence calculation in both studies, as the forty-five noma cases reviewed by Denloye et al [17] were used against the total number of children that presented to the referral centre within the study period.
In further comparison of our findings with previous reports from other countries within the noma belt of the world, our estimated incidence was lower than the case incidence reported from Niger Republic (1.34 per 1000) and Senegal (0.7–1.2 per 1000) by Barmes et al [15] among children aged 0–6 years. However, this comparison may be flawed since at the time of data extrapolation, a seemly unrealistic mortality rate of 70% was utilized for the incidence estimation in their study; this observation was also highlighted by the reports of Fieger et al [10]. Subsequently, incidence estimates adjusted to a more accurate noma mortality rate of 90% resulted in incidence estimates between 1.2 to 4.2 cases per million in Dakar, Senegal among children aged 0–9 years of age [18], which ranks lower than the calculated incidence in our study.
The geographic distribution of noma is commonly represented figuratively on world maps by the WHO and its Regional offices, and titled “Noma in the world” [13]. These maps, which were first published in 1994, were made to depict reported cases of the disease across various parts of the world based on available data at the time of publication; hence, providing a diagrammatic panorama of the current noma situation in the world. The last update of these maps (published around 2009) [19] showed then recent noma case observations in 67.9% of African countries, no recent reported cases in two countries (DR Congo and Morocco) and sixteen “noma free” countries (Swaziland, Lesotho, Liberia, Sierra Leone, Guinea Bissau, Congo, Gabon, Libya, Tunisia, Equatorial Guinea, Burundi, Rwanda, Eritrea, Comoros, Mauritius and Western Sahara) within the African continent. Other affected continents included Asia (India, Pakistan, Myanmar), South America (Colombia, Guyana, Suriname, Argentina, Paraguay and Uruguay), with sporadic reports in Oceania, Europe and North America. With no recent map publication for over a decade, there is no current ‘snapshot’ or precise description of noma case observations worldwide, as well as no performance indicators of preventive strategies targeted at disease prevalent areas. Therefore, an update of these ‘noma maps’ by relevant monitoring stakeholders based on recent data/reports from experts within the last decade is urgently required in Africa and indeed globally to allow for knowledge acquisition on current ‘high risk’ areas and concomitant implementation of primary and secondary preventive approaches in these regions.
The pattern of noma/post-noma soft tissue defects in north central Nigeria includes varying degree of deformities involving the nose, upper lip, left cheek, right cheek and lower lip in decreasing order of occurrence. Our finding corroborates the pattern of noma presentation observed in Dakar Senegal in which the upper lip then cheek were quoted as the two most common sites affected by the noma defect; although, cases involving the nose were not cited in their report [18]. In contrast, the site distribution of defects in our study varies from the recent observations of Adeniyi and Awosan in Sokoto, north west Nigeria where lesions involving the cheeks were mostly seen, followed by upper and lower lips with the nose being the least affected site of soft tissue defects [20].
Farley et al [9] in a case-control study involving 74 cases and 222 controls in north west Nigeria, associated “caretaker” (i.e third party carer) as a factor that may influence the risk of developing noma. This was supported by Adeola et al [21] as reported in a case series involving five subjects managed for acute phase of noma in north west Nigeria. In that study, they asserted there was increasing number of noma cases in the region associated with lack of direct maternal care after children were weaned. This observation was not common in our study where only 19.2% of noma patients were being cared for by extended relatives around the time of disease onset. Although this proportion may be considerable, it was not significant enough to corroborate the observation of the earlier authors. Hence, it may require further studies to determine the plausibility of this assertion. Another important observation is the travel distance to access health care for patients in the north central sub-region. In this study, subjects had to travel about 125km on the average to access care from our treatment teams at the secondary health centres. In fact, 76.9% of them had not presented to any referral centre previously during the course of active disease or after its remission. Our experience in the region attributes this observation to factors such as subject’s preference for self-medication or traditional medical alternatives, lack of primary/secondary health centres within close proximity to residence, presence of primary centres but lack of required expertise and facilities for treatment, available facilities and expertise within state of residence but lack of finances to offset the cost of treatment and additional cost of transportation.
Since the extrapolations in this study were based on data records obtained from four states within the sub-region, the non-availability of records from Plateau, Benue and Kwara states represents a limitation to our incidence estimation; although, most areas are largely sub-urban/urban with tertiary health institutions in each of the states mentioned. Also, the possibility that some of the subjects may have developed the disease in these neighbouring north central states previously and relocated to the state of encounter prior to the outreach cannot be totally ruled out. Another limitation is the inability to explore in details, some major risk factors for noma due to the retrospective nature of the study design. Furthermore, the use of non-probability (convenience) sampling methods to arrive at the sample size employed for the analysis of associated risk factors for noma in this study may have introduced selection bias.
Considering that our study sought out to primarily determine the estimated incidence of noma in north central Nigeria, the results that have emanated are specific to characterizing the burden of noma in this zone. However, the statistical methods for the calculations done were adapted from the WHO Oral Health Unit and could as well be applied to other regions of the country or in other locations within the noma belt.
The estimated incidence of noma in north central Nigeria is 8.3 cases per 100000 population. Although the noma scourge is deemed prevalent in north west Nigeria and Sokoto state in particular, substantial number of cases is being encountered in the north central zone. Hence, efforts should be intensified in terms of public awareness, establishment of new primary health centres in deficient councils/wards, and education of community health workers in existing primary health care centres on disease identification (possibly primary care) in order to facilitate presentation of sufferers to appropriate referral centres within the north central zone. With a prevalence of 1.6 per 100000 population at risk and majority living with post-noma defects, it is clear that attention to surgical rehabilitation in the region is also suboptimal. It is therefore imperative in the absence of any health facility solely dedicated to the management and rehabilitation of noma patients in the region (unlike the northwest); that existing secondary health centres and nongovernmental organizations in the zone be better equipped to mitigate the disease burden and provide standard care for noma cases and survivors, especially as the poverty index of the zone and country is increasing. We further recommend that maps denoting noma occurrence in Africa and globally be updated according to recent available data so as to reflect current disease distribution and enable targeted preventive strategies in identified ‘high risk’ nations.
|
10.1371/journal.pntd.0001678 | Ecological Niche Modeling to Estimate the Distribution of Japanese Encephalitis Virus in Asia | Culex tritaeniorhynchus is the primary vector of Japanese encephalitis virus (JEV), a leading cause of encephalitis in Asia. JEV is transmitted in an enzootic cycle involving large wading birds as the reservoirs and swine as amplifying hosts. The development of a JEV vaccine reduced the number of JE cases in regions with comprehensive childhood vaccination programs, such as in Japan and the Republic of Korea. However, the lack of vaccine programs or insufficient coverage of populations in other endemic countries leaves many people susceptible to JEV. The aim of this study was to predict the distribution of Culex tritaeniorhynchus using ecological niche modeling.
An ecological niche model was constructed using the Maxent program to map the areas with suitable environmental conditions for the Cx. tritaeniorhynchus vector. Program input consisted of environmental data (temperature, elevation, rainfall) and known locations of vector presence resulting from an extensive literature search and records from MosquitoMap. The statistically significant Maxent model of the estimated probability of Cx. tritaeniorhynchus presence showed that the mean temperatures of the wettest quarter had the greatest impact on the model. Further, the majority of human Japanese encephalitis (JE) cases were located in regions with higher estimated probability of Cx. tritaeniorhynchus presence.
Our ecological niche model of the estimated probability of Cx. tritaeniorhynchus presence provides a framework for better allocation of vector control resources, particularly in locations where JEV vaccinations are unavailable. Furthermore, this model provides estimates of vector probability that could improve vector surveillance programs and JE control efforts.
| Japanese encephalitis virus (JEV) is transmitted predominately by the mosquito, Culex tritaeniorhynchus. The primary reservoirs of the virus are wading birds, with swine serving as amplifying hosts. Despite the development of a JEV vaccine, people remain unvaccinated in endemic countries and are susceptible to JEV infection. The distribution of the JEV vector(s) provides essential information for preventive measures. This study used an ecological niche modeling program to predict the distribution of Cx. tritaeniorhynchus based on collection records and environmental maps (climate, land cover, and elevation). The model showed that the mean temperatures of the wettest quarter had the greatest impact on the model. Of the 25 countries endemic for Japanese encephalitis (JE) endemic countries, seven possessed greater than 50% land area with an estimated high probability of Cx. tritaeniorhynchus presence. Our model provides a useful tool for JEV surveillance programs that focus on vector control strategies.
| Japanese encephalitis virus (JEV), the causative agent of Japanese encephalitis (JE), is an arbovirus that belongs to the family Flaviviridae and is endemic to Southeast and Northeast Asia, the Pacific Islands, and northern Australia (Figure 1) [1]. The primary vector of JEV is Culex tritaeniorhynchus Giles, but other Culex species (e.g., Culex annulirostris, Culex vishnui Theobald, Culex bitaeniorhynchus Giles, and Culex pipiens Linnaeus) have also been implicated as important viral transmitters [2], [3], [4], [5]. The larval habitat of Cx. tritaeniorhynchus is primarily low lying flooded areas containing grasses and flooded rice paddies, but this species can also be found in urban environments in close proximity to human populations [6]. Within the past 40 years, rice agriculture in JEV endemic countries has increased by 20%, thereby expanding Cx. tritaeniorhynchus habitat and increasing human risk of exposure to vector populations [7].
Swine, including domestic and feral pigs, serve as amplifying hosts of JEV in endemic areas. The proximity of human populations to pig farms, sties or feral pig populations increases the risks of JEV exposure [8], [9]. Ardeid birds (large wading birds) are an important JEV reservoir and can spread JEV to new regions through their northern migration to breeding and feeding grounds in the spring and southern return in the fall [3]. Additional animals have been identified as host species for JEV, including domesticated animals (chickens, goats, cows, and dogs), as well as bats, flying foxes, ducks, snakes and frogs. However, these are considered dead-end hosts as they infrequently develop sufficient viremias to infect mosquito vectors [10], [11], [12], [13].
Despite the introduction of an effective vaccine to the public in the mid-1900s, JEV remains the leading cause of viral encephalitis globally [14]. Comprehensive vaccination programs in Japan, Republic of Korea (ROK), Brunei, Australia, and Malaysia have significantly reduced the number of human cases [15]. Rare occurrences of neurological complications associated with the mouse-brain derived JEV vaccine interrupted vaccination programs in some regions, initiating concerns of the reemergence of JEV in an unvaccinated and non-immune population [16], [17]. The prevalence of JE is higher in countries with lower socioeconomic status, when compared to more affluent neighboring countries, indicating the importance of economic and social stability as additional risk factors that impact the transmission and prevalence of JE in non-immune populations [15].
Recent developments in the field of ecological niche modeling and the development of global environmental data sets have resulted in the ability to predict the distribution of vector populations that directly relate to transmission of viruses, parasites, and fungal pathogens and impact on animal and human health. Modeling to estimate the distribution of disease vectors provides useful information in disease-endemic areas, in addition to predicting how anthropogenic changes to the environment will affect disease presence [18], [19], [20], [21].
In the current study, the Maxent ecological niche modeling program was utilized to model the distribution of the primary vector of JEV, Cx. tritaeniorhynchus [22]. The resulting vector habitat suitability map was compared to the reported locations of JE human cases and the current status of established JE vaccination programs by country. Our ecological niche model can be used by public health officials and government agencies in endemic regions to guide implementation of comprehensive vaccination programs, vector control strategies, and public health awareness campaigns.
Geographical coordinates of known Cx. tritaeniorhynchus records were identified by performing a literature search in PubMed for all previous field collection studies. When exact geographical coordinates were not provided, locations were approximated by searching for the given city, town, or village using Google Earth software. Further geographical data points for the distribution of Cx. tritaeniorhynchus were obtained through MosquitoMap (http://www.mosquitomap.org/), a database of spatial data points of mosquitoes that is maintained by the Walter Reed Biosystematics Unit, Smithsonian Support Center, Silver Hill, MD. Additional Cx. tritaeniorhynchus collection data were obtained from Force Health Protection and Preventive Medicine, 65th Medical Brigade, Yongsan Army Garrison, ROK.
In previous modeling work in the ROK [20], we found that a large number of collection records in a limited geographical area biased the model. As a result of the large number of collection sites for the ROK (96 unique locations), we reduced the number of records to 23 by deleting all but one randomly selected record per administrative district.
Approximate locations of known human JE cases were determined using locations provided in ProMED mail reports (www.promedmail.org) from 1994 through 2010 (Figure 2). Additional locations of confirmed JE cases were also determined through a PubMed literature search. Exact geographical coordinates were not reported for most documented human cases and were therefore extrapolated using the Google Earth software to obtain the latitude/longitude coordinates of the reported city, town, or village in which JE was documented.
JEV vaccination program information was obtained from “Japanese Encephalitis Morbidity, Mortality and Disability: Reduction and Control by 2015” published in 2009 by the Program for Appropriate Technology in Health (PATH), Armed Forces Research Institute of Medical Sciences, and BIKEN [15]. Additionally, JEV vaccination programs information was also obtained from the WHO/IVB database [23]. Countries lacking a JEV vaccination program were identified using information in the above mentioned publications and confirmed with additional literature searches. A summary of these data are listed in Table 1.
One kilometer resolution climate and elevation data were obtained from WorldClim (http://www.worldclim.org/bioclim). The WorldClim organization has processed 50 years of ground-based weather measurements to produce mean monthly minimum and maximum temperatures and precipitation in a grid format at several different resolutions. The data were further processed to produce bioclimatic variables (e.g., mean temperatures of the wettest quarters). For this project, the highest resolution data available from WorldClim (approximately 1 km) were downloaded. In addition to bioclimatic variables, global elevation data obtained from WorldClim was re-sampled to 1-km resolution from NASA's Shuttle Radar Topography Mission (SRTM). Descriptions of the bioclimatic and elevation variables used for this study are listed in Table 2. To better understand the effect of each environmental variable on Cx. tritaeniorhynchus distribution, the values of each environmental layer at each site were extracted using ArcGIS (ESRI, Redlands, California, www.esri.com). This allowed for a comparison to the known environmental and distribution limitations for Cx. tritaeniorhynchus in the literature.
A map of rice growing areas was created by processing GeoCover-LC (Land Cover) data from MDA Information Systems, Inc. (http://www.mdafederal.com/geocover/geocoverlc). GeoCover was created by processing Landsat Thematic Mapper images to create land cover maps for most areas of the world. Each pixel within the GeoCover-LC represents 30 by 30 meters. To convert the image to a resolution that could be used in the Maxent model, ArcGIS was used to count the number of rice pixels within each square kilometer (33 by 33 pixels). Then, a rice percentage was calculated for each square kilometer (number of rice pixels divided by total number of pixels in 1 km) and stored in a final output image.
The Maxent 3.2.1 modeling program (http://www.cs.princeton.edu/~schapire/maxent/) was utilized to model the distribution of Cx. tritaeniorhynchus based on previously obtained geographical locations. Maxent utilizes a maximum entropy algorithm to analyze values of environmental layers, such as temperature, precipitation, and elevation, at known locations of species occurrence (collection records) to estimate the probable range of the species over a geographic region [22], [24]. This model is based on presence-only data instead of presence/absence data due to the lack of available absence data. Although absence data can be informative for modeling, ecological niche models based on presence-only data are useful in regions with limited collection data [22]. Without absence data, the true probability of presence cannot be modeled. In Maxent, which uses presence only data, the species distribution is output as an estimated probability map [25].
The Maxent program calculates the importance of environmental variables in developing predictive species distribution models by using the jackknife test of variable importance. The jackknife test runs the model 1) once with all variables, 2) dropping out each variable in turn, and 3) with a single variable at a time. Variables are considered import if they produce high training gains when used alone in a model. A variable is also important if the training gain is low when the variable is removed from the model [22].
Maxent utilizes two approaches to validate the accuracy of the model. The first method randomly selects occurrence points to be withheld from the model building to use as testing points. Using multiple definitions, a set of thresholds split the continuous probability values of the model into ‘predicted presence or absence’ categories. Maxent then calculates the p-value based on the null hypothesis that testing points will be predicted as “present” no better than by a random model. The second method calculates the Area Under the Curve (AUC) of the receiver operator characteristic (ROC), a graphical depiction of the sensitivity versus one (1) minus the specificity of the model often used to validate ecological niche models [22], [26]. The AUC indicates whether the model predicts species location better than a random distribution. AUC values of ≤0.5 indicates a random distribution and AUC values >0.9 indicates high reliability of the model [22]. To determine the best combination of environmental data for modeling, the model was run four times using different sets of input layers each time: 1) bioclimatic layers, elevation and rice crop data, 2) bioclimatic layers and elevation data, 3) bioclimatic layers only, and 4) elevation data only.
A total of 139 unique sites of documented Cx. tritaeniorhynchus geographical locations were utilized to construct the ecological niche model (Figure 3). Of the 139 total points, 105 (76%) were randomly designated as training points in order to build the model and 34 (24%) points were used to test the model. The model was run four times using different combinations of environmental layers (Table 3). Statistical results indicate that the most accurate model included bioclimatic layers and elevation (Table 3), and therefore this model was used in all subsequent analyses. Statistical evaluation showed the model to have a high accuracy, with the AUC>0.9 and low p-values. The model is available to view or download from www.vectormap.org.
In order to evaluate the contribution of each environmental variable to the model, Maxent utilizes a jackknife test, which indicated that the annual precipitation (bio12) environmental layer is the environmental variable with the highest gain when used in the model by itself. The Maxent program also calculates a percent contribution for each variable in the model. The annual precipitation variable contributed 16.2% of the information used by the model, another indication that it is an important environmental factor for estimating the distribution of Cx. tritaeniorhynchus (Table 2). The mean temperature of the wettest quarter variable (bio08) contributed the highest percentage (21.7%) of the information to the model. Elevation was also an important variable, contributing 9.6% to the model. From the jackknife test, if elevation data were removed from the model, the overall training gain would decrease the most, indicating the elevation variable contained the most unique information of the variables in the Cx. tritaeniorhynchus distribution model.
The values of each environmental variable at each recorded location of occurrence were extracted using ArcGIS (Table 1). For example, the known locations for Cx. tritaeniorhynchus used in the model fell within 0 and 838 meters of elevation. This is consistent with the published reports that Cx. tritaeniorhynchus is rarely collected above 1,000 meters [27], [28].
Ninety-six reported JE case locations were identified in endemic regions (Figure 2). ArcGIS analysis categorized human JE cases based on the estimated probability of Cx. tritaeniorhynchus presence (Figure 4). Human JE cases were identified at locations with a range of estimated probability of vector presence, including regions with 25% or less estimated probability. However, the majority (>75%) of human JE cases were reported from regions with greater than 25% estimated probability of Cx. tritaeniorhynchus presence. Limited availability of location data of human JE cases greatly impacts any associations between areas of high estimated vector probability and disease. For instance, the lack of human JE cases in other regions of estimated high probability of vector presence could be due to lack of reporting, improper diagnosis, or due to successful prevention strategies.
ArcGIS analysis determined the approximate percentage of each country with >25% probability of Cx. tritaeniorhynchus presence based on the Maxent model (Table 1). Of the 25 endemic countries, seven possessed >50% of their land area with a higher probability of Cx. tritaeniorhynchus presence. Three countries (Bhutan, Pakistan, and Russia) possessed <1% of their total country area with a 25% probability of Cx. tritaeniorhynchus presence.
In this study, a statistically significant ecological niche model for Cx. tritaeniorhynchus was developed using mosquito presence records, climate, and elevation variables. Locations of human cases of JE generally fell within the higher probability areas of Cx. tritaeniorhynchus (Figure 4). Regions of estimated high probability of Cx. tritaeniorhynchus presence (Figure 3) are representative of preferred environments, based on temperature, precipitation and elevation where Cx. tritaeniorhynchus habitats occur. This model serves as a tool to fill in knowledge gaps regarding Cx. tritaeniorhynchus and can be utilized by health care professionals and policy officials in endemic regions to help guide the development and implementation of disease mitigating strategies in endemic regions.
The Maxent program identifies important environmental variables that are major contributors to the vector distribution model. Based on the jackknife test of variable importance, the annual precipitation (bio12) is an important contributor to the model. Additionally, the mean precipitation of the wettest quarters (bio08) and elevation also contributed greatly to the model for distribution of Cx. tritaeniorhynchus (Table 2). Previous studies that aimed to identify favorable ecological conditions of mosquitoes found that the optimal temperature of JEV vectors is between 22.8 and 34.5°C [29]. The importance of temperature during the wet season in the model is attributed to temperatures and flooded habitats that are optimal for larval development and adult survival. Locations in which the temperatures do not fall into the optimal range during the rainy season may therefore experience fewer mosquitoes, despite harboring the appropriate habitat. Temperature also plays a role in disease transmission rates, as higher temperatures increase the rates of virus replication and dissemination, while decreasing the time from mosquito infection to transmission of the virus to animal and human hosts [30].
Sampling bias is an issue that affects the accuracy of the model as the model was developed using existing data from the literature and VectorMap. Therefore, some regions have not been sampled in the study area and some have been oversampled. Cx. tritaeniorhynchus data for China were very limited (Figure 2), which may mean that some potential environmental conditions of Cx. tritaeniorhynchus were not represented in the model, in particular, the cooler Northeast region of China. Because modeling was limited to Cx. tritaeniorhynchus, there is a potential that for some regions, other primary or secondary vectors, i.e., Cx. annulirostris, Cx. bitaeniorhynchus, and Cx. vishnui, may predominate and maintain transmission of JEV in these areas. Collection records of Cx. tritaeniorhynchus were obtained spanning many decades and at different times during the year, furthering the impact of sampling bias on our model. In addition, the density of Cx. tritaeniorhynchus was not collected in this study and is an important limitation as vector abundance plays a crucial role in disease transmission. Further collection studies are therefore needed to determine the abundance of vector species in addition to presence in endemic regions.
Low-lying flooded areas containing grasses, including rice paddies, are the primary larval habitats for Cx. tritaeniorhynchus. An increase in the amount of flooded rice field habitat has shown to be positively correlated with increases in adult populations of Cx. tritaeniorhynchus in the ROK [31]. Although the rice map derived from the GeoCover Land Cover map (Figure 5) does generally match the predicted occurrence of Cx. tritaeniorhynchus, there are some areas where the model predicts the presence of the mosquito, yet no rice crops were mapped. For some areas, rice may not have been identified correctly on the satellite images, since agricultural areas were limited or were adjacent to other predominant habitats. For example, rice is produced in Nepal [32], but no rice fields were identified by GeoCover in Nepal, since the identification of small rice fields in mountainous areas on satellite images can be difficult. Alternatively, this shows that environments other than rice fields are suitable habitat for Cx. tritaeniorhynchus.
The predicted probability of Cx. tritaeniorhynchus presence values were used to determine the percentage of a country at high risk (greater than 25% probability) for vector presence (Table 1). Many Asian countries have high percentages of their total land area with a >25% probability for the presence of Cx. tritaeniorhynchus. Cambodia, the ROK, Sri Lanka, and Thailand have over 75% of their land area with a >25% probability of Cx. tritaeniorhynchus presence. Countries demonstrating >50% of their total land area and with Cx. tritaeniorhynchus occurrence >25% probability includes: Bangladesh, East Timor, and Vietnam. However, some countries may have small areas of vector habitat close to large populations that can result in outbreaks despite low percentage estimated probability in the region overall. Further, this analysis does not take into account country size or vector abundance, which would also impact disease transmission. Although additional factors contribute to JE disease risks, the distribution of the vector populations within a country is a valuable data set when considering the necessity of vaccination and other health risk reduction programs.
Human JE cases were categorized based on the estimated probability of vector presence at the reported location (Figure 4). Interestingly, a portion of human cases were reported from regions with 25% or less estimated probability of Cx. tritaeniorhynchus presence. JE cases not falling within high probability pixels could have been acquired in nearby locations. Even for precisely located case data, the high resolution of the model (one kilometer pixels) increases the likelihood that the predicted Cx. tritaeniorhynchus location does not match disease acquisition location as many people travel more than a kilometer in the course of a typical day. Alternatively, additional factors other than the presence of Cx. tritaeniorhynchus may be important when determining the risk of disease. For instance, other JEV vectors may dominant in these regions estimated with low Cx. tritaeniorhynchus presence. JE is one of many febrile illnesses that affect human populations in Asia. Difficulties arise in diagnosis of JE in patients based on symptoms alone that range from mild to very severe, with laboratory tests required for confirmation. Obtaining geographical data of where human cases were acquired is made difficult due to lack of confident diagnoses, patient travel history, and spatial data. The lack of precision of the reported case locations may also contribute to lower numbers of JE cases falling within high probability Cx. tritaeniorhynchus pixels. Identification of human JE cases in this study is extremely limited and does not represent all human cases in JE endemic regions. In many cases, only a village or city name was given for reported cases. A previous study to model the distribution of Cx. tritaeniorhynchus to predict JE in the Republic of Korea found human cases to occur in areas of high estimated probability of vector presence [20]. This study, however, utilized intensive vector collection methods and JE case data were obtained from the Korea Centers for Disease Control resulting in an overall more extensive and accurate model. This illustrates the need for increased surveillance of vector and human JE cases in order to generate more accurate risk models for JE. In order to evaluate the impact of vector presence on the risk of JE in humans, comprehensive efforts to identify specific locations of both symptomatic and asymptomatic JE cases across endemic regions are needed.
Ecological niche modeling inherently possesses limitations in that it makes predictions based solely on environmental variables that impact larval development and adult survival. Other important factors that influence vector distributions include: vector control strategies, public health campaigns, socioeconomic status, human population densities, anthropogenic changes to land (creation of vector habitat), vector species competition, and predator influences on their potential distribution and population densities. Further, the use of WorldClim data may underestimate or ignore environmental variables that occur during a short time period or transient habitat suitable for the vector to survive. Incorporation of these variables will undoubtedly increase the validity of the model. These factors are also important to take into consideration when implementing mosquito control initiatives and vaccination campaigns.
The reemergence of JEV remains possible due to multiple factors. Increases in the pig farming industry, modification and expansion of arable lands for wetland rice farming, and a fraction of the population unvaccinated/non-immune, in combination with optimal climatic conditions, contribute to the potential for periodic outbreaks of JE as the one observed in the ROK [9]. Genotype analyses of circulating JEV strains identified the reemergence of genotype V, which was unseen in Asia for over 50 years [33], [34]. The identification of emerging/reemerging JEV strains is important for vaccine development and the implementation of effective vaccination programs. Increased surveillance in areas with known vector populations and additional risk factors, such as reservoir and amplifying hosts, will aid in the identification of circulating JEV strains as well as strains that are emerging in novel human and vector populations. Understanding the vector distribution is a key step to effectively understanding JEV risks and also to preventing additional outbreaks of JE in endemic countries.
|
10.1371/journal.pntd.0001987 | Imipramine Is an Orally Active Drug against Both Antimony Sensitive and Resistant Leishmania donovani Clinical Isolates in Experimental Infection | In an endeavor to find an orally active and affordable antileishmanial drug, we tested the efficacy of a cationic amphiphilic drug, imipramine, commonly used for the treatment of depression in humans. The only available orally active antileishmanial drug is miltefosine with long half life and teratogenic potential limits patient compliance. Thus there is a genuine need for an orally active antileishmanial drug. Previously it was shown that imipramine, a tricyclic antidepressant alters the protonmotive force in promastigotes, but its in vivo efficacy was not reported.
Here we show that the drug is highly active against antimony sensitive and resistant Leishmania donovani in both promastigotes and intracellular amastigotes and in LD infected hamster model. The drug was found to decrease the mitochondrial transmembrane potential of Leishmania donovani (LD) promastigotes and purified amastigotes after 8 h of treatment, whereas miltefosine effected only a marginal change even after 24 h. The drug restores defective antigen presenting ability of the parasitized macrophages. The status of the host protective factors TNF α, IFN γ and iNOS activity increased with the concomitant decrease in IL 10 and TGF β level in imipramine treated infected hamsters and evolution of matured sterile hepatic granuloma. The 10-day therapeutic window as a monotherapy, showing about 90% clearance of organ parasites in infected hamsters regardless of their SSG sensitivity.
This study showed that imipramine possibly qualifies for a new use of an old drug and can be used as an effective orally active drug for the treatment of Kala-azar.
| The disease Kala-azar or visceral leishmaniasis is still a big problem in the Indian subcontinent. The antimonials were used for the chemotherapy of Kala-azar but with time its efficacy has reduced dramatically. The newer version of orally active drug miltefosine has been introduced, but its efficacy has decreased considerably as relapse cases are on the rise. Other drugs like liposomal form of amphotericin B is expensive and the patients require hospitalization. Thus there is a genuine need for an orally active antileishmanial drug. There are reports that the cationic amphiphilic molecule, imipramine, a drug used for the treatment of depression in humans, kills the promastigotes of Leishmania donovani. We tested the efficacy of imipramine in experimental infection in hamster and mouse model. Our study showed that the drug is highly effective against antimony sensitive and antimony resistant Leishmania donovani infected hamsters as well as mouse and offered almost sterile cure.
| The disease visceral leishmaniasis or Kala-azar is caused by the protozoan parasite Leishmania donovani (LD) and is widening its base in different parts of the world [1], [2]. Pentavalent antimonial or SSG, which has long been the first line drug, is no longer recommended for use as high levels of resistance in the Indian subcontinent have been reported [3]. Other drugs like miltefosine (hexadecylphosphocholine, a polyene antibiotic) and amphotericin B (an anti-fungal agent) are in current clinical use. As miltefosine is orally active, it offers advantages in terms of reduced hospitalization but cannot be used during pregnancy and lactation [2]. Amphotericin B and its liposomal form are to be administered as an infusion and therefore the patients require hospitalization [4]. Unfortunately, treatment failure cases to miltefosine [5] and amphotericin B [6] are emerging, which raises serious concerns for their future use. There is a genuine need for an orally active and affordable drug for the treatment of relapsed Kala-azar cases.
Imipramine, N-(γ-dimethylaminopropyl)-iminodibenzyl HCl, is a tricyclic antidepressant and belongs to the broad class of cationic amphiphilic drugs. The tricycle consists of two benzene rings fused with a seven member heterocycle. Imipramine is FDA (Food and Drug Administration) approved drug for treating depression and paediatric nocturnal enuresis [7], and is sometimes used off-label to treat chronic pain in combination with other pain medications [8]. The dose range for treating depression is 100–200 mg daily and the recommended use for enuresis is 10–75 mg daily [9]. The selection of imipramine for therapy of experimental visceral leishmaniasis is based on the following past observations by others: (i) the drug alters the proton motive force of LD's membrane [10], (ii) inhibits trypanothione reductase, an enzyme upregulated in SSG resistant LD parasites [11], (iii) an effective immunomodulator as it induces the production of TNF-α, an important cytokine for antileishmanial defense [12], (iv) cationic properties favor its absorption by phagocytic cells and accumulation in phagolysosomal bodies [13], and (v) its metabolite desipramine is as effective as the parent drug against LD promastigotes [14]. These compelling attributes of imipramine towards Leishmania parasites led us to test its efficacy directly on LD and also in experimental infection induced by recent clinical isolates of SSG-S and SSG-R LD parasites with miltefosine as a reference oral drug.
In this investigation, we endeavored to study the effect of oral administration of imipramine in LD infected hamster model. Our study done in hamster model very clearly showed that this drug is highly active in vitro as well as in vivo. Furthermore it plays a strong immunomodulatory role which also favored parasite clearance. Thus imipramine may be used orally in the treatment of visceral leishmaniasis. To our knowledge this is the first report on the therapeutic efficacy of imipramine in experimental visceral leishmaniasis.
BALB/c mice (Mus musculus) and hamsters (Mesocricetus auratus) were maintained and bred under pathogen free conditions. Use of both mouse and hamster was approved by the Institutional Animal Ethics Committees of Indian Institute of Chemical Biology, Kolkata, India. All experiments were performed according to the National Regulatory Guidelines issued by CPSEA (Committee for the Purpose of Supervision of Experiments on Animals), Ministry of Environment and Forest, Government of India.
All parasites for this study were received from European Union KaladrugR project consortium. These parasite samples are fully anonymized and study with these parasites is approved by Institutional Review Board of Institute of Medical Sciences, Benaras Hindu University, Varanasi, India. The details of the patients and the treatment profile of the patients from whom Leishmania donovani (LD) parasites were derived have been published previously [15]. Clonal population of LD parasites MHOM/IN/10/BHU816/1 (BHU 816) and MHOM/IN/09/BHU777/0 (BHU 777) are SSG sensitive (SSG-S) and strains MHOM/IN/09/BHU575/0 (BHU 575), MHOM/IN/10/BHU782/0 (BHU 782), MHOM/IN/10/BHU814/1 (BHU 814) and MHOM/IN/10/BHU872/6 (BHU 872) are SSG resistant (SSG-R). LD promastigotes were maintained in M199 medium (Sigma Aldrich, St. Louis, MO) supplemented with 10% heat inactivated FBS (Gibco), 100 IU/mL of penicillin and 100 µg/mL of streptomycin (Gibco) in a 22°C room as described elsewhere [15].
Imipramine hydrochloride (Sigma Aldrich, St. Louis, MO) and miltefosine (Kindly provided by Aeterna Zentaris GmbH (Germany) to the KaladrugR project consortium, batch#1149149) solutions were prepared at 1 mg/ml in PBS (Sigma Aldrich, St. Louis, MO), followed by sterile filtration using 0.22 µM filters (Milipore) as and when required.
PECs were harvested from BALB/c mice by lavage, 48 h after i.p. injection of 2% (w/v) soluble starch (Sigma Aldrich, St. Louis, MO). For convenience, PECs of BALB/c mice were defined as MΦ. MΦ were harvested on sterile 22 mm square coverslips (Bluestar, India) in 35 mm disposable petriplates (Tarsons, India) at a density of 105/cover slip in RPMI 1640 medium (Sigma Aldrich, St. Louis, MO) supplemented with 10% heat inactivated FBS, 100 IU/mL of penicillin, and 100 µg/mL of streptomycin, i.e. RPMI complete medium. The cells were left to adhere for 48 h at 37°C under 5% CO2 before infection. The MΦs were infected with stationary phase promastigotes at a ratio of 1∶10 [14], [15]. After incubating the cultures at 37°C and 5% CO2 overnight or for 4 h, non-phagocytosed promastigotes were washed off with serum free medium RPMI 1640 and treatment provided as described [16].
MΦs were harvested on a 96-well tissue culture plate (BD Biosciences) in RPMI complete media and left to adhere for 48 h at 37°C under 5% CO2. Successive increasing concentrations of imipramine were added in triplicate and incubated for 24 h. After completion of incubation, MTT (Sigma Aldrich, St. Louis, MO) was added and incubated for 4 h at room temperature. Solublizing agents [0.04 N HCl (Merck) in isopropanol (Merck)] were added after incubation and the optical density (OD) was measured after 30 min in a plate reader at 570 nm. The relative number of live cells was determined based on the optical absorbance of the treated and untreated samples and of blank wells, as described previously [17].
Day 5 culture of parasites was used to determine the drug efficacy (IC50) to kill promastigotes using MTT [18]. The LD parasites were plated on the 96-well cell culture plates at a density of 105 cells/well and kept in presence of imipramine for 48 h. Results were expressed as the concentration that inhibited parasite growth by 50% (IC50). Analysis was carried out using Graphpad Prism5 software (version 5.03).
In order to determine EC50 (Efficacy against intracellular amastigote), the drug was serially diluted in RPMI complete medium over six concentrations in triplicate at each concentration. Stock solutions and dilutions were freshly prepared for each use. Infected MΦs were incubated with drug dilutions for another 24 h at 37°C and under 5% CO2. Untreated MΦs received medium alone and intracellular parasites were enumerated.
At the endpoints, the coverslips were washed with PBS, dried, fixed with 100% methanol (Merck), stained with 10% Giemsa (Sigma Aldrich, St. Louis, MO) and examined microscopically. One hundred MΦs/coverslip were scored and the amastigotes were enumerated [19]. The average of three untreated cultures was taken as 100% control against which the percentage inhibition of infected MΦs in treated cultures was calculated. The 50% effective concentration (EC50) of imipramine for each of the isolates was estimated as described elsewhere [19], [20].
The JC-1 dye (Molecular Probes, Eugene, OR) has been used routinely to monitor the mitochondrial potential [21]. The monomeric form has an emission maximum at 527 nm. The dye at higher concentrations or potentials forms red fluorescent J-aggregates with an emission maximum at 590 nm. The ratio of this red/green (λ590/λ527) fluorescence is known to depend only on the membrane potential. A working solution of JC-1 was therefore prepared as per manufacturer's instruction. Imipramine and miltefosine treated and untreated LD promastigotes were incubated with the JC-1 working solution for 25 min in a 96-well plate and washed. Cell pellets were resuspended in assay buffer and analyzed under a fluorescent plate reader (Fluorescence plate reader LS 55, Perkin Elmer).
Double staining for annexin V fluorescein isothiocyanate (FITC)-PI was performed with the Annexin-V apoptosis detection kit (Molecular Probes, Eugene, OR) [22]. In brief, untreated, imipramine-treated, or miltefosine-treated promastigotes were washed twice in cold PBS and centrifuged at 3000 rpm for 10 min. The pellets were resuspended in 100 µL of annexin V-FITC in the presence of PI according to the instructions of the manufacturer. After 15 min of incubation in the dark, the intensity of annexin V-FITC labeling was recorded on a flow cytometer (FACSARIA II, Becton Dickinson, San Diego, CA) and analyzed with FACSDIVA software, version 6.1.1; the percentage of positive cells was then assessed.
The membrane fluorescence and lipid fluidity of MΦ under parasitized condition as well as after imipramine treatment were measured following the method described by Shinitzky and Inbar [23]. Briefly, the fluorescent probe DPH (Molecular Probes, Eugene, OR) was dissolved in tetrahydrofuran (Merck) at 2 mM concentration. To 10 ml of the rapidly stirring PBS solution (pH 7.2), 10 µL of 2 mM DPH solution was added. For labeling, 106cells were mixed with an equal volume of DPH in PBS (cf 1 µM) and incubated for 2 h at 37°C. Thereafter the cells were washed thrice and resuspended in PBS. The DPH probe bound to the membrane of the cell was excited at 365 nm and the intensity of emission was recorded at 430 nm in a spectrofluorometer (LS 55, Perkin Elmer). The FA value was calculated using the equation: FA = [(III−I⊥)/(III+2I⊥)], where III and I⊥ are the fluorescent intensities oriented, respectively, parallel and perpendicular to the direction of polarization of the excited light [23].
MΦs, also defined as antigen presenting cells (APCs), were harvested from peritoneal cavity of mice at 106 cells/well in a 48 well tissue culture plate, then incubated for 24 h with specific peptide Lambda repressor λR12–26 (GenScript, USA) and T cell hybridoma 9H3.5 (kind gift of Professor Malcolm Gefter, Massachusetts Institute of Technology, Cambridge, Massachusetts) in complete RPMI medium in a 37°C incubator. The culture supernatants were analyzed for the presence of IL 2 using mouse IL 2 ELISA kit (BD Biosciences, San Diego, CA) as per manufacturer's instruction.
To monitor the level of reactive oxygen species (ROS, including superoxide, hydrogen peroxide, and other reactive oxygen intermediates), the cell-permeable, non polar, H2O2-sensitive probe H2DCFDA (Molecular Probes, Eugene, OR) was used [24]. The extent of H2O2 generation was defined as the extent of ROS generation for convenience. For each experimental sample, fluorometric measurements were performed in triplicate and the results were expressed as the mean fluorescence intensity per 106 cells. Nitric oxide (NO) generation was monitored by using the Griess reagent (Molecular Probes, Eugene, OR) as described previously [25], and the results are expressed in µM nitrite.
To infect hamsters (6 weeks old), two SSG-S (BHU 777 and BHU 816) and two SSG-R (BHU 575 and BHU 814) LD amastigotes were purified as described [26] and inoculated (107 parasites in 200 µL) via intracardiac routes as described previously [27].
Imipramine is usually used in human at a dose of 100–200 mg/day (average of 150 mg/day) for the treatment of depression [9], [28]. Considering the average human body weight of 60 kg, the effective dose is 2.5 mg/kg/day. Based on the dose equivalence between human and rodents [29], the dose of imipramine in mouse and hamsters would be 41 and 25 mg/kg/day respectively. In our investigation, the highest dose used was 5 mg/kg/day both in mouse and hamsters which is effectively 12.3 and 7.4 times lower than the equivalent human dose. Miltefosine is used in human at the dose 2.5 mg/kg [5]. Based on dose equivalent formula, we converted the normal human dose to hamster equivalent dose. Thus we treated hamsters with the maximum dose 17.5 mg/kg which is ∼7 times high then the normal human dose.
The 8-week infected hamsters (i.e. 14-week old hamsters) were randomly divided into four groups (groups I to IV). Group I received only saline, groups II to IV received imipramine at the dose levels of 0.05, 0.5 and 5 mg/kg/day respectively for 4 weeks by oral route using a feeding needle as described by others [30]. Miltefosine treatment was carried out in 8-week infected hamsters for 4 weeks at a dose of 17.5 mg/kg/day orally. Two days after the completion of treatment, hamsters were sacrificed to determine splenic and hepatic parasite burdens by stamp smear method as described elsewhere [27], [31], as well as by the serial dilution method [27].
Blood was collected from hamsters and mice as described previously [27] and kept overnight at 4°C; serum was prepared by centrifugation.
Soluble Leishmanial Antigen (SLA) was prepared from stationary phase LD promastigotes of LD following the published protocol [27]. Briefly, leishmanial lysate from washed promastigotes (109/ml) was prepared by several cycles (minimum six) of freezing (−70°C) and thawing (37°C) followed by 5 min incubation on ice. Partially lysed promastigotes were then disrupted in a sonicator (Misonex, Farmingdale, NY) thrice for 30 s each and centrifuged at 10,000 rpm for 30 min at 4°C. The supernatant containing soluble antigen was collected and the protein concentration was determined by Bradford Protein Assay method (Bio-Rad, Herculis, CA). The prepared antigen was stored at −70°C until further use.
Splenocytes from different experimental groups of hamsters were prepared after Ficoll (Sigma Aldrich, St. Louis, MO) density gradient centrifugation and then suspended in complete RPMI medium. Cells were plated in triplicate at a concentration of 105 cells/well in 96-well plates and allowed to proliferate for 3 days at 37°C in a 5% CO2 incubator either in the presence or absence of SLA (5 µg/ml) (29). For ConA (Sigma Aldrich, St. Louis, MO) induced proliferation, the mitogen was added at a concentration of 5 µg/mL as described previously [27]. Cells were treated with MTT (0.5 mg/mL) 4 hr before harvest as described previously [18] and incubated again at the same condition for 4 more hrs. MTT crystals were then solubilised using Isopropanol-HCl mixture (0.04%) and the absorbance at 570 nm was read at an ELISA plate reader (DTX 800 multimode detector, Beckman Coulter, California).
Serum samples were obtained from different groups of hamsters and mice (five animals per group), and analyzed to determine the parasite SLA-specific antibody titer. 96-well Enzyme-Linked ImmunoSorbent Assay (ELISA) plates were coated with SLA (2 µg/ml) in PBS for overnight at 4°C. The plates were blocked with 5% FCS in PBS at room temperature for 1 h to prevent nonspecific binding. Sera from different groups of hamsters were added at various dilutions, and incubated for 2 h at room temperature. These were diluted 10−1, 10−2, and 10−3 times for the determination of IgG1 and 10−3, 10−4, and 10−5 times for IgG2. Biotin-conjugated mouse anti-hamster IgG1 (BD Biosciences, San Diego, CA) and mouse anti-Armenian and anti-Syrian hamster IgG2 (BD Biosciences, San Diego, CA) were added and incubated for 1 h at room temperature; this was followed by 1 h of incubation with the detection reagent (streptavidin-conjugated horseradish peroxidase). As a peroxide substrate in citrate buffer (0.1 M, pH 4), TMB (Sigma) was added along with 0.1% H2O2 (Merck) to a 96-well plate, and the absorbance at 450 nm was read with an ELISA plate reader [27].
RNA was isolated from the splenocyte of hamsters using Trizol (Invitrogen) as described previously [27]. The forward and reverse primers were used to amplify cytokine transcripts. All of these hamster-specific primers, except for the inducible NO synthase (iNOS) primer, were originally described by Melby et al. [31]. The following forward and reverse primers were used: for IL 10, forward primer 5′ACAATAACTGCACCCACTTC3′ and reverse primer 5′AGGCTTCTATGCAGTTGATG3′ (432-bp product); for IL 4, forward primer 5′CATTGCATYGTTAGCRTCTC3′ and reverse primer 5′TTCCAGGAAGTCTTTCAGTG3′ (463-bp product); for interferon gamma (IFN γ), forward primer 5′GGATATCTGGAGGAACTGGC3′ and reverse primer 5′CGACTCCTTTTCCGCTTCCT3′ (309-bp product); for tumor necrosis factor alpha (TNF α), forward primer 5′GACCACAGAAAGCATGATCC3′ and reverse primer 5′TGACTCCAAAGTAGACCTGC3′ (695-bp product) and for transforming growth factor beta (TGF-β), forward primer 5′CCCTGGAYACCAACTATTGC3′ and reverse primer 5′ATGTTGGACARCTGCTCCAC3′ (310-bp product). To obtain specific amplification for iNOS, the following specific primers were used (6): forward primer 5′ GCAGAATGTGACCATCATGG3′ and reverse primer 5′CTCGAYCTGGTAGTAGTAGAA3′ (198-bp product). For hypoxanthine phosphoribosyl transferase (HPRT) amplification the following primers were used [31]: forward primer 5′ATCACATTATGGCCCT CTGTG3′ and reverse primer 5′CTGATAAAATCTACAGTYATGG3′ (125-bp product). Degenerate bases are indicated above by International Union of Pure and Applied Chemistry designations (Y = C or T; R = A or G).
Details of the procedure have been described previously [27]. Densitometry analyses were done using the ImageJ software (v1.41o), ethidium bromide staining, and visualization under a UV transilluminator. For densitometric calculations, the same band area was used to determine band intensity and normalized for HPRT.
To evaluate long-term therapeutic ability, normal hamsters, infected hamsters and imipramine treated infected hamsters (30 hamsters per group) were used to study survival kinetics as described previously [32].
Spleens and livers were fixed in 10% formalin (Merck) and embedded in paraffin. Tissue sections (5 µm) were stained with hematoxylin-eosin to study their microarchitecture by light microscopy. Photomicrographs were taken with a Nikon Eclipse E200 microscope.
The statistical significance of differences between groups was determined by the unpaired two-tailed Student's t test. Statistical significance was defined as a P value of <0.05 and the results were expressed as averages and standard deviations of triplicate measurements.
Details on the clinical isolates used in this investigation in terms of their sensitivity to Sodium stibogluconate (SSG) have already been published [15]. Out of these, two SSG sensitive (BHU 777 and BHU 816) and four SSG resistant strains (BHU 575, BHU 782, BHU 814 and BHU 872) were selected for this investigation. These isolates were subjected to imipramine treatment in vitro and ex vivo to measure the IC50 and EC50 (Table 1). It was observed that regardless of difference in SSG sensitivity, there was no significant difference in IC50 or EC50 (Table 1). For convenience, the rest of the study was carried out with two SSG-S (BHU 777 and BHU 816) and two SSG-R (BHU 814 & BHU 575) isolates, and were defined as BHU 816(S), BHU 777(S), BHU 814(R) and BHU 575(R) respectively.
The transmembrane potential (ΔΨm) was evaluated using JC-1, a lipophilic cationic dye as described [21]. For this investigation the drugs imipramine and miltefosine were used at a concentration of 75 µM [33] and 40 µM [22] respectively. We observed that both the SSG-R and SSG-S strains showed similar sensitivities to imipramine as evident from the significant decrease in ΔΨm after 8 h of treatment (Figure 1A). On the other hand, miltefosine failed to induce any change in ΔΨm at 8 h of treatment (Figure 1A). After 8 h of imipramine exposure to BHU 575(R), 60% of parasites were apoptotic whereas miltefosine induces apoptosis in only 5.5% parasites (Figure 1B). However at 24 h and 48 h after miltefosine treatment the extent of apoptotic BHU 575(R) was 32.8% and 60.7% respectively (Inset Figure 1B). Similar studies were performed with lesion derived purified amastigotes, BHU 575(R) and BHU 777(S) to find that 8 h treatment with imipramine, but not miltefosine induced a significant decrease in ΔΨm (Figure 1C).
The replication of intracellular LD in the presence of imipramine was studied in in vitro infected MΦ. It was observed that intracellular LD replication was inhibited very efficiently as a function of the imipramine concentration regardless of the SSG sensitivity (Figure 2). The dose required to clear 100% of the intracellular parasites was around 60 µM of imipramine. To show that 60 µM of the drug has no toxic effect; MΦs were incubated with increasing concentration of imipramine. It was observed that almost 100% MΦs remained viable upto 90 µM imipramine (Figure 2, inset).
It is known that infected MΦs are more fluid than their normal counterpart, and this is associated with defective T cell stimulating ability [32]. For convenience BHU 575 (R) infected MΦ were defined as MΦ-575 (R). To show that imipramine restores membrane rigidity, we treated MΦ-575 (R) with increasing dose of imipramine and observed that there was a gradual increase in fluorescence anisotropy (FA) value in a dose dependent manner (Figure 3A). To show that imipramine treatment restores the antigen presenting ability, MΦ-575 (R) were used as antigen presenting cells (APC) with and without imipramine treatment. This showed that the T-cell stimulating ability of MΦ-575 (R) is improved as a function of imipramine concentration as evident from the increase in resulting IL-2 production from I-Ad restricted T-cell hybridoma (Figure 3B).
ROS (Reactive Oxygen Species) and NO (Nitric oxide) are two very important leishmanicidal molecules [34]. Generation of these molecules was found to be enhanced in a time and dose dependent manner in imipramine treated MΦs (Figure 4). ROS generation reached a plateau at around 8 h in the presence of 75 µM imipramine treatment (Figure 4A), whereas maximum NO generation was observed after 20 h exposure at the same concentration of imipramine (Figure 4B).
The effect of graded doses of orally administered imipramine on the splenic and hepatic parasite load in infected hamsters was investigated. Hamsters were infected with BHU 816(S), BHU 777(S), BHU 814(R), or BHU 575(R) LD isolates. We performed a microscopic evaluation of stamp smears and limiting dilutions to detect parasites in tissue samples of infected organs. Eight-week infected hamsters were divided into 4 groups (I–IV) for a given isolate. Groups I–IV received imipramine at doses of 0, 0.05, 0.5 and 5 mg/kg/day respectively for 4 weeks. The results were expressed as total parasite load in terms of LDU. There was no clearance of splenic and hepatic parasite load in group ΙΙ whereas about 50% clearance was observed in group ΙΙΙ animals and there were no detectable parasites in group ΙV animals (Figure 5 A–D). The organ parasite clearance essentially showed similar trends after imipramine treatment for all the isolates regardless of their SSG sensitivity.
To show that group IV hamsters were indeed infected with LD, antileishmanial antibodies were measured in the animals. The presence of antileishmanial IgG2 antibodies in the sera of these animals was detected together with a marginal increase in anti IgG1 titer (Figure 5 A–D c).
We used miltefosine as a reference drug and studied its effect on the organ parasite clearance in a similar set up. Here also eight-week infected hamsters were subjected to miltefosine treatment (17.5 mg/kg/day) orally for 4 weeks and splenic and hepatic parasite load were determined 2 days after completion of the last treatment dose (Figure 5 A–D Group V). The dose of miltefosine was selected as described elsewhere [5]. This showed that miltefosine treated hamsters infected with either BHU 575(R) or BHU 814(R) had low level of residual parasites in the spleen and liver (Figure 5 A–B) whereas miltefosine treated hamsters infected with BHU 816(S) or BHU 777(S) showed no residual parasites (Figure 5 C–D). The presence of residual parasites was further confirmed by limiting dilution experiments with spleen tissue (Figure S1).
To study the status of antileishmanial T cell repertoire in infected and imipramine treated infected hamsters, splenocytes were purified and stimulated either with SLA or ConA. The hamsters were infected either with SSG-S (BHU 816 and BHU 777) or SSG-R (BHU 814 and BHU 575) LD parasites. After completion of imipramine treatment in 8 week infected hamsters, animals were sacrificed and splenocytes were prepared. Fixed concentrations of SLA and ConA were used to stimulate splenocytes as described previously [27]. Splenocytes of group ΙΙ (received 0.05 mg/kg body weight) hamsters failed to mount any antileishmanial immune response but responded well to non specific mitogen ConA regardless of the phenotype of the input parasites for infection (Figure 6). The SLA specific proliferation was marginally improved in group ΙΙΙ (received 0.5 mg/kg body weight/animal), but was further improved in group ΙV (receiving 5 mg/kg body weight/animal). The antileishmanial T cell response was essentially similar regardless of the phenotype of SSG sensitivity. The response to the non specific mitogen ConA remained unaltered in infected and in imipramine treated animals regardless of the dose of imipramine (Figure 6).
The ability of cytokine and iNOS gene expression in infected hamsters and imipramine treated infected hamsters was studied by profiling cytokine gene expression (Figure 7A). The results generated from a densitometry analysis of each hamster were collectively expressed as mean±sd for each cytokine, and the statistical significance between groups was determined (Figure 7A). Comparative cytokine analysis showed that the expression of IFN-γ,TNF-α and iNOS transcripts were 1.43, 1.23, and 1.35 times higher, respectively, in imipramine treated hamsters than in infected hamsters, whereas the levels of TGF-β and IL-10 transcripts were 1.65 and 1.13-fold lower, respectively. Studies of the ratio of IFN-γ to TGF-β or IL-10 revealed that the IFN-γ/TGF-β ratio was 1.78-fold greater in imipramine treated hamsters than in infected hamsters (Figure 7B). Similarly, the IFN-γ/IL-10 ratio was 1.47-fold greater in imipramine treated animals than in infected ones (Figure 7B).
To show that imipramine treated hamsters are also protected in the long run, we studied the survival kinetics of infected hamsters and imipramine treated infected hamsters (5 mg/kg/day for 4 weeks), using normal hamsters as control. In each group 30 hamsters were used. Hamsters were infected at 6 weeks and infection was allowed to proceed for another 8 week, i.e. before initiating any treatment. We observed that 80% of the infected hamsters survived up to 14 weeks, 60% up to 18 weeks, 20% up to28 weeks, and the rest died by 34 weeks. On the other hand, 90% of imipramine treated infected hamsters remained alive until the termination of the experiment, i.e. 44 weeks (Figure 8). Remarkably, amastigotes could not be detected by microscopy in impressions of Giemsa-stained tissue stamp smears of spleen and liver or by limiting dilution experiments at 34th week and also at 44th week. The organ weights had returned to near normal (Table 2 lower panel), and high titers of antileishmanial IgG2 persisted at 34th week (Figure 8 inset).
Evolution of granuloma formation in infected and imipramine treated infected hamsters was studied. Liver section of LD infected hamsters showed immature granuloma formation associated with Kupffer cells surrounded by less number of infiltrating lymphocytes (Figure 9A). High resolution figure of the same showed the presence of parasitized Kupffer cells (Figure 9B) inside the cell assembly. Imipramine treated infected liver tissue shows fair number of lymphocyte infiltration in periportal area and the presence of fair number of mature and uniform granuloma (Figure 9C). Parasites could not be seen in these mature granulomas.
The organ weight at 18 weeks increased significantly in untreated infected animals as compared to uninfected animals, but the 4 week imipramine treatment in infected group showed only a marginal decrease in organ weight (Table 2). After 34 weeks, the organ weight continued to increase marginally in the untreated infected group whereas in the imipramine treated group the organ weight was reduced by 50% as compared to the 18 week time point. However, at 44 weeks the organ weight of the imipramine treated infected group was similar to that of the age matched normal (Table 2).
In an effort to find new orally active chemotherapeutics for visceral leishmaniasis, we evaluated the potential of the existing antidepressant drug imipramine against SSG-R and SSG-S LD parasites. The potent in vitro activity of imipramine against intracellular amastigotes (EC50 = 16.2 µM) [14] as well as promastigotes (Low IC50 values, Table 1) coupled with the absence of obvious cytotoxicity on host MΦ formed the basis for advanced exploration of this promising lead.
Our study showed that imipramine decreases the mitochondrial transmembrane potential of SSG-S and SSG-R LD promastigotes significantly by 8 h of treatment which continued to decrease with time. In contrast, miltefosine showed only a marginal change in the mitochondrial transmembrane potential only at 24 h of treatment. Similar observation was made in imipramine or miltefosine treated purified amastigotes. Oddly enough, imipramine induced 60% of apoptosis of LD promastigote at 8 h whereas miltefosine at that stage induced only 5.5% apoptosis. Similar level of apoptosis by miltefosine was noted at 48 h. This clearly indicates that imipramine induces apoptosis in LD parasites much faster than miltefosine. This is in agreement with reports that apoptosis and change in mitochondrial potential are linked phenomenon [35], [36].
LD infection is associated with increase of membrane fluidity [37] as well as defective antigen presentation by APC's [37]. Therefore the antigen presenting ability of imipramine treated MΦs (also defined as APCs) with murine T cells was studied. We observed that successive doses of imipramine restored membrane rigidity, which was in turn coupled with improved antigen presentation ability (Figure 3A & 3B). This may be attributed to the fact that due to the clearance of intracellular parasites upon imipramine treatment, MΦ may regain its normal fluidity. Due to structural similarities to some extent (fused ring structure with side chain), imipramine can mimic cholesterol which acts as a cementing molecule to pack the lipid bilayer [38]. This may be another reason for the restoration of membrane fluidity. Imipramine has already been reported to mimic the action of cholesterol to regulate protein synthesis in SREB (Sterol Regulatory Element Binding) protein synthesis pathway [39].
We have observed that imipramine induces production of leishmanicidal molecules such as superoxide and nitric oxide in MΦ (Figure 4A & 4B). This observation complements other studies that imipramine induced production of TNF-α [12], an important cytokine for antileishmanial defense. The therapeutic role of imipramine in experimental VL infection may be most likely be attributed to a direct leishmanicidal activity both in vitro and in vivo, its capacity to modulate the host immune response in favour of the host and its ability to induce reactive oxygen species generation on MΦ.
Since imipramine at a dose of 60 µM completely clears the intracellular amastigotes in an in vitro macrophage system (Figure 2), we tested the efficiency of imipramine in an in vivo model. At a dose of 5 mg/kg/day for 4 weeks, the drug indeed clears >99.5% parasites in hamster infected with either SSG-R or SSG-S LD. We chose to administer oral treatment of imipramine for 4 weeks based on the fact that in humans miltefosine is given orally for 4 weeks [40]. The drug imipramine was found to be equally effective in a murine model where infection was induced by SSG-R parasites (Unpublished observation). The in vivo dose of miltefosine was determined based on the animal equivalent of the human dose as described elsewhere [29]. On the other hand, miltefosine treatment for 4 weeks at the dose 17.5 mg/kg/day showed complete clearance of SSG-S but not SSG-R parasites from the spleen and liver of infected hamsters (Figure 5 A–D Group V). The residual SSG-R parasites after miltefosine treatment, even though their number was low, raise concern of cross resistance between miltefosine and SSG. This is perhaps not so surprising since infection with SSG-R parasites upregulates an ABC transporter in the host cells [14] that regulates efflux of both the drugs [17], [41].
Recently there is a great deal of interest for short course of combination treatment regimens [42]. As such, a 10-day treatment protocol has been proposed as an ideal short term regimen by the Drugs for Neglected Diseases Initiative (DNDi) [43]. The DNDi also aims to study new indications for existing medicines in the field of the most neglected diseases. In tune with these recommendations, we also tested the efficacy of imipramine as a monotherapy for a shorter version of treatment, i.e. for 10 days. We again opted to select the dose that provided maximum protection, i.e. 5 mg/kg/day. This regimen showed to clear about 90% of the organ parasites in infected hamsters (Data not shown). Importantly, there is still opportunity to increase the dose of the drug based on the rodent equivalent of human dose [27], also to opt for combination treatment. Imipramine at 5 mg/kg/day for 4 weeks does not affect the hepatic enzymes or creatinine levels in hamsters (Unpublished observation), suggesting that the dose might even be increased further.
Elicitation of effective T cell based host immune response defines the success of antileishmanial chemotherapeutics [44]. Disease severity in experimental animal models infected with SSG-R strains was associated with significantly hampered antigen presentation; antigen-specific T cell activation, low expression of IL-12, TNF-α and IFN-γ, and upregulation of suppressive cytokines IL-10 and TGF-β in murine and hamster models respectively. We therefore tested the immunological parameters associated with successful chemotherapy in SSG-R LD infected animals treated with various therapies.
Enhanced antigen presentation by imipramine treated APCs are also reflected in antigen specific expansion of T cell repertoire in vivo. The skewing of the T cell repertoire towards a Th1 type population is substantiated by the elevated level of IFN γ mRNA expression in splenocytes derived from imipramine treated hamsters.
In vivo treatment with imipramine markedly increased the level of IL-12 in mRNA. The established phase of VL is associated with deactivation of MΦ with severely reduced capacity for production of inflammatory mediators like IL-12 and TNF-α. IL-12 interacts with T cells and induces the initiation and maintenance of Th1 responses via IFN-γ production. IFN-γ and TNF-α are often reported to act synergistically to activate iNOS for the production of NO, the leishmanicidal effector molecule [45]. Strong IL-12 driven IFN-γ coupled with TNF-α triggering by imipramine treatment suggests that these cytokines might be acting in concert to produce NO to effectively kill the parasites.
A growing body of literature correlates IL-10 and TGF-β with susceptibility to Leishmania infection [46]–[54]. Imipramine caused strong suppression of IL-10 and TGF-β production that correlated with successful resolution of infection. Recent reports suggest that resistant parasites modulate the host immunity to exacerbate the ongoing disease pathogenicity [55]. TGF-β is implicated as an important contributor to disease susceptibility or resistance to Leishmania by direct MΦ deactivation [56] and also by increased production of IL-10 [57]. TGF-β being a pleotropic cytokine also suppresses IFN-γ-induced MHC class II expression by inhibiting class II transactivator mRNA [58]. We found significant down regulation of TGF-β mRNA expression in imipramine treated hamsters compared to infected controls. While imipramine treatment attenuates TGF-β expression, it could be causally related to a simultaneous inhibition of IL-10 production [46] with concurrent rescue of MΦ deactivation, thus implying the importance of both the cytokines in disease progression. We like to emphasize the possible role of TGF-β in the outcome of LD infection in hamsters because the IFN-γ/IL-10 ratio changed (1.47 fold) compared to the IFN-γ/TGF-β ratio (1.78 fold) upon imipramine treatment.
The study of the survival kinetics of infected hamsters showed clearly that imipramine treatment increased the life expectancy of the infected hamsters with 90%. In the remainder 10% of the hamsters, death occurred the early time point, the cause of which is not clear. The remaining 90% remained healthy until the termination of the experiment, i.e. 44 weeks post infection. We determined the organ weight (spleen and liver) at 34 weeks and 44 weeks (Table 2) and the parasite burden of the imipramine treated group. Surprisingly, parasites could not be detected in imipramine treated hamsters at the 34th nor at the 44th week and the organ weights were close to normal at 44 weeks post infection. These hamsters at the time of termination of experiment failed to show any parasites but displayed the presence of antileishmanial antibodies, indicating that they were indeed exposed to parasites.
Efficient immune response in the liver depends on the formation of granulomas [59], which is associated with the resolution of hepatic parasite burden [60], [61]. It is well documented that only mature granuloma can develop efficient leishmanicidal mechanism to kill parasites whereas developing immature granulomas lack that efficiency [62], [63]. Parasite killing within the granulomas requires infiltrating monocytes and TNF α [64], although their formation is independent of TNF α family of cytokines [65]. Our previous study with KMP-11 vaccinated hamsters reveals well formed granuloma formation and absence of LD infected Kupffer cells [27]. Our present study showed matured sterile granuloma formation in 4 week of imipramine treated infected hamsters, which was absent in the 12 week infected group (Figure 9) and is associated with the protection. It may be recalled that enhanced maturation of granulomas represents a marker of vaccine induced protection [66].
Orally administered imipramine is rapidly absorbed in the gastro-intestinal tract [9]. Imipramine is a lipophilic compound, binds to albumin, and attains a peak plasma concentration within 2–6 h. This tertiary amine is typically metabolized by demethylation to the secondary and active metabolite, desipramine [9]. Both imipramine and its metabolite desipramine have been found to be equally effective against LD promastigotes [14].
Resistance to SSG is a major problem in the Indian subcontinent and MΦs upregulate both MRP-1 and P-gp upon infection with SSG-R LD leading to efflux of antimonials [15]. Furthermore, circulatory monocytes of kala-azar patients harboring SSG-R LD show over expression of P-gp and MRP-1 [67]. SSG in combination with pharmacological inhibitors of MRP-1 and P-gp favors killing of intracellular SSG-R LD [68]. The tricyclic imipramine is lipophilic and possesses a positive charge due to the nitrogen atom, characteristics that are important to affect the function of P-gp [69] In MDR gene transfected and also in human AML cells ex vivo, such drugs reverse the multidrug resistance phenotype [69]. Thus imipramine will offer an additional advantage since it is a selective P-gp inhibitor. Furthermore, cationic amphiphilic drugs that is basic (pKa 7–8) concentrates on lysosomes [13]. Imipramine being a tertiary amine and weak base will remain as positively charged molecular entity in the body fluids, leading to an affinity towards lysosomes [13], [70]. This unique property is of importance because phagolysosomes constitute the home for intracellular Leishmania parasites.
In conclusion, our study clearly indicated that imipramine is more effective than miltefosine and has a strong potential to be considered as an orally active, highly effective, very cheap, affordable chemotherapeutic agent against kala-azar either alone or in combination.
|
10.1371/journal.pcbi.1002979 | Processing of Multi-dimensional Sensorimotor Information in the Spinal and Cerebellar Neuronal Circuitry: A New Hypothesis | Why are sensory signals and motor command signals combined in the neurons of origin of the spinocerebellar pathways and why are the granule cells that receive this input thresholded with respect to their spike output? In this paper, we synthesize a number of findings into a new hypothesis for how the spinocerebellar systems and the cerebellar cortex can interact to support coordination of our multi-segmented limbs and bodies. A central idea is that recombination of the signals available to the spinocerebellar neurons can be used to approximate a wide array of functions including the spatial and temporal dependencies between limb segments, i.e. information that is necessary in order to achieve coordination. We find that random recombination of sensory and motor signals is not a good strategy since, surprisingly, the number of granule cells severely limits the number of recombinations that can be represented within the cerebellum. Instead, we propose that the spinal circuitry provides useful recombinations, which can be described as linear projections through aspects of the multi-dimensional sensorimotor input space. Granule cells, potentially with the aid of differentiated thresholding from Golgi cells, enhance the utility of these projections by allowing the Purkinje cell to establish piecewise-linear approximations of non-linear functions. Our hypothesis provides a novel view on the function of the spinal circuitry and cerebellar granule layer, illustrating how the coordinating functions of the cerebellum can be crucially supported by the recombinations performed by the neurons of the spinocerebellar systems.
| The movement control of the brain excels in the seamless coordination of our multi-segmented limbs and bodies and in this respect the brain widely outperforms the most advanced technical systems. So far, however, there is little knowledge about the neuronal circuitry mechanisms by which this coordination could be achieved. The present paper makes a synthesis of some recent findings of cerebellar neuronal circuitry functions and spinocerebellar systems to introduce a novel hypothesis of how the cerebellar and spinal cord neuronal networks together establish signals that form a basis for coordination control in the mammalian central nervous system. The hypothesis takes into account some recent, surprising findings about cerebellar granule cell function and explains some long-standing enigmas concerning the structure of and information mediated by the spinocerebellar pathways. It describes some interesting parallels between the spinocerebellar network and Artificial Neural Networks (ANNs), and capitalizes on some of the major conclusions from ANN studies to explain the biological observations.
| The coordination of our multi-segmented bodies and limbs is an unsolved computational challenge that is dealt with seamlessly by the neuronal circuitries of our brains. Within the brain, the cerebellum is considered the most important structure for coordination [1], [2] but we know very little about the mechanisms that could underlie this aspect of cerebellar function. The spinocerebellar and spino-reticulo-cerebellar systems are major sources of mossy fiber (MF) inputs for the widespread cerebellar regions with direct connections to the motor systems, the corticospinal, rubrospinal, reticulospinal, tectospinal and/or vestibulospinal tracts [3]–[13] and the corresponding regions of the cerebellum are implicated in limb coordination in man [13]. The neurons of origin of these systems are located in the spinal cord, where they act as components of the spinal motor circuitry (Figure 1) or receive input directly from such neurons. These systems comprise the spinocerebellar tracts (SCTs), i.e. the ventral spinocerebellar tract (VSCT) and its components including the spinal border cells (SBCs), the components of the dorsal spinocerebellar tract (DSCT), the rostral spinocerebellar tract (RSCT) as well as the spino-reticulocerebellar tracts (SRCTs) providing MFs to the cerebellum via the lateral reticular nucleus (LRN) [14]–[18] (Fig. 1). All of these systems receive sensory feedback either directly or mediated via spinal interneuron systems. They also receive input from supraspinal motor centers, again either directly or mediated via spinal interneuron systems. Single SCT and SRCT neurons can vary with respect to the relative weights of the sensory feedback and motor command components. Considering the influence from motor command systems, it is reasonable to assume that the MFs of these systems carry information to the cerebellum regarding on-going movements, rather than transmitting information resulting from passive sensory stimulation [19]. However, why this combination of different sensory and motor signals occurs before the level of the cerebellum and how it is used by the cerebellum is not well understood.
A second issue of cerebellar function is the tonic inhibition of granule cells (GrCs) in the mature mammalian cerebellum. The inhibition of GrCs in the adult cerebellar cortex, supposedly mostly due to Golgi cell release of GABA, primarily consists in tonic and slowly modulated inhibition whereas fast inhibitory postsynaptic potentials are weak or absent [20], [21]. A couple of observations indicate that GrCs are designed to be tonically inhibited. The tonic GrC inhibition is to a large extent mediated via alpha-6-containing GABA(A) receptors [22], [23], which is a type of receptor that is characterized by long-lasting inhibitory effects and is present at uniquely high concentrations in the cerebellar granule layer [24]–[26]. The second observation is that even when the alpha-6 subunit is knocked-out using genetic engineering, GrCs seem to compensate for this loss of hyperpolarisation by increasing the expression of potassium conductances [27]. At the same time as the tonic component of the GrC inhibition develops, the traditional fast inhibitory response is gradually lost [20], [28]. Accordingly, GrC responding with burst responses to skin stimulation show little sign of Golgi cell inhibition, even though the afferent Golgi cells are relatively strongly activated by the same stimulation [21]. However, although the contribution to the immediate, fast information processing may be modest, Golgi cells can still contribute to cerebellar processing by regulating the difference between the resting membrane potential and spike firing threshold of the GrCs by modulating the level of tonic inhibition. This arrangement could result in that the GrC resting membrane potentials are distributed across the population of cells. Consistent with this idea, the resting membrane potentials of GrCs in vivo [29] had a mean value of −64 mV, but reached at least as low as −80 mV and as high as −40 mV.
A third conspicuous observation is that for GrCs receiving cutaneous input, the 3–5 MF synapses that the GrC receives have been reported to carry functionally equivalent inputs. In a study focused on the cuneocerebellar tract [30], in which the individual MFs carry submodality-specific input from small receptive fields primarily from the distal forelimb, the GrCs were found to sample submodality- and receptive field-specific inputs [21], i.e. information that was equivalent to that of the individual MFs. In an extended study, where cutaneous inputs mediated via the SCT and SRCT systems were included, the MF inputs to individual GrCs were found to be similarly coded, i.e. they originated not only from same receptive field and same submodality, they were also found to originate from precerebellar cells that coded for that particular skin input in the same way [31]. This ‘similar coding’ principle of MF to GrC innervation was suggested to be due to the fact that different afferent pathways process the skin afferent input in different ways, with different degrees of involvement of other modalities and/or descending motor commands. In order to preserve the information generated by the afferent systems similarly coded MFs needed to converge on the same GrCs [19]. This view is supported by numerous anatomical studies, showing that different afferent pathways have focal rather than diffuse termination patterns in the cerebellar cortex [6], [32]–[34] and that different afferent pathways have complementary distributions in the GrC layer [35]–[37]. Hence, at least for the cerebellar regions involved in limb control, MFs carrying functionally equivalent inputs preferably target the same set of GrCs. A consequence is that the probability of finding GrCs sampling functionally disparate input would be expected to be low. Altogether, the findings of these studies were at odds with previous theoretical predictions that GrCs were expected to integrate MF information from widely separate, functionally disparate sources [38], [39], and presented a challenge to our understanding of the cerebellar granule layer and thereby of the cerebellar cortex in general.
In the present paper, we seek a theoretical explanation for how these three observations can be understood in terms of the function of the spinal and cerebellar neuronal circuitry and how they could provide for the integration across input dimensions necessary to achieve coordination. In the initial part of the paper, we account for the main settings of our hypothesis to describe how sensorimotor functions can be generated in the SCT and SRCT MF systems and how the cerebellum could integrate this information to approximate useful functions. We continue by exploring how non-linear interactions between input dimensions could be handled in the spinal circuitry and the cerebellum and the limitations that may apply to these systems in this respect. For this purpose we construct a simple static model and use a concrete example of non-linear interactions between input dimensions used in the performance evaluation of previous cerebellar models.
As the Purkinje cells (PCs) constitute the only output from the cerebellar cortex, their responses to stimuli reaching the cerebellum through the MFs determine what the cerebellar cortex is able to do, and what type of transformations of the input signals that are possible. While Albus [38] viewed the PC as a binary perceptron for classification, he also acknowledged that the PC is able to modulate its relatively high spontaneous firing frequency, creating in principle a continuous output signal. The spontaneous firing rate of simple spikes can be modulated in both excitatory and inhibitory directions by specific inputs, in a manner consistent with their generation by excitatory parallel fiber (PF) and inhibitory interneuronal inputs, respectively, and with these inputs being summed primarily in a linear fashion [40]. Since the molecular layer interneurons are innervated by the parallel fibers and in turn inhibit the PC, the weights between PFs and PCs are allowed to become negative [40], [41]. The bidirectional plasticity and the complementary location of the receptive fields in PCs and interneurons confirm this assumption [42], [43]. Eq. (1) describes a simplified relationship between PF and PC activity, and forms the basis of the later mathematical models of this paper, where the properties of the neurons in the granular layer will also be taken into consideration.(1)where is the PC activity, the activity of the GrC, is the spontaneuos activity of the PC and is the synaptic efficacy between the GrC and the PC. Note than is allowed to be negative since the GrC input can be mediated through inhibitory interneurons [40], [41].
A distinct feature of GrCs is the marked difference between the resting membrane potential and the threshold potential for spike firing [21], [29], [44]. Once the spike firing threshold is crossed, however, the input-output relationship of the GrCs is approximately linear [29], [45]. The activity of a GrC can be approximated to be,(2) represents the synaptic weight between the MF and the GrC and is the distance to the spike firing threshold from the GrC resting membrane potential. The firing threshold is constructed through the use of a ramp function , which is defined to be equal to whenever is equal to or lower than , but equal to whenever is larger than . When the synaptic input depolarizes the neuron above , the GrC spike output will also start to increase in proportion to the synaptic input. Eqs. (1) & (2) assumes that the MF input to the cerebellum is rate encoded, consistent with existing in vivo studies of MFs in awake animals during behaviour [46]–[49], and during fictive locomotion [50]–[52].
As pointed out above, previous observations indicate that an individual GrC sample functionally equivalent input from all of its incoming MFs [21], [31]. As individual MFs branch to contact up to thousands of GrCs [33], the GrC layer would be expected to contain a highly redundant representation of MF inputs, which is also suggested by a systematic investigation of MF receptive fields [53]. Due to neural noise and the limitations of spike encoded transmission of information [54], some redundancy in the GrC representation would be expected to be needed to average out the noise. The very large number of GrCs innervated by the same MF does however suggest that the GrC population by some means recode the incoming MF signals, rather than just compensate for the inherent noise in the afferent signal. In the following section, we propose a solution for how useful expansion recoding can be performed by a redundant population that sample functionally equivalent input. In this view, the role of the Golgi cell to GrC tonic inhibition, potentially supplemented by other sources of GABA input as described in the discussion, is to allow the GrC responses within a redundant population to become sufficiently varied, hence playing a crucial role of the proposed expansion recoding.
Using a population of redundant GrCs, which sample functionally equivalent MF input and differ with respect to the separation between the resting and the spike firing threshold, it becomes possible for the PC to combine the inputs from these GrCs into a piecewise-linear (PL) approximation of an arbitrary smooth function, see Figure 2A. In this case, even if each GrC in the population would receive exactly the same MF inputs, the diversified thresholds in the population of GrCs would be a useful feature for the PC that integrates these signals. Since the GrC thresholds directly correspond to the knots of the PL-approximation (see Figure 2A), the number of GrCs is directly related to the approximation error, which for PL approximations is bounded by , where is the number of knots or GrCs. In other words, as new knots are introduced, the upper bound of the approximation error will shrink in proportion to the number of new knots.
To further develop this reasoning, we proceed by considering a simplified example of input from knee skin afferents, for which a relatively straight-forward relationship between knee-joint angle and firing frequency of single afferents on the hairy skin of the thigh has been reported [55]. In this study, it was shown that skin afferents with receptive fields located at different distances from the knee joint all could code for the knee angle but with different intensity. Considering this example, it is conceivable that in addition to the tonic inhibition, GrC populations with related but not completely equivalent inputs could help differentiate the firing thresholds with respect to the knee joint angle over a larger input region. The combined effect of both Golgi cell inhibition and related receptive fields is illustrated in Figure 2B. In accordance with this view, the distribution of the active range, with respect to joint angle within a population of muscle-spindle afferents, has been found to cover the entire range of joint-angle positions investigated, while any individual afferent from the population had a much smaller active range [56]. The properties of the one-dimensional example in Figure 2B do, however, not tell how the one-dimensional arrangement could be expanded to integrate signals related to more than one input dimension and approximate complex, possibly non-linear interactions between them. To approach this issue, we consider in addition to the skin stretch sensors another type of sensor, the Ib afferents, the firing frequency of which is related to muscle force [57]. Note that even though we are presenting these ideas using sensory inputs, the following line of reasoning would also apply to information from the motor command domain, but since the exact content and format of these signals is less well known it would make for a worse example. For example, skin stretch might well be replaced with motor command in Figure 2C below.
The naïve approach to multidimensional input would be to simply superimpose one PL approximation along each dimension, see Figure 2C. Each point in the input space of the combined approximation, would map to the sum of all the separate one-dimensional approximations. Relating this approach to the cerebellar structure, it would correspond to each GrC receiving signals from a single input dimension. The PC, working as a linear integrator of GrC inputs, would then integrate information from the input dimensions available in the GrC population contacting this PC via the parallel fibers, hence integrating the information into multiple combined PL-approximations of these input dimensions. But, a presumed core cerebellar function is to improve coordination, which depends on interactions between adjacent limb segments, between multiple modalities or submodalities and between motor command and sensory feedback, i.e. interactions that are most likely non-linear. It is not obvious to what extent the naïve approach could be used to approximate such non-linear interactions. Given the description of the MF-GrC system so far, it is possible to interpret Eqs. (1) and (2) as a radial basis-function network with PL radial basis functions [58]. Light and Cheney [59] showed that such a network cannot approximate arbitrary multivariate functions to arbitrary precision with no recombination of the input signals prior to the hidden layer (‘hidden layer’ corresponds to a layer of neurons located between the input layer and the output layer of the system, i.e. in our case the GrCs). Consequently, if the cerebellum is to be able to approximate arbitrary functions, it is not enough to have raw afferent signals as MF content as in the naïve approach described above – there is a need for the signals to be recombined.
In contrast to the naïve approach, the most general approach would be to allow each GrC to receive any combination of all the inputs that reach the cerebellum via the MFs as illustrated in Figure 3A. It is possible to formalize such a multidimensional PL approximation into the following equation, in order to investigate its properties:(3)where is the number of GrCs and is the number of raw signals transmitted through the MFs. As opposed to in Eq. (2), is not a single synaptic efficacy of the MF GrC transmission, but the total transmission efficacy between the raw input signal and the GrC, which includes contributions from one or more MF-to-GrC synapses. We use to denote signals that have not been recombined, e.g. sensory signals related to a single input dimension. It should be noted that Eqs. (2) and (3) are not in contrast to each other, since in Eq. (2) the efficacies between the raw signals and the MF activities are not mentioned. In other words, any situation that can be described by Eq. (3) can also be true for Eq. (2) and Eq. (1) by the correct choice of efficacies between raw signals and the MF activities.
Eq. (3) has close links and mathematical similarities to several concepts in multivariate regression and in particular artificial neural nets (ANN), see Bishop [60]. It is a special case of a feed-forward ANN with a single hidden layer, where each hidden unit would correspond to a GrC, and all hidden units have ramp activation functions. In Eq. (3) and Figure 3A, as in a regular feed-forward ANN, each hidden unit (i.e. GrC) is innervated by all units in the input layer (i.e. raw input signals). Marr and Albus assumed such multimodal recombination as an essential part of their models, in order to expand the input onto a higher dimensional space enabling non-linear classification. It is known that in theory such networks can approximate arbitrary functions [61], but in practice they often require either a very large amount of neurons or extremely large synaptic weights to approximate high dimensional or complex function surfaces [62].
Available evidence from the mammalian cerebellum, however, does not support any substantial multimodal recombination within the granule layer (see Introduction) and the linear integrative and firing properties of the individual GrCs combined with the rate coding in the SCT/SRCT systems (see above) do not seem well suited for calculating non-linear functions. But extensive multimodal recombination of inputs does occur in the major MF pathways, i.e. in the neurons of origin of the SCTs and SRCTs [8], [14]–[17], [49], [63]–[70], and they are a major source of integrated sensorimotor information to the cerebellum. The biological observations consequently suggest an alternative view of the spinocerebellar network structure (Figure 3B). Notably, in this view, the multimodal recombination that is necessary in the Marr-Albus type of classifier is retained but the recombination is placed prior to the granule layer.
The recombination of two or more input dimensions can be viewed as a projection in the input space defined by the available input dimensions. The term projection is used because a recombination of the input variables, provided that it is a linear recombination, can be seen as a projection along a straight line. It should be noted that even though we in the account below primarily discuss linear projections, also non-linear projections are valid. In fact, non-linear recombinations could differentiate the input in addition to the firing thresholds of the GrCs. In specific cases, such non-linear recombinations have been shown to improve the quality of approximations containing non-linear interactions between input dimensions [71]. Since the SCT/SRCT neurons are located within a complex network of spinal interneurons, non-linear functions can potentially be generated here.
If the available recombinations of the input dimensions are present already at the stage of the MFs (Figure 3B), then the number of MFs imposes a restriction on how many projections that can be represented in the system. The number of GrCs is much higher than the number of MFs [33], [40], but as noted above, if the PCs are to be able to make PL approximations along each projection, a population of GrCs is needed to represent each type of sensorimotor combination provided by the MF systems (i.e., to create a large number of knots in the PL approximation) (Figure 2). In any case, as we will see below, the number of GrCs is far too low to be able to accommodate all possible recombinations in the system. Hence, the cerebellum is limited to working with a restricted set of recombinations, or projections. In the case of a restricted number of projections in the available input space, there exist regression models which allow us to further analyze the properties of such a model [72]. Eq. (3) can be changed into the more restricted form(4)where is the number of projections, determines the direction of the projection and the number of GrCs along projection . Before we apply the model in Eq. (4) to the double joint arm we will describe to what extent the ‘curse of dimensionality’ [73] affects and probably shapes the properties of the projections available in these systems.
To illustrate the curse of dimensionality and its implications for the amount of sensorimotor functions that could be represented in the cerebellum, all inputs to the system is first assumed to be limited to a finite number of values. Since there is a physical limit upon most biological signals, the number of finite values () of the input could for example be chosen such that we can achieve a sufficiently good approximation error according to the bound, , that was presented earlier. The size of the space spanned by one such discretized input could be considered to be . If we introduce another input or dimension to this space, it will become a square with size . By adding another input dimension, the space will become a cube with size .
Hence, the size of the input space grows exponentially with the number of input dimensions. Furthermore, each GrC, through the PF to PC synaptic weight, can only determine the value of the approximation at a single point in the input space [58] (Figures 2,3). If the number of dimensions increase, and the size of the input space grow, the average distance between randomly placed points in that space grows as , where d is the number of dimensions and N is the number of points spread randomly across the entire space [60]. Hence, as more and more dimensions are introduced to the input space, the number of approximation points or GrCs has to grow exponentially in order to maintain the average distance between these points. Formally, the root mean squared error (RMSE) can be shown to be related to the average distance between points and is bounded by [74]. The exponential growth of the space spanned by all input dimensions and all the resulting implications are what Bellman coined as the curse of dimensionality [73].
In a biological system, one sensory receptor sampling information from one submodality at one locus in the body would correspond to one dimension. In the extreme case, each sensor is considered to sample unique information and therefore represents a unique input dimension. In this case, the total input space would be of astronomical size due to the described exponential growth per dimension. It is completely unfeasible that such a space can be represented to any detail within the brain. To illustrate the exponential growth in numbers even in a simplified system, let us consider a set of receptors where each receptor samples information from one submodality around a single joint. The hand alone has approximately 20 degrees of freedom. In order to encode the entire input space (for simplicity, we only consider static configurations of the hand) with the MF-GrC population in the human (assuming 10 billion GrCs are devoted to hand control, which is likely to be a huge overestimate) the system would be limited to about 3 knots along each input dimension. Hence, even with these crude levels of accuracy in terms of sensory input (and disregarding the role of the motor commands), the number of GrCs needed reaches astronomical figures. If we add the wrist and hence 2 more degrees of freedom to the input, the number of GrCs required for the same crude accuracy would rise to 100 billion GrCs (compared to the estimates of a total of 70–100 billion GrCs in the human cerebellum [75]), illustrating the exponential growth. Again, it is easier to discuss the input dimensions in terms of sensor signals, which we know relatively well what they code for. Nonetheless, it should be recalled that we in addition have to take into account the motor command, which is likely to represent another multi-dimensional set of signals whose functions and interdependencies with the sensor signals would also need to be approximated.
The bounds upon the approximation error we have discussed so far are the worst case scenarios, in which we are assuming that the cerebellum needs to approximate arbitrary functions. As described above, it would require all possible recombinations to be present at level of the granule layer, requiring an enormous amount of GrCs, most of which would remain superfluous if there were no non-linear interactions between the recombined dimensions that actually needed to be approximated. In contrast, if it was sufficient to only use specific projections in the input space, selected to enable the system to approximate specific non-linear functions, the number of MFs and GrCs needed could possibly be substantially reduced. For example, input dimensions that never interact (that we will refer to as ‘superfluous projections’) would not need to converge at all, and it would be possible to approximate additive functions of two or more input dimensions even if they did not converge until the PC layer. Interestingly, Barron [74] showed that it is possible to construct a feed forward ANN that reduces the bound upon the approximation error from to , showing that it in principle is possible to reduce the number of needed hidden units in the ANN to the point where the curse of dimensionality is completely avoided but the precision of the approximation is still maintained. While it is unlikely that all requirements to reach the second bound can be fulfilled in the biological system, the results of Barron still demonstrates that finding optimal recombinations of inputs can be used to reduce the number of projections available in the system while maintaining acceptable levels of the approximation errors. Given that it is known that the SCT/SRCT systems represent a limited number (in relative terms) of combinations of sensor and motor signals, this viewpoint has a high biological relevance. Such recombination would correspond to that the SRCT/SCT systems have selected to represent specific projections within the input space. Consequently, the structure of the spinocerebellar network prior to the granule layer needs to be considered in order to explain how the curse of dimensionality can be at least partially lifted.
To illustrate how the coordination of a multi-segmented limb could work in principle, we consider the simplified example of controlling a planar double joint arm. The planar double joint arm has previously been used to evaluate biophysically detailed models of cerebellar function [76], [77], also including feasible neural delays [78]. Note that the kinematic variables in this example could reside both in the domain of the sensors and in the domain of the motor command – this is not relevant in this example, though, as we only intend to illustrate in principle how projections can be useful in approximating non-linear functions across different dimensions. This example serves the additional purpose of illustrating that a limited number of appropriate projections can go a long way in capturing the interdependencies between different input dimensions. Thereby, this example ties back to the SCT/SRCT systems where more selective, rather than unlimited, combinations of inputs are available. Position, speed and acceleration, i.e. the variables represented in our example, from multiple arm-joints do indeed seem to be represented in individual MFs of the arm-controlling regions of the cerebellum [46] although this does not mean to imply that these are the only variables represented or that the terms in Eq. (5) correspond to the exact interdependencies represented within the spinocerebellar system.
To enable accurate control of the arm, the movement of both joints needs to be coordinated due to interaction torques arising from inertial, centripetal and Coriolis forces [79]. These interactions lead to non-linear terms in the inverse equations of motion of the arm, i.e. the transformation from joint angles, velocities and accelerations to joint torques, which depends on two or three kinematic variables of the arm joints. The three types of non-linear dependencies among the different terms in the inverse dynamics of the planar double joint arm are listed below,(5)where is the joint angle, and are joint velocities and accelerations, respectively, and the subscripts and denote the elbow and shoulder joints, respectively. The terms are simplified compared to those in the inverse dynamics [79], but retain all non-linear interactions. In particular, all constant coefficients are removed and all are replaced with to ease visual comparisons of the results. Variables without subscripts indicate that the term is present within the inverse dynamics with both a shoulder and an elbow variant.
In order to explore how the neuronal circuitry could be used to construct approximations of these equations, we apply the basic static model from Eq. (4) to the non-linear terms in Eq. (5). In particular, we investigate how the input to the cerebellum through the recombination of the input variables influences the quality, or the root mean squared error (RMSE), of the approximations that were constructed.
We also investigated whether selecting particular combinations of input enables better approximations, or if it is sufficient to choose the recombinations at random. By creating approximations of the terms in Eq. (5) with a varied number of random projections as input to the granule layer, it is possible to investigate how reliable an approach using a random selection would be and how many different projections such a method would need to use. To evaluate the results in the case of random projections of input, we compare the approximations using random projections situation with approximations where the directions of the projections were also optimized, instead of chosen at random. Having the optimal projections also allow us to compare approximations using a varied number of GrCs. They can also be used to explore if the optimal projections or at least quality-wise similar (‘good enough’) projections are included in the set of random projections that were used in the previous step. Finally, it also allows us to explore if the results using fewer different recombinations could be significantly improved or compensated by increasing the number of GrCs.
All approximations are constructed by minimizing the error function in Eq. (6), which calculates the RMSE of the approximation compared to the approximated function over a grid of points covering the input space.(6)where is the approximated function or target function, the approximation and the are placed on an equidistant grid covering the input space.
In order to capture the relevant shapes of the functions surfaces of the terms in Eq. (5), the input to the trigonometric functions range between and . The RMSE is evaluated as the proportion of the maximum RMSE obtained with Eq. (6) having . All two-dimensional approximations were constructed using the Levenberg-Marquardt algorithm [80] using the implementation in the Matlab optimization toolbox. Matlabs Nelder-Mead algorithm [81] was found to handle the larger amount of unknowns better than the Levenberg-Marquardt algorithm, due to the need for Jacobians in the Levenberg-Marquardt method, and was used to create the three-dimensional approximations. It is important to note that we use general approximation methods with the intent to prove the possibility to construct approximations with the model rather than to develop a method to construct such approximations.
We explored how well non-linear functions could be captured using a limited number of projections, each represented in a limited number of GrCs, in a model of the SCT/SRCT systems and the cerebellar cortex to which the constraints described above applied. The functions that the model was assigned to approximate were a couple of central functions for monitoring intersegment dependencies in a multi-segmented limb (see ‘Application’), i.e. functions that have to be monitored somewhere in a system designed to achieve coordination control. Given what is known of the SCT and SRCT synaptic inputs and signals (see preceding sections), the signals necessary for approximating such functions might well be generated in these systems. Figure 4A illustrate how well the approximations created with the model in Eq. (4) could capture the non-linearities in Eq. (5) using a limited number of GrCs and a limited number of projections. The approximations with the lowest RMSE (relative to the actual function) using 60 GrCs along 1 and 2 projections, respectively, are compared to the actual functions (Figure 4A, right column). As it is hard to illustrate approximations with three regressors, the three-dimensional term , is exchanged with together with in Figure 4A. The colors illustrate the shape of the surface in a similar fashion as in Figure 2B, but in contrast to Fig. 2B the projections are not perpendicular to each other as in the naïve case with raw inputs, but indicated by the broken lines. Due to the symmetry of , approximations with equal RMSEs can be obtained by rotating the projections making them close to equal to the projections used for the optimal approximation of . The best found approximation of using two projections deviates with a considerably higher RMSE and the projections nearly overlap.
In a next step, we tested the effects of increasing the number of projections available to the system beyond two projections. Figure 4B illustrates that on average, using random projections, RMSE values improved substantially when moving from two to three projections. Notably, the approximation of , which was hard to approximate using two projections, could now be approximated with much better RMSE values. Figure 4B also indicates the RMSE values obtained for non-recombined raw signals (only two projections, corresponding to the two axes in the panels of Figure 4A), which generally lead to poorer results. Figure 4B in addition indicates separately the projections leading to the lowest RMSE values (i.e. the projections illustrated as dashed lines in Fig. 4A). This illustrates that out of a set of random projections, it is generally possible to find close to optimal recombinations (or, rather, ‘good enough’ recombinations, see Discussion).
In mathematical terms, already the relatively simple double joint arm gives rise to several functions that cannot be approximated using raw (i.e. non-recombined) inputs to the GrCs. E.g. for and , both the worst case using a single projection and the approximations using the raw inputs reached 100% of the maximum RMSE, i.e. a flat function surface equivalent to . The average and worst case RMSE of the approximations was substantially improved as the number of available projections was increased (Figure 4B). The best obtained RMSE did however stay relatively constant as the number of projections was increased to two or more.
Concerning , the approximation using both raw signals had no better RMSE than the approximation using a single ideal projection of both signals. Again, when the number of projections in the system is increased, the average and worst case RMSE was substantially reduced also in this case. At the same time, the best obtained RMSE did not improve as the number of projections was increased to three or more.
Common for all three functions was that low RMSEs could be obtained with only two or three projections (Figure 4B) and that the average RMSE of the approximations could be reduced by a rather modest increase in the number of projections, especially approximating . The above analysis indicates that recombinations of sensorimotor signals describing the non-linear dependencies between limb segments is necessary at the level of the cerebellar granule layer in order to create approximations of the functions in Eq. (55). At the same time, several GrCs can receive the same MF input without impairing the quality of the approximation in any significant way, allowing the recombination to take place already prior to the granule layer.
While it was possible to find projections trough random selection that yielded good approximations of the 2-dimensional terms, the same is not true for the three-dimensional term (see Figure 5). It should be possible to find approximations that reach the same RMSEs as with its two-dimensional counterpart (see Figure 4B), but the constructed approximations indicate that random selection of projections is not sufficient to reach the same low RMSEs, even with a relatively high number of projections. However, by also including the projection directions to the approximation algorithm to search for the optimal projections, three-dimensional approximations with RMSEs comparable with their two-dimensional counterpart were found, even with a relatively low number of projections (see Figure 6). These observations indicate that the higher the number of dimensions, the worse the random projections would work. Hence, in particular for high-dimensional inputs, it would be highly inefficient for the spinocerebellar systems, and the cerebellar granule layer, to rely on random projections.
It should be noted that the algorithm did not converge by its own measure when Nelder-Mead was used, but had to be limited to a maximum number of iterations. Also, the algorithm did converge on local minima solutions and had to be restarted around the found minima to reach the solutions that are presented in Figure 6. There is no formal guarantee that the found solutions lie close to the global optima, and should be considered as upper bounds as better solutions could exist.
It is also interesting to note that the approximations using optimized projections had close to the same RMSE as the most accurate approximations created with random projections. By comparing the RMSEs of the approximations using two and three projections in Figure 6, it is also evident that it is not enough to increase the number of GrCs in the approximation using two projections to reach the RMSE of the approximation obtained using three projections. However, as long as the number of projections is sufficient, additional projections does not seem to increase the accuracy as can be seen comparing the RMSEs of the approximations using 4 and 8 projections in Figure 6. Altogether, these findings (i.e. from Figures 4–6) indicate that while it was possible to construct approximations of all the non-linear terms in Eq. (5) with small RMSEs, the properties of the function that is approximated determines how many projections that are necessary and how they should be selected.
Recombination of sensory and motor command signals in the spinocerebellar systems is likely a crucial circuitry feature to allow the cerebellum to be able to perform a variety of motor control functions such as coordination. Here we made a theoretical analysis of how this information can be used, based on some novel constraints coming out of in vivo analyses of cerebellar GrC function and the properties of the spinocerebellar systems. Important conclusions are:
A main focus of our hypothesis is that the spinal circuitry makes a number of sensorimotor functions, useful in the task of coordination, available to the cerebellar cortex through its MF projections via the SCT/SRCT systems. A second component of the hypothesis is how this information could be received by the cerebellar granule layer. Here, our hypothesis and the constructed model rests on a number of experimental observations in vivo, which suggest that rate coding is the predominant form of coding in MFs, that GrCs receive input from functionally similar MFs and that the dominant mode of control of excitability in GrCs is through tonic inhibition, presumably mainly through Golgi cells. One of our main assumptions is that these observations apply. On the view of the cerebellar cortex, the hypothesis extends the adaptive filter hypothesis [40], [82] and addresses the some of the questions raised in a recent review [40] related to the role of the granule layer and the MF signaling that follow from the described in vivo findings. Naturally, in cases where the MF systems would operate primarily through spike time coding and where inhibitory synaptic inputs from Golgi cells to GrCs primarily operates through fast, phasic IPSPs, very different functional interpretations of the function of the granule layer would apply. Under these circumstances other functional models of the granule layer, similar to those presented in other reviews and models [83], [84], would become more likely. At present, however, support for our assumptions from in vivo recordings seems relatively strong. All recordings of MFs in awake animals during natural behavior seem to support the interpretation that MFs operate with rate coding, in particular for the spinocerebellar systems and within the limb control zones that we are considering here [46], [49]–[52], [85], [86] but also for MFs in the oculomotor controlling regions of the cerebellum [48] and quite possibly also for vestibulocerebellum [47]. Also Golgi cells seem to follow the rate coding principle in relation to controlled movement [48], [47]. Similarly, although intracellular recordings in GrCs in vivo are rare, studies of their inhibitory responses in vivo are rarer still, but those available indicate that a prominent tonic inhibition is existent [21], that in the adult animal fast IPSPs are difficult to detect [29], whereas tonic inhibition under Golgi cell control is demonstrable [88], and that even in the juvenile animal (which is remarkable considering that full maturation of the Goc-grc inhibition does not occur until adulthood, see Introduction) more than 98% of the inhibitory charge is carried by tonic inhibition [89]. Hence, although the predominance of tonic inhibition in the Golgi cell to GrC inhibition is still controversial, so far all in vivo studies available support this idea.
Notably, another controversy with respect to GrC function, which revolves around the question whether single MFs on high-frequency repetitive activation can cause the GrC to fire or not [40], [90], does not affect the functional view presented here. As long as the MF – GrC transmission has a threshold this functional view would not change. In fact, a varying threshold across the population of GrCs, which is beneficial for our present model, could possibly in some cases result in that single MFs would have such heavy influence on GrC firing also in vivo, although this remains to be shown [40].
Note that as long as the inhibition exerted on GrCs is predominantly tonic, the exact timing of Golgi cell spikes, although regulated through an intricate set of mechanisms [91], has a comparably small impact in this model.
In the present hypothesis, Golgi cells or GABAergic control of GrCs have the important function of diversifying the GrC thresholds to maximize the ability of the PCs to make piecewise-linear approximations of any function that is useful for fulfilling their roles in motor control. Although the source of GABA that sustain the tonic inhibition of GrCs is somewhat controversial, and may include for example glial sources [92], assuming that there is a contribution from GABA released by Golgi cells has some interesting consequences. By regulating Golgi cell activity, it might then be possible to give the same GrC different thresholds in different contexts. This would depend on the fact that MFs representing a given input have focal terminations [35], [37], [53], which should create clusters of GrCs with functionally similar inputs, and that Golgi cells inhibit primarily local clusters of GrCs [12], [93]. Hence, by modulating the Golgi cell spike firing to different depths [88], which would result in a modulated tonic inhibition [87], the cerebellar cortex could assign different specific roles to the same GrC in different contexts, even though the input it receives from the MFs are still coding for the same function. Exploring these issues, and the mechanisms of how the cerebellar cortical circuitry could regulate the thresholds of specific GrC clusters in different contexts, are important outstanding issues for future cerebellar research.
Although it is known that the spinal neurons of origin of the SCT and SRCT pathways integrate sensory feedback with descending motor command signals, the idea that this integration is used to provide the cerebellum with sensorimotor functions or projections into sensorimotor space is a fundamental assumption in the current hypothesis. This assumption has not been extensively explored, but there is certainly data that is consistent with this idea [94], [95].
As argued in the Model and Hypothesis section, in relation to the potential number of projections that could be formed within the complex sensorimotor space of our bodies, the cerebellum and the number of GrCs that are available poses a severe limit on the number of projections that can be represented. It follows that the selection of these projections is an important step for the organism so that the most useful calculations, relative to the sensorimotor apparatus available, is made available to the cerebellum.
How these projections are established and configured are crucial outstanding issues, but a number of observations suggest that the spinal cord is an appropriate place to take the decisions of which projections that are relevant. Simulations using a realistic network structure and local sensory feedback patterns have shown that the spinal circuitry does provide a number of useful sensorimotor functions that can be played on by using motor commands [96]. The connectivity of the circuitry is established through plasticity processes whose outcome depends on the configuration of the sensorimotor apparatus [97] and the existence of spontaneous movements [98]. In other words, the development of the spinal circuitry is adapted to the development of the sensorimotor apparatus and the correspondence between movement and sensory feedback. It is likely that individual neurons of the spinal circuitry during the development ‘finds’ appropriate combinations of sensors and motor command signals, associations that are helpful for brain movement control by providing such assistive sensorimotor functions. The close correspondence between the sensor signals and motor command signals in the individual neurons could be a means for the spinocerebellar system to avoid superfluous recombinations (which the system can certainly not afford, see argumentation under ‘Limited number of GrCs’) and that ‘good enough’ recombinations (close to optimal) might be formed with a high probability. Describing the processes that establish the precise patterns of recombinations made available by individual spinal interneurons could be a particular intriguing example of advanced learning processes within the central nervous system and is another crucial outstanding issue for future research.
|
10.1371/journal.pntd.0000096 | Pediatric Measles Vaccine Expressing a Dengue Antigen Induces Durable Serotype-specific Neutralizing Antibodies to Dengue Virus | Dengue disease is an increasing global health problem that threatens one-third of the world's population. Despite decades of efforts, no licensed vaccine against dengue is available. With the aim to develop an affordable vaccine that could be used in young populations living in tropical areas, we evaluated a new strategy based on the expression of a minimal dengue antigen by a vector derived from pediatric live-attenuated Schwarz measles vaccine (MV). As a proof-of-concept, we inserted into the MV vector a sequence encoding a minimal combined dengue antigen composed of the envelope domain III (EDIII) fused to the ectodomain of the membrane protein (ectoM) from DV serotype-1. Immunization of mice susceptible to MV resulted in a long-term production of DV1 serotype-specific neutralizing antibodies. The presence of ectoM was critical to the immunogenicity of inserted EDIII. The adjuvant capacity of ectoM correlated with its ability to promote the maturation of dendritic cells and the secretion of proinflammatory and antiviral cytokines and chemokines involved in adaptive immunity. The protective efficacy of this vaccine should be studied in non-human primates. A combined measles–dengue vaccine might provide a one-shot approach to immunize children against both diseases where they co-exist.
| Dengue is a tropical emerging disease that threatens one-third of the world's population, mainly children under the age of 15. The development of an affordable pediatric vaccine that could provide long-term protection against all four dengue serotypes remains a global public health priority. To address this challenge, we evaluated a strategy based on the expression of a minimal dengue antigen by live attenuated measles vaccine (MV), one of the most safe, stable, and effective human vaccines. As a proof-of-concept, we constructed a MV vector expressing a secreted dengue antigen composed of the domain III of the envelope glycoprotein (EDIII), which contains major serotype-specific neutralizing epitopes, fused to the ectodomain of the membrane protein (ectoM) from DV-1, as an adjuvant. This vector induced in mice durable serotype-specific virus-neutralizing antibodies against DV1. The remarkable adjuvant capacity of ectoM to EDIII immunogenicity was correlated to its capacity to mature dendritic cells, known to initiate immune response, and to activate the secretion of a panel of cytokines and chemokines determinant for the establishment of specific adaptive immunity. Such strategy might offer pediatric vaccines to immunize children simultaneously against measles and dengue in areas of the world where the diseases co-exist.
| Dengue fever is a mosquito-borne viral disease that threatens the health of a third of the world's population. During the last twenty years, the four serotypes of dengue virus spread throughout the tropics due to the presence of the mosquito vector Aedes aegypti in all urban sites and to the major demographic changes that occurred in these regions. This global re-emergence shows larger epidemics associated with more severe disease [1]. Dengue is a major worldwide public health problem with an estimated 100 million annual cases of dengue fever (DF) and 500,000 annual cases of dengue hemorrhagic fever (DHF), the severe form of the disease, resulting in about 25,000 fatal cases, mainly in children under the age of 15. Although global prevention appears the best strategy to control dengue expansion, there is still no licensed vaccine available.
Dengue viruses (DV) are enveloped, positive-stranded RNA viruses (Flaviviridae family). Four antigenically distinct viral serotypes exist (DV1-4). The surface of virions is composed of the major envelope glycoprotein (E) and a small membrane protein (M). Very little information is available concerning the role of the 75-amino acid long M protein. We previously reported that ectopic expression of the 40-residue intraluminal ectodomain of M (referred hereafter as ectoM) is able to induce apoptosis in mammalian cells, suggesting that M might play an important role in the pathogenicity of flaviviruses [2]. The envelope E protein, which is exposed on the surface of viral particles, is responsible for virus attachment and virus-specific membrane fusion. Anti-E antibodies inhibit viral binding to cells and neutralize infectivity. A primary DV infection is believed to induce life-long immunity to the infecting serotype, while heterologous cross-protection against other serotypes lasts only a few weeks, allowing re-infection by another serotype. A number of clinical and experimental data demonstrated the implication of the immune response in the pathogenesis of severe forms of dengue, possibly through an antibody-dependant enhancement (ADE) phenomenon based on the cross-reactivity of DV antibodies [3],[4]. The molecular structure of the ectodomain of E glycoprotein has been determined [5]. It is folded in three distinct domains I, II and III. The C-terminal immunoglobulin-like domain III (EDIII) can be independently folded as a core protein through a single disulfide bond and contains major serotype-specific neutralizing epitopes [6]. On the opposite, epitopes inducing antibodies that cross-react between serotypes have been located within the domain II, which contains the fusion peptide [7]. Therefore, EDIII has emerged as an antigen of choice to develop a dengue vaccine eliciting serotype-specific rather than cross-reactive antibodies. Indeed, recent studies have demonstrated that immunization with EDIII, either encoded by a plasmid or as a recombinant protein in fusion with a bacterial carrier, elicited neutralizing antibodies to DV [8],[9],[10],[11],[12].
A preventive dengue vaccine needs to protect unexposed individuals against all four serotypes of DV. It must be tetravalent, safe for 9–12 months children and provide long-lasting protective immunity. It must be produced at low cost and scaled up at million doses. To address these challenges, we evaluated the immunogenicity of a live recombinant vector derived from pediatric measles vaccine (MV) expressing a DV antigen designed to induce neutralizing and non cross-reactive antibodies. MV vaccine is a live-attenuated negative-stranded RNA virus proven to be one of the safest, most stable, and effective human vaccines developed so far. Produced on a large scale in many countries and distributed at low cost through the Extended Program on Immunization (EPI) of WHO, this vaccine induces life-long immunity after a single injection [13],[14],[15] and boosting is effective. We previously developed a vector derived from the live-attenuated Schwarz strain of MV [16] that expressed stably different proteins from HIV and induced strong and long-term specific humoral and cellular immune responses [17],[18]. Based on this approach a program of clinical trials was initiated in collaboration with an industrial vaccine manufacturer with funding from the EC, to evaluate the safety and immunogenicity in humans of MV encoding an HIV antigen. We also demonstrated that recombinant MV could protect against flaviviruses, since MV expressing the secreted form of the E protein from West Nile virus (WNV) induced sterilizing humoral immunity against WNV in a mouse model [19].
In the present work, we evaluated the immunogenic potential of a MV vector expressing a DV1 soluble antigen composed of the EDIII fused with the ectoM. In a mouse model of MV infection this vector induced serotype-specific, virus-neutralizing antibodies against DV1. Consistent with this observation, we showed that infection of human monocyte-derived dendritic cells (DCs) resulted in up-regulation of co-stimulatory molecules as well as robust secretion of cytokines and chemokines that are identified as playing a pivotal role in establishment of anti-viral immune responses.
Vero (African green monkey kidney) cells were maintained in DMEM-Glutamax (Gibco-BRL) supplemented with 5% heat-inactivated fetal calf serum (FCS, Invitrogen, Frederick, MD). Helper 293-3-46 cells (a gift from M. A. Billeter, Zurich University) used for viral rescue [20] were grown in DMEM/10% FCS and supplemented with 1.2 mg of G418/ml. The human monocytic cell line U937 (ATCC CRL 1593, American Type Culture Collection, Rockville, Md.) was maintained in complete RPMI (Gibco-BRL) supplemented with 10% FCS (Invitrogen), sodium pyruvate, non-essential amino acids, penicillin G (100 IU/ml), and streptomycin (100 µg/ml). Clinical-grade DCs were prepared as described elsewhere [21],[22]. DCs were maintained in AIMV medium containing 500 U/mL GM-CSF (Gentaur, Brussels, Belgium) and 50 ng/mL IL-13 (Peprotech, Tebu-bio, Rocky Hill, NJ).
The plasmid pTM-MVSchw, which contains an infectious MV cDNA corresponding to the anti-genome of the Schwarz MV vaccine strain, has been described elsewhere [16]. The genomic RNA of DV-1 strain FGA/89 [23] (Genbank accession number AF 226687) was extracted from purified virions and reverse transcribed using Titan One-Step RT-PCR kit (Roche Molecular Biochemicals) according to the manufacturer's instructions. The coding sequence for PrM/E (amino acids 1-395) was cloned into pMT/Bip/V5-His A plasmid (kindly provided by Erika Navarro-Sanchez) and used as a template for further cloning. A PCR fragment encoding the EDIII (aa 295-394) from the E protein was amplified by High Fidelity Polymerase (PCR expand High Fidelity, Invitrogen) using the forward primer 1EDIII 5′-AATTAAGATCTAAAGGGATGTCATATGTGATGTG-3′ containing a BglII restriction site (underlined), and the reverse primer 2EDIII 5′-TTAAGCGGCCGCTATCGCTTGAACCAGCTTAGTTTC-3′ containing a NotI restriction site (underlined) and a stop codon (in bold). The sequence encoding the EDIII from the E protein (aa 295-394) linked to the ectodomain of the membrane M protein (aa 1-40) by the original furin-like cleavage site RRDKR, was generated as follows. The FGA/89 EDIII (Genbank accession no. AF 226687) was amplified using the forward primer 1EDIII and the reverse primer 3EDIII 5′-CGGAACGTTTGTCTCGTCGGAACCAGCTTAGTTTCAAAGC-3′ containing the reverse complement sequence of the furin site of the DV ectoM protein (underlined) and in 3′ the reverse complement sequence of the 3′ EDIII end. The PCR product was used as primer and template to amplify by a second PCR the chimeric sequence EDIII-ectoM (Genbank accession no. CS479843) using 1EDIII primer and the reverse primer 5EDIII 5′-TTAAGCGGCCGCTATCATGGGTGTCTCAAAGCCCAAG-3′ that contains a NotI restriction site (underlined) and a stop codon (in bold). The cloned sequences respect the “rule of six”, which stipulates that the number of nucleotides into the MV genome must be a multiple of 6 [24]. A shuttle plasmid (pTRE2-ssCRT) containing the human calreticulin signal sequence was generated by transferring the calreticulin-derived endoplasmic reticulum targeting signal sequence from the pEGFP-RE vector (Clontech) to the pTRE2-Hyg plasmid (Clontech). The DV-1 cDNAs were introduced into pTRE2-ssCRT using BglII/NotI digestion. After sequencing, the 384 bp coding for the EDIII and 516 bp coding for the EDIII-ectoM antigens were inserted into BsiWI/BssHII-digested pTM-MVSchw-ATU2, which contains an additional transcription unit (ATU) between the phosphoprotein (P) and the matrix (M) genes of the Schwarz MV genome [16],[18],[19]. The resulting plasmids were designated as pTM-MVSchw-EDIII and pTM-MVSchw-EDIII-ectoM.
Rescue of recombinant Schwarz MV from the plasmids pTM-MVSchw-EDIII and pTM-MVSchw-EDIII-ectoM was performed as previously described [16] using the helper-cell-based rescue system described by Radecke et al [20] and modified by Parks et al. [25]. The titers of MV-EDIII and MV-EDIII-ectoM were determined by an endpoint limit-dilution assay on Vero cells. The TCID50 was calculated by use of the Kärber method.
The EDIII and EDIII-ectoM PCR products described above were cloned into pMT/Bip/V5-His A plasmid (Invitrogen) between BglII and NotI restriction sites. The clones were validated by sequencing. Drosophila S2 cells (Invitrogen) were transfected by these plasmids using the Calcium Phosphate Transfection Kit (Invitrogen). Transfected cells were selected by adding 25 µg/ml blasticidin. The EDIII and EDIII-ectoM protein production was induced by adding 750 µM CuSO4. Cell culture supernatant was filtered on 0.2 µM filters before concentration on 10,000-MWCo Vivaspin columns (Vivasciences) eluted with PBS. Recombinant proteins were semi-quantified by Western blot using the MAb 9D12 reactive to EDIII from DV [26].
The following DV strains from the collection of Institut Pasteur were used: strain FGA/89 French Guiana for serotype DV1 [27], Jamaica/N.1409 for DV2 [28], PaH881/88 Thailand for DV3 and 63632/76 Burma for DV4. The E. coli strain BL21(DE3) and SB medium have been described [29]. Plasmids pLB11, pLB12, pLB13, pLB14, coding respectively for EDIII from DV serotypes 1, 2, 3, 4 (residues 296-400) with an hexahistidine tag in C-term, under control of the T7 promoter and pelB signal sequence, were constructed by insertion of RT-PCR products obtained with primers specific for EDIII, in plasmid pET20b+ (Novagen) (O. Lisova et al., in preparation). The DV-EDIII-H6 recombinant proteins were produced from these plasmids in E. Coli BL21(DE3). Bacteria were grown at 24°C in SB medium with ampicillin (200 µg/mL) until A600nm = 1.5 to 2.0 and then induced for 2.5 hours with 1 mM IPTG to obtain the expression of the recombinant genes. The purification of DV-EDIII-H6 proteins from bacteria's periplasmic fluid was performed by chromatography on a NiNTA resin (Qiagen, Hilden) and concentration determined by absorbance spectrometry as previously described [30]. The protein fractions were analyzed by SDS-PAGE in reducing conditions. The fractions that were homogeneous at >95%, were pooled, dialyzed against 50 mM Tris-HCl, pH 8.0, 50 mM NaCl, snap frozen in liquid nitrogen, and stored at −80°C.
A synthetic 40-residue long peptide corresponding to the ectoM sequence from FGA/89 strain (SVALAPHVGLGLETRTETWMSSEGAWKQIQKV ETWALRHP) was synthesized with a purity of at least 80% (Genecust, France).
Protein lysates from Vero cells or U937 cells infected with recombinant viruses were fractionated by SDS-PAGE gel electrophoresis and transferred to cellulose membranes (Amersham Pharmacia Biotech). DV1 EDIII (5 ng) produced in drosophila cells was loaded as a positive control. The blots were probed with a murine monoclonal antibody mAb4E11 directed against the E EDIII of DENV1 (Hybridoma cells producing 4E11 raised against the envelope protein of DV1 were kindly provided by Dr Morens [30]). A goat anti-mouse immunoglobulin G (IgG)-horseradish peroxidase (HRP) conjugate (Amersham) was used as a secondary antibody. Peroxidase activity was visualized with an enhanced chemiluminescence detection kit (Pierce).
Cells were recovered by pipetting, centrifuged for 3 min at 1200 rpm, washed once in PBS and resuspended in FITC-labeled annexinV/propidium iodide (PI) according to the manufacturer's instructions (Becton Dickinson, Apoptosis Detection kit). Labeled cells were analyzed by flow cytometry using a FacsCalibur (BD Biosciences, San Diego, CA), with CellQuest software (Becton Dickinson, Lincoln Park, NJ). The percentage of apoptotic cells was determined as the percentage of AnnexinV positive and propidium iodide negative cells.
Immunofluorescence staining was performed on infected cells, as described elsewhere [31]. Cells were probed with mouse anti DV-1 EDIII 4E11 antibody, mouse anti DV-1 HyperImmune Ascitic Fluid [32] or rabbit anti human MHC-II dimer (kindly provided by Neefjed J.) antibodies. Cy3-conjugated goat anti mouse IgG antibody Cy3 conjugated (Jackson Immunoresearch laboratories), FITC-conjugated goat anti-mouse IgG antibody (Chemicon), or and FITC-conjugated goat anti-rabbit IgG antibody (Amersham Pharmacia Biotech) were used as secondary antibodies respectively.
Cell-surface staining was performed at 4°C for 30 minutes using anti-CD86-PE (BD Pharmingen), CD83-APC (BD Pharmingen), CD80-PE-Cy5 (Immunotech, Marseille) in 1% BSA and 3% human serum-PBS. Isotype-matched mAbs were used as negative controls. Labeled cells were analyzed by flow cytometry using a FacsCalibur (BD), with FlowJo software (Tree Star, Ashland, OR).
Supernatants were harvested after 16 h or 24 h of DC incubation with MV-EDIII or MV-EDIII-ectoM at an MOI of 1. Aliquots of 200–300 µl were stored at −80°C. Production of cytokines/chemokines was analyzed in 50 µL supernatant with a human cytokine 25-plex antibody bead kit (Biosource, CA, WA, cat. LHC0009) which measures IL-1α, IL-1Ra, IL-2, IL-2R, IL-4, IL-5, IL-6, IL-7, IL-8, IL-10, IL-12p40/70, IL-13, IL-15, IL-17, TNF-α, IFN-α, IFN-β, GM-CSF, MIP-1α, MIP-1β, IP-10, MIG, Eotaxin, RANTES and MCP-1 by using a Luminex 100 instrument (Luminex Corp., Austin, TX, USA).
CD46-IFNAR mice susceptible to MV infection were produced as previously described [16]. Mice were housed under specific pathogen-free conditions at the Pasteur Institute animal facility. Six-week-old CD46-IFNAR mice were inoculated intraperitoneally (i.p.) with 104 or 105 TCID50 of recombinant MV. Boosting was performed using 10 µg of recombinant EDIII-ectoM protein in Alugel adjuvant. To detect the anamnestic response generated by immunization, immunized mice were i.p. inoculated with 107 FFU of live FGA/NA d1d variant of DV1 (Genebank accession number AF 226686). This strain was previously generated by adaptation of a clinical isolate to growth in newborn mouse brain [23]. All experiments were approved and conducted in accordance with the guidelines of the Office of Laboratory Animal Care at Pasteur Institute.
To evaluate the specific antibody responses, mice were bled via the periorbital route at different time after inoculation. Sera were heat inactivated at 56°C for 30 min and the presence of anti-MV antibodies was detected by ELISA (Trinity Biotech). HRP-conjugated anti-mouse immunoglobulin (Jackson Immuno Research) was used as secondary antibody. Anti-DV antibodies were detected by ELISA using 96-wells plates coated with either highly purified FGA/89 DV1 particles, recombinant EDIII proteins from DV1, DV2, DV3, DV4 produced in E. Coli. or synthetic ectoM peptide. HRP-conjugated anti-mouse immunoglobulin was used as secondary antibody. The endpoint titers of pooled sera were calculated as the reciprocal of the last dilution giving twice the absorbance of sera from MV inoculated mice that served as negative controls.
Anti-DV neutralizing antibodies were detected by a focus reduction neutralization test (FRNT) on Vero cells previously described [19] using 50 FFU of Vero-adapted DV1 Hawaï (WHO reference strain, Genbank accession no. AF226687), DV2 Jamaica (Genbank accession no. M20558), DV3 H97 (WHO reference strain, Genbank accession no. M93130) or DV4 63632. The endpoint titer was calculated as the highest serum dilution tested that reduced the number of FFU by at least 50% (FRNT50) or 75% (FRNT75). For the neutralization tests in presence of recombinant EDIII or synthetic peptides ectoM peptides serum samples were pre-incubated (in 50 µl medium) with 5 µg, 500 ng or 50 ng of recombinant EDIII (produced in Drosophila S2 cells) or synthetic ectoM peptide before performing FRNT.
We cloned from DV1 viral RNA the sequence encoding the EDIII (E295-394) fused in C-term to the ectoM (M1-40), using as a linker the original prM/M furin-like cleavage site RRDKR. This combined dengue antigen was cloned downstream the cellular calreticulin (ss-CRT) signal peptide sequence in order to allow the disulfide bond formation and therefore the correct folding of EDIII and to address the antigen in the secretion pathway. As a control of proper folding and effective secretion of EDIII, we also generated a similar construct without ectoM. The resulting ss-CRT-EDIII-ectoM and ss-CRT-EDIII constructs were inserted into MV vector (pTM-MVSchw plasmid), which contains an infectious MV cDNA corresponding to the anti-genome of the Schwarz MV vaccine strain [16] (Figure 1A). The recombinant measles viruses MV-EDIII-ectoM and MV-EDIII were rescued by transfecting the pTM-MVSchw-EDIII-ectoM and pTM-MVSchw-EDIII plasmids into helper cells and propagation on Vero cells, as previously described [16].
We analyzed the expression of DV antigens by recombinant MV in infected Vero cells by immunofluorescence using a monoclonal neutralizing anti-DV1 EDIII antibody (4E11, [33]) and anti-DV1 Hyper Immune Ascitic Fluid (HMAF) (Figure 1B). In both cases, the antigens were clearly detected indicating that the EDIII was expressed and that the epitope of 4E11 neutralizing antibody was present and accessible. The presence of EDIII in lysates and supernatants of infected Vero cells was further confirmed by Western blot using 4E11 antibody (Figure 1C). DV1 recombinant EDIII polypeptides (rEDIII and rEDIII-ectoM) secreted from stable S2 cell lines were used as positive controls (Figure 1D). DV1 EDIII antigen was detected both in lysates and supernatants from cells infected by MV-EDIII-ectoM and MV-EDIII vectors. The EDIII was clearly detected in unconcentrated supernatant, despite that the supernatant volume was 100 times larger than the lysate volume. Thus, secretion was efficient. The intracellular EDIII shows a higher molecular weight on the western blot than EDIII secreted in the supernatants, because of the presence of the peptide signal, which was cleaved during secretion. Similarly, the positive control (rEDIII from S2 cells) has a higher molecular weight because it contains a poly-histidine tag. Cells infected by MV-EDIII-ectoM produced cleaved EDIII and uncleaved EDIII-ectoM antigens both in cell lysates and medium, indicating that the furin-like cleavage site was accessible. Taken together, these data show that MV-EDIII-ectoM vector is able to produce secreted forms of EDIII, EDIII-ectoM, and by assumption, ectoM (not detected because of the lack of specific antibodies to ectoM).
We analyzed the replication of MV-EDIII-ectoM and MV-EDIII viruses on Vero cells using the same MOI (0.01) than for MV production. The growth kinetics of both recombinant viruses were similar to that of control MV and the final titer was slightly higher for MV-EDIII-ectoM (Figure 2A). We then investigated whether the presence of ectoM in MV-EDIII-ectoM virus could increase apoptosis of infected cells [2]. Since human monocytes constitute a determinant target of MV infection for initiation of immune responses, we addressed this question by infecting human monocytic cells (U937 leukemic monocyte lymphoma cell line). Growth kinetics of MV, MV-EDIII and MV-EDIII-ectoM in U937 cells (MOI 1) show that these cells are permissive to MV infection and that MV-EDIII-40 growth was slightly delayed as compared to MV and MV-EDIII (Figure 2B). We quantified apoptotic cells (annexin V positive/propidium iodide negative cells) after infection at different MOI 0.1, 1 and 10. While we did not observe apoptosis up to 42 hours post-infection when using MOIs of 0.1 or 1 with the 3 viruses, increasing the MOI to 10 with MV-EDIII-ectoM virus eventually induced apoptosis in 15% of cells (Figure 2C). Apoptosis was related to the activation of caspase 3 pathway (not shown) and was dependent on virus replication, since UV inactivated virus did not trigger apoptosis.
We examined the ability of MV-EDIII-ectoM recombinant virus to raise specific anti-DV1 neutralizing antibodies in genetically modified mice susceptible to MV infection [34]. These mice express CD46, the human receptor for vaccine MV strains, and lack the INF-α/β receptor (IFNAR) [16],[18],[35],[36]. They have previously been used as a model to evaluate the immunogenicity of recombinant MV [16],[17],[18],[19],[36]. Six-week-old CD46-IFNAR mice received two intraperitoneal (ip) injections within one month of either 104 or 105 TCID50 of MV-EDIII-ectoM. As a control, CD46-IFNAR mice were immunized with MV-EDIII and empty MV vector. Specific antibody responses were analyzed by ELISA one month after the second injection (Table 1). All immunized mice raised antibodies to MV at similar titers. Specific anti-DV1 and anti-rEDIII antibodies were mounted in mice immunized with MV-EDIII-ectoM (titers 3,000 and 10,000 respectively). Surprisingly, no anti-DV antibodies were detected in sera of mice immunized with MV-EDIII. An ELISA test using ectoM as coated viral antigen showed that immune sera had no detectable levels of anti-M antibodies. We evaluated the anti-DV1 neutralizing activity of immune sera by using a focus reduction neutralization test (FRNT) that allows to determine the highest serum dilutions able to reduce by at least 50% or 75% the number of DV1 focus forming units (FFU) on Vero cells. Again, whereas immunization with MV or MV-EDIII did not induce neutralizing antibodies to DV1, immunization by MV-EDIII-ectoM raised FRNT50 titers to 320 and FRNT75 titers to 40 (Table 1). Altogether, these data show that EDIII-ectoM is able to elicit humoral immunity to DV.
Although CD46-IFNAR mice used in this study did not allow to evaluate the protection conferred by immunization, we tested the ability of live DV1 peripheral inoculation to stimulate long-term anamnestic humoral response against MV-EDIII-ectoM. To assess first the susceptibility of CD46-IFNAR mice to DV1 replication, we inoculated mice intraperitoneally with 107 FFU of DV1 strain FGA/NA d1d. No symptoms or mortality were observed and virus replication could not be detected in mice serum by direct plaque assay on mosquito cells (not shown). However, all groups of mice seroconverted after live DV1 inoculation (Table 1). Interestingly, the anamnestic memory induced by immunization with recombinant MV-EDIII-ectoM was remarkably boosted by live DV1 inoculation, whereas animals immunized with MV-EDIII or empty MV did not show any boost (Table 1). Both ELISA and FRNT neutralizing titers against DV1 were strongly increased in mice immunized with MV-EDIII-ectoM (30–100 times increase), showing evidence of an efficient anamnestic response.
To test the longevity of this memory, another group of mice was primed with two injections of MV-EDIII-ectoM vector (Table 2). Six months later, they were boosted by injecting 5 µg of adjuvanted rEDIII-ectoM protein purified from supernatants of transfected drosophila S2 cells. Protein boost increased the neutralizing titer from 10 to 200. However, the titer decreased rapidly to 40, indicating a transient boost. As a control, mice inoculated only with the recombinant protein remained negative even after three injections (not shown). At 9 months post priming, mice were i.p. inoculated with 107 FFU of live FGA/NA d1d DV1. One month after DV1 inoculation, we again observed a 100 times increase in the level of antibodies to DV1 and to EDIII as well as DV1 neutralizing titers (Table 2). This experiment shows that neutralizing antibodies to DV EDIII induced by MV-EDIII-ectoM are efficiently boosted upon live DV exposure 9 months after priming, and demonstrates the induction of a durable B-cell memory. However, the protection against DV infection needs to be evaluated in a more appropriate non-human primate model.
To assess the DV serotype specificity of antibodies induced, we tested mice sera from experiment presented in table 2 by ELISA against rEDIII proteins from DV1, 2, 3 and 4, respectively. We also evaluated the anti-DV1, anti-DV2, anti-DV3 and anti-DV4 neutralizing activity of immune sera by FRNT50. Antibodies induced by recombinant MV-EDIII-ectoM were specific to DV1 and did not cross-react with the EDIII from the other serotypes of DV (Table 3). This confirms that EDIII antigenic surface is serotype specific. To determine whether the antiviral neutralizing activity induced by MV-EDIII-ectoM was specifically directed against the EDIII, even after live DV1 inoculation, we performed neutralization tests in presence of increasing concentrations of either rEDIII protein produced in drosophila cells or synthetic ectoM peptide (Figure 3). Increasing concentrations of rEDIII protein strongly reduced the antiviral activity of sera collected before and after live DV1 inoculation, whereas ectoM peptide was ineffective. This experiment demonstrates that the neutralizing antibodies induced by immunization were specifically directed against EDIII epitopes essential for virus infectivity. Moreover, the antiviral activity of sera collected after live DV1 inoculation was also completely inhibited by rEDIII protein, indicating that live DV1 inoculation increased the level of antibodies already induced by immunization, but did not raise new antibodies directed to other neutralizing epitopes than EDIII.
Altogether, these experiments demonstrate that recombinant MV-EDIII-ectoM virus induced specific antibodies to DV1 EDIII that did not cross-react with other DV serotypes and that neutralized specifically DV1 infection. The EDIII alone expressed by recombinant MV was poorly immunogenic, although it was expressed and secreted at levels similar to EDIII-ectoM. The presence of the ectoM was determinant to its immunogenicity, raising the question of the mechanism of this effect. Does the presence of ectoM in C-term of the EDIII sequence improve the conformation of EDIII to make it biologically active, or adjuvant its immunogenicity through an indirect mechanism? The EDIII secreted from cells infected by MV-EDIII was recognized by neutralizing antibody 4E11, which binds to the active receptor-binding form of EDIII [33] The same EDIII sequence secreted by drosophila cells was able to efficiently compete with the neutralizing activity of antibodies induced by the MV-EDIII-ectoM virus, thus suggesting that EDIII was biologically functional, at least able to present neutralizing epitopes. To address the question of adjuvantation, we compared in vitro the effect of MV-EDIII-ectoM and MV-EDIII infection on human immature monocyte-derived DCs (MDDCs) in terms of activation/maturation and cytokine/chemokine secretion.
Upon virus infection, immature dendritic cells (DCs) undergo maturation, and transport the virus to regional lymph nodes, where viral antigens are presented to lymphocytes to initiate immune response [37]. DCs are permissive to MV infection [38] leading to the up-regulation of co-stimulatory molecules [39],[40],[41]. To evaluate in vitro the effect of recombinant MV-DV on DC activation, we cultivated human immature monocyte-derived DCs (MDDCs) in presence of MV-EDIII-ectoM or MV-EDIII viruses. After 17 hours of infection, DCs expressed the DV1 EDIII as detected by immunofluorescence (Figure 4). We then analyzed the kinetics and levels of expression of costimulatory molecules on the surface of DCs. The three viruses MV-EDIII-ectoM, MV-EDIII and MV promoted the up-regulation of CD86, CD83 and CD80 molecules as compared to mock-treated DCs (Figure 5). Remarkably, MV-EDIII-ectoM up-regulated these molecules more extensively and at earlier time points than MV-EDIII and MV. Viral replication and de novo synthesis of EDIII-ectoM were required since UV-inactivated recombinant viruses or synthetic ectoM peptide had no effect (not shown). To determine if the increased capacity of MV-EDIII-ectoM to activate DCs correlated with phenotypic functional changes, we analyzed the secretion of 23 cytokines/chemokines in the supernatant of DCs cultivated for 16h and 24h in presence of MV-EDIII-ectoM or MV-EDIII. As a control, DCs were mock-infected or cultivated in presence of LPS. We found that infection with MV-EDIII or MV-EDIII-ectoM induced a panel of cytokines and chemokines consistent with other reports on DC infection by MV (Table 4) [42]. Remarkably, some cytokines and chemokines were significantly enhanced and/or induced at earlier time points by MV-EDIII-ectoM as compared to MV-EDIII (Table 4, Figure 6). Among them, IFN-α (1000 pg/ml at 16 h), IL1RA (750 pg/m), IL4 (13 pg/m) and the proinflammatory cytokines IL-6 (1250 pg/m) and TNF-α (1700 pg/m) were induced much more rapidly after MV-EDIII-ectoM infection and at levels 8–10 times higher than after MV-EDIII infection (Figure 6). Such a strong enhancement in production of these cytokines is expected to accelerate the establishment of immune response and to favor humoral immunity. Similarly, remarkable levels of MIP-1α chemokine were induced by MV-EDIII-ectoM at early time points (9,000–20,000 pg/ml). A similar robust and early secretion was observed for RANTES, MIP-1β and MCP-1α. The production of these chemokines by maturing DCs may promote the recruitment of other antigen presenting cells (APC) such as immature DCs to enhance and sustain antigen sampling, and polarization of the immune response [43],[44],[45]. Thus, adding the ectoM protein in fusion with EDIII resulted in a stronger and faster maturation of human DCs and activated the secretion of higher levels of inflammatory and antiviral cytokines as well as chemokines determinant for the establishment of specific immune responses.
The objective of this study was to evaluate the immunogenicity of a dengue vaccine candidate based on a pediatric measles vaccine expressing a minimal dengue antigen. This strategy provides a recombinant vaccine that might protect children simultaneously from measles and dengue and that might be affordable to populations through the EPI program in the regions affected both by dengue and measles infections. An efficient pediatric dengue vaccine is supposed to elicit durable protective humoral immune responses against all four dengue serotypes without risk of ADE [46]. Regarding this objective, we assembled covalently the antigenic domain III from the DV1 envelope E glycoprotein and the pro-apoptotic ectodomain of DV-1 M protein to generate a dengue combined antigen, EDIII-ectoM. In the fusion construct, the N-terminal calreticulin peptide signal sequence directs EDIII-ectoM to the secretory pathway. The furin-dependent cleavage site of prM/M which links ectoM to EDIII allows the processing of the antigen by specific proteases throughout the Golgi apparatus. Expressed by recombinant MV vector, the EDIII-ectoM antigen induced in mice susceptible to MV specific antibodies to DV1 EDIII that did not cross-react with other DV serotypes and that neutralized DV1 infection in vitro. Immunization primed a long-term memory that was vigorously boosted when animals were inoculated with live DV.
Although DV disease pathogenesis and protection mechanisms are not fully clarified, disease severity is correlated with viremia levels and neutralizing antibody is generally used as a marker of vaccine effectiveness [47]. Experimental mouse models of DV infection have been reported showing that adult AG129 mice, which are deficient for IFN α/β/γ receptors develop a dose-dependant transient viremia after peripheral injection of unadapted or mouse-adapted DV, whereas A129-IFNAR mice, which are deficient only for IFN-α/β receptor are less sensitive to DV infection [48],[49],[50]. However, AG129 mice are not sensitive to MV infection and the prototype DV1 Hawaï strain did not replicate in these mice [50]. Suckling mice develop lethal encephalitis after DV intracerebral inoculation, but in our study mice were 3–4 month-old after two MV immunizations and intracranial inoculation could not be performed. Moreover, this model is far from the human situation since DV does not infect the nervous system, nor lead to encephalitis in humans. The CD46-IFNAR mouse model sensitive to MV infection that we used did not allow documenting in vivo protection from DV replication. Therefore, to demonstrate the induction of anamnestic neutralizing antibody response upon live DV exposure, mice were peripherally inoculated with DV a long time after immunization. These experiments showed that neutralizing antibodies induced by immunization with MV-EDIII-ectoM were strongly boosted by live DV inoculation, thus suggesting a protective capacity. Indeed, the available vaccines against yellow fever, Japanese encephalitis and tick-borne encephalitis viruses have proven that anamnestic neutralizing antibodies play an essential role for protection against flaviviral infections [47].
The EDIII without ectoM was poorly immunogenic in the context of MV expression. It appeared, therefore, that the 40-residue long ectodomain of M plays a critical role in its immunogenicity. DV EDIII has been previously shown to be immunogenic in the form of recombinant chimeric proteins [8],[9],[10],[12] or expressed from a plasmid [51] or from adenovirus vector [52],[53]. To determine whether the EDIII sequence inserted into MV vector was able to present neutralizing epitopes, we produced in E.coli recombinant EDIII proteins from DV1, 2, 3 and 4 corresponding to the same sequence and we coated plates with these proteins. Tested by ELISA on these plates, a neutralizing HMAF specific to DV1 recognized specifically the DV1-EDIII, but not the other serotypes (data not shown), indicating its specificity. This DV-1 HMAF recognized also specifically by immunofluorescence the DV1 EDIII expressed in cells infected by MV-EDIII, indicating the capacity of EDIII to expose serotype-specific epitopes. The neutralizing monoclonal antibody 4E11 recognized also the EDIII expressed by MV-EDIII infected cells, indicating that the epitope specific of this antibody is accessible within the EDIII expressed by MV. This epitope has been mapped and shown to be exposed on the native form of EDIII [33]. Furthermore, we expressed the same EDIII sequence as a secreted protein by drosophila cells and showed that it was able to efficiently compete with the neutralizing activity of antibodies induced by the MV-EDIII-ectoM virus. Altogether, these observations suggest that EDIII expressed by MV was able to present a conformationally active neutralizing epitope. Recent studies evaluating the immunogenicity of West-Nile virus (WNV) EDIII showed that a high amount of EDIII was necessary to induce neutralizing antibodies, while EDIII fused to TLR ligands was immunogenic and conferred protection at lower doses [54],[55]. Therefore, the low immunogenicity of DV EDIII in our hands might be due to the lower amount of antigen expressed by recombinant MV as compared to the high protein or DNA doses administered by others. To increase the level of expression by MV, EDIII can be cloned upstream the N gene, as MV genes are expressed as a gradient from the 3′ to the 5′ end of the genome.
This small ectoM protein, which is highly conserved among the four serotypes of DV, has pro-apoptotic properties [2]. High titers of MV-EDIII-ectoM induced apoptosis of infected U937 monocyte-like cells that was not observed at standard titers. This critical point in terms of safety needs to be evaluated further in the development of this vaccine candidate. Indeed, recombinant MV vector has to keep the high safety level of standard MV vaccine. However, this property might be determinant to the immunogenicity of EDIII because apoptotic infected cells express Toll-like receptor (TLR) ligands that increase the cross-presentation of viral epitopes by antigen presenting cells [56],[57]. Indeed, we observed that ectoM, in the context of MV replication, increased human DCs maturation and triggered the release of cytokines and chemokines determinant for the establishment of specific adaptive immunity. Therefore, its capacity to adjuvant EDIII might be still more efficient in humans than in CD46-IFNAR mice. Further studies are needed to address the mechanism of action at the molecular level.
In conclusion, we have produced a minimal antigen from DV1 able to induce long-term specific neutralizing antibodies to DV1 with no cross-reactivity with other serotypes. We have shown that the remarkable adjuvant capacity of ectoM to EDIII immunogenicity was correlated to its capacity to mature primary DCs and to activate the secretion of a panel of proinflammatory and antiviral cytokines, as well as numerous chemokines determinant for the establishment of specific adaptive immunity. The immunogenicity of this antigen was demonstrated through its expression by a recombinant MV vector, thus making the proof-of-concept of this strategy for dengue vaccine development. Using MV as a vaccination vector presents a number of advantages : vaccination against measles is mandatory, vaccine strains are genetically stable, MV does not recombine or integrate genetic material, and vaccine does not persist or diffuse. MV-specific CD8 T cells and IgG are detected in vaccinees up to 25–34 years after a single MV vaccination [14] and boosting increases this memory [15]. Using MV as a recombinant vaccine to immunize simultaneously against measles and dengue might be particularly attractive in areas where both diseases threaten children every year, such as Africa and South America. Taking advantage of the capacity of MV vector to express large amounts of heterologous genetic material very stably [58], we generated tetravalent dengue antigenic constructs inserted into single MV vectors that are currently characterized. Such a strategy should avoid the stability and interference problems encountered with tetravalent formulation of four attenuated viruses, as well as the reactogenicity problems [59]. These new candidates will be evaluated in a much more appropriate non-human primate model. |
10.1371/journal.pcbi.1005536 | A predictive coding account of bistable perception - a model-based fMRI study | In bistable vision, subjective perception wavers between two interpretations of a constant ambiguous stimulus. This dissociation between conscious perception and sensory stimulation has motivated various empirical studies on the neural correlates of bistable perception, but the neurocomputational mechanism behind endogenous perceptual transitions has remained elusive. Here, we recurred to a generic Bayesian framework of predictive coding and devised a model that casts endogenous perceptual transitions as a consequence of prediction errors emerging from residual evidence for the suppressed percept. Data simulations revealed close similarities between the model’s predictions and key temporal characteristics of perceptual bistability, indicating that the model was able to reproduce bistable perception. Fitting the predictive coding model to behavioural data from an fMRI-experiment on bistable perception, we found a correlation across participants between the model parameter encoding perceptual stabilization and the behaviourally measured frequency of perceptual transitions, corroborating that the model successfully accounted for participants’ perception. Formal model comparison with established models of bistable perception based on mutual inhibition and adaptation, noise or a combination of adaptation and noise was used for the validation of the predictive coding model against the established models. Most importantly, model-based analyses of the fMRI data revealed that prediction error time-courses derived from the predictive coding model correlated with neural signal time-courses in bilateral inferior frontal gyri and anterior insulae. Voxel-wise model selection indicated a superiority of the predictive coding model over conventional analysis approaches in explaining neural activity in these frontal areas, suggesting that frontal cortex encodes prediction errors that mediate endogenous perceptual transitions in bistable perception. Taken together, our current work provides a theoretical framework that allows for the analysis of behavioural and neural data using a predictive coding perspective on bistable perception. In this, our approach posits a crucial role of prediction error signalling for the resolution of perceptual ambiguities.
| In bistable vision, perception spontaneously alternates between two different interpretations of a constant ambiguous stimulus. Here, we show that such spontaneous perceptual transitions can be parsimoniously described by a Bayesian predictive coding model. Using simulated, behavioural and fMRI data, we provide evidence that prediction errors stemming from the suppressed stimulus interpretation mediate perceptual transitions and correlate with neural activity in inferior frontal gyrus and insula. Our findings empirically corroborate theorizations on the relevance of prediction errors for spontaneous perceptual transitions and substantially contribute to a longstanding debate on the role of frontal activity in bistable vision. Therefore, our current work fundamentally advances our mechanistic understanding of perceptual inference in the human brain.
| During bistable perception, observers experience fluctuations between two mutually exclusive interpretations of a constant ambiguous input. Remarkably, percepts evoked by ambiguous stimuli usually closely resemble the experience of unambiguous objects and thus illustrate the constructive nature of perception. However, the mechanisms driving transitions in bistable perception remain poorly understood.
Previous neuroimaging work [4, 5, 6, 7, 8, 9, 10] has sought to distill the neural processes underlying bistable perception by recurring to a ‘replay’ condition, in which physical stimulus changes mimic the perceptual alternations induced by ambiguous stimuli. This approach revealed a right-lateralized assembly of fronto-parietal areas whose activity is specifically enhanced during endogenously evoked transitions (ambiguity) as compared to exogenously evoked transitions (replay) [4, 5, 7, 9].
However, the functional role of fronto-parietal areas in bistable perception is a matter of ongoing debate. According to one view, transitions in bistable vision are primarily a result of adaptation and inhibition within visual cortex, while switch-related activations in fronto-parietal areas reflect a mere ‘feedforward’ consequence of neural events at sensory processing levels [6, 10]. Another view proposes that fronto-parietal areas may be involved in stabilizing and destabilizing perception, thus causally contributing to perceptual switching via ‘feedback’ mechanisms [4, 5, 11, 7]. Here, we sought to resolve this debate by using model-based fMRI to empirically test a theoretical model that has the potential to integrate these two seemingly contradictory views of perceptual bistability.
From a theoretical perspective, endogenous transitions might be explained by framing perception as an inferential process generating and testing hypotheses about the most likely causes of sensory stimulation [12, 13, 14]. Such processes can be elegantly implemented by hierarchical predictive coding [15, 16, 17]. Here, ‘predictions’ encoded at higher levels are compared against ‘sensory input’ represented at lower levels, while a mismatch between the two elicits a prediction error, updating higher-level predictions [15]. Such belief-updating schemes can be translated onto Bayes’ rule, where prior distributions (‘predictions’) are combined with likelihood distributions (’sensory input’) into posterior distributions in a sequential manner [16, 18].
Here, we tested whether this framework provides a mechanistic explanation for perceptual transitions and related neural activity during bistable perception. We devised a computational model that formalizes perceptual decisions (i.e., decisions that define the content of conscious perception, as indicated by participants’ response) to be performed on the basis of posterior probability distributions. This model is a modification of an approach introduced by [19], who propose that perceptual time-courses during bistable perception result from samples drawn subsequently from a posterior distribution. The authors implement a memory decay favoring recent over older samples as well as stationary prior capturing the effect of context on bistable perception. Our model, in turn, posits that the shape of the posterior distribution changes dynamically over time in response to prediction errors emerging from the currently suppressed interpretation of the ambiguous input. Importantly, this model has the potential to integrate feedforward and feedback mechanisms in bistable perception: The prediction errors arising from sensory processing levels may be propagated up to higher-level brain areas in a feedforward fashion. The registration of prediction errors in higher-level brain areas leads to an updating of predictions that may in turn drive perceptual switching through a feedback mechanism.
To test this hypothesis, we began with data simulations to establish that our model’s predictions match the key characteristics of perceptual bistability. We proceeded by fitting our model to behavioural data from a fMRI experiment on bistable perception [7].
In this experiment, participants viewed a Lissajous figure [42] rotating either clockwise (as viewed from above, i.e. movement of the front surface to the left) or counter-clockwise (vice versa) and indicated their current perception via button-presses. Participants were presented with alternating blocks of ambiguous and disambiguated Lissajous figures: In the ambiguous condition, we presented bistable Lissajous figures which elicited spontaneous (endogenous) alternations in perception. In the disambiguated (’replay’) condition, we mimicked the endogenous perceptual time-course by introducing exogenous perceptual switches. Ambiguous and disambiguated stimuli were constructed by presenting two Lissajous figures separately to the two eyes: In the ambiguous condition, both eyes received identical stimulation. In the replay condition, the two Lissajous figures were slightly phase-shifted against each other, biasing perception in the direction of the phase shift.
Having inverted our predictive coding approach based on behavioural data from this experiment, we investigated whether our model accurately explains individual perceptual time-courses during ambiguous and replay stimulation.
In a supplementary analysis (see S2 Text), we furthermore compared our model to three established models of bistable perception: Firstly, we tested an oscillator model [1], which is based on mutual inhibition between to competing neural populations coding for the alternative perceptual outcomes during bistable perception. Here, the currently dominant population suppresses activity in the alternative population. However, due to adaptation in the dominant population, this relation reverses over time, leading to regular oscillations in perception. Secondly, we constructed a noise-driven attractor model of bistable perception [2]. In this framework, internal and external sources of noise trigger transitions between two stable states in an attractor network, representing the two perceptual interpretations associated with a bistable stimulus. Thirdly, we tested an intermediate model [3], which contains both adaptive processes and noise. We validated our approach against these models by the use of Bayesian Model Comparison [20].
We then conducted a model-based fMRI-analysis [21] based on the predictive coding model to test whether prediction errors account for transition-related neural activity during bistability. Additionally, we compared the model-based fMRI analysis with conventional fMRI analyses using a Posterior Probability Map (PPM) approach [22].
Our Bayesian modelling approach draws on the view that perception is an inferential process in which perceptual decisions are based on posterior distributions [13]. According to Bayes’ rule, the posterior combines information in the current sensory data (likelihood) with information from previous visual experience (prior) in a probabilistically optimal manner. Crucially, this posterior at a given moment becomes a prior for the current perceptual decision, which entails a prediction error signal that influences on the prior at the next moment. Hence, the posterior not only provides the basis for current perception, but also shapes future perception.
In line with previous theorizations [12], we reasoned that the ambiguous likelihood provides equally strong sensory evidence for two different percepts. We further hypothesized that the current percept establishes an implicit prior belief about similar percepts in the future, thereby contributing to stability of visual perception. The application of Bayes’ rules combines the likelihood for ambiguous stimuli with the stability prior into a posterior that represents stronger evidence for the dominant percept, but still contains residual evidence for the suppressed percept. While the stronger evidence for the dominant percept will again favor this percept for the upcoming perceptual decision, the residual evidence for the suppressed percept is equivalent to a prediction error that leads to an update of the stability prior.
Over time, the stability prior is weakened and the posterior shifts towards the suppressed percept, paralleled by an escalating prediction error. When the residual evidence for the suppressed percept equals the evidence for the dominant percept, the prediction error reaches a maximum and a perceptual transition is most likely to occur. Once such a transition has occurred, the process starts over again, minimizing the current prediction error.
Please note that our approach was influenced by the work of [19], who argue that bistable perception is a product of Bayesian decision making in ambiguous sensory environments. They study the effects of viewpoint context on perception of the Necker Cube and propose that bistable perception arises from sampling a bimodal posterior distribution. Here, the sample with the highest ‚weight’ determines the content of conscious perception. Key elements of their model are (1), a stationary prior, whose precision reflects interindividual differences in the effects of viewpoint context on perception of the Necker Cube and (2), a memory decay that discounts the weight associated with a sample drawn from the posterior distribution by its age and influences on the length of individual phase durations.
In contrast to [19], our model does not assume a specific memory decay process, but controls the length of phase durations by means of the dynamically updated stability prior. In analogy to the stationary viewpoint prior in [19], our model captures the influence of additional sensory evidence on perceptual decisions using a ‚stereodisparity’ distribution, whose precision determines the effectiveness of disambiguation.
Please refer to to the mathematical appendix (see S1 Text) for a complete description, to Fig 1 for a step-by-step illustration of our approach and to Table 1 for a summary of model parameters and quantities. For computational expediency, we assume Gaussian probability distributions defined by mean and variance (or inverse precision).
To test whether our model is able to reproduce the temporal dynamics of bistable perception, we used it to generate perceptual time-courses from some ambiguous visual input such as the Lissajous figure. We assumed a sampling rate of 0.33 Hz, which was chosen to be close to the average overlap frequency in the behavioural experiment (see below), and simulated for a total of 6 * 105 seconds. To model the ambiguous visual input, the impact of the stereodisparity weight was suppressed by setting μstereo = 0.5 and πstereo = 0. We further assumed fixed values for the precision πinit, which was set to 3.5 to match the posterior parameter value from our behavioural modelling (see Modelling analysis of behavioural data).
To examine whether our prediction error model might account for bistable perception and associated neural activity in human observers, we used data from an fMRI experiment applying the Lissajous figure. Results from conventional analyses but not from behavioural modelling or model-based fMRI (see below) have been reported previously [7].
We recorded BOLD images by T2-weighted gradient-echo echo-planar imaging (FOV 192, 33 slices, TR 2000 ms, TE 30 ms, flip angle 78°, voxel size 3 x 3 x 3 mm, interslice gap 10 percent) on a 3T MRI scanner (Tim Trio, Siemens). The number of volumes amounted to 402 (0.15 Hz and 0.2 Hz) or 415 (0.12 Hz) volumes, respectively. We used a T1-weighted MPRAGE sequence (FOV 256, 160 slices, TR 1900 ms, TE 2.52 ms, flip angle 9°, voxel size 1 x 1 x 1 mm) to acquire anatomical images.
Image preprocessing (standard realignment, coregistration, normalization to MNI stereotactic space using unified segmentation, spatial smoothing with 8 mm full-width at half-maximum isotropic Gaussian kernel) was carried out with SPM8 (http://www.fil.ion.ucl.ac.uk/spm/software/spm8).
To probe whether our predictive coding model might explain perceptual time-courses during bistable perception in human observers, we fitted our model to the behavioural data collected during the fMRI experiment. We optimized our model for the prediction of perceptual outcomes, i.e. on the perception of clockwise or counter-clockwise rotation as indicated by the individual participants. To this end, participants’ responses were aligned to the overlapping stimulus configurations of the Lissajous figure (’overlaps’). This refers to timepoints during presentation when fore- and background of the stimulus cannot be discerned (i.e. depth-symmetry) [25, 26]. Depending on the rotational speed of the stimulus and the associated ‘overlap’ frequency, sampling rates varied across participants between 0.24 Hz and 0.40 Hz (see above). We first constructed models incorporating all combinations of the likelihood weight ‘stereodisparity’ and prior ‘perceptual stability’, yielding a total of 4 behavioural models (behavioural model 1: no stereodisparity, no perceptual stability; behavioural model 2: no stereodisparity, perceptual stability; behavioural model 3: stereodisparity, no perceptual stability; behavioural model 4: stereodisparity, perceptual stability) to be compared. The respective precision of these distributions was optimized for the prediction of perceptual outcomes based on posterior distributions using a free energy minimization approach [27]. This method minimises the surprise about the individual participants’ data, thereby maximising log-model evidence.
For model inversion, precisions were modelled as log-normal distributions. πinit and πstereo were either estimated as free parameters (πinit: prior mean of log(3) and prior variance of 5; πstereo: prior mean of log(5) and prior variance of 5) or fixed to zero (thereby effectively removing the distribution from the model). We kept ζ, which represents the inverse decision temperature in the response model represented by Equation 11 (see Mathematical Appendix, S1 Text), fixed to 1, since we did not have a particular a-priori hypothesis regarding this parameter. Please note that when choosing ζ as a free parameter (prior mean of log(1), prior variance of 1), results remained almost identical. Parameters were optimised using quasi-Newton Broyden-Fletcher-Goldfarb-Shanno minimisation as implemented in the HGF4.0 toolbox (TAPAS toolbox, http://www.translationalneuromodeling.org/hgf-toolbox-v3-0/).
After identifying the optimal model using Random Effects Bayesian model selection [20], as implemented in SPM12 (http://www.fil.ion.ucl.ac.uk/spm/software/spm12/), we analyzed its posterior parameters with regard to the respective precision of the prior distributions using classical frequentist statistics. Since parameters were estimated in log-space, we report the geometric mean (i.e. the arithmetic mean in log-space).
In a supplementary analysis (see S2 Text), we further compared the explanatory power of our predictive coding model with established models of bistable perception. To this end, we implemented models of bistable perception belonging to three different classes ([1] as an example of so-called oscillator models based on mutual inhibition and self-adaptation between two competing neuronal populations, [2] as a representative of noise-driven attractor models and [3] as an intermediate model), which can be fitted to experimental data. We conducted a Random Effects Bayesian Model Comparison [20] between the established models and our predictive coding model in order to probe the validity of our approach.
To examine the neural correlates of prediction error time-courses from our model, we conducted model-based fMRI analyses [21] in SPM12. We adopted a general-linear-model-(GLM-) approach, constructing a total of three models:
The design matrix of the first GLM (the ‘PE model’) represented prediction error trajectories timepoint by timepoint. To this end, the regressor ‘transitions’ and the regressor ‘overlaps’ were modelled as stick functions. Furthermore, we extracted the individual ‘Prediction Error’ time-course for every participant and run and used its absolute value as a parametric modulator for the regressor ‘overlaps’.
In order to enable a comparison to the conventional approach of analysing fMRI data on bistable perception, we constructed a second GLM that dissociated between transition-related activity specific to bistable perception and the replay condition [4, 5, 6, 7, 9, 10]. In addition to the regressor ‘overlaps’, the design matrix of this ‘Conventional model’ contained ambiguous and replay transitions represented by stick functions.
To further investigate the specificity of the prediction error trajectories and their neural correlates, we constructed a third GLM that took into account the presence of ambiguity inherent to the bistable condition. The design matrix contained the regressors ‘transitions’ as well the regressor ‘overlaps’ modelled as stick functions. Here, however, we used a box-car function being 1 for ambiguous and 0 for ‘replay’ blocks as a parametric modulator of the regressor ‘overlaps’. Hence, this ‘Block model’ only differs from the ‘PE model’ in the values of the parametric modulator and serves to investigate whether correlations with the prediction error (which we assumed to be higher in the bistable condition) merely correspond to ambiguity per se.
All further analyses were conducted for all models in parallel: regressors were convolved with the canonical hemodynamic response function as implemented in SPM12. We added six rigid-body realignment parameters as nuisance covariates and applied high-pass filtering at 1/128 Hz.
In a first step, we tested which of the three models accounted best for the measured BOLD signal. Therefore, we conducted a voxel-wise model comparison of the ‘PE model’ with the ‘Conventional model’ and the ‘Block model’, as described in [22]. In brief, this technique uses Bayesian statistics for the construction of ‘Posterior Probability Maps’ (PPMs) and ‘Exceedance Probability Maps’ (EPMs), which enable the calculation of log-evidence maps for each participant and model separately. On a second level, these log-evidence maps can be combined, thereby enabling voxel-wise model inference at the group level. Using the ‘Bayesian 1st level’ procedure for model estimation, we constructed log-evidence maps for every participant and model separately and compared the ‘PE model’ to the other models on a group level using exceedance probabilities computed with Random Effects analyses.
In a second step, we aimed to identify regions in which prediction error trajectories (‘PE model’), ambiguity per se (‘Block model’) or ambiguous as compared to replay transitions (‘Conventional model’) were correlated with the recorded BOLD signals. To this end, we estimated single-participant statistical parametric maps, then created contrast images for the parametric regressor against baseline (‘PE model’ and ‘Block model’) or ambiguous against replay transitions (‘Conventional model’). These were entered into voxel-wise one-sample t-tests at the group level. Voxels were considered statistically significant if they survived family-wise-error (FWE) correction for all voxels in the brain at p < 0.05. Anatomic labeling of cluster peaks was performed using the SPM Anatomy Toolbox Version 1.7b [28].
In order to further visualize our results, we extracted eigenvariate time-courses (without adjustment for effects of interest) from spherical ROIs (radius: 3 mm) around peak voxels from clusters for the contrast ‘Prediction Error vs baseline’ (thresholded at p < 0.05) corresponding to left IFG (peak voxel: [-54 2 22]), right IFG (peak voxel: [51 8 10]), left insula (peak voxel: [-30 20 10]) and right insula (peak voxel: [33 23 7]). These time-courses were extracted for ambiguous stimulation only. The time-courses for all perceptual phases were aligned with the respect to the end of the perceptual phase and averaged within and across observers.
To test whether our predictive coding model was able to reproduce perceptual switching in bistable perception, we used the model to generate perceptual time-courses during simulated viewing of an ambiguous stimulus.
The distribution of perceptual phase durations followed a sharp rise and slow fall (Fig 2) typical for bistable stimuli [29, 30]. Mean and median simulated phase durations were 10.40 and 10.00 seconds, closely matching the results from behavioural analysis (see Modelling analysis of behavioural data). As illustrated by exemplary time-courses of model parameters, the prediction error PE (Fig 2A) increases over time while one percept is dominant and is reduced once a new percept is adopted, reflecting the accumulation of evidence from the suppressed percept. The variance (1/πstability) of the prior ‘perceptual stability’ (Fig 2C) increases over a perceptual phase as a function of the prediction error. In line with the hypothesized role of prediction errors in driving perceptual transitions, the prediction error PE and, hence, the variance 1/πstability are maximal when the posterior P(θ > 0.5) relaxes to 0.5 (Fig 2B), thereby increasing the probability of a new perceptual transition.
To investigate whether our model is able to explain the dynamics of perceptual bistability in human observers, we fitted our model to behavioural data collected from 20 healthy participants during an fMRI experiment, in which participants viewed ambiguous and unambiguous (replay) versions of a rotating Lissajous stimulus. As reported previously, perceptual transitions occurred on average every 9.3 seconds in the ambiguity condition and neither block-by-block ratings nor debriefing after the experiment revealed differences in perceived appearance between the ambiguity and the replay condition [7].
We first performed a model comparison with other models that lacked the key conceptual elements of our model. By eliminating either the likelihood weight ‘stereodisparity’ or the prior ‘perceptual stability’ or both from the model, we constructed three additional models which we compared to our model using Random Effects Bayesian Model Selection. Our model (i.e. behavioural model 4) was identified as a clear winning model with a protected exceedance probability of 99.96%, demonstrating that the incorporation of both the likelihood weight ‘stereodisparity’ and the prior ‘perceptual stability’ best explained participants’ perception.
From this model, we extracted the parameters for πinit and πstereo and averaged across runs and participants (Fig 3A). We predicted average prediction errors to be lower in replay as compared to the ambiguous condition, since the presented stereodisparity reduces the ambiguity left in the experimental display, and hence, the residual evidence for suppressed percept. Consistently, mean prediction errors were significantly higher in the ambiguous condition than in the replay condition (0.36 +/- 0.03 vs. 0.26 +/- 0.02, mean +/- s.e.m., p < 10−6, t19 = 7.06, two-sample t-test, Fig 3B), providing support for a correct implementation of our predictive coding model.
Given that πinit describes the strength of the initial stabilization after a switch in perception, we expected this parameter to be related to the frequency of perceptual transitions. In line with this, model parameter estimates πinit were negatively correlated with perceptual transition frequencies across participants (ρ = −0.88, p < 10−7, Pearson correlation, Fig 3C), providing a sanity check for model fit. Notably, this correlation was also significant when we correlated model parameter estimates for πinit averaged over run 1 and 2 with perceptual transition frequencies from run 3 (ρ = −0.76, p = 10−4, Fig 3D), corroborating that our model successfully accounted for observers’ perception evoked by an ambiguous stimulus.
We furthermore validated our approach by comparing our predictive coding model to established models of bistable perception from three different classes: oscillator models [1], attractor models [2] and intermediate models [3] (see Supplementary Methods in S2 Text). Data simulations indicated that all established models, similar to our predictive coding model, were able to produce spontaneous transitions in perception and a typical gamma-like distribution of perceptual phase durations (see Supplementary Results and Fig. A-C in S2 Text). Fitting of the behavioural data further showed that both the oscillator and the intermediate, similar to our predictive coding model, adequately accounted for the observers’ perceptual decisions during bistable perception (see Supplementary Results and Fig. D-I in S2 Text). In order to validate our approach, we conducted a Bayesian Model Comparison, which showed that our predictive coding model compared to these established models was best in explaining the behavioural data collected during this experiment (see Fig. J in S2 Text).
Please note that we did not carry out these analysis to demonstrate a superiority of our approach over these earlier models, which were initially conceived mainly for binocular rivalry and not for the prediction of behavioural responses during presentation of the Lissajous figure (a specific type of structure-from-motion stimulus). On the contrary, we aimed at probing the validity of our approach and tried to ascertain that the predictive coding approach was at least equivalent to existing models of bistable perception.
One central aim of our study was to gain mechanistic insight into the neural processes underlying transition-related activity during bistable perception. We therefore performed both a model-based fMRI analysis suitable to identify the neural correlates of modelled prediction errors (‘PE model’), and, for the purpose of comparison, a conventional analysis (‘Conventional model’) dissociating between ambiguous and replay transitions as well as a ‘Block model’ accounting for effects of ambiguity per se.
To test the validity of these models, we first searched for voxels that were more active during visual stimulation as compared to baseline (‘overlaps vs. baseline’). For the ‘PE model’, this analysis revealed significant clusters (p < 0.05, FWE-corrected across the whole brain) bilaterally in middle occipital cortex (right: [39 -9 1], T = 10.21; left: [-30 -94 1], T = 13.30), in V5/hMT+ (right: [45 -70 1], T = 11.61; left: [-45 -73 4], T = 14.09), as well as in superior parietal cortex (right: [27 -49 58], T = 10.26; left: [-36 -46 -61], T = 8.62). The same analyses for the ‘Conventional model’ and the ‘Block model’ yielded virtually identical results (see Tables 2–4), confirming the comparability between all three models.
We then investigated which voxels were more active during perceptual transitions as compared to baseline (‘transitions vs. baseline’, Fig 4A): For the ‘PE model’, we found significant activations of motor-related areas in left precentral gyrus ([-36 -16 67], T = 12.23) extending to left postcentral gyrus ([-63 -19 25], T = 8.62) as well as significant clusters in regions previously associated with transition-related activity during bistable perception: right inferior frontal gyrus ([54 17 13], T = 7.96), right inferior parietal lobulus (54 -37 52, T = 9.32) and right middle frontal gyrus ([39 44 31], T = 7.57). Additional clusters were located in bilateral posterior-medial frontal gyrus (right: [6 2 67], T = 9.50; left: [-6 2 55], T = 12.63). Again, repeating this analysis for the ‘Block model’ and the ‘Conventional model’ yielded qualitatively very similar results as in the ‘PE model’ (see Tables 5–7), thereby providing further evidence for the validity and comparability of all three models.
To formally test whether the modelled prediction error explains the BOLD signal better than the conventional comparison of ambiguous with replay perceptual switches (‘Conventional model’), or the mere ambiguity of the visual display (the ‘Block model’), we performed a PPM analysis [22] to compute voxel-wise exceedance probability maps for the ‘PE model’ against the ‘Conventional model’ and the ‘Block model’ (Fig 4C). We restricted this analysis to areas of the fronto-parietal cortex, which be delineated by intersecting the statistical-parametric maps for ‘transitions vs. baseline’ thresholded at p < 0.05 FWE for all three models considered. Remarkably, when applying a conservative threshold of an exceedance probability of γ = 99% and a cluster size of n > 10 voxels, we found clusters in right insula ([39 26 -2]) and right inferior frontal gyrus ([51 14 1]) to show strong evidence for the ‘PE model’ as compared to the ‘Block model’ and the ‘Conventional model’. Additional clusters were located in right posterior medial frontal gyrus ([6 5 49]) as well as left precentral gyrus ([-36 -16 52]).
Conversely, for the exceedance probability map of the ‘Conventional model’ compared against ‘Block model’ and ‘PE model’, no voxels survived the conservative threshold used in the main experiment. For the exceedance probability map of the ‘Block model’ compared against the ‘Conventional model’ and ‘PE model’, we found clusters in bilateral inferior parietal lobule at an exceedance probability of 99% and a cluster size > 10.
For our central analysis aimed at identifying the neural correlates of modelled prediction errors, we searched for voxels in which BOLD activity was related to the parametric modulator of the ‘PE model’ that encoded prediction error trajectories from our Bayesian model of bistable perception (Fig 4B). We found significant clusters (p < 0.05, FWE-corrected across the whole brain) in bilateral insulae (right: [33 23 7], T = 7.24; left: [-30 20 10], T = 7.88) and bilateral inferior frontal gyri (right: [51 8 10], T = 6.89; left: [- 54 2 22], T = 6.67). These regions are located in close anatomical proximity to frontal regions previously suggested to mediate perceptual transitions in bistable perception [4, 5, 7].
In order to further visualize the correlation between modelled prediction error and BOLD activity, we extracted eigenvariate time-courses from right insula, left insula, right IFG as well as left IFG and averaged across perceptual phase durations and observers. As expected, these time-courses showed a gradual increase towards a transition in perception (Fig 5), nicely mirroring the build-up of prediction error during a perceptual phase.
In this work, we present a Bayesian predictive coding model for bistable perception, which rests on the basic assumption that prediction errors are elicited by the unexplained alternative interpretation of an ambiguous stimulus and represent the driving force behind perceptual transitions during bistable perception. We found that this model is able to reproduce key temporal characteristics of human bistable perception and that it explains observers’ behaviour during a bistable perception experiment. Our central finding shows that modelled prediction errors correlate with BOLD activity in bilateral insulae and bilateral inferior frontal gyrus. Remarkably, our PPM analysis revealed that modelled prediction errors best accounted for BOLD activity as compared to mere occurrence of endogenous perceptual transitions or ambiguity of the visual display in these frontal regions. Hence, our current results suggest that prediction errors might provide the mechanistic basis for perceptual switching in bistable perception and offer a novel interpretation of frontal activity in bilateral insulae as well as the right inferior frontal gyrus during bistable perception.
The functional significance of enhanced frontal brain activity for transitions during bistability as compared to an unambiguous control condition is a matter of ongoing debate: Some authors proposed that non-sensory higher-level brain regions are actively implicated in resolving the perceptual conflict during bistable perception, thus mediating perceptual transitions [4, 31, 5, 11, 7]. Others have argued that perceptual conflicts are resolved primarily in sensory brain areas and that activity in frontal and parietal regions reflects the registration and/or report of perceptual transitions, rather than their cause [6, 8, 10]. For a detailed discussion of this debate, see “Brascamp, Sterzer, Blake and Knapen, Multistable perception and the role of frontoparietal cortex in perceptual inference, Annual Review of Psychology, in press.”
Here, we provide further evidence for an active implication of frontal regions in bistable perception by functionally relating these regions to a prediction error signal. Hence, our work is in line with hybrid models that suggest bistable perception to arise from an interplay between lower-level sensory and higher-level non-sensory areas [32, 12, 11]. In this context, it might be speculated that prediction errors are computed in frontal regions based on feedforward signals from visual and parietal cortex; and that these prediction errors, in turn, modulate activity in visual cortex via feedback signals.
In addition to the prediction error, the stability prior represents an essential element of our predictive coding model of bistable perception, since its initial precision determines the frequency of perceptual transitions. The notion of such a stability prior is supported by experimental work on serial dependence in visual perception: In an orientation-judgement task, [33] showed that perceived orientation was biased by recently observed stimuli and reasoned that the visual system might use past experiences as predictors of present perceptual decisions, thereby incorporating representations of the continuity of the visual environment. Corroborating these results in a fMRI experiment, [34] found that orientation signals in early visual cortex were biased towards previous perceptual decisions. At this point, however, we can only speculate about the neural correlates of the stability prior from our model: In recent work on the role of parietal cortex in bistable perception, [35] and [9] have proposed a functional segregation of the superior parietal lobulus (SPL), which they deduced from differential effects of grey matter volume on perceptual dominance durations and analyses of effective connectivity on the basis of fMRI. By interpreting their results in a Bayesian framework, the authors argued that posterior SPL might represent a prediction error, while the anterior SPL would entertain a perceptual prediction.
A key advantage of our predictive coding model of bistable perception is that it allows us to treat ambiguous and replay stimulation within the same framework. By formalizing the disambiguating factor as a weight on a bimodal likelihood distribution, such models can be used to investigate perceptual transitions under varying degrees of ambiguity, thus dissolving the artificial dichotomy between the two conditions. Hence, such models provide a new perspective on how the brain might resolve perceptual conflicts despite the ambiguity inherent in every sensory signal and offer a generic tool for quantifying the contribution of different contextual factors on perceptual outcomes.
The major strength of predictive coding models for bistable perception, however, lies in their ability to parsimoniously link different levels in the description of perceptual dynamics in ambiguous visual environments: On a computational level, prediction errors constitute the driving force behind perceptual transitions and are substantially reduced by additional sensory information (such as stereodisparity) during replay. On a neural level, casting frontal activity during rivalry in terms of prediction error signals nicely relates to increased transition-related activity [4, 5, 9] and connectivity [7]. On a theoretical level, viewing perceptual transitions as means of reducing prediction errors places bistable perception in the context of Bayesian theories of the brain [16, 36, 27, 37], and in particular the free-energy principle [13]. According to the latter, agents strive for a reduction of their model’s free energy, which translates onto a minimization of squared prediction errors in predictive coding schemes. When sensory information is constantly ambiguous, one possibility to reduce free energy is to update beliefs about the world, which ultimately corresponds to the adoption of a new percept.
However, given that the Lissajous differs in some aspects from other types of bistable stimuli, one has to consider important limitations regarding the generalization of our findings: While being physically ambiguous for all angles of rotation, transitions almost exclusively occur at overlapping stimulus configurations, which is similar to the behaviour of some types of random dot kinematograms [26] or intermittent presentation of bistable stimuli [38] and accompanied by a reduced incidence of mixed percepts or incomplete transitions. Since these phenomena are present in many other forms of bistable perception and significantly affect frontoparietal activity during perceptual transitions [6], our current imaging results can only be interpreted in relation to the specific stimulus used here.
A similar limitation applies to the behavioural modelling presented in this manuscript: Previous work on computational modelling of bistable perception has focused on a variety of mechanisms at the heart of spontaneous perceptual transitions: Oscillator models have focused on mutual inhibition between two competing neuronal populations combined with slow adaptation of the currently dominant population [1]. [39] have studied the differential effects of short and long interruptions in intermittent bistable perception for binocular rivalry and structure-from-motion and presented a model based on adaptive processes, cross-inhibition and neural baseline levels. Importantly, this model also accounts for the possibility of voluntary control via attentional processes interacting with early processing stages.
Alternative approaches view noise as the underlying cause of perceptual transitions [2]. Importantly, models belonging to this class have also taken account of the aforementioned mixed percepts and incomplete transitions during binocular rivalry [40].
Further models have related transitions in perception to a combination of adaptation and noise [3]. In this vein, [41] have argued for a neurodynamic mechanism at the bifurcation between adaptation- and noise-driven processes to be the basis for perceptual transitions during binocular rivalry.
The majority of the models mentioned above has been developed for continuous presentation of binocular rivalry or ambiguous structure-from-motion, while [39] have also studied paradigms with intermittent presentation. As noted above, such stimuli differ significantly from the Lissajous figure used in our current study, which shares aspects with intermittent stimulation due to the existence of overlapping configurations facilitating transitions in perception. Hence, future theoretical and empirical work is needed to probe our modelling approach on paradigms such as binocular rivalry and ambiguous structure-from-motion for both continuous and intermittent presentation and to extend the predictive coding model in order to account for top-down attentional control as well as interactions at earlier processing stages.
Taken together, our current work provides theoretical and empirical evidence across different levels for a driving role of prediction errors in bistable perception, thereby shedding new light on an ongoing debate about the neural mechanisms underlying bistable perception and, more generally, opening up a novel computational perspective on the mechanisms governing perceptual inference.
|
10.1371/journal.pcbi.1004842 | Efficient Coalescent Simulation and Genealogical Analysis for Large Sample Sizes | A central challenge in the analysis of genetic variation is to provide realistic genome simulation across millions of samples. Present day coalescent simulations do not scale well, or use approximations that fail to capture important long-range linkage properties. Analysing the results of simulations also presents a substantial challenge, as current methods to store genealogies consume a great deal of space, are slow to parse and do not take advantage of shared structure in correlated trees. We solve these problems by introducing sparse trees and coalescence records as the key units of genealogical analysis. Using these tools, exact simulation of the coalescent with recombination for chromosome-sized regions over hundreds of thousands of samples is possible, and substantially faster than present-day approximate methods. We can also analyse the results orders of magnitude more quickly than with existing methods.
| Our understanding of the distribution of genetic variation in natural populations has been driven by mathematical models of the underlying biological and demographic processes. A key strength of such coalescent models is that they enable efficient simulation of data we might see under a variety of evolutionary scenarios. However, current methods are not well suited to simulating genome-scale data sets on hundreds of thousands of samples, which is essential if we are to understand the data generated by population-scale sequencing projects. Similarly, processing the results of large simulations also presents researchers with a major challenge, as it can take many days just to read the data files. In this paper we solve these problems by introducing a new way to represent information about the ancestral process. This new representation leads to huge gains in simulation speed and storage efficiency so that large simulations complete in minutes and the output files can be processed in seconds.
| The coalescent process [1, 2] underlies much of modern population genetics and is fundamental to our understanding of molecular evolution. The coalescent describes the ancestry of a sample of n genes in the absence of recombination, selection, population structure and other complicating factors. The model has proved to be highly extensible, and these and many other complexities required to model real populations have successfully been incorporated [3]. Simulation has played a key role in coalescent theory since its beginnings [2], partly due to the ease with which it can be simulated: for a sample of n genes, we require only O(n) time and space to simulate a genealogy [4].
Soon after the single locus coalescent was derived, Hudson defined an algorithm to simulate the coalescent with recombination [5]. However, after some early successes in characterising this process [6, 7] little progress was made because of the complex distribution of blocks of ancestral material among ancestors. Some years after Hudson’s pioneering work, the study of recombination in the coalescent was recast in the framework of the Ancestral Recombination Graph [8, 9]. In the ARG, nodes are events (either recombination or common ancestor) and the edges are ancestral chromosomes. A recombination event results in a single ancestral chromosome splitting into two chromosomes, and a common ancestor event results in two chromosomes merging into a common ancestor. Analytically, the ARG is a considerable simplification of Hudson’s earlier work as it models all recombination events that occurred in the history of a sample and not just those that can potentially affect the genealogies. Many important results have been derived using this framework, one of which is particularly significant for our purposes here. Ethier and Griffiths [10] proved that the expected number of recombination events back to the Grand MRCA of a sample of n individuals grows like eρ as ρ → ∞, where ρ is the population scaled recombination rate. In this paper we consider a diploid model in which we have a sequence of m discrete sites that are indexed from zero. Recombination occurs between adjacent sites at rate r per generation, and therefore ρ = 4Ne r(m − 1). The Ethier and Griffiths result implies that the time required to simulate an ARG grows exponentially with the sequence length, and we can only ever hope to simulate ARGs for the shortest of sequences.
This result, coupled with the observed poor scaling of coalescent simulators such as the seminal ms program [11] seems to imply that simulating the coalescent with recombination over chromosome scales is hopeless, and researchers have therefore sought alternatives. The sequentially Markov coalescent (SMC) approximation [12, 13] underlies the majority of present day genome scale simulation [14–16] and inference methods [17–19]. The SMC simplifies the process of simulating genealogies by assuming that each marginal tree depends only on its immediate predecessor as we move from left-to-right across the sequence. As a consequence, the time required to simulate genealogies scales linearly with increasing sequence length. In practice, SMC based simulators such as MaCS [14] and scrm [16] are many times faster than ms.
The SMC has disadvantages, however. Firstly, the SMC discards all long range linkage information and therefore can be a poor approximation when modelling features such as the length of admixture blocks [20]. Improving the accuracy of the SMC can also be difficult. For example, MaCS has a parameter to increase the number of previous trees on which a marginal tree can depend. Counter-intuitively, increasing this parameter beyond a certain limit results in a worse approximation to the coalescent with recombination [16]. (The scrm simulator provides a similar parameter that does not exhibit this unfortunate behaviour, however.) Incorporating complexities such as population structure [21], intra-codon recombination [22] and inversions [23] is non-trivial and can be substantially more complex than the corresponding modification to the exact coalescent model. Also, while SMC based methods scale well in terms of increasing sequence length, currently available simulators do not scale well in terms of sample size.
We solve these problems by introducing sparse trees and coalescence records as the fundamental units of genealogical analysis. By creating a concrete formalisation of the genealogies generated by the coalescent process in terms of an integer vector, we greatly increase the efficiency of simulating the exact coalescent with recombination. In the section Efficient coalescent simulation, we discuss how Hudson’s classical simulation algorithm can be defined in terms of these sparse trees, and why this leads to substantial gains in terms of the simulation speed and memory usage. We show that our implementation of the exact coalescent, msprime, is competitive with approximate simulators for small sample sizes, and is faster than all other simulators for large sample sizes. This is possible because Hudson’s algorithm does not traverse the entire ARG, but rather a small subset of it. The ARG contains a large number of nodes that do not affect the genealogies of the sample [24], and Hudson’s algorithm saves time by not visiting these nodes. This subset of the ARG (sometimes known as the ‘little’ ARG) has not been well characterised, which makes analysis of Hudson’s algorithm difficult. However, we show some numerical results indicating that the number of nodes in the little ARG may be a quadratic function of the scaled recombination rate ρ rather than an exponential.
Generating simulated data is of little use if the results cannot be processed in an efficient and convenient manner. Currently available methods for storing and processing genealogies perform very poorly on trees with hundreds of thousands of nodes. In the section Efficient genealogical analysis, we show how the encoding of the correlated trees output by our simulations leads to an extremely compact method of storing these genealogies. For large simulations, the representation can be thousands of times smaller than the most compact tree serialisation format currently available. Our encoding also leads to very efficient tree processing algorithms; for example, sequential access to trees is several orders of magnitude faster than existing methods.
The advantages of faster and more accurate simulation over huge sample sizes, and the ability to quickly process very large result sets may enable applications that were not previously feasible. In the Results and Discussion we conclude by considering some of these applications and other uses of our novel encoding of genealogies. The methods developed in this paper allow us to simulate the coalescent for very large sample sizes, where the underlying assumptions of the model may be violated [25–27]. Addressing these issues is beyond the scope of this work, but we note that the majority of our results can be applied to simulations of any retrospective population model.
In this section we define our encoding of coalescent genealogies, and show how this leads to very efficient simulations. There are many different simulation packages, and so we begin with a brief review of the state-of-the-art before defining our encoding and analysing the resulting algorithm in the following subsections.
Two basic approaches exist to simulate the coalescent with recombination. The first approach was defined by Hudson [5], and works by applying the effects of recombination and common ancestor events to the ancestors of the sample as we go backwards in time. Events occur at a rate that depends only on the state of the extant ancestors, and so we can generate the waiting times to these events efficiently without considering the intervening generations. This contrasts with time-reversed generation-by-generation methods [28–31] which are more flexible but also considerably less efficient. The first simulation program published based on Hudson’s algorithm was ms [11]. After this, many programs were published to simulate various evolutionary complexities not handled by ms, such as selection [32–35], recombination hotspots [36], codon models [37], intra-codon recombination [22] and models of species with a skewed offspring distribution [38]. Others developed user interfaces to facilitate easier analysis [39, 40].
The second fundamental method of simulating the coalescent with recombination is due to Wiuf and Hein [24]. In Wiuf and Hein’s algorithm we begin by generating a coalescent tree for the left-most locus and then move across the sequence, updating the genealogy to account for recombination events. This process is considerably more complex than Hudson’s algorithm because the relationship between trees as we move across the genome is non-Markovian: each tree depends on all previously generated trees. Because of this complexity, exact simulators based on Wiuf and Hein’s algorithm are significantly less efficient than ms [16, 41]. However, Wiuf and Hein’s algorithm has provided the basis for the SMC approximation [12, 13], and programs based on this approach [14–16] can simulate long sequences far more efficiently than exact methods such as ms. Very roughly, we can think of Wiuf and Hein’s algorithm performing a depth-first traversal of the ARG, and Hudson’s algorithm a breadth-first traversal. Neither explore the full ARG, but instead traverse the subset required to contruct all marginal genealogies.
Recently, Hudson’s algorithm has been utilised in cosi2 [35], which takes a novel approach to simulating sequences under the coalescent. The majority of simulators first generate genealogies and then throw down mutations in a separate process. In cosi2 these two processes are merged, so that mutations are generated during traversal of the ARG. Instead of associating a partial genealogy with each ancestral segment, cosi2 maps ancestral segments directly to the set of sampled individuals at the leaves of this tree. When a coalescence between two overlapping segments occurs, we then have sufficient information to generate mutations and map them to the affected samples. This strategy, coupled with the use of sophisticated data structures, makes cosi2 many times faster than competing simulators such as msms [34]. The disadvantage of combining the mutation process with ARG traversal, however, is that the underlying genealogies are not available, and cosi2 cannot directly output coalescent trees.
Many reviews are available to compare the various coalescent simulators in terms of their features [42–47]. Little information is available, however, about their relative efficiencies. Hudson’s ms is widely regarded as the most efficient implementation of the exact coalescent and is the benchmark against which other programs are measured [13–16, 41, 47]. However, for larger sample sizes and long sequence lengths, msms is much faster than ms. Also, for these larger sequence lengths and sample sizes, ms is unreliable and crashes [15, 47]. Thus, msms is a much more suitable baseline against which to judge performance. The scrm simulator is the most efficient SMC based method currently available [16].
There has been much recent interest in the problem of representing large scale genetic data in formats that facilitate efficient access and calculation of statistics [53–55]. The use of ‘succinct’ data structures, which are highly compressed but also allow for efficient queries is becoming essential: the scale of the data available to researchers is so large that naive methods simply no longer work.
Although genealogies are fundamental to biology, there has been little attention to the problem of encoding trees in a form that facilitates efficient computation. The majority of research has focused on the accurate interchange of tree structures and associated metadata. The most common format for exchanging tree data is the Newick format [56], which although ill-defined [57] has become the de-facto standard. Newick is based on the correspondence of tree structures with nested parentheses, and is a concise method of expressing tree topologies. Because of this recursive structure, specific extensions to the syntax are required to associate information with tree nodes [58, 59]. XML based formats [57, 60] are much more flexible, but tend to require substantially more storage space than Newick [57]. Various extensions to Newick have been proposed to incorporate more general graph structures [61–64], as well as a GraphML extension to encode ARGs directly [65]. Because Newick stores branch lengths rather than node times, numerical precision issues also arise when summing over many short branches [65].
General purpose Bioinformatics toolkits such as BioPerl [66] and BioPython [67] provide basic tools to import trees in the various formats. More specific tree processing libraries such as DendroPy [68], ETE [69], and APE [70] provide more sophisticated tools such as visualisation and tree comparison algorithms. None of these libraries are designed to handle large collections of correlated trees, and cannot make use of the shared structure within a sequence of correlated genealogies. The methods employed rarely scale well to trees containing hundreds of thousands of nodes.
In this section we introduce a new representation of the correlated trees output by a coalescent simulation using coalescence records. In the Tree sequences subsection we discuss this structure and show how it compares in practice to existing approaches in terms of storage size. Then, the Generating trees subsection presents an algorithm to sequentially generate the marginal genealogies from a tree sequence, which we compare with existing Newick-based methods. Finally, in the Counting leaves subsection we show how the algorithm to sequentially visit trees can be extended to efficiently maintain the counts of leaves from a specific subset, and show how this can be applied in a calculation commonly used in genome wide association studies.
The primary contribution of this paper is to introduce a new encoding for the correlated trees resulting from simulations of the coalescent with recombination. This encoding follows on from previous work in which trees are encoded as integer vectors [49, 50], but makes the crucial change that tree vectors are sparse. Using this encoding, the effects of each coalescence event are stored as simple fixed-size records that provide sufficient information to recover all marginal genealogies after the simulation has completed. This approach leads to very large gains in simulation performance over classical simulators such as ms, so that the exact simulation of genealogies for the coalescent with recombination over chromosome scales is feasible for the first time. We have presented an implementation based on the sparse tree encoding called msprime, which is faster than all other simulators for large sample sizes. This simulator supports the full discrete population structure and demographic event model provided by ms along with variable recombination rates. We plan to include populations evolving in continuous space [86–88] and gene conversion [89] in subsequent releases.
Coalescence records also lead to an extremely compact storage format that is several orders of magnitude smaller than the most compact method currently available. Despite this very high level of compression, accessing the genealogical data is very efficient. In an example with 100,000 samples, we saw a roughly 40,000-fold reduction in file size over the Newick tree encoding, and a greater than million-fold decrease in the time required to iterate over the genealogies compared to several popular libraries. This efficiency is gained through very simple algorithms that we have stated rigorously and unambiguously, and also analysed in terms of their computational complexity. Being able to process such large sample sizes is not an idle curiosity; on the contrary, we have a pressing need to work with such datasets. We envisage three immediate uses for our work.
Firstly, sequencing projects are being conducted on an unprecedented scale [90–95], and the storage and analysis of these data pose serious computational challenges. Sophisticated new methods are being developed to organise and analyse information on this immense scale [53–55]. Developers have struggled to generate simulated data on a similar scale [53, 54], as present day simulators perform poorly on these huge sample sizes. Using msprime, the time required to generate genome scale data for hundreds of thousands of samples is reduced from weeks to minutes.
Secondly, prospective studies such as UK Biobank [96, 97] are collecting genetic and high-dimensional phenotypic data for hundreds of thousands of samples. The key statistical method to interrogate such data is the genome wide association study (GWAS) [98], and large sample size has been identified as the single most important factor in determining the power of these studies [83]. Simulation plays a key role in GWAS, and typically proceeds by superimposing the disease model of interest on haplotypes obtained via various methods [99]. Because the accurate modelling of linkage disequilibrium is essential in disease genetics [100], recombination must be incorporated. Resampling methods [83, 101–103] generate simulated haplotypes based on an existing reference panel, and provide a good match to observed linkage patterns. However, there is some bias associated with this process, and there are statistical difficulties when the size of the sample required is larger than the reference panel. Other methods obtain simulated haplotypes from population genetics models via forwards-in-time [104, 105] or coalescent [106, 107] simulations. None of these methods can efficiently handle the huge sample sizes required, however. A simulator for high dimensional phenotype data based on msprime could alleviate these performance issues and be a key application for the library.
Thirdly, today’s large sample sizes provide us with an unprecedented opportunity to understand the history and geographic structure of our species. Aside from its intrinsic interest, correctly accounting for population stratification is critical for the interpretation of association studies [108, 109], particularly for rare variants [110, 111]. Researchers are seeking to understand fine scale population structure using methods based on principal component analysis [112], admixture fractions [113–115], length of haplotype blocks [116–118] and allele frequencies [119]. To date, it has been challenging to assess the accuracy of these methods, as simulations struggle to match the required sequence lengths and sample sizes. Furthermore, methods based on the SMC approximation [17, 18] have been tested using SMC simulations out of necessity, making it difficult to assess the impact of the approximation on accuracy. Simulations of the exact coalescent with recombination at chromosome scales for large sample sizes and arbitrary demographies will be an invaluable tool for developers of such methods.
As we have demonstrated, the tree sequence structure leads to very efficient algorithms, and allows us to encode simulated data very compactly. We would also wish to encode biological data in this structure so that we can apply these algorithms to analyse real data. However, to do this we must estimate a tree sequence from data, which is a non-trivial task. Nonetheless, there has been much work in this area [120] with several heuristic [121] and more principled approaches that may be adopted [19, 122]. Using the PBWT [53] to find long haplotypes (which will usually correspond to long records) seems like a particularly promising avenue.
Finally, an interesting issue arises when we consider the problem of inferring a tree sequence from data. Suppose we have observed a set of haplotypes resulting from a coalescent simulation with infinite sites mutations occurring at a very high rate. Under these conditions, the underlying tree sequence can be recovered exactly from the data, but the corresponding ARG (i.e., the specific realisation of the ARG that was traversed by Hudson’s algorithm) cannot. For example, a recombination may have occurred during the simulation that was immediately followed by a common ancestor event involving the same lineages. These nodes in the ARG can have no effect on the data, and are therefore unobservable. To put this in another way, there is no observable information in an ARG that is not in a tree sequence. Given this representational sufficiency and the storage and processing efficiencies demonstrated in this article, we would argue that a tree sequence is a more natural and powerful representation of observed genetic variation than an ARG.
|
10.1371/journal.pgen.1008371 | RNAi-mediated depletion of the NSL complex subunits leads to abnormal chromosome segregation and defective centrosome duplication in Drosophila mitosis | The Drosophila Nonspecific Lethal (NSL) complex is a major transcriptional regulator of housekeeping genes. It contains at least seven subunits that are conserved in the human KANSL complex: Nsl1/Wah (KANSL1), Dgt1/Nsl2 (KANSL2), Rcd1/Nsl3 (KANSL3), Rcd5 (MCRS1), MBD-R2 (PHF20), Wds (WDR5) and Mof (MOF/KAT8). Previous studies have shown that Dgt1, Rcd1 and Rcd5 are implicated in centrosome maintenance. Here, we analyzed the mitotic phenotypes caused by RNAi-mediated depletion of Rcd1, Rcd5, MBD-R2 or Wds in greater detail. Depletion of any of these proteins in Drosophila S2 cells led to defects in chromosome segregation. Consistent with these findings, Rcd1, Rcd5 and MBD-R2 RNAi cells showed reduced levels of both Cid/CENP-A and the kinetochore component Ndc80. In addition, RNAi against any of the four genes negatively affected centriole duplication. In Wds-depleted cells, the mitotic phenotypes were similar but milder than those observed in Rcd1-, Rcd5- or MBD-R2-deficient cells. RT-qPCR experiments and interrogation of published datasets revealed that transcription of many genes encoding centromere/kinetochore proteins (e.g., cid, Mis12 and Nnf1b), or involved in centriole duplication (e.g., Sas-6, Sas-4 and asl) is substantially reduced in Rcd1, Rcd5 and MBD-R2 RNAi cells, and to a lesser extent in wds RNAi cells. During mitosis, both Rcd1-GFP and Rcd5-GFP accumulate at the centrosomes and the telophase midbody, MBD-R2-GFP is enriched only at the chromosomes, while Wds-GFP accumulates at the centrosomes, the kinetochores, the midbody, and on a specific chromosome region. Collectively, our results suggest that the mitotic phenotypes caused by Rcd1, Rcd5, MBD-R2 or Wds depletion are primarily due to reduced transcription of genes involved in kinetochore assembly and centriole duplication. The differences in the subcellular localizations of the NSL components may reflect direct mitotic functions that are difficult to detect at the phenotypic level, because they are masked by the transcription-dependent deficiency of kinetochore and centriolar proteins.
| The Drosophila Nonspecific Lethal (NSL) complex is a conserved protein assembly that controls transcription of more than 4,000 housekeeping genes. We analyzed the mitotic functions of four genes, Rcd1, Rcd5, MBD-R2 and wds, encoding NSL subunits. Inactivation of these genes by RNA interference (RNAi) resulted in defects in both chromosome segregation and centrosome duplication. Our analyses indicate that RNAi against Rcd1, Rcd5 or MBD-R2 reduces transcription of genes involved in centromere/kinetochore assembly and centriole replication. During interphase, Rcd1, Rcd5, MBD-R2 and Wds are confined to the nucleus, as expected for transcription factors. However, during mitosis each of these proteins relocates to specific mitotic structures. Our results suggest that the four NSL components work together as a complex to stimulate transcription of genes encoding important mitotic determinants. However, the different localization of the proteins during mitosis suggests that they might have acquired secondary “moonlighting” functions that directly contribute to the mitotic process.
| The spindle is a microtubule (MT)-based highly dynamic molecular machine that mediates chromosome segregation during cell division. Spindle formation requires many proteins that specifically bind and/or regulate MT assembly and dynamics, and also some proteins that are associated with the chromatin during interphase [1]. These latter proteins dissociate from the chromosomes upon mitotic entry and return to the nucleus during telophase; they have therefore functional roles in both interphase chromatin and in mitotic spindle assembly. Examples of these proteins are the components of the human KAT8-associated nonspecific lethal (KANSL) complex, which includes at least seven proteins: KANSL1, KANSL2, KANSL3, MCRS1, PHF20, WDR5 and the MOF/KAT8 histone acetyltransferase. The KANSL complex localizes in the nucleus of interphase cells where it regulates transcription of a specific set of genes and contributes to stem cell identity [2, 3]. Mutations in KANSL1 dominantly induce the Koolen-de Vries syndrome characterized by mental retardation and peculiar facial features, and mutations in KANSL2 have been associated with intellectual disabilities [4–6]. Studies on human cells and Xenopus egg extracts have shown that during mitosis, KANSL1, KANSL3 and MCRS1 re-localize from the chromatin to the MT minus ends of the mitotic spindle, playing essential roles in spindle assembly and chromosome segregation [7, 8]. Also WDR5 associates with the spindle during mitosis, and its RNAi-mediated depletion leads to spindle defects and chromosome misalignment [9]. Thus, the subunits of the KANSL complex, which is restricted to the nucleus during interphase, after the nuclear envelope breakdown redistribute to spindle structures where they are thought to play mitotic functions.
The Drosophila Nonspecific Lethal (NSL) complex is the fly counterpart of the KANSL complex. It includes seven conserved subunits: Nsl1/Wah (KANSL1), Dgt1/Nsl2 (KANSL2), Rcd1/Nsl3 (KANSL3), Rcd5 (MCRS1), MBD-R2 (PHF20), Wds (WDR5) and Mof (MOF/KAT8). MBD-R2, Rcd5, Nsl1, and Mof colocalize in the interbands of polytene chromosomes of third instar larvae, and MBD-R2 physically interacts with the histone modifying complexes Trx/MLL and the Nup98 nucleoporin [10, 11]. Wds is not only a member of the NSL complex, but is also part of several chromatin complexes including the ATAC histone acetyltransferase and the Trx/MLL histone methyltransferase [9, 12, 13]. The NSL complex associates with the promoters of more than 4,000 housekeeping genes, indicating that it acts as a major transcriptional regulator [10, 12, 14].
As in the case of their human counterparts, depletion of the NSL complex members results in mitotic defects. Genome-wide RNAi screens performed in S2 cells showed that Dgt1, Rcd1 and Rcd5 control mitotic centrosome behavior. Dgt1 (diminished γ-tubulin 1) is required for γ-tubulin recruitment, Rcd1 (Reduction in Cnn dots 1) for centriole duplication, and Rcd5 (Reduction in Cnn dots 5) for both centriole duplication and pericentriolar material (PCM) recruitment at the centrosomes [15, 16]. However, the mechanisms through which Dgt1, Rcd1 and Rcd5 regulate centriole duplication and centrosome maturation are currently unknown. Another RNAi-based screen in S2 cells showed that Rcd1 and MBD-R2 are required for mitotic chromosome segregation, but the mechanisms leading to this phenotype are also unknown [17].
Here, we analyzed the mitotic phenotypes caused by Rcd1, Rcd5, MBD-R2 or Wds depletion in greater detail. Cells depleted of these proteins showed a common defect in centrosome duplication, confirming the centrosome phenotype described earlier for Rcd1 and Rcd5 RNAi cells [16]. However, we found that Rcd1, Rcd5 and MBD-R2 RNAi cells, and to a lesser extent wds RNAi cells, also exhibit defect in chromosome alignment and segregation. Accordingly, Rcd1-, Rcd5- and MBD-R2-depleted cells displayed reduced levels of both Cid and Ndc80. Cid is a centromere-specific histone variant homologous to CENP-A that is required for kinetochore assembly [18], and Ndc80 is the protein that directly mediates MT attachment to the kinetochore [19]. We generated stable S2 cell lines for the inducible expression of Rcd1-GFP, Rcd5-GFP, MBD-R2-GFP or Wds-GFP. We show that these tagged proteins largely rescue the mitotic phenotype caused by RNAi-mediated depletion of their endogenous counterparts, and localize to the nucleus of interphase cells, as expected for transcription factors. However, during mitosis each protein relocalizes to a specific set of mitotic structures, including the centrosomes, the kinetochores and different regions of the midbody. Our results suggest that the common mitotic phenotypes generated by the depletion of the four NSL components are primarily due to reduced transcription of genes encoding centromere/kinetochore components and genes required for centriole duplication. The different mitotic localizations of Rcd1, Rcd5, MBD-R2 and Wds further suggest that these proteins might have acquired some direct mitotic roles.
We first analyzed the centrosomal phenotype of S2 cells exposed to Rcd1, Rcd5, MBD-R2 or wds double stranded RNAs (dsRNAs) for 5 days. To check for RNAi efficiency we performed Western blotting of cell extracts using antibodies directed to Rcd1, MBD-R2 or Wds. These assays showed that the antibodies specifically recognize bands of the expected molecular weights (MWs) that are strongly reduced after RNAi (Fig 1A–1C). The efficiency of RNAi against Rcd5 was demonstrated by RT-qPCR, which revealed a drastic reduction of the Rcd5 transcript level (Fig 1D). RNAi cells were fixed and stained for DNA, tubulin and the centrosomal marker Spd-2 [20, 21]. We examined only cells that appeared to have the basic karyotype of the S2 cells (~ 12 chromosomes [22]) and did not consider cells that are clearly “polyploid” that are common in S2 cell cultures. However, several S2 cells with the basic karyotype have either a single centrosome or more than two centrosomes (Fig 1E).
To ascertain whether depletion of Rcd1, Rcd5, MBD-R2 or Wds affects centrosome structure and duplication, we scored the Spd-2 signals in prometaphase and metaphase cells from 3 independent experiments. In this analysis, we counted cells showing no centrosomes, a single centrosome, two centrosomes, or more than 2 centrosomes. We did not subdivide the cells with more than 2 centrosomes in subclasses (i.e. cells with different centrosome numbers), because multiple centrosomes tend to overlap or cluster at the spindle poles and are therefore difficult to count. In addition, we did not take into account cells showing very small Spd-2 signals, because it was unclear whether they represented examples of PCM fragmentation or were produced by immunostaining artifacts.
In Rcd1, Rcd5, MBD-R2 and wds RNAi cells, the frequencies of cells showing 0–1 centrosomes were significantly increased compared to controls, although in Wds-depleted cells this increase was more modest that in the other RNAi cells (Fig 1F). The frequency of cells with more than two centrosomes was either significantly decreased (Rcd1, Rcd5, MBD-R2 RNAi cells) or unchanged (wds RNAi) compared to control (Fig 1F). We also measured the fluorescence intensity of Spd-2 signals in prometaphase and metaphase cells showing two centrosomes; we found that in Rcd1, MBD-R2 and wds RNAi cells the average fluorescence of these signals is not significantly different from controls, while in Rcd5-depleted cells it was slightly but significantly higher (p < 0.01; Wilcoxon Signed-Rank Test) than in control (Fig 1G and S1 Data). Consistent with these findings, we found that the asters of these RNAi cells are indistinguishable from those of control cells. We note that the results on Rcd1 RNAi cells are fully consistent with those of Dobbelaere et al. [16], who showed that depletion of Rcd1 reduces the centrosome number without affecting PCM recruitment. Dobbelaere et al. also showed that Rcd5 depletion negatively affects centrosome duplication but also reduces Cnn recruitment at centrosomes. We confirmed that Rcd5 deficiency reduces centrosome number but we found that it leads to a small increase in the fluorescence intensity of centrosome-associated Spd-2. We do not know the reasons for this apparent discrepancy but it is possible that the centrosomes of Rcd5-depleted cells are defective in Cnn recruitment but have normal or slightly increased ability to associate with Spd-2.
In Drosophila S2 cells, which often contain multiple centrosomes and can divide even in the absence of centrosomes, centrosome loss could arise either through defective centrosome separation during prophase or defective centriole duplication during interphase [16, 23]. Failure in centrosome separation in prophase would lead to formation a bipolar spindle with two centrosomes (each containing a pair of centrioles) at one pole and no centrosome at the other. Because S2 cells can assemble fully functional monastral spindles [24], a defect in prophase centrosome separation is expected to ultimately lead to an increase in both cells with zero and with more than two centrosomes. In the case of defective centriole duplication during interphase, the expectation is instead an increase in cells with 0–1 centrosomes and a decrease in cells with more than 2 centrosomes. Thus, our results (Fig 1F) are consistent with the second alternative, and suggest that depletion of the NSL components primarily impairs centriole duplication.
To confirm and extend these results, we counted the number of centrioles in interphase cells depleted of Rcd1, Rcd5, MBD-R2 or Wds. As centriole marker we decided to use Asterless (Asl), a protein that is thought to link the centriole wall with the PCM [25–27]. However, preliminary tests with antibodies directed to Asl or to other centriolar components revealed that in our hands they do not stain the centrioles in all interphase cells. We thus constructed an S2 cell line that expresses Asl-GFP under the control of the copper-inducible Metallothionein A (MtnA) promoter, and used it in the RNAi experiments against the NSL genes. As Asl overexpression causes centriole overduplication during the S phase [26], we limited the impact of this event by treating both RNAi and control cells with CuSO4 for only 12 hours before fixation (the cell cycle of S2 cells lasts approximately 24 hours). We stained control and RNAi cells with an anti-GFP antibody and scored them for the number of Asl signals present in interphase cells (Fig 1H). We did not take into account cells with zero signals because they could be cells in which the Asl-GFP expression was not sufficiently strong to mark the centrioles. RNAi-mediated depletion of Rcd1, Rcd5, MBD-R2 or Wds resulted in significant increases in interphase cells showing a single Asl signal compared to control. We also found significant decreases in cells with more than 2 signals in Rcd1, MBD-R2 and wds RNAi cells; the frequency of Rcd5 RNAi cells with multiple centrioles was also reduced compared to control but not significantly (p = 0.07) (Fig 1H and 1I). Collectively, these results reinforce the conclusion that the NSL components are required for centriole duplication.
An analysis of mitotic division in several independent experiments revealed that Rcd1-, Rcd5- and MBD-R2-depleted cells exhibit very similar phenotypes. They display an approximately ten-fold reduction in the frequency of anaphases compared to controls and very high frequencies of PMLES (prometaphase-like cells with elongated spindles) figures (Fig 2A–2C). PMLES (formerly called pseudo-ana-telophases, abbreviated with PATs [17]) are peculiar mitotic figures that contain late anaphase/early telophase-like spindles associated with chromosomes that are improperly comprised of both sister chromatids and are usually scattered along the spindles; in some cases PMLES exhibit central spindle-like structures and irregular cytokinetic rings (Fig 2B and S1 Fig; see also refs. [17, 28]). Notably, several PMLES exhibit arched spindles, which are presumably the consequence of an excessive spindle elongation within the constraints imposed by the plasma membrane. Despite their ana-telophase-like spindle structure, PMLES contain high levels of Cyclin B, which is normally degraded at the beginning of anaphase (Fig 2A and 2B). Thus, PMLES appear to be in a pre-anaphase stage as far as sister chromatid separation and Cyclin B degradation are concerned, but they are nevertheless permissive of typical telophase events, such as central spindle assembly and initiation of cytokinesis.
To further characterize the PMLES, we measured the length of their spindle axis and compared it with the length of the other mitotic figures. In Rcd1, Rcd5 and MBD-R2 RNAi cells the prometaphase and metaphase spindles were morphologically normal but slightly longer (~10%) than their control counterparts (Fig 2D, S2 Data). PMLES spindles of RNAi cells were instead substantially longer than prometaphase/metaphases spindles (~ 60%) and anaphases spindles (~ 25%) (Fig 2D and 2E). The average length of the PMLES spindles was either slightly shorter or similar to the length of the spindles of control telophases (Fig 2E), indicating that most PMLES can attain the maximum spindle elongation that is normally achieved by S2 cells. The slight increase in the prometaphase and metaphase spindle length observed in RNAi cells is likely to reflect the presence of some cells in the initial stages of evolution towards a PMLES configuration.
Although Rcd1, Rcd5 and MBD-R2 RNAi cells exhibit very low anaphase frequencies (Fig 2C), they show relatively high frequencies of mitotic figures with telophase-like spindles, characterized by the presence of a constricted central spindle and decondensed chromosomes at the cells poles; about a third of these cells displayed lagging chromosomes between the cell poles (Fig 2B and S1D Fig). Because the chromosomes in these cells are decondensed, it is not possible to discern whether they contain one or two sister chromatids. Nonetheless, we favor the idea that many of the “telophases” observed in Rcd1-, Rcd5- and MBD-R2-depleted cells are in fact PMLES that managed to progress further through the mitotic process and undergo chromosome decondensation as normally occurs in telophase. Regardless of the nature of these telophase-like cells, it is clear that Rcd1, Rcd5 and MBD-R2 are all required for sister chromatid separation and chromosome segregation.
Cells depleted of Wds displayed a mitotic phenotype qualitatively similar but quantitatively milder than that observed in Rcd1-, Rcd5- or MBD-R2-deficient cells (Figs 1F and 2C-2E). Thus, although the degree of RNAi-mediated depletion of the Wds protein is comparable to that observed for the other NSL proteins (Fig 1C), in wds RNAi cells both the centrosome and chromosome segregation phenotypes are substantially milder than those observed in Rcd1, Rcd5 or MBD-R2 RNAi cells.
Collectively our findings indicate that Rcd1-, Rcd5- and MBD-R2-depleted cells are severely defective in chromosome segregation. Depletion of Wds also perturbed chromosome segregation but caused a relatively mild defect. This chromosome segregation phenotype cannot be ascribed to centrosome defects, as abundant evidence indicates that Drosophila mitotic spindle assembly and functioning does not require the centrosomes [15, 17, 29–31].
We have previously observed frequent PMLES in S2 cells depleted of the centromere-specific histone H3 Cid/Cenp-A or the kinetochore components Ndc80, Nuf2 and Kmn1 [17, 32]. PMLES have been also observed in S2 cells depleted of Mast/Orbit that has a role in MT-kinetochore attachment [33], or depleted of the Klp67A kinesin-like protein, which represses MT plus end growth and is required for proper MT binding to kinetochores [32]. More recently, PMLES were observed in S2 cells depleted of the Sf3A2 and Prp31 splicing factors that have direct roles in MT-kinetochore interactions [28]. These results suggest that the PMLES found in Rcd1, Rcd5 and MBD-R2 RNAi cells are a consequence of a defective kinetochore-MT attachment.
To test this possibility, we first analyzed the effects of Rcd1, Rcd5, MBD-R2 or Wds depletion on the intracellular concentration of the core centromere component Cid/CenpA that is required for kinetochore assembly [18], and Ndc80 that is directly responsible for MT-kinetochore attachment [19, 34–36]. Western blotting analysis showed that in Rcd1-, Rcd5- and MBD-R2-depleted cells Cid is considerably reduced compared to controls (32%, 36% and 40% of the control level, respectively); the level of Cid was also reduced in Wds-depleted cells but to a lesser extent (78%) (Fig 3A, S3 Data). In addition, Western blotting on extracts from Rcd1, Rcd5, and MBD-R2 RNAi cells showed substantial reductions of Ndc80 compared to mock-treated cells (32%, 35% and 52% of the control level, respectively), while in wds RNAi cells the Ndc80 level was comparable to that of controls (Fig 3A, S3 Data).
To extend the analysis at the subcellular level we focused on Rcd1, Rcd5 and MBD-R2 RNAi cells, which exhibit strong reductions in the Cid content compared to both control and Wds-depleted cells. The Cid protein is detectable both at the centromeres of mitotic chromosomes and in interphase nuclei; nuclei of control cells exhibit 3–6 foci corresponding to clustered centromeres (Fig 3B). In Rcd1, Rcd5 and MBD-R2 RNAi cells, the frequencies of interphase nuclei devoid of Cid signals were significantly higher than in controls (Fig 3B and 3C). We also examined prometaphases and metaphases with a basic karyotype for the presence of Cid signals. In general, RNAi cells displayed Cid signals of lower intensity compared to controls; they also showed a great variability in the number of detectable signals, ranging from zero to more than 20 signals (in a cell with 12 chromosomes the maximum number of Cid signals is 24). In Rcd1-, Rcd5- and MBD-R2-depleted cells, the frequencies of cells with 0–9 signals were drastically increased compared to controls, in which this type of cells are virtually absent (Fig 3C).
The similarity of the mitotic phenotypes observed in Rcd1-, Rcd5- and MBD-R2-depleted cells might be a consequence of an interdependence of these proteins. Indeed, it has been previously shown that depletion of Rcd5 in salivary glands leads to a severe reduction in Rcd1, while Rcd1 depletion results only in a slight reduction in Rcd5 [10]. Similar results were obtained in SL-2 cells, where RNAi-mediated depletion of Rcd5 caused a reduction in Rcd1, whereas Rcd1 depletion did not substantially affects the Rcd5 level [10]. In both salivary glands and SL-2 cells, MBD-R2 depletion did not affect the stability of the other members of the complex [10], while the Wds levels were similar to those of controls in Rcd1, Rcd5 or MBD-R2 RNAi cells [10]. We found that RNAi against Rcd5 leads to a small reduction in Rcd1 also in the S2 cell line used here. Rcd5 RNAi cells displayed a small reduction in MBD-R2, while MBD-R2 depletion did not substantially affect the Rcd1 level (Fig 4А). Consistent with these results, RT-qPCR showed that RNAi against each of the four NSL genes studied here does not substantially affect transcription of the others, which, with the exception of wds upon Rcd5 RNAi, are expressed at slightly higher rates compared to control (Fig 4B). These results suggest that the phenotypes observed in Rcd1-, Rcd5-, MBD-R2- and Wds-depleted cells are largely due to the deficiency of each individual factor and do not reflect interdependencies among these proteins.
In the attempt of detecting a mitotic localization of the NSL subunits, we first tested the commercial anti-MBD-R2 and anti-Wds antibodies and our anti-Rcd1 mouse antibody. All these antibodies specifically recognize bands of the expected MWs that are strongly reduced in RNAi cells (Figs 1A–1C and 4A). However, immunostaining using the same antibodies showed that the anti-MBD-R2 and anti-Wds antibodies stain weakly only the interphase nuclei but do not decorate any mitotic structure, while the anti-Rcd1 did not work at all in indirect immunofluorescence.
We then generated stable cell lines expressing Cherry-tubulin and any of Rcd1-GFP, Rcd5-GFP, MBD-R2-GFP or Wds-GFP tagged proteins, all under the control of the copper-inducible MtnA promoter. We exposed these cells for 12–14 hours to different concentrations of CuSO4 (0.1, 0.25, 0.4 and 0.5 mM) and examined them under a confocal fluorescence microscope. All GFP-tagged proteins showed strong nuclear signals but also displayed clear but different mitotic localizations (Figs 5 and 6). In prometaphase and metaphase cells, Rcd1-GFP and Rcd5-GFP proteins were no longer associated with chromatin, but accumulated at the centers of the asters/centrosomes and were occasionally weakly enriched at the spindle area. In anaphase and telophase cells, the accumulation at the centrosomes was reduced compared to the previous mitotic phases. In telophase cells, Rcd1-GFP and Rcd5-GFP concentrated in the daughter nuclei as expected for transcription factors, but were also enriched at the midbody, the structure that connects the two daughter cells during late telophase and cytokinesis (Fig 5). The midbody contains bundled antiparalled MTs with their plus ends overlapping at the center of the structure. The MT overlapping area is often dark (dark zone) after staining with anti-tubulin antibodies because it is enriched in proteins that block antibody binding to tubulin [37]. We see a dark zone also in living cells expressing Cherry-tubulin, most likely because the same proteins that prevent antibody binding quench the Cherry-tubulin fluorescence. Interestingly, while Rcd1 was excluded from the dark zone and enriched at both sides of this region, Rcd5 was specifically accumulated in the dark zone at center of the midbody (Fig 5).
After CuSO4 induction, MBD-R2-GFP and Wds-GFP were strongly enriched in interphase nuclei but showed very different localization patterns in mitotic cells. MBD-R2-GFP localized exclusively at the chromosomes and did not show accumulations at either the centrosomes or the midbody (Fig 6A). In contrast, Wds-GFP was enriched at several mitotic structures (Fig 6B and 6C). In all mitotic phases, Wds-GFP accumulated in a discrete region of a specific chromosome. This GFP signal was also detected in both telophase and interphase nuclei of living cells, and was sufficiently strong to be detected in fixed cells with well spread chromosomes (Fig 6B and 6C). This allowed us to localize the Wds-GFP accumulation on a specific region of an acrocentric chromosome characterized by a highly DAPI-fluorescent pericentric heterochromatin (Fig 6C). The DAPI staining pattern of this chromosome and the localization of the GFP signal along the chromosome suggest that Wds might associate with the nucleolus organizer of the X chromosome in both mitotic cells and interphase nuclei [22]. We also observed an enrichment of Wds-GFP at the centrosomes; this enrichment was clearly visible in most prophase, prometaphase and metaphase cells, but was hardly detectable in anaphases and telophases. In addition, in most late prometaphase and metaphase cells, Wds-GFP was enriched at structures that are likely to correspond to the centromeres/kinetochores (Fig 6B). This localization is transient, and was never observed in the other mitotic phases. In all telophase cells, Wds-GFP was accumulated at the midbody dark zone.
The cells shown in Figs 5 and 6 were treated for 12–14 hours with 0.4 mM CuSO4, as this is the optimal concentration for a clear visualization of both the GFP-tagged protein and Cherry-tubulin. However, we were able to see mitotic accumulations of GFP-tagged proteins also after 12–14 hours induction with 0.1 mM CuSO4. Importantly, at all CuSO4 concentrations, we consistently observed the same localization patterns as those shown in Figs 5 and 6. Thus, the mitotic localization of each protein appears to be a characteristic feature of the protein independent of its intracellular concentration.
To obtain further insight into this issue, we also performed a Western blotting analysis to determine the levels of the endogenous and the GFP-tagged forms of Rcd1, MBD-R2 and Wds after CuSO4 induction (S2 Fig). We could not carry out this analysis for cells expressing Rcd5-GFP due to the unavailability of an anti-Rcd5 antibody. These experiments showed that MBD-R2-GFP and its endogenous counterpart were expressed at similar levels after induction with 0.1 or 0.25 mM CuSO4, while induction with 0.5 mM CuSO4 led to a limited MBD-R2-GFP overexpression (S2A Fig, S4 Data). In MtnA-Rcd1-GFP-bearing cells, the endogenous and the GFP-tagged protein were expressed at similar levels after induction with 0.1 mM CuSO4, but the GFP protein was 2.5- and 3.9-fold more abundant than the normal protein after induction with 0.25 and 0.5 mM CuSO4, respectively (S2B Fig, S4 Data). Wds-GFP was expressed at relatively high levels at all CuSO4 concentrations, with the tagged protein showing expression levels 3-4-fold higher than that of the corresponding endogenous protein (S2C Fig, S4 Data). These results strongly suggest that Rcd1 and MBD-R2 localizations are independent of the concentration of the individual proteins.
The lack of biochemical data on Rcd5-GFP and the fact the Wds-GFP is 3 to 4 times more abundant than the endogenous protein do not permit us to exclude that the mitotic localizations of these proteins could be partially affected by their intracellular quantity. However, an analysis of live cells expressing either Rcd5-GFP or Wds-GFP strongly suggests that this is not the case. In cells induced with 0.1 mM CuSO4, there is great variability in the levels of the GFP-proteins expressed by the individual dividing cells. Nonetheless, regardless their degree of GFP fluorescence, all mitotic cells showed the same Rcd5-GFP- or Wds-GFP-specific accumulations (see Figs 5 and 6), suggesting that the mitotic localization of these proteins is largely independent of their intracellular concentration.
We also fixed the cells treated for 12 hours with either 0.1 or 0.5 mM CuSO4 and stained them with anti-GFP and anti-tubulin antibodies. Cells expressing Rcd1-GFP, Rcd5-GFP, MBD-R2-GFP or Wds-GFP fixed with standard formaldehyde- and paraformaldehyde-based procedures (see Materials and Methods) showed an evident GFP staining of interphase nuclei. However, the localization of these proteins on the mitotic apparatus was fixation-dependent but independent of the concentration of CuSO4. In Rcd1-GFP and Rcd5-GFP expressing cells, we did not detect any clear accumulation of the tagged proteins at either the centrosomes or the central spindles (S3 Fig). This prevented double immunostaining experiments aimed determining whether Rcd1 and Rcd5 precisely colocalize with the centrosomes. Previous work has shown that both Rcd1 and Rcd5 are required for centriole duplication in S2 cells but failed to detect accumulations of these proteins at the centrosomes [16]. However, in these studies cells expressing GFP-tagged Rcd1 or Rcd5 were fixed with 4% paraformaldehyde, a treatment that likely disrupted centrosomal localization of the GFP-tagged proteins [16]. In contrast, fixed cell expressing MBD-R2 GFP showed a strong and specific enrichment of the tagged protein at the chromosomes (S3 Fig). Lastly, after fixation, Wds-GFP was consistently enriched at a specific chromosomal region in metaphase and anaphase cells and at the dark zone of the midbody during telophase. The fixation-resistant localization of Wds on a discrete chromosomal region is a likely example of mitotic chromosome bookmarking. Such bookmarking occurs when transcription factors remain associated with chromosomes during mitosis so as to facilitate reactivation of a subset of genes in the subsequent cell cycle [38].
The observation that the GFP-tagged components of the NSL complex exhibit different localization patterns during mitosis raises the question of whether these proteins have the same functions as their non-tagged counterparts. To address this question we performed RNAi using dsRNAs that target only the 5ʹ and 3ʹ untranslated regions (UTRs) of the Rcd1, Rcd5, MBD-R2 and wds endogenous genes (see S1 Table) but not the coding sequences (CS) of the GFP-tagged transgenes. We specifically asked whether the expression of the GFP-tagged NSL proteins rescues the mitotic effects caused by treatments with the corresponding UTR dsRNAs. In cells expressing the GFP-tagged proteins, dsRNAs targeting of CS resulted in very strong phenotypic effects comparable to those observed in normal cells treated with the same CS dsRNAs (compare Fig 2C with S2 Table). dsRNAs targeting the UTR sequences were less effective than CS dsRNAs in inducing mitotic defects (Fig 2C, S2 Table), consistent with the fact that the CS used for RNAi are considerably longer than the corresponding UTRs (S1 Table). However, when RNAi with UTR dsRNAs was performed in cells expressing the corresponding GFP-tagged proteins (induced by 0.1 mM CuSO4; see Materials and Methods) the mitotic effects were substantially milder than those observed in cells that express only the endogenous proteins (S2 Table). In summary, the data reported in the S2 Table show that the GFP-tagged forms of Rcd1, Rcd5, MBD-R2 and Wds rescue the phenotypic defects caused by depletion of the endogenous proteins. This suggests that the GFP-tagged NSL components are largely functional and supports the view that their different localizations during mitosis reflect the normal localizations of their untagged counterparts.
The finding that Rcd1, Rcd5, MBD-R2 and Wds exhibit different localization patterns during mitosis, and yet cause very similar phenotypes when depleted, suggests the hypothesis that these proteins might act together in regulating the expression of mitotic genes. Previous work has shown that at least four components of the NSL complex (Nsl1, Rcd1, Rcd5 and MBD-R2) bind the active promoters of more than 4,000 constitutively expressed genes, suggesting that NSL acts as complex to specifically upregulate this type of genes [10, 12, 14]. However, it appears that NSL depletion results in diminished expression of only a subset of the genes to which it is bound [12].
To address the possibility that depletion of the NSL components affects mitotic gene transcription, we exploited the published ChIP and gene expression datasets generated in S2 cells [12, 14] to ask whether the NSL complex binds the promoters and regulates the expression of genes encoding (i) centromere and kinetochore proteins (cid, Cenp-C, Mis12, Nnf1a, Nnf1b, Kmn1/Nsl1, Ndc80, Nuf2, Spc25/Mitch and Spc105R/KNL1), (ii) factors that mediate centriole duplication (ana2, asl, SAK, Sas-4, Sas-6) and, and (iii) components of the spindle assembly сheckpoint (SAC) machinery (Mad1, mad2, Bub1, Bub3, BubR1, Zw10, rod, Zwilch, cmet, nudE). We examined the SAC genes not only as a term of comparison with centromere/kinetochore and centriole duplication genes but also to gather information on whether the SAC is compromised in the absence of functional NSL complex. We found that the promoters of all these genes are bound by at least two components of the NSL complex (Rcd1 and MBD-R2; Table 1 and S3 Table). In addition, interrogation of published datasets [12] revealed that RNAi-mediated Nsl1 depletion results in reduced transcription of most of these genes (Table 1). Among the centromere/kinetochore genes, the strongest reductions in transcription were observed for Nnf1b, Mis12 and cid, which were transcribed at 6.1%, 6.9% and 17.5% of the control level, respectively. The genes specifying SAC functions were generally under-transcribed, with rod and Zwilch showing particularly reduced transcription levels (below 30% the control level), suggesting the SAC could be partially compromised in cells depleted of the NSL components. Finally, all genes required for centriole duplication showed reduced transcription, with Sas-6, Sas-4 and asl transcripts reduced to 9.5%, 26.3%, and 30.3% of those of controls, respectively (Table 1).
These data prompted us to determine the transcription levels of cid, Ndc80, Nnf1b, Mis12 and Spc25/Mitch in cells depleted of the NSL components. RT-qPCR showed that the transcripts of all these genes are substantially reduced in Rcd1, Rcd5 and MBD-R2 RNAi cells, but only weakly reduced in wds RNAi cells. In cells depleted of Rcd1, Rcd5 or MBD-R2, the Nnf1b and Mis12 transcripts were below
(1) Cen, centromere; Kin, kinetochore; SAC, spindle assembly checkpoint; Ced, centriole duplication. (2) ChIP data on Rcd1 and MBD-R2 are from [14]; ChIP data on Nsl1 are from [12]. We did not include Kmn2 in the table because it was not included in the analyses of [14] and [12]. Transcription factor binding to the gene promoter region is indicated by +; for a quantitative analysis of promoter binding see S3 Table; NA, data not available. (3) Gene expression data are from [12]. (4) The values reported are the means of a number (indicated between brackets) of independent experiments.
30% of the control level, while the cid transcripts were approximately 50% of the control level; the Spc25/Mitch and Ndc80 transcripts were roughly 70% of those of controls (Table 1). Importantly, the reductions of the transcript levels detected by RT-qPCR, although quantitatively lower, are absolutely proportional to those previously observed by microarray experiments (Table 1 and [12]). The transcription level of Ndc80 detected by RT-qPCR does not match the strong reduction of the protein seen by Western blotting (Fig 3A). However, it is possible that this discrepancy is due to the downregulation of Nuf2 transcription (Table 1), which might affect the quantity of the Ndc80 protein, as Nuf2 and Ndc80 are mutually dependent for their stability [34, 39].
We also determined whether RNAi against Rcd1, Rcd5, MBD-R2 and wds affects the transcription of asl, Sas4 and Sas-6 (Table 1). Sas-6, SAK and Ana2 form a conserved core module required for Drosophila centriole duplication [40, 41]. Consistent with the published ChIP data (Table 1), we found that in Rcd1, Rcd5 and MBD-R2 RNAi cells the Sas-6 transcripts are substantially reduced compared with controls, ranging from 20% to 37% of the control level (Table 1). Reductions in the asl and Sas-4 transcripts were less pronounced, ranging from 40–56% and 63–66% of the control levels, respectively. In wds RNAi cells the abundance of Sas4 and Sas-6 was slightly reduced while the level of asl transcripts was not affected (Table 1). Thus, our direct measure of the transcript abundance determined by RT-qPCR is fully consistent with both the published datasets and with our hypothesis that both the centriole duplication and chromosome segregation phenotypes are caused by reduced transcription of specific mitotic gene sets.
The finding that Rcd1, Rcd5, MBD-R2 and Wds exhibit different localization patterns during mitosis, and yet cause very similar phenotypes when depleted raised the hypothesis that these proteins regulate mitosis by controlling the transcription of mitotic genes. This hypothesis is strongly supported by previous [12, 14] and current (see Table 1) findings indicating that Rcd1, Rcd5 or MBD-R2 depletion results in a substantial downregulation of several genes involved in centromere/kinetochore assembly and centriole duplication. This hypothesis is further corroborated by the observation that Wds depletion, which causes a limited reduction in mitotic gene transcription, also results in a milder mitotic phenotype compared to those caused by Rcd1, Rcd5 or MBD-R2 deficiency.
It has been previously shown that the NSL complex specifically binds the promoters of most housekeeping genes and activates a large subset of these genes. It has been further shown that subunits of the NSL complex co-localize at promoters of the target genes, and that the complex acts as a single functional unit [10, 14]. Consistent with these results, we found that RNAi-mediated silencing of Rcd1, Rcd5 or MBD-R2 results in identical defects in sister chromatid separation. These defects lead to PMLES, which have been previously observed in cells defective in MT-kinetochore interactions [17, 28, 32, 33]. Analysis of published datasets revealed that transcription of both cid and several kinetochore protein-coding genes is downregulated in Nsl1-depleted cells (Table 1; see [12]). Moreover, we showed that in Rcd1, Rcd5 and MBD-R2 RNAi cells the transcription levels of cid, Mis12 and Nnf1b, and to lesser extent, those of Ndc80 and Spc25/Mitch, are reduced with respect to controls. The same RNAi cells displayed fewer Cid signals and reduced levels of the Cid and Ndc80 proteins compared to control. Thus, these results collectively suggest that Rcd1, Rcd5 and MBD-R2 work together in interphase to regulate proper transcription of multiple genes encoding centromere and kinetochore components. We propose that reduced transcription of these genes disrupts proper kinetochore assembly, impairing kinetochore-MT interaction. There are at least two considerations that support this interpretation. First, there is a clear hierarchy in recruitment of the kinetochore proteins. Cid is required for recruitment of all kinetochore proteins; localization of the Mis12 complex (Mis12, Nnf1a, Nnf1b and Knm1) and Spc105R/KNL1 are interdependent, while recruitment of the Ndc80 complex (Ndc80, Nuf2, Spc25R/Mitch) requires both Spc105R/KNL1 and the Mis12 complex. Second, even components of the same complex such as Ndc80 and Nuf2 are mutually dependent for their stability/localization [34–36, 42]. These complex dependency relationships suggest that even relatively modest reductions in kinetochore proteins can generate synthetic effects leading to kinetochore dysfunction.
We have shown that Rcd1 and Rcd5 accumulate at the centrosomes while MBD-R2 localization is restricted to the chromosomes. Nonetheless, depletion of each of the three NSL components leads to a clear defect in centrosome duplication. Thus, it is unlikely that this centrosomal phenotype is caused by a direct effect of the NSL complex proteins on centrosome behavior. Here again, the most likely explanation is that defective centrosome duplication is due to reduced transcription of multiple genes required for proper centriole structure and duplication, such as ana2, asl, SAK, Sas-4 and Sas-6 [23, 40, 41, 43, 44]. The products of these genes also show dependencies in their recruitment at centrioles. For example, the SAK kinase recruits the centriole cartwheel components Sas-6 and Ana2 that are required for recruitment of Sas-4, which in turn recruits Asl [41, 44], suggesting that multiple, even modest, depletions of these proteins can impair the centriole duplication machinery. Thus, we propose that the centrosome duplication phenotype elicited by RNAi against Rcd1, Rcd5 or MBD-R2, is due to reduced transcription of target genes required for centriole duplication.
Mitotic defects caused by depletion of transcription factors have been previously observed both in flies and mammals. For example, mutations in genes encoding subunits of the Drosophila TFIIH transcription complex have been previously shown to disrupt mitosis. The TFIIH holo-complex is comprised of two subcomplexes, a 7-proteins core complex (XPB, XPD, p62, p52, p44, p34 and p8) and the CAK (Cdk7, CycH and MAT1) complex [45, 46]. Mutations in marionette (mrn) that encodes Drosophila p52 cause defects in mitotic chromosome condensation and integrity in larval brains [47]. Recently it has been also shown that embryos produced by Drosophila females depleted of p8, p52, XPB or Cdk7 often exhibit mitotic defects; these defects include poorly condensed chromosomes, disorganized spindles, isolated centrosomes and chromosomes not associated with the spindle. These mitotic aberrations have been attributed to transcriptional downregulation of a set of critical genes. Specifically, embryos from p8 mutants showed reduced transcription of 104 genes that encode factors involved in DNA replication or mitosis [48].
Another interesting example pointing to an involvement of transcription factors in mitotic regulation is provided by studies on the human MMXD complex, which includes the MIP18, MMS19 and XPD proteins; XPD is also a component of TFIIH complex and is responsible for Xeroderma pigmentosum (XP). MMS19 or MIP18 knockdown moderately reduces the XPD level but does not affect the levels of the other TFIIH subunits, suggesting that the MMXD complex can function independently of the TFIIH complex. MIP18, MMS19 and XPD localize to the mitotic spindle of human cells and their siRNA-mediated depletion results in monopolar and multipolar spindles; similar aberrant spindles were also observed in cells from XP patients [49]. However, the molecular mechanisms leading to this mitotic phenotype are not fully understood, and both a direct mitotic role of the MMXD complex and possible subtle defects in global transcription were considered [49]. Strong mitotic defects have also been observed in human cells upon siRNA-mediated inactivation of ERG, a transcription factor of the Erg family. However, in this case the mitotic defects have been attributed to failure to degrade the Aurora A and B mRNAs, and to a consequent excess of these mitotic kinases [50].
While our data suggest that the chromosome segregation and the centrosome duplication phenotypes result from reduced transcription of mitotic genes during interphase, the localization of the NSL proteins to mitotic structures raises the possibility that these proteins might also play direct roles during cell division. A direct mitotic role of the KANSL complex components has been suggested in several studies on human cells. In HeLa cells, MCRS1 (Rcd5), KANSL1 (Nsl1/Wah) and KANSL3 (Rcd1/Nsl3) co-localize with the centrosomes at the center of the asters. KANSL1 and KANSL3 bind the MT minus ends, while MCRS1 does not directly bind MTs but is recruited at the minus ends by KANSL3 [7, 8]. RNAi-mediated silencing of the MCRS1, KANSL1 or KANSL3 resulted in very similar phenotypes consisting of misaligned chromosomes, a prolonged arrest in a prometaphase-like state, often followed by mitotic catastrophe [7, 8]. It has been suggested that MCRS1, KANSL1 and KANSL3 stabilize MTs, favor chromosome-driven MT formation, and promote correct kinetochore fiber dynamics [8]. WDR5, the human orthologue of Wds, has been also implicated in mitosis. WDR5 associates with the mitotic spindles [9] and is particularly enriched at the midbody dark zone [51]. Most interestingly, WDR5 depletion results in many cells that are highly reminiscent of the PMLES observed here. In most WDR5-depleted cells, the chromosomes did not properly align in the metaphase plate but are dispersed throughout unusually long anaphase-like spindles without undergoing sister chromatid separation, suggesting a defect in kinetochore-MT connection [9]. It has been also reported that RNAi-mediated knockdown of WDR5 impairs completion of cytokinesis leading to multinucleated cells [51].
We have shown that MBD-R2 remains associated with the chromosomes throughout mitosis, just like some components of the TFIIH complex that localize to the chromosomes of dividing nuclei in Drosophila embryos [48]. However, we found that Rcd1 (KANSL3) and Rcd5 (MCRS1) are enriched at the centrosomes and at the midbodies, with Rcd5 accumulating in the midbody dark zone and Rcd1 excluded from this region but enriched at its sides. Finally, we showed that during mitosis Wds remains associated with a specific chromosome region that is likely to correspond to the nucleolus organizer; in addition, it accumulates at the centrosomes, the kinetochores and the midbody dark zone. These results raise the question of whether the accumulations of Rcd1, Rcd5 and Wds at the centrosomes/asters and the midbody reflect direct roles of these proteins during mitosis. It is unlikely that Rcd1, Rcd5 or Wds localization at the centrosomes is related to centrosome duplication, because defects in this process have been also observed in cells depleted of MBD-R2, which fails to accumulate at the centrosomes. It is also unlikely that these three proteins are required for aster formation, as in Rcd1-, Rcd5- or Wds-depleted cells these structures are morphologically normal. Postulating centrosomal functions for Rcd1, Rcd5 or Wds is extremely difficult because the centrosomes contain hundreds of proteins that do not appear to serve canonical centrosome functions, and it remains to be determined whether these proteins fulfill centrosome-related regulatory functions [52, 53]. Postulating direct roles of Rcd1, Rcd5 and Wds at the midbody is even more difficult. Their enrichment at the midbody suggests an involvement in cytokinesis. However, we did not notice any significant increase in the frequency of binucleated cells after Rcd1, Rcd5 or wds RNAi, and none of these genes was detected in genome-wide RNAi-based screens aimed at identifying Drosophila genes required for cytokinesis [54, 55]. Moreover, proteomic analyses have shown that the midbody contains many proteins with no obvious roles in the execution of cytokinesis, and it is currently unknown whether they regulate some aspects of the process [56].
Studies on the human homologues of Rcd1, Rcd5 and Wds have shown that depletion of these proteins leads to chromosome segregation defects reminiscent of those observed in Drosophila cells [7, 8, 9]. These studies did not address the roles of Rcd1, Rcd5 and Wds on centrosome duplication and did not check whether their depletion leads to a reduced transcription of critical mitotic genes. However, it has been reported that KANSL3 (Rcd1) and MCRS1 (Rcd5) bind to MTs and physically interact with the TPX2 and MCAK spindle proteins, suggesting a direct participation of both KANSL components in the mitotic process [7, 8]. It has been also shown that WDR5 (Wds) interacts with the mitotic kinesin Kif2A and with the MLL complex that binds to MTs and regulates spindle assembly and chromosome segregation [9]. Although these findings do not exclude that downregulation of KANSL3, MCRS1 and WDR5 lowers the expression of a number of mitotic genes, they suggest that these KANSL subunits play direct mitotic roles. Likewise, we cannot exclude that Rcd1, Rcd5, MBD-R2 and Wds have some minor direct roles in chromosome segregation in Drosophila S2 cells. We have shown that Wds localizes to the kinetochores and it is possible that also small, cytologically undetectable, amounts of Rcd1, Rcd5, and MBD-R2 localize and function at kinetochores.
The human WDR5 protein localizes to dark zone of the midbody and is required for abscission during cytokinesis [51]. It is conceivable that Rcd1, Rcd5 and Wds play some functions in cytokinesis, and that these functions cannot be detected because they are masked by the chromosome segregation phenotype leading to PMLES. Alternatively, Rcd1, Rcd5 and Wds might play roles in cytokinesis that do not lead to a complete failure of the process; for example they could regulate the timing of the events underlying cytokinesis either in its early or late stages, such as central spindle assembly and MT severing during abscission.
In conclusion, our data suggest that depletion of components of the NSL complex negatively affects centriole duplication and kinetochore assembly through downregulation of genes required for these processes. However, we have also shown that Rcd1, Rcd5 and Wds accumulate at the centrosomes and the midbody suggesting possible moonlighting functions for these proteins during mitosis. It is generally accepted that protein moonlighting occurred through the gradual transition from the original function to a novel function, an evolutionary process that involves the coexistence of two functions in the same protein. Because these functions should not be conflicting it has been also posited that evolution of moonlighting functions is favored when they are exerted in different cellular compartments [57, 58]. The components of the conserved NSL/KANSL complex are therefore in a very favorable situation to evolve secondary mitotic functions, as transcription is strongly reduced during cell division [59, 60]. In both Drosophila and human cells, the NSL/KANSL components appear to localize to mitotic structures. However, the localizations and functions of the NSL subunits are not identical to those of their KANSL counterparts. We would like to speculate that the components of the KANSL and NSL complexes are both evolving towards the acquisition of direct mitotic functions. However, while in human cells the functions of these proteins are integral to the mitotic process, in Drosophila they are not yet essential for mitosis. Although, further analyses are required to compare the mitotic functions of the KANSL and NSL complexes, the current results suggest their components are evolving secondary mitotic functions that are partially different.
All DNA templates for synthesis of dsRNAs specific to the CS or the UTRs of the MBD-R2, Rcd1, Rcd5 and wds genes were amplified by PCR from a pool of cDNAs obtained from ovaries of 3-day-old wild-type females and from 0–2 hour wild-type embryos (for the primer sequences used, see S1 Table) The PCR products were purified using spin columns (BioSilica; http://biosilica.ru/). Synthesis of dsRNAs was done as described earlier [61], with the following minor modifications. Heating of the synthesized dsRNAs to 65°C and the subsequent slow cooling to room temperature were done before treatment with DNaseI; also, the phenol/chloroform extraction was omitted.
S2 cells free from mycoplasma contamination were cultured in 39.4 g/L Shields and Sang M3 Insect medium (Sigma) supplemented with 0.5 g/L KHCO3 and 20% heat-inactivated fetal bovine serum (FBS) (Thermo Fisher Scientific) at 25°C. S2 cells expressing GFP-tagged proteins were cultured in 39.4 g/L Shields and Sang M3 Insect medium supplemented with 2.5 g/L bacto peptone (Difco), 1 g/L yeast extract (Difco) and 5% heat-inactivated FBS at 25°C. RNAi treatments were carried out as follows. 1×106 cells were plated in 1 ml of serum-free medium in a well of a six-well culture dish (TPP) and 30 μg of CS dsRNA or 40 μg of UTR dsRNA (S1 Table) was added to each well. After a 1 hour incubation, 2 ml of the medium supplemented with 20% or 5% heat-inactivated FBS was added to each well and cells were grown for 3 days. After that, the second dose of the same dsRNA (30 μg of CS dsRNA or 40 μg of UTR dsRNA) was added to each sample and cells were grown for 2 additional days. In the case of the rescue experiments shown in S2 Table, together with the second dose of dsRNA, we added CuSO4 to the final concentration of 0.1 mM in the medium. Control S2 cell samples were prepared in the same way, but without addition of dsRNA.
Gene-specific primers were designed by using Primer-BLAST (https://www.ncbi.nlm.nih.gov/tools/primer-blast/) or Primer3 (http://bioinfo.ut.ee/primer3-0.4.0/primer3/) software; primer sequences are provided in S4 Table. For each primer pair, the efficiency was determined by construction of a standard curve using dilutions of the cDNA prepared from S2 cells according to [62] (S4 Table). Total RNA was isolated from control and dsRNA-treated S2 cells using RNAzol RT reagent (MRC) according to the manufacturer’s instructions. Genomic DNA was eliminated using the RapidOut DNA Removal Kit (Thermo Fisher Scientific). Reverse transcription was performed with the RevertAid reverse transcriptase (Thermo Fisher Scientific) using 2 μg of total RNA in the presence of 2 U/μl of RNaseOut Recombinant RNase Inhibitor (Thermo Fisher Scientific). qPCR was carried out using BioMaster HS-qPCR SYBR Blue (2×) reagent kit (Biolabmix; http://biolabmix.ru/en/) and CFX96 Real-Time PCR Detection System (Bio-Rad). We used the following thermal cycling conditions: 5 minutes at 95°C, followed by 39 cycles of 15 seconds at 95°C, 30 seconds at 60°C, and 30 seconds at 72°C. Data were collected during each extension phase. Negative control templates (water and cDNA synthesized without reverse transcriptase) were included in each run. Measurements of gene expression were done at least in two biological replicates, each with three (or, in the case of the negative controls, in two) technical replicates. The relative mRNA quantification was determined using the ΔΔCq method. mRNA expression levels were normalized to those of the housekeeping gene RpL32.
A 488-aa portion of Rcd1 (corresponding to amino acids 133–620 of GenPept accession no. NP_610927.3) was expressed as GST-fusion in E.coli and subsequently purified as described in [63]. The purified GST-Rcd1 fusion protein was used to immunize mice. Immunization was performed at the Center for Genetic Resources of Laboratory Animals, Institute of Cytology and Genetics SB RAS. Polyclonal antibodies were affinity purified from serum as reported previously [63].
S2 cells were collected by centrifugation and pellets were lysed in either RIPA buffer (Sigma) containing 1× Halt Protease and Phosphatase Inhibitor Cocktail (Thermo Fisher Scientific) or in lysis buffer (50 mM Hepes-KOH pH 7.6, 1 mM MgCl2, 1 mM EGTA, 1% Triton X-100, 45 mM NaF, 45 mM β-glycerophosphate, 0.2 mM Na3VO4) in the presence of a cocktail of protease inhibitors (Roche). Cell extracts were pelleted at 15,000g for 15 minutes at 4°C and the supernatants were analyzed by Western blotting. Lysates were run on an 8% or a 10% SDS-PAGE and transferred to an Amersham Protran Supported 0.45 μm Nitrocellulose Blotting Membrane (GE Healthcare) by wet or semi-dry transfer. Membranes were blocked for 30 minutes in 2% dry milk in PBT (PBS with 0.1% TritonX-100). Membranes were incubated overnight using following primary antibodies: mouse anti-Rcd1 (1:500, this study), rabbit anti-MBD-R2 (1:1000, Novus Biologicals 49940002), rabbit anti-Wds (1:1000, Novus Biologicals 40630002), rabbit anti-Ndc80 (1:1000; a gift of M. Goldberg, Cornell University), rabbit anti-Cid (1:500; Active Motif 39713), mouse anti-α-tubulin (1:5000, Sigma T6199), rabbit anti-beta-actin, (1:1000, Invitrogen PA5-16914), mouse anti-Lamin Dm0 (1:3500, Developmental Studies Hybridoma Bank ADL67.10) and mouse anti-β-actin-HRP-conjugated (1:5000, Santa Cruz Biotechnology sc-47778 HRP). The non-HRP-conjugated primary antibodies were detected with HRP-conjugated anti-mouse or anti-rabbit IgGs, using either the ECL detection kit (GE Healthcare) or the Novex ECL Chemiluminescent Substrate Reagent Kit (Thermo Fisher Scientific) following the manufacturer’s protocols.
Full-length CS of MBD-R2 (nucleotides (NTs) 123–3629, GenBank accession no. NM_169461.3, but with 1433T>C, 1436G>C and 1893G>A NT substitutions), Rcd1 (NTs 295–3492, GenBank accession no. NM_137083.4), Rcd5 (NTs 101–1834, GenBank accession no. NM_139595.3, but with 748C>A NT substitution), wds (NTs 300–1382, GenBank accession no. NM_080245.5, but with 698T>C and 827C>T NT substitutions) and asl (NTs 98–3079, GenBank accession no. NM_141300.2, but with 376C>G, 2046A>G, 2872C>T and 3041A>G NT substitutions) were cloned in a piggyBac transposon-based plasmid vector upstream of and in frame with the enhanced GFP (for simplicity, referred to as GFP) CS. The plasmids also contained a blasticidin-resistance cassette and the sequence encoding mCherry-αTub84B (Cherry-tubulin) fluorescent fusion protein. The expression of all fluorescent fusion proteins is under the control of the copper-inducible MtnA promoter. S2 cells co-transfected with a plasmid encoding the fluorescent fusion proteins and a plasmid encoding piggyBac transposase were cultured in 39.4 g/L Shields and Sang M3 Insect medium supplemented with 2.5 g/L bacto peptone, 1 g/L yeast extract, 5% heat-inactivated FBS and 20 μg/ml blasticidin (Sigma) for two weeks at 25°C. The antibiotic was then removed from the culture medium. All cells were free from mycoplasma contamination. To induce expression of fluorescent fusion proteins, cells were grown in the presence of different CuSO4 concentrations (0.1, 0.25, 0.4 or 0.5 mM) for 12–14 hours before in vivo analysis or fixation.
All procedures were performed at room temperature. 2×106 S2 cells were centrifuged at 800g for 5 minutes, washed in 2 ml of PBS (Sigma), and fixed for 10 minutes in 2 ml of 3.7% formaldehyde in PBS. Fixed cells were spun down by centrifugation, resuspended in 500 μl of PBS and placed onto a clean side using Cytospin 4 cytocentrifuge (Thermo Fisher Scientific) at 900 rpm for 4 minutes. The slides were immersed in liquid nitrogen, washed in PBS, incubated in PBT (PBS with 0.1% TritonX-100) for 30 minutes and then in PBS containing 3% BSA for 30 minutes. The slides were then immunostained using the following primary antibodies, all diluted in PBT: mouse anti-α-tubulin (1:500, Sigma T6199), rabbit anti-Spd-2 (1:4,000, [21]), rabbit anti-Cid (1:300, Abcam ab10887), rabbit anti-CycB (1:100, [64]), and mouse anti-Rcd1 (1:50, this study). Primary antibodies were detected by incubation for 1 hour with goat FITC-conjugated anti-mouse IgG (1: 30, Sigma F8264) or goat Alexa Fluor 568-conjugated anti-rabbit IgG (1: 350, Invitrogen A11077).
In the attempt to stain GFP-tagged proteins with anti-GFP antibodies, cells expressing these proteins were collected as described above, and fixed (i) for 10 minutes in 2 ml of 3.7% formaldehyde in PBS or (ii) for 15 minutes in 4% paraformaldehyde in PBS, or (iii) for 10 minutes with 8% formaldehyde (methanol-containing) in PBS (Sigma). The slides obtained as described above were then stained with mouse anti-α-tubulin (1:500, Sigma T6199) and either with rabbit anti-GFP (1:200, Invitrogen A11122) or chicken anti-GFP (1:200, Invitrogen PA1-9533), which were detected by Alexa Fluor 568-conjugated goat anti-mouse IgG (1:300, Invitrogen A11031), Alexa Fluor 488-conjugated goat anti-rabbit IgG (1:300, Invitrogen A11034) or Alexa Fluor 488-conjugated goat anti-chicken IgG (1:300, Invitrogen A11039), respectively. These procedures stained the interphase nuclei but did not result in clear immunostaining of the spindle-associated GFP-tagged proteins.
All slides were mounted in Vectashield antifade mounting medium with DAPI (Vector Laboratories) to stain DNA and reduce fluorescence fading. Images were obtained on ZeissAxioImager.M2 using an oil immersion EC Plan-Neofluar 100x/1.30 lens (Carl Zeiss) and captured by 506 mono (D) High Performance camera (Carl Zeiss).
The spindle length in S2 cells was measured with the ZEN 2012 (Carl Zeiss) software, using the “Spline curve” tool and measure function. We considered only cells that did not appear to be polyploid with respect to the basic karyotype of S2 cells. To measure the spindle length in cells at different mitotic stages we drew a freehand line between the two poles along the spindle axis. The data obtained for each spindle type (prometaphase/metaphase; anaphase, telophase and PMLES) were compared using the Wilcoxon Signed-Rank Test and plotted using the BoxPlotR program (http://shiny.chemgrid.org/boxplotr/).
Cells carrying a transgenic construct encoding Cherry-tubulin and either MBD-R2-GFP, Rcd1-GFP, Rcd5-GFP or wds-GFP were grown for 12–14 hours in the presence of different concentrations of CuSO4 (0.1, 0.25, 0.4 or 0.5 mM). 500 μl aliquots of cell suspensions (5×105 cells/ml) were then transferred to cell chambers (Invitrogen A-7816) containing coverslips treated with 0.25 mg/ml concanavalin A (Sigma-Aldrich C0412) placed on the bottom of the chambers. Observations were performed between 20 and 120 minutes after cell plating in the chamber at a Zeiss LSM 710 confocal microscope, using an oil immersion 100×/1.40 plan-apo lens and the ZEN 2012 software.
To estimate the promoter binding by Rcd1/Nsl3 and MBD-R2, we first calculated the genome-wide distributions of these proteins. BAM files with Illumina sequencing reads obtained in ChIP-seq profiling of Rcd1 and MBD-R2 (and the corresponding Input) in S2 cells aligned to the Drosophila melanogaster genome release 5 (dm3) [14] were downloaded from ArrayExpress (https://www.ebi.ac.uk/arrayexpress/) (accession number: E-MTAB-1085). The data were transformed to the BED format using convert2bed in BEDOPS toolkit (version 2.4.35) (http://bedops.readthedocs.io/en/latest/index.html) [65] and genomic positions of aligned reads were converted to Drosophila melanogaster genome release 6 (dm6) using UCSC LiftOver tool (http://genome.ucsc.edu/cgi-bin/hgLiftOver). Next, we divided the genome (only sequences of chromosomes X, 2L, 2R, 3L, 3R and 4 were taken for the analysis) into bins of equal size (100 bp) and counted the number of reads in each bin. Then, we converted the counts in reads per million (RPM) values and calculated log2-transformed MBD-R2/Input and Rcd1/Input ChIP-seq ratios. Only bins with finite values were used for the further analysis (1,124,416 bins for log2(MBD-R2/Input) and 1,125,776 bins for log2(Rcd1/Input)).
To estimate the promoter binding by Nsl1 we first calculated the genome-wide distribution of the protein. Scaled log2(ChIP/Input) microarray-based data for two replicates of Nsl1 in S2 cells [12] were downloaded from GEO (https://www.ncbi.nlm.nih.gov/geo/) (accession number: GSE30991). Genomic positions of ChIP microarray probes were converted from Drosophila melanogaster genome release 4 (dm2) to 6 (dm6) using FlyBase Drosophila Sequence Coordinates Converter (http://flybase.org/convert/coordinates) and the log2(ChIP/Input) ratios from the replicates were averaged.
We next retrieved gene annotation data from Ensembl BioMart release 91 (http://www.ensembl.org/index.html). Promoters were arbitrary defined as gene regions spanning from -1000 to +101 bp relative to the transcription start sites (TSS). To measure the levels of Rcd1, MBD-R2 and Nsl1 binding to promoters, we identified all 100-bins (in the case of Rcd1 and MBD-R2) or microarray probes (in the case of Nsl1) that overlap with each promoter by 1 bp or more. Then, for each promoter, we averaged the log2 ChIP values of such bins or microarray probes. If there were more than one TSS for a gene, their log2 ChIP values were averaged as well. The exact values obtained are reported in S1 Table; in Table 1 we indicate with a “+” symbol all promoters that are enriched in Rcd1, MBD-R2 or Nsl1 ChIP samples compared to the rest of the genome (in nearly all cases, the promoter sequences analyzed are within the 5% of the most Rcd1-, MBD-R2- or Nsl1-enriched genomic sequences).
To measure the effects of Nsl1 deficiency on gene expression, normalized log2-transformed microarray-based data for gene expression in GST-RNAi (control; in three replicates) or nsl1 RNAi (in two replicates) S2 cells [12] were downloaded from GEO (accession number: GSE30991) and the replicates were averaged. Next, we identified microarray probes that belong to each gene, averaged their values, and calculated the percentage of gene expression in Nsl1-depleted cells compared to control.
|
10.1371/journal.pcbi.1000502 | Fast Mapping of Short Sequences with Mismatches, Insertions and Deletions Using Index Structures | With few exceptions, current methods for short read mapping make use of simple seed heuristics to speed up the search. Most of the underlying matching models neglect the necessity to allow not only mismatches, but also insertions and deletions. Current evaluations indicate, however, that very different error models apply to the novel high-throughput sequencing methods. While the most frequent error-type in Illumina reads are mismatches, reads produced by 454's GS FLX predominantly contain insertions and deletions (indels). Even though 454 sequencers are able to produce longer reads, the method is frequently applied to small RNA (miRNA and siRNA) sequencing. Fast and accurate matching in particular of short reads with diverse errors is therefore a pressing practical problem. We introduce a matching model for short reads that can, besides mismatches, also cope with indels. It addresses different error models. For example, it can handle the problem of leading and trailing contaminations caused by primers and poly-A tails in transcriptomics or the length-dependent increase of error rates. In these contexts, it thus simplifies the tedious and error-prone trimming step. For efficient searches, our method utilizes index structures in the form of enhanced suffix arrays. In a comparison with current methods for short read mapping, the presented approach shows significantly increased performance not only for 454 reads, but also for Illumina reads. Our approach is implemented in the software segemehl available at http://www.bioinf.uni-leipzig.de/Software/segemehl/.
| The successful mapping of high-throughput sequencing (HTS) reads to reference genomes largely depends on the accuracy of both the sequencing technologies and reference genomes. Current mapping algorithms focus on mapping with mismatches but largely neglect insertions and deletions—regardless of whether they are caused by sequencing errors or genomic variation. Furthermore, trailing contaminations by primers and declining read qualities can be cumbersome for programs that allow a maximum number of mismatches. We have developed and implemented a new approach for short read mapping that, in a first step, computes exact matches of the read and the reference genome. The exact matches are then modified by a limited number of mismatches, insertions and deletions. From the set of exact and inexact matches, we select those with minimum score-based E-values. This gives a set of regions in the reference genome which is aligned to the read using Myers bitvector algorithm [1]. Our method utilizes enhanced suffix arrays [2] to quickly find the exact and inexact matches. It maps more reads and achieves higher recall rates than previous methods. This consistently holds for reads produced by 454 as well as Illumina sequencing technologies.
| Since the 454 pyrosequencing technology [3] has been introduced to the market, the need for algorithms that efficiently map huge amounts of reads to reference genomes has rapidly increased. Later, high throughput sequencing (HTS) methods such as Illumina [4] and SOLiD (Applied Biosystems) have intensified the demand. The development of read mapping methods decisively depends on specifications and error models of the respective technologies. Unfortunately, little is known about specific error models, and models are likely to change as manufactures are constantly modifying chemistry and machinery. Increasing the read length is a key aim of all vendors — tolerating a trade-off with read accuracy. In a recent investigation on error models of 454 and Illumina technologies, it has been shown that 454 reads are more likely to include insertions and deletions while Illumina reads typically contain mismatches [5],[6]. Currently available read mapping programs are specifically designed to allow for mismatches when aligning the reads to the reference genome. Most of the programs, e.g. MAQ [7], SOAP [8], SHRiMP [9] or ELAND (proprietary), use seeding techniques that gain their speed from pre-computed hash look-up tables. Some of these programs, in particular SOAP and MAQ, are specifically designed to map short Illumina or SOLiD reads. Longer sequences cannot be mapped by these tools. The matching models of MAQ, ZOOM [10], SOAP, SHRiMP, Bowtie [11], and ELAND focus on mismatches and largely neglect insertions and deletions. Indels are only considered during subsequent alignment steps but not while searching for seeds. With indels accounting for more than two thirds of all 454 sequencing errors, this is a major shortcoming for these kinds of reads [5]. Only PatMaN [12] and BWA [13] are able to handle a limited number of indels.
Mapping is aggravated by the manufacturers' overestimation of their read accuracies. While an overall error rate of 0.5% has been observed for 454, the error rate increases drastically for reads shorter than 80 bp and longer than 100 bp [5], leading to considerably larger error frequencies in real-life datasets. This implies that, sequencing projects aiming to find short transcripts such as miRNAs lose a substantial fraction of their data, unless a matching strategy is used that takes indels into account. In Illumina reads, error rates of up to 4% have been observed [6]. This differs significantly from Illumina's specification. Compared to 454, the frequency of indels is significantly lower. Moreover, differences between reads and reference genome might also occur due to genomic variations such as SNPs. We present a matching method that uses enhanced suffix arrays to compute exact and inexact seeds. Sufficiently good seeds subsequently trigger a full dynamic programming alignment. Our method is insensitive to errors and contaminations at the ends of a read including 3′ and 5′ primers and tags. The results section describes the basic ideas and an evaluation of our segemehl software implementing our method. The technical details of the matching model are described in the Methods section at the end of this contribution.
A read aligner should deliver the original position of the read in the reference genome. Such a position will be called the true position in the following. Optimally scoring local alignments of the read and the reference genome can be used to obtain a possible true position, but because an alignment of the read with the reference genome at the true position does not always have an optimal score according to the chosen scoring scheme, this method does not always work. Nevertheless, there are no better approaches available unless further information about the read is at hand.
We present a new read mapping approach that aims at finding optimally scoring local alignments of a read and the reference genome. It is based on computing inexact seeds of variable length and allows to handle insertions, deletions (indels; gaps), and mismatches. Throughout the document the notion of differences refers to mismatches, insertions and deletions in some local alignment of the read and the reference genome, irrespective of whether they arise from technical artifacts or sequence variation. A single difference is either a single mismatch, a single character insertion or a single character deletion. Although not limited to a specific scoring scheme, we have implemented our seed search model in the program segemehl assigning a score of 1 to each match and a score of −1 to each mismatch, insertion or deletion. Our matching strategy derives from a simple and commonly used idea. Assume an optimally scoring local alignment of a read with the reference genome with exactly two differences. If the positions of the differences in the alignment are sufficiently far apart, we can efficiently locate exact seeds which in turn may deliver the position of the optimal local alignment in the reference genome. Likewise, if the distance between the two differences is small, two continuous exact matches at the ends of the read possibly allow to map the read to this position. To exploit this observation, the presented method employs a heuristic based on searches starting at all positions of the read. That is, for each suffix of the read the longest prefix match, i.e. the longest exact match beginning at the first position of the suffix with all substrings of the reference genome is computed. If the longest prefix match is long enough that it only occurs in a few positions of the reference genome, it may be feasible to check all these positions to verify if the longest prefix match is part of a sufficiently good alignment. While this approach works already well for many cases, we need to increase the sensitivity for cases where the computation of the longest prefix match fails to deliver a match at the position of the optimally scoring local alignment. This is the case when a longer prefix match can be obtained at another position of the reference genome by exactly matching characters that would result in a mismatch, insertion or deletion in the optimal local alignment (cf. Fig. 1). Therefore, during the computation of each longest prefix match we check a limited number of differences by enumerating at certain positions all possible mismatches and indels (cf. Fig. 2).
To efficiently compute the longest prefix matches, we exploit their properties for two consecutive suffixes of a read, i.e. for two suffixes starting at position i and i+1. If the suffix starting at position i has a longest prefix match of length ℓ, then the suffix starting at position i+1 has a longest prefix match of length at least ℓ−1. For example, assume a read ACTGACTG. If the second suffix has a longest prefix match of length 4, i.e. CTGA, with the reference genome, we immediately see that the third suffix has a longest prefix match not shorter than 3—because we already know that the substring TGA exists in the reference genome. Using an enhanced suffix array of the reference sequence, we can easily exploit this fact and determine the longest prefix match of the next suffix without rematching the first ℓ−1 characters. Likewise, the enumeration of mismatches and indels is also restricted to the remaining characters of the suffix in our model.
For each suffix of a read, we thus obtain a set of exact matches and alternative inexact matches and their respective positions in the reference sequence. These exact and inexact matches act as seeds. If a seed occurs more than t times in the reference genome, then it is omitted, where t is a user specified parameter (segemehl option –maxocc). The heuristics rigorously selects the exact or inexact seed with the smallest E-value, computed according to the Blast-statistics [14]. If this E-value is smaller than some user defined threshold (segemehl option -E), the bitvector algorithm of [1] is applied to a region around the genomic position of the seed to obtain an alignment of the read and the reference sequence. While the score based search for local alignment seeds controls the sensitivity of our matching model, the bitvector alignment controls its specificity: if the alignment has more matching characters than some user specified percentage a of the read (segemehl option -A) the corresponding genomic position is reported (see Methods).
The computation of the longest prefix match is implemented by a top-down traversal of a conceptual suffix interval tree, guided by the characters of the read. The suffix interval tree is equivalent to a suffix trie (see Methods). The traversal delivers a matching stem. Note that for the DNA alphabet there are at most four edges outgoing from each node of the suffix interval tree. To introduce mismatches, the traversal is simply continued with alternative edges, i.e. edges diverging from the matching stem. To introduce insertions, the traversal is not regularly continued, but characters of the read are skipped. Deletions are simulated by skipping nodes of the suffix interval tree and continuing the search at their child nodes (see Methods). We refer to these alternative paths that branch off from the matching stem as branches. The maximum number of branches to be considered is controlled by the seed differences threshold k (segemehl option -D). Note, that while matching character by character along a suffix of a read, the number of branches is expected to decrease quickly.
segemehl constructs indices either for each chromosome of a genome and the matching is performed chromosome-wise or, depending on the available RAM, chromosomes are combined to larger sequences. Compared to other methods, the index structure used by segemehl is significantly larger. For example, the enhanced suffix array of human chromosome 1 occupies approximately 3 GB of space. As it is stored on disk, the index only needs to be computed once. The construction of the index requires linear time. For example, on a single CPU, the construction of the complete enhanced suffix array for human chromosome 1 takes approximately 15 minutes. For our comparison, we ran segemehl with maximum occurrence parameter t = 500. The maximum E-value for seeds was set to 0.5 and minimum identity threshold to a = 85% which corresponds to a maximum of ⌈0.15·m⌉ differences in an alignment of the read of length m.
We compared segemehl to Bowtie v0.9.7 with option –all, BWA v0.2.0, MAQ v0.7.1, PatMaN v1.2.1 and SOAP v1.11 with option –r 2. MAQ and SOAP are based on ungapped alignments which are computed by hash lookups [7],[8],[13]. Due to length restrictions, MAQ is limited to Illumina (and SOLiD) reads. It additionally takes quality scores into account. The quality values needed by MAQ were, for all nucleotides, uniformly set to a value corresponding to the error rate. Bowtie [11] and BWA [13] index the reference genome with the Burrows-Wheeler transform. BWA allows a limited number of indels. PatMaN [12] matches the reads by traversing a non-deterministic suffix automaton constructed from the reference genome. Except for PatMaN, all programs only report matches with the smallest edit distance. BWA and Bowtie each need about 10 minutes to build their index. The fastq files needed by MAQ are built in approximately 2 minutes. PatMaN and SOAP require no indexing steps. The options for the other programs were chosen so as to achieve results similar to segemehl. For our comparison, we performed tests on simulated as well as real-life read data sets. For the simulation we generated read sets representing different error rates, types and distributions. We used three distinct error sets, one containing only mismatches, one containing only indels and a last one representing reads with mismatches and indels at a ratio of 1∶1. Additionally, different error distributions were used to model error scenarios such as terminal contamination (e.g. linker, poly-A tails) or decreasing read quality. We chose uniform, 5′, 3′ and terminal error distributions.
Each simulated dataset contained 500 000 simulated reads, each of length 35 bp, sampled from a 50 MB large region of the human genome (chromosome 21). We introduced errors to each simulated read according to previously defined rates, error types and distributions. For the 50 MB region we constructed the indexes required for segemehl and Bowtie. For MAQ we constructed the index for the read set under consideration. Index construction took approximately one minute for Bowtie and BWA. The construction for the enhanced suffix array for segemehl took 3.5 minutes. The binary fastq files for MAQ were created in about 20 seconds.
We ran segemehl with seed differences threshold k = 0 and k = 1. For k = 0, only exact seeds are computed and for k = 1 seeds with at most one difference are computed. All programs were executed single-threaded on the same machine. The results for a uniform error distribution for mismatches only as well as for mismatches and indels are shown in Fig. 3. We measured the performance in terms of running time (Fig. 3 (A)) and recall rates, i.e. the percentage of reads mapped to the correct position. segemehl has recall rates of more than 95% (k = 1) and 80% (k = 0) in each setup with not more than two errors in the reads. With four uniformly distributed errors in the reads, the recall rate drops below 80% (k = 1) and 50% (k = 0), respectively. Hence, for k = 1 segemehl outperforms all other methods in terms of recall rates. For reads containing only mismatches and k = 0, segemehl is comparable to other methods (Fig. 3 (B)) while it has a significantly better recall rate as soon as insertions and deletions are involved (Fig. 3 (C)). As expected, the recall rate of most short read aligners drops if insertions and deletions are introduced into the reads. The running time of segemehl for k = 0 is comparable to other short read aligners. For k = 1, the running time increases by a factor of 10.
In contrast to Bowtie, BWA, MAQ, and SOAP, segemehl reports, by default, multiple matches for a read within the reference genome if the corresponding alignments have an E-value smaller than some user defined threshold. This behavior leads to an increase in the running time and a decrease in specificity. Compared to PatMaN, which is also able to report multiple matches, segemehl can cope with more than two differences and still is on average faster by a factor of 1.7 (k = 1) and 14 (k = 0). As expected, the worst segemehl results are seen for high error rates with a uniform error distribution (Fig. 4). Terminal, 3′ and 5′ error distributions yield better results, suggesting that segemehl implements a robust method that is insensitive to leading and trailing contaminations. Next, we compared segemehl, Bowtie and MAQ on two real-life data sets. We used Bowtie with option –all and MAQ with option –C 513 as suggested in the manuals to achieve maximum sensitivity. segemehl's sensitivity was controlled by option –M 500 to omit all seeds occurring more than 500 times in the reference sequence.
The data set ERR000475 of 20 million Illumina reads (length 45) for H. sapiens was downloaded from the NCBIs Short Read Archive (http://www.ncbi.nlm.nih.gov/Traces/sra/). The second data set comprised about 40 000 short 454 reads from the arabidopsis mpss plus database (http://mpss.udel.edu/at/). The average length of the 454 reads was 23 bp.
We partitioned the 454-set into subsets of equal size, to satisfy input requirements for MAQ. An average quality value was assigned to each base.
Mapping multiple reads to a reference genome is a task which can easily be parallelized. Like all other methods, segemehl offers a parallelization option to run the program on multiple cores. segemehl runs for the ERR000475 dataset were carried out in eight parallel threads on a single machine with two Quadcore CPUs and 16GB of RAM. Seven enhanced suffix arrays were constructed representing the whole human genome. segemehl mapped 92% of the reads to the reference sequence while MAQ mapped 85% without and 89% with quality values. The corresponding values for Bowtie are 81% and 89%. The largest difference between the three tools is for the total number of exact matches. Although MAQ was, according to the manual, running in maximum sensitivity mode, segemehl computes 20 times more matches than MAQ (Tab. 1 (a)). Bowtie reports 2.5 billion matches which is much more than the two other tools. As expected, for the 454-set, the difference among the compared programs is even larger. While Bowtie is able to map 71% of all reads, segemehl achieves 95%. MAQ, a program explicitly designed for Illumina reads, matches 79% of the reads. Interestingly, compared to Bowtie, MAQ reports more matches with two mismatches. segemehl mainly achieves this result by mapping more reads with one or two errors. In fact, by allowing insertions and deletions segemehl doubles the number of reads matched at the unit edit distance of 1 (Tab. 1 (b)).
We have presented a novel read mapping approach that is able to efficiently handle 3′ and 5′ contaminations as well as mismatches, insertions and deletions in short and medium length reads. It is based on a matching model with inexact seeds containing mismatches, insertions and deletions. The sensitivity and specificity of our method is controlled by a maximum seed differences threshold, a maximum occurence threshold, an E-value threshold and an identity threshold. Compared to previous methods, our approach yields improved recall rates especially for reads containing insertions and deletions. Since indels have been reported to be the predominant error type in 454 reads, allowing for indels is most important to achieve a correct mapping. While PatMaN, by default, fully enumerates all matches with up to two differences, segemehl's heuristic reports only best-scoring matches. The price for the gain in sensitivity is an increase in running time: with k = 1 our method is approximately ten times slower than Bowtie, the fastest program in our comparison. As we used enhanced suffix arrays, matching against a large mammalian genome has to be done chromosome by chromosome when off-the-shelf hardware is used. However, the gain in sensitivity for reads with mismatches and the failure of other methods when dealing with indels may be, depending on the users demands, a reasonable trade off for these shortcomings. Our method is not limited to a specific technology or read length. Although quality values are not considered yet, the matching strategy can easily be adapted to evaluate low quality bases specifically. In principle, we show that for k = 0, i.e. exact seeds, our method is sufficiently sensitive to map reads with up to two differences. This is an interesting result since most of the current methods do not tolerate insertions and deletions. In summary, segemehl with k = 0 is among the fastest mapping algorithms. For k = 1, segemehl is able to achieve good recall rates beyond the two error barrier. This is especially interesting since manufacturers try to increase their read lengths at the cost of higher error rates. The increased sensitivity of the presented matching model, along with its ability to handle leading and terminal contaminations is a trade off for the large memory requirements of the enhanced suffix arrays. In the future, compressed index structures like the FM-index [15] may be a suitable framework to implement our matching model with smaller memory requirements.
Our strategy, based on enhanced suffix arrays, aims to find a best local alignment of short reads and reference sequences with respect to a simple scoring system. It does so by determining, for each suffix of the read, the longest prefix occurring as a substring in the reference sequence. This gives a matching backbone, from which a limited number of branches are derived by mismatches, insertions and deletions (Fig. 2). The concept of a matching backbone is equivalent to the concept of matching statistics introduced in [16]. We introduce the concept of matching backbone and branches via a conceptual tree of suffix intervals. Our heuristic approach delivers a small number of inexact seeds of variable length that are subsequently checked by the bitvector algorithm of Myers [1] to verify the existence of alignments with a limited number of differences. First, a short introduction to the basic notions for sequence processing and enhanced suffix arrays will be given, before the concept of suffix intervals is defined. Subsequently, we introduce our new matching strategy.
We consider sequences over the DNA alphabet ΣDNA = {A, C, G, T, N}, where N denotes an undetermined base. In our approach the alignment of N with any character, including N itself, results in a mismatch.
Consider the suffix interval [l‥r, h] for w. A child of [l‥r, h] is a suffix interval [l′‥r′, h+1] satisfying l≤l′≤r′≤r. We call [l′‥r′, h+1] the a-child of [l‥r, h] if there is a character such that [l′‥r′, h+1] is the suffix interval for . Note that for all q, l′≤q≤r′, we have = Ssuf[q][h]. Hence we can easily determine from [l′‥r′, h+1] or split [l‥r, h] into its children. A method computing the a-child of a suffix interval in constant time is described in [2].
Let be a suffix interval. For the empty sequence ε we define . For any character and any sequence u we recursively define
That is, delivers the interval , obtained by greedily matching the characters in v beginning at the suffix interval and q is the length of the matching prefix of v.
Let P denote a sequence of length m neither containing a wildcard symbol N nor the sentinel $. For any i, 0≤i≤m, Pi = P[i‥m−1] denotes the suffix of P beginning at position i. Let ℓi be the length of the longest prefix of Pi occurring as a substring of S. Then P[i‥i+ℓi−1] occurs in S and either i+ℓi = m or P[i‥i+ℓi] does not occur in S. Moreover, there is a sequence of suffix intervals , such that for all q, 0≤q≤ℓi, is the suffix interval for P[i‥i+q−1]. This implies that . We call a matching stem. Obviously, for any i, 0≤i≤m, . For any i, 0≤i≤m and any q, 1≤q≤ℓi, is the -child of where = S[t+q−1] for any . (Note that all suffixes in have the common prefix P[i‥i+q−1] and is the last character of this prefix.) The ℓi-values are determined in the same way as the length-values of the matching statistics, introduced in [16]. Using the suffix link table, the ℓi-values can be computed in O(m) time altogether (cf. [2]).
We now consider the relation of matching stems of two neighboring suffixes Pi−1 and Pi for some i>0. First note that ℓi−1≤ℓi+1. Moreover, for each q, 1≤q≤ℓi−1 we have where = {x+y | x∈M} denotes the elementwise addition for any set M. That is, any suffix in can be found in with offset one.
To allow differences in our matching heuristic, we introduce the concept of matching branches which branch off from sets of the matching stem. We describe the branching in terms of a transformation of some suffix interval .
Let i, 0≤i≤m−1 be arbitrary but fixed. Let q be such that i+q−1<m. Consider some suffix interval such that the unit edit distance of S[suf[l]‥suf[l]+h−1] and P[i‥i+q−1] is exactly d≤k. Then, for the edit operations x∈{MM, I, D}, we define the matching branch as follows:
Any computation of a triple (, q′, d+1) according to these equations is called branching step. The MM-branching step implies a mismatch of a≠P[i+q] (in the reference sequence) with P[i+q] (in the read). The I-branching step implies an insertion of character P[i+q] in the read. The D-branching step implies a deletion of character a∈ΣDNA in the read.
Note that in case some a-child of does not exist, there is no corresponding contribution to the matching branch. We combine the different types of matching branches by defining:
Obviously, any element in can itself be extended by branching from it. To define this, we introduce for all j≥1 the iterative matching branch as follows:
This gives us the matching branch closure , defined by
That is, is the set of matching branches that can be derived by one or more branching steps from (, q, d) (Fig. 6). Of course, since each step increases the difference value d, the number of steps is limited by k – d. Each element is extended by exactly matching P[i+q′‥m−1] against the enhanced suffix array beginning at the suffix interval . That is, we compute .
While we have defined matching branches for any element in a matching stem, we only compute them for a few elements of the matching stem which make up the matching backbone: Let , where is defined by
Thus, for each suffix i, is the position in P from which to continue processing the next suffix. For any , we compute . That is, we omit computing for . This is due to the fact that some of the suffixes in are already included (with offset one) in , see equation (1). All in all, we arrive at a set Q(P, k) of 4-tuples (i, , q, d) such that the unit edit distance of P[i‥i+q−1] and w is d≤k and is the suffix interval for w. The Figure 7 gives pseudocode for computing Q(P, k) (which includes the matching backbone).
Turning to the analysis of the algorithm, first note that
That is, the matching backbone contains at most m+1 elements and thus the statements in the inner loop of the algorithm (Fig. 7) are executed O(m) times altogether. Obviously contains up to 5 elements, contains at most 1 element and contains at most 6 elements. Since there can be k iterations when computing , the size of this set is at most (12)k. Hence the total number of all matching branches is (m+1) · (12)k. Each matching branch is generated from a previously generated element in constant time. Hence the algorithm runs in time proportional to (m+1) · (12)k.
From the matching backbone and from the set of all matching branches we select an element achieving a maximum score according to a simple scoring scheme where a character match scores +1 and a mismatch, an insertion and a deletion scores −1. The maximum score element (i, [l‥r, h], q, d) defines a set of substrings of S which are aligned to P. More precisely, for any j, l≤j≤r, P is matched against the reference substring S[suf[j]−(i+k)‥suf[j]+(m−i+k)] using the bit vector algorithm of Myers [1]. For this, we allow a maximum number of differences, according to the the identity threshold . Myers algorithm runs in O(m/ω · ℓ) time where ℓ = 2k+m+1 is the length of the reference substring and ω is the word size of the machine. As ω = 64 in our implementation, for reads of size up to 64, we have m/ω = 1 and so the algorithm runs in O(m+k) time. Note that this running time is independent of a. In summary, by specifying k along with some E-value [14] we set the thresholds to search for local alignment seeds. Subsequently, we use Myers algorithm to discards all seeds that produce poor semi-global alignments, according to parameter a, typically loosely set to values around 80% (which corresponds to ).
|
10.1371/journal.pbio.1000300 | Endogenous Nmnat2 Is an Essential Survival Factor for Maintenance of Healthy Axons | The molecular triggers for axon degeneration remain unknown. We identify endogenous Nmnat2 as a labile axon survival factor whose constant replenishment by anterograde axonal transport is a limiting factor for axon survival. Specific depletion of Nmnat2 is sufficient to induce Wallerian-like degeneration of uninjured axons which endogenous Nmnat1 and Nmnat3 cannot prevent. Nmnat2 is by far the most labile Nmnat isoform and is depleted in distal stumps of injured neurites before Wallerian degeneration begins. Nmnat2 turnover is equally rapid in injured WldS neurites, despite delayed neurite degeneration, showing it is not a consequence of degeneration and also that WldS does not stabilize Nmnat2. Depletion of Nmnat2 below a threshold level is necessary for axon degeneration since exogenous Nmnat2 can protect injured neurites when expressed at high enough levels to overcome its short half-life. Furthermore, proteasome inhibition slows both Nmnat2 turnover and neurite degeneration. We conclude that endogenous Nmnat2 prevents spontaneous degeneration of healthy axons and propose that, when present, the more long-lived, functionally related WldS protein substitutes for Nmnat2 loss after axon injury. Endogenous Nmnat2 represents an exciting new therapeutic target for axonal disorders.
| In a normally functioning neuron, the cell body supplies the axon with materials needed to keep it healthy. This complex logistical activity breaks down completely after injury and often becomes compromised in neurodegenerative diseases, leading to degeneration of the isolated axon. Whilst there are probably many important cargoes delivered from the cell body that isolated axons cannot exist without indefinitely, proteins that are short-lived will be depleted first, so loss of these proteins is likely to act as a trigger for degeneration. Using clues from a mutant mouse whose axons are protected from such degeneration, we have identified delivery of Nmnat2, a protein with an important enzyme activity, as a limiting factor in axon survival. Importantly, Nmnat2 is very labile and its levels decline rapidly in injured axons before they start to degenerate. Even uninjured axons degenerate in a similar way without it. These properties are consistent with loss of Nmnat2 being a natural stimulus for axon degeneration, and it might therefore be a suitable target for therapeutic intervention.
| The endogenous molecular trigger for Wallerian degeneration remains unknown. Recent progress towards understanding how the slow Wallerian degeneration fusion protein (WldS) delays degeneration of injured and sick axons has not addressed this wider question [1]–[8], and this aberrant protein is only expressed in a few strains of mouse, rat, and fly. Knowledge of the normal regulation of axon survival in wild-type animals should not only lead to greater mechanistic insight but could also have important therapeutic implications for axon protection since pharmacological manipulation of endogenous processes is likely to be more achievable than overexpression of exogenous proteins.
Many stresses that induce Wallerian or Wallerian-like degeneration involve a partial or complete block of axonal transport. Since transport is bi-directional, degeneration could be triggered by failed anterograde delivery of essential survival factors or by failed removal of harmful substances by retrograde transport. Defective anterograde transport seems more directly associated with axon loss than dysfunctional retrograde transport [9]–[13]. Therefore, extending an old model [14], we propose a “survival factor delivery hypothesis” of axon degeneration. We suggest that axon integrity requires continuous anterograde delivery of one or more labile, cell body–synthesized survival factors. Other axonal components should be dispensable, more long-lived, or synthesized locally. Once supply is disrupted, following injury or other insult, levels of the limiting survival factor(s) will drop below a critical threshold due to natural turnover, activating an intrinsic axon degeneration program.
This model has several attractions. First, the initial latent phase of Wallerian degeneration [14],[15] would reflect the rate of survival factor turnover before the critical threshold is reached. Second, altered turnover would explain how low temperature and proteasome inhibition extend this latent phase [16]–[18]. Finally, redistribution of the remaining survival factor(s) in the distal stump by axonal transport could underlie the progressive nature of Wallerian degeneration [14],[19],[20].
No such endogenous survival factor has been identified, but the WldS protective mechanism offers important clues. WldS contains the N-terminal 70 amino acids of multiubiquitination factor Ube4b fused, in frame, to NAD+ synthesizing enzyme Nmnat1 [21]. Both regions are required for full WldS function in vivo [4],[5]. The N-terminal VCP binding region probably targets the essential Nmnat activity to a specific subcellular location [2],[5]. Despite being predominantly nuclear [6], recent studies indicate a cytoplasmic and potentially axonal site of action for WldS [2],[3],[7],[8], rekindling interest in its relationship to the earlier model of a putative endogenous survival factor(s) [14]. WldS, an aberrant protein, cannot be one of these factors, but Nmnat1 [22] and the other mammalian Nmnat isoforms (Nmnat2 and Nmnat3 [23],[24]) are candidates since they all possess the same critical enzyme activity and can all delay axon degeneration in primary neuronal culture when expressed exogenously at high levels [1],[25],[26]. Only Nmnat3 has so far been shown to confer robust protection to injured axons in vivo when wild-type proteins (except for a tag used for detection) are overexpressed [2],[4],[8].
In support of the “survival factor delivery hypothesis” we show that briefly suppressing protein synthesis in cell bodies of uninjured primary neuronal cultures induces Wallerian-like degeneration. The ability of a single protein (WldS) to block this suggests that only one or a few critical factors are directly involved. We hypothesized that WldS substitutes for one or more mammalian Nmnat isoforms, so we compared their properties against those predicted for a critical axon survival factor. We reasoned that depletion should trigger Wallerian-like degeneration without injury, its natural half-life should be consistent with the latent phase of Wallerian degeneration (WldS should be much more stable to extend this period), the survival factor should be degraded by the proteasome to explain why proteasome inhibition extends axonal survival, it should be present in axons, and it should significantly prolong injured axon survival when highly overexpressed (to outweigh its short half-life). Nmnat2 uniquely fits this profile, indicating that its depletion after injury is a trigger for Wallerian degeneration and that “dying-back” pathology is likely to reflect defects in Nmnat2 axonal transport or synthesis.
Our main hypothesis predicts that blocking synthesis of one or more putative axon survival factors should trigger Wallerian-like degeneration without injury, similar to that induced by blocking axonal transport [27],[28]. To test this we initially inhibited all protein translation in mouse superior cervical ganglia (SCG) explant cultures, using two unrelated inhibitors, cycloheximide (CHX) and emetine, to rule out nonspecific effects. One µg/ml CHX, which suppresses global protein synthesis by more than 95% [29],[30], not only stopped neurite outgrowth as expected [30],[31] but also induced widespread blebbing of distal neurites (Figure 1A and 1C). Ten µg/ml CHX or 10 µM emetine caused more rapid and extensive blebbing of neurites, presumably due to more complete suppression of protein synthesis, followed by fragmentation and detachment shortly afterwards (Figure 1A and 1C), similar to the degeneration of transected neurites. To test whether the degeneration is Wallerian-like, we used cultures from slow Wallerian degeneration (WldS) mice and found a delay of over 48 h (Figure 1B and 1C). Similar results with rat SCG cultures and mouse dorsal root ganglion (DRG) cultures indicate that these events are not restricted to one species or neuron type (Figure S1). Delayed degeneration in WldS cultures after inhibition of translation also shows that local translation of mRNAs in neurites is unlikely to underlie WldS-mediated axon protection as hypothesized previously [32]. Similarly, localized translation is not required in injured neurites for WldS-mediated protection, and it is also not needed for Wallerian degeneration itself (Figure S2).
Rapid cleavage of neurofilament heavy chain (NF-H) is an early molecular change that occurs as injury-induced Wallerian degeneration is initiated after the latent phase both in vitro and in vivo [6],[18]. We found that this also occurs after protein synthesis suppression in wild-type cultures but not in WldS cultures (Figure 1D). Thus, molecular assays also indicate this degeneration is Wallerian-like.
Importantly, degeneration induced by protein synthesis suppression is not due to loss of neuronal viability but is a much earlier event independent of cell death. Even 7 d after treatment with 1 µg/ml CHX, long after complete degeneration of neurites, many SCG cell bodies retain the ability to re-grow neurites when this reversible inhibitor is removed (Figure 1E). Most cell bodies in 7-d CHX-treated dissociated cultures also excluded Trypan Blue, further indicating neuron viability (unpublished data).
To test directly whether a critical axon survival factor(s) has to be synthesized and delivered from cell bodies, we used compartmented cultures where distal neurites can be treated separately from cell bodies and proximal neurites (Figure 2). Neurites degenerated only when inhibitors were applied to the compartment containing neuronal cell bodies and proximal neurites. Consistent with a previous report [33], translation inhibitors applied only to distal neurites caused no significant degeneration within this timeframe. Indeed, neurites continued to grow (unpublished data). Thus, suppression of protein synthesis in the cell body triggers Wallerian-like neurite degeneration, providing strong support for the survival factor delivery hypothesis and suggesting the survival factor(s) is proteinaceous.
We then investigated the molecular basis of these findings. Because Nmnat1 contributes essential Nmnat enzyme activity to the WldS fusion protein, we reasoned that WldS might protect axons by substituting for injury-induced loss of an endogenous Nmnat activity. Transcripts of all three mammalian Nmnat isoforms are expressed in mouse SCG neurons (Figure S3 and [26]), suggesting each is a reasonable candidate. Moreover, although their predominant localizations are nuclear (Nmnat1), Golgi-associated (Nmnat2) and mitochondrial (Nmnat3) [34], the recent finding that WldS acts at a non-nuclear site despite its nuclear abundance [3] reminds us that low levels of protein can act elsewhere, especially if enzyme activity amplifies the effect. We therefore decided to test whether any of the Nmnat isoforms possess the predicted properties of an endogenous axon survival factor in our model.
The first key prediction is that survival factor depletion should induce Wallerian-like neurite degeneration without injury as levels drop below a critical threshold. We used pools of siRNAs (siNmnat1, 2, or 3) to knock down expression of the murine Nmnat isoforms and confirmed specificity for the appropriate isoform by assessing their ability to prevent expression of N-terminal FLAG-tagged Nmnat (FLAG-Nmnat) proteins in transfected HEK 293T cells and SCG neurons (Figure 3).
To assess the effect of Nmnat isoform knock-down in SCG neurons, we used a microinjection-based strategy (see Figure S4), enabling us to consistently introduce similar amounts of siRNA. Neurons in wild-type dissociated cultures were first injected with each siRNA pool, with DsRed2 expression allowing visualization of injected neurons and their neurites. Of the three Nmnat siRNA pools, only injection of siNmnat2 caused a significant reduction in the percentage of healthy neurites compared to the non-targeting siRNA pool (siControl) (Figure 4A and 4B). Some of the neurites of the siNmnat2-injected neurons already appeared abnormal 24 h after injection, when the entire lengths of the DsRed2-labeled neurites could first be clearly visualized, and almost all showed abnormal morphology or had completely degenerated 72 h after injection. In contrast, injection of siControl, siNmnat1, and siNmnat3 all caused relatively little degeneration (Figure 4A and 4B), and neurites continued to grow (unpublished data). Combined injection of all three Nmnat siRNA pools did not significantly accelerate neurite degeneration relative to siNmnat2 alone (Figure 4C). Thus, Nmnat2 knock-down is sufficient to induce neurite degeneration, whereas knock-down of the other Nmnat isoforms has no clear effect on neurite survival. To confirm that the siNmnat2-induced neurite degeneration is Wallerian-like, we microinjected WldS neurons with siNmnat2 and found degeneration was completely blocked for at least 72 h (Figure 4A and 4D).
To rule out a contribution from any off-target effect of the four individual siRNAs within the siNmnat2 pool, we tested whether they could cause neurite degeneration when injected individually or in non-overlapping sub-pools (Figure S5). One siRNA alone (J-059190-11) and two others in combination (J-059190-10 and J-059190-12) triggered significant neurite degeneration that was similar to that induced by the complete pool. A clear combinatorial effect was also seen as J-059190-11 injected at the concentration it contributes to the siNmnat2 pool caused significantly less neurite degeneration than the pool itself. Together, these observations show that siNmnat2-induced neurite degeneration is due to knock-down of Nmnat2.
The siNmnat2-induced neurite degeneration is distinctive, characterized by the appearance of multiple neuritic DsRed2-containing swellings and a distal-to-proximal “dying-back” progression that appears to be independent of neuronal viability (Figure 4E and 4F). In contrast, the small amount of background neurite degeneration seen with all the siRNA pools (including siControl) coincides with cell death and is faster and morphologically distinct (Figure 4G).
Some loss of neuronal viability occurred in these experiments, irrespective of the siRNA injected, but a small, additional decrease in neuronal viability following siNmnat2 knock-down was also apparent (Figure S6). Even though this reduction in neuronal viability, relative to siControl, was proportionately much smaller than the reduction in neurite survival (Figure S6F), we sought to completely exclude the possibility that cell death might be responsible for the siNmnat2-associated neurite degeneration. We were able to almost completely eliminate neuronal cell death in the siNmnat2 injection experiments in two ways (Figure 5). First, we reduced expression of the fluorescent marker after finding that toxicity was causing the (caspase-independent) background cell death. Second, we found that the small siNmnat2-associated decrease in neuronal viability could be prevented by the pan-caspase inhibitor z-VAD-fmk (Figure 5A), indicating that this death is caspase-dependent. Importantly, the amount of siNmnat2-induced neurite degeneration was unchanged when cell death was reduced in these ways (compare Figure 5C and 5D to Figures 4A, 4B, and S6). This clearly shows that neurite degeneration precedes any associated loss of neuron viability in these experiments. It is also consistent with WldS-mediated protection of neurites (Figure 4D) being able to reduce siNmnat2-associated neuronal loss to control levels (Figure S6C), despite the fact that WldS cannot directly prevent neuronal cell death in SCG cultures [35]. In addition, failure of z-VAD-fmk to prevent siNmnat2-induced neurite degeneration provides further evidence that it is Wallerian-like as Wallerian degeneration has been shown to be unaffected by a range of anti-apoptotic interventions [36]–[38].
Thus, constitutive expression of endogenous Nmnat2 in SCG neurons is required to prevent spontaneous “dying-back” Wallerian-like neurite degeneration. Importantly, these data also indicate that endogenous Nmnat1 and Nmnat3 cannot compensate for loss of Nmnat2, despite the ability of these proteins to protect injured neurites when sufficiently overexpressed [1],[25].
In our model, axon degeneration is initiated when survival factor levels drop below a critical threshold after synthesis or delivery is blocked. If Nmnat2 depletion acts as a trigger for Wallerian degeneration, Nmnat2 half-life should be compatible with the short latent phase of 4–6 h before transected SCG neurites degenerate. WldS, on the other hand, should be more stable to directly substitute for loss of endogenous Nmnat2. A direct comparison of the relative turnover rates of the FLAG-tagged murine Nmnat isoforms and WldS in co-transfected HEK 293T cells (Figure 6A) showed that FLAG-tagged Nmnat2 is turned over rapidly when protein synthesis is blocked with an in vitro half-life of less than 4 h. In contrast, there was minimal turnover of FLAG-tagged WldS, Nmnat1, and Nmnat3 up to 72 h. Similar results were also obtained with C-terminal FLAG-tagged proteins (unpublished data). We also found that proteasome inhibition with MG-132 largely prevented turnover of FLAG-tagged Nmnat2 in these cells for at least 24 h (Figure 6B). Importantly, turnover of endogenous Nmnat2 in SCG explants following protein synthesis inhibition was found to be similarly rapid (Figure 6C).
The half-life of Nmnat2 is also consistent with the time when wild-type SCG neurites become committed to degenerate after inhibition of translation (Figure S7A). Neurites exposed to CHX for just 4 h remain healthy and continue to grow for over 5 d, but they become irreversibly committed to degenerate when exposed to CHX for just 8 h, despite only minimal evidence of degeneration when CHX is removed. Intermediate treatment for 6 h gave a mixed outcome. This suggests that degeneration of these neurites can be prevented by reestablishing synthesis of the labile survival factor(s) providing levels have not dropped below a critical threshold. The precise threshold can only be determined when the duration of downstream events leading to activation and execution of degeneration are better understood. Importantly, WldS expression not only delays the onset of neurite degeneration following protein synthesis suppression, it also delays their commitment to degenerate at least 3-fold (Figure S7B).
Therefore, the half-life of Nmnat2, but not Nmnat1 and Nmnat3, is compatible with its turnover being a trigger for Wallerian degeneration. Furthermore, the longer half-life of WldS is consistent with it substituting for Nmnat2 loss for a prolonged period.
According to our model, the putative axon survival factor should also be present in neurites under normal conditions, and its level in transected neurites should drop significantly prior to initiation of degeneration at 4–6 h. Therefore, we assessed Nmnat2 levels in neurite-only extracts from SCG explant cultures at the time of transection and 4 h afterwards when the gross morphology of the transected neurites still appears relatively normal (Figure 7A). Neurite extracts contained significant amounts of Nmnat2 at the time of transection and this fell to ∼30% of steady-state levels within 4 h. Furthermore, loss of endogenous Nmnat2 occurs before cleavage of NF-H, which accompanies physical break-down of SCG neurites after injury [18] or protein synthesis suppression (Figure 1D), and before β-Tubulin degradation. An increase in Nmnat2 levels in the corresponding cell body/proximal neurite extracts 4 h after separation of their transected distal neurites is also seen. This probably represents accumulation of Nmnat2 in a greatly reduced cellular volume (see Discussion).
Proteasome inhibition modestly extends the latent phase of Wallerian degeneration in SCG explant cultures [18], so we tested whether this correlates with reduced turnover of endogenous Nmnat2 given that FLAG-tagged Nmnat2 is degraded via the proteasome in HEK cells (Figure 6B). Neurites treated with the proteasome inhibitor MG-132 appear relatively normal 8 h after transection, with no associated NF-H cleavage, whereas untreated neurites show extensive physical and molecular signs of degeneration (Figure 7B). We found that loss of Nmnat2 was also significantly reduced by MG-132 at this time (Figure 7B), consistent with depletion of endogenous Nmnat2 being a critical trigger for axon degeneration. The fact that Nmnat2 turnover was not completely prevented might explain why the duration of neurite protection by MG-132 is fairly limited [18], although prolonged proteasome inhibition is also toxic to axons [39].
Nmnat2 loss within 4 h in transected wild-type neurites seems unlikely to be a consequence of axon degeneration, as cytoskeletal proteins and neurite morphology are little altered at this time point (Figure 7A). However, to rule this out conclusively, we assessed Nmnat2 turnover in transected WldS neurites (Figure 7C), which do not degenerate for several days. Nmnat2 levels in WldS neurites fell with a remarkably similar time course to those in wild-type neurites. In contrast, cleavage of NF-H was prevented, showing that proteins that degrade as a consequence of degeneration are stabilized in WldS neurites. As predicted, WldS levels in neurites also remained relatively constant. Indeed, levels of WldS protein are only moderately reduced in neurites 48 h after transection (Figure S8).
Thus, Nmnat2 is rapidly depleted in distal stumps of injured neurites, as a result of natural turnover rather than a consequence of degeneration. This is consistent with Nmnat2 loss triggering Wallerian degeneration. The continued presence of WldS in transected WldS neurites long after Nmnat2 is lost shows that WldS does not act by stabilizing Nmnat2 but instead supports a model in which WldS substitutes for the functionally related Nmnat2.
We also found that an Nmnat2–enhanced green fluorescent protein (eGFP) fusion protein localizes to SCG neurites in highly defined particles shortly after being expressed (Video S1 and Figure 7D). In contrast, eGFP alone showed uniform distribution in neurites (unpublished data). Particles containing Nmnat2-eGFP travel bi-directionally, but the majority move in an anterograde direction (72.2%±3.8% based on particle movements in 18 neurites). The average and maximal velocities of particles moving anterogradely (0.58±0.09 and 1.52±0.12 µm/sec) and retrogradely (0.29±0.06 and 1.18±0.10 µm/sec) are consistent with fast axonal transport. This indicates that Nmnat2 undergoes rapid net anterograde delivery from the cell body to neurites. This is another important prediction of our model, as rapid delivery is needed to replenish constant turnover of Nmnat2 in distal neurites (above).
Finally, if Nmnat2 is an endogenous axon survival factor, overexpression should protect transected neurites by preloading them with increased amounts of the protein. However, due to its relatively short half-life, protection should be highly dose-dependent and prolonged protection might only be achieved with very high levels of Nmnat2. In contrast, relatively long-lived WldS should also confer protection at much lower levels.
We tested the ability of exogenous expression of tagged Nmnat2 and WldS to protect transected neurites in a microinjection-based assay (Figure S9). Dilution of the injected construct allowed controlled amounts to be reproducibly introduced into neurons. At low vector concentration (1 ng/ml), WldS conferred robust protection to neurites for 24 h after cutting, whereas Nmnat2 provided almost no protection (Figure 8A and 8B). In contrast, at 50-fold higher construct concentrations, both Nmnat2 and WldS conferred protection to almost all cut neurites at 24 h (Figure 8A and 8B). Although we used identical expression cassettes to give the best chance of equal expression of the two proteins in this assay, the shorter half-life of FLAG-Nmnat2 probably manifests as a lower steady-state level at the time of cut relative to FLAG-WldS. Indeed, in transfected HEK 293T cells, we found that 2.5 times more FLAG-Nmnat2 construct was required to give steady-state protein levels approximately equal to FLAG-WldS (and the other Nmnat isoforms). Importantly, whilst we found that injection of the FLAG-Nmnat2 construct at 2.5 ng/ml gave slightly increased protection 24 h after cut relative to 1 ng/ml, this was still greatly reduced protection compared to the FLAG-WldS construct at the lower concentration (Figure 8A). Thus exogenous Nmnat2 only confers significant protection of cut neurites when expressed at high levels, consistent with its short half-life, whilst more stable WldS protects even at low levels.
Our results provide direct support for the hypothesis that constant delivery of a labile, cell body–synthesized survival factor is required to stop healthy mammalian axons undergoing Wallerian degeneration. Defects that prevent its delivery, including axon injury [6],[40], axonal transport impairment [27],[41], cell death [35], or disruption of protein synthesis in the cell body (Figures 1 and 2), all trigger WldS-sensitive axon degeneration. We identify Nmnat2 as one such critical axon survival factor, required to maintain normal axon integrity and sufficient to preserve injured ones at high doses. Nmnat2 half-life, uniquely among the three Nmnat isoforms, is consistent with the timing of the latent phase of Wallerian degeneration and commitment to degenerate in primary culture. We also show for the first time that endogenous Nmnat2 is present in neurites, where levels drop rapidly after injury. Importantly, this is not a consequence of neurite degeneration but represents natural turnover prior to activation of degeneration. These findings have significant implications for our molecular understanding of Wallerian degeneration and “dying-back” axonopathies, and for the mechanism by which WldS and other Nmnat isoforms delay axon degeneration.
The most compelling evidence that Nmnat2 is required for maintenance of healthy axons is our observation that siRNA-mediated knock-down of Nmnat2 alone induces neurite degeneration in the absence of injury and that this precedes any effect on neuronal viability. The initiation and progression of this degeneration is clearly slower than that caused by protein synthesis suppression, but this is consistent with the mechanisms involved. The critical rate-limiting factor following translation inhibition is protein half-life, but for siRNA-mediated knock-down additional time is needed for mRNA degradation. Pharmacological inhibition of translation could also result in more efficient and homogenous knock-down. It is also possible that depletion of other axon survival factors after global suppression of protein synthesis may contribute to this difference in timing. Nmnat1 and Nmnat3 seem unlikely to be among them in this experimental system because of their long half-lives, the absence of any clear effect of their siRNAs, and the fact that endogenous levels of both proteins cannot compensate for loss of Nmnat2.
Nmnat2 is a labile protein in HEK cells, in whole SCG explants, and in transected neurites. The rate of Nmnat2 turnover is consistent with the trigger for axon degeneration being depletion below a critical threshold. Nmnat2 falls to barely detectable levels in transected wild-type SCG neurites prior to any significant physical signs of degeneration, which suggests that the critical threshold level of Nmnat2 is quite low. However, the precise threshold level is difficult to determine because the duration of downstream steps needed to bring about degeneration is unknown. Steady-state levels of Nmnat2 in SCG neurites also seem quite low and this could account for the short latent phase between neurite transection and degeneration in these cultures.
Of the three mammalian Nmnat isoforms, Nmnat2 did not initially appear the most obvious candidate for an endogenous axon survival factor, despite being the most abundantly expressed isoform in the nervous system at the mRNA level [23],[34]. First, its predominant Golgi localization seemed inconsistent with an axonal location. However, a recent report shows axons in primary neuronal cultures contain Golgi components [42] and, as with WldS [3], predominant localization may not reflect the site of its axon protective role. We have now clearly detected endogenous Nmnat2 in SCG neurites by immunoblotting and have shown that an Nmnat2-eGFP fusion localizes to distinct, rapidly transported particles in these neurites (Figure 7 and Video S1). It will be interesting to determine the precise nature of these particles. Second, the inability of Nmnat2 to protect 5-d lesioned axons in Drosophila, unlike the other Nmnat isoforms and WldS, initially suggested it was either ineffective or by far the least potent isoform [2]. However, it has more recently been shown that exogenous expression of Nmnat2 can protect injured mammalian axons [26]. We propose that the short half-life of Nmnat2 could provide an explanation for this discrepancy, with the degree of protection being related to the levels of Nmnat2 expression achieved in the different systems. It is also possible that some protection of lesioned Drosophila axons might be evident at a less stringent time point (wild-type fly axons begin to degenerate just 1 d after injury). Thus, a short half-life, one of the most critical inherent properties of the endogenous survival factor in our model, might mask the capacity of exogenous Nmnat2 to protect in some situations. Conversely, greater stability probably underlies the ability of exogenous Nmnat1 and Nmnat3 to protect injured axons/neurites more robustly in this and other in vivo and/or in vitro situations [1],[2],[8],[25].
Our data suggest that WldS may protect axons by directly substituting for loss of endogenous Nmnat2 after injury or other stresses. This is based on three principal observations. First, WldS is inherently more stable than Nmnat2, decaying less in 48 h than Nmnat2 does in 4 h after neurite transection (Figures 6, 7, and S8). Importantly, continued degradation of Nmnat2 in WldS neurites rules out an alternative hypothesis, that WldS could delay axon degeneration by stabilizing Nmnat2. Second, WldS and Nmnat2 share the same enzyme activity, which is required for their ability to protect axons [5],[26],[43]. Third, both are present in axons (Figure 7, Video S1, and [3]), and the presence of WldS in microsome fractions [3],[8] is consistent with a possible shared localization with Nmnat2 in Golgi, or Golgi-derived structures in axons [42].
The ability of exogenous nuclear Nmnat1 and mitochondrial Nmnat3 to confer axon protection in a number of situations outwardly seems to contradict the claim that Nmnat localization is actually important, but there is increasing evidence to support it. First, endogenous Nmnat1 and Nmnat3 (which do appear to be expressed in SCG neurons; Figure S3) cannot compensate for loss of Nmnat2 (Figure 4), probably as a result of strict compartmentalization. Alternatively, this could simply reflect the relative contributions of each isoform to total basal Nmnat activity in these axons. Second, redistribution of predominantly nuclear WldS and Nmnat1 into the cell body and axon enhances their ability to delay Wallerian degeneration [3],[7],[8]. Finally, Nmnat1 and Nmnat3 only confer protection when overexpressed. This appears to be accompanied by significant mis-localization (Figure 3B, unpublished observations, and [8]), which may cause a serendipitous increase in effective Nmnat levels in the relevant axonal location. The ability of barely detectable extra-nuclear WldS to protect injured WldS mouse axons suggests that minimal mis-localization of relatively stable Nmnat1 and Nmnat3 may be sufficient to confer strong protection. The absence of significant axon protection in transgenic mice expressing Nmnat1 in neurons at similar levels to WldS in WldS neurons [4],[8] suggests either that Nmnat1 localization is more rigorously controlled in vivo or that Nmnat1 overexpression is greater in vitro. Importantly, Nmnat1 can only protect mammalian axons in vivo when specifically mutated to cause mis-localization [7].
The main known function of the mammalian Nmnat isoforms is NAD+ biosynthesis, and the ability of Nmnat1, Nmnat2, and WldS to delay Wallerian degeneration requires Nmnat enzyme activity [1],[5],[26],[43]. NAD+ production may therefore underlie the ability of endogenous Nmnat2 to prevent spontaneous axon degeneration. However, there is much disagreement over the ability of NAD+ to protect axons directly [1],[4],[8],[44],[45], or even its involvement at all [43]. Indeed, siRNA-mediated knock-down of Nampt, the rate-limiting enzyme upstream of Nmnat in the NAD+ salvage pathway, does not itself trigger axon degeneration despite a substantial 70%–90% reduction in NAD+ levels, leading to the suggestion that an alternative Nmnat metabolite may be involved [43].
Regarding downstream events, the rapid initiation and progression of Wallerian degeneration is more consistent with an active degeneration program than passive degeneration resulting simply from loss of an essential metabolic activity. Recently, dual leucine kinase (DLK) and JNK signalling have been implicated in regulating Wallerian degeneration of DRG neurites [46]. We have found the same JNK inhibitor (SP600125) used in that study also significantly delays Wallerian degeneration of SCG neurites (unpublished data). The localization of Nmnat2 in defined particles in axons and the role it plays in them could now be key to identifying additional downstream events.
Whilst neurite degeneration in primary neuronal cultures is a useful model of in vivo axon degeneration, high levels of protein overexpression can give misleading outcomes (discussed above) and other differences need to be considered. For example, there is a much longer latent phase before fragmentation of axons in transected sciatic nerves in vivo (36–40 h [15]) than for transected SCG neurites in culture (∼8 h). This could reflect differences in the half-life of Nmnat2 in vivo and in culture (for which there is some precedent [47]), steady-state levels of Nmnat2, or the involvement of additional factors that are more critical for axon survival in vivo. However, it would be somewhat surprising if Nmnat2 did not play a critical role in vivo based on its rapid turnover and it being required for neurite survival in vitro. Other Nmnat isoforms remain candidates in vivo, particularly Nmnat3 as its mitochondrial localization makes its presence in axons likely. Indeed, a contribution from other molecules could help to explain the longer latent phase in vivo.
It will also be interesting to see whether endogenous Nmnat proteins are involved in axon survival in non-mammalian organisms such as Drosophila. Whilst loss of the single Drosophila Nmnat homolog causes degeneration of photoreceptors, this appears to be a more general effect on neuronal viability, rather than axon health, and does not require its NAD+-synthesizing activity [48]. This contrasts with the protection against axon degeneration by mammalian Nmnat isoforms and WldS, which does require enzyme activity [1],[5],[26],[43]. Neuronal viability could therefore be dependent on a reported Nmnat-associated chaperone activity [49], with axons having a more specific dependency on enzyme activity. Thus, it is possible that the small decrease in neuron survival associated with Nmnat2 knock-down in SCG neurons (Figure 5A) could be due to loss of chaperone activity, although it is not yet known whether Nmnat2 possesses this activity like the other mammalian isoforms [49].
Loss of Nmnat2 could also underlie “dying-back” axon degeneration in disease. Due to its rapid turnover, Nmnat2 might fail to reach distal axons in sufficient quantities when axonal transport is either pathologically compromised [12],[27],[50] or slows during normal ageing [51]. Impairment of protein synthesis would be predicted to have a similar outcome, which could explain axon degeneration accompanying viral infections as the cellular protein synthesis machinery is overwhelmed [52]. More subtle effects on protein synthesis resulting from cell body defects, such as vacuolization, could underlie Wallerian-like “dying-back” axon degeneration and/or neuromuscular junction loss in slowly developing, chronic diseases like ALS [53],[54], even in the absence of neuronal loss. Our model would additionally explain why the longest axons are often most susceptible in disease. The ability of some larger mammals to support very long axons (up to several meters in some cases) raises the intriguing possibilities that Nmnat2 is inherently more stable in larger species or that chaperones stabilize it during transport.
We propose that increasing Nmnat2 stability or its delivery to axons could have important therapeutic implications for these and other disorders characterized by Wallerian-like degeneration. Both treatments should delay the point at which axons become committed to degenerate (like WldS). Such therapies might be particularly effective when axonopathy results from a short-term impairment (e.g., of cell body metabolism, axonal transport, glial support, etc.) lasting just a few hours to a few weeks. Examples include Taxol-induced neuropathy, relapsing-remitting multiple sclerosis, some viral disorders, and stroke. Axons could be saved permanently if the degeneration commitment point is delayed long enough for the causative defect to be removed or to abate naturally. Although WldS mice have already been shown to be resistant to Taxol-induced neuropathy [41], developing therapies based on the WldS neuroprotective mechanism has been limited by the technical challenge of introducing exogenous WldS (or other stable Nmnat isoforms). In contrast, pharmacological manipulation of endogenous Nmnat2 should be more feasible.
Finally, the increase in Nmnat2 levels in SCG cell bodies/proximal neurite stumps that we observed shortly after transection of their neurites is also intriguing (Figure 7A). The simplest explanation is that this represents accumulation of Nmnat2 in a reduced cellular volume following neurite removal while synthesis continues at pre-injury levels. However, the possibility that this could represent a stress response cannot be completely excluded at this time, especially in light of the recent report that the Drosophila Nmnat isoform can act as a chaperone [49]. Irrespective of the mechanism involved, this increase in Nmnat2 levels might nevertheless facilitate subsequent neurite regeneration.
In summary, we propose a model in which sustained expression and anterograde delivery of Nmnat2 is required to prevent activation of an intrinsic axon degeneration program. Degeneration is triggered when synthesis and/or delivery of Nmnat2 is disrupted and rapid turnover causes its level to drop below a critical threshold. We additionally propose that the relatively stable WldS fusion protein delays axon degeneration by directly substituting for loss of Nmnat2 and that localization may be an important factor. Endogenous Nmnat2 represents an exciting target for therapeutic manipulation.
Expression vectors encoding FLAG-tagged murine Nmnat isoforms and WldS were generated by amplification of the full coding region of each gene by Reverse Transcriptase PCR (RT-PCR) (see below) from 1 µg total RNA from wild-type and WldS mouse brain. Products were cloned into pCMV Tag-2B (Stratagene) to generate FLAG-Nmnat/WldS expression vectors or pEGFP-N1 (BD Biosciences Clontech) to generate a Nmnat2-eGFP expression vector. Sequencing (Cogenics) was performed to confirm the absence of PCR errors. Other plasmids used were pDsRed2-N1 for expression of variant Discosoma red fluorescent protein (DsRed2) and pEGFP-C1 for expression of eGFP (both BD Biosciences Clontech). Dharmacon ON-TARGETplus SMART pools of siRNA (Thermo Scientific) specifically targeted against mouse Nmnat1 (L-051136-01), Nmnat2 (L-059190-01), or Nmnat3 (L-051688-01) were used in this study. Dharmacon ON-TARGETplus siControl non-targeting siRNA pool (D-001810-10) was used as a control in experiments. Each pool consists of 4 individual siRNAs. The siRNAs making up the ON-TARGETplus Nmnat2 SMART pool (J-059190-09, -10, -11, and -12) were also tested individually or in subpools.
Total brain RNA was extracted using TRIzol reagent (Invitrogen), and RNA from dissociated SCG neuronal cultures was isolated using RNeasy columns (Qiagen). One µg of brain RNA and 30% of that recovered from SCG cultures was reverse transcribed into cDNA using Superscript II (both Invitrogen). Control samples without reverse transcriptase were processed simultaneously to rule out DNA contamination of samples. Standard PCR amplification was performed using REDTaq DNA polymerase (Sigma). Primers used for detection of Nmnat isoform transcripts in SCG neuron RNA were as follows: Nmnat1 5′-ttcaaggcctgacaacatcgc-3′ and 5′-gagcaccttcacagtctccacc-3′, Nmnat2 5′-cagtgcgagagacctcatccc-3′ and 5′-acacatgatgagacggtgccg-3′, Nmnat3 5′-ggtgtggaggtgtgtgacagc-3′ and 5′-gccatggccactcggtgatgg-3′. Products were sequenced to confirm correct amplification.
1,000× aqueous stock solutions of emetine (dihydrochloride hydrate) and CHX in DMSO (both Sigma) were diluted 1∶1000 in culture media to give final concentrations indicated (1 µg/ml CHX = 3.5 µM). InSolution MG-132 (Calbiochem) was diluted to 20 µM. MG-132 was added to SCG explant cultures 3 h prior to neurite transection. This pre-treatment is required to see neurite protection in these cultures [18]. Media was changed once with addition of fresh inhibitors when cultures were treated for more than 5 d. CHX-containing media was completely removed and replaced with media containing no CHX in experiments involving temporary suppression of protein synthesis.
Microinjection was performed on a Zeiss Axiovert 200 microscope with an Eppendorf 5171 transjector and 5246 micromanipulator system and Eppendorf Femtotips. Plasmids and siRNAs were diluted in 0.5× PBS and passed through a Spin-X filter (Costar). The mix was injected directly into the nuclei of SCG neurons in dissociated cultures. ON-TARGETplus siRNA pools were injected at a concentration of 100 ng/µl and individual siRNAs or sub-pools as indicated in the text, pDsRed2-N1 at 50 ng/µl, pEGFP-C1 at 10 ng/µl, the Nmnat2-eGFP expression construct at 50 ng/µl, and FLAG-Nmnat/WldS expression constructs or FLAG-empty control (pCMV Tag-2B) at 10 ng/µl for siRNA-mediated knock-down analysis by immunostaining (Figure 3B) and at 1, 2.5 or 50 ng/µl in neurite transection experiments (Figure 8 and Figure S9). Seventy to 150 neurons were injected per dish. Injection of relatively few neurons per dish facilitated visualization of individual labelled neurites as neurites tend to cluster together in bundles. For detection of FLAG-tagged protein expression by immunostaining, neurons were fixed with 4% paraformaldehyde (20 min), permeabilized with 1% Triton X-100 in PBS (10 min), blocked in 50% goat serum in PBS containing 1% BSA (30 min), and stained using monoclonal M2 anti-FLAG (Sigma) (1∶400 in PBS, 1% BSA for 1 h) and an Alexa568-conjugated secondary antibody (1∶200 in PBS, 1% BSA for 1 h). Cells were mounted in Vectashield containing DAPI (Vector Laboratories) for counterstaining of nuclei. For comparing the quantification of neuronal viability based on gross morphology with other indicators of health (Figure 5B), cultures were incubated with 1 µg/ml propidium iodide (Invitrogen) for 15 min and were then fixed with 4% paraformaldehyde (20 min) before being mounted in Vectashield containing DAPI.
Neurites were cut with a disposable scalpel roughly half-way between their cell bodies and their most distal ends. Where applicable, inhibitors of translation or vehicle (DMSO) were added to the media less than 10 min before transection. Uncut neurites treated with DMSO continue to grow normally (unpublished data). Microinjection of a row of cell bodies in dissociated SCG cultures enabled neurites to be cut so that all injected cell bodies and their proximal neurites were located on the opposite side of the cut site to the distal stumps (Figure S9).
Bright-field and fluorescence images were captured on an Olympus IX81 inverted fluorescence microscope using a Soft Imaging Systems F-View camera linked to a PC running the appropriate imaging software. Wherever possible, images of the same field of neurites or neuronal cell bodies were captured at the indicated time points after initial manipulation. Images were processed using Adobe Photoshop Elements 4.0. The intensity of FLAG immunostaining relative to eGFP fluorescence in individual injected neurons (Figure 3) was quantified using ImageJ software. Images were captured for analysis using identical microscope settings between samples for each channel. Time-lapse images of Nmnat2-eGFP transport were acquired 6 h after injection of the expression vector using an Olympus CellR imaging system comprising IX81 microscope linked to a Hamamatsu ORCA ER camera and a 100×1.45 NA apochromat objective. Cultures were maintained at 37°C in a Solent Scientific environment chamber. Wide-field epifluorescence images were captured at 2 Hz for 1 min. ImageJ software plug-ins were used for analysis of the stacks (kymograph generation and analysis of particle velocities) and conversion of an image stack into an annotated movie (Video S1).
|
10.1371/journal.pgen.1001171 | Nasty Viruses, Costly Plasmids, Population Dynamics, and the Conditions for Establishing and Maintaining CRISPR-Mediated Adaptive Immunity in Bacteria | Clustered, Regularly Interspaced Short Palindromic Repeats (CRISPR) abound in the genomes of almost all archaebacteria and nearly half the eubacteria sequenced. Through a genetic interference mechanism, bacteria with CRISPR regions carrying copies of the DNA of previously encountered phage and plasmids abort the replication of phage and plasmids with these sequences. Thus it would seem that protection against infecting phage and plasmids is the selection pressure responsible for establishing and maintaining CRISPR in bacterial populations. But is it? To address this question and provide a framework and hypotheses for the experimental study of the ecology and evolution of CRISPR, I use mathematical models of the population dynamics of CRISPR-encoding bacteria with lytic phage and conjugative plasmids. The results of the numerical (computer simulation) analysis of the properties of these models with parameters in the ranges estimated for Escherichia coli and its phage and conjugative plasmids indicate: (1) In the presence of lytic phage there are broad conditions where bacteria with CRISPR-mediated immunity will have an advantage in competition with non-CRISPR bacteria with otherwise higher Malthusian fitness. (2) These conditions for the existence of CRISPR are narrower when there is envelope resistance to the phage. (3) While there are situations where CRISPR-mediated immunity can provide bacteria an advantage in competition with higher Malthusian fitness bacteria bearing deleterious conjugative plasmids, the conditions for this to obtain are relatively narrow and the intensity of selection favoring CRISPR weak. The parameters of these models can be independently estimated, the assumption behind their construction validated, and the hypotheses generated from the analysis of their properties tested in experimental populations of bacteria with lytic phage and conjugative plasmids. I suggest protocols for estimating these parameters and outline the design of experiments to evaluate the validity of these models and test these hypotheses.
| CRISPR is the acronym for the adaptive immune system that has been found in almost all archaebacteria and nearly half the eubacteria examined. Unlike the other defenses bacteria have for protection from phage and other deleterious DNAs, CRISPR has the virtues of specificity, memory, and the capacity to abort infections with a virtually indefinite diversity of deleterious DNAs. In this report, mathematical models of the population dynamics of bacteria, phage, and plasmids are used to determine the conditions under which CRISPR can become established and will be maintained in bacterial populations and the contribution of this adaptive immune system to the ecology and (co)evolution of bacteria and bacteriophage. The models predict realistic and broad conditions under which bacteria bearing CRISPR regions can invade and be maintained in populations of higher fitness bacteria confronted with bacteriophage and narrower conditions when the confrontation is with competitors carrying conjugative plasmids. The models predict that CRISPR can facilitate long-term co-evolutionary arms races between phage and bacteria and between phage- rather than resource-limited bacterial communities. The parameters of these models can be independently estimated, the assumptions behind their construction validated, and the hypotheses generated from the analysis of their properties tested with experimental populations of bacteria.
| For many species of bacteria, adaptive evolution is through the expression of chromosomal and extrachromosomal (plasmid- and prophage - borne) genes or clusters of genes (pathogenicity and nicer islands) acquired by horizontal gene transfer (HGT) from the same or even quite distant species [1], [2]. Thus, on first consideration it may seem that bacteria and their accessory genetic elements would have mechanism to promote the acquisition, incorporation and expression of genes from without. And, indeed there are mechanisms like integrons [3]–[7] that appear to have that function. On the other side, DNA acquired from external sources may be deleterious. This is certainly the case when that DNA is borne on lytic bacteriophage, but also for plasmids that engender fitness costs [8], [9] or chromosomal DNA from the wrong source [10], [11]. To deal with these contingencies, it would seem that bacteria would have mechanisms to protect themselves against infection by deleterious foreign DNA [12]. And indeed there are systems like restriction-modification (restriction endonucleases) which appear to have that role [13], [14].
The most recently discovered mechanism postulated to provide bacteria immunity to infectious genetic elements are Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR). For recent reviews see [15], [16]. CRISPR is particularly intriguing because of its ubiquity, appearing in ∼90% and ∼40% of archaeal and eubacterial sequenced genomes, respectively, and because of the adaptive mechanism by which it provides immunity to infections by a virtually indefinite diversity of bacteriophage and plasmids. DNA from infecting phage and plasmids is incorporated into the CRISPR array. Through a yet to be fully elucidated mechanism, bacteria abort the replication of infecting phage [17] or the establishment of conjugative plasmids [18] bearing copies of the DNA incorporated into their CRISPR arrays, also see [19]. Further support for CRISPR being an adaptive immune system that is maintained because it protects bacteria from infection with phage comes from studies of the community ecology of bacteria and phage; DNA in the CRISPR regions of the bacteria from those communities corresponds to that in the co-existing phage [20]–[23]. For an intriguing perspective on CRISPR as a witness to the coevolutionary history of bacteria and phage, see [24].
CRISPR-mediated immunity has been likened to a Lamarckian mechanism [25], because the selection pressure, the infecting phage and plasmids, determine the genotype. This analogy however does not account for the evolution and maintenance of the machinery responsible for taking up the infecting phage and plasmid DNA and the mechanism employed to prevent the replication or establishment of infecting genetic elements with those sequences. Under what conditions will adaptive immunity to phage and plasmid infection be the selection pressure responsible for establishing and maintaining CRISPR-mediated immunity in populations of archeae and bacteria? What about other mechanisms of resistance, like structural modification blocking phage adsorption (envelope resistance) and restriction-modification? How do these mechanisms interact with CRISPR – acquired immunity and contribute to its establishment and maintenance?
To address these questions and provide a framework and hypotheses for their study experimentally, I use mathematical models of the population dynamics of bacteria, phage and plasmids to explore the conditions under which a CRISPR–like adaptive immune mechanism will provide bacteria a selective advantage in competition with bacteria without this immune system. The results of the numerical analysis of the properties of these models suggest that with bacterial replication and phage infection parameters in realistic ranges, there are broad but not universal conditions where a CRISPR–like adaptive immune system can be favored and will be maintained in populations of bacteria confronted with lytic phage. While this model predicts conditions where CRISPR-mediated immunity will be favored when bacteria compete with populations bearing conjugative plasmids, these conditions are relatively restrictive. The parameters of these models can be independently estimated, the validity of the assumptions behind their construction and the hypotheses generated from the analysis of the properties can be tested in experimental populations of bacteria with lytic phage and conjugative plasmids. Procedures for doing these experiments are outlined and their potential outcomes described and/or speculated upon. Also discussed are the broader implications of CRISR-mediated adaptive immunity to the population and evolutionary biology and ecology of bacteria and phage.
Both the lytic phage and conjugative plasmid models used here assume a chemostat-like habitat. The bacteria grow at a rate that is a monotonically increasing function of the concentration of a limiting resource, R µg/ml [26].where Vi hr−1 is the maximum growth rate of the ith strain of bacteria and k the concentration of the resource when the growth rate is half its maximum value (the “Monod constant”). The populations are maintained in a vessel of unit volume, (1ml) into which medium containing the limiting resource from a reservoir where it is maintained at a concentration A µg/ml flows in at a rate w per hour. Excess resource and wastes are removed from the vessel at the same rate. As in [27], the rate of uptake of the resource by the bacteria is proportional to the density, the resource concentration-dependent growth rates of the different populations of bacteria and a conversion efficiency parameter, e µg/per cell.
The model developed here is an extension of that in [28]. There are four populations of bacteria. Two are sensitive to the phage, N, non–CRISPR and C, CRISPR and two that are either fully resistant (envelope resistance), or immune because of CRISPR, NR and CR, respectively. The variables N, C, NR and CR are the both the densities (bacteria per ml) of these populations and used as their designations. There is one population of phage, with density and designation, P particles per ml.
The phage adsorb to the N and C and CR bacteria with rate constants, δN and δC (ml per phage per cell per hour) respectively. Phage do not adsorb to bacteria with envelope resistant, i.e. the NR cells. To account for a possible multiplicity of infection (MOI) effect on survival of phage-infected CR, the effective killing rate constant for phage adsorption to CRISPR can be an increasing function of the ratio of free phage and CR cells, M = P/CR.(1)where δMIN and δMAX are the minimum and maximum adsorption rates. The parameter x is a coefficient (0≤x≤1) that specifies the magnitude of the MOI effect, q is the MOI where the adsorption rate is half its maximum value and n is an exponent which contributes to the shape of the distribution. At low multiplicities, δCR (M) the CRISPR cells would be effectively immune (resistant) (Figure 1). At high multiplicities, however, immune CRISPR cells can be overburdened by phage, their immunity would be overridden, and the phage would replicate, killing the cells. On the other side, we assume that the phage are removed from the population by adsorption to immune CRISPR cells at the maximum adsorption rate, δMAX.
For convenience I neglect the latent periods of the phage infection but assume that the phage have potentially different burst sizes, βN, βC, and βCR particles per cell, for N, C and CR cells, respectively.
Phage-immune CRISPR cells, CR are produced from C at a rate proportional to the rate at which the phage adsorb to them and a constant m (0≤m≤1) which is the probability that a phage infection will be aborted and a CRISPR strain will be produced. At a rate v per cell per hour, CRISPR lose their immunity, CR→C. For the N and C populations the loss of the adsorbed phage is subsumed in the value of the burst size (which is one less than the number of phage produced). For the CR population, the loss of the phage due to adsorption is specifically considered because only a small fraction of the adsorbed phage replicate when the MOI is low.
In Table 1, I separately define these parameters and in Figure 2, illustrate the interactions between the different populations of bacteria and the phage. The equations for this model follow.
The model developed here is an extension of that in [29]. There are five bacterial populations. Two populations do not code for CRISPR, N and NP, and three populations code for CRISPR, C and CP and CX. The NP and CP populations bear the conjugative plasmid and CX, carries CRISPR and plasmid sequences that make it completely immune to the receipt of these plasmids. Plasmids are transferred by conjugation at rates proportional to the product of the densities of the plasmid-bearing and plasmid-free populations and rate constants, γNN, γNC, γCN and γCC (ml per cell per hour) respectively for the transfer of the plasmid from NP to N, NP to C, CP to N and CP to C., respectively. Plasmids are lost by vegetative segregation at rates τN and τC per cell per hour, with NP→N and Cp→C. C are converted to CX at a rate proportional to the rate at which C acquires the plasmid and a probability m (0≤m≤1). Cx lose the CRISPR plasmid immunity region and become C at rate ν per cell per hour. Each of the cell lines, have a maximum growth rate, VN, VNP, VC, and VCP, and VX per hour. In Figure 3, I illustrate the interactions between the different cell lines in this model, and, in Table 2, I separately define the parameters and variables. The equations for this model are:
For the numerical solutions to these equations (computer simulations) I use a differential equation-solving software package, Berkeley Madonna. For the phage simulations there is a refuge density, below which the phage are unable to adsorb to the bacteria. The purpose of this is to control the system from oscillating without limits, see [30]. In these simulations, if the phage density falls below 10−1 particles per ml, the phage are considered to be lost. Copies of these simulations are available online, www.eclf.net/programs.
The bacterial growth, resource-uptake, phage adsorption parameters and burst sizes used in these simulations (Table 1) are in a range similar to that which we observed for E. coli and the phages T2 and T7 [28], [31].
In accord with [34], conjugative plasmids will be maintained as long as the rate of infectious transfer exceeds the rates of loss of the plasmid due to selection against the cells carrying it, vegetative segregation, and the rate of flow through the chemostat. In terms of the above parameters, the plasmid will be maintained in an N-NP population as long as(2)where N* is the density of plasmid-free cells at the chemostat equilibrium. For example, if VN = 1.0, VNP = 0.95, w = 0.2, τN = 10−3, the plasmid will be maintained in a population of density N* = 108 as long as γNN>1.1×10−10. If the plasmid augments the growth rate (which in this model is the sole parameter of cell fitness) of the bacteria that carry it, VNP>VN, as we would anticipate for antibiotic resistance encoding plasmids in the presence of the selecting antibiotic, bacteria bearing the plasmid will be able to invade even without transfer, as long as the segregation rate, τN, is sufficiently small.
It has been less than eight years since the ubiquitous clusters of palindromic repeats now known as CRISPR first acquired this moniker [35]. Although there had been compelling circumstantial evidence that CRISPR was part of an adaptive immune system that provides protection against infecting phage and plasmids, it has been less than four and three years respectively since the publication of the first direct (read experimental) evidence that CRISPR can provide immunity to infection by lytic phage [17] and conjugative plasmids [18].
In the course of this time a great deal has been learned about the molecular biology of CRISPR and the mechanisms by which it provides adaptive immunity to plasmid and phage infection. But there remain many unanswered questions about these processes. Most important for this consideration is a dearth of the quantitative information needed to understand the population dynamics of CRISPR-mediated adaptive immunity and thereby the conditions for the establishment and maintenance of CRISPR in bacterial populations. To my knowledge, this study is the first formal consideration of these dynamics.
The models developed in this report incorporate what has been learned about CRISPR-mediated adaptive immunity to phage and conjugative plasmids, primarily from the studies of Barrangou and colleagues [17] and Marraffini and Sontheimer [18], into models of the population dynamics of lytic phage [28] and conjugative plasmids [29]. Although they may appear complex, at best they are simplistic caricatures of interactions between these infectious genetic elements and bacteria with CRISPR-mediated adaptive immunity. These models are not intended or anticipated to be numerically precise analogs of these processes and dynamics.
The role of these mathematical models is similar to that of the diagrammatic models (cartoons) used to illustrate the molecular basis and mode of action of CRISPR, i.e., to provide a framework for understanding these processes, designing experiments, and interpreting their results. In this case, these experiments are on population and evolutionary dynamics of bacteria with CRISPR-mediated immunity confronted with lytic phage and competing bacteria bearing conjugative plasmids. The purpose of these models for this experimental enterprise is: (i) to identify and, in a quantitative way, evaluate the role of the different factors (parameters) contributing to these dynamics and the conditions for the establishment and maintenance of CRISPR in bacterial populations, and (ii) to generate hypotheses about these dynamics and existence conditions that can be tested (and rejected) in experimental populations.
The results of the analysis of the properties of the phage - CRISPR model are consistent with the proposition that in the presence of lytic bateriophage there are broad conditions under which a CRISPR–like adaptive immune system can become established and will be maintained in bacterial populations. With population densities, growth rates, and phage infection parameters in realistic ranges, these models predict that despite a growth rate disadvantage, bacteria with CRISPR–like acquired immunity to infecting phage will increase in frequency when initially rare and will be maintained. The necessary condition for this is that the phage population continues to persist at a sufficiently high density for CRISPR-mediated adaptive immunity to overcome an intrinsic disadvantage associated with the costs of carrying and expressing these genes.
When will the phage maintain their populations at sufficient levels for this outcome? With the parameters used to address this question, the phage will be maintained under broad conditions, but may eventually be lost if a population with envelope or other resistance ascends to dominance. I emphasized the word may for two reasons. The first is theoretical, if the relative growth rate of the resistant population is adequately low, the phage and thereby CRISPR will be maintained. The second is empirical, even when resistant bacteria dominate experimental populations of bacteria and phage, in general the phage continue to be maintained [30], [31], [36].
The CRISPR plasmid model predicts that because of the immunity to infection with conjugative plasmids, a lower growth rate (Malthusian fitness) CRISPR population can become established and will be maintained when competing with bacteria with a greater Malthusian fitness but bearing deleterious (fitness-reducing) conjugative plasmids. Although these conditions are met with plasmid fitness costs in the range estimated for “laboratory” plasmids [9], [37], it is not clear that naturally occurring plasmids would be as burdensome as those maintained in the Lab. The greater the Malthusian fitness burden attributed to the plasmid, the greater the advantage of CRISPR-mediated immunity.
The rate constants of plasmid transfer used in these simulations are those for plasmids with permanently derepressed conjugative pili synthesis. Wild type conjugative plasmids are more likely to be repressed for the production of these transfer organelles and would have substantially lower rates of transmission than plasmids that are permanently derepressed for plasmid transfer [38], [39]. Indeed, it is not clear whether in natural populations conjugative plasmids that engender fitness cost can be maintained by transfer alone. Their persistence may require periodic episodes where bacteria carrying them have an advantage [34], [40], but also see [41]. If the rate of infectious transfer is not sufficient to maintain deleterious plasmid in a population and they persist by continually or periodically enhancing the cells Malthusian fitness, immunity to these plasmids would not be sufficient to maintain CRISPR-encoding cells that have an intrinsic fitness disadvantage.
It would be nearly impossible to determine whether the quantitative conditions predicted by these models for the establishment and maintenance of CRISPR-mediated immunity are met in natural populations. On the other hand, the values of the parameters of these models can be estimated and the validity of the assumptions behind their construction and hypotheses generated from the analysis of their properties can be tested in laboratory culture using CRISPR–positive and CRISPR–negative bacterial constructs, phage and plasmids of the types used respectively by Barrangou and colleagues [17] and Marraffini and Sontheimer, [18] in chemostat culture.
In this report, I elected to restrict the model and its analysis to the simplest cases with lowest realistic number of states of bacteria, phage and plasmids. I have done so because at this time these minimum number of states models and the predictions generated from their analysis are more amenable to evaluating and testing experimentally than models with more states of bacteria, phage and plasmids. Moreover, these tests, and particularly the population dynamic experiments, should indicate the importance of the generation of additional population states by mutation, like host range phage and host range plasmids, are to these dynamics. Be that as it may, I also realize that this minimum number of states model will not account for what may turn out to be the most important contributions of CRISPR-mediated immunity to the ecology as well as the population and evolutionary biology of bacteria and phage.
|
10.1371/journal.pmed.1002204 | The Subclonal Architecture of Metastatic Breast Cancer: Results from a Prospective Community-Based Rapid Autopsy Program “CASCADE” | Understanding the cancer genome is seen as a key step in improving outcomes for cancer patients. Genomic assays are emerging as a possible avenue to personalised medicine in breast cancer. However, evolution of the cancer genome during the natural history of breast cancer is largely unknown, as is the profile of disease at death. We sought to study in detail these aspects of advanced breast cancers that have resulted in lethal disease.
Three patients with oestrogen-receptor (ER)-positive, human epidermal growth factor receptor 2 (HER2)-negative breast cancer and one patient with triple negative breast cancer underwent rapid autopsy as part of an institutional prospective community-based rapid autopsy program (CASCADE). Cases represented a range of management problems in breast cancer, including late relapse after early stage disease, de novo metastatic disease, discordant disease response, and disease refractory to treatment. Between 5 and 12 metastatic sites were collected at autopsy together with available primary tumours and longitudinal metastatic biopsies taken during life. Samples underwent paired tumour-normal whole exome sequencing and single nucleotide polymorphism (SNP) arrays. Subclonal architectures were inferred by jointly analysing all samples from each patient. Mutations were validated using high depth amplicon sequencing.
Between cases, there were significant differences in mutational burden, driver mutations, mutational processes, and copy number variation. Within each case, we found dramatic heterogeneity in subclonal structure from primary to metastatic disease and between metastatic sites, such that no single lesion captured the breadth of disease. Metastatic cross-seeding was found in each case, and treatment drove subclonal diversification. Subclones displayed parallel evolution of treatment resistance in some cases and apparent augmentation of key oncogenic drivers as an alternative resistance mechanism. We also observed the role of mutational processes in subclonal evolution.
Limitations of this study include the potential for bias introduced by joint analysis of formalin-fixed archival specimens with fresh specimens and the difficulties in resolving subclones with whole exome sequencing. Other alterations that could define subclones such as structural variants or epigenetic modifications were not assessed.
This study highlights various mechanisms that shape the genome of metastatic breast cancer and the value of studying advanced disease in detail. Treatment drives significant genomic heterogeneity in breast cancers which has implications for disease monitoring and treatment selection in the personalised medicine paradigm.
| We understand very little about the genomic changes that take place in advanced breast cancers, particularly in the most advanced disease that results in death.
Seeking this information is expected to provide important insights into cancer biology and help us understand how advanced breast cancer evolves over time and the impact this has on the natural history of the disease.
We conducted rapid autopsies on four patients that died of advanced breast cancer and extensively profiled multiple lesions, which were compared to biopsies taken whilst patients were alive, where possible.
We aimed to reconstruct the relationship between different metastatic lesions and understand the tumour cell subpopulations that make up metastatic lesions.
Our findings reveal significant heterogeneity and evolution over time, which was shaped by treatment, mutational processes, and the interaction of these factors with alterations in cancer-promoting driver genes.
Our findings reveal the complexity of the cancer genome in advanced disease and highlight the importance of ongoing monitoring and reassessment of the genomic profile if this is to be used to guide therapy.
Detailed autopsy studies in small numbers of patients can yield valuable insights into cancer biology.
Detecting small populations of cancer cells from actual patient specimens is difficult, and the findings in this study could be limited by poor sensitivity to detect rare populations.
| Heterogeneity in the natural history of advanced cancers has long been noted. The advent of cancer genomics has revealed that significant heterogeneity may exist both between and within lesions in the same patient. Since that time, autopsy studies have found a great variety of genomic heterogeneity in multiple cancer types. Evolution over time has also been documented, with varying influences of therapy in shaping the subclonal architecture of advanced disease. At the same time, the clinical significance of heterogeneity is yet to be established.
Patients die of metastatic disease, but little is known about the biology of this late-stage, lethal process. Genomics is a mature and robust platform for querying this biology. In breast cancer, although many thousands of primary tumours have been characterised in detail, such comprehensive data do not exist for metastatic disease, particularly in the most advanced and lethal disease. Shah et al. studied genomic heterogeneity in a case of lobular breast cancer that recurred 9 years after initial diagnosis. This single case showed genomic evolution over time from primary to metastatic disease [1]. Ding et al. analysed a chemoresistant metastatic basal-like breast cancer and showed low divergence between primary and metastatic lesions, with 48 of 50 mutations in common between sites [2].
More recently, the focus has shifted to understanding heterogeneity at the subclonal level. Complex but reproducible subclonal dynamics were revealed by Eirew et al. in a large study of treatment-naïve breast cancer xenografts analysed at the single-cell level [3]. Murtaza et al. combined multi-region sequencing at autopsy with circulating DNA (ctDNA) measurements, finding that ctDNA captured some of the underlying subclonal dynamics [4]. Juric et al. used sequential biopsies and samples from autopsy to study the evolution of resistance to a PI3K inhibitor in a case of hormone-positive breast cancer, finding evidence for multiple resistance mechanisms evolving simultaneously in spatially distinct sites (termed convergent evolution) [5]. Yates et al. performed a combination of whole genome and targeted sequencing on 50 breast cancer cases and inferred subclonal populations, finding variable intra-lesional heterogeneity [6]. Low prevalence or “minor” subclones in primary tumours could be seen giving rise to recurrent disease and convergent evolution occurred both late and early in the disease course. The profound influence of treatment on the subclonal structure of breast cancer has also been delineated in human epidermal growth factor receptor 2 (HER2)-positive tumours using a method that visualises driver mutation heterogeneity in different tumour subpopulations [7]. This study also found an association between heterogeneity and poor clinical outcome. Determining the genomic changes that define subclonal populations allows the evolutionary history of a cancer to be inferred, as was performed by Gao et al. in a single-cell whole genome sequencing study of triple negative breast cancers [8]. Well-demarcated subclonal populations could be defined by copy number alterations (CNA) without the presence of populations showing intermediate copy number states. These findings are most consistent with punctuated evolution of neoplastic phenotypes early in the natural history of a tumour rather than gradual accumulation of genomic alterations. These studies paint an expanding picture of the challenges and opportunities in understanding the complexities of advanced disease. Many questions remain about the relationship between breast cancer’s capacity to evolve towards heterogeneous states and clinical outcome.
To understand the biology of lethal metastatic disease, our institution has implemented a prospective community-based rapid autopsy program (CASCADE) in which patients consent whilst alive to allow tissue donation after death [9]. In this study, we attempted to understand the evolutionary history of four patients with lethal breast cancer who participated in the CASCADE program: an aggressive and treatment-resistant triple negative breast cancer where the patient died 12 months from diagnosis (TN1); an oestrogen-receptor (ER)-positive HER2-negative breast cancer with late relapse 7 years from diagnosis of early-stage disease (ER1); a de novo metastatic ER-positive, HER2-negative breast cancer (ER2); and an ER-positive, HER2-negative breast cancer in a young patient with multiple instances of discordant responses to treatment between metastatic lesions (ER3). By interrogating the cancer genome at multiple metastatic sites, including metastatic biopsies taken during life, we aimed to infer the subclonal structure and evolutionary history of the disease for each case as it progressed from the primary tumour to lethal metastatic dissemination.
As breast cancer is relatively treatment responsive compared to other tumour types, we focussed our analysis on understanding subclonal composition and how this evolves over time under the influence of therapy. The first case was recruited in 2013, at which time methods for performing subclonal inference on multiple samples were not mature. By the time the final case was recruited in 2015, the field had advanced considerably, which made this approach feasible, as will be detailed below. All four cases were analysed concurrently.
Patients were recruited by their treating clinicians to CASCADE, a prospective community-based rapid autopsy program [9] conducted in association with the Kathleen Cuningham Foundation Consortium for Research into Familial Breast Cancer (kConFab) and approved by the Human Research Ethics Committee of the Peter MacCallum Cancer Centre, Melbourne (HREC approval numbers: CASCADE 13/122, kConFab 92/97 and 11/102). Any patient with advanced breast cancer was eligible. All patients provided written informed consent.
The workflow for the CASCADE program is shown in Fig 1. Once recruited, the CASCADE coordinator communicates with the patient, their family, and the associated health care providers regarding the patient’s status and ongoing willingness to participate. Following the patient’s death, the next of kin specifies when the body can be transferred for autopsy. Autopsies were conducted by a specialist forensic pathologist assisted by the research team. Tumour tissue was harvested from multiple representative metastatic sites. Each lesion was divided into portions that were immediately snap frozen in liquid nitrogen and also fixed in formalin for subsequent paraffin embedding. Frozen samples were used for sequencing where possible. Prospectively collected fresh frozen samples were available for some cases. For all fresh tissues, frozen sections were reviewed by a pathologist to confirm the presence of tumour, quantify necrosis, and estimate tumour cellularity. Premorbid formalin-fixed, paraffin-embedded (FFPE) archival samples from primary tumours and metastatic lesions were also obtained.
DNA was extracted from sections cut from frozen tissues or FFPE blocks. For three of the cases, ER1–ER3, whole exome libraries were prepared using the Roche-NimbleGen SeqCap EZ exome version 3. For case TN1, libraries were prepared using the Illumina Nextera Rapid Exome. All libraries were sequenced on the Illumina Hiseq platform. TN1 had DNA submitted for single nucleotide polymorphism (SNP) array analysis with the Illumina Human Omni 2.5S beadchip, and other cases had DNA submitted for SNP array analysis on the Affymetrix Genome Wide SNP 6.0 array. Blood was available for all cases for reference germline DNA.
After adapter trimming with cutadapt [10], raw sequence data were aligned to the human reference genome GRCh37 with BWA [11]. Two samples for TN1 exhibited low sequencing yield and a large number of likely artefactual mutations due to FFPE processing, and were excluded from further analysis. Fresh frozen samples were sequenced to a mean depth of 105, with FFPE samples sequenced to a mean depth of 75. Following alignment, BAM files were processed according to GATK best practices [12]. Variants were called using MuTect [13] with default parameters. Indels were called with VarDict [14] and pindel [15].
Called variants were annotated with ANNOVAR [16]. Variants were filtered as follows: exclusion of variants in 1000 Genomes and the Exome Sequencing Project [17,18]; variants with an allele frequency ≥2% or with 2 or more supporting reads in the normal; and allele frequency <5% in fresh tissues or <10% in FFPE tissues. Indels were manually reviewed and excluded if adjoining >4 homopolymer runs or repetitive regions.
Copy number was called from whole exome sequencing (WES) data using the facets package, which also provided purity and ploidy estimates [19]. For the Illumina bead chips, raw data were tQN normalised [20], and segmented allele-specific copy number was called using OncoSNP [21]. For the Affymetrix SNP chips, samples were processed using the Aroma package, applying TumorBoost and PSBCS [22,23]. Output of facets was compared to the SNP chip data, and, in some cases, the parameters were adjusted to exclude less likely combinations of copy number and purity. As SNP arrays were not available for the FFPE samples, whole exome copy number was used for all analyses. Copy number segmentation determined from SNP arrays was used as orthogonal validation of WES copy number calls from facets.
To determine the functional significance of novel nonsynonymous variants, the following steps were taken: search COSMIC [24] and TCGA [25] for previous reports (across all tumour types); prioritise variants with a high CADD score [26]; map variants to functional domains of the gene, if possible; check if variant falls in a mutational hotspot using MutationAligner [27] and Cancer HotSpots [28]; and review literature (using GeneRIF queried via MyGene.info [29]). Annotation of CNA was limited to genes found to be recurrently altered in the literature [30,31].
superFREQ is a cancer exome clonality inference tool that takes advantage of multiple samples from the same individual by tracking both single nucleotide variants (SNV) and CNA across samples [32]. This allows the detection of highly subclonal somatic mutations that are present at higher cell fraction in other samples. An advantage of superFREQ is that it estimates and propagates statistical and systematic error sources throughout the analysis, thus decreasing the number of false positives and allowing downstream analysis of uncertainties. It also accepts aligned sequence data as input directly. Briefly, the pipeline performs the following steps: (1) GC bias correction, (2) differential coverage analysis with limma-voom [33], (3) examines variant positions shared between individual samples, flagging variants for base quality, mapping quality, strand bias, stuttering, or other artefacts detected in the pool of normal control samples, (4) summarises differential coverage and SNPs for each gene, (5) recursively clusters neighbouring genes with sufficiently similar differential coverage or SNP frequencies until a segmentation of the genome is achieved for each sample, (6) summarises consensus coverage and SNP frequencies for each segment and renormalizes, taking accuracy of coverage and SNP frequency into account, (7) calls CNAs and clonality in each segment based on coverage and SNP frequency, (8) calculates clonalities of somatic SNVs using local CNA, (9) tracks clonalities of SNVs and CNAs over multiple samples to determine if the same or different alleles are gained/lost between samples, (10) clusters mutations (SNVs or CNAs) with similar clonalities in all samples into clones, (11) sorts clones into a tree structure, with smaller clones being assigned as subclones when the sum of the clonalities of disjoint subclones cannot be larger than the clonality of the containing clone. Not all mutations are allocated to subclones by superFREQ, usually due to highly deranged copy number, which does not permit the cancer cell fraction to be reliably determined. Others have noted this problem with subclonal reconstruction [6].
Case TN1 displayed a high degree of copy number variation, including large swathes of loss of heterozygosity (LOH) affecting over 50% of the genome. Accurate clonal inference with superFREQ was not successful. As an alternative, mutation clonality was determined using PyClone [34], with allele-specific copy number provided by the R package facets [19]. Mutation genotypes were built using “parental_copy_number” mode. For mutations in regions of LOH, genotypes consistent with LOH were assigned a prior weight of 10, and other genotypes down weighted to 0.1. All other genotype prior weights were left as 1. From the PyClone output, the median clonality for each mutation (or cancer cell fraction) was clustered via affinity propagation using the Schism package [35]. Mutation clusters were reviewed to separate agglomerated private subclones if needed, and then the reviewed clusters were supplied to Schism’s genetic algorithm to construct a multi-sample phylogeny. To validate this approach, high depth validation sequencing of mutations found in more than one sample was subjected to the same process to check the consistency of the tree with the WES data.
From each case, mutations and indels were selected for high depth validation based on their contribution to the subclonal phylogeny and biological interest. An Ion AmpliSeq custom panel was designed for the variants of interest in each case, and multiplex PCR was performed as per the manufacturer’s protocol. Libraries were sequenced on the Ion Torrent PGM sequencer (Life Technologies) to a median depth of 770. Wild-type and mutant reads from validation loci were extracted from BAM files using the GenomicAligments R package [36]. Taking loci where there were at least 20 reads, a mutation was considered validated if there was a significant deviation from an expected error rate of 1/200 using the p-binomial test [1].
Mutational signatures were ascertained in a two-step process. Somatic mutations were filtered to have allele frequency >10% in FFPE samples to avoid detecting formalin fixation artefact and >5% in fresh samples. To increase detection power and avoid spurious signature detection, all unique mutations from each patient were pooled together and used as input to deconstructSigs [37]. “default” normalisation was used as per the authors’ instructions, with a minimum signature contribution of 0.06. The 30 signatures found in the latest COSMIC classification were used [38,39]. The contribution of the restricted set of signatures found in the pooled mutation set was then examined in each sample. To test the robustness of signature detection, pooled mutations were randomly downsampled to between 10% and 90% of the original mutations, and deconstructSigs was re-run on 1,000 such downsampled sets.
ctDNA was assayed as previously described [40]. For immunohistochemistry, heat-induced antigen retrieval was performed in 1× citrate buffer (Thermo Scientific). Samples were blocked in 2% bovine serum albumin in tris-buffered saline/Polysorbate 20 solution and endogenous peroxidase inactivated in 1.5% H2O2. Samples were incubated with primary antibodies, including AGF2 (Abcam), CHD4 (Abcam), and Cullin1 (Cell Signalling Technology). Biotinylated species-specific secondary antibodies were used at 1:300 (Dako) followed by the avidin-biotin-complex (ABC) method prior to visualisation with 3,3′-Diaminobenzidine (DAB) chromogen (Dako). Bright-field microscopy was performed on an Olympus BX-51 microscope. Murine inguinal mammary fat pads were used as normal control tissue.
Three ER-positive cases (denoted ER1, ER2, and ER3) and one triple negative case (denoted TN1) were analysed. Cases were recruited sequentially, and no cases were excluded from the analysis. In all cases, primary tumours were available as FFPE samples. Original pathology reports describing the macro-dissection of the primary tumours were used to select FFPE blocks in spatially distinct regions for each patient. In ER2, ER3, and TN1, metastatic biopsies taken during life were also available for analysis, as well as ctDNA for ER2. To study heterogeneity and evolution in detail, 8 (ER1), 13 (ER2), 16 (ER3), and 15 (TN1) samples were sequenced from each patient. Samples are summarised in S1 Table.
Patient TN1 was a woman diagnosed with a locally advanced triple negative breast cancer not associated with a BRCA1 or BRCA2 germline mutation, at age 39. She had a clinical response to a third-generation standard adjuvant chemotherapy regimen and underwent mastectomy, which showed significant residual disease. Shortly thereafter, she developed a solitary bony metastasis, followed by a liver metastasis and then multiple additional sites of metastatic disease across lung, liver, and brain before dying of liver failure less than 12 months from first diagnosis. She was treated with trastuzumab briefly when the breast primary was found to have focal HER2 positivity, which was not seen again in other lesions. In contrast, patient ER1 relapsed with metastatic disease 7 years after her primary diagnosis at age 42 and underwent a series of endocrine therapies before developing liver metastases and dying of liver failure after surviving 8 years with metastatic disease. Patient ER2 had de novo metastatic disease diagnosed at age 35 and survived 3 years, receiving multiple lines of chemotherapy and endocrine therapy. Patient ER3 did not receive surgical or medical treatment after initial diagnosis of early stage disease at age 36 due to personal circumstances. She presented 18 months later with metastatic disease and displayed discordant responses to therapy during her disease course, which ran 4 years in duration before death.
Of 52 samples, three samples from TN1 were sequenced but not used in further analysis, as they failed quality control (a breast core biopsy and biopsy of a femoral lesion taken premortem, and a subcarinal lymph node from autopsy). One additional sample, a brain metastasis from TN1, failed to sequence. Quality control metrics along with per sample validation rates are shown in S2 and S3 Tables. For mutations detected with an allele frequency ≥10% from WES, high depth validation rates for FFPE samples ranged from 95%–100% and from 97%–100% for fresh samples. When mutations with an allele frequency ≥5% were used, 44/45 samples had validation rates of 95% or greater, with one sample from TN1 having a rate of 94%. Importantly, high depth orthogonal validation did not affect the structure of the subclonal tree. Three samples could not undergo validation due to insufficient DNA (1 pre-chemotherapy breast biopsy for ER2 and the 2 primary samples for ER3).
Inferring subclones and constructing subclonal phylogenies is a complex problem, requiring high sequencing depth and/or multiple samples to increase sensitivity and reduce the space of possible tree reconstructions [41–43]. By utilising multiple samples from each patient, we were able to construct consistent subclonal phylogenies using mutation and copy number information (Figs 2–5). Private subclones (that is, subclones that are found only in one lesion) are not displayed in these figures for clarity, as they do not contribute to the core tree structure; therefore, all clones displayed were found in at least two spatially distinct sites.
All ER-positive cases had founding clones containing known driver alterations. These founding clones developed secondary alterations before giving rise to diverse subclones across multiple metastatic sites. By way of example, in ER1 shown in Fig 2B, we inferred the presence of an intermediate blue clone not found in isolation. This follows from (1) the absence of this clone in the primary tumours, indicating the blue clone is a true subclone; (2) the absence of the orange subclone in two samples where the red subclone has a high clonality, indicating the red subclone did not arise in the orange subclone; and (3) two different subclones (red and orange) with the blue clone as an ancestor. This intermediate blue clone was marked by an FGFR4 tyrosine kinase domain mutation and a splice site mutation in MGA, which counter-regulates MYC activity. The blue clone also shows deletions in genomic regions containing known tumour suppressor genes RB1, SPEN, CASP9, and FANCA. These mutations and CNAs likely conferred metastatic potential, as all subsequent metastatic subclones arose from this intermediate clone, as described above. The FGFR4 mutation could not be detected in three spatially separated samples of the primary tumour, even with high depth validation sequencing, although this does not rule out very low prevalence subclones at the time of diagnosis. Emergence of an unheralded “lethal subclone” following treatment has been documented in many tumour types, including breast cancer, but here it associated with delayed disease relapse.
ER2 and ER3 display a different relationship between subclonal structure and metastatic disease. For ER2, a biopsy of the breast mass was the index lesion, and the breast primary was removed after chemotherapy, followed by a liver biopsy after several lines of therapy. All premortem samples displayed a similar subclonal structure, with a subclone (orange) marked by a frameshift deletion in the cancer testis antigen MAGEA3 and a splice site mutation in ROCK1 that was subsequently detected at multiple sites in the liver at autopsy. The absence of this subclone at other metastatic sites, however, shows there was early divergence and parallel evolution of at least two other subclones (blue and red). ctDNA for the AKT1 (E17K) mutation and three ESR1 mutations (D538G, S463P, E380Q) were assayed in plasma taken at the time of liver biopsy and subsequently after death. Fig 6F shows all these mutations were detectable at both time points (albeit at very low levels for the first time point), even though the liver biopsy WES data did not reveal any ESR1 mutations.
ER3 was a case of delayed presentation, with 2 years elapsing between initial detection of early stage disease and subsequent metastatic disease, without intervening treatment. No samples could be obtained from the original diagnosis. In a similar fashion to ER1, the primary subclones have a linear monophyletic relationship with the metastatic subclones. This case is distinguished, however, by an ovarian lesion that contains the founding clone only. This ovarian lesion was discovered incidentally during a therapeutic oophorectomy prior to any anticancer therapy. Hence, despite lacking the more complex subclonal structure seen in other lesions, this disease possessed the ability to metastasise very early. Whether this ovarian metastasis would have resulted in clinically significant disease is unknown. The rest of the lesions must have arisen from dissemination of a different subclone, which had acquired the SPEN K1838X nonsense mutation and deletion of the negative modulator of oestrogen signalling DUSP22. These alterations occurred while the patient was receiving endocrine therapy.
TN1 displayed both short linear evolution and early divergence. Fig 5 shows the subclonal structure of the primary is found in all metastatic lesions via the green clone, but the blue and orange clone diverged from a common ancestor derived from an earlier progenitor of the green clone. Fig 5C shows the private subclones, some of which also emerged from the common ancestor. It is noteworthy that this case, with the most aggressive disease history with no response to three different standard chemotherapy regimens, had the fewest common subclones. S1 Fig demonstrates that CDH4, which contains a truncal mutation, is expressed at the protein level.
A key question is whether metastatic disease from one lesion can cross-seed anatomically distant sites. This is difficult to distinguish from a subclinical low prevalence clone seeding two different metastatic sites, a disseminative pattern rather than cross-seeding. All cases showed spatially distinct lesions with very similar subclonal structures (e.g., liver lesions in ER1; para-aortic nodal metastases in ER2; brain, lung, and liver metastases in ER3; lung metastases in TN1). This is suggestive of metastatic cross-seeding. In addition, ER1, ER2, and TN1 showed recurrent subclonal mixtures, raising the possibility of seeding by polyclonal clusters sustained by clonal cooperation, in which two subclones help maintain each other’s survival and confer novel phenotypic traits [3,44]. For such patterns to arise from dissemination alone rather than cross-seeding, polyclonal seeding, or cooperation, multiple waves of dissemination by different subclones from a primary tumour to the same metastatic site would be required.
In contrast to subclonal mutations, copy number changes between samples were relatively stable (Fig 7) for ER1, ER2, and TN1. Areas of LOH were particularly consistent across samples. ER1 and ER2 displayed relatively few CNA, with overall 80%–85% of coding genes unaffected. ER3 showed widespread copy number derangement affecting 40%–60% of coding genes, particularly on chromosomes 8 and 20. The ploidy of ER3 increased from early to late stage disease; the primary, ovarian metastasis, and lymph node biopsies all displayed diploid genomes, with autopsy samples showing triploidy and tetraploidy in some cases (mean ploidy across samples 3.2). There was inter-lesion copy number heterogeneity for ER3. This is reflected in the relative abundance of subclonal copy number events in ER3 called by superFREQ, with a mean of 22 copy number events per subclone for ER3, compared to 7 and 11 for ER1 and ER2, respectively. Chromosomal instability (CIN) was therefore an ongoing process contributing to subclonal evolution in ER3. TN1 showed the most deranged genome, which also had a mean ploidy of 3.9, consistent with genome doubling, with 40%–50% of genes subject to LOH and some form of copy number derangement affecting 70%–80% of genes. Subclonal copy number could not be called satisfactorily for TN1. Although less deranged, in ER1, there was evidence that the primary event was amplification of 1p, prior to driver mutations in PIK3CA (Fig 2B).
These findings are consistent with the growing body of work that extensive CIN takes place early in the evolution of a tumour. In particular, a variety of data from single-cell sequencing of tumours and normal cells has shown tumours contain only a few copy number states, without a large number of intermediates, suggesting that large-scale copy number changes appear to occur in a punctuated fashion rather than accumulating gradually over time [8]. A limitation of studying this question in bulk sequencing is that detecting subclonal copy number changes with high sensitivity is difficult, particularly in specimens with CIN and suboptimal purity.
These cases present a unique opportunity to study clonal evolution during therapy. Breast cancer is the prototypical “treatable” solid tumour for which patients receive multiple lines of therapy that usually result in a partial response or stable disease.
There was clear evidence that treatment alters subclonal structure. Disease at autopsy showed subclones that contained mutations conferring chemotherapy resistance. In ER1 one such mutation, TYMS V106M, is at a highly conserved residue in the binding site of thymidylate synthase, adjacent to two mutations known to cause resistance to 5FU in vitro (Fig 2B, yellow branch) [45].
Mutations in the ligand binding domain of ESR1 have recently been discovered as a common mechanism of resistance to endocrine therapy in metastatic breast cancer [46]. ER2 displayed three different ESR1 mutations all in different subclones. Furthermore, an ESR1 S463P mutation occurred in a subclone that already harboured an ESR1 E380Q, presumably in a different allele. As ESR1 mutations can be treated with alternative endocrine therapy and represent a tractable form of endocrine resistance, identifying them in patients failing endocrine therapy is of some importance [47]. ER2 had a liver biopsy while alive that did not show any ESR1 mutations, highlighting the difficulty in managing advanced heterogeneous disease. At the time of the liver biopsy, circulating cell-free DNA in plasma showed detectable low levels of all three ESR1 mutations (Fig 6F), which increased dramatically by the time of autopsy.
ER3 demonstrated discordant clinical responses on several occasions, with disease in the lung responding while disease in the liver progressed and left and right hilar lymph nodes responding differentially to chemotherapy. Fig 5C shows the divergent subclonal structure of lung, liver, and hilar lymph nodes, although this may be a consequence of heterogeneous treatment responses rather than the cause. Fig 7 also shows that differences in copy number can be discerned between sample pairs in ER3, more so than the other cases, as discussed in the previous section.
We found additional evidence of this effect when considering the evolutionary trajectory in the context of dominant oncogenic drivers. In ER2 and ER3, there were multiple subclones with alterations in genes that interact with the truncal drivers in each case, AKT1 and TP53, respectively (Fig 8). For ER2, alterations occurred in genes involved in the regulation of phosphatidylinositide phohsphates and PI3K signalling (PLPP4, PNPLA6, and PIK3R5), negative regulators of AKT1 signalling (PLEKHO1, PHLPP1, PID1, and USP12), and downstream effectors of AKT1 (MFN1). For ER3, alterations occurred in coactivators and facilitators of TP53 activity (BACH2, ANKRD11, TP53I11, MYO10, WDR3, and NDUFAF6), negative regulators of TP53 function (G3BP2), and downstream effectors (CD82). These findings suggest that resistance to chemotherapy agents may arise from augmentation of existing oncogenic or tumour suppressor signalling pathways rather than direct resistance through altered drug targets or drug metabolism. This could be one explanation why increasing lines of treatment become progressively less effective, despite having non-overlapping mechanisms of action.
Due to relatively low numbers of mutations, clear mutational signatures could be established only in ER2 and TN1. The pattern of mutational signatures in these two cases were robust and maintained a stable prevalence even when the starting mutation pool was down sampled to 10% of the total.
TN1 had an unusual composition of mutational signatures (Fig 6A–6C). There was a strong signal from signature 17 in the COSMIC classification [38,39]. This signature has been reported in breast cancer as well as oesophageal, liver, lung, and stomach cancers, melanoma, and B-cell lymphoma [38]. The aetiology is unknown. As signature 17 produces a characteristic T > G transversion that has relatively little overlap with other signatures, the contribution of the signature to the mutations in subclones can be tracked. Mutations arising from this signature were found primarily in metastatic lesions from autopsy, comprising between 9%–45% of the private mutations in four spatially distinct lesions in the lung and liver. The signature was not active in the primary lesions or premortem liver biopsy. These private signature 17 mutations have subclonal fractions in the 0.2–0.6 range and cluster with other mutations to form private subclones. Assuming that it is extremely unlikely a mutational process could cause the same mutation in multiple cells within a single lesion, each private subclone with a relatively high subclonal fraction of the signature 17 mutations must have arisen from a very restricted population of single cells that expanded in each lesion. In addition, as a mutational process per se confers no proliferative advantage, the prevalence of this signature may represent cells that were able to survive a treatment bottleneck, sustaining DNA damage in the process. The only treatment the patient received after the liver biopsy (which did not contain signature 17) was a pan aurora kinase inhibitor. There was no evidence of response to this therapy. It is unclear how an aurora kinase inhibitor could cause this mutational signature. Signatures 24 and 29 also show activity in TN1 samples. Signature 29 has been associated with the mutagenic effect of tobacco chewing on the oral mucosa, and, although distinct from the classic smoking associated signature 4, there is a high degree of overlap. Signature 24 is reported to occur with aflatoxin exposure. The significance of these carcinogen-induced signatures in a triple negative breast cancer is unclear.
ER2 was dominated by signatures 2 and 13 in the COSMIC classification [39], which both relate to the well-known phenomenon of APOBEC deaminase activity (Fig 6D). This signature was present consistently in both early and late clones and allows us to study how APOBEC may contribute to the evolution of breast cancer. Key driver mutations AKT1 E17K, SPEN, and the ESR1 E380Q mutation are consistent with those induced by APOBEC. However, the ARID1A and the TP53 truncating mutations have a low probability of arising from the APOBEC signature, and the ESR1 S463P and D538G mutations are not consistent with APOBEC. In this case, therefore, APOBEC-related activity alone was able to furnish most of the putative driver mutations and resulted in an ESR1 mutation that causes resistance to endocrine therapy but could not explain all of the ESR1 mutations.
In this study, we analyse four cases chosen to represent difficult clinical scenarios in the contemporary management of advanced breast cancer. We took the approach of extensively sampling metastatic disease at the time of autopsy, performing whole exome sequencing and SNP arrays coupled with high depth validation of discovered mutations. We also analysed samples taken from the primary tumour and, where available, metastatic lesions biopsied whilst patients were still alive. In line with other work utilising the CASCADE program, we found studying advanced disease at the time of death to be highly informative [48]. We found significant heterogeneity present across multiple metastatic sites, and by performing subclonal inference, it was possible to understand the key processes that drive tumour evolution over time. We have shown several novel findings, including how treatment shapes clonal evolution, the importance of mutational processes over a disease course, and augmentation of oncogenic signalling as a mechanism of treatment failure.
Late relapse is a significant problem for ER-positive disease, which continues to show a decline in survival beyond 10 years from diagnosis. In case ER1 with a long clinical latency, we observed that the subclone giving rise to metastatic disease was not detectable at diagnosis with standard sequencing and sampling measures. This has important implications for genomic determinism, or the expectation that genomic assays from a single time point are able to predict clinical outcome and guide therapy. In this case, additional oncogenic drivers from mutations and CNA accumulated before metastatic disease occurred. The long time to recurrence may be because micrometastatic disease remained dormant until acquiring additional drivers, or that the intermediate clone was present early in the natural history but was suppressed by tamoxifen. Although these possibilities cannot be distinguished here, other studies have found evidence that the primary tumour population contains low-frequency subclones that may give rise to metastatic or recurrent disease [49]. Our results are similar to those obtained in another case of recurrent disease after a long latency in lobular breast cancer [1]. Assuming a model in which infrequent subclones may give rise to eventual relapse 5 years or more after initial therapy, prolonging adjuvant endocrine therapy is expected to be beneficial, as has been shown in clinical trials [50,51]. Detecting and eliminating these rare subclones is a difficult proposition, but targeting known clonal driver alterations could be one strategy. These driver alterations could also be used to monitor for disease recurrence via ctDNA.
To our knowledge, ER2 is the first case of de novo metastatic ER-positive/HER2-negative disease to undergo longitudinal sampling of metastases and autopsy along with ctDNA assessments. In contrast to ER1, ER2 showed less divergence between the primary and metastatic lesions. The most likely explanation for this pattern is that metastatic potential was available early in the evolutionary history of this case. Another explanation for this could be reseeding of the primary tumours from a metastatic niche, as has been reported elsewhere [52]. This case did display subclonal patterns consistent with metastatic reseeding between para-aortic nodal sites. That this mechanism exists in breast cancer highlights the difficulties in accurately profiling the cancer genome, because even a lesion once biopsied may significantly change its subclonal structure. Circulating tumour DNA is one method to avoid sampling bias, as was illustrated in this case, in which biopsy of a liver lesion did not reveal any of the three ESR1 mutations that were detectable in plasma at the time.
In addition to ESR1, all three ER-positive cases showed alterations in SPEN after exposure to endocrine therapy. SPEN has recently been implicated as a novel tumour suppressor that regulates cell proliferation and inhibits oestrogen receptor downstream signalling, with a role in resistance to tamoxifen [53]. The truncal location of SPEN alterations in three different cases suggests this is a bona fide tumour suppressor and mediator of resistance to endocrine therapy.
In ER3, there was the unexpected finding that the earliest subclone was able to metastasise to the ovary. This case is also notable for the widespread copy number derangement that was present in all samples, and, similar to TN1, this may have conferred the metastatic phenotype. ER3 displayed ongoing copy number changes that segregated with subclones in our analysis. This was in contrast to the other ER-positive cases, in which mutational processes dominated. The presence of CIN has been shown in vitro to be associated with drug resistance, and there is some evidence that aneuploidy is an adaptive response in lower eukaryotes [54,55]. There is increasing evidence for punctuated evolution of copy number in cancers [8,56]. This fits well with our data for ER1, ER2, and TN1. The situation for ER3 is less clear cut, as there is ongoing acquisition of copy number changes beyond increases in ploidy. In some cancers, therefore, “active” CIN is likely to be an important mechanism driving evolution, treatment resistance, and heterogeneity. Notably, during treatment with tamoxifen and chemotherapy, ER3 showed discordant responses between lesions in liver and lung. This could be explained by the significant copy number heterogeneity seen in this case, although it is not possible to trace a premortem lesion on imaging to a lesion at autopsy.
Understanding mutational signatures provided valuable insights for ER2 and TN1. In TN1, the late emergence of signature 17 implies rapid rescaling of the subclonal structure of several metastatic lesions. The aetiology of signature 17 is unknown, but the activity of this signature in unrelated subclones at different anatomical sites suggests it can be chemically induced. Furthermore, the presence of apparent carcinogen-induced mutational signatures in TN1 also raises questions about the aetiology of aggressive triple negative breast cancers. In ER2, APOBEC was responsible for many important alterations of functional significance, including the key founding driver mutation. It has been previously noted that APOBEC may give rise to clonal and subclonal driver mutations [57,58], but we demonstrate here that this mutational activity is maintained during treatment and throughout the natural history of the disease. Arresting the activity of APOBEC may be a potential strategy to restrain progression and evolution of APOBEC-enriched cancers.
We identified a potentially novel mechanism of broad treatment resistance to chemotherapy, which arises from augmentation of existing oncogenic signalling. This implies that oncogenic signalling remains essential even in heavily pretreated disease and raises the possibility that combining targeted therapies with chemotherapy may severely restrict the fitness landscape that a tumour can access to achieve treatment resistance. The superior efficacy of such combination therapy is well known, and this approach has been used with great success in HER2-positive breast cancer, for example [59]. Other studies have shown that driver alterations influence response to chemotherapy [7]. Our findings extend these concepts to define augmented oncogenic signalling as a resistance mechanism that is widespread in advanced disease and may be therapeutically tractable.
There are several limitations to this study. WES alone may result in poor resolution to detect subclones, which could underestimate the subclonal diversity present. In contrast to mutations, detection of subclonal copy number events remains difficult, and important subclonal amplifications or deletions may have been missed. Newer technologies, such as single-cell approaches or long read sequencing, will be required to overcome these limitations. In addition, we did not analyse structural variants, the transcriptome, the epigenome, or the proteome, which along with noncoding elements such as long noncoding RNAs or microRNAs could make important contributions to subclonal evolution and phenotypic diversity not captured by WES. Although we have analysed a small number of cases in detail, it is unclear whether our findings are representative of the broader patient population.
In conclusion, we demonstrate the feasibility and value of subclonal inference in understanding the biology and evolutionary history of lethal breast cancer. This approach also provided insight into difficult clinical scenarios in breast cancer. Extension of the whole exome approach to whole genome, transcriptome, and methylome studies, as well as more novel single-cell and long read sequencing technology, is expected to provide further insights, particularly when used in the setting of rapid autopsy studies, which afford the unprecedented ability to sample multiple evolutionary trajectories comprehensively. It is notable that each case studied was unique in the processes that ultimately resulted in death. It is unclear if patterns will be found that can generalize across patient subgroups: for this, we will need large cohorts for which we can track genomic evolution from diagnosis to death. To that end, our prospective rapid autopsy program, which continues to accrue in breast cancer and other cancer types [9], as well as other international efforts will be essential to help us understand how cancer disseminates and ultimately becomes resistant to treatment [60].
|
10.1371/journal.pgen.1008035 | Polygenic adaptation: From sweeps to subtle frequency shifts | Evolutionary theory has produced two conflicting paradigms for the adaptation of a polygenic trait. While population genetics views adaptation as a sequence of selective sweeps at single loci underlying the trait, quantitative genetics posits a collective response, where phenotypic adaptation results from subtle allele frequency shifts at many loci. Yet, a synthesis of these views is largely missing and the population genetic factors that favor each scenario are not well understood. Here, we study the architecture of adaptation of a binary polygenic trait (such as resistance) with negative epistasis among the loci of its basis. The genetic structure of this trait allows for a full range of potential architectures of adaptation, ranging from sweeps to small frequency shifts. By combining computer simulations and a newly devised analytical framework based on Yule branching processes, we gain a detailed understanding of the adaptation dynamics for this trait. Our key analytical result is an expression for the joint distribution of mutant alleles at the end of the adaptive phase. This distribution characterizes the polygenic pattern of adaptation at the underlying genotype when phenotypic adaptation has been accomplished. We find that a single compound parameter, the population-scaled background mutation rate Θbg, explains the main differences among these patterns. For a focal locus, Θbg measures the mutation rate at all redundant loci in its genetic background that offer alternative ways for adaptation. For adaptation starting from mutation-selection-drift balance, we observe different patterns in three parameter regions. Adaptation proceeds by sweeps for small Θbg ≲ 0.1, while small polygenic allele frequency shifts require large Θbg ≳ 100. In the large intermediate regime, we observe a heterogeneous pattern of partial sweeps at several interacting loci.
| It is still an open question how complex traits adapt to new selection pressures. While population genetics champions the search for selective sweeps, quantitative genetics proclaims adaptation via small concerted frequency shifts. To date the empirical evidence of clear sweep signals is more scarce than expected, while subtle shifts remain notoriously hard to detect. In the current study we develop a theoretical framework to predict the expected adaptive architecture of a simple polygenic trait, depending on parameters such as mutation rate, effective population size, size of the trait basis, and the available genetic variability at the onset of selection. For a population in mutation-selection-drift balance we find that adaptation proceeds via complete or partial sweeps for a large set of parameter values. We predict adaptation by small frequency shifts for two main cases. First, for traits with a large mutational target size and high levels of genetic redundancy among loci, and second if the starting frequencies of mutant alleles are more homogeneous than expected in mutation-selection-drift equilibrium, e.g. due to population structure or balancing selection.
| Rapid phenotypic adaptation of organisms to all kinds of novel environments is ubiquitous and has been described and studied for decades [1, 2]. However, while the macroscopic changes of phenotypic traits are frequently evident, their genetic and genomic underpinnings are much more difficult to resolve. Two independent research traditions, molecular population genetics and quantitative genetics, have coined two opposite views of the adaptive process on the molecular level: adaptation either by selective sweeps or by subtle allele frequency shifts (sweeps or shifts from here on).
On the one hand, population genetics works bottom-up from the dynamics at single loci, without much focus on the phenotype. The implicit assumption of the sweep scenario is that selection on the trait results in sustained directional selection also on the level of single underlying loci. Consequently, we can observe phenotypic adaptation at the genotypic level, where selection drives allele frequencies at one or several loci from low values to high values. Large allele frequency changes are the hallmark of the sweep scenario. If these frequency changes occur in a short time interval, conspicuous diversity patterns in linked genomic regions emerge: the footprints of hard or soft selective sweeps [3–6].
On the other hand, quantitative genetics envisions phenotypic adaptation top-down, from the vantage point of the trait. At the genetic level, it is perceived as a collective phenomenon that cannot easily be broken down to the contribution of single loci. Indeed, adaptation of a highly polygenic trait can result in a myriad of ways through “infinitesimally” small, correlated changes at the interacting loci of its basis (e.g. [1, 7, 8]. Conceptually, this view rests on the infinitesimal model by Fisher (1918) [9] and its extensions (e.g. [10]). Until a decade ago, the available moderate sample sizes for polymorphism data had strongly limited the statistical detectability of small frequency shifts. Therefore, the detection of sweeps with clear footprints was the major objective for many years. Since recently, however, huge sample sizes (primarily of human data) enable powerful genome-wide association studies (GWAS) to resolve the genomic basis of polygenic traits. Consequently, following conceptual work by Pritchard and coworkers [7, 11], there has been a shift in focus to the detection of polygenic adaptation from subtle genomic signals (e.g. [12–14], reviewed in [15]). Very recently, however, some of the most prominent findings of polygenic adaptation in human height have been challenged [16, 17]. As it turned out, the methods are highly sensitive to confounding effects in GWAS data due to population stratification.
While discussion of the empirical evidence is ongoing, the key objective for theoretical population genetics is to clarify the conditions (mutation rates, selection pressures, genetic architecture) under which each adaptive scenario, sweeps, shifts—or any intermediate type—should be expected in the first place. Yet, the number of models in the literature that allow for a comparison of alternative adaptive scenarios at all is surprisingly limited (see also [18]). Indeed, quantitative genetic studies based on the infinitesimal model or on summaries (moments, cumulants) of the breeding values do not resolve allele frequency changes at individual loci (e.g. [19–22]). In contrast, sweep models with a single locus under selection in the tradition of Maynard Smith and Haigh [3], or models based on adaptive walks or the adaptive dynamics framework (e.g. [23–25]) only allow for adaptive substitutions or sweeps. A notable exception is the pioneering study by Chevin and Hospital [26]. Following Lande [27], these authors model adaptation at a single major quantitative trait locus (QTL) that interacts with an “infinitesimal background” of minor loci, which evolves with fixed genetic variance. Subsequent models [28, 29] trace the allele frequency change at a single QTL in models with 2-8 loci. Still, these articles do not discuss polygenic adaptation patterns. Most recently, Jain and Stephan [30, 31] studied the adaptive process for a quantitative trait under stabilizing selection with explicit genetic basis. Their analytical approach allows for a detailed view of allele frequency changes at all loci without constraining the genetic variance. However, the model is deterministic and thus ignores the effects of genetic drift. Below, we study a polygenic trait that can adapt via sweeps or shifts under the action of all evolutionary forces in a panmictic population (mutation, selection, recombination and drift). Our model allows for comprehensive analytical treatment, leading to a multi-locus, non-equilibrium extension of Wright’s formula [32] for the joint distribution of allele frequencies at the end of the adaptive phase. This way, we obtain predictions concerning the adaptive architecture of polygenic traits and the population genetic variables that delimit the corresponding modes of adaptation.
The article is organized as follows. The Model section motivates our modeling decisions and describes the simulation method. We also give a brief intuitive account of our analytical approach. In the Results part, we describe our findings for a haploid trait with linkage equilibrium among loci. All our main conclusions in the Discussion part are based on the results displayed here. Further model extensions and complications (diploids, linkage, and alternative starting conditions) are relegated to appendices. Finally, we describe our analytical approach and derive all results in a comprehensive Mathematical Appendix (S2 Appendix). For the ease of reading, we have tried to keep both the main text and the Mathematical Appendix independent and largely self-contained.
In the current study, we aim for a “minimal model” of a trait that allows us to clarify which evolutionary forces favor sweeps over shifts and vice versa (as well as any intermediate patterns). For shifts, alleles need to be able to hamper the rise of alleles at other loci via negative epistasis for fitness, e.g. diminishing returns epistasis. Indeed, otherwise one would only observe parallel sweeps. Negative fitness epistasis is frequently found in empirical studies (e.g. [33]) and implicit to the Gaussian selection scheme (e.g. [26, 30, 31]). More fundamentally, diminishing returns are a consequence of partial or complete redundancy of genetic effects across loci or gene pathways. Adaptive phenotypes (such as pathogen resistance or a beneficial body coloration) can often be produced in many alternative ways, such that redundancy is a common characteristic of beneficial mutations.
As our basic model, we focus on a haploid population and study adaptation for a polygenic, binary trait with full redundancy of effects at all loci. We assume a non-additive genotype-phenotype map where any single mutation switches the phenotype from its ancestral state (e.g. “non-resistant”) to the adaptive state (“resistant”). Further mutations have no additional effect. On the population level, adaptation can be produced by a single locus where the beneficial allele sweeps to fixation, or by small frequency shifts of alleles at many different loci in different individuals—or any intermediate pattern. The symmetry among loci (no build-in advantage of any particular locus) and complete redundancy of locus effects provides us with a trait architecture that is favorable for collective adaptation via small shifts—and with a modeling framework that allows for analytical treatment. The same model has been used in a preliminary simulation study [6]. In the context of parallel adaptation in a spatially structured population, analogous model assumptions with redundant loci have been used [34–36]. In a second step, we extend our basic model to relax the redundancy condition, as described below.
Consider a panmictic population of Ne haploids, with a binary trait Z (with phenotypic states Z0 “non-resistant” and Z1 “resistant”, see Fig 1). The trait is governed by a polygenic basis of L bi-allelic loci with arbitrary linkage (we treat the case of linkage equilibrium in the main text and analyze the effects of linkage in S1 Appendix, Section A). Only the genotype with the ancestral alleles at all loci produces phenotype Z0, all other genotypes produce Z1, irrespective of the number of mutations they carry. Loci mutate at rate μi, 1 ≤ i ≤ L, per generation (population mutation rate at the ith locus: 2Ne μi = Θi) from the ancestral to the derived allele. We ignore back mutation. The mutant phenotype Z1 is deleterious before time t = 0, when the population experiences a sudden change in the environment (e.g. arrival of a pathogen). Z1 is beneficial for time t > 0. The Malthusian (logarithmic) fitness function of an individual with phenotype Z reads
W ( Z ) = { s d Z for t < 0 s b Z for t ≥ 0 . (1)
Without loss of generality, we can assume Z0 = 0 and Z1 = 1. We then have W(Z0) = 0. Furthermore, W(Z1) = sd < 0, respectively W(Z1) = sb > 0, measure the strength of directional selection on Z (e.g. cost and benefit of resistance) before and after the environmental change. For the basic model, we assume that the population is in mutation-selection-drift equilibrium at time t = 0.
We extend the basic model in several directions. This includes linkage (S1 Appendix, Section A), alternative starting conditions at time t = 0 (S1 Appendix, Section B), diploids (S1 Appendix, Section C), and arbitrary time-dependent selection s(t) (S2 Appendix, Section M.1). Here, we describe how we relax the assumption of complete redundancy of all loci. Diminishing returns epistasis, e.g. due to Michaelis-Menten enzyme kinetics, will frequently not lead to complete adaptation in a single step, but may require multiple steps before the trait optimum is approached. In a model of incomplete redundancy, we thus assume that a first beneficial mutation only leads to partial adaptation. We thus have three states of the trait, the ancestral state for the genotype without mutations, Z0 = 0 (non-resistant), a phenotype Zδ = δ (partially resistant) for genotypes with a single mutation, and the mutant state Z1 = 1 (fully resistant) for all genotypes with at least two mutations, see Fig 1(b). For diminishing returns epistasis, we require 1 2 ≤ δ < 1. The fitness function is as in Eq (1). A model with asymmetries in the single-locus effects is discussed in S1 Appendix, Section D.
For the models described above, we use Wright-Fisher simulations for a haploid, panmictic population of size Ne, assuming linkage equilibrium between all L loci in discrete time. Selection and drift are implemented by independent weighted sampling based on the marginal fitnesses of the ancestral and mutant alleles at each locus. Due to linkage equilibrium, the marginal fitnesses only depend on the allele frequencies and not genotypes. Ancestral alleles mutate with probability μi per generation at locus i. We start our simulations with a population that is monomorphic for the ancestral allele at all loci. The population evolves for 8Ne generations under mutation and deleterious selection to reach (approximate) mutation-selection-drift equilibrium. Following [6, 37], we condition on adaptation from the ancestral state and discard all runs where the deleterious mutant allele (at any locus) reaches fixation during this time. (We do not show results for cases with very high mutation rates and weak deleterious selection when most runs are discarded). At the time of environmental change, selection switches from negative to positive and simulation runs are continued until a prescribed stopping condition is reached.
We are interested in the genetic architecture of adaptation—the joint distribution of mutant frequencies across all loci—at the end of the rapid adaptive phase. Following [31], we define this phase as “the time until the phenotypic mean reaches a value close to the new optimum”. Specifically, we stop simulations when the mean fitness W ¯ in the population has increased up to a proportion fw of the maximal attainable increase from the ancestral to the derived state,
W ( Z 1 ) - W ¯ W ( Z 1 ) - W ( Z 0 ) = f w . (2)
For the basic model with complete redundancy, this simply corresponds to a residual proportion fw of individuals with ancestral phenotype in the population. Extensions of the simulation scheme to include linkage or diploid individuals are described in S1 Appendix, Sections A and C.
Parameter choices: Unless explicitly stated otherwise, we simulate Ne = 10 000 individuals, with beneficial selection coefficients sb = 0.1 and 0.01, combined with deleterious selection coefficients sd = −0.1 and sd = −0.001 for low and high levels of SGV, respectively. (The corresponding Wrightian fitness values used as sampling weights in discrete time are 1 + sb and 1 + sd.) We investigate L = 2 to 100 loci. We usually (except in S1 Appendix, Section D) assume equal mutation rates at all loci, μi = μ and define Θl = 2Ne μ as the locus mutation parameter. Mutation rates are chosen such that Θbg ≔ 2Ne μ(L − 1) (the background mutation rate, formally defined below in Eq (10)) takes values from 0.01 to 100. We typically simulate 10 000 replicates per mutation rate and stop simulations when the population has reached the new fitness optimum up to fw = 0.05. In the model with complete redundancy, we thus stop simulations when the frequency of individuals with mutant phenotype Z1 has increased to 95%. Different stopping conditions are explored in S1 Appendix, Section G.
We partition the adaptive process into two phases (see Fig 2 for illustration). An initial stochastic phase, governed by selection, drift, and mutation describes the origin and establishment of mutant alleles at all loci. We call mutants “established” if they are not lost again due to genetic drift. The subsequent deterministic phase governs the further evolution of established alleles until the stopping condition is reached as described above. While mutation and drift can be ignored during the deterministic phase, interaction effects due to epistasis and linkage become important (in our model, they enter, in particular, through the stopping condition). We give a brief overview of our analytical approach below; parameters are summarized in Table 1. A detailed account with the derivation of all results is provided in the Mathematical Appendix S2 Appendix.
During the stochastic phase, we model the origin and spread of mutant copies as a so-called Yule pure birth process following [38] and [39]. The idea of this approach is that we only need to keep track of mutations that found “immortal lineages”, i.e. derived alleles that still have surviving offspring at the time of observation (see Fig 2 for the case of L = 2 loci). Forward in time, new immortal lineages can be created by two types of events: new mutations at all loci start new lineages, while birth events lead to splits of existing lineages into two immortal lineages. For t > 0 (after the environmental change), in particular, new mutations at the ith locus arise at rate Neμi per generation and are destined to become established in the population with probability ≈ 2sb. Similarly, birth of new immortal lineages due to split events in the Yule process occur at rate sb (because the selection coefficient measures the excess of births over deaths in the underlying population). For the origin of new immortal lineages in the Yule process and their subsequent splitting we thus obtain the rates
p mut , i ≈ N e μ i · 2 s b = Θ i s b ; p split ≈ s b . (3)
Extended results including standing genetic variation and time-dependent fitness are given in the Appendix. Assume now that there are currently {k1, …kL}, 0 ≤ kj ≪ Ne mutant lineages at the L loci. The probability that the next event (which can be a split or a mutation) occurs at locus i is
k i · p split + p mut , i ∑ j = 1 L ( k j · p split + p mut , j ) = k i + Θ i ∑ j = 1 L ( k j + Θ j ) . (4)
Importantly, all these transition probabilities among states of the Yule process are constant in time and independent of the mutant fitness sb, which cancels in the ratio of the rates. As the number of lineages at all loci increases, their joint distribution (across replicate realizations of the Yule process) approaches a limit. In particular, as shown in the Appendix, the joint distribution of frequency ratios xi ≔ ki/k1 in the limit k1 → ∞ is given by an inverted Dirichlet distribution
P inDir [ x | Θ ] = 1 B [Θ] ∏ j = 2 L x j Θ j - 1 ( 1 + ∑ i = 2 L x i ) - ∑ i = 1 L Θ i (5)
where x = (x2, …, xL) and Θ = (Θ1, …, ΘL) are vectors of frequency ratios and locus mutation rates, respectively, and where B [ Θ ] = ∏ j = 1 L Γ ( Θ j ) ∑ j = 1 L Γ ( Θ j ) is the generalized Beta function and Γ(z) is the Gamma function. Note that Eq (5) depends only on the locus mutation rates, but not on selection strength.
After the initial stochastic phase, the dynamics of established mutant lineages that have evaded stochastic loss can be adequately described by deterministic selection equations. For allele frequencies pi at locus i, assuming linkage equilibrium, we obtain (consult S2 Appendix, Section M.1, Eq (M.2a), for a detailed derivation)
p ˙ i = p i ( W ( Z 1 ) - W ¯ ) = s b p i ( Z 1 - Z ¯ ) , (6)
where W ¯ and Z ¯ are population mean fitness and mean trait value. For the mutant frequency ratios xi = pi/p1, we obtain
x ˙ i = d d t ( p i p 1 ) = p ˙ i p 1 - p i p ˙ 1 p 1 2 = 0 . (7)
We thus conclude that the frequency ratios xi do not change during the deterministic phase. In particular, this means that Eq (5) still holds at our time of observation at the end of the rapid adaptive phase. This is even true with linked loci. Finally, derivation of the joint distribution of mutant frequencies pi (instead of frequency ratios xi) at the time of observation requires a transformation of the density. In general, this transformation depends on the stopping condition fw and on other factors such as linkage. Assuming linkage equilibrium among all selected loci, we obtain (see S2 Appendix, Theorem 2, Eq (M.20))
P f w [ p | Θ ] = δ ∏ j = 1 L ( 1 - p j ) - f w B [ Θ ] ∏ j = 1 L p j Θ j - 1 ( ∑ i = 1 L p i ) - ∑ i = 1 L Θ i ( ∑ j = 1 L f w p j 1 - p j ) (8)
for p = (p1, …, pL) in the L-dimensional hypercube of allele frequencies. The delta function δX restricts the distribution to the L − 1 dimensional manifold defined via the stopping condition f w = ∏ j = 1 L ( 1 - p j ). Further expressions, also including linkage, are given in S2 Appendix and in S1 Appendix, Section A. In general, the joint distribution corresponds to a family of generalized Dirichlet distributions.
We assess the adaptive architecture not as a function of time, but as a function of progress in phenotypic adaptation, measured by fw, Eq (2). Hence, fw rather than time t plays the role of a dynamical variable in the joint distribution, see Eq (8). In the special case fw → 0 (i.e. complete adaptation, enforcing fixation at at least one locus), this distribution is restricted to a boundary face of the allele frequency hypercube and Eq (8) reduces to the inverted Dirichlet distribution given above in Eq (5). In the Results section below, we assess our analytical approximations for the joint distributions of adaptive alleles, Eqs (5) and (8), and discuss their implications in the context of scenarios of polygenic adaptation, ranging from sweeps to small frequency shifts.
While the joint distribution of allele frequencies, Eq (8), provides comprehensive information of the adaptive architecture, low-dimensional summary statistics of this distribution are needed to describe and classify distinct types of polygenic adaptation. To this end, we order loci according to their contribution to the adaptive response. In particular, we call the locus with the highest allele frequency at the stopping condition the major locus and all other loci minor loci. Minor loci are further ordered according to their frequency (first minor, second minor, etc.). The marginal distributions of the major locus or kth minor locus are 1-dimensional summaries of the joint distribution. Importantly, these summaries are still collective because the role of any specific locus (its order) is defined through the allele frequencies at all loci. This is different for the marginal distribution at a fixed focal locus, which is chosen irrespective of its role in the adaptive process, e.g. [26, 28, 29].
Concerning our nomenclature, note that the major and minor loci do not differ in their effect size, as they are completely redundant. Still, the major locus is the one with the largest contribution to the adaptive response and would yield the strongest association in a GWAS case-control study.
In the following, we analyze adaptive trait architectures in three steps. In the Section Expected allele frequency ratio, we use the expected allele frequency ratio of minor and major loci as a one-dimensional summary statistic. Subsequently, in Section Genomic architecture of polygenic adaptation, we analyze the marginal distributions of major and minor loci for a trait with 2 to 100 loci. Finally, in Section Relaxing complete redundancy, we investigate the robustness of our results under conditions of relaxed redundancy. Further results devoted to diploids, linkage, asymmetric loci, and alternative starting conditions are provided in S1 Appendix.
For our biological question concerning the type of polygenic adaptation, the ratio of allele frequency changes of minor over major loci is particularly useful. With “sweeps at few loci”, we expect large differences among loci, resulting in ratios that deviate strongly from 1. In contrast, with “subtle shifts at many loci”, multiple loci contribute similarly to the adaptive response and ratios should range close to 1. Our theory (explained above) predicts that these ratios are the outcome of the stochastic phase, and their distribution is preserved during the deterministic phase. They are thus independent of the precise time of observation. For our results in this section, we assume that the mutation rate at all L loci is equal, Θi ≡ Θl, for all 1 ≤ i ≤ L. This corresponds to the symmetric case that is most favorable for a “small shift” scenario. Results for asymmtric mutation rates are reported in Appendix S1 Appendix, Section D.
Consider first the case of L = 2 loci. There is then a single allele frequency ratio “minor over major locus”, which we denote by x. For two loci, the joint distribution of frequency ratios from Eq (5) reduces to a beta-prime distribution. Conditioning on the case that the first locus is the major locus (probability 1/2 for the symmetric model), we obtain for 0 ≤ x ≤ 1,
P β ′ [ x | Θ l ] = 2 Γ ( 2 Θ l ) ( Γ ( Θ l ) ) 2 x Θ l - 1 ( 1 + x ) - 2 Θ l , (9)
Fig 3 compares the expectation of this analytical prediction with simulation results for a range of parameters for the strength of beneficial selection sb and for the level of standing genetic variation (SGV implicitly given by the strength of deleterious selection sd before the environmental change). There are two main observations. First, the simulation results demonstrate the importance of the scaled mutation rate Θbg ≡ Θl (for two loci). Low Θbg leads to sweep-like adaptation (heterogeneous adaptation response among loci, E[x] ≪ 1), whereas high Θbg leads to shift-like adaptation (homogeneous response, E[x] near 1). Second, the panels show that the selection intensity given by sd and sb has virtually no effect. Both results are predicted by the analytical theory (Eq (9)). In S1 Appendix, Section A, we further show that these results hold for arbitrary degrees of linkage (including complete linkage).
For more than two loci, L > 2, one-dimensional marginal distributions of the joint distribution, Eq (5), generally require (L − 1)-fold integration, which can be complicated. However, it turns out that the key phenomena to characterize the adaptive architecture can still be captured by the 2-locus formalism, with appropriate rescaling of the mutation rate. For the general L-locus model, we broaden our definition of the summary statistic x above to describe the allele frequency ratio of the first minor locus and the major locus. To relate the distribution of x in the L-locus model to the one in the 2-locus model, we reason as follows: For small locus mutation rates Θl, the order of the loci is largely determined by the order at which mutations that are destined for establishment originate at these loci. I.e., the locus where the first mutation originates ends up as the major locus and the first minor locus is usually the second locus where a mutation destined for establishment originates. The distribution of the allele frequency ratio x is primarily determined by the distribution of the waiting time for this second mutation after origin of the first mutation at the major locus. In the 2-locus model, this time will be exponentially distributed, with parameter 1/Θl. In the L-locus model, however, where L − 1 loci with total mutation rate Θl(L − 1) compete for being the “first minor”, the parameter for the waiting-time distribution reduces to 1/(Θl(L − 1)). We thus see from this argument that the decisive parameter is the cumulative background mutation rate
Θ b g = ( L - 1 ) Θ l (10)
at all minor loci in the background of the major locus. In Fig 3 (orange dots) we show simulations of a L = 10 locus model with an appropriately rescaled locus mutation rate Θl → Θl/9, such that the background rate Θbg is the same as for the 2-locus model. We see that the analytical prediction based on the 2-locus model provides a good fit for the 10-locus model. A more detailed discussion of this type of approximation is given in S1 Appendix, Section F.
While the distribution of allele frequency ratios, Eqs (5) and (9), offers a coarse (but robust) descriptor of the adaptive scenario, the joint distribution of allele frequencies at the end of the adaptive phase, Eq (8), allows for a more refined view. In contrast to the distribution of ratios, the results now depend explicitly on the stopping condition (the time of observation) and on linkage among loci. We assume linkage equilibrium in this section and assess the mutant allele frequencies when the frequency of the remaining wildtype individuals in the population has dropped to a fixed value of fw = 0.05. In S1 Appendix, Section G, we complement these results and study the changes in the adaptive architecture when fw is varied.
Fig 4 displays the main result of this section. It shows the marginal distributions of all loci, ordered according to their allele frequency at the time of observation (major locus, 1st, 2nd, 3rd minor locus, etc.) for traits with L = 2, 10, 50, and 100 loci. Panels in the same row correspond to equal background mutation rates Θbg = (L − 1)Θl, but note that the locus mutation rates Θl are not equal. The figure reveals a striking level of uniformity of adaptive architectures with the same Θbg, but vastly different number of loci. For Θbg ≤ 1 (the first three rows), the marginal distributions for loci of the same order (same color in the Figure) across traits with different L is almost invariant. For large Θbg, they converge for sufficiently large L (e.g. for Θbg = 10, going from L = 10 to L = 50 and to L = 100). In particular, the background mutation rate Θbg determines the shape of the major-locus distribution (red in the Figure) for high p → 1 − fw = 0.95 (the maximum possible frequency, given the stopping condition). For Θbg < 1, this distribution is sharply peaked with a singularity at p = 1 − fw, whereas it drops to zero for high p if Θbg > 1 (see also the analytical results below).
As predicted by the theory, Eq (8) and below, simulations confirm that the overall selection strength does not affect the adaptive architecture (see S1 Fig for comparison of simulation results for sb = 0.1 and sb = 0.01). As discussed in S1 Appendix, Section A, sufficiently tight linkage does change the shape of the distributions. Importantly, however, it does not affect the role of Θbg in determining the singularity of the major-locus distribution. This confirms the key role of the background mutation rate as a single parameter to determine the adaptive scenario in our model. While Θbg = 1 separates architectures that are dominated by a single major locus (Θbg < 1) from collective scenarios (with Θbg > 1), the classical sweep or shift scenarios are only obtained if Θbg deviates strongly from 1. We therefore distinguish three adaptive scenarios.
To complete our picture of adaptive architectures, we investigate the robustness of our model assumption against relaxation of redundancy. As explained above (Model extensions and Fig 1), we implement diminishing returns epistasis, such that an individual with a single mutation has fitness δsb/d, while individuals carrying more than one mutation have fitness sb/d. With small deviations from complete redundancy (e.g. δ = 0.9, stopping at 5% ancestral phenotypes, see Fig S2 Fig) we obtain basically no differences in the genomic patterns of adaptation. With larger deviations (e.g. δ = 0.5) quantitative differences appear. However, the qualitative picture concerning the scenario of polygenic adaptation remains the same.
Fig 5 shows the marginal frequency distributions of major and minor loci for a trait with relaxed redundancy with δ = 0.5 that is sampled when the population has accomplished 95% of the fitness increase on its way to the new optimum, Eq (2). Given the fitness function, this is not possible with adaptation at only a single locus. At least two loci are needed. The Figure compares the simulation data for the relaxed redundancy model (colored dots) and the full redundancy model (dots in back and gray). As in Fig 4, traits in the same row have the same background mutation rate Θbg. However, the background rate for the model with relaxed redundancy is redefined as
Θ b g relax = ( L - 2 ) Θ l , (15)
where Θl is the locus mutation rate (equal at all loci). We thus define the background rate, more precisely, as the combined population-scaled mutation rate of all loci that are not essential to accomplish adaptation of the phenotype and, thus, are truly redundant. With this choice, the adaptive architecture of the relaxed redundancy model reproduces the one of the model with full redundancy—up to a shift in the number of the loci due to an extra locus that is needed for adaptation with relaxed redundancy. The Figure captures this by comparing traits with relaxed redundancy with L = 3, 4, 11, and 101 loci to fully redundant traits with one fewer locus. The inset figures in the column for L = 4 loci show the same scenario, but with an averaged marginal distribution for the two largest loci with relaxed redundancy (in green).
In summary, our results show that relaxing redundancy leads to qualitatively similar results, but with a reduced “effective” background mutation rate that only accounts for “truly redundant” loci.
Traits with a polygenic basis can adapt in different ways. Few or many loci can contribute to the adaptive response. The changes in the allele frequencies at these loci can be large or small. They can be homogeneous or heterogeneous. While molecular population genetics posits large frequency changes—selective sweeps—at few loci, quantitative genetics views polygenic adaptation as a collective response, with small, homogeneous allele frequency shifts at many loci. Here, we have explored the conditions under which each adaptive scenario should be expected, analyzing a polygenic trait with redundancy among loci that allows for a full range of adaptive architectures: from sweeps to subtle frequency shifts.
For any polygenic trait, the multitude of possible adaptive architectures is fully captured by the joint distribution of mutant alleles across the loci in its basis. Different adaptive scenarios (such as sweeps or shifts) correspond to characteristic differences in the shape of this distribution, at the end of the adaptive phase. For a single locus, the stationary distribution under mutation, selection, and drift can be derived from diffusion theory and has been known since the early days of population genetics (S. Wright (1931), [32]). For multiple interacting loci, however, this is usually not possible. To address this problem for our model, we dissect the adaptive process into two phases. The early stochastic phase describes the establishment of all mutants that contribute to the adaptive response under the influence of mutation and drift. We use that loci can be treated as independent during this phase to derive a joint distribution for ratios of allele frequencies at different loci, Eq (5). During the second, deterministic phase, epistasis and linkage become noticeable, but mutation and drift can be ignored. Allele frequency changes during this phase can be described as a density transformation of the joint distribution. For the simple model with fully redundant loci, and assuming either LE or complete linkage, this transformation can be worked out explicitly. Our main result Eq (8) can be understood as a multi-locus extension of Wright’s formula. For a neutral locus with multiple alleles, Wright’s distribution is a Dirichlet distribution, which is reproduced in our model for the case of complete linkage, see S1 Appendix, Section A. For the opposite case of linkage equilibrium, we obtain a family of inverted Dirichlet distributions, depending on the stopping condition—our time of observation.
Note that (unlike Wright’s distribution) the distribution of adaptive architectures is not a stationary distribution, but necessarily transient. It describes the pattern of mutant alleles at the end of the “rapid adaptive phase” [30, 31], because this is the time scale that the opposite narratives of population genetics and quantitative genetics refer to. In particular, the quantitative genetic “small shifts” view of adaptation does not talk about a stationary distribution: it does not imply that alleles will never fix over much longer time scales, due to drift and weak selection. On a technical level, the transient nature of our result means that it reflects the effects of genetic drift only during the early phase of adaptation. These early effects are crucial because they are magnified by the action of positive selection. In contrast, our result ignores drift after phenotypic adaptation has been accomplished—which is also a reason why it can be derived at all.
To capture the key characteristics of the adaptive architecture, we dissect the joint distribution in Eq (8) into marginal distributions of single loci. As explained at the start of the results section, these loci do not refer to a fixed genome position, but are defined a posteriori via their role in the adaptive process. For example, the major locus is defined as the locus with the highest mutant allele frequency at the end of the adaptive phase. (Since all loci have equal effects in our model, this is also the locus with the largest contribution to the adaptive response, but see S1 Appendix, Section D.) This is a different way to summarize the joint distribution than used in some of the previous literature [26, 28, 29], which rely on a gene-centered view to study the pattern at a focal locus, irrespective of its role in trait adaptation. In contrast, we use a trait-centered view, which is better suited to describe and distinguish adaptive scenarios. For example, “adaptation by sweeps” refers to a scenario where sweeps happen at some loci, rather than at a specific locus. This point is further discussed in S1 Appendix, Section F, where we also display marginal distributions of Eq (8) for fixed loci.
The theme of “competition of a single locus with its background” relates to previous findings by Chevin and Hospital (2008) [26] in one of the first studies to address polygenic footprints. These authors rely on a deterministic model of an additive quantitative trait to describe the adaptive trajectory at a single target QTL in the presence of background variation. The background is modeled as a normal distribution with a mean that can respond to selection, but with constant variance. Obviously, a drift-related parameter, such as Θbg, has no place in such a framework. Still, there are several correspondences to our result on a qualitative level. Specifically, a sweep at the focal locus is prohibited under two conditions. First, the background variation (generated by recurrent mutation in our model, constant in [26]) must be large. Second, the fitness function must exhibit strong negative epistasis that allows for alternative ways to reach the trait optimum—and thus produces redundancy (due to Gaussian stabilizing selection in [26]). Finally, while the adaptive trajectory depends on the shape of the fitness function, Chevin and Hospital note that it does not depend on the strength of selection on the trait, as also found for our model.
A major difference of the approach used in [26] is the gene-centered view that is applied there. Consider a scenario where the genetic background “wins” against the focal QTL and precludes it from sweeping. For a generic polygenic trait (and for our model) this still leaves the possibility of a sweep at one of the background loci. However, this is not possible in [26], where all background loci are summarized as a sea of small-effect loci with constant genetic variance.
This constraint is avoided in the approach by deVladar and Barton [42] and Jain and Stephan [31], who study an additive quantitative trait under stabilizing selection with binary loci (see also [43] for an extension to adaptation to a moving optimum). These models allow for different locus effects, but ignore genetic drift. Before the environmental change, all allele frequencies are assumed to be in mutation-selection balance, with equilibrium values derived in [42]. At the environmental change, the trait optimum jumps to a new value and alleles at all loci respond by large or small changes in the allele frequencies. Overall, [42] and [31] predict adaptation by small frequency shifts in larger parts of the biological parameter space relative to our model. In particular, sweeps are prevented in [31] if most loci have a small effect and are therefore under weak selection prior to the environmental change. This contrasts to our model, where the predicted architecture of adaptation is independent of the selection strength. Thus, in our model, weak selection does not imply shifts. This difference can at least partially be explained by the neglect of drift effects on the starting allele frequencies in the deterministic models. In the absence of drift, loci under weak selection start out from frequency x0 = 0.5 [42]. In finite populations, however, almost all of these alleles start from very low (or very high) frequencies—unless the population mutation parameter is large (many alleles at intermediate frequencies at competing background loci are expected only if Θbg ≫ 1, in accordance with our criterion for shifts). To test this further, we have analyzed our model for the case of starting allele frequencies set to the deterministic values of mutation-selection balance, μ/sd. Indeed, we observe adaptation due to small frequency shifts in a much larger parameter range (S1 Appendix, Section B).
Generally, adaptation by sweeps in a polygenic model requires a mechanism to create heterogeneity among loci. This mechanism is entirely different in both modeling frameworks. While heterogeneity is (only) produced by unequal locus effects for the deterministic quantitative trait, it is (solely) due to genetic drift for the redundant trait model. Since both approaches ignore one of these factors, both results should rather underestimate the prevalence of sweeps. Indeed, heterogeneity increases for our model with unequal locus effects (see S1 Appendix, Section D).
Both drift and unequal locus effects are included in the simulation studies by Pavlidis et al (2012) [28] and Wollstein and Stephan (2014) [29]. These authors assess patterns of adaptation for a quantitative trait under stabilizing selection with up to eight diploid loci. However, due to differences in concepts and definitions there are few comparable results. In contrast to [31] and to our approach, they study long-term adaptation (they simulate Ne generations). In [28, 29], sweeps are defined as fixation of the mutant allele at a focal locus, whereas frequency shifts correspond to long-term stable polymorphic equilibria [29]. With this definition, a shift scenario is no longer a transient pattern, but depends entirely on the existence (and range of attraction) of polymorphic equilibria. A polymorphic outcome is likely for a two-locus model with full symmetry, where the double heterozygote has the highest fitness. For more than two loci, the probability of shifts decreases (because polymorphic equilibria become less likely, see [44]). However, also the probability of a sweep decreases. This is largely due to the gene-centered view in [28], where potential sweeps at background loci are not recorded (see also S1 Appendix, Section F).
We have described scenarios of adaptation for a simple model of a polygenic trait. This model allows for an arbitrary number of loci with variable mutation rates, haploids and diploids, linkage, time-dependent selection, new mutations and standing genetic variation, and alternative starting conditions for the mutant alleles. Its genetic architecture, however, is strongly restricted by our assumption of (full or relaxed) redundancy among loci. In the haploid, fully redundant version, the phenotype is binary and only allows for two states, ancestral wildtype and mutant. Biologically, this may be thought of as a simple model for traits like pathogen or antibiotic resistance, body color, or the ability to use a certain substrate [45, 46].
Our main motivation, however, has been to construct a minimal model with a polygenic architecture that allows for both sweep and shifts scenarios—and for comprehensive analytical treatment. One may wonder how our methods and results generalize if we move beyond our model assumptions.
Key to our analytical method is the dissection of the adaptive process into a stochastic phase that explains the origin and establishment of beneficial variants and a deterministic phase that describes the allele frequency changes of the established mutant copies. This framework can be applied to a much broader class of models. Indeed, in many cases, the fate of beneficial alleles, establishment or loss, is decided while these alleles are rare. Excluding complex scenarios such as passage through a fitness valley, the initial stochastic phase is relatively insensitive to interactions via epistasis or linkage. We can therefore describe the dynamics of traits with a different architecture (e.g. an additive quantitative trait with equal-effect loci under stabilizing selection) within the same framework by coupling the same stochastic dynamics to a different set of differential equations describing the dynamics during the deterministic phase.
This is important because, as described above, the key qualitative results to distinguish broad categories of adaptive scenarios are due to the initial stochastic phase. This holds true, in particular, for the role of the background mutation rate Θbg. We therefore expect that these results generalize beyond our basic model. Indeed, we have already seen this for our model extensions to include diploids, linkage, and relaxed redundancy. Vice-versa, we have seen that other factors, such as alternative starting conditions for the mutant alleles, directly affect the early stochastic phase and lead to larger changes in the results. As shown in S1 Appendix, Section B, however, they can be captured by an appropriate extension of the stochastic Yule process framework.
Several factors of biological importance are not covered by our current approach. Most importantly, this includes loci with different effect sizes and spatial population structure. Both require a further extension of our framework for the early stochastic phase of adaptation. Unequal locus effects (both directly on the trait or on fitness due to pleiotropy) are expected to enhance the heterogeneity in the adaptive response among loci, as confirmed by simulations of a 2-locus model in S1 Appendix, Section D. The opposite is true for spatial structure, as further discussed below.
Although our assumptions on the genetic architecture of the trait (complete redundancy and equal loci) are favorable for a collective, shift-type adaptation scenario, we observe large changes in mutant allele frequencies (completed or partial sweeps) for major parts of the parameter range. A homogeneous pattern of subtle frequency shifts at many loci is only observed for high mutation rates. This contrasts with experience gained from breeding and modern findings from genome-wide association studies, which are strongly suggestive of an important role for small shifts with contributions from very many loci (reviewed in [1, 15, 47–49], see [12, 50, 51] for recent empirical examples). For traits such as human height, there has even been a case made for omnigenic adaptation [8], setting up a “mechanistic narrative” for Fisher’s (conceptual) infinitesimal model. Clearly, body height may be an extreme case and the adaptive scenario will strongly depend on the type of trait under consideration. Still, the question arises whether and how wide-spread shift-type adaptation can be reconciled with our predictions. We will first discuss this question within the scope of our model and then turn to factors beyond our model assumptions.
|
10.1371/journal.pntd.0007535 | Chronic hepatitis B virus infection drives changes in systemic immune activation profile in patients coinfected with Plasmodium vivax malaria | Plasmodium vivax and Hepatitis B virus (HBV) are globally outspread in similar geographic regions. The concurrence of both infections and its association with some degree of protection against symptomatic and/or severe vivax malaria has been already described. Nevertheless, data on how host response to both pathogens undermines the natural progression of the malarial infection are scarce. Here, a large cohort of vivax malaria and HBV patients is retrospectively analyzed in an attempt to depict how inflammatory characteristics could be potentially related to the protection to severe malaria in coinfection.
A retrospective analysis of a databank containing 601 individuals from the Brazilian Amazon, including 179 symptomatic P. vivax monoinfected patients, 145 individuals with asymptomatic P. vivax monoinfection, 28 P. vivax-HBV coinfected patients, 29 HBV monoinfected subjects and 165 healthy controls, was performed. Data on plasma levels of multiple chemokines, cytokines, acute phase proteins, hepatic enzymes, bilirubin and creatinine were analyzed to describe and compare biochemical profiles associated to each type of infection.
Coinfected individuals predominantly presented asymptomatic malaria, referred increased number of previous malaria episodes than symptomatic vivax-monoinfected patients, and were predominantly younger than asymptomatic vivax-monoinfected individuals. Coinfection was hallmarked by substantially elevated concentrations of interleukin (IL)-10 and heightened levels of C-C motif chemokine ligand (CCL)2. Correlation matrices showed that coinfected individuals presented a distinct biomarker profile when compared to asymptomatic or symptomatic P. vivax patients, or HBV-monoinfected individuals. Parasitemia could distinguish coinfected from symptomatic or asymptomatic P. vivax-monoinfected patients. HBV viremia was associated to distinct inflammatory profiles in HBV-monoinfected and coinfected patients.
The findings demonstrate a distinct inflammatory profile in coinfected patients, with characteristics associated with immune responses to both pathogens. These host responses to P. vivax and HBV, in conjunction, could be potentially associated, if not mainly responsible, for the protection against symptomatic vivax malaria.
| The determinants of the diverse clinical presentations of Plasmodium vivax malaria are not completely understood. Previous studies have reported that P. vivax-HBV coinfection is associated with increased odds of presenting with asymptomatic malaria, but little is known about the immune mechanisms driving such association. To illuminate host pathways associated with protection against malaria, we analyzed multiple cytokines, chemokines and acute phase proteins in groups of patients from the Brazilian Amazon with different presentations of vivax malaria monoinfection, HBV monoinfection, and P. vivax-HBV coinfection. The results indicate that coinfection is hallmarked by a conjunction of immune responses, related to each one of the monoinfections, that results in a balanced inflammation associated with clinical immunity and absence of symptoms. In biological terms, the readouts are that the combined responses to each pathogen would induce a distinct profile of systemic immune activation, with the hallmarked activity of IL-10, a classical immunoregulatory cytokine, in confluence mainly with CCL2 and IL-4 activity. These multiple pathways would prevent the unbalanced proinflammatory activity associated with symptomatic and/or severe vivax malaria. Moreover, these findings highlight the importance of the immune system in driving disease presentation, raise discussion of immunotherapy in vivax malaria, and how these approaches have the potential to influence clinical outcomes.
| Malaria still rises major concerns in public health worldwide. The burden caused by the disease is noticeable, as it leads to more than 200 million cases and billions of dollars invested each year [1]. Despite all the investments and increased interest in the pursuit of new interventions [1,2], there was an elevation in the number of estimated cases in the successive years of 2016 and 2017 [1]. Moreover, P. vivax, which is the most widespread of the five main species of Plasmodium [1,3,4], have been increasingly associated with severe disease presentations and mortality [1,5–8].
Hepatitis B virus (HBV) infections are no less of a problem, with more than 250 million chronic cases estimated in 2015 [9]. Incidence of HBV has been reduced since the introduction of the vaccine, however approximately 815,000 deaths were accountable to HBV infections and its chronic complications in 2016 [10]. Both hepatitis B and vivax malaria are mainly outspread in tropical countries [1,9], and there is overlapping occurrence of these diseases [11].
HBV-associated tissue damage is described to be directly related to the host inflammatory response against infection [12,13]. The immune responses in chronic HBV infections are characterized by decreased T-cell proliferation potential and exhaustion [14–18]. Although not completely understood, these events are thought to be related to a higher release of HBsAg particles (hence, viral load) in the circulation, expression of co-inhibitory receptors, and production of IL-10 [14–21]. In vivax malaria, intensity of immune activation is associated with worse clinical outcomes [5,6,22–24]. On the converse, cases of asymptomatic P. vivax infection are hallmarked by a less pronounced pro-inflammatory response, with increased IL-10 levels in peripheral blood [6,11,22]. Thus, at first glance, both HBV and P. vivax infections seem to drive similar profiles of systemic inflammation in distinct clinical settings. Nevertheless, no previous study has performed a detailed characterization of systemic immune activation profile in HBV-malaria comorbidity.
Another similarity between HBV and malarial biology is the participation of the liver as a key organ part of the immunopathogenesis in both infections [4,7]. Notably, severe vivax malaria is associated with remarkable hepatic involvement [5–7,22,25], whereas tissue damage is determinant for the presentation of cirrhosis and hepatocellular carcinoma in chronic HBV infections [12,16,26]. Counterintuitively, HBV infection has been shown to lead to a distinct systemic inflammatory response in Plasmodium infections, resulting in increasing odds for asymptomatic malaria [11]. The present study expands the current knowledge as it examines in detail a rich interplay of cytokines, chemokines and acute phase proteins in a large number of patients infected with P. vivax, HBV or both. These analyzes demonstrate the nuances of different inflammatory responses in confluence, which culminates in an intense but balanced immune response to both pathogens, with key participation of relevant biomarkers as TNF-α, IL-4, IL-10 and CCL2.
Written informed consent was obtained from all participants or their legally responsible guardians, and all clinical investigations were conducted according to the principles expressed in the Declaration of Helsinki. The project was approved by the institutional review board of the Faculdade de Medicina, Faculdade São Lucas, Rondônia, Brazil, where the study was performed.
The present study is based on analyses performed retrospectively in databank containing immunological, clinical and epidemiological data from 601 subjects, including uninfected controls, recruited between 2006 and 2007 from the state of Rondônia, in the Brazilian Amazon. Multiple investigations have been reported from the project which this study is a part of [5,6,11,22–25,27–30]. Patient investigation included both active case detection in the municipalities of Buritis and Demarcação (Rondônia, Brazil) and passive case detection from individuals who sought care at Brazilian National Foundation of Health (FUNASA) diagnostic centers or at the municipal hospital in Buritis (Rondônia, Brazil).
Malaria diagnosis was conducted through microscopic examination of thick smears and nested polymerase chain reaction (PCR) evaluation in whole blood samples (20mL), with control for cross-contamination, performed at the Instituto Gonçalo Moniz (Fiocruz-BA), Salvador, Bahia, Brazil, as previously reported [5,6,22–24]. Individuals who tested positive through PCR evaluation and persisted without the presentation of fever (axillary temperature >37.8°C) and/or sweating, chills, jaundice, myalgia, arthralgia, asthenia, nausea, and emesis for 30 days were considered asymptomatic. Patients, which parasitological tests were positive, presenting any symptom listed above, were considered symptomatic. HBV diagnosis was conducted employing the AXSYM automatic ELISA system (Abbott, Wiesbaden, Germany), HBSAg, HBeAg, total anti-HBS, total anti-HBc, anti-HBc IgM and anti-HBe IgG were screened, according to the most updated protocols published by the Brazilian Ministry of Health at the time of study enrollment, and no acute HBV infection was detected (HBSAg+, anti-HBS-, anti-HBc IgM+). All the measurements were performed right at the study enrollment and diagnosis of malarial and/or HBV infection, meaning that the collections were performed before the initiation of antimalarial or HBV-specific therapy. Information regarding the number of previous malaria episodes and years that the patients resided in the area at the time of study enrollment were obtained directly from the patients in the interview part of the medical examination. For the present study patients with both symptomatic (n = 179) and asymptomatic (n = 145) P. vivax monoinfection, ongoing HBV infection (n = 29), concurrent P. vivax and HBV infections (n = 28) and healthy controls (n = 165, from which 152 had all the epidemiological data available) were included. The exclusion criteria for the present study were: patients with documented P. falciparum or HIV infections, tuberculosis, cancer, or use of immunosuppressant drugs. For the analyses of biochemical markers, patients presenting P. vivax monoinfection who were previously infected by HBV were excluded, in order to avoid interferences on the inflammatory profile. In addition, for part of these analyses, P. vivax monoinfected individuals, independently of clinical status (symptomatic or asymptomatic), were considered as a single group (n = 268), to compare and attest if the factors involved in the clinical presentation from coinfected subjects would also differ P. vivax-HBV coinfection from the P. vivax monoinfection overall. Clinical, demographic and epidemiological characteristics of the participants included in the current study are described in Tables 1 and 2 and S1.
Plasma levels of cytokines IL1-β, IL-4, IL-6, IL-10, IL-12p70, IFN-γ, tumor necrosis factor (TNF)-α, C-C motif chemokine ligand (CCL)2, CCL5, C-X-C motif chemokine ligand (CXCL)9, and CXCL10 were measured using the Cytometric Bead Array—CBA (BD Biosciences Pharmingen, San Diego, CA, USA), according to the manufacturer’s protocol. The measurements of aspartate amino-transferase (AST), alanine amino-transferase (ALT), total bilirubin, direct bilirubin, creatinine, fibrinogen and C-reactive protein (CRP) were performed at the Pharmacy School of the Federal University of Bahia and at the clinical laboratory of Faculdade São Lucas.
The median values with interquartile ranges (IQR) were used as measures of central tendency and dispersion. Chi-square test was used to compare frequencies between the study groups. Continuous variables were compared between the study groups using the Mann-Whitney U test (2-group comparisons), or the Kruskall-Wallis test with Dunn’s multiple comparisons ad hoc test (between 3 or more groups). Hierarchical cluster analyzes were performed using the Ward’s method with bootstrap (100X). Spearman tests were performed to analyze correlations and to build the correlation matrices, which assessed markers in each study group. Only correlations with Spearman rank (r) values above 0.6 were plotted in the matrices. A p-value below 0.05 after adjustment for multiple measurements (false discovery rate of 1%) was considered statistically significant. The statistical analyzes were performed using Graphpad Prism 7.0 (GraphPad Software Inc., San Diego, CA, USA), and JMP 12.0 (SAS, Cary, NC, USA).
The baseline characteristics of the study population are shown in Table 1. The study groups were similar with regard to sex. Among P. vivax-infected individuals, asymptomatic patients were older than symptomatic patients and those coinfected with HBV (median age: 43yrs, IQR: 34–52 vs. 29yrs, IQR: 19–42 vs. 31yrs, IQR: 23-46yrs, respectively) (Table 1). In addition, asymptomatic malaria patients were older than healthy endemic controls but with similar median age than those with HBV monoinfection (Table 1). Of note, referred number of previous malaria episodes was lower in patients presenting with symptomatic malaria compared to those with asymptomatic malaria, malaria-HBV coinfection and those with HBV monoinfection (Table 1).
Individuals with symptomatic P. vivax infection more frequently referred that they had lived for shorter time in the malaria endemic area when compared with the other clinical groups (Table 1). As expected according to previous reports [11], parasitemia, assessed in thick blood smears, was substantially lower in individuals with HBV-malaria comorbidity compared to those with symptomatic P. vivax monoinfection (median: 753 parasites/μL, IQR: 444.3–4,262 vs. 6,324, IQR: 913.5–60,623, P = 0.0004), whereas asymptomatic malaria patients predominantly did not exhibit detectable number of parasites in peripheral blood using microscopic examination (Table 1). In addition, frequency of P. vivax-HBV coinfection was significantly higher in asymptomatic individuals (n = 25). Overall, 18 (13.04%) asymptomatic and 38 (21.23%) symptomatic vivax malaria patients presented serological status compatible with previous history of HBV infections (HBSAg-, anti-HBS+, anti-HBc+). Furthermore, all cases of HBV infection, with or without malaria co-infection, presented serological status of chronic infection. Among the 179 symptomatic vivax malaria patients, eighteen presented severe/complicated vivax malaria, and six individuals eventually died from the disease. Detailed information on symptoms presented by individuals from each clinical group is available in S1 Table.
Median values of all biochemical markers per group were log-transformed and z-score normalized for hierarchical cluster analysis. Using this approach, three clusters of markers were identified (Fig 1A). Fold differences of the circulating levels of all biomarkers were then calculated to assess which parameters were differentially expressed in all four main subpopulations against healthy controls (Fig 1A). Asymptomatic vivax malaria patients presented a similar number of variables with significant concentrations increases (mainly IFN-γ, IL-10 and direct bilirubin) and decreases (such as CXCL10, IL-1β, IL-4 and indirect bilirubin). Patients with HBV monoinfection presented the same number of variables which concentrations were increased (mainly IFN-γ, TNF-α, IL-6 and CXCL9), when compared to asymptomatic vivax malaria monoinfection, but only significant decrease of one variable (also IL-4). Furthermore, symptomatic vivax patients presented augmented levels of almost every analyzed analyte, except for IL-10, CCL5, CXCL10 and specifically CCL2, which levels were diminished in comparison to uninfected controls (Fig 1A). P. vivax-HBV coinfected individuals presented multiple significant elevations in biomarker values when compared to healthy controls, although not as many as found in symptomatic malaria monoinfection. Noteworthy, coinfected patients exhibited an impressive 22-fold elevation in IL-10 levels when compared to healthy controls, and remarkable decreases in IL-8 concentrations (Fig 1A). Other significant differences are illustrated in Fig 1A. Fig 1B and 1C show Venn’s diagrams to further illustrate and depict these differences and similarities initially demonstrated by the subpopulations. These results delineate the systemic inflammatory profile associated with this comorbid condition. Detailed information of the laboratorial results and analysis in the subpopulations are available in Table 2.
Fold differences of the circulating levels of all biomarkers were also calculated to assess which parameters were differentially expressed in coinfected individuals in comparison to other main study groups (coinfected vs. asymptomatic or symptomatic P. vivax monoinfected patients, and HBV monoinfected individuals). Patients with malaria-HBV coinfection presented elevated concentrations of multiple variables such as IFN-γ, TNF-α, IL-4, IL-10, CCL2, and reduced levels of IL-12 and creatinine levels when compared to those with asymptomatic vivax malaria monoinfection (Fig 2A). When compared to symptomatic vivax malaria patients (Fig 2A,), coinfected individuals presented elevated levels of IFN-γ, IL-10 and CCL2, and diminished plasma concentrations of multiple variables as TNF-α, IL-6, IL-12, and CRP. When compared to those with HBV monoinfection, coinfected patients presented elevated levels of IL-4, IL-10, CCL2, CRP, fibrinogen, and direct bilirubin, and reduced concentrations of TNF-α and IL-8 (Fig 2A). Other significant differences are illustrated in Fig 2A.
Thus, in summary, IL-10 and CCL2 were the only variables which coinfected patients presented with elevated concentrations in comparison to all the other three main study groups. Then, considering the immunoregulatory nature of IL-10 and the dimension of its elevations in coinfected individuals, the next step was to analyze the behavior of the biomarkers in comparison to IL-10 levels in all main study groups. Coinfected individuals presented reduced IL-10 ratios for all variables (S1 Fig), with the exception of IFN-γ, IL-4 and CXCL10 (Fig 2B). The IFN-γ/IL-10 and CXCL10/IL-10 ratios could not distinguish coinfected and asymptomatic vivax patients. In addition, HBV-monoinfected and P.vivax-HBV coinfected patients could not be distinguished by their IL-4/IL-10 ratio values. These results highlight similar biosignatures that may be reminiscent from each respective infection in P.vivax-HBV coinfected individuals.
Multiple correlation matrices were inputted into a network analysis to assess the profile of associations between cytokine levels in each study subpopulation (Fig 3A). It was noticeable the decreased number of significant connections (which represent statistically significant correlations) in the network of asymptomatic vivax patients (Fig 3A) when compared to the networks calculated from the other groups. This tendency is also maintained when such networks were compared to that from uninfected controls (S2 Fig). In addition, the correlation matrix of individuals with symptomatic P. vivax monoinfection showed an increase of significant positive connections between the biochemical parameters (Fig 3A). Interestingly, while displaying an increased number of significant positive correlations between variables, P. vivax-HBV coinfected individuals (Fig 3A) also exhibited the tendency of negative connections presented by HBV-monoinfected patients (Fig 3A). HBV patients presented negative correlations between IL-4 and IFN-γ, IL-1β, and IL-12p70 concentrations, between CCL2 and IFN-γ or IL-1β levels, and between total bilirubin and creatinine levels. In coinfected individuals, IL-4 and IL-12p70 levels were again negatively correlated, but also between IL-8 and AST, fibrinogen, direct bilirubin, and total bilirubin concentrations. In addition, IFN-γ was only positively correlated with CXCL9 and CXCL10, whereas IL-1β levels were negatively correlated with concentrations of CCL5, ALT and total bilirubin (Fig 3A).
The further step was to analyze the correlations between the biomarkers and viral load, which has been previously associated with a downregulated proinflammatory response in individuals chronically infected with HBV [14,15]. Overall, concentrations of ten biomarkers were significantly correlated with viremia levels in patients with HBV monoinfection or in those with HBV-malarial coinfection. Furthermore, TNF-α did not presented the same correlation pattern in the two groups, which was the case of IFN-γ concentrations, for example (Fig 3B). In HBV-monoinfected individuals, TNF-α levels was not correlated with the viral load, while presenting negative correlation with viremia in coinfected patients. The other significant correlations to viral load are shown in S2 Fig.
We next performed additional analyses in which symptomatic and asymptomatic P. vivax monoinfected subjects were considered as a single group (S3 Fig). When compared with uninfected controls, monoinfected P. vivax individuals presented the characteristic significant reduction of CXCL10 and CCL2 levels and increases of IL-10 and TNF-α levels (S3A Fig) found in previously separated groups (Fig 1A). In addition, when compared against coinfected patients, monoinfected P. vivax individuals presented significant elevations of AST, ALT, CRP, IL-8 and IL-12 (S3A Fig), and significant reductions in IFN-γ, CXCL10, CCL2, IL-4 and IL-10 (S3A Fig). The IFN-γ/IL-10 and CXCL10/IL-10 ratios could distinguish coinfected patients from both groups of P. vivax and HBV-monoinfected patients (S3B Fig). S3C Fig shows the distribution of the patients based on their parasitemia values. When compared with those presenting HBV coinfection, P. vivax-monoinfected patients presented with significantly reduced parasite counts (S3D Fig).
In the present study, we performed novel analyses of multiple inflammatory biomarkers related to key immune and inflammatory responses associated with disease progression in the context of HBV and Plasmodium vivax infections. These expanded analyses provide deeper comprehension of the immune response against P. vivax-HBV coinfection, which culminates with reduced odds of severe disease and progression of vivax malaria [11].
In the study population, asymptomatic vivax and coinfected patients presented distinct epidemiological profiles. Elevated number of previous malaria episodes and more advanced age are well-known to be associated with milder and asymptomatic vivax malaria [6,11,31]. However, while referring a similarly increased number of previous malaria episodes, coinfected patients were predominantly younger than asymptomatic vivax patients (Table 1). Furthermore, coinfected individuals presented similar median age to symptomatic P. vivax malaria patients (Table 1). Moreover, asymptomatic monoinfected vivax patients predominantly presented with undetectable parasitemia examined by thick smears, whereas coinfected individuals, predominantly asymptomatic from a malarial perspective, presented an elevated values of parasite counts (S3C Fig). In fact, coinfected individuals presented significantly increased parasite counts when compared to P. vivax monoinfection (S3D Fig). Hence, these distinct epidemiological and serological characteristics of asymptomatic vivax and coinfected patients argues that other factors may have influenced malarial presentations in coinfection with HBV. Moreover, the parasitemia results are in line with previous hypothesis that mainly the host response to infection, and not the parasite load alone, are responsible for clinical presentations in vivax malaria [5,6,22–24]. Multiple cytokines, chemokines and acute phase proteins were then profiled to further analyze the mechanisms associated with disease presentation in P. vivax-HBV coinfected patients. Coinfection was hallmarked by extremely elevated concentrations of IL-10, as well as heightened levels of CCL2, in comparison to the distinct clinical presentations of P. vivax infections or HBV monoinfection.
IL-10 is an immunoregulatory cytokine and its levels were previously reported to be closely associated with disease progression and outcomes in both hepatitis B and vivax malaria. Patients with severe vivax malaria have been shown to present unbalanced concentrations of IL-10 against levels of proinflammatory biomarkers, when compared to individuals with uncomplicated vivax malaria [5,6,22]. In addition, these individuals with uncomplicated or asymptomatic P. vivax infections have been shown to express relatively augmented concentrations of IL-10 when compared to those with symptomatic or severe vivax malaria [5,6,31,32]. In viral infections, IL-10 levels are associated with diminished T-cell activation, which may already start to occur rapidly after infection [20]. Furthermore, HBV actively suppresses immune responses [14] and augmented IL-10 levels are closely associated with viral persistence [16–18,33]. Herein, as expected, HBV-infected patients presented increased IL-10 levels when compared to uninfected controls. Furthermore, only symptomatic vivax malaria patients could not be distinguished from uninfected controls by their IL-10 levels. Although coinfected patients presented almost an 8-fold increase in IL-10 concentrations when compared to asymptomatic vivax-monoinfected individuals, they presented undistinguishable values of IFN-γ/IL-10 ratios, which highlights a similar tendency of immune balance in this aspect of inflammatory response. A higher baseline concentration of IL-10, associated with the condition of antiviral response, and hence a distinct overall inflammatory profile, could then be responsible for the difference in absolute IL-10 levels identified between coinfected and asymptomatic vivax patients. Both study groups also presented similar CXCL10/IL-10 ratio values. CXCL10 is an IFN-γ induced protein which acts in chemotaxis, apoptosis, cell proliferation and angiogenesis [34]. Thus, similar results of CXCL10/IL-10 and IFN-γ/IL-10 are not surprising. Even with these solid findings, experimental models are still necessary to further define and ratify whether these similarities were directly carried from responses associated to HBV persistence and T-cell exhaustion, or from the antimalarial response, or from concurrent responses to both pathogens.
CCL2, the other biomarker which concentrations are augmented in coinfected individuals in comparison to all other study groups, is an important chemokine involved in recruitment of monocytes [35,36] and NK cells [36]. CCL2 is reported to be produced by hepatocytes under acute HBV stress [37] and was previously associated with uncomplicated P. vivax infections [31]. However, without the weight of an acute and heavy antiviral response, an infection with a pathogen known to induce acute hepatocyte damage as P. vivax [6] could possibly trigger this local chemokine production. ALT and AST levels could reinforce this hypothesis as they were found to be augmented in symptomatic vivax malaria patients when compared to HBV-monoinfected individuals (Table 2), while only being correlated to viremia in coinfected P. vivax-HBV patients and not in HBV-monoinfected subjects (S2 Fig). Herein, coinfected individuals, which presented predominantly asymptomatic malarial infection, had a 2.2-fold increase in CCL2 concentrations when compared to patients with symptomatic P. vivax infections. Therefore, these results may reinforce the association of CCL2 with uncomplicated malaria.
It is also reported that CCL2 influences and directs CD4+ T lymphocytes to a more biased response towards IL-4 production [38]. Herein, coinfected individuals presented significant elevations IL-4 levels when compared to healthy controls (Fig 1A), asymptomatic vivax or HBV-monoinfected subjects (Fig 2A). Although IL-4 levels could not distinguish coinfected and symptomatic vivax malaria patients, correlations with IL-4 concentrations were completely different in both study groups. IL-4 concentrations were negatively correlated to IL-12p70 levels in coinfected patients, while being positively correlated with multiple other proinflammatory cytokines in symptomatic vivax patients (Fig 3A). Hence, these antagonic tendencies suggests that different mechanisms, and not just the antimalarial response in this case, could be responsible for the elevation of IL-4 concentrations in coinfected and symptomatic vivax malaria patients. In addition, IL-4/IL-10 ratio values could not distinguish coinfected and HBV-monoinfected individuals (Fig 2B). This similar profile displayed in both groups of patients infected by HBV suggests that antiviral or responses associated with hepatocyte stress, possibly under CCL2 influence in this hypothesis, could be responsible for these elevations of IL-4 levels in coinfected individuals. In practical terms, this immune response of coinfected individuals with augmented IL-4 concentrations happens without much proinflammatory pressure, as IL-10 immunoregulatory mechanisms should be expected to bring them a more balanced inflammatory response, oppositely to what occurs in symptomatic vivax individuals (Fig 1A). This augmented production of IL-10 alongside IL-4 heightened levels can directly downregulate key proinflammatory cytokines such as TNF-α [39,40], and thus have a protective effect against severe malaria presentations. Herein, TNF-α concentrations were found to be significantly reduced in coinfected individuals, when compared to both HBV or symptomatic vivax malaria patients (Fig 2A). Furthermore, TNF-α concentrations were negatively correlated to viremia only in patients with P. vivax-HBV coinfections (Fig 3B), which could be read as a possible effect of these previously reported mechanisms in the patients from the present study. These results are compatible with the hypothesis that coinfection drives reduction of systemic inflammation, which we previously published [11]. Therefore, these confluent events from responses to both pathogens (increased production of IL-10, CCL2 protective role in malaria, as well as combined effects of IL-4 and IL-10) could enable the host to respond properly without unbalanced inflammation. Thus, this proper response creates an environment unfavorable for the Plasmodium to thrive and induce detectable symptoms.
Our study presented some limitations. We did not have data available from follow-up of the HBV-infected patients and their antiviral treatments, as they were referred to a specialized service. Although we collected information regarding number of previous malaria episodes, these data were expressed by the patients, and not extracted from official documents of the health centers. Thus, this fact further limits the analysis and evaluation of relapses in the study patients, and if these events could be related to an association between the chronic HBV infection and hypnozoite activation. Experimental models and biopsies would have helped with the assessment of T-cell exhaustion, the impact of liver involvement into inflammatory responses, and cytokine evaluation at tissue level. Therefore, further longitudinal and experimental studies are still necessary to completely understand the events associated with P. vivax-HBV coinfection. However, despite some limitations, the present study was successful to analyze several biomarkers and their associated biomechanisms, and link them to the known protective effect of chronic HBV infections in vivax malaria.
In conclusion, the results presented here represent a translation of an increased demand and pressure caused by the acute P. vivax infection on the immune system of a chronically HBV-infected host. Hence, there is an augmented presence of inflammatory biomarkers as IFN-γ and CRP, counterbalanced with the immunoregulatory mechanisms discussed here. In summary, coinfection was hallmarked by substantially increased levels of IL-10 and augmented concentrations of CCL2. CCL2 is expressed by hepatocytes during acute injury, reportedly leads to IL-4 increases, while IL-10 is directly related to viral persistence and T-cell exhaustion, and both cytokines are associated with protection in P. vivax infections. Thus, these results argue that distinct mechanisms associated with antiviral and antimalarial activity are due to changes in cytokine balance, and lead to the known increased odds of asymptomatic vivax malaria in coinfected HBV-P. vivax patients. This knowledge of responses to both pathogens counteracting proinflammatory responses helps to depict the pathophysiology associated with the coinfection, and could prove relevant to future studies and approaches with immunotherapy in cases of severe malaria or HBV infection.
|
10.1371/journal.pgen.1005604 | A Follicle Rupture Assay Reveals an Essential Role for Follicular Adrenergic Signaling in Drosophila Ovulation | Ovulation is essential for the propagation of the species and involves a proteolytic degradation of the follicle wall for the release of the fertilizable oocyte. However, the precise mechanisms for regulating these proteolytic events are largely unknown. Work from our lab and others have shown that there are several parallels between Drosophila and mammalian ovulation at both the cellular and molecular levels. During ovulation in Drosophila, posterior follicle cells surrounding a mature oocyte are selectively degraded and the residual follicle cells remain in the ovary to form a corpus luteum after follicle rupture. Like in mammals, this rupturing process also depends on matrix metalloproteinase 2 (Mmp2) activity localized at the posterior end of mature follicles, where oocytes exit. In the present study, we show that Mmp2 activity is regulated by the octopaminergic signaling in mature follicle cells. Exogenous octopamine (OA; equivalent to norepinephrine, NE) is sufficient to induce follicle rupture when isolated mature follicles are cultured ex vivo, in the absence of the oviduct or ovarian muscle sheath. Knocking down the alpha-like adrenergic receptor Oamb (Octoampine receptor in mushroom bodies) in mature follicle cells prevents OA-induced follicle rupture ex vivo and ovulation in vivo. We also show that follicular OA-Oamb signaling induces Mmp2 enzymatic activation but not Mmp2 protein expression, likely via intracellular Ca2+ as the second messenger. Our work develops a novel ex vivo follicle rupture assay and demonstrates the role for follicular adrenergic signaling in Mmp2 activation and ovulation in Drosophila, which is likely conserved in other species.
| Ovulation is the process of releasing fertilizable oocytes from the ovary and is essential for metazoan reproduction. Our recent work has demonstrated principles governing ovulation process that are highly conserved across species, such that both mammals and Drosophila utilize matrix metalloproteinase (Mmp) to degrade extracellular matrix and weaken the follicle wall for follicle rupture. However, a fundamental question remaining in the field is how Mmp activity is precisely regulated during ovulation. This paper reports that Drosophila octopamine (OA), the insect equivalent of norepinephrine (NE), is the signal to induce Mmp activity through activating its receptor Oamb on mature follicle cells and that this may induce ovulation. These findings allow us to develop the first ex vivo follicle rupture assay for Drosophila, which gives us unprecedented ability to characterize the entire follicle rupturing process ex vivo and to identify essential factors for ovulation. Furthermore, we show that NE partially fulfills OA’s role in inducing follicle rupture ex vivo, indicating that follicular adrenergic signal is a conserved signal to regulating Mmp activity and ovulation. Our work not only sheds light on the long-standing question of Mmp regulation, but also may lead to a better understanding of Mmp and NE linked pathological processes including cancer metastasis and polycystic ovary syndrome.
| Ovaries in organisms ranging from humans to insects are extensively innervated [1–4], and neuronal inputs likely play important roles in ovarian physiology [5]. In mammals, ovaries are predominantly innervated by sympathetic fibers from the ovarian plexus nerve and the superior ovarian nerve [6], which release norepinephrine (NE) locally and contribute to follicle development [7]. Deregulation of sympathetic nerve outflow to ovaries is associated with polycystic ovary syndrome (PCOS), a common endocrine disorder leading to anovulatory infertility [8,9]. Despite the apparent importance of sympathetic innervation, however, it is not yet clear how the neuronal modulators/transmitters released from nerve termini affect ovulation [10–16].
In Drosophila and other insects, the biogenic monoamines tyramine (TA) and octopamine (OA) act as functional counterparts to mammalian epinephrine and norepinephrine and regulate a variety of behaviors, including the fight-or-flight response, motivation, aggression, and reproduction [17,18]. Analogous to the adrenergic innervation in mammalian ovaries, Drosophila octopaminergic neurons innervate ovaries and the female reproductive tract (Fig 1A; [3,19,4]). OA released from these neurons is essential for ovulation, as mutations that disrupt the enzymes required for OA synthesis, tyrosine decarboxylase 2 (Tdc2) and tyramine β-hydroxylase (TβH), completely block ovulation [20–22].
Four OA receptors have been identified in Drosophila: Oamb, Octβ1R, Octβ2R, and Octβ3R. Oamb is most closely related to mammalian α-adrenergic receptors, and the other three to β-adrenergic receptors [17,23]. Recent work demonstrated that Oamb and Octβ2R are important in egg laying and ovulation [24–26]. Oamb is widely expressed in the female reproductive system, including the ovary, with strongest expression observed in the oviduct [24]. It is currently believed that OA activates receptors in the oviduct, inducing oviduct contraction and secretion, which ultimately regulates ovulation through an unknown mechanism [19,27,25]. In addition to OA signaling, ovulation in Drosophila is affected by female reproductive gland secretions [28] and by mating, which increases the ovulation rate by stimulating afferent nerve activity in the female reproductive tract [29–33,4]. In particular, Ovulin transferred into the female reproductive tract after mating was recently shown to increase octopaminergic signaling and relax oviduct muscle [34], consistent with the role of OA signaling in regulating muscle contraction. It is, however, not clear whether OA plays any direct roles in the ovary to control ovulation.
In addition to above important work on Drosophila ovulation (also see review [35]), recent studies from our lab also showed significant conservation of the basic cellular and molecular mechanisms of ovulation from flies to mammals. Drosophila female contains two ovaries that are connected by the oviduct. Each ovary is organized into ovarioles, which have mature follicles (stage-14 egg chambers) at the posterior end toward the oviduct (Fig 1A; [36]). Each mature follicle has one layer of epithelial follicle cells surrounding the oocyte. During ovulation, posterior follicle cells are first trimmed to break the follicle-cell layer and to allow the oocyte to be released into the oviduct. The rest of the follicle cells remain at the end of the ovariole and form a corpus luteum [37]. Similar to vertebrate ovulation [38–40], the entire follicle rupture requires matrix metalloproteinase 2 (Mmp2), a proteolytic enzyme expressed in posterior follicle cells of mature egg chambers but only activated during follicle rupture [37]. It is not yet clear what signals control Mmp2 activity, but it is clear that studying this question in Drosophila could yield important insights into the fundamental mechanism of ovulation.
Here, we developed the first ex vivo assay for follicle rupture in Drosophila and used it to investigate the role of octopaminergic signaling in this process. We found that OA directly activates its receptor Oamb on mature follicle cells to induce the breakdown of posterior follicle wall and ovulation. In addition, NE could partially substitute for OA, indicating an evolutionary conserved role for follicular adrenergic signaling in ovulation. Finally, we demonstrated that follicular adrenergic signaling activates Mmp2 activity to control ovulation via the intracellular Ca2+ as the second messenger. This is the first demonstration of a direct role of a neuromodulator in the control of follicle rupture during ovulation.
Octopaminergic neurons innervate ovarioles extensively [21], and OA receptor Oamb is transcribed in mature follicle cells according to in situ hybridization [24], microarray analysis (S1 Fig; [41]), and the expression of R47A04-Gal4 [42], an Oamb enhancer element-regulated Gal4 driver, in mature follicle cells [37]. We examined whether OA activates Oamb directly in mature follicle cells to induce follicle rupture. Mature follicles with an intact layer of follicle cells marked by R47A04-Gal4 were isolated from ovaries (see methods) and cultured with OA or control media (Fig 1A). After three hours, follicles in control medium maintained an intact follicle-cell layer (Fig 1B). In contrast, about 80% of the follicles cultured with 5 μM of OA had shed their follicle-cell layer to the dorsal appendage at the anterior tip of the oocytes (Fig 1C); some were completely detached from the oocyte and floating in the medium. This phenomenon of shedding the follicle-cell layer, which we called follicle rupture in our ex vivo culture system, is reminiscent of what occurs during the ovulation process in vivo [37]. The percentage of ruptured follicles with OA stimulation increased dramatically in the first two hours and reached a plateau at about three hours (Fig 1D). Extending the culture period neither increased the percent of ruptured follicles to 100% in the OA medium, nor allowed follicles in the control medium to reach the same level of rupture as OA-stimulated follicles (Fig 1D).
To validate that the follicle rupture in our ex vivo assay mimics the in vivo process, we video-recorded the entire rupturing process (Fig 1E and S1 Movie). We found that posterior follicle cells were first trimmed, as we previously observed in vivo [37]. The remaining follicle-cell layer was then squeezed toward the anterior dorsal appendage (Fig 1E and S1 Movie). The entire rupturing process took 13.1 ± 5.0 minutes (S1 Table), resembling the estimated in vivo ovulation time of 11.2 ± 2.5 minutes (Table 1; [37]). Each mature follicle initiated the follicle rupture asynchronously, likely reflecting their asynchronous developmental stages; however, the kinetics of all ruptures was similar, with a very slow initial speed (Fig 1F). It took about 10 minutes to rupture through the posterior half of the oocyte, but only four minutes for the rest of the area (Fig 1E and 1F).
All data are mean ± 95% confidence interval. Student's T-test was used for egg laying, Chi-square test was used for egg distribution, and Z Score test was used for egg laying time assuming normal distribution
To further examine the quality of ex vivo ruptured oocytes, we determined whether these oocytes were activated. Mature oocytes released into the oviduct are activated and resistant to bleach treatment because their egg shells are hardened through cross-linking [43]. This activation process can also be mimicked in vitro by culturing oocytes in hypotonic activation buffer [44,45]. Using the established bleach assay (see methods), we found that oocytes from our ex vivo assay dissolved immediately after bleach treatment (n = 96), indicating that they were not fully activated and their eggshells were not hardened. However, treatment with hypotonic activation buffer for 15 minutes can efficiently activate these ruptured oocytes (95%, n = 150; S2A and S2B Fig), indicating these oocytes from our ex vivo system are of good quality and responsive to egg activation stimuli.
OA-induced follicle rupture is dose-dependent. Stimulation with 1 μM of OA had a minimal effect on follicle rupture, while stimulation with 20 μM of OA reached the maximal effect (Fig 1G), which led us to use 20 μM for all the following experiments. In contrast, stimulation with 20 μM of tyramine (TA), the immediate precursor of OA, had a much weaker effect on follicle rupture (Fig 1H), consistent with a previous report that OA, but not TA, is responsible for inducing ovulation [20]. Since NE is the counterpart of OA in mammals, we tested whether NE can also induce follicle rupture in our ex vivo assay. NE had only a minimal effect at lower doses (Fig 1I). Higher doses of NE could induce follicle rupture (Fig 1I), likely reflecting a differential binding properties of OA and NE to their respective receptors [18]. Nevertheless, these data suggest that OA and NE play a conserved role in regulating follicle rupture. In summary, we developed the first ex vivo assay to study follicle rupture in Drosophila, and our data suggest that OA is sufficient to induce follicle rupture in the absence of the oviduct and muscle function, as these tissues were excluded from our culture assay (68 mature follicles examined and none had ovariole muscle; S3A and S3B Fig).
To identify the receptor responsible for OA/NE-induced follicle rupture, we focused on Oamb, which is essential for ovulation [24] and is the most highly expressed OA receptor in mature follicles (S1 Fig). We verified the requirement of Oamb in ovulation with a new mutant allele (OambMI12417), in which a MiMIC vector with a splice acceptor [46] was inserted in the coding intron of Oamb gene to disrupt the correct mRNA splicing (S4 Fig). Females bearing this mutant allele laid significantly fewer eggs and took a much longer time to ovulate an egg (Table 1). We then isolated mature follicles from these females and applied OA stimulation ex vivo. Oamb mutant follicles showed severe defects in OA-induced follicle rupture compared to control follicles (Fig 2A, 2B and 2E). In addition, the Oamb mutation abolished the NE-induced follicle rupture (Fig 2C–2E). The defective response of Oamb mutant follicles to OA/NE stimulation is not likely due to defective OA signaling in the oviduct or other organs, because follicles from TβH or Tdc2 mutant females are fully competent to OA/NE-induced follicle rupture (Fig 2F and 2G). These data indicate that Oamb in mature follicles is likely responsible for OA/NE-induced follicle rupture.
To test if Oamb functions directly in mature follicle cells, we knocked down Oamb specifically in these cells with RNA interference (RNAi) and then performed OA stimulation ex vivo. Oamb knockdown in mature follicle cells with R47A04-Gal4 severely disrupted OA-induced follicle rupture (Fig 2H–2J). Since R47A04-Gal4 is regulated by an Oamb enhancer element [42], it could potentially be expressed in other Oamb-expressing cells, which may facilitate follicle maturation and ovulation. To exclude this possibility, we identified another Gal4 driver (R44E10-Gal4) expressed in mature follicle cells (S5B–S5D Fig). Compared to R47A04-Gal4, which is only expressed in late stage-14 follicles (S5A Fig), R44E10-Gal4 was expressed in all stage-14 follicles, slightly earlier than R47A04-Gal4. R44E10-Gal4 was not expressed in any tissues in the lower reproductive tract, nor in the neurons innervating the reproductive tract (S5B, S5E and S5F Fig). Like mature follicles isolated using R47A04-Gal4, follicles isolated using R44E10-Gal4 were also responsive to OA/NE-induced follicle rupture (S5G and S5H Fig). In addition, mature follicles with R44E10-Gal4 driving OambRNAi showed similar unresponsiveness to OA or NE stimulation (Fig 2K–2M). Taken together, these data suggest that follicular Oamb is required for OA/NE-induced follicle rupture ex vivo.
To determine whether follicular adrenergic signaling is required for ovulation in vivo, we first analyzed the fecundity of females lacking follicular Oamb. Follicular Oamb-knockdown females with either R47A04-Gal4 or R44E10-Gal4 drivers laid significantly fewer eggs than control flies (Fig 3A and Table 1). The egg-laying defect is not caused by oogenesis problems, as mature follicles are abundant in these ovaries. In fact, Oamb-knockdown flies generally had more mature follicles in their ovaries (Fig 3B), indicating an ovulation defect. Indeed, Oamb-knockdown flies had a much longer ovulation time compared to control flies but did not show defects in transporting ovulated eggs into the uterus or ejecting them out of the uterus (Fig 3C and Table 1). These data strongly suggest that follicular Oamb is required for ovulation in vivo.
All data are mean ± SD. Student’s T-test was used.
Trimming of posterior follicle cells is essential for ovulation and precedes follicle rupture [37]. We investigated the role of follicular adrenergic signaling in this trimming process. Posterior trimmed follicles were readily observed in the ovaries of control females six hours after mating, and they account for 9% of the total mature follicles in each female (Fig 3D and 3F and Table 2), consistent with our previous analysis [37]. In contrast, the percentage of posterior trimmed follicles was reduced three fold in females lacking follicular Oamb (Fig 3E and 3F and Table 2), indicating its essential role in follicle trimming. This is consistent with our observation that posterior follicle cells remain intact in Oamb-knockdown follicles even after three hours of OA stimulation ex vivo (Fig 2I and 2L). Furthermore, the percentage of trimmed follicles also decreased in flies that lacked the ability to produce OA; we saw a reduction to 2.4% and 0.4% in TβH and Tdc2 mutant females, respectively (Fig 3G–3L and Table 2). This reduction of trimmed follicles was not only observed in mated females, but also in virgin females (Table 2). Taken together, these data suggest that follicular adrenergic signaling is required for posterior follicle cell trimming.
The crucial role of Mmp2 in trimming of posterior follicle cells [37] prompted us to investigate the relationship between follicular adrenergic signaling and Mmp2 activity. It is unlikely that adrenergic signaling regulates Mmp2 expression, as Mmp2 was readily detected in the posterior follicle cells of TβH mutants (S6A and S6B Fig). To test whether OA regulates Mmp2 activity, we examined gelatinase enzymatic activity in the OA-induced ex vivo ovulation assay using in situ zymography [37,38]. About 20% of mature follicles cultured in a control medium had gelatinase activity at their posterior end (Figs 4A, 4C, S6C and S6G). In contrast, more than 70% of mature follicles stimulated with OA had gelatinase activity (Figs 4B, 4C, S6D and S6G). The entire eggshells of ruptured oocytes were coated with Mmp-activated gelatin-fluorescein (Figs 4B and S6D), as we observed in vivo [37]. In addition, OA-induced gelatinase activity was blocked in mature follicles with Oamb knockdown or misexpression of Timp, an endogenous inhibitor of Mmp2 [47], in follicle cells (S6E–S6G Fig). These data indicate that OA-Oamb signaling is sufficient to induce Mmp2 activation.
To determine whether Mmp2 activity is required for OA-induced follicle rupture, we isolated mature follicles containing follicle cell-specific Mmp2 knockdown and cultured them in the OA medium. These follicles did not respond to OA stimulation, and their posterior follicle cells remained intact (Fig 4D–4I). In addition, Mmp2 knockdown in follicle cells also abolished the NE-induced follicle rupture (Fig 4F and 4I). Furthermore, misexpression of Timp in mature follicle cells completely prevented follicle rupture ex vivo (Fig 4F and 4I). Therefore, Mmp2 activity in mature follicle cells is essential for OA/NE-induced follicle rupture ex vivo, consistent with its essential role in follicle trimming and ovulation in vivo [37].
To confirm that Mmp2 acts downstream of adrenergic signaling in follicle trimming and rupture, we attempted to rescue the defect of follicle rupture in Oamb mutant flies with ectopic expression of Mmp2 in mature follicle cells. Oamb mutant females had two intact ovaries, which contain a large number of mature follicles (Fig 4J). In contrast, follicular misexpression of Mmp2 in Oamb mutant females caused the breakdown of the ovariole muscle sheath and the release of mature follicles into the abdominal cavity (Fig 4K). Further examination of these released follicles demonstrated that 99% of them (n = 70) had no follicle-cell covering, similar to follicles released upon misexpression of Mmp2 in Oamb heterozygous or wild-type females (Fig 4L; [37]). Therefore, Mmp2 is sufficient to induce follicle rupture in the absence of adrenergic signaling. Together, our data indicate that follicular adrenergic signaling activates Mmp2 to control follicle trimming and ovulation.
OA-Oamb interaction can induce transient increase of intracellular Ca2+ concentration ([Ca2+]i) [23]. To determine whether OA evokes Ca2+ signaling in mature follicle cells to induce follicle rupture, we first monitored the [Ca2+]i using a genetically encoded calcium sensor (see methods). Fluorescent intensity of the calcium sensor expressed in mature follicle cells rose significantly around six minutes after OA administration in our ex vivo culture system (S7 Fig and S2 Movie). To determine whether Ca2+ is required for OA-induced follicle rupture, we pretreated mature follicles with BAPTA-AM, an intracellular Ca2+ chelator, before OA stimulation. Two hundred μM BAPTA-AM treatment significantly perturbed the OA-induced follicle rupture (Fig 5A–5C). To determine whether Ca2+ is sufficient to induce follicle rupture, we stimulated mature follicles with ionomycin, a potent ionophore for increasing [Ca2+]i. Ionomycin is potent to induce follicle rupture even at 5 μM concentration (Fig 5D–5F), lower than the dose typically used in the field [48]. Taken together these data suggest that the increase of [Ca2+]i is both necessary and sufficient to induce follicle rupture.
To further test whether Ca2+ is the second messenger of follicular adrenergic signaling for Mmp2 activation and follicle rupture, we set to examine whether ionomycin is sufficient to induce rupture of follicles lacking follicular Mmp2 or Oamb, which do not respond to OA stimulation. Ionomycin only partially induces follicle rupture when Mmp2 is knocked down in mature follicle cells and is not able to induce any rupture when Timp is overexpressed (Fig 5G–5I). In contrast, ionomycin is able to induce follicle rupture in both control and Oamb mutant follicles at the equal efficiency (Fig 5J–5L). All these data indicate that Ca2+ acts downstream of Oamb but upstream of Mmp2 during follicle rupture. Together, we conclude that follicular adrenergic signaling activates Mmp2 to control follicle trimming and ovulation likely via intracellular Ca2+ (Fig 5M).
Ovulation, an essential step in metazoan reproduction, has been extensively studied in mammals over the past several decades [49–51]. However, progress in the field has been hindered by the limited ability of mammalian model systems to be genetically manipulated. Thus it is still unclear how follicles break their wall in a highly regulated spatio-temporal manner to allow release of oocytes. The model organism Drosophila offers a wealth of tools for genetic manipulation, but to date, few specific readouts for Drosophila ovulation has been developed. Previous studies of Drosophila ovulation have used readouts such as egg laying, percentage of females with eggs in the reproductive tract, or egg retention [21,25–27,33]. We recently combined these parameters to estimate ovulation time [28,37]. In the present study, we developed the first ex vivo follicle rupture assay in Drosophila and demonstrated that OA-induced follicle rupture in this assay is similar to the rupturing process in vivo. This assay gave us the unprecedented ability to visualize the entire process of follicle rupture and quantify its kinetics. Further genetic evidence illustrated that genes required for ex vivo follicle rupture are also required for in vivo ovulation, including Oamb and Mmp2. Our ex vivo assay represents a simple, specific, and reliable method for measuring rupturing ability of mature follicles. In conjunction with the powerful genetic tools available in Drosophila, this ex vivo assay will allow genetic screens to identify candidate genes involved in follicle rupture, thus opening new avenues for ovulation research.
Octopamine, a biogenic amine derived from tyrosine, has been identified as essential for ovulation in Drosophila [20]. The major source of OA is octopaminergic neurons innervating the female reproductive system, and previous studies showed that restoring TβH specifically in these neurons rescues the ovulation defect caused by TβH mutation [21]. Due to its effects on muscle contraction, OA was proposed to regulate ovulation by inducing the contraction of ovarian muscle and relaxation of oviduct muscle [3,19,25,27,34].
Ovarian smooth muscle contraction was also proposed to regulate ovulation in mammals in the early 1980’s [11,52,53]. However, subsequent work suggest that ovulation requires the active proteolytic degradation of the follicle wall rather than passive muscle contraction [54,40,55]. At least three families of proteolytic enzymes are involved in this process, including matrix metalloproteinases [56,57]. Pharmacological blockage of any of these enzymes results in inhibition of follicle rupture.
Our recent work suggested that Drosophila also requires proteolysis for breaking the follicle wall and ovulation [37], and in this way shares similarities with mammalian ovulation at both the cellular and molecular level [28,37]. These new insights into Drosophila ovulation process lead to the speculation that octopaminergic signaling may play a direct role on the follicle in controlling ovulation in addition to its role on muscle contraction. Here, we demonstrate that OA-Oamb signaling in mature follicle cells directly regulates follicle wall degradation, follicle rupture, and ovulation by activating key enzyme Mmp2. Furthermore, our pharmacological data suggest that OA-Oamb signaling likely fulfill these functions via intracellular Ca2+ as the second messenger. However, it is still unclear how OA-Oamb-Ca2+ regulates Mmp2 activity. Lacking a method to detect Mmp2 protein prevents us to test whether OA-Oamb-Ca2+ regulates Mmp2 protein secretion. The Mmp2::GFP fusion allele we previously generated [37] is good to detect Mmp2::GFP expression but not good to track Mmp2 secretion because Mmp2::GFP fusion proteins are not properly processed and secreted to the extracellular space (S8 Fig) and Mmp2::GFP homozygous flies are lethal as Mmp2 mutant females do [47]. Alternatively, Ca2+ signaling may indirectly regulate Mmp2 activity via its inhibitor or other regulatory processes. Despite that, it is intriguing that [Ca2+]i also rises after NE and gonadotropin stimulation in human granulosa cells [58] and that perfusion of a Ca2+ chelator in rabbits significantly reduces gonadotropin-induced ovulatory efficiency [59]. Given adrenergic innervation of ovaries observed throughout metazoans, it is plausible to speculate that follicular adrenergic signaling plays conserved roles in regulating Mmp activity and ovulation (See below).
Adrenergic innervation of the ovary has long been found in mammals including humans. The role of adrenergic signaling in ovulation has been studied as early as the 1970’s. The neurotransmitter norepinephrine (NE) reaches the highest level in peripheral plasma during ovulation [60] and is enriched in the follicular fluid of preovulatory follicles compared to in peripheral plasma in healthy women [16,61,62]. Functional adrenergic receptors are expressed in mammalian ovarian follicular cells [13,58,63]. Ovarian perfusion of adrenergic agonists or antagonists influences the ovulation rate in rabbits and rats [12,14]. It has been speculated that adrenergic signaling regulates ovulation by stimulating muscle contraction or by increasing production of reactive oxygen species [16,53]. In contrast to this view, ovarian sympathetic denervation does not affect ovulation in rabbits and rats [10,15]; instead, it rescues ovulation defect in a rat model of PCOS [64,65], which is associated with increased sympathetic inputs to the ovary [8,9]. It is not clear why a discrepancy exists between the effects of surgical denervation and of pharmacological agents. Thus, no consensus has been reached in regard to the role of ovarian adrenergic signaling in mammalian ovulation.
Instead of regulating ovarian smooth muscle contraction, the results of the present study suggest an alternative pathway for ovarian NE to regulate ovulation. NE likely activates adrenergic receptors in granulosa and theca cells (equivalent to Drosophila follicle cells) in mammalian periovulatory follicles, which activates Mmp enzymatic activity at the apex [38], where mature oocytes rupture through. A surgical denervation may cause tissue damage and activate Mmps directly, bypassing the requirement of follicular adrenergic signaling. Future studies, using both mammalian and Drosophila genetic tools, will identify fundamental mechanisms of adrenergic signaling in ovulation.
Flies were reared on standard cornmeal-molasses food at 25°C unless otherwise indicated. OambMI12417 is a MiMIC line inserted in the coding intron of both Oamb spicing isoforms (S4 Fig) [46], and OambMI12417/Df(3R) BSC141 was used to characterize the Oamb mutant phenotype. TbHM18 [20] and Tdc2RO54 [22] were kindly provided by Dr. Mariana Wolfner. All RNAi-knockdown experiments were performed at 29°C with UAS-dcr2 to increase the efficiency of RNAi. R47A04-Gal4 (Oamb) and R44E10-Gal4 (lilli) from the Janelia Gal4 collection [42] were used for misexpressing genes or RNAi in mature follicle cells. The following RNAi or overexpressing lines were used: UAS-OambRNAi1 (V2861) and UAS-OambRNAi2 (V106511) from the Vienna Drosophila Resource Center; UAS-OambRNAi3 (B31233) and UAS-OambRNAi4 (B31171) from the Bloomington Drosophila Stock Center; UAS-Mmp2RNAi [66]; UAS-Mmp2 [47]; and UAS-GCaMP5G [67]. UASpGFP-act79B; UAS-mCD8-GFP[37] was used to analyze Gal4 expression in both germline and somatic cells, as well as neurons. UAS-GFPnls and UAS-RFP were used for follicle isolation. Control flies were derived from specific Gal4 drivers crossed to Oregon-R or yv; attP2 (B36303). The Mmp2::GFP fusion allele in the Mmp2 endogenous locus was used for detecting Mmp2 protein expression [37].
For the ex vivo follicle rupture assay, 4–6-day-old virgin females were used to isolate mature follicles, and follicle cells were fluorescently labeled using R47A04-Gal4 or R44E10-Gal4. Ovaries were dissected in Grace’s medium and ovarioles were separated from each other using forceps. This process will partially break the ovariole muscle sheath and release mature follicles. Mature follicles with an intact follicle-cell layer and completely dissociated from younger follicles were immediately transferred to new Grace’s medium to minimize their exposure to endogenous biogenic amines during dissection. With this method, we can isolate about 10 mature follicles/female and isolated mature follicles are no longer surrounded by ovariole or oviduct muscle sheaths (S3 Fig). Within one hour, isolated mature follicles were subsequently cultured in culture media (Grace’s medium, 10% fetal bovine serum, and 1X penicillin/streptomycin) supplemented with the indicated concentration of OA, TA, NE (Sigma), or ionomycin (dissolved in ethanol; Cayman Chemical). For chelating intracellular Ca2+, isolated mature follicles were treated with BAPTA-AM (dissolved in DMSO; Cayman Chemical) for 30 minutes before OA culture. All cultures were performed at 29°C, the same condition as flies were maintained, to enhance Gal4/UAS expression. About 25–30 follicles were used for each culture group and the percentage of ruptured follicles was then calculated as one data point. Typically three-six replicates were used for each genotype or treatment; data were represented as mean percentage ± standard deviation (SD); and Student’s T-test was used for statistic analysis. Ruptured follicles were defined as those losing more than 80% follicle-cell covering. With the exception of Fig 1D, all data were collected at the end of the three-hour culture.
For Ca2+ imaging and follicle rupture kinetics, video images were captured at 0.2 frame/second (FPS) with a sCOMS camera (PCO.Edge) installed in a Leica MZ10F fluorescent stereoscope. To examine the kinetics of follicle rupture, mature follicles were cultured in 20 μM of OA medium for 20 minutes at 29°C before recording. Unruptured follicles were then transferred into a home-made slide for video recording at room temperature. Each ruptured follicle was analyzed frame-by-frame manually to determine the ruptured distance between the posterior tip of the oocyte and the posterior leading edge of the follicle-cell layer using ImageJ. The percent of ruptured distance was then calculated as the ruptured distance divided by the length of the oocyte from the anterior to posterior tip. Because of the asynchronous onset of follicle rupture, data were normalized at the time point when follicles reach 50% ruptured distance.
In situ zymography for detecting gelatinase activity was performed as previously reported with minor modifications [37]. 50 μg/ml of DQ-gelatin conjugated with fluorescein (Invitrogen) was added into the culture media with or without OA for three hours. After a quick rinse, mature follicles with posterior fluorescent signal were directly counted. For egg activation, ruptured oocytes were treated with hypotonic activation buffer [45] for 15 minutes and treated with 50% bleach for three minutes. The number of unbroken oocytes was counted.
Egg laying, ovulation time, and follicle cell trimming were performed as previously described [28,37]. In brief, 4–6-day-old virgin females fed with wet yeast for one day were used. For egg laying, five females were housed with ten Oregon-R males in one bottle to lay eggs on grape juice-agar plates for two days at 29°C. After egg laying, ovaries were dissected and mature follicles in these ovaries were counted. The number of eggs on the plates was then counted, which was used to calculate the average time for laying an egg (egg-laying time). The egg-laying time was partitioned into the ovulation time and the uterus time (the time egg spent in the uterus and during oviposition). The partition ratio was determined based on the percentage of females having eggs in the uterus at six hours after mating. To do so, ten virgins were placed in a vial with 15 Oregon R males for six hours at 29°C, frozen for 4.5 minutes at -80°C, and then dissected to examine the reproductive tract. For follicle cell trimming, virgin or mated females were frozen for 4.5 minutes at -80°C, and ovary pairs were dissected, fixed, stained with DAPI, and mounted carefully to preserve the posterior end of mature follicles. Trimmed follicles were defined as more than a quarter of oocytes at the posterior end lacking follicle cell covering. Normalized trimming follicles were then calculated by the number of trimming follicles divided by the number of mature follicles in each female.
Immunostaining was performed following a standard procedure [68], including fixation in 4% EM-grade paraformaldehyde for 15 minutes, blocking in PBTG (PBS+ 0.2% Triton+ 0.5% BSA+ 2% normal goat serum), and primary and secondary antibody staining. For Mmp2::GFP localization, dissected tissues were stained in primary antibodies for 45 minutes at 4°C before the fixation treatment followed with the secondary antibody staining. Mouse anti-Hnt (1:75; Developmental Study Hybridoma Bank) and rabbit anti-GFP (1:4000; Invitrogen) were used as primary antibodies, and Alexa 488 goat anti-rabbit and 546 goat anti-mouse (1:1000, Invitrogen) were used as secondary antibodies. Images were acquired using a Leica TCS SP8 confocal microscope or Leica MZ10F fluorescent stereoscope with a sCOMS camera (PCO.Edge), and assembled using Photoshop software (Adobe, Inc.).
|
10.1371/journal.pntd.0005389 | Favipiravir pharmacokinetics in Ebola-Infected patients of the JIKI trial reveals concentrations lower than targeted | In 2014–2015, we assessed favipiravir tolerance and efficacy in patients with Ebola virus (EBOV) disease (EVD) in Guinea (JIKI trial). Because the drug had never been used before for this indication and that high concentrations of the drugs were needed to achieve antiviral efficacy against EBOV, a pharmacokinetic model had been used to propose relevant dosing regimen. Here we report the favipiravir plasma concentrations that were achieved in participants in the JIKI trial and put them in perspective with the model-based targeted concentrations.
Pre-dose drug concentrations were collected at Day-2 and Day-4 of treatment in 66 patients of the JIKI trial and compared to those predicted by the model taking into account patient’s individual characteristics. At Day-2, the observed concentrations were slightly lower than the model predictions adjusted for patient’s characteristics (median value of 46.1 versus 54.3 μg/mL for observed and predicted concentrations, respectively, p = 0.012). However, the concentrations dropped at Day-4, which was not anticipated by the model (median values of 25.9 and 64.4 μg/mL for observed and predicted concentrations, respectively, p<10−6). There was no significant relationship between favipiravir concentrations and EBOV viral kinetics or mortality.
Favipiravir plasma concentrations in the JIKI trial failed to achieve the target exposure defined before the trial. Furthermore, the drug concentration experienced an unanticipated drop between Day-2 and Day-4. The origin of this drop could be due to severe sepsis conditions and/or to intrinsic properties of favipiravir metabolism. Dose-ranging studies should be performed in healthy volunteers to assess the concentrations and the tolerance that could be achieved with high doses.
ClinicalTrials.gov NCT02329054
| In 2014–2015, the JIKI trial was conducted in Guinea to test favipiravir tolerance and efficacy in patients with Ebola virus disease (EDV). The main results of the trial were previously published without drug concentrations which were not available at the time of publication. The purpose of this study was to report favipiravir concentrations achieved in participants in the JIKI trial and to compare them with the targeted concentrations. We analyzed drug concentrations obtained at Day-2 and Day-4 and compared them to the targeted concentrations. At Day-2, favipiravir concentrations were significantly below but still close to the targeted concentration. At Day-4, a significant and unanticipated drop of concentrations as compared to Day-2 was observed. The origin of the lower-than-targeted concentrations and the unexpected drop could be due to severe sepsis conditions and/or to intrinsic properties of favipiravir metabolism. No significant correlation was found between the drug exposure and the virological response, indicating that it is possible that the favipiravir concentrations in the JIKI trial were not sufficient to strongly inhibit the viral replication. These findings suggest the necessity of performing dose-ranging studies with high doses of favipiravir in healthy volunteers to inform any further development of favipiravir for treatment of EVD.
| The 2014–2016 Ebola virus disease (EVD) outbreak in West Africa has been the deadliest occurrence of the disease since its discovery in 1976. Between January 2014 and June 2016, the World Health Organization reported 28,616 EVD cases, of which 11,310 were fatal [1]. In September 2014, at the peak of the outbreak, World Health Organization launched a fast-track process to identify potential anti-Ebola drugs and established three criteria for a drug to be acceptable as a candidate for clinical trials, namely i) availability of safety data in humans ii) evidence from preclinical studies of in vivo efficacy against Ebola virus (EBOV) iii) sufficient drug supply.
Favipiravir, a RNA polymerase inhibitor, approved in Japan to treat non complicated influenza infection, met all three criteria [2]. First the drug demonstrated antiviral activity against EBOV both in vitro (with a drug EC50 found between 10.8 μg/mL and 63 μg/mL) and in vivo in mice models [3,4]. Second it had been already safely administered to more than 2000 healthy volunteers and patients worldwide [5] and its pharmacokinetics (PK) was therefore well characterized for the influenza dosing. Briefly, the drug is a small and relatively hydrophilic molecule, with a protein bound fraction of 54% and a distribution volume between 15 and 20 liters [6]. Administered orally, the drug is rapidly absorbed with a tmax ranging from 0.5 to 1 hour and a bioavailability close to 100% [6]. The main elimination pathway involves hepatic metabolism by aldehyde oxidase, and marginally xanthine oxidase, producing a hydrophilic and inactive metabolite M1, which is eliminated in the urine [6]. Favipiravir inhibits aldehyde oxidase, leading to time- and dose-dependent pharmacokinetics [6].
In November 2014, our group decided to perform a historically-controlled, single-arm proof-of-concept trial to assess the tolerance and efficacy of favipiravir in patients with EVD in Guinea (JIKI trial) [7]. Launching an emergency trial in the midst of such an historical outbreak posed many human, logistical, ethical and scientific challenges. Among those was the choice of the dosing regimen to be used against EBOV, which has been explained prior to the trial implementation [8]. In brief the dosing regimen was found such that it achieves safely and rapidly free average concentration comparable to that obtained in mice successfully treated while maintaining free minimal concentrations higher than the drug EC50. Because the pharmacokinetics is nonlinear, the search for an optimal dosing regimen was based on a pharmacokinetic model developed by the manufacturer. Using this model a dosing regimen of 1,200 mg every twelve hours was proposed for the maintenance dose, with a loading dose of 6,000 mg (2,400; 2,400; 1,200 mg) on the first day. This dosing regimen was predicted to achieve stable concentrations after 48 hours, with median total trough (pre-dose) and average concentrations in plasma of 57.0 and 83.3 μg/mL, respectively [8]. One important aspect regarding this model is that it had been developed using data collected in studies in which the highest maintenance dose received was 800 mg twice a day and the largest treatment duration was 5 days. Doses in children were derived from adult doses and adjusted for body weight [7–9].
Overall, the JIKI trial results showed that mortality was strongly associated with baseline viremia. The results provided no evidence that favipiravir monotherapy at this dose might have a favorable benefit/risk ratio in patients with very high viral load at onset, but that it would merit future research in patients with a cycle threshold (Ct) ≥ 20, corresponding to a viral load below 107.7 genome copies/mL [7]. We previously published the trial results without reporting drug concentrations because they were not available at the time of the publication.
Here we report the results of the concentrations of favipiravir that were measured in patients of the JIKI trial. We compare them to the concentrations predicted by the model before the trial [8] and we analyse the possible association between drug plasma concentrations, viral loads and biochemical/haematological parameters.
Three ethics committees were approached, namely, the institutional review board of the Institut National de la Santé et de la Recherche Médicale (Inserm, France), the Médecins Sans Frontières International Ethics Committee, and the Guinean Comité National d’Éthique pour la Recherche en Santé. All three committees commented on the protocol and approved the final version and further amendments. Even though not asked for formal approval, because it was neither the sponsor nor investigator of the study, the WHO Ethics Research Committee received the trial protocol and provided important advice that helped improve it.
The design of the JIKI trial has been previously reported [7]. The inclusion criteria in the JIKI trial were the following: age ≥1 year, body weight ≥10kg, EVD confirmed by positive RT-PCR, no pregnancy, ability to take oral drug, oral or signed informed consent. In this PK sub-study, we included only patients of the JIKI trial that did not receive convalescent plasma prior to treatment, as we did for the main analysis [7], and who had at least one blood sample after the first day of treatment with sufficient volume to assess the favipiravir concentration.
All participants received standard of care and favipiravir. Favipiravir (Toyama-Chemical, 200 mg tablets) was given orally. The treatment started as soon as the consent was obtained (Day-0) and was administered for ten days. The adult dose was 6,000 mg at Day-0 (first dose: 2,400 mg; second dose eight hours after the first dose: 2,400 mg; third dose eight hours after the second dose: 1,200 mg) and 2,400 mg (1,200 mg every 12 hours) from day 1 to day 9. For children, the dose was adjusted on body weight [9].
Blood samples were taken at Day-0 (baseline), Day-2, Day-4, end of symptoms, Day-14 and Day-30. Favipiravir total concentration was measured at Day-2 and Day-4 from plasma or serum samples collected less than one hour before the first favipiravir intake of the day, i.e., between 11 and 12 hours after the last drug intake.
All samples were immediately decanted. EDTA, heparin or dry tubes were divided into aliquots, frozen at -20°C, and shipped to the INSERM Jean Mérieux biosafety level 4 laboratory in Lyon. In this laboratory, they were heated at 60°C for one hour to inactivate EBOV then refrozen (-20°C) and transferred to another INSERM laboratory in Marseilles for the drug concentration measurement, using a validated procedure (S1 Text). Previous study on plasma samples collected in nonhuman primates has shown that the inactivation process by heating did not significantly impact the quantification of plasma favipiravir concentrations (S2 Text). Serum and plasma samples were analysed using the same assay technique that had been validated for plasma samples. Both types of concentrations were referred as plasma concentrations in the following.
EBOV viraemia (molecular viral load) was immediately assessed at the onsite laboratories in four centers of the JIKI trial using a semi-quantitative RT-PCR assay (RealStar Filovirus Screen RT-PCR Kit 1.0, Altona Diagnostics). The results were expressed in terms of Ct, whose value is inversely proportional to viral load. An increase of 3 units in Ct scale corresponds approximately to a 1-log decline in viral load, therefore Ct unit corresponds to log scale for the viremia [7]. The Ct cut-off value for positivity was <40.
Biochemical and haematological parameter assays were performed using either the Piccolo Xpress (Abaxis) or the i-STAT (Abbott Laboratories) point-of-care system. We used here the results of the parameters that were available in most patients and were the most plausible to affect the drug pharmacokinetics, namely creatinine, sodium, albumin and haemoglobin.
All deaths during the trial were attributed to EVD and patients who were discharged were considered as patients who survived. Criteria for discharge were the absence of fever and significant symptoms for four consecutive days, ability to feed and walk independently, and two consecutive negative blood EBOV RT-PCR tests [7].
Of the 126 patients included in the JIKI trial, 10 were not included in the PK sub-study because they also received convalescent plasma. In addition, 21 patients died before the PK sampling time at Day-2 and 29 patients did not have enough plasma sample volume for drug concentration measurement. Thus a total of 66 patients were analysed in the PK sub-study (see flowchart in Fig 1), out of which 46 survived and 20 died. The median time from favipiravir initiation to death was 5 days (min-max: 2–17), with 8 patients who died between Day-2 and Day-4, 11 who died between Day-4 and Day-7 and one who died at Day-17. Patients took favipiravir as per the protocol and two missed doses were reported, both occurring the first day of drug initiation.
The characteristics of the patients included in the analysis are given in Table 1 and are compared with those of patients included in the main analysis of the JIKI trial but not included in the PK sub-study. The patients included in this sub-PK analysis had significantly higher Ct values, lower creatinine, CRP values at baseline and significantly lower mortality rate than those who were not included (30% vs 82%, p<10−7). This is due in particular to the fact that the 21 patients who died before the PK sampling time at Day-2 could not be included in this sub-PK analysis (Fig 1).
Among the 66 patients included in the PK sub-study, there were five adolescents (14–17 years old) and a 5-year old child. The child had a weight of 14 kg at inclusion and received 600/400/200 mg at Day-0 followed by a maintenance dose of 200 mg thrice a day. He had negative malaria test, initial EBOV Ct value of 19.9, initial EBOV viral load of 9.1 log10 copies/mL and no biochemical or haematological parameters before or during treatment. Three adolescents, aged 14, 14 and 17 and weighting 42, 31 and 48 kg, respectively, received loading doses of 1,600/1,600/800, 1,200/1,200/600 and 2,000/2,000/1,000 mg, respectively. Two adolescents, aged 15 and 16 years and weighting more than 50 kg received the adult dose. Among these 6 patients, only the 15-year old adolescent did not survive the infection.
Overall, 94 favipiravir trough concentrations were collected, 44 at Day-2 and 50 at Day-4. Among these 94 concentrations, 67 were obtained from plasma samples and 27 were from serum samples, and concentration in plasma and serum samples were largely similar (S3 Text). The median sampling time for the Day-2 measurement was 2.6 days after treatment initiation (min-max: 1.6–2.9) and the median sampling time for the Day-4 measurement was 4.6 days after treatment initiation (3.3–7.6). At Day-2, the median observed trough concentration was 46.1 μg/mL (23–106.9) and was close to the targeted trough concentration (57 μg/mL). However the concentrations dropped at Day-4 and the median concentration was 25.9 μg/mL (0–173.2) (Fig 2). This trend was observed in both plasma and serum samples (S3 Text). In patients having measurements at Day-2 and Day-4, the median reduction was -19.8 μg/mL (-54.6–-1.7) and was significantly different from 0 (p<10−5).
Next, we refined our comparisons by adjusting the model’s predictions for the individual dosing regimen, sampling time, age and weight for the predicted concentrations (Fig 3 and Table 2). In the 5 patients for whom the information was not available (Table 1), a weight of 70 kg was assumed. At Day-2 the median observed concentration was equal to 46.1 μg/mL and the median predicted concentration was equal to 54.3 μg/mL (p = 0.01). At Day-4, the difference was more pronounced, with a median observed concentration of 25.9 μg/mL compared to a median predicted concentration of 64.4 μg/mL (p<10−6). While the model predicted a modest median increase in concentrations equal to 5.1 μg/mL between Day-2 and Day-4, the drug concentrations actually had a marked median decrease equal to -19.8 μg/mL (p<10−8).
One patient for whom no initial Ct value was available was not included in this sub-study and two PK measurements did not have corresponding Ct values. Overall 65 patients with 90 simultaneous measurements of favipiravir concentrations and Ct values were included in the analysis (41 and 49 observations at Day-2 and Day-4, respectively). Regardless of the day considered, no significant relationship between the EBOV viral decline in plasma (increase in Ct value) and favipiravir concentrations could be established (Fig 4) and no significant association between mortality and drug concentrations was found (Table 3).
Longitudinal evolution of the biochemical and haematological parameters are displayed in S1 Fig. Albumin concentrations and haemoglobin levels decreased in most patients, while median sodium increased over time. For creatinine levels, two patterns were observed: in most patients, creatinine level decreased during treatment but in a subset of patients, creatinine increased strongly over time (S1 Fig). We found no significant correlation between the drug concentrations and any of the biochemical parameters at Day-2 or Day-4 (S2 Fig).
We reported here the favipiravir plasma concentrations obtained in 66 patients of the JIKI trial. The main finding of our analysis was that favipiravir concentration was significantly below but still close to the predicted value target concentration at Day-2, but decreased by nearly 50% between Day-2 and Day-4. Consequently, the concentrations at Day-4, with a median value of 25.9 μg/mL, were much below the predicted model-based value. With a protein binding of 54%, free favipiravir trough concentrations at Day-4 remained thus slightly larger than the in vitro EC50 reported in Oestereich et al., which was estimated at 10.5 μg/mL [4], but lower than those reported in another publication, where the EC50 of favipiravir was found larger than 31 μg/mL[3]. The conclusion was similar when model predictions were adjusted for individual dosing regimen or individual characteristics. In particular, these low concentrations were not due to a lack of compliance, as only two doses were not taken and both occurred on the first day of treatment initiation. Of note the assay technique was only validated for plasma samples and we did not distinguish these two types of drug concentrations. Yet serum and plasma are close matrices and similar values were observed in both plasma and serum samples, including the strong reduction in drug concentrations between day 2 and day 4 (S3 Text). Lastly, we did not find any significant correlation between the drug exposure and the virological response. Taken together, these results indicate that it is possible that the favipiravir concentrations in the JIKI trial were not sufficient to strongly inhibit the viral replication.
Yet this conclusion should be nuanced and taken cautiously for several reasons. First the analysis relied only on plasma favipiravir concentrations and intracellular concentrations of the active phosphorylated moiety were not available. For instance, intracellular concentrations of HIV nucleotide reverse transcriptase inhibitors were associated with antiviral efficacy, but not with the plasma concentrations of the corresponding nucleoside analogue [10]. Second our analysis only relied on pre-dose concentrations but other PK factors that could not be determined here (e.g., AUC, time above EC50 or EC99) could be a better marker of drug efficacy. Third, the fact that no significant correlation was found between the drug exposure and the virological response could also be due to the delay between infection and treatment initiation. For instance viral dynamic modelling shows that a drug affecting viral replication, such as favipiravir, will only have a limited impact on viraemia if treatment is initiated after the viraemia peak, regardless of drug efficacy [11]. Lastly, our study included only patients with drug measurements at Day-2 or after, which excluded the most severe patients who died before Day-2. Thus the patients analysed here are not representative of the JIKI population study and this is why they differed in terms of mortality, initial Ct, viral load or biochemical parameters (Table 1).
Two non-exclusive explanations for the lower-than-predicted concentrations and the unexpected drop between Day-2 and Day-4 can be proposed, namely the effect of the disease/treatment on the drug pharmacokinetics and the non-linear PK of favipiravir which has never been documented at this dosing regimen.
Many disease symptoms can affect the drug pharmacokinetic processes and lead to reduced concentrations. For instance reduced plasma favipiravir concentrations and altered kinetics of absorption and elimination were observed in a hamster model of arenaviral haemorrhagic fever [12]. Here in absence of frequent data points and historical data with the same dosing regimen, the effect of the disease cannot be evaluated. Obviously, the disease symptoms such as dehydration, diarrhoea, vomiting, and reduction of gut perfusion can hamper or modify favipiravir absorption. Likewise, the disease symptoms could also affect the bioavailability and the hepatic first pass, in particular through an increase in the activity of the main metabolic enzyme of favipiravir (aldehyde oxidase) with temperature [13]. In the JIKI trial, only 30 episodes of vomiting were reported within 30 minutes of drug intake, which represent 2% of the overall number of drug intakes during the trial [7]. Altered pharmacokinetics could also involve the distribution volume of favipiravir, which may be increased in Ebola patients due to treatment or to the disease itself and may explain at least in part the reduced plasma concentrations. Favipiravir’s apparent volume of distribution ranges from 15 to 20 L and is likely restricted to vascular and extra vascular fluids [5,6]. This distribution volume can be influenced by two factors, namely change in volume of body fluid and/or favipiravir protein binding. In this sub-study of the JIKI trial, 89.4% of patients received IV fluid rehydration during treatment [7], which may be responsible for haemodilution. Here, the modest decline in haemoglobin does not suggest a massive haemodilution but some vascular leakage resulting from infusion of a large volume of rehydration fluid [14] or from the disease [15] cannot be ruled out and could affect to some extent the favipiravir volume of distribution. However, such effects of infusion or disease are unlikely to solely explain the 50% reduction of favipiravir concentrations at Day-4 which would require a doubled volume of distribution. Another possible alteration of the drug PK might be due to the reduction in albumin concentrations. With a protein binding of 54%, a mean decline of about 20% in albumin levels between Day-0 and Day-4 observed in this study is unlikely to have a major effect on favipiravir distribution or elimination. Lastly, liver failure could impact favipiravir concentrations [16,17] but this should rather favour drug accumulation than accelerate elimination. Of note, no significant correlation was found between biochemical and haematological parameters and drugs concentrations but the number of observations available was limited (S2 Fig).
The other main cause of these lower-than-predicted concentrations could be the fact that the model used to predict the drug exposure in the JIKI trial was based on data collected in a very different context. Indeed, the drug was historically developed against influenza virus and the model was therefore developed using data collected with much lower doses of favipiravir (at most 800 mg BID) for shorter period of time (at most 5 days). Favipiravir is known to have non-linear pharmacokinetics due to its inhibitory effect on its main metabolic enzyme, aldehyde oxidase [5,6], which is also known to have several genetic polymorphisms with different catalytic activities [18]. The fact that the non-linearity of favipiravir pharmacokinetics was evaluated at doses much lower than those used in the JIKI trial [5] and that only few data on patients of African ethnicity were previously available made it complicated to predict the exposure of favipiravir with high doses in the JIKI study population. In addition, reduction in the drug concentrations over 14 days of treatment was also observed in uninfected non-human primates receiving high doses of favipiravir [19], suggesting that reduction in drug concentrations over time may be an unanticipated feature of the drug that is independent of the disease [19].
In conclusion, we have demonstrated that favipiravir plasma concentrations decreased with time and were likely too low in most patients. We advocate for a dose-ranging study on healthy volunteers to assess the pharmacokinetics and the tolerance of higher dosing regimen.
|
10.1371/journal.ppat.1000636 | Evasion by Stealth: Inefficient Immune Activation Underlies Poor T Cell Response and Severe Disease in SARS-CoV-Infected Mice | Severe Acute Respiratory Syndrome caused substantial morbidity and mortality during the 2002–2003 epidemic. Many of the features of the human disease are duplicated in BALB/c mice infected with a mouse-adapted version of the virus (MA15), which develop respiratory disease with high morbidity and mortality. Here, we show that severe disease is correlated with slow kinetics of virus clearance and delayed activation and transit of respiratory dendritic cells (rDC) to the draining lymph nodes (DLN) with a consequent deficient virus-specific T cell response. All of these defects are corrected when mice are treated with liposomes containing clodronate, which deplete alveolar macrophages (AM). Inhibitory AMs are believed to prevent the development of immune responses to environmental antigens and allergic responses by interacting with lung dendritic cells and T cells. The inhibitory effects of AM can also be nullified if mice or AMs are pretreated with poly I:C, which directly activate AMs and rDCs through toll-like receptors 3 (TLR3). Further, adoptive transfer of activated but not resting bone marrow–derived dendritic cells (BMDC) protect mice from lethal MA15 infection. These results may be relevant for SARS in humans, which is also characterized by prolonged virus persistence and delayed development of a SARS-CoV-specific immune response in individuals with severe disease.
| Severe Acute Respiratory Syndrome (SARS) occurred in human populations in 2002–2003 and was caused by a novel coronavirus (CoV). Human SARS was characterized by prolonged virus excretion, lymphopenia and delayed adaptive immune responses in patients with severe disease. Recently, small animal models have been developed that mimic some of the features of the human disease. Specifically, BALB/c mice infected with mouse-adapted SARS-CoV develop severe respiratory disease. Here, we show that the T cell response is defective in these mice and that this results from inefficient activation of the initial immune response to the virus. This defect can be corrected by several treatments, including depletion of inhibitory macrophages from the lungs and direct activation of respiratory dendritic cells, important in initiating the immune response or transfer of activated dendritic cells prior to infection. All of these modalities result in improved initiation of the immune response and an enhanced anti-virus T cell response. Inefficient activation of the immune response may play a role in human SARS, and our results suggest possible strategies that might be used to develop novel anti-viral therapies.
| The lung is exposed to many challenges, both environmental and pathogenic. Defense of this portal must be tightly regulated so that appropriate immune responses to pathogens are mounted but responses to innocuous antigens are minimized. Alveolar macrophages (AM) play a central role in maintaining this immunological homeostasis [1],[2],[3]. In the lung, resident AMs are continuously encountering inhaled substances due to their exposed position in the alveolar lumen, but they are kept in a quiescent state. They function poorly as accessory cells for in vitro T cell activation [4],[5] and in many situations actively suppress the induction of adaptive immunity through their effects on alveolar and interstitial DCs and T cells [6],[7],[8]. In vivo elimination of alveolar macrophages using clodronate-filled liposomes (CL) leads to overt inflammatory reactions to otherwise harmless particulate and soluble antigens [9]. Alveolar macrophages adhere closely to alveolar epithelial cells (AECs) at the alveolar wall and are separated by a distance of only 0.2–0.5 µm from rDCs [6]. In macrophage-depleted mice, DCs have enhanced antigen-presenting function [6]. It has been estimated that the pool of murine alveolar macrophages can process up to 109 intratracheally injected bacteria before there is “spillover” of bacteria to DCs and before adaptive immunity is induced [10].
Although the importance of such mechanisms to control undesirable responses to inert environmental antigens is self-evident, it is also axiomatic that countermeasures must be available to allow reversal of this inhibition after challenge with inhaled pathogenic (notably microbial) antigens. During infection with respiratory pathogens, such as influenza virus, antigen is acquired by respiratory dendritic cells (rDCs) and these cells must be sufficiently activated to overcome anti-inflammatory factors in the lungs. These rDCs then migrate to the lung draining lymph nodes (DLN) to initiate an antiviral CD8 T cell response [11],[12]. After the interaction of naive T cells with such antigen-bearing DCs, CD8 and likely CD4 T cells undergo activation and division in the DLNs and migrate into the lungs to eliminate virus-infected cells, leading to resolution of the infection [13],[14],[15]. Recently, a secondary peripheral interaction of CD8 T cells with antigen-bearing rDCs in the lung was found important for effective antiviral immunity [16]. Overall rDC activation is a prerequisite for initiation and maintenance of the immune response.
Patients with the Severe Acute Respiratory Syndrome (SARS), caused by a novel coronavirus (SARS-CoV), developed mild to fatal pulmonary disease, with a mortality incidence of 10% [17]. Patients with worse outcomes generally exhibited a more protracted clinical course, characterized by the development of Adult Respiratory Distress Syndrome (ARDS), as well as lymphopenia, neutrophilia and prolonged cytokine production [17],[18],[19],[20]. Virus could be detected in nasopharyngeal aspirate and feces for as long as 21 days after disease onset [19],[21]. Delayed virus clearance may have resulted from suboptimal T and B cell responses; suboptimal neutralizing antibody responses are detected in patients with severe disease [17],[18],[19],[20]. Numerous studies demonstrated that SARS-CoV infection fails to activate macrophages and dendritic cells. Although these cells can be infected, they are functionally impaired: antiviral cytokines such as type I interferon were not expressed and endocytic capacity (antigen capture) was compromised ([22],[23],[24],[25],[26],[27],[28] and reviewed in [29]). These unusual findings raised the possibility that initial infection with the virus resulted in delayed or suboptimal activation of the innate immune system. Inefficient activation of rDCs might be unable to counter the potent anti-inflammatory factors that are normally present in the lung, resulting in both a deficient T cell response and delayed kinetics of virus clearance.
Recently, rodent-adapted strains of SARS-CoV, which cause mild to fatal respiratory disease, were developed in several laboratories [30],[31]. Here, we demonstrate that lethal disease in mice infected with a mouse-adapted strain of SARS-CoV (MA15) can be prevented if AMs with anti-inflammatory properties are depleted from the lung prior to infection. Treatment with toll-like receptor (TLR) agonists to activate rDCs or transfer of activated bone marrow-derived dendritic cells (BMDC) also prevents a lethal outcome. Together, these results demonstrate that SARS-CoV, by inefficiently activating the immune system, uses a novel mechanism to evade immune recognition.
SARS-CoV infection results in inefficient activation of macrophages and DCs in vitro [22],[23],[24],[25],[26],[27],[28] and slow virus clearance and a prolonged clinical course in humans [17],[18],[19]. Similarly, MA15 infection in vitro did not result in upregulation of CD86 on AM (Fig. S2, Gating shown in Fig. S1 A). To determine whether inhibitory AMs play a role in MA15-mediated severe lung disease, we depleted these cells by intranasal administration of clodronate liposomes (CL). CL are useful for depletion of AM, and to a lesser extent, alveolar/airway DCs [9], but intranasal administration does not affect the level of circulating macrophages [32]. As a control, we treated mice with PBS as described previously [33].
BALB/c mice were treated with 75 µl of CL or PBS intranasally (i.n.) and total lung cells were harvested after enzymatic digestion. After 24 h, there was a decrease of AMs (CD11c+CD11b−siglec F+ [34]) in the lung, both in frequency (>70%) and absolute number (from 5–6×104 to 1–2×104 cells/lung), in CL, but not PBS-treated mice (Fig. S3 A and B). By 48 h, approximately 90% of AMs in the lung were depleted (Fig. S3 A and B).
To determine whether there was a change in clinical disease after AM depletion, BALB/c mice were treated with 75 µl of CL and infected i.n. with 3×104 PFU of MA15 virus. Mice were monitored daily for weight loss and mortality. At this virus dosage, control mice lost more than 20% of their body weight and 60%–70% of them died (Fig. 1 A), generally from day 6 to day 8 post infection (p.i.). Depletion of AM before inoculation (at day −1 and day −2) completely protected mice from this lethal infection and animals rapidly regained their body weight (Fig. 1 A). AM depletion at day 2 p.i. was not protective and may have resulted in more severe disease, as observed also in influenza A virus-infected mice [16]. Of note, 6 week old C57Bl/6 mice are resistant to MA15 infection and treatment at day −1 or 2 with clodronate had no effect on the clinical course in these mice (data not shown).
Clodronate treatment resulted in enhanced kinetics of virus clearance, with virus cleared from all treated but not control BALB/c mice by day 7 p.i. (Fig. 1 B). We next examined lung sections for changes in histology. There were no histological differences in the lungs between CL-treated and control mice at day 0, indicating that depletion of AMs did not result in significant inflammatory cell recruitment to the lung. From day 2 p.i., PBS-treated mice developed a rapidly progressive interstitial pneumonia with extensive edema and damage to bronchiolar and alveolar epithelial cells (Fig. 1 C). Inflammatory infiltrates were consistently identified from days 2-to 6 p.i. CL-treated mice had a much better outcome with less destruction of the pulmonary architecture, but extensive alveolar, interstitial and perivascular inflammatory cell infiltration (Fig. 1 C, day 4 and day 6). Total lung cell numbers are shown in Fig. 2 A. Clodronate treatment, by removing AM, also altered the inflammatory milieu of the lungs. As a consequence, levels of pro-inflammatory cytokines and chemokines, such as IL-1β, Il-6, IL-12, CCL2 and CCL3 increased within 24 hours of CL, but not PBS treatment, prior to virus infection. By day 2 p.i., levels of these cytokines were generally similar in CL and PBS-treated mice, consistent with the notion that a delayed, and possibly dysregulated, immune response contributed to severe disease in control mice (Table S1).
Infection with respiratory viruses such as influenza A virus and respiratory syncytial virus (RSV) results in recruitment of CD11c+MHC II+ DCs to the lung [12],[35],[36],[37]. Unlike these infections, recruitment of inflammatory cells, including DCs, to the lung is impaired in MA15-infected mice (Fig. 2 A). The total lung cell number increased slightly, but there was no appreciable change in numbers of the respiratory dendritic cells (rDC) in control mice. Clodronate treatment resulted in enhancement of inflammatory cell recruitment to the lung (Fig. 1 C and 2 A), with a nearly tenfold increase in numbers of rDCs within 6 days (Fig. 2 A). For these experiments, we distinguished two populations of rDCs: alveolar/airway dendritic cells (aDC: CD11c+CC11b−MHC II+) and interstitial dendritic cells (iDC: CD11c+CD11b+MHC II+) using the gating strategy shown in Fig. S1. By day 4 p.i., the frequencies of MHC IIhigh/CD86+ and MHC IIhigh/CD40+ aDC and iDC increased significantly in drug-treated mice but only modestly on iDC and not at all on aDC in PBS-treated mice Over the next few days aDCs and iDCs remain activated in CL-treated mice but mostly returned to a baseline state in control mice (Fig. 2 B and C). Concomitant with this recruitment and activation of rDCs, we also observed enhanced rDC migration to draining lymph nodes (DLN), using a tracking method in which rDCs are labeled in the lung by i.n. inoculation of carboxyfluorescein diacetate succinimidyl ester (CFSE) (see Materials and Methods and Fig. S1 B for gating) [12]. In all mice, rDC migration to the DLNs peaked at 18 h p.i., but migration was accelerated by treatment with clodronate. After 48 hours the frequency and number of CFSE+ rDCs in the DLNs decreased suggesting that the first 48 h p.i. were most important period for rDC migration. There was also a two-three fold increase in total cell numbers in the DLNs (Fig. 2 D). Collectively, these results show that DCs remained activated for longer times in the lung and exhibited enhanced migration to DLNs after CL treatment. A consequence of the increase in both numbers of rDCs and the frequency that was activated was a 30–50 fold increase in total activated DCs in the lung.
Since enhanced rDC migration to DLNs is predicted to result in enhanced virus-specific T cell responses, we next examined the magnitude of total and MA15-specific T cell responses in the lungs of CL treated and control infected mice. Clodronate treatment resulted in greater numbers of activated CD8 and CD4 T cells in the MA15-infected lung (Fig. S4 A and B), compared to PBS treatment, as determined by CD43 (clone 1B11) expression. The latter is upregulated on activated effector T cells [38],[39].
To assess effects on MA15-specific T cell responses, we initially identified a set of H-2d-restricted virus-specific CD4 and CD8 T cells epitopes using lung derived cells harvested from infected mice and a peptide library covering all four structural proteins (S, N, M, E) of SARS-CoV. Several IFN-γ inducing CD8 and CD4 epitopes in the spike (S) and nucleocapsid (N) proteins (S366–374, S521–529, S1061–1071 and N353–370) were identified (manuscript in preparation). Some of these epitopes were described previously, but S521 and S1061 epitopes were newly discovered. Of note, all other previously described H-2d-restricted T cell epitopes were not recognized by lung-derived T cells in our assays [40],[41],[42]. These previous reports identified T cell epitopes using adenovirus vectors or DNA constructs expressing single SARS-CoV proteins, or isolated peptides. We speculate that the numbers of T cells recognizing these previously described epitopes are present at very low levels in infected mice compared to the immunodominant epitopes that we identify, possibly because of differences in antigen presentation between infected and immunized mice.
Using these epitopes, we found that AM-depleted mice exhibited earlier and more robust virus-specific T cell responses, as measured by intracellular cytokine staining (ICS) for IFN-γ, whereas control mice had almost no virus-specific T cell responses at days 6 and 7 p.i. (Fig. 3 A and B). PBS-treated mice that survived until day 8 p.i. mounted virus-specific T cell responses in the lung, but at a level that was much less than observed in CL-treated mice. We confirmed that these cells were functional using in vivo cytotoxicity assays. Naïve splenocytes were costained with PKH26 and CFSE, pulsed with MA15-specific CD8 T cell peptides and adoptively transferred i.n. into mice 12 h before harvest of total lung cells. Robust CD8 T cell cytotoxic responses were observed in AM-depleted mice, with 40%–50% killing of virus-specific targets. By comparison, only about 5% of target cells were lysed in control mice (Fig. 3 C).
Results thus far suggest that inhibitory macrophages are dominant in MA15-infected lungs. In support of this, AM were only transiently and slightly activated, as measured by CD86 and CD40 expression, after infection with MA15 (Fig. 4 A). F4/80, considered a marker for macrophage maturation and phagocytosis [43], was present at lower levels on AMs harvested from uninfected mice compared to macrophages isolated from other sites (e.g., peritoneal macrophages [44], Fig. S5) and was not upregulated after MA15 infection (Fig. 4 A). Further, surface levels of CD200R, important in maintaining lung homeostasis, were higher on AM than peritoneal macrophages [44] (Fig. S5) and were not significantly downregulated after infection (Fig. 4 A), indicating that AMs continued to be inhibitory even after the onset of the infection. The number and frequency of AMs increased at day 2 before returning to baseline by day 6 p.i. in control mice but, as expected, remained low throughout the infection after clodronate treatment, (Fig. 4 B).
Mature “resting” AMs are able to suppress in vitro proliferation of homologous T-cells, and freshly isolated rDCs are poor antigen-presenting cells, consistent with a suppressive state [6],[45]. To confirm the inhibitory properties of AMs, we isolated aDCs from total lung cells and cultured them in vitro for 24 h in the presence and absence of AMs. When cultured in the absence of AMs, aDC upregulated expression of CD86, MHC II and CD40. Co-culture with AMs prevented CD86 and MHC class II, and to a lesser extent, CD40 upregulation (Fig. 4 C).
The prolonged presence of AMs in MA15-infected lungs suggested that AMs not only inhibited rDCs activation, and thereby delayed DC migration from lung to lymph nodes, but also inhibited the function of anti-virus T cells in the lung. To examine this possibility, we co-cultured AMs and T cells in vitro. Concanavalin A (Con A) and soluble anti-CD3 (sCD3) antibody treatment of lung cells resulted in proliferation of both CD4 and CD8 T cells as measured by CFSE dilution. This proliferation was almost completely inhibited by co-culture with purified AMs at a ratio of 10∶1 (10 T cells∶1 AM) (Fig. 4 D). Of note, endogeous AMs were removed from the lung cell preparations by incubation in a tissue culture plate for 2 h (90% depletion, measured by flow cytometry). In the absence of this prior incubation, no robust proliferation was observed. To assess the effect of AM on virus-specific T cells, we isolated CD8 T cells from MA15-infected, CL-treated mouse lungs at day 8 p.i. using microbeads and stained them with CFSE. Cells were then stimulated for 72 hours with lung cells or splenocytes that were pulsed with three MA15-specific CD8 T cell peptides (S366/S521/S1061) with or without AMs. Although only about 30% of CD8 T cells were MA15-specific, proliferation of CD8 T cells was clearly detected. When co-cultured with AMs, CD8 T cell proliferation was totally inhibited (Fig. 4 E). Thus, AMs inhibited both nonspecific and specific CD8 T cell proliferation. However, AM co-culture in vitro did not inhibit IFN-γ expression after stimulation with MA15-specific peptides (Fig. S6 A), consistent with previous data, showing that AMs did not inhibit IL-2 secretion by Con A-stimulated T cells [45]. Further, when AMs and T cells were separated by a transwell during co-culture, no significant decrease of proliferation was observed as measured by CFSE dilution (Fig. S6 B) suggesting that AM inhibition of T cell proliferation required direct cell contact.
The results described above raised the possibility that direct activation of rDCs in the lung or adoptive transfer of activated DCs to the lung would bypass AM inhibitory function. Signaling through Toll-like receptors (TLR) results in a series of signaling events that leads to the induction of an acute inflammatory response. Ligand binding to TLRs also results in dendritic cell maturation, which is necessary for the initiation of adaptive immune responses [46],[47],[48]. Previous reports showed that Poly I:C or CpG treatment protected animal from lethal virus infection, but the mechanism of protection was not investigated in those studies [49],[50]. In preliminary experiments, we treated mice with ligands for several TLRs, including poly I:C (TLR3), LPS (TLR4), CpG (TLR9), R837(TLR7), R848 (TLR7/8), Pam3CSK4 (TLR1/2), and Pam2CSK4 (TLR2/6). We observed that treatment with poly I:C (Fig. 5 A) and, to a lesser extent, CpG (data not shown), but not the other TLR ligands, protected mice from lethal disease. Consequently, additional analyses were performed after treatment with poly I:C and as a control, LPS since both are widely used to stimulate macrophages and DCs [51]. Poly I:C (20 µg/mouse)-treated mice lost about 10% of their original weight but quickly recovered within 7 days. The LPS-treated group (5 µg/mouse), lost more than 20% of their weight with death occurring in all mice within 6–7 days. Virus titers were higher at day 5 in the lungs of these mice compared to mice treated with poly I:C (Fig. 5 B).
Poly I:C, and to a much less extent LPS treatment resulted in enhanced CD86 and CD40 upregulation on AMs (Fig. 5 C) and rDCs (Fig. S7). Treatment with both TLR agonists resulted in a modest increase in F4/80 and a small decrease in CD200R expression (Fig. 5 C). Consistent with the results obtained after clodronate treatment (Fig. 3 A), poly I:C treatment resulted in an earlier and more robust antigen-specific T cell responses than observed in PBS (Fig. 3 A) or LPS-treated mice (Fig. 5 D, E). Nearly twenty fold more MA15-specific T cells were detected in the lungs of poly I:C treated mice compared to LPS recipients at day 7 p.i. and these cells were functional in in vivo killing assays (Fig. 5 E and F). To determine whether poly I:C or LPS directly activated AMs, AMs were isolated and stimulated in vitro with both agonists. After 24 h stimulation, poly I:C but not LPS treatment resulted in a pronounced upregulation of CD86 (Fig. 6 A). Further, treatment with poly I:C but not LPS partially reversed the ability of AM to inhibit CD8 T cell proliferation after stimulation with Con A or sCD3 (Fig. 6 B). These results indicate that poly I:C can abrogate AM inhibitory function both in vivo and in vitro, by directly activating AMs and rDCs.
Given these results, direct delivery of activated DCs to the lungs might overcome AM-mediated inhibition. Activated DCs exhibit an enhanced ability to migrate to DLNs and to stimulate CD8 T cell proliferation and IFN-γ expression [52],[53],[54],[55]. Since AMs were unable to inhibit costimulatory molecule expression on previously activated DCs (Fig. 7 A), we next assessed whether adoptively transferred activated DCs could bypass AM inhibition and protect mice from a lethal MA15 infection. For this purpose, bone marrow cells were harvested from naïve mice, and DCs selectively cultured by treatment with GM-CSF plus IL-4 for 6 days [56]. BMDCs were then activated with either LPS or poly I:C, which resulted in enhanced CD86 and MHC class II expression on BMDCs (Fig. 7 B). As expected, MA15 was unable to activate these cells. 3×105 activated or resting BMDCs were transferred to mice i.n. 18 h prior to infection. Mice that received BMDCs activated with either poly I:C or LPS were protected from a fatal outcome, although they still lost about 15% of their body weight. In marked contrast, recipients of resting BMDC were not protected (Fig. 7 C). Further, higher virus titers were detected in the lungs at day 5 mice that received resting BMDC as opposed to activated BMDC (Fig. 7 D). BMDC migration from the lungs to DLNs was accelerated by prior activation. More CFSE+ activated BMDC than resting BMDC accumulated in the DLNs of recipient mice (Fig. 8 A and B) and additionally, the total number of cells in the DLNs was increased dramatically by activated BMDC transfer (Fig. 8 B). Consistent with enhanced rDC migration to the DLNs, recipients of activated BMDCs developed more robust CD4 and CD8 T cell responses in the lungs when compared to those that received resting BMDC (Fig. 8 C and D). Nearly tenfold more MA15-specific T cells were detected in the lungs of activated BMDC compared to resting BMDC recipients at day 7 p.i. and these cells were functional in in vivo killing assays (Fig. 8 E). Collectively, these results indicate that adoptive transfer of activated BMDCs to the lung amplified virus specific T cell responses, cleared virus earlier and protected mice from lethal MA15 infection.
The pathogenesis of SARS in patients that exhibit more severe disease is not well understood but includes slow virus clearance and a prolonged clinical course [19],[21],[57],[58]. The results presented herein suggest that this severe disease may occur in part because infected individuals do not mount an appropriate anti-virus T cell response. Anti-virus CD8 T cells are critical for virus clearance in mice infected with other pathogens, such as influenza A virus and LCMV [15],[59], so it is not unexpected that they are necessary for resolution of infection with SARS-CoV. While lymphopenia is associated with a worse prognosis in SARS patients [17],[18],[19], no prior studies, to our knowledge, has shown that this poor prognosis results, in part, from a sub-optimal CD8 T cell response. This defect in development of a protective T cell response occurs because the virus does not reverse the anti-inflammatory state that is naturally present in the uninfected lung. These results are consistent with in vitro studies in which the SARS-CoV is able to infect but can not activate human DCs or macrophages [22],[23],[24],[25],[26],[27],[28]. This may occur, in part, because coronaviruses, including SARS-CoV, are “invisible” to cellular sensors in some cell types [29].
Alveolar macrophages play a central role in maintaining immunological homeostasis [1],[2] and actively suppress the induction of adaptive immunity through their effects on alveolar and interstitial DCs and T cells [6],[7],[8]. Several molecules, including nitric oxide, TGF-β and CD200R have been implicated in AM suppressive function. These molecules have either short half lives or require cell-to-cell contact [1],[44],[60],[61]. Consistent with this, AMs are separated by a distance of only 0.2–0.5 µm from rDCs in the lung [10]. Our results also suggest that cell contact or close proximity to target cells is required, because AMs were unable to suppress T cell proliferation if separated from responders by a transwell membrane (Fig. S6 B). In another mechanism that maintains an anti-inflammatory state in the lungs, AMs ingest and process innocuous antigen and bacteria before they can reach and activate rDCs [10]. AM depletion results in enhanced antigen-presenting function by rDCs [6] and in increased ability to lyse influenza A virus-infected cells [62].
These reports indicate that countering the quiescent, anti-inflammatory state of AM is critical for developing a protective immune response; our results indicate that infection with SARS-CoV reverses this quiescent state inefficiently. We used three approaches to support this conclusion. First, pre-treatment of MA15-infected mice with clodronate depleted AM, resulting in enhanced activation and migration of rDCs, which in turn led to the development of a vigorous and protective virus-specific T cell response in the lung (Fig. 2 and 3). The activation and migration of rDCs at early times p.i. are critical for the timely initiation of anti-SARS-CoV T cell responses. Consistent with this, treatment with clodronate at day 2 p.i. was not protective (Fig. 1), because rDC migration to the DLNs is largely complete by 48 hours p.i. ((Fig. 2 D) and [12]). Depletion at day 2 p.i. resulted in more severe disease, suggesting that in SARS-CoV-infected mice, virus-specific T cells require additional DC stimulation in the lungs, as occurs in influenza A-infected animals [16].
Second, activation of AMs and rDCs in situ via treatment with TLR agonists also circumvented the anti-inflammatory state of the lung. Our results showed that only poly I:C, a TLR3 agonist, and to a lesser extent CpG, a TLR9 agonist, were able to perform this function. TLR7 is primarily located on plasmacytoid DCs and the inability of R848 to protect mice indicates that activation of these cells was insufficient to induce a protective immune response. Poly I:C, which activated AMs and rDC in vivo (Fig. 5 C and S7) and in vitro (Fig. 6 A), protected animals from lethal MA15 infection. The ability of poly I:C to stimulate rDC activation and migration has been described previously [12], and is likely to explain its protective ability. It should be noted that poly I:C treatment also induced type 1 IFN expression in the lung. This may also have contributed to the protective effect of poly I:C, but this is not likely to be the major effect because SARS-CoV is only modestly sensitive to IFN treatment of cultured cells or of mice [63],[64]. In addition, CL treatment did not induce type 1 IFN in the lungs, showing that IFN induction is not required for protection (data not shown).
LPS, which is a TLR4 agonist, was unable to protect mice from lethal disease. We considered the possibility that LPS might have toxic effects unrelated to TLR4 binding, but treatment with monophosphoryl lipid A (MPLA), a derivative of LPS that is a TLR4 agonist but is less toxic [65],[66], was also not protective (data not shown). Our results are consistent with a recent study that showed that TLR4 ligation contributed to worse outcomes in several models of lung injury [67]. TLR4 ligation, in the absence of treatment with specific agonists, did not contribute to worsened disease in MA15-infected BALB/c mice since infection of TLR4−/− BALB/c mice did not result in significant differences in clinical disease when compared to wild type BALB/c mice (data not shown).
Third, we showed that adoptive transfer of activated but not resting BMDCs bypassed AM-mediated suppression and protected mice from lethal disease (Fig. 7). While DC maturation makes these cells the most potent in antigen presentation in an animal, it also results in the loss of ability to take up antigen. However, antigen macropinocytosis is transiently stimulated after activation [52], possibly explaining how transferred BMDC could acquire SARS-CoV antigen for presentation to T cells in the DLNs. Alternatively, mature DCs are able to uptake antigen for cross-presentation [68]. Activated BMDCs preferentially migrated to the DLNs (Fig. 8 A and B) and initiated a protective T cell response in the lungs (Fig. 8 C–E). This transfer was successful because inhibitory AMs cannot reverse prior rDC activation (Fig. 7 A). All of these three experimental interventions resulted in enhanced rDC migration to the DLNs, enhanced MA15-specific T cell responses at the site of infection, the lungs, and improved outcomes.
It is notable that virus-specific T cells are also critical for virus clearance in C57BL/6 mice, which are resistant to MA15 infection. Six week old mice deficient in recombination activating enzyme activity 1 (RAG1−/−) on a C57Bl/6 background do not clear virus when measured at 9 days [69] or even 21 days p.i. (data not shown), yet remain completely asymptomatic. On the other hand, mice with Severe Combined Immunodeficiency Syndrome (SCID) on a BALB/c background, which, like RAG1−/− mice, are genetically unable to mount a T cell response, develop clinical disease that is more severe than that observed in wild type BALB/c mice. All SCID mice succumb to the infection (data not shown), compared to a 60–70% mortality rate in BALB/c mice that are infected with the same dosage of virus (Fig. 1 A). Collectively, these results show that an optimal T cell response is required for virus clearance but that strain-specific components of the initial immune response, not yet defined, are critical for preventing clinical disease in resistant strains.
An outstanding question is why SARS-CoV does not activate AMs and rDCs in BALB/c mice. As described above, SARS-CoV does not efficiently activate human DCs or macrophages. We have also shown that MA15 does not efficiently induce costimulatory molecule upregulation on murine rDCs or AM in vivo and/or in vitro (Fig. 2 C, 4 A and S2). However, while most viruses have mechanisms to evade host recognition sensors, they still efficiently induce an immune response. For example, successful resolution of influenza A virus infections requires activation of immune responses via TLR7, RIG-I and NLR (NOD-like receptors) inflammasome pathways [70], even though influenza A virus encodes an immune-evading protein, nsp1 [71]. HSV, lymphocytic choriomeningitis virus, hepatitis C virus, RSV and human cytomegalovirus are recognized via TLR2-dependent mechanisms while the RSV F protein activates cells via a TLR4-dependent mechanism [37]. Some viruses, such as vaccinia virus, directly inhibit TLR expression, confirming the importance of these molecules in virus recognition by the host [72]. TLR signaling is also important for SARS-CoV recognition by the innate immune system, since C57BL/6 mice, which are very resistant to the virus, become susceptible when MyD88 is genetically deleted [69]. The precise TLR or other receptor required for protection in C57BL/6 mice is not known at present. Why this same pathway is not efficiently induced in BALB/c mice after MA15 infection will be an area of future investigation.
In conclusion, we have shown that lethal disease in mice infected with a mouse-adapted strain of SARS-CoV (MA15) is correlated with a lack of activation of AMs and rDCs. Further, lethal disease can be prevented if AMs with anti-inflammatory properties are depleted from lungs prior to infection. Depletion results in enhanced DC recruitment to the lung and accelerated migration to DLN, and a more vigorous anti-SARS-CoV T cell response. Treatment with TLR agonists to activate AMs and rDCs or transfer of activated BMDCs also prevents a lethal outcome. Together, these results demonstrate that SARS-CoV, by “hiding” from the immune system, uses a novel mechanism to evade immune recognition in mice. The pathogenesis of SARS in humans may involve similar stealth mechanisms.
Pathogen-free BALB/c mice were purchased from the National Cancer Institute (Frederick, MD). Mice were maintained in the animal care facility at the University of Iowa. Animal studies were approved by the University of Iowa Animal Care and Use Committee. African Green monkey kidney-derived Vero E6 cells were grown in Dulbecco's modified Eagle's medium (DMEM, GIBCO, Grand Island, NY) supplemented with 25 mM HEPES and 10% fetal bovine serum (FBS) (Atlas Biologicals, Fort Collins, CO). Mouse-adapted SARS-CoV (MA15) was a kind gift from Dr. Kanta Subbarao (N.I.H., Bethesda, Maryland) [30]. Virus was passaged once on Vero E6 cells.
Mice were lightly anesthetized with isoflurane and infected intranasally (i.n.) with 3×104 PFU of MA15 virus in 25 µl of DMEM medium. Mice were monitored for weight loss and mortality daily. All work with MA15 virus was conducted in the University of Iowa Biosafety level 3 (BSL3) Laboratory Core Facility. To obtain lungs for virus titers, animals were sacrificed at the indicated time points post-infection (p.i.) and lungs were removed into phosphate buffered saline (PBS). Tissues were homogenized using a manual homogenizer, and titered on Vero E6 cells. For plaque assays, cells were fixed with 10% formaldehyde and stained with crystal violet three days post-infection. Viral titers are expressed as PFU/g tissue.
A peptide library, covering all 4 structural proteins of SARS-CoV was provided by BEI Resources (Manassas, VA). Virus-specific peptides were synthesized by BioSynthesis Inc. (Lewisville, TX). TLR agonists poly I:C, Monophosphoryl Lipid A (MPLA), CpG, Imidazoquinoline compound (R837 and R848), Pam3CSK4 and Pam2CSK4 were purchased from Invivogen (San Diego, CA). LPS was purchased from Alexis Biochemicals (Farmingdale, NY).
Alveolar macrophage depletion was performed by treatment with liposomes containing dichloromethylene bisphosphonate (clodronate). Clodronate was a gift from Roche Diagnostics GmbH (Mannheim, Germany), and it was encapsulated in liposomes as described earlier [9],[33]. At the indicated times, mice were anesthetized by intraperitoneal injection of 2% avertin and administered 75 µl of clodronate liposomes, or PBS i.n.
Animals were anesthetized and transcardially perfused with PBS followed by zinc formalin. Lungs were removed, fixed in zinc formalin, and paraffin embedded. Sections were stained with hematoxylin and eosin.
Mice were anaesthetized with 100 µl pentobarbital (50 mg/ml, Lundbeck Inc., Deerfield, IL) at the indicated time points. The lung vascular bed was flushed via the right ventricle with 5 ml PBS to eliminate any blood and lungs and draining lymph nodes were then removed. Lungs were cut into small pieces and digested in HBSS buffer containing 2% FCS, 25 mM HEPES, 1 mg/ml Collagenase D (Roche, Indianapolis, IN) and 0.1 mg/ml DNase (Roche) for 30 min at RT. Lymph nodes were minced and pressed though a wire screen. Particulate matter was removed with a 70 µm nylon filter to obtain single-cell suspensions. Cells were enumerated by 0.2% trypan blue exclusion.
CFSE (Molecular Probes, Eugene, OR) was dissolved at 25 mM in DMSO stored at −20°C until use. The CFSE stock solution was diluted in DMEM media to a concentration of 8 mM and then administered i.n. (50 µl/mouse) following anesthesia with isoflurane [12].
The following monoclonal antibodies were used for these studies: rat anti-mouse CD3 (145-2C11), rat anti-mouse CD4 (RM4-5), rat anti-mouse CD8β (53-6.7), rat anti-mouse CD11b (M1/70), hamster anti-mouse CD11c (HL3), rat anti-mouse CD16/32 (2.4G2), rat anti-mouse Siglec F (E50-2440), mouse anti-mouse I-Ad (AMS-32.1), all from BD Bioscience (San Diego, CA); rat anti-mouse IFN-γ (XMG1.2), anti-mouse F4/80 (BM8), rat anti-mouse CD40 (1C10), all from eBioscience (San Diego, CA); rat anti-mouse CD43 (1B11, Biolegend, San Diego, CA); rat anti-mouse CD200R (OX-110, Serotec, Raleigh, NC).
For surface staining, 106 cells were blocked with 1 µg anti-CD16/32 antibody and 1% rat serum, stained with the indicated antibodies, and then fixed using Cytofix Solution (BD Biosciences). For intracellular cytokine staining (ICS), cells were cultured at 1×106 per 96-well at 37°C for 6 h or the indicated time period in the presence of brefeldin A (BD Biosciences). Cells were then labeled with surface antibodies, fixed/permeabilized with Cytofix/Cytoperm Solution (BD Biosciences) and labeled with anti-IFN-γ antibody. All flow cytometry data were acquired on a BD FACSCalibur or an LSR II (BD Biosciences) flow cytometer with CellQuest (BD Biosciences) and were analyzed using FlowJo software (Tree Star, Inc. Ash, OR).
In vivo cytotoxicity assays were performed on day 6 after MA15 infection, as previously described [73]. Briefly, splenocytes from naive mice were costained with PKH26 (Sigma-Aldrich, St. Louis, MO) and either 1 µM or 100 nM CFSE (Molecular Probes, Eugene, OR). Labeled cells were then pulsed with the indicated peptides (3 µM) at 37°C for 1 h and 5×105 cells from each group were mixed together (1×106 cells in total). Cells were transferred i.n. into mice and at 12 h after transfer, total lung cells were isolated. Target cells were distinguished from host cells on the basis of PKH26 staining and from each other on CFSE staining. After gating on PKH26+ cells, the percentage killing was calculated as previously described [73].
AMs were obtained from uninfected lungs as previously described [74]. Briefly, lungs were inflated with warm PBS containing 0.2% BSA and 12 mM lignocaine (Sigma-Aldrich, St. Louis, MO) via cannulation of the trachea, and were lavaged at least 6 times. Cells were collected by centrifugation, resuspended in RPMI 1640 (Gibco, Grand Island, NY) containing 10% FCS (Atlanta, Lawrenceville, GA) and cultured at 4×104 in each 96-well for 48 h before use in the presence or absence of stimulators [75].
To demonstrate inhibition of polyclonal T cell proliferation, 4×105 splenocytes or lung cells (after AM-depletion by attachment to plates for 2 h at 37°C) were labeled with 1 µM CFSE and added to wells, stimulated with 2.5 µg/ml Con A (Sigma) or 1 µg/ml soluble CD3 (eBioscience) and cultured with AMs at a ratio of 10∶1 for 72 h. For inhibition of virus-specific CD8 T cell proliferation, lung CD8 T cells were purified from AM-depleted, MA15-infected animals at day 8 p.i. using CD8 Microbeads (Miltenyi Biotec, Cologne, Germany) at day 8. Splenocytes pulsed with 1 µM peptides or CD8 T cell-depleted lung cells were added as APCs and cultured with AMs at a ratio of 10∶1 for 72 h. Cells were then harvested, stained with antibodies and subjected to flow cytometric analysis.
aDC population were purified from the lungs of naïve BALB/c mice by FACS sorting based on their expression of CD11c+MHC II+CD11b− (Fig. S1) and enriched to about 80% purity.
Bone marrow-derived DCs (BMDC) were generated as previously described [56]. Briefly, red blood cell-depleted BM cells were plated at a density of 1×106/ml in RP10 (RPMI with 10% fetal calf serum, 1.0 mM HEPES, 0.2 mM L-glutamine, 0.05 mM gentamicin sulfate, 1% penicillin- streptomycin, 1 mM sodium pyruvate, and 0.02 mM 2-mercaptoethanol) supplemented with 1,000 U/ml recombinant granulocyte-macrophage colony stimulating factor (BD Pharmingen) and 50 U/ml recombinant interleukin-4 (eBioscience). Cells were incubated for 6 days, with 75% medium replacement every 2 days. At day 6, BMDCs were stimulated with or without 20 µg/ml Poly I:C or 1 µg/ml LPS for 18–24 h. CD11c microbeads and a Miltenyi autoMACS magnetic cell sorter (Miltenyi Biotec, Cologne, Germany) were used to purify CD11c+ DCs. Purity was confirmed by flow cytometry. BALB/c mice were lightly anesthetized with isoflurane and 3×105 BMDCs in 75 µl PBS were adoptively transfer i.n. 18 h before MA15 infection.
A 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 (SEM). P values of <0.05 were considered statistically significant.
|
10.1371/journal.pcbi.1004960 | Theoretical Insights into the Biophysics of Protein Bi-stability and Evolutionary Switches | Deciphering the effects of nonsynonymous mutations on protein structure is central to many areas of biomedical research and is of fundamental importance to the study of molecular evolution. Much of the investigation of protein evolution has focused on mutations that leave a protein’s folded structure essentially unchanged. However, to evolve novel folds of proteins, mutations that lead to large conformational modifications have to be involved. Unraveling the basic biophysics of such mutations is a challenge to theory, especially when only one or two amino acid substitutions cause a large-scale conformational switch. Among the few such mutational switches identified experimentally, the one between the GA all-α and GB α+β folds is extensively characterized; but all-atom simulations using fully transferrable potentials have not been able to account for this striking switching behavior. Here we introduce an explicit-chain model that combines structure-based native biases for multiple alternative structures with a general physical atomic force field, and apply this construct to twelve mutants spanning the sequence variation between GA and GB. In agreement with experiment, we observe conformational switching from GA to GB upon a single L45Y substitution in the GA98 mutant. In line with the latent evolutionary potential concept, our model shows a gradual sequence-dependent change in fold preference in the mutants before this switch. Our analysis also indicates that a sharp GA/GB switch may arise from the orientation dependence of aromatic π-interactions. These findings provide physical insights toward rationalizing, predicting and designing evolutionary conformational switches.
| The biological functions of globular proteins are intimately related to their folded structures and their associated conformational fluctuations. Evolution of new structures is an important avenue to new functions. Although many mutations do not change the folded state, experiments indicate that a single amino acid substitution can lead to a drastic change in the folded structure. The physics of this switch-like behavior remains to be elucidated. Here we develop a computational model for the relevant physical forces, showing that mutations can lead to new folds by passing through intermediate sequences where the old and new folds occur with varying probabilities. Our approach helps provide a general physical account of conformational switching in evolution and mutational effects on conformational dynamics.
| The role of protein biophysics is increasingly recognized in the study of evolution, and the study of protein biophysics has also benefitted from evolutionary information [1–4]. Emerging from a more physical perspective of molecular evolution is the realization that natural selection can act on a nonsynonymous mutation as long as it modifies the conformational distribution, even if it leaves the folded structure of a protein unchanged and maintains the original biological function. For instance, if the mutation stabilizes a nonnative “excited” conformational state which is structurally distinct from native, this state can potentially serve an additional “promiscuous” biological function which is then subject to natural selection [5]. This effect, demonstrated experimentally [6], is a direct consequence of the ensemble nature of protein conformations and follows simply from the principle of Boltzmann distribution [7,8]. Similarly, even if the most stable structure of a protein is robust against a mutation, the protein’s functional structural dynamics can be modulated by the mutation, which should then also be subjected to natural selection [5,9]. In this way, positive selection of an excited conformational state favors mutations that gradually increase the stability of the excited state, so that it finally becomes the new dominant native structure or one of two (or more) native structures with comparable stabilities in a “bi-stable” (or “multi-stable”) protein. Protein sequences interconnected by mutations and encoding for the same folded structure form neutral networks [10]. Bi-stability was predicted to occur at the intersection of neutral networks [8,10].
Consistent with theory [7,8,11–14], some phylogenetically reconstructed ancestral proteins are bi-stable [15]. Although there is no direct measurement to date of a gradually shifting conformational equilibrium for a set of naturally occurring amino acid sequences traversing two neutral networks, recent advances in NMR spectroscopy allow mutational changes in the stability of nonnative excited states to be detected [16]. A handful of conformational switches and bi-stable sequences have now been designed in the laboratory [17–19]. Among them, the one that is most extensively characterized is the set of designed mutant sequences that span the human serum albumin-binding and IgG-binding domains of Streptococcus protein G [19,20]. The wildtype sequences of these proteins, termed GAwt and GBwt respectively, are of equal length (56 residues) in the experimental system. GAwt and GBwt have only 16% sequence identity with very different folded structures. GAwt folds to a three-helix bundle (3α), whereas folded GBwt is a helix packing against a four-stranded β-sheet (4β+α). By carefully selecting amino acid substitutions, Alexander et al. created mutant sequence pairs with 30%, 77%, 88%, 95%, and 98% identity while still maintaining the original different folds. A single L45Y substitution separates the pair of mutants GA98 and GB98 with 98% identity. L45Y switches the dominant fold of GA98 from that of GA (3α) to that of GB (4β+α) for GB98 [19,20]. As suggested by theory [7,8] and by molecular dynamics simulations of the unfolded states of the GA88/GB88 pair [21], appreciable excited-state populations for the alternative fold should be present in the GA/GB mutants with 95%, 88%, or even 77% identity. Ligand binding data provide evidence that GA98 and another mutant GB98-T25I that also adopts the 3α GA fold have excited-state populations of the alternative GB fold. However, GB98-T25I is the only mutant for which the alternative fold is directly observable by NMR [22], as nonnative populations lower than ~1% are currently difficult to detect experimentally. By simulating the folding energy landscapes of the mutants, the goal of the present computational analysis is to gain physical insights into the mechanism of the GA/GB conformational switch, including how it might evolve via a gradual increase in stability of the alternate fold as the mutants approach the switch.
The most direct method of molecular simulation is to use a completely general physics-based potential. Such an approach has succeeded recently in showing that it is computationally possible for a series of mutants of a 16-residue peptide to undergo an α to β switch [14]. Owing perhaps to the limitations of molecular dynamics forcefields [23,24], folding simulations with fully transferrable potentials have not reproduced much of the switching behavior of the larger GA/GB system [25,26], although complementary theoretical methods have made useful progress. For instance, some of the GA/GB mutants can be assigned to their correct native folds by various threading approaches [8,27] or a “confine-and-release” simulation algorithm applied to the GA88/GB88 and GA95/GB95 pairs [28], suggesting that the forcefields used in these techniques may be quite adequate. But the conformations sampled by these techniques are limited only to those very similar to the GA and GB folded structures [8,27], or at best include also a highly confined set of conformations between them [28]. As such, the rather restricted conformational sampling in these techniques can mask possible shortcomings of the forcefields, e.g., by missing low-energy conformations that the techniques fail to sample. Therefore, to address fundamental physics of the GA/GB system, as for any protein folding study, it is necessary to employ self-contained explicit-chain models that extensively sample both the folded and unfolded conformations [29].
One class of self-contained models proven useful in biomolecular studies is the Gō-like explicit-chain structure-based models (SBMs). These models are native-centric in that the only contacts favored by the potential are those that exist in the known native structures [29–32]. Most SBMs studied to date are single-basin in that they target a single native structure; but dual- and multi-basin SBMs can be constructed to fold to two or more native structures. The latter approach has been employed to analyze the conformational switches between different functional states of a protein [33–36]. A prime example is the large-scale allosteric conformational transition between the open and close forms of adenylate kinase [34,37]. Recent applications of dual-basin all-atom SBMs to the GA/GB system suggest that the conformational preferences of some of the mutants can be rationalized to an extent by their differences in steric packing [38,39]. However, the effects of nonnative interactions that are not present in either the GA or GB folds are not considered in these SBMs; but nonnative interactions are important for protein evolution because they may lead to detrimental aggregation [40–42]. In any event, the extent to which these dual-basin SBMs are generalizable is not clear. They have only been applied to a small number of mutants, viz., GA95/GB95 and GA98/GB98 in ref. [38] and GA98/GB98 in ref. [39]. Moreover, in some cases, it appears necessary to single out contacts involving the mutated residues for ad hoc treatment [38].
To delineate the utility and limitation of common physical notions in accounting for experimental GA/GB observations, we introduce a model that combines a SBM potential with a physics-based transferrable all-atom potential. Going beyond prior efforts that considered only two or four sequences, our model is applied coherently to an extensive set of twelve GA/GB sequence variants covering the 3α and 4β+α folds. Favorable nonnative contacts are possible in our formulation because of the transferrable terms. This “hybrid” modeling approach recognizes that current knowledge of protein energetics is not sufficiently adequate—thus the need for a native-centric bias—yet at the same time posits that physical nonnative effects should manifest at least as a perturbation [43]. Within this conceptual framework, the transferrable component represents what we believe we know physically, whereas the SBM component represents the extent of our ignorance, which we should aim to eliminate in the future. To tackle the GA/GB system, we generalize the well-studied hybrid approach for a single-basin SBM [43–50] to one based upon a dual-basin SBM [33–36,51]. The formalism is general, however, and thus should be applicable also to conformational switches other than GA/GB.
As detailed below, the GA/GB switching predicted by our model agrees with experiment. Moreover, the robustness and physicality of our predictions are buttressed by control simulations indicating a lack of folding of decoy protein sequences with folded structures very different from that of either GA or GB. Interestingly, refinements of the transferrable component in our potential to better account for the π-interactions of aromatic residues [52] leads to a sharper conformational switch, suggesting that incorporation of more accurate descriptions of the physical interactions can lead to tangible improvement of the model under the present framework.
As noted above, SBMs are valuable conceptual tools; but SBMs and hybrid models are admittedly interim measures. Ultimately, one wishes to simulate biomolecular processes using a completely transferrable physical potential. With this in mind, to maximize the physical content, our hybrid model was constructed with a native-centric, structure-based component as nonspecific and as unimposing as we found technically possible. For example, in contrast to previous all-atom SBMs for GA/GB [38,39] that enforce detailed native biases on dihedral angles and inter-atom distances [32], the SBM component of our hybrid model constrains only the Cα-Cα distances between residues that are at least three sequence positions apart. The rest of the interactions—including local backbone preferences and sidechain excluded volume—are provided entirely by the transferable component. The SBM component of our model is sequence independent, in that the same native-centric potential applies to all GA/GB variants (Fig 1). In this way, the spatially coarse-grained SBM component serves merely to enable folding to the GA or GB native structures in an unbiased manner, all the while reducing as much as possible any artefactual memory of the sequence-structure relationship of any particular sequence. Accordingly, the differences in population in the two alternate folds for different sequences are determined solely by the physical transferable potential that admits nonnative as well as native interactions.
As described in Methods, the present sequence-independent SBM component is based on the consensus Cα-Cα native contact maps for GA and GB. Each consensus map was constructed using the four PDB structures for GA or GB variants for which experimental folded structures are available (Fig 2a and 2b). The consensus map contains only the native contacts common to all four PDB structures. Two residues of a given PDB structure are defined to form a native contact if the closest distance between any two non-hydrogen sidechain atoms, one from each residue, does not exceed 6 Å. Here the SBM energy for each consensus residue-residue native contact is constructed as a multi-Gaussian well potential [53], wherein the position of the minimum for each of the wells is determined by the four defining PDB structures. In most cases, the individual minima fuse into a single wider well because they are in close proximity (Fig 2c), although in some cases they retain their distinct minima when there are larger variations in contact distances among the PDB structures (Fig 2d). The potentials for all contacts in the two consensus native contact maps (Fig 2e) are provided in S1 Fig and S2 Fig.
Summing the energy terms for individual consensus native GA contacts gives the overall native-centric potential EA for GA and EB for GB, the strengths of which are given, respectively, by εA and εB (Methods). A bi-stable SBM potential, ESBM, is then obtained by combining EA and EB. The multi-Gaussian contact potentials here ensure that all native conformers used as input for the SBM potential are at an energy minimum of the same depth (εA or εB) for a given fold. This approach captures the salient features of the two folds while allowing sufficient flexibility to accommodate variations in backbone and sidechain configurations among different GA/GB sequences.
To achieve an unbiased baseline sampling of the GA and GB folds, the SBM energy scales εA and εB are expected to be somewhat different and thus a calibration is necessary. Indeed, it has long been known from the study of single-basin SBMs that imposing a single SBM energy scale for different native structures would result in a spurious correlation between folding temperature and native contact density that is not observed experimentally [54]. For our system, the GA fold was found to be only slightly more dominant in test simulations using min(EA) = min(EB) and the GB fold was only slightly more dominant for εA = εB. (Supporting Information S1 Text and S3a and S3b Fig and S3c and S3d Fig, respectively), whereas εA = 0.96εB allows for unbiased baseline sampling to produce results consistent with experiment. To minimize native-centricity as much as possible, we have examined the effect of different overall SBM interaction strengths and arrived at a workable value of εB = −0.37 (S1 Text, S4 Fig and S5 Fig). This strength corresponds to a weak native bias relative to the transferrable component, yet strong enough to guide folding. Under εB = −0.37, on average only less than one third (18.9/60.2 = 0.31) of the stabilization of GB98 is contributed by the SBM component ESBM (S6 Fig). The rest (69%) is contributed by the transferrable Etrans. Further analyses in S1 Text and S7 Fig–S11 Fig, including Hamiltonian replica exchange simulations (S9 Fig and S10 Fig), indicate that the GA98/GB98 switching behavior is robust over values of εB ranging from −0.30 to −0.50 (S7 Fig and S8 Fig), and that folding and switching are observed only when neither ESBM nor Etrans vanishes (S11 Fig). We adopt for Etrans the implicit-solvent all-atom potential developed at Lund University (available as PROFASI), which accounts for backbone, non-bonded excluded-volume, hydrogen-bonding, charged and hydrophobic side chain interactions in a physical manner [55,56].
With a SBM component providing minimally necessary restriction on the accessible conformational space, the transferable component of our hybrid model modulates the stability of the native and unfolded populations. Using the progress variables QA and QB and the simulation procedure described in Methods, the present modeling setup correctly identifies the native basin of 12 sequence variants of the GA/GB system (Fig 3a). The variables QA≡EA/95εA and QB≡EB/137εB are continuum versions of the discrete native contact fraction Q commonly used in protein folding studies [57,58]. For the energy landscapes in Fig 3a, the GA and GB native basins are situated, respectively, at QA≈0.9, QB≈0.15 and QA≈0.3, QB≈0.85; whereas the basin for the common unfolded state is centered at QA≈0.4, QB≈0.15. The dual native-bias of the SBM notwithstanding, Fig 3a shows that the transferable component is sufficiently strong to capture the physical mutational effects, resulting in significant shifts in populations and, in the case of GAwt, GBwt and GA30/GB30, virtual depopulation of the entire alternate native basin.
We computed a free energy difference ΔF(GA-GB) ≡ −ln(PA/PB) between the GA and GB folds for all the sequence variants (Fig 3b), where PA and PB are the populations of the two native basins defined, respectively, by QA ≥ 0.6, QB < 0.6 and QB ≥ 0.6, QA < 0.6. Thus, a negative ΔF(GA-GB) favors GA whereas a positive ΔF(GA-GB) favors GB. The replica-exchange simulation results in Fig 3b show that the single L45Y mutation from GA98 to GB98 entails a small yet appreciable shift in favor of GB, a robust finding corroborated by constant-temperature simulations (S12 Fig). The aromatic Y45 partakes in a hydrophobic cluster in GB but apparently contributes little to stability in GA [22]. In the present transferrable potential, the hydrophobicity-based non-bonded energy term is mostly responsible for favoring this Y45-containing hydrophobic GB cluster because the strength of the term scales with the number of contacting nonpolar atoms, and aromatics provide large contact areas [55]. The three mutations separating GA95 and GB95 result in a more notable population shift. In addition to L45Y, the other two amino acid substitutions are I30F leading from GA95 to GA98 and L20A leading from GB98 to GB95. Notably, the phenylalanine substitution of I30F fits into the hydrophobic core of both GA (partially buried) and GB (almost fully buried).
As the sequence separation between the pair is further widened (GA88/GB88, GA77/GB77, GA30/GB30, and GAwt/GBwt differ by 7, 13, 39, and 47 mutations respectively, Fig 1), the GA/GB free energy difference increases. The value of ΔF(GA-GB) increases rather smoothly from GAwt to GBwt as expected. The only exception is the step from GA88 to GA95, for which there is a decrease in GB propensity instead of the expected increase (Fig 3b). As mentioned above, for the GA30/GB30 pair, and the GAwt/GBwt pair that shares only 16% of their amino acids, the preference for the dominant native structure is so strong that only the fringe but not the bottom of the alternate native basin was sampled (Fig 3a). These free energy shifts are echoed by the balance between transition frequencies to and from the native basins along Monte Carlo simulation trajectories. Using a three-state division of the QA/QB energy landscape into unfolded (U), GA, and GB regions, a gradual shift from U↔GA to U↔GB transitions is concomitant with the sequence variation from GAwt to GBwt (S1 Text and S13 Fig).
We also compared the experimental and simulated melting temperatures of the GA/GB variants (Fig 3c and 3d). Because the model potential in the present hybrid GA/GB model lacks cooperativity-enhancing desolvation barriers [59,60] and neglects temperature dependence in the solvent-mediated interactions [29,61,62], simulated and experimental Tms are not directly comparable. For example, as suggested by related kinetic trends in other protein folding models [29], insufficient folding cooperativity in the present hybrid model likely caused the simulated Tm range to be narrower than that observed experimentally (the Tm ratios of GB98 over GA77 is 0.99 for simulation and 0.88 for experiment; see Fig 3d). Nonetheless, for the sequence variants from GA77 to GB77, the correlation between simulated and experimental Tm is reasonably good. The consistency in Tm trend for seven of these eight variants is apparent in the comparison using a normalized non-absolute temperature scale (Fig 3c) as well as in the scatter plot for absolute temperatures (Fig 3d). The steady drop in experimental Tm from GA77 to GB98 was captured very well by simulation (Fig 3c). The outlier GB88 is known to be very unstable experimentally (Tm ≈ 44°C). Curiously, this effect is also reflected in our model, albeit to an exaggerated degree.
Combined structure-based clustering of the simulated GA98 and GB98 conformations allows for an analysis of likely kinetic events during bi-stable folding (Methods). The centroid positions of 50 conformational clusters on the QA/QB landscapes are shown in Fig 4 together with the outlines of the bi-stable GA98 free energy landscape, which is quite similar to that of GB98 (Fig 3a). The size of a cluster is the number of sampled conformations that are within a certain degree of structural similarity among themselves. Each centroid conformation is a representative of all the conformations in a given cluster. Fig 4 shows that the centroid positions cover most accessible regions of the free energy landscape. Naturally, the unfolded state harbors the majority of clusters because unfolded conformations are structurally most diverse. The most extended conformations are positioned in the bottom-left region with small QA and QB values as expected (cluster no. 7). Under our model potential, there is a significant bias in favor of helical structures instead of unstructured coils in the unfolded ensemble.
As has been demonstrated, kinetic information can be gleaned from features on low-dimensional free energy landscapes determined solely by equilibrium sampling of one or two progress variables [63–65]. In using QA/QB landscapes for kinetic inference, we are following this tradition. It should be noted, however, that not all kinetic properties, especially those related to kinetic trapping, are deducible from low-dimensional landscapes [45,50]. For instance, not all structurally similar conformations based on the superposition-map measure and indicated by connecting lines in Fig 4 are readily accessible to one another kinetically. Therefore, here we qualify the “transition state” and “intermediate states” suggested by free energy landscape features as “putative”. With this caveat in view, we identify the conformations around the 0.66 < QA < 0.74, 0.12 < QB < 0.22 bottleneck region as the putative transition state for GA folding. Likewise, we identify the conformations around the two bottleneck regions around 0.3 < QA < 0.55, 0.35 < QB < 0.43 and 0.28 < QA < 0.4, 0.58 < QB < 0.66 as two putative transition states for GB folding (yellow boxes in Fig 4), and the local-minima region between the latter two transition states as a putative GB intermediate state.
Along the QA direction at QB ≈ 0.15, a simple folding transition via a compact transition state TS-GA is apparent in Fig 4. This putative process starts from an extended, mostly disordered state (cluster no. 7). Subsequently, more helices form and the chain first collapses into a loose arrangement of three helices around TS-GA and then proceeds to form the ordered native GA structure, with cluster no. 43 and adjacent clusters differing only by their disordered termini.
Folding along QB at QA ≈ 0.35 is more complex. Fig 4 suggests that the second (C-terminal) β-hairpin is formed upon reaching the first GB transition state TS1-GB, but at this stage the rest of the protein chain is still relatively open. The GB intermediate state that follows consists mainly of a variety of conformations with the second β-hairpin aligned with the N-terminal β-strand. TS1-GB encompasses more conformational diversity than the single centroid conformation might convey. When we partition the conformational ensemble in this region into two or more clusters (S14 Fig), alternative pathways across this transition region appear possible. One of the alternate pathways may entail a “mirrored” version of the second β-hairpin collapsing and accumulating as an “off-pathway” intermediate (see, e.g., the centroid conformation of cluster no. 12 in S15 Fig). As such, conformations with this topology likely constitute a kinetic trap that requires significant unfolding before folding to the GB native state can proceed.
Direct transition from an “on-pathway” intermediate to native GB is expected for those conformations with native-like orientation of the terminal secondary structure elements. To reach the second putative GB transition state TS2-GB, excess helical structure needs to be converted into the fourth β-strand. The chain then proceeds to sample different near-native orientations of the central helix relative to the β-sheet, and attempt packing of the hydrophobic core before finally arriving at the GB native state (cluster no. 4).
A detailed analysis of the population shift caused by the L45Y mutation in the conformational clusters in Fig 4 indicates that L45Y can start biasing in favor of the GB structure even when the folding is in its early stage (S1 Text and S15 Fig). In this process, the aromatic-aromatic Y45-F52 interaction, which is more frequent in GB98 than in GA98, is seen as playing a significant role in the GB-favoring effect of L45Y (S16 Fig).
As a test of the robustness of our hybrid model, we challenged it by several other sequences from the PDB that have the same 56-residue chain length as the GA/GB sequences but with native folds different from either GA or GB. The same GA/GB SBM was applied with each sequence’s Lund potential used as the transferrable component. The goal is to ascertain whether these decoy sequences would mistakenly adopt the GA or GB fold. Seven of the decoy sequences tested behaved reassuringly. Despite the GA/GB SBM, they did not populate either of the GA/GB native basin, even though some of their native conformations have secondary structures similar to those of GA or GB (Fig 5a–5g). This result shows that Etrans can override ESBM, underscoring that the transferrable physical potential plays a highly significant, if not dominant, role in our model.
Among the decoys tested, serine protease inhibitor infestin 4 is an interesting exception because its native structure is not similar to GA but it populates the GA basin (Fig 5h); but the bulk of its conformations remain unfolded. In this regard, depopulation of both native basins is remarkable for the double helical Ral binding domain because its helical secondary structures are similar though not identical to that of GA (Fig 5i).
Finally, to test whether our model can fold a non-GA/GB sequence if its native fold is essentially identical to either GA or GB, we considered a modified 56-residue version of Protein L (Methods). Protein L has only ~ 16% sequence identity with GBwt but adopts the overall GB fold experimentally. Reassuringly, our simulation shows that the modified Protein L sequence is compatible with the GB basin but not the GA basin (Fig 5j).
Apart from decoys, we also challenged our formulation with an alternative structure switch in the GA/GB system discovered more recently. Experiments indicate that the T25I mutant of GB98 reverts back to the helical structure of the GA folds, but with an additional L20A mutation can be restored to the GB fold [22]. Our simulations show a high degree of bi-stability for these two sequences as for GA98 and GB98. Nonetheless, we also found a small free energy difference that is consistent with the experimentally observed native structures of these two variants (Fig 6a and 6b). Another pair of possible GA/GB switch sequences that came to our attention was proposed recently [66], but the predicted switching behavior has not been confirmed by experiment or investigated by explicit-chain modeling. Our simulations here are in agreement with the predictions in finding that sequence “S2” prefers the GA fold while “S1” prefers the GB fold (Fig 6c and 6d). The free energy differences for these two alternative switch mutations are provided in S17 Fig. Our results suggest that GB98-T25I,L20A and S1 favor GB via different mechanisms. GB98-T25I,L20A predominantly stabilizes the entire unfolded state and parts of the GB native state yet leaving the native GA basin appreciably populated (Fig 6b), whereas the S2 to S1 mutation P54V clearly destabilizes the GA fold (Fig 6c).
The analysis of the L45Y mutation in S15 Fig reveals that a major part of its stabilizing effect on the GB fold is through enabling the aromatic-aromatic Y45-F52 interaction in GB98. In view of this observation and the general importance of π-related interactions in biomolecular processes [49,67], we constructed a rudimentary π-π interaction potential for F and Y residues (Methods). Our goal here is to explore how an orientation-dependent interaction between aromatics that goes beyond simple hydrophobic effects may affect the behavior of the GA/GB conformational switch, although a comprehensive study of aromatic interactions is beyond the scope of this work. By using three geometric variables for two neighboring aromatic rings (Fig 7a), we derived an empirical π-π potential [68] for F and Y from PDB statistics (Fig 7b). When this π-π potential replaces the simpler hydrophobic interactions among F and Y residues in the original Lund potential, the effect of L45Y is affected appreciably (Fig 7c).
We define a difference landscape for the original Lund potential (Fig 7c, left) as the difference between the GA98 and GB98 panels in Fig 3a. The difference landscape for the modified transferrable potential (Fig 7c, right) is similarly defined using the QA/QB landscapes of GA98 and GB98 in S18 Fig that incorporates our π-π potential. In the Lund potential (Fig 7c, left), L45Y stabilizes the unfolded and GB intermediate states rather homogeneously (stabilization indicated by blue coloring). The GA native basin is destabilized (red coloring), but so are parts of the GB native basin. In contrast, with the π-π potential (Fig 7c, right), L45Y has a stronger impact. It now destabilizes most of the unfolded state and parts of the GA native basin whereas the stabilization focuses more on the intermediate and native basins of GB. Although the present π-π potential is rudimentary, this comparison suggests that orientation-dependent π-π interactions likely play a significant role in the experimental sharpness of the GA/GB conformational switch.
To recapitulate, we showed that a coarse-grained Cα SBM in combination with an all-atom transferable potential correctly identifies the native state of an extensive set of GA and GB sequence variants. As shown above, our hybrid model is well suited for exclusively selecting the correct native state for GA/GB pairs of up to 77% identity. At higher sequence similarity, both folds were populated in our simulations; but a clear preference consistent with experiment was observed.
Beside this overall success, two findings from our investigation are of experimental relevance: (i) existence of an equilibrium intermediate for GB folding (Fig 3a, GB panels); and (ii) a critical role of the second β-hairpin in the GB folding pathway (Fig 4 and S15 Fig). On both counts, our model results are in general agreement with experimental findings (see below), lending additional credence to our contention that the present hybrid model is capable of capturing essential physics of GA/GB bi-stability and the GA98/GB98 conformational switch.
Firstly, our prediction of a GBwt (also called protein G or GB1) intermediate is in line with several [69–72] though not all [73] simulation studies. Experimental evidence for a GB folding intermediate was presented, but there is no clear consensus yet regarding the existence and/or nature of a GB intermediate–unlike the generally recognized two-state nature of GA folding. Two early experiments oncluded that GBwt folding is two-state [74,75]. In contrast, another early continuous-flow ultrarapid mixing experiments on GBwt suggested a native-like intermediate [76], but this conclusion was disputed [77]. A later FRET study also found an intermediate near the urea denaturation midpoint of GBwt [78]. A subsequent equilibrium GBwt unfolding experiment showed two-state behavior; but the kinetic chevron rollover was indicative of an intermediate [79]. The latter finding is in line with a recent experimental and molecular dynamics study showing that GBwt folding is three-state [80]. As for GB variants, one study found that GB88 and GA88 are two-state folders [21]. However, an investigation on a different set of variants GA30/GB30, GA77/GB77, and GA88/GB88 supported three- and two-state folding, respectively, for all GB and GA variants [81]. Taken together, recent evidence appears to be somewhat more preponderant on the existence, rather than non-existence, of a GB folding intermediate; and is unequivocally affirmative of the two-state nature of GA folding. This trend is reflected by our simulated free energy landscapes in Fig 3a.
Secondly, Fig 4 and S15 Fig suggest that the second β-hairpin is critical and more important than the first β-hairpin in GB folding. Although this finding was deduced from an analysis of GA98 and GB98 clusters, it is likely applicable to other GB variants, including GBwt, because of the similarity among their free energy landscapes (Fig 3a). Indeed, NMR experiments on peptides from GBwt found that, in isolation, the second β-hairpin is much more stable than both the helix and the first β-hairpin. It forms a stable, native-like β-hairpin with its three aromatic residues W43, Y45, and F52 forming a cluster stabilized by both hydrophobic and (probably π-related) polar interactions [82]. In contrast, the first hairpin was found to be mostly flexible in isolation and not native-like [83]. Hydrogen exchange experiments on the entire GBwt protein also revealed an early folding state with the second β-hairpin having the highest protection factors, whereas the helix has a lower and the first hairpin has the lowest [77,84]. Based on Φ-value analysis for a single transition state, another study also pointed to the presence of the second β-hairpin in the GBwt transition state [74]. Taken together, the experimental data summarized above provide support for a critical role of Y45-F52 in favoring early formation of the second β-hairpin and its partial collapse together with the helix, as suggested by our simulation (compare TS1-GB in Fig 4 and S14 Fig).
In this regard, some differences between the folding transition states of GB variants and that of GBwt were reported. In particular, Φ-value analysis [85] has found that the first transition state in GB30 is more sensitive to mutations in the second β-hairpin whereas GB88 is more sensitive in the first hairpin [81]. Nonetheless, the same set of data for GB88 is suggestive of native-like transition-state contacts, such as I6-T53, that are between strands at the two termini because some of their residues have high Φ-values (e.g., 0.48 for I6 and 0.42 for T53). If this is indeed the case, the experimental data is not inconsistent with our simulation result suggesting that the anti-parallel alignment of the termini is an early rate-limiting event for GB folding (Fig 4 and S15 Fig).
Taking all the evidence presented together, the performance of our model suggests that the remarkable GA/GB bi-stability phenomenon can be rationalized to a significant extent by specific hydrophobic interactions, though our physical understanding is still far from complete. As discussed above, future improvement in matching theory with experiment should be sought by enhancing folding cooperativity and increasing sharpness of the conformational switch in our model. One possible direction is to incorporate desolvation barriers in the transferrable potential because this is a robust physical feature of solvent-mediated interactions that have a significant impact on folding cooperativity [29]. Another direction, which was initiated with some success here, is to devise a more accurate description of aromatic interactions [67]. In this respect, a natural next step is to extend our model π-π interactions to encompass Trp and to adopt a more comprehensive account of the relative position and orientation of interacting aromatic sidechains that goes beyond the three variables in Fig 7.
Despite the simplicity of the Lund potential, it has succeeded in folding several smaller proteins [55,86] and the 92-residue Top7 [87]. However, in long unbiased folding simulations using only the Lund potential with no SBM, we were unable to observe stable native-like conformations of GA/GB variants, indicating that as-yet-unknown energetic contributions, in addition to those in the Lund potential, are needed for a complete physical account. The GA/GB system is a useful benchmark for testing forcefields and simulation techniques. Recent success in using all-atom explicit-water molecular dynamics to simulate folding of a number of small proteins is remarkable [88–90]. However, despite the notable advance and ongoing force-field improvement [23,91], no ab initio forcefield to date has been able to fold the GA/GB variants correctly [25].
In this context, hybrid modeling is a highly useful interim approach to gain physical insight into protein folding energetics, effects of mutations, and to assist in protein design. Owing to its reliance on SBMs, this approach is limited to proteins with known structures. Nonetheless, for many globular proteins, the native structure is either known or can be inferred through homology or sequence-based statistical models [92,93], and are therefore amenable to hybrid modeling. Common approaches to estimate mutational ΔΔG [94] only consider the known native structure with little or no regard to the unfolded state and folding dynamics. Hybrid models can address this shortcoming by providing testable predictions about the mutational effects on the entire free energy landscape. Indeed, because of its computational tractability, hybrid models can facilitate efficient development and testing of physically more accurate transferrable potentials, and thus can contribute to an ultimate elimination of the current necessity for SBMs.
As described above in Results, we derived for the native-centric SBM component of our hybrid model two consensus native contact maps that capture the general features of the GA and GB folds by using PDB structures for four GA sequence variants and four GB sequence variants (Fig 2a and 2b). The sequences and their corresponding structures (in parentheses) are GAwt (2FS1), GA88 (2JWS), GA95 (2KDL), GB98 (2LHC), GBwt (1PGA), GB88 (2JWU), GB95 (2KDM), and GB98 (2LHD). All of these PDB structures except the x-ray structure for GBwt were determined using NMR and contain multiple model structures. For simplicity, we used only the first model in each NMR PDB file in our analysis. Assuming that these consensus contact maps provide a reasonable coverage of the structural variations in the GA/GB system, we apply these maps to sequence variants GA30, GB30, GA77, and GB77 as well, since no detailed structural data were available for the latter four sequences [20].
We introduce EA and EB as the individual native-centric potential energy functions for the GA and GB folds, respectively. EA and EB depend on the Cα-Cα distances rij for all residue pairs i,j that belong to the given consensus native contact map via the following Gaussian form [53]:
EA=εA∑i,jnA[∏sns(1−e−(rij−dij(s))2/2w2)−1],
and a similar equation for EB with all instances of “A” replaced by “B”. Here the summation over i,j for EA and EB runs over, respectively, all nA = 95 and nB = 137 contacts in the consensus contact maps for GA and GB. The product over s takes into account the multiple native distances
dij(s)
for residue pair i,j in the ns = 4 PDB structures contributing to the consensus map. The strength of EA or EB is given, respectively, by εA or εB, which corresponds to the well depth for a single native contact. The w parameter that controls well width is set at 0.5 Å. In the present study, this formulation leads to a wide potential well for an overwhelming majority of consensus contacts. Because in most cases the native Gaussian wells for individual structures overlap considerably, we observe only minor barriers between individual Gaussian minima among all the consensus native potentials shown in S1 Fig and S2 Fig. In Fig 2c and 2d, examples of the consensus potential
Eij=∏sns{1−exp[−(rij−dij(s))2/2w2]}
for an individual contact (black curves) are provided together with the corresponding energy term
1−exp[−(rij−dij(s))2/2w2]
for one of the four contributing PDB structures (color curves).
The above Gaussian form of the native-centric energy function is more suitable than the Lennard-Jones (LJ) form for our present purpose. As has been noted, it is difficult to produce a viable combined energy function from multiple native-centric LJ functions for multiple structures unless the conformational diversity is approximated by a single centroid structure [95]. LJ potentials are inflexible in their well shape (width). Each inter-residue contact comes with a built-in repulsion term determined by the minimum-energy contact distance in LJ. As a result, multiple instances of the same contact at varying distances can lead to occlusion of the shorter-range contact by the repulsion of the longer-range contact if the LJ form is used instead of the Gaussian form to construct a combined energy function in accordance with the equation above (S1 Fig and S2 Fig).
As outlined above, the total potential energy Etotal is the sum of a native-centric component and a transferrable component, viz.,. Etotal = ESBM + Etrans. The dual-basin native-centric SBM component ESBM is constructed simply as ESBM = EA + EB. Aiming to increase the weight of the transferrable component in our model potential, we did not employ the more native-specific prescription of logarithmic mixing in ref. [35] for ESBM. For the transferrable component Etrans, we adopt the Lund potential: Etrans = Elocal + EEV + EHB + ESC + EHP, where the energy terms on the right are for local backbone interactions (Elocal), non-bonded excluded volume (EEV), hydrogen bonds (EHB), charged (ESC) and hydrophobic (EHP) sidechain interactions. Bond lengths and bond angles are kept constant, as described by the original authors [55]. Dimensionless energy units are used in our simulations with Boltzmann constant kB effectively set to unity.
We use a Monte Carlo (MC) [96] package [56] from Lund University to conduct parallel tempering (temperature replica exchange) MC simulations [97]. MC chain moves included backbone and side chain rotations as well as biased Gaussian steps [98]. All simulations were initialized from random chain conformations and time propagated in units of MC cycles. Each cycle consisted of a number of elementary conformational MC updates scaled to the number of rotational degrees of freedom of the simulated protein chain so that on average all degrees of freedom were perturbed once per cycle. For example, for GA98 and GB98 these numbers of degrees of freedom were 283 and 282, respectively.
Initially, parallel tempering simulations were performed over 32 replicas per simulation over a wide temperature range. This is then followed by a second simulation using a finer temperature grid around the melting (unfolding) temperature, Tm, determined as the temperature at which the heat capacity function
CV(T)=1kBT2(⟨Etotal2⟩T−⟨Etotal⟩T2)
computed from the first set of simulations attains its maximum. Here T is absolute temperature of the simulation, Etotal is the total energy defined above, and <…>T denotes conformational averaging at T. The refined temperature grid was tuned to ascertain sufficient replica exchange acceptance probability around Tm (~99%). Replica exchange was attempted every 5,000 MC cycles, the first 30% (3.0×106 MC cycles) of every trajectory was excluded from analysis. Populations simulated at different temperatures were reweighted to Tm using WHAM [99]. In select instances, 128 constant-T simulations at Tm with increased sampling were conducted to corroborate parallel tempering results (S5 Fig and S12 Fig). In view of the need for a high computational throughput for varying input parameters and sequences, most simulations were terminated after 107 MC cycles (~2.8×109 elementary MC updates). We verified that the resulting simulated ΔF(GA‒GB) for GA98 and GB98 is reasonably robust in longer simulations.
From the replica exchange simulations around Tm for GA98 and GB98, for each sequence we randomly sampled 20,000 conformations obtained at the two sequences’ respective Tms. These conformations were combined into a single pool of 40,000 conformations for clustering analysis. Each conformation in the pool was represented as a (4×56)-dimensional vector. The first 56 and second 56 components of this vector are the distances between the Cα atoms in the given conformation and the corresponding Cα atoms, respectively, of an optimally superposed GA98 PDB structure (2LHC) and an optimally superposed GB98 PDB structure (2LHD). Similarly, the third 56 and fourth 56 components of the (4×56)-dimensional vector are the distances between the Cβ atoms in the given conformations and the corresponding Cβ atoms in the optimally superposed PDB structures, respectively, for GA98 and GB98. Structural superpositions were optimized using the MDtraj [100] implementation of Theobald’s algorithm for RMSD calculations [101]. The (4×56)-dimensional distance vectors were then clustered by the k-means algorithm [102] with k = 50 chosen as the number of clusters. Cluster centroids are defined as actual conformations situated closest to the cluster centers in the (4×56)-dimensional space.
We define a distance measure between the centroids of two conformational clusters as the Cartesian distance between the centroids’ (4×56)-dimensional vectors normalized by (4×56)1/2. We refer to this distance measure as RMSDsm because it is the root mean square difference of the centroids’ superposition maps, RMSDsm. The latter is defined for any pair of conformations Cμ and Cν as
RMSDsm(Cμ,Cν)=14×56∑i=14×56(di(μ)−di(ν))2
where
di(μ)
and
di(ν)
are the components of the (4×56)-dimensional vectors representing, respectively, conformations Cμ and Cν. RMSDsm was first used in the general clustering for all conformations. For Fig 4, the general definition was applied to pairs of cluster centroids, wherein only pairs with RMSDsm ≤ 5.75 Å are shown by connecting lines. This threshold was chosen solely for the presentational purpose of not obstructing the visualization in Fig 4 yet providing as much information as possible about the structural relationships between clusters that share a reasonable degree of geometric similarity.
The sequence of the modified version of protein L in Fig 5 was obtained by first structurally aligning its PDB structure (2PTL) with that of GB1 (1PGA) and then removing the unaligned N- and C-terminal tails. Internal loop residues 12, 40, 41, and 42 were also removed and a glycine was inserted between residues 23 and 24. This procedure led to the following sequence used in Fig 5: VTIKANLIFANSTQTAEFKGTFAEKATSEAYAYADTLKKEYTVDVADKGYTLNIKF.
Interactions between aromatic residues in the Lund potential are treated only by its hydrophobic side chain potential [55]. To explore possible π-interactions that are not hydrophobic in nature but are nonetheless known to play significant structural roles in biomolecules [49,67,103,104], we modified the Lund potential for Phe and Tyr, replacing their contact-area-dependent hydrophobic interactions by an orientation-dependent potential. This rudimentary π-π potential is parametrized by three geometric variables r,θ,φ characterizing the relative position and orientation of two aromatic rings (Fig 7a). There is one Trp in the GA/GB sequences (W43); but for simplicity we restrict our exploration to Phe and Tyr, leaving the treatment of the geometrically more complex Trp to future studies. The present π-π interaction is derived as a statistical potential from a PDB data set obtained through the PDB-SELECT [105] repository at http://swift.cmbi.ru.nl/gv/select/index.html. The sequence similarity cut-off was 30%, R-factor cutoff was 0.21, and resolution cut-off was 2.0 Å. The dataset contained 9,796 protein crystal structures (created on January 26, 2013). For all the observed F-F, Y-Y, and F-Y contact pairs in this data set, the number of occurrences P(r,θ,φ) of r,θ,φ were distributed into bins of size 0.3 Å for r between r = 3 Å and 12 Å and bins of size 3° for θ,φ between θ,φ = 0° and 90°. Based on this statistics and following Procacci and coworkers [68], we define a rudimentary π-π interaction energy Eππ(r,θ,φ) = −εππ{1+ln[P(r,θ,φ)/Pmax]/|ln(Pmin/Pmax)|} for each of the three residue type pair F-F, Y-Y, or F-Y, where Pmax and Pmin are, respectively, the maximum and minimum non-zero values of P(r,θ,φ) among all the bins for a given pair. We further set Eππ = 0 for all r,θ,φ bins that received zero entry from the PDB data set. In this way, for εππ > 0, the present π-π potential is an attractive interaction that varies between Eππ = −εππ and 0 (Fig 7b). Here we use εππ = 1.5 for all three residue type pairs.
|
10.1371/journal.pgen.1006962 | A Becn1 mutation mediates hyperactive autophagic sequestration of amyloid oligomers and improved cognition in Alzheimer's disease | Impairment of the autophagy pathway has been observed during the pathogenesis of Alzheimer’s disease (AD), a neurodegenerative disorder characterized by abnormal deposition of extracellular and intracellular amyloid β (Aβ) peptides. Yet the role of autophagy in Aβ production and AD progression is complex. To study whether increased basal autophagy plays a beneficial role in Aβ clearance and cognitive improvement, we developed a novel genetic model to hyperactivate autophagy in vivo. We found that knock-in of a point mutation F121A in the essential autophagy gene Beclin 1/Becn1 in mice significantly reduces the interaction of BECN1 with its inhibitor BCL2, and thus leads to constitutively active autophagy even under non-autophagy-inducing conditions in multiple tissues, including brain. Becn1F121A-mediated autophagy hyperactivation significantly decreases amyloid accumulation, prevents cognitive decline, and restores survival in AD mouse models. Using an immunoisolation method, we found biochemically that Aβ oligomers are autophagic substrates and sequestered inside autophagosomes in the brain of autophagy-hyperactive AD mice. In addition to genetic activation of autophagy by Becn1 gain-of-function, we also found that ML246, a small-molecule autophagy inducer, as well as voluntary exercise, a physiological autophagy inducer, exert similar Becn1-dependent protective effects on Aβ removal and memory in AD mice. Taken together, these results demonstrate that genetically disrupting BECN1-BCL2 binding hyperactivates autophagy in vivo, which sequestrates amyloid oligomers and prevents AD progression. The study establishes new approaches to activate autophagy in the brain, and reveals the important function of Becn1-mediated autophagy hyperactivation in the prevention of AD.
| Accumulation of amyloid β peptides (Aβ) is a major cause of Alzheimer’s disease (AD). Although many efforts have been made, no effective therapies are available to cure AD. Autophagy is a stress-induced pathway nerve cells use to dispose damaged structures, and may be a strategy to eliminate Aβ aggregation. However, direct evidence of autophagic disposal of Aβ is lacking. Here we described a new AD mouse model that shows high autophagy activity even under non-autophagy-inducing conditions. In these mice, we engineered a single mutation into a key autophagy gene Becn1, which disrupts an inhibitory binding and leads to constitutively active autophagy in brain. Compared to regular AD mice, the autophagy-hyperactive AD mice are protected from amyloid accumulation, memory deficits and high mortality. We also isolated autophagic vesicle from these mice by an antibody, and found that Aβ oligomers are incorporated inside the vesicles. Thus, we obtained direct evidence for the first time that oligomerized amyloids are substrates of autophagy. These findings are important, as we demonstrated the reversal of AD progression by modulating the activity of a single autophagy gene. Our findings also revealed the therapeutic potential of brain-permeable autophagy-inducing chemicals in the prevention of AD by reducing intracellular amyloids.
| Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by protein aggregation and deposition, leading to progressive neuronal loss and cognitive decline among elderly populations [1]. Amyloid plaques and neurofibrillary tangles are the two primary hallmarks of AD pathology, and aging is a major known risk factor of the disease [2, 3]. Amyloid plaques are formed by amyloid-β (Aβ) peptides, generated by sequential enzymatic cleavages of amyloid precursor protein (APP) at the plasma membrane [4, 5]. Besides the well-recognized extracellular deposition of Aβ, recent studies also revealed the accumulation of intracellular pools of Aβ in AD brain. Intracellular Aβ can be generated at the trans-Golgi network and endoplasmic reticulum as part of the secretory pathway, or be reuptaken by neurons and glial cells from the secreted extracellular pools [6, 7]. Although many therapeutic efforts have been made to eliminate Aβ aggregation and deposition at either the synthesis or the degradation stage, no effective therapies are available so far to cure AD, and the mechanism driving the neurodegenerative progression remains unclear [8].
Autophagy is an evolutionarily conserved lysosomal catabolic pathway regulated by autophagy-related (ATG) proteins [9, 10]. Autophagy is induced by stress conditions such as hypoxia, starvation or oxidative stress [11, 12]; upon autophagy induction, autophagosomes sequester cytoplasmic components and fuse with lysosomes to generate autolysosomes, in which degradation of the autophagic cargos occurs [13, 14]. Although many studies have reported the roles of autophagy in the elimination of wasteful components, including protein aggregates, the relationship between autophagy and AD progression is complex. Several lines of evidence suggest an impairment of the autophagy pathway in the pathogenesis of AD. Brain from AD patients shows an abnormal accumulation of autophagic vacuoles and a reduction in the level of Beclin 1/BECN1, an essential autophagy protein and ortholog of ATG6 [15, 16]. However, direct evidence of autophagosome-mediated degradation of Aβ or APP in brain is lacking. Paradoxically, autophagy has been reported to promote, rather than reduce, the production of Aβ. Knockout (KO) of an essential autophagy gene Atg7 specifically in forebrain excitatory neurons of AD mice decreases extracellular amyloid plaque formation, which is due to reduced processing and secretion of Aβ; however, these Atg7 KO mice have exacerbated memory deficits [17], suggesting that the intracellular level of amyloids, which may be regulated by autophagy, may play a key role in cognitive impairment in AD. It is also under debate whether the level of the precursor protein APP is directly regulated by autophagy in either rodent brain or primary neurons [16–19]. On the other hand, enhancing lysosomal degradation capacity by genetic deletion of Cystatin B, a suppressor of lysosomal cysteine proteases, or use of autophagy-inducing chemicals such as a phytochemical Rg2 or the mTOR inhibitor rapamycin, reduces amyloid burden and memory deficit in mouse models of AD [20, 21, 22]. However, the mechanism of these compounds remains enigmatic. In addition, although knockout of autophagy genes leads to neurodegeneration [15, 23, 24], it is unknown whether physiologically increased basal autophagy prevents neurotoxicity of Aβ and has beneficial effects in protecting against Alzheimer’s-like diseases.
Thus, to directly assess the function of physiological enhancement of autophagy in vivo, we generated and characterized a unique mouse model of constitutively active autophagy caused by a single knockin mutation (F121A) in Becn1. We crossed these autophagy-hyperactive mice with the 5XFAD transgenic AD mice, which overexpress a combination of 5 familial Alzheimer’s disease (FAD) mutations in human APP and human PS1 (presenilin 1) proteins and show early amyloid deposition beginning at 2 months of age and cognitive decline at 6 months of age [25]. We demonstrated that elevated basal autophagy targets Aβ oligomers, and significantly reduces the accumulation of Aβ, but not APP. Genetic hyperactivation of autophagy also ameliorates neuronal dysfunction and enhances survival in AD mice. In addition to genetic activation of autophagy, we also found that autophagy hyperactivation either pharmacologically by a novel compound ML246 or physiologically by voluntary exercise protects AD mice from amyloid deposition and memory loss. Overall, this study provides the first evidence that hyperactive autophagy caused by a single mutation in Becn1 sequesters amyloids and restores memory in AD, and also establishes the first genetic model of constitutively active autophagy as a useful in vivo tool to study autophagy in different diseases.
To study how autophagy physiologically regulates the progression of Alzheimer’s disease (AD), we generated a new knock-in mouse model with hyperactive autophagy, by genetically disrupting the nutrient-regulated interaction between BECN1 and its inhibitor BCL2 (Fig 1A). Reversible BECN1-BCL2 binding is an important regulatory mechanism of autophagy induction [26]. When nutrients are abundant, BECN1 is bound and inhibited by BCL2, an anti-apoptotic and anti-autophagy protein. In response to stress such as starvation, BECN1 is released from the inhibitory binding of BCL2 for autophagy function [27, 28]. The BCL2 binding site in human BECN1 is reported as F123 [27]. We found that F121 in the BH3 domain of mouse BECN1 is the corresponding conserved residue of human F123. Thus, we proposed that mutating the residue F121 (TTT) to an alanine (A, GCT) disrupts BECN1-BCL2 binding and leads to constitutive activation of BECN1 and autophagy in mice (Fig 1A). We then generated a global knock-in mouse line (Becn1FA/FA) (S1A–S1C Fig), and found that the homozygous Becn1FA/FA mice are viable, fertile, of normal size and weight, and display normal histology in major organs under normal housing conditions. The BECN1 protein expression level in Becn1FA/FA mice is also comparable to that in WT mice in multiple major organs, including brain, heart, skeletal muscle, liver and pancreas (S2 Fig). Co-immunoprecipitation analysis showed that in Becn1FA/FA mice, there is much less interaction between BECN1 and BCL2 in both skeletal muscle and brain than in WT mice (Fig 1B), suggesting that the F121A mutation significantly weakens BCL2 binding to BECN1 in vivo.
To determine whether these mice have hyperactive autophagy, we crossed them with the GFP-LC3 autophagy reporter mice [29]. Upon autophagy induction, diffusely distributed autophagosome marker protein LC3 (LC3-I) is conjugated to phosphatidylethanolamine to form lipidated LC3 (LC3-II), which specifically associates with autophagosomal membranes and can be resolved by western blot or visualized as fluorescent puncta. We found that under non-autophagy-inducing conditions (fed and resting), Becn1FA/FA knock-in mice exhibit a higher number of GFP-LC3 puncta (autophagosomes) in both skeletal muscle (Fig 2A) and brain (Fig 2B) than wild-type (WT) mice, which reaches similar levels under autophagy-inducing conditions (90-min treadmill exercise or 48-h starvation). These data suggest that Becn1FA/FA mice show high basal autophagy. To determine that the increment of autophagosomes in Becn1FA/FA mice is due to elevated autophagic flux, rather than a block in autophagosome degradation, we analyzed the autophagy flux by inhibiting lysosomal degradation using the lysosomal inhibitor chloroquine. In skeletal muscle of Becn1FA/FA mice, chloroquine injection led to more accumulation of LC3 and GFP-LC3 puncta compared to WT mice, measured by western blot analyses and microscopy, respectively (Fig 2C and S3 Fig). In addition, compared to WT mice, Becn1FA/FA mice showed a lower level of p62, an autophagy cargo protein, in skeletal muscle, which was rescued by chloroquine treatment (Fig 2C). These data suggest that the Becn1F121A mutation in mice leads to higher autophagic degradation of LC3 and p62. Altogether, we conclude that mutating F121 to A disrupts BECN1-BCL2 binding and constitutively activates autophagy in mice, thus providing a novel mouse model with hyperactive autophagy as a useful tool to analyze the physiological effects of autophagy upregulation in vivo.
To determine the effects of autophagy activation on AD, we crossed the Becn1F121A mice with the 5XFAD mice, an amyloid mouse model used in AD research. 5XFAD mice demonstrate early and aggressive phenotypes of intraneuronal Aβ42 aggregates, β-amyloid plaques and neurodegeneration, and represent a good model for our study. We analyzed the amyloid burden in the resulting 5XFAD Becn1FA/FA mice by dot blot assays, ELISA and microscopy at the age of 6 months (Fig 3A–3D). The dot blot assay has been previously validated to measure levels of Aβ42 in APP transgenic mouse [30]. We found that 5XFAD Becn1FA/FA mice show lower levels of both soluble and insoluble Aβ42 in the brain than the 5XFAD mice by dot blot assays (Fig 3A and 3B), whereas expression of the precursor APP remained unaffected (S4A Fig), suggesting that Becn1FA/FA-mediated hyperactive autophagy downregulates the levels of Aβ42, but not of APP. This is also confirmed by ELISA analyses on the level of total brain Aβ42 (Fig 3C). Furthermore, staining of amyloid plaques by Thioflavin S (Fig 3D) or Aβ42 antibody (S7A Fig) showed that there is a significant reduction of amyloid plaques in the cortex and a trend of reduction in the hippocampus of 5XFAD Becn1FA/FA mice.
Importantly, Becn1F121A-induced reduction in Aβ42 is dependent on autophagy but not other pathways that regulate amyloid transport. We found that short-term (7-day) treatment of 5XFAD Becn1F121A mice with SBI-0206965, an autophagy inhibitor blocking the kinase activity of an essential upstream kinase ULK1 [31], abolished the reduction in brain Aβ42 levels by dot blot assays (Fig 3E). Similarly, in HEK293 cells stably expressing APP and Becn1F121A, siRNA knockdown of the essential autophagy gene ATG7 significantly increased the level of intracellular Aβ42 by western blot analysis (S5A Fig). The reduced Aβ42 level is not due to alterations in APP trafficking in Becn1F121A mice, as immunofluorescence microscopy showed no detectable difference in the amount of APP colocalized with Rab5+ early endosomes or Rab7+ late endosomes in primary cortical neurons isolated from PDAPP Becn1+/+ mice and PDAPP Becn1FA/FA mice (PDAPP mice is another amyloid model as described below) (S5B Fig). Furthermore, we also biochemically analyzed APP internalization and trafficking, by biotin protection assays using HEK293 cells stably expressing APP and WT Becn1 or Becn1F121A (the endogenous BECN1 was deleted by CRISPR/Cas9) (S5C Fig). After inducing endocytic trafficking of cell surface APP by incubating the cells at 37°C for 5 min or 15 min, we observed a similar level of APP endocytosis (represented by biotinylated APP that is protected from glutathione stripping) in cells expressing WT Becn1 and expressing Becn1F121A. Thus, altogether, these results demonstrated that Becn1F121A does not affect APP trafficking. In addition, 5XFAD Becn1F121A mice expressed a similar level of amyloid receptors in the brain that contribute to the clearance of Aβ, such as LDLR (Low-density lipoprotein receptor) and LRP1 (LDLR-related protein 1), compared to 5XFAD mice expressing WT Becn1 (S5D Fig). Thus, altogether, we conclude that the Becn1F121A knockin mutation reduces amyloid accumulation, and the effect of Becn1F121A on Aβ metabolism is mediated by the hyperactive autophagy activity.
Next, to analyze memory function, we performed Morris water maze tests on WT mice and AD mice with normal or high autophagy. During the visible platform training, all 3 groups of mice showed no significant difference in either escape latency or distance (Fig 3F), suggesting that there was no visual or swimming abnormality among all groups. In contrast, during the hidden platform trials, 5XFAD mice expressing WT Becn1 showed apparent deficiency in memorizing the platform location, while 5XFAD Becn1FA/FA mice had significantly improved performance day by day in both escape latency and distance, similar to WT mice (Fig 3F). These data suggest that the memory impairment caused by Aβ accumulation is ameliorated by the Becn1 F121A mutation. Overall, we conclude that genetic stimulation of basal autophagy mediated by Becn1F121A reduces Aβ42 levels and plaque formation in mouse brain, and improves memory capacity that is impaired by amyloid aggregation in AD.
To fully analyze the function of the Becn1F121A allele in AD, we used another amyloid mouse model, known as PDAPP mice. These mice carry a V717F (Indiana) mutation in APP [32], and exhibit extracellular amyloid deposition starting at 6–9 months of age. The PDAPP mice have been shown to display an increased mortality rate compared to other AD lines [33, 34]. Similar to previous reports, we found that PDAPP mice have higher early mortality than WT mice starting at 2 months of age (Fig 4). We crossed PDAPP mice with either the autophagy-hyperactive Becn1FA/FA mice, or autophagy-deficient Bcl2AAA mice. Bcl2AAA mice contain 3 knock-in alanine mutations (T69A, S70A and S84A) at the phosphorylation residues of BCL2, which block BCL2 phosphorylation and BECN1 release from BCL2 binding; thus, they are opposite to the Becn1F121A mice and show defective autophagy [35]. Notably, homozygous expression of the Becn1F121A mutation decreased mortality in PDAPP mice, while the autophagy-deficient PDAPP Bcl2AAA mice showed a trend of exacerbated mortality compared to the PDAPP mice with normal autophagy (Fig 4). These data suggest a positive impact of hyperactive autophagy mediated by Becn1F121A on the survival of PDAPP Alzheimer’s mice.
To directly address whether intracellular amyloids are efficient autophagic cargos, and degraded by the autophagy machinery upon autophagy hyperactivation, we developed a method to immunoisolate intact autophagosomes from the cortex of 5XFAD Becn1FA/FA mice expressing the autophagosome marker GFP-LC3. After sequential centrifugation and immunoprecipitation by anti-GFP antibody and magnetic beads, the purity of autophagosomes was validated by co-isolation of a known autophagy cargo p62 but not a cytosolic enzyme GAPDH (Fig 5). We found that Aβ42 oligomers, including trimers, pentamers and higher-molecular weight fibrils or fibril intermediates (of size between 100 kD and 250 kD), but not monomers, are co-immunoprecipitated and concentrated with autophagosomes from autophagy-hyperactive mice (Fig 5). Thus, these data suggest biochemically that intracellular Aβ oligomers are cargos of autophagy, and are sequestered and cleared by hyperactive autophagy in brain.
In addition to Becn1F121A-mediated genetic activation of autophagy, we decided to further study whether stimulating autophagy pharmacologically is also protective against neurodegenerative progression. We recently identified a brain-penetrable autophagy-inducing small molecule ML246 (metarrestin)[36] (Fig 6A), and analyzed its effects on the clearance of aggregate-prone proteins in vitro and in vivo. For in vitro analyses, we utilized the HEK293 cell line stably expressing APP (APP-HEK293), in which the produced Aβ molecules are efficiently secreted, to study the effect of ML246 on amyloid metabolism. Via dot blot assays, we found that ML246 treatment for 24 h significantly reduced the level of secreted Aβ in the conditioned media (Fig 6B). In addition, cultured WT primary cortical neurons treated with the conditioned media from ML246-treated APP-HEK293 cells underwent a lower level of apoptotic cell death than those treated with media from vehicle-treated APP-HEK293 cells (Fig 6C). These results demonstrate that ML246 reduces amyloid production and secretion in vitro. Importantly, siRNA knockdown of the essential autophagy gene ATG7 in APP-HEK293 cells reversed the ML246-mediated reduction of both secreted Aβ42 (Fig 6B) and apoptotic neuronal death (Fig 6C), suggesting that the effect of ML246 in amyloid metabolism is autophagy-dependent.
In addition to the amyloid cell model, we also found that ML246 promotes the removal of intracellular aggregates formed by polyglutamine (polyQ)-expansion proteins. We used HeLa cell lines stably expressing tetracycline-repressible expanded polyQ-repeat protein HTT (huntingtin) as a model, HTT65Q and HTT103Q [37]. In contrast to the HTT protein with the normal number of glutamine repeats (HTT25Q), HTT65Q and HTT103Q formed insoluble polyQ aggregates larger than 0.2-μm in diameter, which can be detected by filter trap assay. We discovered that the accumulation of both HTT65Q and HTT103Q aggregates is decreased after ML246 treatment for 24 h, whereas knockdown of ATG7 prevents this reduction (S6A Fig), suggesting that ML246 reduces intracellular protein aggregation, and this effect is dependent on the autophagy activity. Fluorescence imaging further confirmed that ML246 administration decreased the number of cells positive for HTT aggregates, which is also in an ATG7-dependent manner (S6B Fig). Thus, these data indicate that the autophagy pathway stimulated by ML246 promotes the clearance of aggregate-prone proteins (including both amyloid and polyQ expansion proteins) in vitro, and ML246 can be used as a candidate compound for in vivo analyses in AD mouse models.
Accordingly, we investigated the function of ML246-induced autophagy in amyloid accumulation and cognitive function in 5XFAD mice. Via dot blot assays, we found that compared to the ones treated with vehicle, 6-month old 5XFAD mice treated with ML246 for 5 weeks showed decreased levels of both soluble (Fig 6D) and insoluble Aβ42 in brain (Fig 6E). The expression of the precursor APP was not affected (S4B Fig), supporting the hypothesis that the level of Aβ42, but not APP, is regulated by autophagy. Notably, the effect of ML246 was abolished in the autophagy-deficient 5XFAD Becn1+/- KO mice (Fig 6D and 6E), further supporting that ML246-induced reduction of Aβ42 in vivo is autophagy-dependent.
Moreover, in addition to pharmacological approaches, physical exercise has recently been demonstrated as a fast and robust physiological method to induce autophagy in various tissues, including brain [38]. Intriguingly, previous studies indicated that aerobic exercise decreases amyloid load in AD mouse models [39–42], and is also associated with a lower risk of cognitive decline among elderly populations [43–46]. Thus, we hypothesized that exercise-induced autophagy may represent a cellular mechanism underlying the neuroprotective effects of exercise in AD brain. To test this hypothesis, we housed 2-month old 5XFAD mice individually with access to a running wheel for 16 weeks. Through dot blot assays on brain lysates, we found that 5XFAD mice subject to 16 weeks of voluntary running have significantly lower levels of both soluble (Fig 7A) and insoluble (Fig 7B) Aβ42 in brain than those housed under resting conditions (without running wheels), suggesting that physical exercise decreases the amyloid burden in AD mouse brain. Fluorescence microscopy also shows a significant reduction of amyloid plaques after exercise training, and a trend of plaque reduction after ML246 treatment, stained by thioflavin S or Aβ42 antibody in brain of 5XFAD mice, especially in the cerebral cortex (Fig 7C and S7A Fig). Similar to ML246 treatment, APP expression is not affected by exercise (S4C Fig). In comparison, exercise failed to reduce amyloid accumulation in the autophagy-deficient 5XFAD Becn1+/- KO mice (Fig 7A and 7B), suggesting that the autophagy pathway is required for the effects of exercise on amyloid accumulation.
Finally, to analyze whether the autophagy-inducing compound ML246 has the potential to improve the cognitive function of Alzheimer’s mice, we performed Morris water maze tests on 6-month old 5XFAD mice injected with ML246 daily for 5 weeks. We found that ML246 treatment, as well as 16-week voluntary exercise, improved the performance of 5XFAD mice during the hidden platform trials, compared to the vehicle-treated resting mice at the same age (Fig 7D). These data suggest that similar to exercise, the autophagy inducer ML246 ameliorates memory impairment in AD mice.
The role of autophagy in amyloid production and clearance has been unclear. In this study, we generated a mouse model with hyperactive autophagy by knocking-in a point mutation F121A to Beclin 1/Becn1, and found that Becn1F121A-mediated autophagy hyperactivation reduces brain amyloid accumulation, ameliorates cognitive deficits, and improves survival rates in Alzheimer’s mouse models.
BECN1 is a core component of the type III phosphatidylinositol-3-kinase (PI3K) complex, and is key for the initiation of autophagosome biogenesis [26]. Lentiviral overexpression of Becn1 has been shown to reduce APP levels in cultured CHO cells or decrease amyloid deposition in AD mouse brain [15, 16]. Yet it is unclear whether Becn1 overexpression represents a physiological method for autophagy activation. Thus, we designed a strategy to constitutively activate autophagy in vivo by preventing BECN1 from binding with its inhibitor BCL2. Under nutrient rich conditions, BECN1 is bound and inhibited by BCL2; whereas in the presence of stress such as nutrient starvation and exercise, BCL2 is phosphorylated and released from BECN1, which activates autophagy [28] and represents a physiological regulatory mechanism of the function of Becn1 in autophagy. In our new knock-in mouse model, the introduction of the F121A mutation in Becn1 (F121A) disrupts the BCL2 binding site, resulting in the constitutive activation of BECN1 in autophagy that is no longer regulated by stress. In skeletal muscle and brain of the Becn1F121A mice, the autophagy levels under basal conditions are as high as those obtained after physical exercise or starvation in WT mice. Thus, we consequently crossed these mice with amyloid mouse models, including 5XFAD and PDAPP mice, to study the function of Becn1-mediated autophagy in AD. Given the roles of autophagy in a broad spectrum of diseases, this new mouse model can be a useful genetic tool to study the physiological effects of autophagy hyperactivation in multiple diseases.
We found that Becn1F121A-mediated autophagy hyperactivation decreases Aβ levels and improves memory in 5XFAD mice. Yet how the autophagy pathway downregulates intracellular Aβ still remains mysterious. Besides the plasma membrane, APP also localizes to the secretory pathway (such as the trans-Golgi network and endoplasmic reticulum), endosomes, lysosomes and mitochondria. We do not know whether it is the intracellular Aβ, or the extracellular secreted pool taken back up by cells, that is regulated by autophagy in Becn1FA/FA mice. Several studies also suggest that BECN1 promotes internalization and lysosomal trafficking of the precursor protein APP. In cultured neuronal and HEK293 cell lines, BECN1 has been reported to promote endocytosis and endolysosomal and autolysosomal proteolysis of plasma membrane APP [47]. The adaptor protein AP2 seems to interact with LC3 to target APP to autophagosomes [19]. However, whether APP trafficking and degradation depends on other key components in the autophagy machinery is not known, and whether the process of autophagosome-mediated APP degradation occurs in AD mouse brain or neurons is still under debate [18]. Our data argue against a role of autophagy in regulating the levels of APP, since we found that Becn1F121A does not alter the level, internalization, or trafficking of APP in mouse brain, primary cortical neurons, or cell lines (S4 and S5B and S5C Figs), suggesting that the effect of Becn1F121A on amyloid metabolism is not through the regulation of APP.
The role of autophagy in the regulation of Aβ is more complex. On one hand, Becn1 has been shown to be important for the phagocytosis and autophagic degradation of extracellular Aβ by cultured microglial cells [48, 49], and Becn1-deficient mice showed impaired Aβ clearance [49], which is consistent with our findings. On the other hand, autophagy is suggested to facilitate Aβ processing and secretion from neurons, using neuroglioma cell lines [50] and tissue-specific Atg7 KO mice in excitatory forebrain neurons [17]. Thus, our autophagy-hyperactive AD mouse model is useful to assess the overall readout of autophagy activation on Aβ levels in vivo (Figs 3 and 4). Using this model system, we biochemically detected Aβ oligomers in purified intact autophagosomes (Fig 5), suggesting that autophagy plays a direct role in brain amyloid clearance. We propose a model in which autophagic degradation of Aβ occurs in both neurons and glial cells, where neuronal autophagy mainly degrades de-novo processed Aβ, whereas autophagy in glia removes Aβ reuptaken from the extracellular space (S7B Fig). As future directions, it will be interesting to determine whether the endocytic reuptake and trafficking machinery in neurons or glia is required for autophagic degradation of Aβ, and to discover what receptors are involved in the autophagosomal recognition of Aβ42-containing secretory or endocytic vesicles.
Furthermore, besides hyperactivating autophagy by genetic factors, we analyzed the effects of ML246, a novel autophagy-inducing compound that can pass the blood-brain barrier [36], on Aβ accumulation and cognition in AD mice. Pharmacological strategies to modulate autophagy have been recently proposed in the prevention of neurodegenerative diseases [51]. Most autophagy inducers that have been tested are based on inhibiting the autophagy suppressor mTOR, such as the well-known mTOR inhibitor rapamycin [21, 52, 53], which seems to be effective to decrease Aβ levels and prevent cognitive impairment in AD mice when used at early stages prior to the formation of extracellular plaques. Here we showed that ML246 is able to decrease protein aggregates in cultured cells, and reduce Aβ levels and ameliorate memory deficit in 5XFAD mice, and notably, we started compound treatment at the age of 4–5 months when amyloid deposition has already been documented in this AD mouse model [25]. Thus, ML246 is a new autophagy activator of neuroprotective function and potential use in AD treatment, although the signaling pathways and mechanisms by which ML246 induces autophagy are yet to be determined.
In addition to genetic and compound inducers of autophagy, we also studied whether activating autophagy by physiological methods prevents AD. Starvation and exercise are the best-known physiological inducers of autophagy in vivo [29, 35, 54]. Interestingly, although starvation induces detectable formation of autophagosomes in neurons of 3-month 5XFAD mice after 48 h [55, 56], it seems ineffective in removing intra-neuronal or extracellular Aβ [56], likely due to insufficient degradation of Aβ-containing autolysosomes after short-term starvation. In comparison, exercise, either forced exercise by treadmill or voluntary exercise by running wheel, has been recently shown to increase the autophagy flux in various tissues, including skeletal muscle and cerebral cortex in mice [35, 38, 54, 57]. Thus, we investigated the effects of physical exercise in AD, and demonstrated that 4 months of voluntary running exerted positive effects on animal behavior and amyloid pathology in brain of 5XFAD mice. It should be noted that we started voluntary exercise at the age of 2 months prior to any detectable cognitive impairment. Physical exercise has previously been suggested to play a beneficial role against cognitive decline in AD [39, 41, 42], but the molecular mechanism remains unknown. We showed for the first time that compared with AD mice with normal autophagy activity, exercise is not able to reduce amyloid deposition in brain of autophagy-deficient AD mice, suggesting that exercise-induced autophagy may be an important mechanism mediating some of the beneficial effects of exercise on AD [35, 54], although exercise may affect other pathways that also contribute to the exercise-mediated neuroprotective effects. Intriguingly, treadmill exercise does not seem to be as effective as voluntary wheel running to prevent neurodegeneration [40]. It is likely that the stress associated with forced running on the treadmill exerts detrimental effects on animal behavior and disease pathogenesis. Finally, considering that one key problem in AD is the late diagnosis of the disease that significantly reduces the effectiveness of subsequent treatments [58, 59], voluntary exercise should be considered as an important component in modern lifestyle to effectively induce autophagy and prevent cognitive decline as a non-pharmacological intervention.
In summary, in this study we developed 3 new strategies to potently activate autophagy in the brain, genetic (by the Becn1F121A mutation), pharmacological (by ML246), and physiological (by voluntary exercise). Using the different approaches, we provided evidence that autophagy induction ameliorates amyloid pathology and reduces cognitive deficits in 5XFAD mice. Our data revealed the potential of autophagy stimulation in lowering toxic aggregate-prone proteins and improving neuronal functions for the treatment of AD.
All animal experiments have been approved by the Northwestern University Institutional Animal Care and Use Committee (IACUC) (Protocol number: IS00004749). All mice were housed on a 12-h light/dark cycle, and male mice were used for behavioral analyses. All mice were in the C57BL/6 background except PDAPP mice. The PDAPP mice were generated in a C57BL/6 and DBA2 mixed genetic background and have been backcrossed with C57BL/6 mice for 8 generations prior to analyses. GFP-LC3 transgenic, Becn1+/- KO, Bcl2AAA, PDAPP and 5XFAD mice have been previously described [25, 29, 32, 35, 60]. For the construction of a mouse strain with the F121A knock-in allele in Becn1, BAC clones (Incyte) were screened for the presence of Becn1. The Becn1 BAC clone was subcloned into the pVB vector and the F121 (TTT) in exon 6 of Becn1 was replaced by A (GCT). A neomycin resistance marker flanked by LoxP sites was inserted between exons 7 and 8. The resulting targeting construct, pVBKI-Becn1, was linearized by I-CeuI digestion and electroporated into 129 Sv/J × C57BL/6J hybrid ES cells, and 36 h later, clones were selected with neomycin, and screened by Southern blot analysis with the probes indicated in S1 Fig. DNA was digested with HindIII, and electrophoretically separated on a 0.8% agarose gel. After transfer to a nylon membrane, the digested DNA was hybridized with a probe targeted against the 5’ region (Probe 1):
TCCAGTGATGATGGTGGTGCTGATAATAATAGGGATGTTTTCATTACCAAAGATAGATGTTGTAGCTTGATTTTCTTTTGTGGGGAGCAGGTCATTGTCAAGTAGAAGTTACTGACTTGGGAGAGGATCCCAAGGGACCCTAGTACAAAATAGAGAAAACGGATGGGTGGAAAGGGAAAGAAGCCTAGGAGGGAGACATGGTCACACACCAGTGGCACAGCATCCTGGGGAAAGCGCTGGCCTCATCCCTGAGATTTACCTTGCCTGAGCAATACGGGAGGATTTATCCGAGTGACTGCTGTCACTGGGAAAAGCGAACCTTAAGTGGGTTGGGGGCTGTTAATTCTAGCATGCAAGGCCAGAGAAAACCTGCAAAGAAGCAAAAGAGGCAGGCAGCTGAAGCCAGTGTGTTCAAAATGTTGAACATAAATGTTCTAGAACTGTTGATGATAGGCAGTTCTGGTACTGACAGGCCCACCGATTTCTT.
The positive clones were further confirmed by Southern blot analyses using a 3’ probe. DNA was digested with HindIII, and electrophoretically separated on a 0.8% agarose gel. After transfer to a nylon membrane, the digested DNA was hybridized with a probe targeted against the 3’ region (Probe 2):
TGCCTTTCTCTCTGCTCTGTGAGTTAGGGGTGCCTAGGCAGACAGTGAAGAGTACTGTAGCCTTCACTCCCTCCTGTGTGGGTGTGTCCTCTCCTGTCCTGTACTCTGCCATGACAATGAGGCTCTTGTGACAGCCTTTGATTTTAGGCTTTCAAGCAAATCCAAAATACACTAGCGGTAATTCTTTGCCAGGCGTTCTTTATTAGATAAAGTGACGTGAATGGTCTCATGATCAAGTCCCTGCCCATTTGCCTGAACTGACTTAGGTTGGCTCTGTTACTAATGAGCTCTGCTATGTCCACCTGCAGGATGGACGTGGAGAAAGGCAAGATTGAAGACACTGGAGGCAGTGGCGGCTCCTATTCCATCAAAACCCAGTTTAACTCGGAGGAGCAGTGGACAAAAGCGCTCAAGTTCATGCTGACC.
The positive knock-in clones were tested for normal karyotype and used to inject blastocysts from C57BL/6J donors. Mice with germline transmission were bred to mice expressing Cre from the CAG promoter (gift of Eric Olson, UT Southwestern Medical Center) to remove the neomycin cassette. Offspring were genotyped for the presence of the knock-in allele by PCR with the following primers: 5' primer: GGCAGTAGTATAATGTCTGCTCCAG; knock-in 3' primer: TCTAATTCCATCAGAAGCTGACTCT; wild-type 3' primer: TGGGTCTCTCATTGCATTTATTTAT. Using this scheme, the knock-in Becn1F121A allele was identified by a PCR product of 650 bp, and the wild-type allele was identified by a PCR fragment of 320 bp. Becn1F121A mice were backcrossed for more than 10 generations to C57BL/6J mice (Jackson Laboratories).
For the generation of 5XFAD; Becn1FA/FA mice, heterozygous 5XFAD transgenic mice were bred to homozygous Becn1 knock-in (Becn1FA/FA) mice to obtain 5XFAD; Becn1FA/+ mice, which were bred to the Becn1FA/FA or Becn1+/+ littermate mice to produce the 5XFAD; Becn1FA/FA and 5XFAD; Becn1+/+ offspring. For the generation of PDAPP; Becn1FA/FA mice, heterozygous PDAPP transgenic mice were bred to homozygous Becn1 knock-in (Becn1FA/FA) mice, and the offspring were bred to the Becn1FA/FA or Becn1+/+ littermate mice to produce the PDAPP; Becn1FA/FA and PDAPP; Becn1+/+ mice. Similarly, PDAPP transgenic mice were crossed with homozygous Bcl2AAA mice to produce the PDAPP; Bcl2AAA mice.
HeLa cell lines were obtained from ATCC, and HeLa cells conditionally expressing CFP-tagged Huntingtin with polyQ repeats were from A. Yamamoto (Columbia University). The HEK293 cell line stably expressing APP (APP-HEK293 cells) was generated by recombinant adenovirus encoding WT human APP under the control of the CMV promoter. Cells were cultured in DMEM medium (Gibco, 11995073) supplemented with 10% FBS. Tetracycline-free FBS was used for HeLa cells stably expressing Huntingtin (Takara Bio USA, 631107), and regular FBS was used for all other cells (HyClone, SH30070.03HI).
Cortical neurons were isolated from E16.5 mouse embryos via dissociation in 0.25% trypsin at 37°C. Neurons from each single embryo were separately plated at the density of 105 cells per well on culture slides (4 well-culture slide) coated with 100 μg/ml poly-L-lysine in borate buffer (50 mM boric acid, 12.5 mM borax). Neurons were plated in neurobasal media (Gibco 21103–049) supplemented with 2% B-27 (Gibco 17504–044), 500 μM glutamine (Invitrogen), 10% horse serum and 2.5 μM glutamate. After 2 h, media was replaced with growth media (neurobasal media with 2% B-27 and 500 μM glutamine).
Mouse tissues (muscle and brain) were homogenized in lysis buffer containing 50 mM Tris (pH 7.5), 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, halt proteinase inhibitor cocktail (ThermoFisher Scientific), and halt phosphatase inhibitor cocktail (ThermoFisher Scientific), and subjected to immunoprecipitation with anti-BCL2 monoclonal antibody conjugated agarose beads (Santa Cruz Biotechnology 7382 AC). Eluates were separated by SDS-PAGE and detected by anti-BCL2-HRP antibody (C2 Santa Cruz Biotechnology, 1:500) and anti-BECN1 antibody (Santa Cruz Biotechnology, 1:200) using the ONE-HOUR Western Detection Kit (GenScript Corporation) following the manufacturer’s instruction.
Cortex samples from 12-week old 5XFAD; Becn1FA/FA; GFP-LC3 mice were dissected and homogenized in 1 ml cold lysis buffer pH 7.4 containing 250 mM sucrose, 1 mM EDTA, 10 mM HEPES, halt proteinase inhibitor cocktail (ThermoFisher Scientific), and halt phosphatase inhibitor cocktail (ThermoFisher Scientific), using a Dounce tissue grinder (Wheaton). The lysate was then passed 15 times through 27-gauge needle. GFP-based immunoisolation was performed using Dynabeads Protein G (ThermoFisher Scientific). The lysate was centrifuged at 1,000 x g for 10 min at 4°C. The post-nuclear supernatant fraction was centrifuged at 20,000 x g for 20 min and the supernatant fraction was discarded to remove residual cytosolic GFP-LC3 [61]. The pellet fraction was resuspended in 250 μl lysis buffer and was incubated for 2 hours at 4°C with 40 μl of Dynabeads, preincubated O/N with GFP-antibody (Sigma, G1544). The beads were then washed 4 times with wash buffer (150 mM NaCl, 250 mM sucrose, 1 mM EDTA, 10 mM HEPES) using the magnetic Separator DynaMagTM-2 (ThermoFisher Scientific). Immunoprecipitates were eluted with lysis buffer containing 1X sample buffer and analyzed by SDS-PAGE.
For in vivo use, ML246 was dissolved in a solvent containing 5% of NMP, 20% of PEG400 and 75% of 10% HP-β-CD in water, and injected intraperitoneally at the dosage of 5 mg/kg body weight for 5 weeks, 5 days per week. To measure the autophagy flux in vivo, chloroquine was dissolved in PBS and injected intraperitoneally at the dosage of 50 mg/kg, To inhibit autophagy in vivo, SBI-0206965 (Adooq Bioscience; A15795) was dissolved in PBS containing 50% DMSO, and injected intraperitoneally into mice at the dosage of 2 mg/kg body weight once per day for 7 days. Mice were sacrificed and tissues were collected 4 h after the last drug injection. For cell culture use, ML246 was dissolved in 100% DMSO and used at the concentration of 0.5 μM.
For acute exercise studies, 8-week old mice (wild-type and Becn1FA/FA mice crossed to GFP-LC3 transgenic mice) were acclimated and trained on a 10° uphill Exer 3/6 open treadmill (Columbus Instruments) for 2 days. On day 1 mice ran for 5 min at 8 m/min and on day 2 mice ran for 5 min at 8 m/min followed by another 5 min at 10 m/min. On day 3, mice were subjected to a single bout of running starting at the speed of 10 m/min. Forty minutes later, the treadmill speed was increased at a rate of 1m/min every 10 min for 30 min, and then increased at rate of 1 m/min every 5 min for 20 min, so that the mice ran for a total of 90 minutes of exercise and 1070 meters of running distance. For long-term exercise, 2-month old 5XFAD mice were single-housed in a cage containing a running wheel (11.4 cm diameter) for a total of 4 months. The running capacity of mice was monitored by an odometer connected to the wheel.
For animal behavior, 6-month old mice were tested. The Morris water maze test consists of two sections: the visible platform testing and hidden platform testing. During the tests, mice were placed in the water tank filled with opaque water (maintained at 25°C), with their heads facing toward the tank wall. In the visible platform section, a black platform extending 2 cm above the water level was used for these trials. For each trial, the platform was randomly positioned, and the mouse was placed in the tank at different start positions. The trial was stopped after the mouse found and climbed onto the platform, and the escape latency was recorded. The trial was stopped if the mouse did not climb onto the platform in 60 s, and the experimenter guided it to the platform. Mice were tested for 4 days with eight trials per day. In the hidden platform section, a transparent platform underneath the water level was used instead of the black one during all trials, mice were tested with a fixed platform location over 5 days period with six trials per day, and they were allowed to search the platform in 60 s. In the tests, two parameters were evaluated: the trail duration (s) and distance to the platform (m).
Snap frozen hemi-brain were homogenized in 800 μl phosphate-buffered saline (PBS; Sigma-Aldrich, D8537) with 1% Triton X-100 supplemented with halt proteinase inhibitor cocktail (ThermoFisher Scientific), and halt phosphatase inhibitor cocktail (ThermoFisher Scientific). Protein concentration was quantified using BCA Assay (Pierce). For Aβ42 dot blots, 10 mg/ml brain homogenates were extracted in guanidine buffer (5 M guanidine-HCl [GuHCl], 50 mM Tris HCl pH 8.0) overnight at room temperature. One μl of sample was spotted in triplicate on 0.2 μm nitrocellulose membrane, and dried for 1 h at 37°C. The membrane was stained with Ponceau S, and the dot blot signal on the membrane was detected by immunostaining with Aβ42 antibody (Invitrogen, 700254, 1:1000) and HRP-conjugated secondary antibody (Santa Cruz Biotechnology, sc2004, 1:2000). Aβ42 signals were normalized to the Ponceau S staining. To separate soluble and insoluble Aβ fractions, 10 mg/ml of the total homogenated brains were centrifuged at 14000 rpm, 4°C for 30 min. The supernatant (soluble fraction) was used directly for dot blot assays. The pellet (insoluble fraction) was extracted in guanidine buffer overnight at room temperature, and used in dot blot analyses. To measure Aβ levels in conditioned media of APP-HEK293 cells, media of 72-h cell culture was collected, mixed with 4X sample buffer (50 mM Tris-HCl pH6.8, 2% SDS, 10% glycerol, 1% β-mercaptoethanol, 12.5 mM EDTA, 0.02% bromophenol blue), and boiled at 95°C for 10 min. One μl of each sample was spotted on nitrocellulose membrane for dot blot analysis.
GuHCl extracted brain samples prepared in the same way as dot blot assays were diluted 1:1000, and ELISA analyses of Aβ42 were performed according to manufacturer’s instructions (Thermo Fisher Scientific, KHB3441).
Paraformaldehyde-fixed brain tissues were sectioned at 30 μm thickness. Free-floating sagittal sections were immnunostained with 1% thioflavin S (Sigma-Aldrich, 230456) for 20 minutes. Additional sections were immunostained with Aβ antibody (Invitrogen, 700254, 1:500) and Alexa Fluro 594 goat anti rabbit (ThermoFisher Scientific, A11012). Sections were mounted on slides with mounting medium containing DAPI (Vectashield) and then analyzed by fluorescence microscopy under the 10x objective.
Cortical neurons derived from PDAPP Becn1+/+ and PDAPP Becn1FA/FA embryos were grown on poly-L-lysine coated culture slides. Cells (9 DIV) were then fixed in 4% paraformaldehyde and permeabilized with 0.3% Triton X-100. Slides were blocked for 1 h in PBS containing 1% BSA and 2% normal goat serum and then incubated overnight at 4°C with primary antibodies: anti-APP (Biolegend; 803001) and anti-Rab5 (Cell Signaling Technology; 3547) or anti-Rab7 (Cell Signaling Technology; 9367). After washing, slides were incubated with species-specific Alexa-dye conjugated secondary antibodies for 1 h at room temperature. Slides were sealed with coverslip using mounting medium containing DAPI (Vectashield) and then analyzed by confocal microscopy. Confocal images were collected on Nikon A1 microscope using a 60x oil immersion objective lens and NIS Elements software. The Mander’s colocalization coefficient and the fluorescence intensity profile were generated using the NIS Element software.
For assessment of autophagy in vivo, 8-week old male WT and Becn1FA/FA mice crossed to GFP-LC3 mice were exercised for 90 minutes, or starved for 48 h, and then anaesthetized by isoflurane and perfused with 4% PFA. Brain and muscle samples were fixed in 4% PFA overnight, 15% sucrose for 4 h and 30% sucrose overnight before frozen sections were prepared. The number of GFP-LC3 puncta per unit area of tissue was quantified by fluorescence microscopy. Autophagy in vivo was also analyzed by western blot analysis of brain tissue extracts with antibodies against LC3 and p62/SQSTM1 (see below for details).
Cell or mouse muscle and brain extracts were prepared in lysis buffer containing 50 mM Tris (pH 7.4), 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, halt proteinase inhibitor cocktail (ThermoFisher Scientific) and halt phosphatase inhibitor cocktail (ThermoFisher Scientific), and subjected to western blot analysis with anti-LC3 (Novus Biologicals, NB100-2220), anti-SQSTM1 (Abnova, H00008878-M01), anti-Aβ42 (Invitrogen; 700254), anti-APP (Biolegend; 803001), HRP-conjugated GFP antibody (Santa Cruz Biotechnology, sc9996), anti-HA (Cell Signaling Technology, C29F4), anti-ATG7 (Sigma Aldrich, A2856), anti-LDLR (Abcam, ab52818), anti-LRP1 (Abcam, ab92544), and anti-ACTB/β-actin-HRP (Santa Cruz Biotechnology, sc47778 HRP) antibodies. The band intensity was analyzed using the ImageJ software.
HeLa cells stably expressing CFP-HTT25Q, CFP-HTT65Q and CFP-HTT103Q were treated with 100 ng/ml tetracycline (IBI Scientific, IB02200) for 48 h, or 0.5 μM ML246 for 24 h with control or ATG7 siRNA (GE Dharmacon ON-TARGETplus control or ATG7 SMARTpool siRNA) for 48 h. Cells were then collected and lysed in lysis buffer containing 0.5% NP-40 at 4°C for 30 min. After centrifugation, the pellet was digested with 0.5 mg/ml DNaseI (in 20 mM Tris-Cl, pH 8.0) for 1 h at 37°C, and dissolved into lysates containing insoluble aggregates by 2% SDS, 50 mM DTT and 20 mM EDTA. The lysates were then added onto 0.2 μm nitrocellulose membrane that was pre-equilibrated with 2% SDS/TBS for 30 min, and were filtered through the membrane by gentle vacuum using the Bio-Dot SF microfiltration apparatus (Bio-Rad). The signal was detected by immunostaining with the HRP-conjugated GFP antibody (Santa Cruz Biotechnology, sc9996, 1:1000).
APP-HEK293 cells were cultured in 6-well dishes for 24 h and then the media was replaced with neuronal culture media. After 48 h, conditioned media was collected and used to treat primary cortical neurons (12 DIV) cultured on poly-L-lysine coated slides for another 24 h. Apoptotic neurons were detected by the In Situ Cell Death Detection Kit, TMR red (Roche, cat. # 12156792910) according to the manufacturer’s instructions. Nuclei were stained using the mounting medium containing DAPI (Vectashield). Quantification of red TUNEL-positive neurons was done using the NIS Elements software.
The gRNA sequence against human BECN1 genome in exon1 (5’-ggacacgagtttcaagatcctgg-3’; underline indicates the protospacer adjacent motif) was designed using the CRISPR Design tool (http://crispr.mit.edu:8079) [62], which contained a Sau3AI restriction enzyme site at the Cas9 cutting position on its gRNA sequence. Annealed oligonucleotides were inserted into the pSpCas9(BB)-2A-puro (PX459) V2.0 vector (Addgene, #62988). The plasmid was transfected into APP-HEK293 cells using lipofectamine 3000 (Thermo Fisher Sciencetific), and cells were selected for 72 h using DMEM supplemented with 2 μg/ml puromycin. Genome editing efficiency and protein expression levels were confirmed by Sau3AI enzymatic digestion and western blotting, respectively.
Mouse wild-type (WT) Becn1 or Becn1F121A mutant cDNA was sub-cloned into the pCDH-CMV-MCS-EF1-GreenPuro vector (System Biosciences, Palo Alto, CA, USA) using XbaI and BamHI restriction sites. Lentivirus encoding Becn1 or Becn1F121A was produced by co-transfection of packing plasmids, pCMV-VSV-G (Addgene, #8454) and psPAX2 (Addgene, #12260) into HEK293 FT cells. The resulting lentivirus encoding Becn1 or Becn1F121A was used to infect BECN1 KO cells at the multiplicity of infection of 1 for 24 h in the presence of 10 μg/ml polybrene (Santa Cruz Biotechnology). Infected cells were selected and maintained in DMEM supplemented with 2 μg/ml puromycin (Thermo Fisher Scientific).
The biotinylation procedure was modified from a previously reported protocol [63]. HEK293 cells stably expressing APP were grown to 90% confluency on gelatin-coated 6 cm2 dish, washed with ice-cold PBS, and incubated in 0.3 mg/ml disulfide-cleavable biotin (EZlink Sulfo-NHS-SS-Biotin, Thermo Scientific) in PBS at 4°C for 30 min. Cells were then washed with cold PBS and returned to warm medium at 37°C, and incubated for 5 or 15 min. Cells labeled “Total” were left on ice in PBS. Cells labeled “Stripping” were also left on ice in PBS and then stripped as described below. The remaining cell-surface biotinylated APP was stripped in 50 mM glutathione, 0.3 M NaCl, 75 mM NaOH, 10% FBS at 4°C for 40 min. Glutathione was then quenched with 50 mM iodoacetamide and 1% bovine serum albumin in PBS at 4°C for 15 min. Proteins were extracted in lysis buffer containing 0.1% sodium dodecyl sulfate (SDS), 0.5% sodium deoxycholate, 1% Triton X-100, 100 mM NaCl, 2 mM EDTA and 50 mM Tris-HCl 7.4 supplemented with protease and phosphatase inhibitor cocktail (Thermo Scientific), and supernatant was collected by centrifugation at 10,000 xg for 10 min at 4°C. Biotinylated APP were isolated using streptavidin-agarose (Millipore) at 4°C for 2 h. Precipitates were washed four times with wash buffer containing 0.1% SDS, 1% Triton X-100, 100 mM NaCl, 2 mM EDTA and 50 mM Tris-HCl 7.4, and proteins were eluted in SDS sample buffer by boiling.
P ≤ 0.05 was considered statistically significant in two-tailed, unpaired Student’s t-tests for detection of differences between two experimental groups; Two-way ANOVA approach was used for comparison among multiple groups. Statistics on the survival study was done by the log-rank test. Figures are depicted as mean ± SEM.
|
10.1371/journal.pgen.1003580 | Environmental Dependence of Genetic Constraint | The epistatic interactions that underlie evolutionary constraint have mainly been studied for constant external conditions. However, environmental changes may modulate epistasis and hence affect genetic constraints. Here we investigate genetic constraints in the adaptive evolution of a novel regulatory function in variable environments, using the lac repressor, LacI, as a model system. We have systematically reconstructed mutational trajectories from wild type LacI to three different variants that each exhibit an inverse response to the inducing ligand IPTG, and analyzed the higher-order interactions between genetic and environmental changes. We find epistasis to depend strongly on the environment. As a result, mutational steps essential to inversion but inaccessible by positive selection in one environment, become accessible in another. We present a graphical method to analyze the observed complex higher-order interactions between multiple mutations and environmental change, and show how the interactions can be explained by a combination of mutational effects on allostery and thermodynamic stability. This dependency of genetic constraint on the environment should fundamentally affect evolutionary dynamics and affects the interpretation of phylogenetic data.
| Epistatic interactions limit the number of adaptive trajectories to peaks on evolutionary fitness landscapes, and may therefore hamper the progress of evolution. Recent research has focused on adaptive landscapes in one constant environment. However, adaptive evolution is generally known to occur in variable, heterogeneous environments. Here, we have constructed fitness landscapes of three inverse lac repressor variants in two contrasting environments. We find that the epistatic interactions between the pairs of mutations are profoundly altered upon an environmental change. We develop a new graphical method to analyze the underlying higher-order interactions between genetic changes and the environment, and explain the complex environmental dependencies in terms of simple molecular mechanisms. Our results show that the information about epistatic interactions acquired in one environment does not inform on the true limitations of adaptive evolution. We argue that this dependency of genetic constraints on the environment will have important effects on the progress of adaptation in heterogeneous environments, and will affect our ability to establish realistic genealogies from the phylogenic record.
| As pointed out by Sewall Wright in the 1930's, the genetic makeup of a biological system should determine not only current functionality but also affect future evolutionary change [1]. How the present genetic architecture constrains future adaptive evolution is now starting to be addressed experimentally [2]–[4]. By systematically constructing single-mutant neighbors and assaying their function or fitness, proteins ranging from TEM β-lactamase [3] to steroid receptors [5] have been shown to exhibit sign epistasis, in which one mutation can be beneficial or deleterious depending on the presence of another mutation. Sign epistasis by itself does not imply evolutionary constraint, as the interacting mutations may simply not play a role in adaptation. However, when mutations essential for functional innovation exhibit sign-epistasis, constraints emerge for evolutionary trajectories that depend on fixing one adaptive mutation after another by positive selection [6]. For sign-epistatic interactions, the number of such adaptive trajectories is reduced. Two mutations may also be deleterious individually but jointly beneficial, as observed for mutations in the regulator MTH1 and glucose transporters HXT6/HXT7 in Saccharomyces cerevisiae [7] and between argH12 and pyrA5 mutants leading to arginine and pyrimidine deficiency in Aspergillus niger [8]. Such reciprocal sign epistasis is a necessary condition for multiple peaks in the fitness landscape [9], which can completely block evolutionary trajectories in which mutations are fixed one-by-one by positive selection. Because of this ability to arrest, delay, and divert evolution, genetic interactions have been speculated to play a central role [10] in speciation [11], [12], the maintenance of biodiversity [13], and developmental evolution [14], [15].
So far, epistastic interactions have been studied predominantly for environments that are constant in time and favor a single function or phenotype. However, natural environments are characterized by irregular temporal changes, which in turn impose temporally changing demands on the expressed phenotypes. Indeed, the complexity of regulatory systems is considered to have evolved in response to environmental heterogeneity [16], [17]. Experimentally, mutations are commonly observed to have different effects in different environments [18]–[20]. For example, in Escherichia coli the fitness effects of single Tn10 transposon insertion mutations [21]and mutations conferring resistance to bacteriophages λ and T4 have been shown to depend on the genetic background and the environment [22]. Correlations exist between epistatic interactions in plant viruses and their hosts [23], and trade-offs have been observed between the effect of mutations in the presence of certain types or concentrations of antibiotics in Escherichia coli [24], [25] and Pseudomonas aeruginosa [26].
These observations raise the question to which extent constraints themselves change when the environment changes. If mutations essential to functional innovation exhibit sign-epistatic interactions that are modulated by environmental change, adaptive trajectories will be drastically affected. For instance, evolutionary change hampered by adaptive valleys in one environment could be opened up to positive selection in another. Conversely, trajectories that can be positively selected for in constant environments [2], [3] could be blocked by environment-induced sign epistasis, which could slow down overall evolutionary progress or drive adaptation to dead ends in genotype space. This environmental control over the accessibility of adaptive trajectories goes beyond merely defining a variable selective environment, and would invalidate commonly held assumptions in analyzing the historical evolutionary record by phylogenetic reconstruction (23).
These elementary issues can be readily investigated using a simple phenotype that responds to the environment. We focused on one of the most well-understood model systems for environmentally controlled gene expression, the Escherichia coli lac repressor LacI [27]. We considered the evolutionary transition to a variant that exhibits an altered regulatory response [28]. In the presence of the wild-type repressor, LacIwt, the lac operon is induced by the ligand IPTG, whereas in the presence of the variant LacIinv, expression is suppressed by IPTG. We have previously isolated LacI variants with such inverse phenotypes in evolutionary experiments [28] (Text S1), which serve as a basis to systematically assess how the environment affects epistasis between the mutations required for inversion. We find that the epistasis is highly environment-dependent, which implies that epistasis perceived in a constant environment does not properly inform on the evolutionary constraints in a variable environment. We can explain the generic pattern of higher-order genotype x genotype x environment interactions that is observed in all three variants using a simple model of changes in the allosteric transition and in protein stability.
To investigate the interplay between the environment and epistasis we focused on three inverse LacI variants [28] (Text S1). The three inverse variants each contained three to six point mutations relative to LacIwt. For all variants, three mutations appeared essential for the inverse function, as was determined by engineering lacI variants that contained sub-sets of these mutations. We denote these three inverse variants as LacIinv1 (S97P, R207L, T258A), LacIinv2 (S97P, L307H, L349P) and LacIinv3 (S97P, G315D, P339H). Note that all share the mutation S97P. Next, we constructed all the single and double mutants, and assayed the operon expression phenotypes in the absence of IPTG (Env0) and in the presence of 1 mM IPTG (Env1) (Table S1) using a fluorogenic reporter assay (materials and methods) (Figure 1A). Given the evolutionary objective of inversion, a high operon expression level is favored in Env0, whereas a low expression level is favored in Env1 [28] (Figure 1B).
To compare the epistasis in each environment, we classified the epistatic (genotype x genotype) interactions for all pairs of mutations for each of the three inverse LacI variants. We distinguished three categories: magnitude epistasis (M) - both mutations are either beneficial or deleterious, irrespective of the genetic background, sign epistasis (S) – the effect of one mutation changes sign depending on the genetic background, or reciprocal sign epistasis (R) - both mutations are individually deleterious, but beneficial in combination [4]. Neutral mutations are not positively selected and are thus grouped under deleterious. We find that nine out of the eighteen mutation pairs display the same category in environments Env0 and Env1 (Table 1). For instance, in the P349 background, L307H and S97P exhibit sign epistasis in both environments (Table 1, LacIinv2). Note that for all these nine pairs, the magnitude of the mutational effect does depend on the environment, but the sign does not. For the other nine mutation pairs, the category of epistasis differs between the two environments (Table 1). Some sign epistatic interactions are switched ‘off’ by the addition of IPTG. In the P97 background for instance, IPTG induces a sign change in the effect of R207L; it transforms the sign-epistasis between R207L and T258A in Env0 to magnitude epistasis in Env1 (Table 1, LacIinv1). Sign epistasis is turned ‘on’ between other mutations. For instance, in a P97 background, L349P and L307H exhibit sign epistasis in an environment without IPTG, and reciprocal sign epistasis with IPTG (Table 1, LacIinv2). Thus, environmental signals modulate sign-epistatic interactions between residues involved in the functional inversion of LacI.
The above classification of genetic interactions into categories reveals a dependence on the environment, but it does not offer intuitive insights into their causes. These dependencies may also be viewed as three-way interactions between two genetic changes and one environmental change. Hence, they can be denoted as genotype x genotype x environment interactions, or briefly GxGxE; analogous to two-way GxG interactions between two genetic changes in a single environment, or two-way GxE interactions between one genetic change and one environmental change [17]. To analyze these higher-order interactions, we introduced a graphical method (Figure 2A). Mutations are represented as vectors in a two-dimensional coordinate system, where the axes indicate the corresponding changes in expression phenotype in both environments. A vector pointing to quadrant I signifies functional improvements in both environments, whereas quadrants II and IV denote improvement in one environment and deterioration in the other, and quadrant III denotes deterioration in both. The probability of fixing neutral mutations is low compared to positively selected mutations that confer functional improvements [6], [29]. Mutations that are neutral in both environments therefore correspond to quadrant III, while mutations that are neutral in one environment and beneficial in the other correspond to quadrants II or IV. Thus, mutations in quadrants II and IV indicate sign-changing GxE interactions.
Higher-order interactions between two or more mutations and the environment can be visualized by sets of paths composed of two or more mutational vectors (Figure 2). The two mutational paths from genotype ab to AB (via Ab or via aB) form a four-sided polygon. The polygon is a simple parallelogram in the absence of any genetic interactions, which may occur either without (Figure 2B) or with GxE interactions (Figure 2C). Deviations from the parallelogram indicate genetic interactions, or epistasis. Vectors at opposing sides of the polygon that have different angles but point in the same quadrant indicate magnitude epistasis. Opposing vectors pointing in different quadrants indicate sign-epistatic interactions (GxG, Figure 2D), and when the sign change of opposing vectors is conditional on the environment higher-order GxGxE interactions can be observed (GxGxE, Figure 2E). Thus, higher-order interactions between mutations and the environment can be graphically recognized and classified using the mutational vector plots.
We analyzed the interactions for the three LacI variants by displaying the expression data as mutational vectors in Figure 3A, B and C. Because the transition to inversion is characterized by a decreasing operon expression in the presence of IPTG (Env1) and an increasing operon expression in the absence of IPTG (Env0), we plotted 1/expression in Env1 against the expression in Env0, such that the closer the phenotype comes to the objective of inversion, the more it moves towards the upper-right corner of Figure 3.
Inspection of the polygon shapes shows that half (50%) lack the signatures of sign-changing higher-order interactions involving mutation pairs and the environment. For instance in Figure 3C, the opposing red and green vectors in the P97 background point in the same quadrant. The polygon is tilted, with both red vectors pointing in quadrant IV, indicating GxE interactions. However, the other half of the opposing mutational vector pairs in the polygons do not point in the same quadrant, indicating the pervasive presence of higher-order GxGxE interactions. For instance, in the P97 background, the addition of T258A turns the green vector (R207L) from quadrant III to IV, which is caused by the fact that R207L is neutral in the presence of IPTG and the absence of T258A, but increases expression by 20-fold in T258A's presence (Fig. 3A). Another example is the addition of L307H, which rotates the red vector (L349P) from quadrant IV to II in the P97 background, which indicates that the effect of L349P on expression changes sign in both environments due to L307H (Figure 3B).
Overall, the pattern displayed by the three variants in the vector plots (Figure 3A, B and C) is strikingly similar, in contrast to the diverse environmental dependence of epistasis seen in Table 1. The blue vectors initially point predominantly up along the Env1 axis (the expression level decreases with IPTG), as the expression level in Env1 is strongly decreased, but turn diagonally to the upper-right corner when the red and green mutations are added (the expression level increases simultaneously in the absence of IPTG) (Figure 3). On the other hand, the green and red vectors either point downward along the Env1-axis, (expression mainly increases in the presence of IPTG), or to the right along the Env0-axis (expression increases in the absence of IPTG). Mutation S97P appears responsible for this rotation of the red and green vectors: in the LacIwt background they point along Env1, while in the P97 background they point along Env0. In other words, S97P represents a ‘switch’ that changes the interaction of the red and green mutations with the environment. This pattern is identical for all three inverse genotypes; all show a roughly similar rotation for the blue as well as for the red and green vectors. Thus, while the genetic solutions to the phenotypic inversion are different in the three variants, the main features of the underlying map of the interactions between genotypes and the environment are general.
Note that one may also consider the presence of higher-order interactions that are purely genetic. Specifically, such GxGxG interactions arise when the addition of a third mutation changes the category of the two-way epistatic motif. For instance, in the wild type background, both green (L307H) and red (L349P) vectors point downward or are neutral along the Env1 axis (Figure 3B), and hence point to magnitude epistasis. However, upon the application of S97P (Figure 3B, blue vectors), one green and one red vector still points down, but one green and one red vector is rotated upwards. Thus, L307H and L349P display reciprocal sign epistasis in the presence of P97, and hence their three-way interaction in Env1 cannot be captured by two-way epistasis alone. Note that this GxGxG interaction itself may in turn be dependent on the environment, indicating GxGxGxE interactions. Among other things, the presence of higher-order genetic interactions illustrates that conclusions on the accessibility of a genotype must be carefully considered. This is particularly relevant when it is unclear to what extent the mapped genotype space fully determines the considered function, as an untested mutation could open up mutational pathways to selection, which otherwise may have been considered blocked [30]. The principle of such effects of higher-order genetic interactions have previously been captured [3], [4], [7], [15], [31] when mapping a larger landscape and assessing the mutational pathways within it. Nonetheless, the explicit presence of GxGxG interactions underscores the care that must be taken when formulating conclusions about selection and constraint from fitness landscapes.
The results also underscore that mechanisms that are comparatively simple on the molecular level, can give rise to GxE interactions. For instance, in the P97 background, L307H has the simple mechanistic effect of generally increasing expression both in the presence and absence of IPTG. In terms of selection, this change is beneficial in one environment (in the absence of IPTG), and deleterious in the other (in the presence of IPTG). Hence, L307H gives rise to a GxE interaction, a trade-off. Given the generic purpose of regulatory functions to modulate biological functions in response to input signals, one can expect such trade-offs that originate from simple molecular mechanisms to be rather generally present.
The observed generality of the genotype-environment interaction maps (Figure 3) suggests that they result from a generic structural cause. However, the positions of the mutated residues within the LacI crystal structure do not directly reveal generic features, as they appear scattered throughout the structure, with different locations for the different variants (Figure S2). Also, the mutations are not positioned at obvious functional sites such as the DNA or ligand binding regions. Alternatively, the origin of the interactions may be rooted in the mechanism of inversion, which has been speculated to be based on two effects [28], [32]. First, the allosteric transition from high to low operator affinity is thought to be impeded by S97P, as P97 cannot form the transient bond with K84′ and V94′ [33], which in turn locks the structure in the DNA-bound confirmation [34], [35]. Second, the response to inducer is assumed to be inverted through changes in the thermodynamic stability of the protein: the additional two mutations in each variant would lower the stability in the absence of IPTG, which would confer an increased expression level in Env0, while the binding of the ligand IPTG to LacI would confer a stabilizing effect that conserves a low expression level in Env1. Our experiments showed that in a LacIwt background, S79P lowers expression in Env1 to repressed levels while maintaining a relatively low expression level in Env0. Thus, these data are indeed consistent with the proposed locking of LacI in the DNA-bound confirmation.
The data further show that expression in Env1 varies along the mutational trajectories from LacIwt to LacIinv (Figure 4A). In contrast, in Env0, the trajectories to inversion show a generic increasing trend in the expression level; all first mutations yield little to no changes, while second and third show increasingly large expression increases (Figure 4B). The pattern of changes in expression level in both environments is consistent with stability-decreasing mutations, as: 1) correlation between the stability and the expression level should be stronger in Env0, as the ability to tightly bind DNA in that environment is dependent on structural stability, in contrast with the ability to efficiently release from the DNA in Env1, and 2) it has been argued that protein function is robust against initial stability decreases, but can be expected to deteriorate when accumulated mutations drive the system across their so-called stability threshold [36]–[38]. We investigated the destabilizing effect of the mutations by analyzing the stability changes due to amino acid substitutions in silico with FoldX [39], [40]. In the absence of IPTG (Env0), FoldX indeed showed significant stability decreases for most (8 out of 11, Table S2) of the studied mutants, including S97P. The expression measurements suggest that in particular S97P brings LacI to the edge of the stability threshold, as subsequent mutations strongly increase expression (Figure 3, Env0). Thus the S97P substitution acts as a switch that systematically alters the phenotypic effect of the other mutations.
While we have addressed the central features of the interaction map, various more detailed interactions between mutations and the environment remain to be explained mechanistically. However, overall the analysis indicates that the combined effects of two independent and simple molecular mechanisms can explain complex higher-order GxGxE interactions between multiple mutations and the environment.
Recent systematic reconstructions of evolutionary intermediates have provided a first view on adaptive landscapes and the causes of evolutionary constraint [4]. Sign epistatic interactions between mutations have been shown to limit the number of mutational trajectories that can be followed under positive selection in constant environments [2], [3]. Directed evolution experiments revealed evolutionary constraints that delay or prevent adaptation [15], [28], and measured trade-offs between environments indicated how such constraints affect selection in variable environments [28], [41]–[43]. Here we investigated how the environment affected the adaptive landscape describing a specific functional innovation, by reconstructing the evolutionary intermediates on route to three different inverse LacI genotypes.
The three evolved genotypes indicated a redundancy within the LacI genetic architecture to develop regulatory functions that respond to the environment, mirroring similar results obtained for microbial populations evolving in constant environments [44]–[46]. We found that a mechanistic model of inversion provided an explanation for the origin of this parallelism. First, a mutation (S97P) blocks the IPTG-induced allosteric transition, and thus affects expression only in the presence of IPTG. Second, the initial mutations have little effect on the ability to repress in the absence of IPTG, while later mutations have a large effect. Third, binding to the ligand IPTG increases the protein stability and hence the ability to repress. Thus, a combination of simple molecular mechanisms can explain the observed complex higher-order interactions between multiple genetic changes and an environmental change.
The data showed that the genetic epistasis in LacI was pervasively dependent on the environment. As the studied genetic changes were not chosen randomly but jointly confer a novel regulatory response, these results inform on constraints in the evolution of a novel biological function. They indicate that limitations in the selective accessibility of trajectories, as detected in a constant environment, not properly inform on evolutionary limitations in the natural variable environment. Due to the environmental dependence of epistasis, some trajectories are closed-off by environmental change while others are opened-up to positive selection. Intriguingly, a consequence of environmental dependence of epistasis is that few mutations are blocked in all environments, and many are positively selected in at least one environment. This suggests that genetic constraints may be more readily overcome in certain variable environments than expected from epistasis detected in constant environments [47], [48].
More generally, the results underscore the complex and diverse roles of the environment in evolutionary dynamics. The environment does not only define a selective pressure on a phenotypic trait or induce a phenotypic change, but also modulates the underlying genetic constraint. This interdependence has a number of consequences. For instance, it affects our ability to understand the evolutionary record as interpreted from extant genetic sequence data. By modulating evolutionary constraint in time, environmental variations can change substitution rates across evolutionary trees [49], [50], referred to as heterotachy, even if selection on a phenotypic trait is constant. It can result in topological inaccuracies in phylogenetic trees [51] such as long-branch biases [52], [53] and a lack of phylogenetic resolution [52], [54] if the underlying adaptive landscapes are shaped differently in each of the environments. This can ultimately affect the predictive power of phylogenetic reconstruction techniques in their use for the prognosis of the emergence and the spread of diseases, such as the spread of the influenza virus [55], where the host can be viewed as a biotic environment [56]. And lastly, it renders a walk on evolutionary branches of life unpredictable and unrepeatable [3], [57], as some adaptive trajectories are constrained in some environments, but not in others.
It will be intriguing to explore the prevalence of the higher-order genotype x genotype x environment interactions in other biological systems. It is not obvious that all biological functions will show such interactions; in particular those specialized to a single environmental factor. On the other hand, the ability to respond to environmental stimuli is one of the defining properties of living systems. Given the inherent interdependency between regulatory systems and the environment, we expect that such insights into the interplay between genetic architecture and the environment will be crucial for a mechanistic understanding of the evolution of biological functions.
Escherichia coli K12 strain MC1061 [58], which carries a deletion of the lac operon was used in all experiments. This strain was obtained from Avidity LLC, Denver CO, USA, as electrocompetent strain EVB100 (containing an additional chromosomal birA). Plasmid pRD007 was constructed based on the pZ vector system [59] and contains LacI, driven by the PLO1-Tet promoter. The reporter plasmid pReplacZ, used for the quantification of LacZ expression, was created by deletion of lacI and Ptrc in pTrc99A [60] followed by insertion of the Plac-lacZ fragment of MG1655 [61].
In all experiments EZ defined rich medium (Teknova, Hollister, CA, USA) with 0.2% glucose and 1 mM thiamine HCL (Sigma) was used. Isopropyl β-D-1-thiogalactopyranoside (IPTG) was purchased from Sigma, and was added to the medium, if applicable, in a 1 mM quantity.
Mutations were introduced into the coding region of lacI by site-directed mutagenesis with the QuickChange II–E Site–Directed Mutagenesis Kit (Stratagene, USA) according to the manufacturer's protocol [28]. Constructs are available upon request.
Cultures were grown at 37°C in a Perkin & Elmer Victor3 plate reader, at 200 µl per well in a black clear-bottom 96 well plate (NUNC 165305). Expression measurements were performed in EZ Rich Defined medium with added 0.2% glucose (Teknova, Hollister, CA, USA, cat. nr. M2105) supplemented with 1 mM thiamine HCl and the appropriate antibiotics for the selective maintenance of plasmid pRD007 and pRepLacZ. Optical density at 600 nm was recorded every 4 min, and every 29 min 9 µl sterile water was added to each well to counteract evaporation. When not measuring, the plate reader was shaking the plate at double orbit with a diameter of 2 mm. Cells were fixed after the cultures had reached an optical density of at least 0.015 and at most 0.07, by adding 20 µl FDG-fixation solution (109 µM fluorescein di-β-D-galactopyranoside (FDG, Enzo Life sciences, NL), 0.15% formaldehyde, and 0.04% DMSO in water). Fluorescence development was measured every 8 min (exc. 480 nm, em. 535 nm), as well as the OD600. Shaking and dispensing conditions were as mentioned above. When cells are not induced with IPTG, directly before or after fixation an appropriate amount of inhibitive IPTG was added. Analysis of the fluorescence trace is as described in [28].
Significance of the phenotypic effect of mutations in LacI was tested with a t-test with Bonferroni correction for multiple comparisons (P<0.05). While the phenotypic effect of S97P in the wild type background in Env0, was not significant in the data set of one inverse Lac variant (LacIinv3), it was significant for the two other variants, and hence S97P was considered significant for the wild type background and Env0.
A FoldX plugin [40](version 1.4.22) in the Yasara software package [62](version 11.11.4) was used for the stability analysis of the single, double and triple (only LacIinv1) mutants on basis of the DNA bound dimeric LacI crystal structure (1EFA) [63], which lacks the tetramerization domain. The structure was minimized without ONPF before addition of the mutations, and the calculation of the stability changes. The stability calculation was performed three times for each mutation, with standard deviations among the calculations smaller than ΔΔG = 0.5 kcal/mol.
|
10.1371/journal.pcbi.1002639 | Prediction of Mutational Tolerance in HIV-1 Protease and Reverse Transcriptase Using Flexible Backbone Protein Design | Predicting which mutations proteins tolerate while maintaining their structure and function has important applications for modeling fundamental properties of proteins and their evolution; it also drives progress in protein design. Here we develop a computational model to predict the tolerated sequence space of HIV-1 protease reachable by single mutations. We assess the model by comparison to the observed variability in more than 50,000 HIV-1 protease sequences, one of the most comprehensive datasets on tolerated sequence space. We then extend the model to a second protein, reverse transcriptase. The model integrates multiple structural and functional constraints acting on a protein and uses ensembles of protein conformations. We find the model correctly captures a considerable fraction of protease and reverse-transcriptase mutational tolerance and shows comparable accuracy using either experimentally determined or computationally generated structural ensembles. Predictions of tolerated sequence space afforded by the model provide insights into stability-function tradeoffs in the emergence of resistance mutations and into strengths and limitations of the computational model.
| Many related protein sequences can be consistent with the structure and function of a given protein, suggesting that proteins may be quite robust to mutations. This tolerance to mutations is frequently exploited by pathogens. In particular, pathogens can rapidly evolve mutated proteins that have a new function - resistance against a therapeutic inhibitor - without abandoning other functions essential for the pathogen. This principle may also hold more generally: Proteins tolerant to mutational changes can more easily acquire new functions while maintaining their existing properties. The ability to predict the tolerance of proteins to mutation could thus help both to analyze the emergence of resistance mutations in pathogens and to engineer proteins with new functions. Here we develop a computational model to predict protein mutational tolerance towards point mutations accessible by single nucleotide changes, and validate it using two important pathogenic proteins and therapeutic targets: the protease and reverse transcriptase from HIV-1. The model provides insights into how resistance emerges and makes testable predictions on mutations that have not been seen yet. Similar models of mutational tolerance should be useful for characterizing and reengineering the functions of other proteins for which a three-dimensional structure is available.
| The relationship between protein sequence and structure is fundamental for protein function, evolution and design [1], [2]. Many sequences are compatible with a given structure and function and thus proteins are often robust to point mutation [3], [4], [5]. The concept of “tolerated sequence space" - the set of sequences that accommodate a given structure and function - has been applied to characterize the emergence of protein families [6], to describe protein interaction specificity [7] and to explain the evolution of new protein functions [8], [9].
Tolerated sequence variability (robustness to mutation) should be an advantage if proteins need to satisfy multiple functional constraints simultaneously. If each constraint can be accommodated by many sequences, it should be easier to find a subset of sequences that satisfy multiple requirements [10]. Moreover, a protein that has many tolerated sequences may be able to accommodate new constraints without abandoning some existing function [8], [11], [12].
An example of this ability of proteins to rapidly adapt to new pressures is the emergence of drug-resistance mutations in pathogens. In many cases, variants of pathogenic proteins that are resistant to inhibitors appear quickly, while still preserving their essential functions for the pathogen. It is likely that some of these mutations are already present in the population as part of naturally occurring nearly neutral sequence variation [13] and are then selected by inhibitor treatment. Thus, the a priori prediction of the tolerated sequence variation of pathogenic proteins would have implications for development of inhibitors against which resistance is less likely to arise quickly [14].
Here we develop and assess a computational approach to predict the tolerated space of single mutations around a given protein sequence. As model systems for validating our approach, we use the protease and reverse transcriptase from HIV-1. With more than 50,000 known sequences and several hundred experimentally determined structures, these two viral proteins are among the best-characterized systems available of tolerated variants around a native sequence. Because protein sequences have been collected before and after viral inhibitor treatment [15], predictions of mutational tolerance can be assessed in both a nearly neutral setting and under selective pressure to evolve resistance mutations. In testing our model for HIV-1 protease mutational tolerance, we also make use of a large-scale mutagenesis experiment which evaluated the in vivo function of roughly 50% of all mis-sense mutations reachable by a single-nucleotide change from a starting consensus sequence [16].
We find that our approach, which employs computational protein design methods in Rosetta [17], recapitulates a substantial fraction of mutations experimentally observed to be tolerated by HIV protease and reverse transcriptase. For accurate predictions, we show that it is critical to treat the protein not as a rigid single structure, but to allow conformational variation to accommodate sequence changes [18], [19], [20]. We show that essentially the same prediction accuracy is achieved when obtaining conformational variation from an ensemble of experimentally determined structures of HIV protease [21] or reverse transcriptase, or from computationally generated conformational ensembles [18], [19], [20], [22]. We thus expect our approach to also be applicable to systems for which there is only one structure known. Computational models of accessible mutational space, such as the one presented here, may prove generally useful for describing the evolvability of proteins by forecasting the emergence of mutations that can enable new protein functions [8].
To predict a protein's tolerance to mutation, ideally all constraints acting on that protein should be modeled explicitly. In addition, accurate predictions of mutational tolerance may require that conformational adjustments in response to mutation be considered. Here we present a methodology that incorporates multiple functional constraints as well as backbone flexibility into RosettaDesign [17] and apply it to the prediction of mutational tolerance. We first consider the viral protein HIV-1 protease, and later extend our results to HIV-1 reverse transcriptase.
HIV-1 protease is an ideal test system for several reasons. First, the mutational tolerance of HIV-1 protease is well characterized: mutations of HIV-1 protease, including those causing resistance to protease inhibitors in HIV treatment, have been extensively documented and are available in the Stanford HIV-1 Drug Resistance Database [15]. Second, HIV-1 protease is under at least three structural and functional constraints that are straightforward to model: (1) the 99-residue protease sequence must adopt a stable fold; (2) the active enzyme is a homodimer, and (3) the dimeric form must bind at least 10 endogenous peptides. Finally, HIV-1 protease is structurally well characterized, with hundreds of crystal structures of native and mutated forms in the apo state or with peptide or inhibitors bound.
Figure 1A outlines the computational strategy for predicting mutational tolerance, starting from three-dimensional structural information on the protein of interest. Figure 1B gives an example of the calculations for one sequence position in HIV-1 protease. We started from the consensus sequence for HIV-1 protease (see Methods), and considered all individual point mutations independently (the simplest model of mutational space around a given sequence). We used RosettaDesign [10], [17], [18] to mutate, in silico, each sequence position to 19 naturally occurring amino acid types (mutations to and from cysteine were excluded; see Methods). For each residue change, the side-chain conformations were optimized around the site of mutation. We then calculated the per-residue energy contribution (termed ERES) of each point mutation using the Rosetta all-atom force field (see Methods). ERES scores were computed with respect to the three functional pressures described above: (1) the stability of the protease fold (ERESFold, Figure 1B, left); (2) the stability of the protease dimer interface (ERESDimer, Figure 1B, middle); and (3) the stability of the binding interactions with endogenous substrate peptides (ERESPeptide, Figure 1B, right). The model has the following key steps and components (Figure 1A):
Evaluating the robustness of a protein to mutation requires accurate distinction between sites that display amino acid variation and ones that do not. Some protein sites are mutation intolerant under neutral conditions but become more tolerant under selective pressure; other sites are intolerant to mutation under both neutral and selective conditions.
Approximately 2/3 of protease sites (63 out of 96) within the Stanford HIV-1 Database sequences [15] appeared largely intolerant to mutation prior to inhibitor treatment (Figure 2A; intolerance to mutation defined as a mutation frequency of <1%). Further, about half of protease sites within the database (43 out of 96) were largely intolerant to mutations under inhibitor treatment (Figure 2B). The neutral and selective models correctly identified the majority of these intolerant protease sites (Figure 2A–B; 45/63 and 31/43, respectively).
Within the database sequences, only a few protease sites displayed high mutational tolerance (Figure 2A–B; 8 and 14 sites, in the absence and presence of inhibitors, respectively; high mutational tolerance defined as a mutation frequency >20%). The neutral and selective models correctly identified over half of these frequently mutated protease sites (Figure 2A–B; 5/8 and 8/14 sites, respectively), including five sites that displayed high mutational tolerance in both a neutral setting and under selective pressure (Figure 2C; 35E, 37N, 62I, 63L, and 77V). Importantly, the individual mutations observed in the Stanford database were also correctly predicted for many of the frequently mutated sites (Figure 2C; bold residues in 4th and 7th columns). Similar results were observed at sites within the database with moderate mutational tolerance; these sites were often correctly predicted by both the neutral and selective models (Figure 2C; 12T, 14K, 18Q, 19L, 20K, 39P, 60D, 61Q, 70K, and 92Q; moderate mutational tolerance is defined as amino-acid variation between 1–20%). Therefore we conclude that the models can, in many cases, recapitulate both protease sites and individual protease mutations that are functionally tolerated (the results for the neutral model are shown in Figure S1).
To quantify the overall ability of the neutral and selective models to recapitulate individual mutations observed in the Stanford database, we used two standard metrics: (1) We computed a Receiver Operating Characteristic (ROC) curve by calculating the true positive rate (TPR) and false positive rate (FPR) of identifying protease mutations observed within the Stanford database above a threshold frequency of 1% and (2) we calculated an Area Under the Curve (AUC) for each ROC plot (Figure 3). Both the neutral and selective models recapitulated many HIV-1 protease database mutations without incorrectly predicting a large number of false positives (Figure 3A and 3E; black curve and black bar). Commonly, a model with no predictive power will have a ROC curve that is a diagonal line and an AUC value of 50%. We chose two naïve mutation tolerance prediction models as additional references. In control model 1, each site can tolerate all mutations that are accessible by a single nucleotide change from the consensus sequence. In control model 2, each site can tolerate amino acid types chemically similar to the native amino acid (see Methods). Control model 1 predicted the majority of the experimentally observed mutations (TPR ∼90%, red triangle in Figure 3A). However, a large number of non-observed mutations were incorrectly predicted as tolerated (∼37% FPR). The computational models had a lower FPR at the same TPR. Control model 2 rarely predicted tolerance to mutations that were not observed within the database (∼11% FPR), but did not capture tolerance to many database mutations (∼60% TPR, blue square in Figure 3A). The computational models ranked more mutations correctly at the same FPRs.
In addition to recapitulating database mutations found in either neutral or selective conditions, our prediction scheme was also successful in recovering literature-documented drug resistance mutations (DRMs) for protease. The comparison between predictions of the neutral and selective models (Figure 2C) yielded 18 sites that showed increase in mutation frequency (rare/moderate to moderate/high), 9 out of which contain previously characterized DRMs (as listed in [23], see circles in Figure 2C). Thus comparing predictions from the neutral and selective models may, in some cases, allow for identification of sites that contain drug resistance mutations.
Overall, the agreement between the individual mutations appearing within the database and the mutations predicted as tolerated by the models was strong (Figures 2C, 3A and Figure S1). Nevertheless, several notable under- and over-predictions were observed.
Under-predictions of mutational tolerance by the neutral model were most notable at 10 sites (Figure 2C, 2nd and 5th columns; 13I, 15I, 16G, 33L, 36M, 41R, 57R, 64I, 89L, and 93I). The same 10 sites were also under-predicted for the selective model, with additional under-predictions occurring at 8 sites (Figure 2C, 3rd and 6th columns; 10L, 20K, 48G, 54I, 73G, 82V, 84I, and 90L). At most of these sites, the specific mutations observed in the Stanford database were correctly identified, but the predicted frequencies of mutation were significantly less than experimentally observed (Figure 2C; 4th and 7th columns). Notably, almost all under-predicted sites contained DRMs (see circles in Figure 2C; exceptions are 15I, 41R and 57R). Under-predictions may result from errors in the Rosetta energy model or from the inability to correctly capture structural changes in response to sequence changes.
Over-predictions of mutational tolerance occurred primarily within the beta-sheet pairing of the dimer interface (1P, 3I, 6W, 98N), three sites in the dimer flaps (45K, 46M and 47I), and several surface sites (21E, 35E, 43K, 55K, 58Q, 65E, 69H, and 72I, Figure 2C). DRMs were relatively rare within sites that were over-predicted, although they did occur at two sites within the protease flaps (46M, 47I) and at surface sites (35E, 43K, 58Q, and 69H; circles in Figure 2C). As with under-predictions, model over-predictions could be due either to inaccuracies of the Rosetta model or additional functional pressures not captured. The high predicted frequency of mutation at sites 46 and 47 likely occurred due to the presence of a clash with one of the modeled substrate peptides at these sites. Thus, predictions at these two sites might be improved if a crystallographic structure of protease bound to this modeled peptide was available. In addition, Rosetta often performed poorly at predicting mutation frequencies at polar exposed sites. This poorer performance highlights known difficulties in accurately modeling the energetics of polar interactions. Furthermore, despite the inclusion of two terms to disfavor mutations away from polar residues, we may not correctly capture other pressures acting particularly on surface residues, such as selection against aggregation.
As described above, we noted several instances where the selective model predictions did not agree with the mutations observed in the HIV-1 protease database sequences. However, we found that some predictions instead agreed with mutations shown to be tolerated in an experimental study of single mis-sense mutations [16] (Figure 2C, bold residues in 8th column). This finding suggests that the selective model might capture protease mutational tolerance not yet observed at high frequency within the database sequences. In support of this idea, we note that three mutations recently identified in the presence of inhibitors (M46V, F53Y, and N83D) [24], [25] were predicted as tolerated by the selective computational model (Figure 2C, 4th column). All three newly identified mutations were not yet found within the protease database sequences at appreciable frequencies.
As described above, differences observed between the selective and neutral models can be used to recapitulate and predict DRMs. In this section we examine in detail the ability of the model to recapitulate tolerance for 71 previously characterized DRMs. We used a list of mutations from [23] and their grouping into major and minor DRMs. Both groups show an increased frequency of mutation after inhibitor treatment, but only major DRMs have been directly implicated in causing resistance to inhibitors.
The selective model permits mutations near the protease inhibitor binding-site by weakening constraints on the protease dimer and substrate-binding interface. We first analyzed whether the selective model predicts tolerance to DRMs located within the inhibitor-binding site. Of the 18 DRMs near the substrate-binding site, 12 were predicted as tolerated by the selective model (Figure 4A; 3 DRMs were disfavored by the model as they required more than a single nucleotide change from the consensus sequence). Not surprisingly, most DRMs within the substrate-binding site were predicted to have mild-to-moderate destabilizing effects on binding of at least one of the 10 endogenous peptide substrates (Figure 4A, red coloring). The three DRMs not identified by the selective model were predicted to highly destabilize binding of at least one peptide (Figure 4A, red boxes; 82L/F, 48V). In contrast, effects on fold and dimer stability of the DRMs within the inhibitor-binding site were predicted as mostly energetically favorable or neutral (Figure 4A, blue and beige coloring; 47A, 48V and 53L are notable exceptions). At least one mechanism to compensate for substrate binding destabilization is known. Peptide sequences cleaved by HIV protease can co-evolve with the appearance of DRMs such that mutations within the cleavage sequences counteract the predicted losses in substrate binding affinity [26], [27], [28]. Although the selective model does not directly mimic this mechanism of co-evolution, it correctly predicted tolerance to most documented DRMs within the protease inhibitor-binding site.
We next examined DRMs known to occur outside of the protease substrate-binding site. Here, the selective model correctly predicted mutational tolerance towards almost all major DRMs and towards the majority of minor DRMs (Figure 4B; 7/8 and 31/45, respectively; note 6 minor DRMs were disfavored by the model). In the cases where the model did not predict a DRM to be tolerated, it was because the mutation was calculated to strongly destabilize the protease fold (Figure 4B, red coloring). These predicted destabilizing effects of some mutations may need to be compensated for by other co-occurring mutations. Consistent with this hypothesis, 4 out of the 12 predicted destabilizing DRMs (close and far from the substrate binding site) occurred in the 53 most statistically significant correlated pairs of mutations observed after protease inhibitor treatment [29]. Even though the selective model currently cannot account for correlated mutations, it nevertheless correctly predicts tolerance towards a considerable number of DRMs outside of the protease-binding site.
We next examined the contribution of stabilizing mutations to DRMs in HIV protease. This analysis was based on a set of 62 out of the 71 documented DRMs, which had a frequency of >0.5% in the Stanford HIV database. In total, 11 of the DRMs were predicted to stabilize the protease fold, both within (30N, 32I, 46I/L, 50L Figure 4A) and outside (35G, 43T, 63P and 71V/I/T, Figure 4B) the binding site. Interestingly, DRMs at sites 30, 32 and 50 are predicted to have a favorable effect on fold stability, and a destabilizing effect on peptide binding. We asked whether DRMs that are predicted to have a fold-stabilizing effect (out of all 62 DRMs that are both documented and predicted) are over-represented relative to any documented protease mutation predicted to have a fold-stabilizing effect (out of all possible protease mutations reachable by a single nucleotide change from the consensus sequence). We found that there is a significant overrepresentation of DRMs that are predicted to be stabilizing (ΔERESFold<0): 17.7% (11/62), in contrast to only 10% (72/705) of all protease mutations observed in the HIV-1 database reachable by a single nucleotide change (p value = 1.43E-7, Mann-Whitney test). One possible reason for the overrepresentation of stabilizing DRMs is that these sites reside in special locations (such as buried sites that generally contribute more to stability) in the protein structure. We thus calculated the percentage of buried and exposed DRMs and compared these values to the percentage of buried and exposed residues of all documented protease mutations (Figure S2). We found no significant difference in the burial of positions at which DRMs appear. In addition, we studied a list of 33 frequent DRMs that often occur in combination (extracted from the Stanford HIV database, Table S5). Assigning our calculated ERESFold scores for these mutations, we found that 22/33 of the co-occurring mutations included a combination of at least one destabilizing and one stabilizing mutation. These analyses suggest that the modeled stabilizing DRMs may play a role in drug resistance by compensating for the destabilizing effects of other mutations.
We next analyzed whether two key features of the model – incorporating multiple constraints and using backbone ensembles – contributed to prediction performance, using the ROC and AUC metrics introduced above. We first asked whether the model we present, which incorporates fold, dimer and peptide constraints for HIV-1 protease, would outperform a simpler model that considers only fold stability. To do so, we recalculated mutational tolerance at every protease site, but this time we used only the scores for each point mutation and we set all the and terms to zero (“single constraint model"). The predictions of mutational tolerance from this single constraint model were less accurate than the original multiple constraint model, at least under selective conditions (Figure 3B and 3E; cyan curves and bars). Thus, incorporating multiple constraints may be particularly useful for modeling selective pressure, because it allows weakening of certain constraints (such as dimer stability and substrate binding) over others.
We next tested how accurately protease mutational tolerance would be predicted if only a single protease structure was used. To do so, we tested a “single structure model", in which we made 263 independent calculations of HIV-1 protease mutational tolerance. In each set of predictions, we used the ERESFold and ERESDimer scores calculated from a single backbone structure rather than finding the minimum and scores calculated over the entire ensemble of structures (identical scores were used in all cases, see Methods). When we compared ROC curves and AUC values obtained from predictions made using single protease structures (Figure 3C grey curves shown for 11 structures; Figure 3E grey bars) to model predictions made using the ensemble of crystal structures (black curves and black bars), we again observed consistently poorer model performance. This suggests incorporating backbone variability by making predictions over an ensemble of backbone structures can be important for correctly predicting protease mutational tolerance.
HIV protease has been particularly well characterized and hundreds of solved crystal structures exist within the Protein Data Bank. Many of these protease crystal structures originally contained point mutations. Thus the improvement seen in predicting mutational tolerance using the ensemble of protease crystal structures could have been influenced by the original presence of these point mutations (while all mutations were computationally reverted to the consensus sequence at the start of our simulations, any backbone structural changes present in the mutated structure remained, see Methods). Furthermore, other proteins may not have comparably large ensembles of experimental structures and thus the method we describe here could, for this reason, be less applicable.
To address both these issues, we next tested whether accurate predictions of mutational tolerance could be made using a computationally generated, rather than an experimentally determined, ensemble of protease backbones. To ensure that the computational ensemble did not contain “structural memory" of point mutations present in the original crystallographic ensemble, we selected as templates 11 protease crystal structures that did not contain mutations from the consensus sequence. From each of the 11 templates, we used a computational method termed “backrub" to generate an ensemble of 400 protease structures [20], [30] with “near-native" backbone conformations (ensemble members had Cα RMSDs of 0.2 to 0.6 Å to the original starting template structure). We then repeated the calculations of mutational tolerance using each computational ensemble as described for the ensemble of experimentally determined structures.
Remarkably, the same crystal structures that had resulted in poorer ROC curves and AUC values when considered as single structures (11 grey curves and grey bar, Figure 3C and 3E) now showed improved results when the structures were used as starting templates for a computationally generated ensemble (11 orange curves and orange bar, Figure 3D and 3E). Furthermore, ROC curves and AUC values for predictions made using computationally generated ensembles were almost identical to those originally made using the ensemble of experimentally determined crystal structures (compare black and orange curves and bars, Figure 3D and 3E). Therefore, while increasing the computational cost linearly with the number of backbones (see Text S1 for estimates on computational time), backbone ensemble calculations can result in considerably better prediction than when using only a single backbone.
To gain insight into how structural flexibility might have resulted in improved predictions of mutational tolerance, we examined the model predictions in more detail. Figure 5 shows two mutations as examples where backbone flexibility appeared to be crucial for correctly predicting tolerance to mutations observed in the Stanford database. When mutations A71V and I93L were individually modeled onto HIV-1 protease fixed backbone structures crystallized in the absence of any mutation, large to moderate clashes resulted (Figure 5, left). In each case, the clashes could be resolved when the same mutation was modeled onto a backbone computationally generated from an unmutated starting structure using the backrub method (Figure 5, middle). The mutations modeled onto the computationally generated backbones had structures and ERESFold scores close to those seen in experimentally determined structures that had originally contained the mutation (Figure 5, right). These results suggest that backrub ensembles, even though they were generated in the absence of mutations, can capture sufficient protein conformational variability to accommodate amino acid changes [20], [30].
Figure S3 confirms that the mutations 71V and 93L were predicted as tolerated when modeled onto either experimental or backrub ensembles, but never when modeled onto a single fixed backbone of the consensus sequence (similar behavior was also observed for mutations 24I and 77I, Figure S3). We note that a few mutations within the Stanford database sequences that had been poorly predicted when using the ensemble of experimentally determined structures were found to be tolerated when using computationally generated ensembles (e.g. 33F and 12S, see Figure S3).
To test the applicability of our model, we chose another protein system, the HIV-1 reverse transcriptase. This RNA-dependent DNA polymerase transcribes the single-stranded retroviral RNA genome into a double-stranded proviral DNA. Reverse transcriptase is a heterodimer built of the p66 subunit (560 residues) and the p51 subunit, which has an identical sequence to the first 440 residues of p66. The unique C-terminal part of the p66 subunit comprises an RNaseH domain. Similar to the protease system, many reverse transcriptase structures have been determined, many pre/post drug treatment mutations are catalogued in the Stanford database [15] and mutational tolerance prediction can be made using both fold and dimer stability as functional constraints. Nonetheless, the reverse transcriptase model has several limitations. There are fewer crystal structures than for protease (see Table S1) and there are stretches of sequence with missing density in these structures. The substrates of reverse transcriptase are DNA/RNA hybrid molecules, for which interaction energy calculations are less established than for protein-protein interactions. We therefore did not consider reverse transcriptase residues in the interface with nucleic acids. In addition, model predictions could not be verified for the RNaseH domain, since mutational data are too sparse in this protein segment (see Methods). In sum, we evaluated our analysis based on an ensemble of 91 structures and 656 of the 1,000 residues in the reverse transcriptase heterodimer (still a much larger number of residues than in HIV protease; note that while some residues are excluded from the analysis, all protein residues present in the structures were used in the calculations). We repeated all mutational tolerance calculations as described for protease, calculating ERESFold and ERESDimer scores for every structure within the ensemble (ERESPeptide could not be computed for reverse transcriptase that does not have peptide substrates). Otherwise the model parameters determined for protease were used unchanged for reverse transcriptase. Mutations were made simultaneously for every shared sequence position in the p51 and p66 subunits, while the p66-specific RNAseH domain sites were mutated only on the p66 subunit. The detailed results for the modeled versus observed mutational tolerance for reverse transcriptase are given in Figure S4 (neutral model) and Figure S5 (selective model).
As was observed for protease, a sizeable number of reverse transcriptase sites have low mutational tolerance, and a rather small number of sites were frequently mutated (224 and 17 sites, see Figure 6A). The neutral model correctly identified the majority of these sites (70% and 53% for the rarely mutated and frequently mutated sites, respectively). In contrast, the performance of the selective model for reverse transcriptase was poorer: 130/201 sites that rarely mutate were correctly predicted, and 14/31 sites that frequently mutate (see Figure 6B). Under-predictions were seen at five out of 17 sites (177D, 211R, 329I, 334Q and 376A), while over-predictions were seen for 42 out of 224 (19%) sites. Thirteen out of these 42 sites are exposed polar residues (as for protease, Rosetta performed poorly at predicting at polar exposed sites).
Many sites with over-predicted mutational tolerance are in protein segments that rarely mutate due to constraints likely not captured in our prediction scheme. For example, 17 of the over-predicted sites are located in the Palm domain (positions 86–119 and 151–244). Within it, sites 88W, 111V, 113D, 116F, 182Q and 233E were shown to be involved in primer loading [31]. Another over-predicted stretch of residues spans positions 216 to 243 (the “primer grip") that is involved in positioning the primer's terminus [32]. This region is almost invariant in the neutral data and is known to mutate after drug treatment (as shown in the selective settings – both in the database and the modeled data). An additional over-predicted segment spans position 251 to 271 (the ‘helix clamp’) that is conserved among other nucleic acid polymerases [33]. Several residues within these regions were not directly in contact with nucleic acid in any of the available structures but were previously shown to be important for the catalytic cycle of reverse transcription, providing a possible explanation for the over-predictions.
As with HIV-1 protease, we calculated ROC curves and AUC values for the reverse transcriptase model predictions to quantify overall performance (Figure 7). The ROC curves show that the computational model correctly ranked many mutations tolerated by HIV-1 reverse transcriptase. AUC values are generally slightly lower for reverse transcriptase than for protease, but exceed 80% (black bars, Figure 7C). In accordance with the results obtained for protease, predictions of mutational tolerance made using any single reverse transcriptase structure were worse than using an ensemble of experimentally determined structures or backrub ensembles computationally generated from a single template structure (grey, black and orange curves and bars in Figure 7A–C). In conclusion, although mutational tolerance predictions for reverse transcriptase were less accurate than for protease, the results still demonstrate reasonable agreement with mutations observed in the database. The application of the model to reverse transcriptase also confirms the notion that using either ensembles of experimentally determined or computationally generated structures improve predictions over using single structures.
Finally, we examined the ability of the selective model to predict DRMs for reverse transcriptase. Nine of the 31 published DRMs were recapitulated by the model. Interestingly, the K103Q DRM, which was not present in the Stanford database (at the time of the database download), was correctly predicted by the model (see Figure 7D). The overall performance of DRM prediction by the selective model was weaker than for protease. Several factors may account for this discrepancy: The number of documented DRMs is lower for reverse transcriptase, and DRM positions are not as evenly distributed over the protein structure as in protease (Figure S6). Hence, down-weighting constraints overall leads to many over-predictions in reverse transcriptase. Moreover, the weights of the model were parameterized using the HIV-1 protease data. In addition, the NNRTI inhibitor binding site is not as close to the dimer interface (where constraints are weakened in the selective model) as in protease. Nonetheless, as for protease, some predictions by the selective model might represent resistance mutations yet to be discovered.
We have shown that an all-atom, computational model that incorporates structural and functional constraints on mutational tolerance is able to predict a considerable fraction of the observed tolerated sequence space of HIV-1 protease and reverse transcriptase. The model uses a previously published and established energy function that, importantly, was complemented with incorporation of protein backbone flexibility. The Rosetta energy function has been shown previously [34] to perform comparably to other methods [35], [36], [37] in predicting changes in protein stability upon point mutations. The model parameters in Equation 1 were optimized using data on HIV-protease by adjusting the relative frequencies with which the various structural and functional constraints operated (as well as a solubility parameter, see Methods). However, there was no explicit parameterization with respect to the actual identities of amino acids selected as tolerated at each protease site, which we use as an evaluation metric in the ROC analysis. Furthermore, after the model parameters were optimized for protease, identical parameters were applied to the much larger protein reverse transcriptase, albeit with a moderate reduction in performance.
The comparison of the predictions of the model with the database mutations, as well as the comparison between the two model systems, reveals both strengths and weaknesses of our approach. Considering multiple structures improved predictions for both proteins. However, developing parameters for one system, such as protease, may make the predictions less applicable to other proteins. This is especially obvious for the dimer and peptide binding constraints used here. These constraints work well for modeling selective pressure in the presence of inhibitors for protease (where inhibitors bind in the peptide binding site, located in the dimer interface), but not for reverse transcriptase. Moreover, the model is likely to fail for regions where important constraints were not included, such as interaction interfaces of reverse transcriptase with nucleic acids. Similarly, HIV protease and reverse transcriptase may bind other partners, as indicated by recent large-scale mapping of interactions of HIV proteins with factors in the human host [38]. If these interactions are not adequately modeled, the sequence constraints they impose may not be correctly captured. Nevertheless, there is overall encouraging agreement between observed and predicted mutational tolerance, suggesting that models similar to the one we developed could be applicable to other proteins.
Errors in the Rosetta energy function are likely responsible for both model over- and under-predictions. This behavior is particularly apparent for mutations to and from polar residues, due to the difficulty of modeling the balance of electrostatic interactions and solvation. Furthermore, the model considers only single, independent mutations, whereas sites that were under-predicted may require the presence of additional compensatory mutations. Correlated mutations occur with both HIV protease and reverse transcriptase drug resistance mutations [39]. For example, the protease mutations 30N and 88D are known to co-vary and while the model predicts mutational tolerance towards both of these mutations, the frequencies predicted are less than seen in the database sequences for each mutation. In such cases, modeling the effects of double mutations may improve predictions. While such a double-mutant analysis is challenging, repeating the selective model calculations using a finite set of double mutations (see Text S1) resulted in predicted increases of individual mutation frequencies at 10 of the previously under-predicted protease sites (10L, 20K, 33L, 36M, 41R, 57R, 64I, 82V, 84I, and 93I; Figure S7).
A strength of our approach is the improved prediction accuracy when using backbone ensembles. This is observed for both protein model systems. Our results underline that incorporating this backbone variability is important for predicting mutational tolerance, particularly when using all-atom force fields that model atomic packing interactions sensitive to precise details and small steric clashes. We show that predictions made from a computationally generated ensemble can be just as accurate as predictions using an ensemble of experimentally determined structures. This finding is notable, as the conformational variation within the ensemble of experimentally determined backbones included changes induced by substrate and inhibitor binding, as well as structural changes in response to single and multiple mutations. In contrast, the computational ensembles were generated from single structures with few or no mutations, to avoid such “structural memory".
The use of crystallographic ensembles to model protein conformational flexibility has been described and shown to be consistent with molecular dynamics simulations [21], elastic network models [40] and protein dynamics detected using nuclear magnetic resonance [41]. In this work, structural variation calculated over the computationally generated backrub ensembles is similar to variation calculated over the ensemble of experimentally determined structures (Figure S8). Previous studies on a variety of other protein systems have found that computationally generated backrub ensembles improve predictions of protein dynamics [19], [42], conformations of single point mutations [20] and sequence diversity in proteins and protein-protein interfaces [18], [22]. Taken together with these results, it is plausible that backrub ensembles sample a significant portion of conformations accessible to proteins.
While the magnitude of stabilizing and destabilizing energetic trade-offs predicted by the model for each individual mutation is only an estimate, the patterns of compensatory effects and functional tradeoffs may nevertheless be informative. There is widespread evidence for the general trend of mutations that confer new functions to destabilize a protein [43], [44]. In contrast, our analysis shows that stabilizing mutations are in fact overrepresented in DRMs relative to all possible mutations in protease that are reachable by a single nucleotide change. Therefore, DRMs arising in viral populations may not conform to the classical definition of ‘new function’ mutations (e.g. in [45]). Instead, a significant fraction of DRMs may belong to the subset of mutations described in [45] that increase protein stability. This is consistent with the hypothesis that several DRMs function to compensate for other co-existing destabilizing mutations that directly affect inhibitor binding.
For HIV-1 protease, the selective model suggests hypotheses about the effects of specific mutations on the stability of the protease fold, dimer interface and substrate binding that, in some cases, can be confirmed using existing experimental data. For example, the model predicted large substrate destabilization effects for I47A and V82A/F/T. These mutations are known to display increased replication in viruses with mutations in either the Nucleocapsid/p1 or p1/p6 cleavage sites [28], [46], [47]. This finding suggests that cleavage site mutations may compensate destabilizing effects at the substrate-binding interface predicted by the model. In another example, incorporating A71V (predicted by the model to stabilize the protease fold) into double and triple mutants with reduced replicative ability (containing either 36I/54V or 36I/54V/82T, all predicted to be destabilizing) has been shown to improve replication to better than wild-type levels [48].
The approach we present here differs from other studies that have characterized the structural [49], [50], functional [51], [52], [53], [54], [55] and energetic effects [55] of HIV-1 protease mutations on inhibitor binding (and similar studies applied on reverse transcriptase [56], [57]). Instead, we make predictions of the mutations tolerated by HIV-1 protease and reverse transcriptase structure and function without explicitly considering inhibitor binding. A possible limitation of this approach is that mutations at sites that directly interact with a protease inhibitor may be under-predicted even in the selective model, since no benefit is given to mutations that specifically destabilize protein-inhibitor interactions. On the other hand, our model should have the advantage of predicting protease mutational tolerance prior to knowledge of any specific inhibitor structures. While some drug resistance mutations are shared among inhibitors, new drug resistance mutations have appeared with the introduction of each clinical drug. The model we present here could be useful in the prediction of yet undiscovered resistance mutations by suggesting mutations structurally and functionally tolerated that would be free to contribute towards the destabilization of new inhibitors.
In conclusion, our results, along with the observation that RosettaDesign simulations using flexible backbone ensembles capture sequence diversity in phage display experiments [18], [22], [58] and protein families [19], suggest that the model presented here for protease and reverse transcriptase may be applicable to other proteins. Moreover, while we initially validated our model using experimentally determined structures of HIV-1 protease and reverse transcriptase solved under a variety of experimental conditions, we have also shown that computationally generated structural variability from a single structure can produce comparable model accuracy. Thus, the model we present here could be extended to predict mutational tolerance in other systems (where only a single structure may be available) to yield insights into the relationship of structure, function, and tolerated sequence space. In practical terms, prediction of the nearly neutral space of sequences consistent with a given structure and function may be exploited in the experimental design and construction of proteins with modified and new properties.
The following HIV-1 protease and reverse transcriptase sequences for subtype B, defined by the Stanford HIV database as the consensus sequence, were used throughout this work. Sequence positions in lower case were excluded from model predictions (see below).
Protease:
PQITLWQRPLVTIKIGGQLKEALLdTGADDTVLEEMNLPGRWKPKMIGGIGGFIKVRYDQILIEIcGHKAIGTVLVGPTPVNIIGRNLLTQIGcTLNF
Reverse transcriptase: pispIETVPVKLKPGMDGPKVkQwpltEEkIKALVEIcTEMEKEGKISKIGPENPYNTPVfAiKKkDSTKWRKlVdFrELnKRTQDFWevqlGiPHPAGLKKKKSVTVLdVGDAyFSVPLDKDFRKYTAFTIPSINNETPGIRYQYNVLPqGWkGSpAIFQSSMTKILEPFRKQNPDIVIYQymddLYVGSDLEIGQHRTKIEELRQHLLRWGFTtpdkkhqkeppflwmGYELHPDKWTVQPIVLPEKDSWTVnDIqkLVGkLnwASQIYAGIKvkQLckLLrGtkAlTEVIPLTEEAELELAENREILKEPVHGVYYDPSKDLIAEIQKQGQGQWTYQIYQEPFKNLKTGkyaRMrGahTNDVKQLTEAVQkIATESIVIWGKTPKFkLPIQkeTWEawwteywqatwipewefvntpplvklwyqlekepivgaetfyvdgaanretklgkagyvtdrgrqkvvsltdttnqktelqaihlalqdsglevnivtdsqyalgiiqaqpdkseselvsqiieqlikkekvylawvpahkgiggneqvdklvsagirkvl
In total, we include 96 out of 99 protease sites, and 328 out of 560 reverse transcriptase sites (note that our algorithm was applied on two chains of each of these proteins, meaning 192 amino acids of protease and 656 amino acids for reverse transcriptase). Not included in the analysis were known catalytic site residues in both proteins (1 in protease and 3 in reverse transcriptase [31]). We also excluded all mutations to and from cysteine as modeling the effect of these mutations can be complicated by disulfide bond formation. For protease, this included mutations at the two naturally occurring cysteine sites (67C and 95C) which were relatively rare and never occurred to any other amino acid type with a frequency >1%. Five mutations to cysteine in the presence of inhibitors were also excluded: L63C 2.8%, N37C 1.3%, G73C 0.9%, I84C 0.4% and V82C 0.1%. For reverse transcriptase, we excluded two naturally occurring cysteines (C38 and C280) which were never documented to mutate to any other amino acid type with a frequency >1%. Eleven mutations to cysteine (pre-drug treatment) were also excluded, only one of which occurred frequently: A33C 0.9%, W88C 0.3%, S162C 20.9%, Y181C 0.4%, T215C 0.3%, S251C 0.2%, Q334C 0.4%, G335C 0.3%, F346 0.4%, A376C 0.1% and S379C 1.9%. Additional residues omitted from the reverse transcriptase analysis are 161 residues with insufficient information on mutational tolerance in the Stanford HIV database (Figure S9), residues in contact with the RNA substrate (46 positions) and regions of missing densities in the available crystal structures (20 positions).
Frequencies of mutated amino acids in protease and reverse transcriptase observed in patients were obtained online from the Stanford HIV drug resistance database (Genotype-Treatment Correlations/Treatment Profiles, see http://hivdb.stanford.edu/cgi-bin/PRMutSummary.cgi and http://hivdb.stanford.edu/cgi-bin/RTMutSummary.cgi). The settings were as follows - reference profile: subtype B untreated, exclude single occurrences: yes, include mixture: no, one mutation per person. We also compiled a sequence set after inhibitor treatment by using similar settings for protease (# of PIs: 1–9, in addition to the profile settings described above) and reverse transcriptase (#NRTI: 1–7, #NNRTI: 1–4).
For protease, 262 dimeric crystal structures (see Table S1) and one NMR minimized model (PDB code: 1BVG) were used as the ensemble of experimentally determined structures. The majority of structures contained 1–7 mutations (223 out of 263). The crystallographic resolution for structures in the ensemble is within the range of 0.84 to 3.1 Angstroms. For the ensemble of experimentally determined structures of reverse transcriptase, we compiled 91 crystal structures (see Table S1) that have a resolution within the range of 1.8 to 3.2 Angstroms and contain 13–24 mutations from the consensus sequence.
All computational simulations were performed using the Rosetta energy function [17], [59], which is dominated by atomic packing, attractive and repulsive Lennard-Jones interactions, an orientation-dependent hydrogen bonding term [59], and an implicit solvation model [60]. The simulations consisted of sampling and scoring side-chains (taken from a rotamer library that included the native amino acid conformations taken from the starting structures and additional rotamers around the chi1 and chi2 side-chain torsion angles [61]) using a Monte-Carlo simulated annealing optimization protocol (“repacking") as described in [62].
In preparation for calculations of fold and dimer stability, all water molecules, heteroatoms, DNA/RNA nucleotides, bound inhibitors or substrates and hydrogens present in the original PDB structures were removed, and hydrogen atoms were added as previously described [59]. An initial round of minimization of the side-chain torsion angles was performed using the Rosetta energy function, keeping all amino acid identities and backbone coordinates fixed. After this initial minimization, all structures containing mutations from the consensus HIV-1 subtype B sequence (see above) were computationally reverted to the consensus sequence. All side-chains that had at least one atom within 4 Å of any mutated residue were repacked and the structures were side-chain minimized a second time.
In preparation for calculations of peptide substrate binding affinity to HIV-1 protease, a protocol identical to that described above, except leaving all bound substrates present, was used for the 19 crystallographic and model structures listed in Table S2. 16 dimeric, crystallographic structures with one of 7 endogenous peptide substrates were used for calculations of substrate binding energy (see Table S2). Structural models for each of the three peptide substrate sequences without experimentally determined structures (CTLNF-PISPI, PQITL-WQRPL and VSFNF-PQITL) were generated by computationally threading each peptide sequence onto each of the known 16 crystallographic structures (sequence positions for which there was missing crystallographic density on any of the 16 peptide template structures were omitted) and performing side-chain minimization and repacking as described above. We selected the structural template for which the resulting Rosetta interface energy (the sum of Rosetta energy terms over all pair-wise interactions between residues l and m, where residue l was located on HIV-1 protease and residue m was located on the bound peptide) was the lowest. 1MT9.pdb was found to be the best template for both peptides CTLNF-PISPI and PQITL-WQRPL, while 1F7A.pdb was selected as the optimal template for VSFNF-PQITL. Peptide interface Rosetta scores for the resulting three models (−18.3 to −25) were within the range observed for 16 crystallographic structures obtained with bound peptides (−18.5 to −33).
Estimates of the effect of mutations on fold stability (ERESFold), dimer interface stability (ERESDimer) and peptide binding (ERESPeptide) were calculated using the RosettaDesign energy function and the computational model outlined in Results. The contribution towards fold stability (ERESFold) of each mutated residue was estimated by recording the sum of inter- and intra-residue Rosetta energy terms (see Text S1 for a detailed explanation of the energy terms). The contribution of each mutated residue to stability of the dimer interface (ERESDimer) was estimated by calculating only inter-chain pair-wise Rosetta energy function contributions between the mutated residue and neighboring residues on the opposite dimer chain (see Text S1 for details). The contribution of the mutated residue towards binding interactions with endogenous substrate peptides (ERESPeptide) was calculated by summing only over pair-wise energy function terms between the mutated residue and the residues of each of 10 bound substrates (as described for the dimer interface).
Each mutation was modeled simultaneously on both chains of HIV-1 protease or reverse transcriptase, and ERES scores from both chains were summed. Note that reverse transcriptase is a heterodimer built of the p66 subunit (560 residues) and the p51 subunit, composed of the first 440 residues of p66 (the sequence-identical parts in p66 and p51 adopt different relative orientations of the constituent domains; if each domain is superimposed separately, the average RMSD is 1.18 Angstrom). The unique C-terminal part of the p66 subunit is the RNaseH domain.
For HIV-1 protease, three of the simulations modeling peptide binding required a portion of the HIV-1 protease sequence itself to be a substrate (this occurred for the transframe region and HIV-1 protease cleave site (TF-PR), the HIV-1 protease and reverse-transcriptase cleavage site (PR-RT) and the auto-proteolysis cleavage site (AutoP); see Table S2). For these simulations, each relevant mutation was modeled simultaneously onto both chains of the HIV-1 protease scaffold as well as onto the peptide backbone.
Optimal values for the six model parameters (WFold, WDimer, WPeptide, FavorNative, FavorPolar and PenaltyPolar→Hydrophobic) were selected using a grid search (see Table S4 for values used) and computing predicted amino acid frequencies over 96 (the catalytic aspartate D25 and cysteine residues C67 and C95 were excluded) HIV protease sites using the minimum ERES scores calculated from the ensemble of experimentally determined structures. For each combination of parameters, each of the 96 HIV-1 residue sites was computationally classified as having either low (1–5%), medium (5–20%) or high (>20%) mutational frequency and the number of residue sites correctly matching the experimentally observed mutational frequency bin was calculated. The percentage of sites correctly determined for each bin was then averaged and used to determine a parameter set for both the “neutral" and “selective" computational models. Both the neutral and selective model parameters were applied unchanged to reverse transcriptase.
Computational ensembles of “near-native" backbones were generated starting from one of 11 crystallographic structures of the HIV-1 protease consensus sequence (1A8G, 1EBY, 1HXW, 1IZH, 1PRO, 1SBG, 1VIJ, 1VIK, 4PHV, 5HVP and 9HVP) and one of five structures of reverse transcriptase (1HNI, 1HPZ, 1IKX, 2B6A and 2BAN) by using the previously described backrub protocol [20], [22]. While the protease structures were selected to have the consensus sequence, this was not possible for reverse transcriptase, as all structures contained at least 13 sequence changes from the consensus. For reverse transcriptase, we therefore chose structures that had among the lowest number of mutations (15 to 22) and in addition did not have any missing backbone density. The backrub protocol consisted of repeatedly selecting Cα atoms of two residues (separated by 1–10 intervening residues), performing a rigid body rotation of the selected protein segment (of up to 40 degrees), optimizing the location of related Cβ and hydrogen atoms and accepting or rejecting the backbone move based on the Rosetta energy function and the Monte Carlo Metropolis criterion. Using the atomic coordinates of each crystallographic structure above as a starting conformation, 100 independent backrub simulations were run at two separate Monte Carlo temperatures (kT = 0.6 and kT = 1.2) for 10,000 moves per simulation. At each temperature, the lowest energy conformation sampled as well as the last conformation accepted during each simulation were saved and used to generate a computational ensemble of 400 backbone conformations per starting crystallographic structure.
The conformational fluctuation within the generated backrub ensembles, compared to the ensemble of experimentally determined structures, is shown in Figure S10. For two structures of HIV-1 protease (1A8G and 1IZH), ensembles of varying sizes (50 to 1000 structures consisting of last or low energy conformations randomly selected from simulations run at the two Monte Carlo temperatures given above) were systematically tested. It was determined that ensemble sizes of 100 structures or greater gave essentially identical results to the ensemble of experimentally determined protease structures with respect to the area under the curve.
ROC curves were computed for each of 1,728 (protease) and 5,904 (reverse transcriptase) possible mutations (18 amino acid types, excluding the native amino acid residue and cysteine, allowable at 96 sites in protease and 328 sites in reverse transcriptase). Residues that were omitted from the ROC analysis are given above. True positive rates (TPR) and false positive rates (FPR) of mutation recovery were calculated by using the parameter values determined above for the “neutral" and “selective" computational models (Table S4) and considering all mutations computationally predicted to occur at frequencies greater than or equal to varying model cutoff values of 30% to 0.00001%. True positives were defined as mutations occurring within the database at >1%. AUC values were calculated for each ROC curve by implementing the trapezoid method.
TPR and FPR rates were also calculated for the set of mutations one nucleotide mutation away from the native codon (Table S3) and for the set of all mutations to amino acid types chemically similar to the native amino acid type (chemically similar groupings were as follows: (A,G,P), (D,E,N,Q), (F,W,Y), (L,I,V,M), (R,K,H), (S,T).
|
10.1371/journal.pgen.1008253 | A Cyclin A—Myb-MuvB—Aurora B network regulates the choice between mitotic cycles and polyploid endoreplication cycles | Endoreplication is a cell cycle variant that entails cell growth and periodic genome duplication without cell division, and results in large, polyploid cells. Cells switch from mitotic cycles to endoreplication cycles during development, and also in response to conditional stimuli during wound healing, regeneration, aging, and cancer. In this study, we use integrated approaches in Drosophila to determine how mitotic cycles are remodeled into endoreplication cycles, and how similar this remodeling is between induced and developmental endoreplicating cells (iECs and devECs). Our evidence suggests that Cyclin A / CDK directly activates the Myb-MuvB (MMB) complex to induce transcription of a battery of genes required for mitosis, and that repression of CDK activity dampens this MMB mitotic transcriptome to promote endoreplication in both iECs and devECs. iECs and devECs differed, however, in that devECs had reduced expression of E2F1-dependent genes that function in S phase, whereas repression of the MMB transcriptome in iECs was sufficient to induce endoreplication without a reduction in S phase gene expression. Among the MMB regulated genes, knockdown of AurB protein and other subunits of the chromosomal passenger complex (CPC) induced endoreplication, as did knockdown of CPC-regulated cytokinetic, but not kinetochore, proteins. Together, our results indicate that the status of a CycA—Myb-MuvB—AurB network determines the decision to commit to mitosis or switch to endoreplication in both iECs and devECs, and suggest that regulation of different steps of this network may explain the known diversity of polyploid cycle types in development and disease.
| Endoreplication is a cell cycle variant that entails cell growth and periodic genome duplication without cell division, and results in large, polyploid cells. Cells switch from mitotic division cycles to endoreplication cycles during development, and also in response to conditional stimuli during wound healing, regeneration, aging, and cancer. Much remains unknown, however, about how mitotic cycles are remodeled into endoreplication cycles, and how similar this remodeling is between induced and developmental endoreplicating cells (iECs and devECs). In the present work, we define a Cyclin A regulated mitotic network in Drosophila whose downregulation promotes the switch from mitotic cycles to endoreplication cycles in both iECs and devECS. Repression of this network in iECs was sufficient to induce endoreplication without reduced expression of E2F-regulated S phase genes that is common among devECs in both flies and mice. Knockdown of downstream cytokinetic proteins, but not kinetochore proteins, were sufficient to induce different types of endoreplication. Altogether our results define a CycA—Myb—AurB network as a key determinant of alternative cell cycles, and provide insight into the regulation of a diversity of polyploid cycle types in development and disease.
| Endoreplication is a common cell cycle variant that entails periodic genome duplication without cell division and results in large polyploid cells [1]. Two variations on endoreplication are the endocycle, a repeated G/S cycle that completely skips mitosis, and endomitosis, wherein cells enter but do not complete mitosis and / or cytokinesis before duplicating their genome again [2]. In a wide array of organisms, specific cell types switch from mitotic cycles to endoreplication cycles as part of normal tissue growth during development [1, 3]. Cells also can switch to endoreplication in response to conditional inputs, for example during wound healing, tissue regeneration, aging, and cancer [1, 4–6]. It is still not fully understood, however, how the cell cycle is remodeled when cells switch from mitotic cycles to endoreplication.
There are common themes across plants and animals for how cells switch to endoreplication during development. One common theme is that developmental signaling pathways induce endoreplication by inhibiting the mitotic cyclin dependent kinase 1 (CDK1). After CDK1 activity is repressed, repeated G / S cell cycle phases are controlled by alternating activity of the ubiquitin ligase APC/CCDH1 and Cyclin E / CDK2 [1]. Work in Drosophila has defined mechanisms by which APC/CCDH1 and CycE / Cdk2 regulate G / S progression, and ensure that the genome is duplicated only once per cycle [7–12]. Despite these conserved themes, how endoreplication is regulated can vary among organisms, as well as tissues within an organism. These variations include the identity of the signaling pathways that induce endoreplication, the mechanism of CDK1 inhibition, and the downstream effects on cell cycle remodeling into either an endomitotic cycle (partial mitosis) or endocycle (skip mitosis) [1, 7]. In many cases, however, the identity of the developmental signals and the molecular mechanisms of cell cycle remodeling are not known.
To gain insight into the regulation of variant polyploid cell cycles, we had previously used two-color microarrays to compare the transcriptomes of endocycling and mitotic cycling cells in Drosophila tissues [13]. We found that endocycling cells of larval fat body and salivary gland have dampened expression of genes that are normally induced by E2F1, a surprising result for these highly polyploid cells given that many of these genes are required for DNA synthesis. Nonetheless, subsequent studies showed that the expression of the E2F-regulated mouse orthologs of these genes is reduced in endoreplicating cells of mouse liver, megakaryocytes, and trophoblast giant cells [10, 14, 15]. Drosophila endocycling cells also had dampened expression of genes regulated by the Myb transcription factor, the ortholog of the human B-Myb oncogene (MYBL2) [13, 16]. Myb is part of a larger complex called Myb-MuvB (MMB), whose subunit composition and functions are mostly conserved from flies to humans [17–21]. One conserved function of the MMB is the induction of periodic transcription of genes that are required for mitosis and cytokinesis [20, 22–26]. It was these mitotic and cytokinetic genes whose expression was dampened in Drosophila endocycles, suggesting that this repressed MMB transcriptome may promote the switch to endocycles that skip mitosis. It is not known, however, how E2F1 and Myb activity are repressed during endocycles, nor which of the downregulated genes are key for the remodeling of mitotic cycles into endocycles.
In addition to endoreplication during development, there are a growing number of examples of cells switching to endoreplication cycles in response to conditional stresses and environmental inputs [1, 5, 6]. We will call these induced endoreplicating cells (iECs) to distinguish them from developmental endoreplicating cells (devECs). For example, iECs contribute to tissue regeneration after injury in flies, mice, humans, and the zebrafish heart, and evidence suggests that a transient switch to endoreplication contributes to genome instability in cancer [4, 6, 27–33]. Cardiovascular hypertension stress also promotes an endoreplication that increases the size and ploidy of heart muscle cells, and this hypertrophy contributes to cardiac disease [29, 34, 35]. It remains little understood how similar the cell cycles of iECs are to devECs.
Similar to the developmental repression of CDK1 activity to promote endocycles, we and others had previously shown that experimental inhibition of CDK1 activity is sufficient to induce endoreplication in flies, mouse, and human cells [28, 36–41]. These experimental iECs in Drosophila are similar to devECs in that they skip mitosis, have oscillating CycE / Cdk2 activity, periodically duplicate their genome during G / S cycles, and repress the apoptotic response to genotoxic stress [13, 36, 42, 43]. In this study, we use these experimental iECs to determine how the cell cycle is remodeled when cells switch from mitotic cycles to endoreplication cycles, and how similar this remodeling is between iECs and devECs. Our findings indicate that the status of a CycA—Myb—AurB network determines the choice between mitotic cycles and endoreplication cycles in both iECs and devECs.
We sought to understand how remodeling of the cell cycle program determines the switch from mitotic cycles to endoreplication cycles, and how similar this remodeling is between iECs and devECs. One challenge to addressing these questions has been obtaining pure populations of cells in different cell cycles, especially for iECs that occur in tissues among a mixed population of cells. As a model for iECs, therefore, we experimentally induced Drosophila S2 cells in culture into endoreplication cycles by knocking down Cyclin A (CycA), which is sufficient to induce endocycles [36, 38, 44]. In Drosophila, CycA / CDK2 is not required for S phase, and it is believed that knockdown of CycA promotes endocycles by inhibiting CycA / CDK1 activity required for mitosis, analogous to the common mechanism of CDK1 inhibition during developmental endocycles in multiple organisms [45]. S2 cells were treated with CycA double-stranded RNA (dsRNA), and compared to a negative control population of mitotic cycling S2 cells that were treated in parallel with GFP dsRNA. Importantly, this permitted a comparison of canonical and variant cell cycles in a pure population of cells of the same cell type. Flow profiling 96 hours after treatment with CycA dsRNA indicated that more than 50% of cells had a polyploid DNA content of ≥ 8C, and a commensurate reduced fraction of cells with diploid 2C and 4C DNA contents (Fig 1A and 1B). These cells had genome doublings of 8C, 16C, and 32C that were multiples of the diploid DNA content, suggesting that they were duplicating their genomes through repeated G / S endocycles (Fig 1A and 1B). In contrast, knockdown of the mitotic Cyclin B (CycB) did not induce cells to endoreplicate, perhaps because of functional redundancy with CycB3 (S1 Fig) [46, 47]. These results confirm previous findings that inhibition of CDK activity through knockdown of CycA is sufficient to induce endoreplication in S2 cells (hereafter CycA dsRNA iEC) [36, 44].
We had previously shown that endocycling cells (G / S cycle) of the Drosophila larval salivary gland and fat body have dampened expression of genes that are normally induced by E2F1 and the MMB transcription factors [13]. To determine if this change in transcriptome signature also occurs in CycA dsRNA iECs, we analyzed the expression of several candidate genes whose expression is induced by E2F1 or MMB. RT-qPCR results indicated that CycA dsRNA iECs had reduced expression of the Myb subunit of the MMB and two genes that are positively regulated by the MMB and essential for mitosis (aurora B and polo) (Fig 1C). In contrast, the expression of three genes normally induced by E2F1 at G1 / S (Cyclin E, PCNA, and dup (fly Cdt1) were similar between CycA dsRNA iECs and mitotic cycling cells (Fig 1C). These results suggest that CycA dsRNA iECs are similar to developmental endocycling cells (devECs) in that they both have reduced expression of MMB-dependent M phase genes, but they differ in that iECs do not have reduced expression of E2F1-dependent S phase genes.
Although CycA dsRNA iECs had lower expression of two MMB-induced genes that are required for mitosis, it was unclear whether dampened MMB activity contributed to the switch to endoreplication. To address this question, we knocked down expression of the Myb subunit of the MMB, which is required to induce the expression of genes for mitosis and cytokinesis [22–24]. Knockdown of Myb inhibited cell proliferation, and resulted in an increase in polyploid DNA content that was similar to that of CycA dsRNA iECs (Fig 2A and 2B, S2 Fig). We then used fluorescence microscopy to further evaluate ploidy and cell cycle in CycA and Myb knockdown cells. S phase cells were detected by incubating in the nucleotide analog EdU for two hours followed by fluorescent click-it labeling, M phase cells detected with antibodies against phospho-histone H3 (pH3), and nuclear DNA labeled with DAPI [48–50]. Treatment of cells with either CycA or Myb dsRNA resulted in a similar frequency and size of large polyploid nuclei, indicating that Myb knockdown induced endoreplication (hereafter Myb dsRNA iEC) (Fig 2C–2F). There was a higher fraction of multinucleate Myb dsRNA iECs (~15%) than CycA dsRNA iECs (~8%), suggesting that Myb knockdown results in a somewhat larger fraction of endomitotic cells than does CycA knockdown (Fig 2G). Approximately 30% of CycA dsRNA iECs and Myb dsRNA iECs incorporated EdU, an S phase fraction that was similar in both mononucleate and multinucleate populations, consistent with periodic duplications of the genome during both endocycles and endomitotic cycles (Fig 2H). Despite this evidence for periodic endoreplication, the fraction of total cells with mitotic PH3 labeling was not decreased after CycA knockdown (~5%), and was slightly increased after Myb knockdown in the mononucleate population (~10%) (Fig 2H). Unlike control mitotic cells, however, the PH3 labeling after CycA and Myb knockdown was diffuse, with little evidence of fully condensed mitotic chromosomes, suggesting that these cells were either arrested or delayed in early mitosis or endomitosis, and are consistent with previous observations of chromosome condensation defects of Myb mutants in vivo [51] (Fig 2C–2E insets). These results indicate that knockdown of Myb is sufficient to induce endoreplication cycles that are similar to those after knockdown of CycA.
The similarity between CycA dsRNA and Myb dsRNA iECs suggested that CycA and Myb may have a functional relationship. It had been shown in human cells that CycA / CDK2 phosphorylates Myb and promotes its activity as transcription factor [52, 53]. These early studies, however, were before the discovery that Myb acts as part of the MMB and the identification of many MMB regulated genes [54, 55]. Moreover, it is not known whether CycA regulation of Myb is conserved in Drosophila. To begin to address this question, we analyzed iECs by Western blotting. The results showed that CycA and Myb dsRNA treatments resulted in the expected lower levels of their respective proteins (Fig 3A). Importantly, both CycA and Myb dsRNA iECs also had greatly reduced levels of CycB protein, consistent with the known requirement of the MMB for transcriptional induction of CycB during mitotic cycles, and further suggesting that CycA knockdown may compromise MMB activity (Fig 3A) [24, 29, 56]. To further address this possibility, we used RT-qPCR to quantify mRNA levels for CycB and other known MMB target genes that function in mitosis or cytokinesis. Knockdown of either CycA or Myb reduced the expression of all these MMB target genes to similar extents (Fig 3B). Knockdown of CycA resulted in reduced Myb mRNA, even though the Western results showed that there was no reduction of Myb protein. This result is consistent with previous reports that the periodic proteolysis of Myb, which normally occurs during mitosis, is absent during endoreplication cycles [57]. In contrast, knockdown of Myb did not reduce levels of either CycA mRNA or protein, suggesting that Myb knockdown is sufficient to induce endoreplication cycles even when CycA protein levels are high (Fig 3A and 3B). These results suggest that CycA complexed with either CDK1 or CDK2, is required for MMB transcriptional activation of M phase genes.
To further evaluate CycA regulation of the MMB, we determined if Myb and CycA physically interact. We used the GAL4 / UAS system to express UAS-CycA with either UAS-Myb-RFP or UAS-RFP in mitotic cycling imaginal discs, immunoprecipitated (IPed) Myb-RFP or RFP with an anti-RFP nanobody, and then blotted for Cyclin A [25, 58]. The results indicated that Myb-RFP, but not RFP alone, co-IPs with CycA (Fig 3C). The IP’ed RFP-Myb protein reproducibly migrated as a cluster of four bands, which could be the result of post-translational modification, although lower molecular weight species specifically recognized by an anti-dsRed antibody suggests some proteolysis had occurred (Fig 3C). In the reciprocal experiment, IP of CycA-HA with HA antibodies co-IPed Myb-RFP but not RFP alone (Fig 3C’). All together, these results are consistent with the hypothesis that during Drosophila mitotic cycles a CycA / CDK complex is directly required for the MMB to induce expression of genes required for M phase, and that in the absence of this activation cells switch to endoreplication cycles.
To further evaluate the relationship between CycA and Myb and gain insight into remodeling of mitotic cycles into endoreplication cycles, we analyzed the global transcriptomes of CycA dsRNA and Myb dsRNA iECs by RNA-Seq. The transcriptome of these two iEC populations were compared to control mitotic cycling S2 cells treated in parallel with GFP dsRNA, all in three biological replicates. Genes were defined as differentially expressed (DE) in iECs if their normalized steady state mRNA levels differed from mitotic cycling cells with a log2 fold change (log2FC) of at least +/- 0.5 and a false discovery rate (FDR) corrected p-value <0.05 [59].
The RNA-Seq results indicated that a switch from mitotic cycles to endoreplication in CycA dsRNA and Myb dsRNA iECs is associated with differential expression of thousands of genes (Fig 4A and 4A’, S1 and S2 Tables). Comparison of the CycA dsRNA and Myb dsRNA iEC transcriptomes revealed that they shared a total of 966 genes that were differentially expressed compared to mitotic cycling controls (698 increased and 268 decreased) (Fig 4B, S3 Table). Permutation testing indicated that this overlap of DE genes was highly statistically significant, with the overlap in upregulated genes being 4.6 fold greater than expected by chance (p<1 x 10−5), and that of downregulated genes being 5.8 fold greater than expected by chance (p<1 x 10−5) (S3A Fig). Analysis of Gene Ontology (GO) biological process categories indicated that the upregulated genes shared by CycA dsRNA iEC and Myb dsRNA iECs were significantly enriched in the categories of immunity, metabolism, and development (q < 5 x 10−4), and that shared down regulated genes also included those for energy metabolism (q ≤ 1 x 10−9) (S4 and S5 Figs, S3 Table) [60].
With respect to cell cycle remodeling, the downregulated genes shared by these two iEC types were significantly enriched for multiple GO categories of mitosis and cytokinesis (q < 1 x 10−9) (Fig 4C, S5 Fig, S4 Table). After removing redundant GO categories, we analyzed the genes from the five most significantly enriched categories. These categories comprise 47 genes with functions in mitosis and cytokinesis that were downregulated by up to several fold in both iEC types (Fig 4C, S4 Table). Georlette and colleagues had previously shown that many of these genes require the Myb subunit of the MMB for their expression in Drosophila Kc cells [24]. The common downregulation of these genes in CycA dsRNA iEC and Myb dsRNA iEC further suggests that CycA is required for the MMB to induce transcription of these mitotic genes, and that downregulation of a subset of the MMB transcriptome in these two iEC types may contribute to the switch from mitotic cycles to endoreplication cycles.
The RNA-seq results, together with our published analysis of devECs, suggested that iECs are similar to devECs in that both have a dampened Myb transcriptome of mitotic genes [13]. However, our previous analysis of devECs used two-color microarrays that had a limited gene set and sensitivity [13]. Therefore, to more fully compare iEC and devEC transcriptomes, we expanded the analysis of devECs with RNA-Seq. Specifically, we used RNA-Seq to compare the transcriptome of endocycling larval salivary glands (SG) to that of mitotic cycling larval brains and discs (B-D) from early third instar larvae, all in three biological replicates. The results indicated that 4,054 genes were upregulated and 4,260 genes downregulated in SG devECs relative to mitotic cycling B-D cells (log2FC at least +/- 0.5 and corrected p-value <0.05) (Fig 4D).
A comparison of SG devEC with CycA dsRNA and Myb dsRNA iECs showed that they had in common 158 genes that are increased and 109 genes that are decreased in expression relative to mitotic cycling cells (Fig 4E, S5 Table). This observed overlap in upregulated genes among all three endoreplicating cell types was 4.3 fold greater than expected by chance (p<1 x 10−5), while the overlap of downregulated genes was 9.1 fold greater than expected by chance (p<1 x 10−5) (S3B Fig). Consistent with our previous array analysis, many genes induced by E2F1 and MMB are expressed at lower levels in endocycling SG devECs relative to mitotic cycling B-D cells [13]. Among the 111 genes that are known to depend on E2F1 for their transcriptional induction in S2 cells, 73 were reduced in expression in SG devEC (Fig 4F, S6 Table) [61, 62]. Fewer E2F1-dependent genes (59) were downregulated in CycA dsRNA iECs, with an overlap of 48 downregulated E2F1-dependent genes in both CycA dsRNA iECs and SG devECs (Fig 4F, S5 Table) [61]. 11 of the 25 E2F1-dependent genes that were downregulated in devECs but not in CycA dsRNA iECs have functions in S phase, including CycE, dup (Cdt1), and PCNA; the three E2F1-regulated genes that RT-qPCR had indicated are not repressed in CycA dsRNA iECs (Fig 1C, S5 Table). Thus, although reduced expression of these E2F1-regulated S phase genes is common in devECs, their repression is not essential for endoreplication [10, 13–15]. Consistent with this idea, only 12 E2F1-dependent genes were commonly downregulated in both iEC types and devECs, and all have functions in mitosis (Fig 4F, Table 1, S6 Table). These 12 E2F1-dependent genes are a subset of the 47 Myb-dependent mitotic genes that we had found are downregulated in iEC, and therefore require both E2F1 and the MMB for their full expression (Fig 4C, S6 Table) [62].
Considering all downregulated genes, the most significantly enriched GO categories shared by iECs and devECs were mitosis and cytokinesis, including all of the 47 Myb-dependent genes that were commonly downregulated between CycA and Myb dsRNA iECs (Fig 4C, S6 Fig, S3 and S4 Tables). Given that CycA / Cdk1 activity is repressed in both CycA dsRNA iEC and SG devECs, the lower expression of these 47 genes in devECs further suggests that their transcriptional induction by the MMB is dependent on CycA (S4 Table) [9, 11, 63]. These genomic results show that while iECs and devECs both have a dampened MMB transcriptome of mitotic genes, repression of E2F1-regulated S phase genes is not essential for endoreplication.
The findings in S2 cells and tissues suggested that downregulation of an MMB transcriptome of mitotic genes promotes endoreplication. It was unclear, however, which of these downregulated genes are key for the decision to switch to endoreplication cycles. To address this question, we took an integrative genetic approach, using a collection of fly strains with GAL4-inducible UAS-dsRNAs to knock down the expression of genes that RNA-Seq had indicated were downregulated in iECs. We used an inclusive criterion and knocked down genes that were downregulated by log2 fold of at least -0.5 in both CycA and Myb dsRNA iECs, but without regard to p value (244 available strains representing 240 genes) [64] (S7 Table). We used dpp-GAL4 to express these dsRNAs along the anterior-posterior compartment boundary of the larval wing disc, and then examined the hair pattern in the central part of adult wings between veins L3 and L4, the region that is the known fate of cells that express dpp-GAL4 [65, 66] (Fig 5A). Each hair on the adult wing represents an actin protrusion from a single cell, and it is known that polyploidization of wing cells results in fewer and larger hairs (Fig 5B) [67–69]. As proof of principle, expression of a UAS-CycAdsRNA along the A/P boundary resulted in a central stripe of longer hairs on the adult wing surface and wing margin between veins L3 and L4, with many of these cells producing clusters of multiple hairs (Fig 5C). Analysis of larval wing discs co-expressing a UAS-mRFP reporter showed that dpp-GAL4; UAS-CycAdsRNA cells at the A/P compartment boundary had larger nuclei and increased DNA content compared to control cells outside of this dpp-GAL4 stripe, confirming that the adult wing phenotype is a result of endoreplication (Fig 5C and 5L, S7B Fig). Knockdown of Myb also resulted in an increased DNA content of wing disc cells and a stripe of larger and more widely spaced adult wing hairs (Fig 5D and 5L, S7C Fig). Although both of these Myb knockdown phenotypes were less severe than that of CycA knockdown, this is likely because this Myb dsRNA is inefficient (S8 Fig). Consistent with this, raising dpp-GAL4; UAS-MybdsRNA larvae at 29 C, a temperature at which transcriptional induction by GAL4 is stronger, resulted in a more severe endoreplication phenotype that was similar to CycA (Fig 5E and 5L). Knockdown of either CycA or Myb resulted in a reduced wing surface area between wing veins L3 and L4, suggesting that growth by an increase in cell size (hypertrophy) was not able to completely recapitulate normal tissue growth by cell proliferation (Fig 5C–5E and 5K). Among the 244 strains tested, 26 resulted in lethality before adulthood, suggesting that their functions are essential (S7 Table).
Among the other 218 crosses that survived to adults, knockdown of five genes reproducibly resulted in a reduction in the area between the L3 and L4 veins and abnormal wing hairs–aurora B (aurB), Incenp, Spc25, tumbleweed (tum), and pavarroti (pav). Of these, three reproducibly had enlarged wing hairs and a corresponding increased DNA content of wing disc cells–aurora B (aurB), tumbleweed (tum), and pavarroti (pav) (Fig 5F–5L, S7D–S7I Fig) [70–74]. Remarkably, all of these genes are either part of the chromosomal passenger complex (CPC) or are downstream effectors of it. The AurB kinase and INCENP are two subunits of the four-subunit CPC complex, which phosphorylates downstream targets to regulate multiple processes of mitosis and cytokinesis [75, 76]. Spc25 is a subunit of the Ndc80 outer kinetochore complex, which is phosphorylated by the CPC to regulate microtubule-kinetochore attachments [77, 78]. The Tum protein is a Rac-GAP protein that is phosphorylated by the CPC and regulates the kinesin Pav for proper cytokinesis [79]. While knockdown of any of these five genes resulted in longer hairs on the wing margin and surface, knockdown of aurB affected hair length primarily in the anterior half of the L3 / L4 intervein region (Fig 5F). This mild phenotype is not unexpected because the UAS-aurBdsRNA-1 transgene in this strain is based on a series of vectors that are not highly efficient for dsRNA expression. The stronger phenotype in the anterior could reflect the influence of patterning signals on a cells propensity to endoreplicate, although it could also be the result of different levels of dpp-GAL4 expression and aurB knockdown in different cells. Expression of a more efficient UAS-aurBdsRNA-2 had resulted in pupal lethality before adulthood, and RT-qPCR indicated that it knocked down aurB mRNA to lower levels than UAS-aurBdsRNA-1 (S8 Fig). Examination of wing disc cells showed that while UAS-aurBdsRNA-1 induced a low level of polyploidy, UAS-aurBdsRNA-2 resulted in very large polyploid cells, suggesting that a strong knockdown of AurB results in high levels of endoreplication (Fig 5L, S7D and S7E Fig). All five of these MMB-regulated genes were expressed at significantly lower levels in both iECs and devECs. The combined genomic and genetic results suggest, therefore, that a dampened CycA—Myb—AurB network promotes a switch from mitotic cycles to endoreplication cycles.
To test whether the status of the CycA—Myb—AurB network determines the decision between mitotic and endoreplication cycles in tissues other than the wing disc, we analyzed the effects in somatic follicle cells of the ovary. These cells have specific advantages for quantifying cell cycle and cell growth. Follicle cells form a regular epithelial sheet that surrounds 15 germline nurse cells and one oocyte in each maturing egg chamber. Their regimented cell cycle programs are well characterized and coupled with stages of oogenesis, dividing mitotically during stages 1–6, undergoing three endocycles during stages 7-10A, and then selectively re-replicating genes required for eggshell synthesis during stages 10B-14 [80–82]. We induced conditional knockdown of the genes identified in the wing screen using the heat-inducible GAL4 / UAS FLP-On system, which results in clonal activation of GAL4 and induction of a UAS-dsRNA and a UAS-RFP reporter in a subset of cells [83]. This conditional knockdown also permitted an analysis of genes whose knockdown resulted in lethality in the wing screen. Three days after heat induction, we quantified the number of cells in the clone, their nuclear size, and their DNA content by measuring DAPI fluorescence. If a gene knockdown induces a switch from mitotic to endoreplication cycles, it should result in clones with fewer cells that have an increase in nuclear size and DNA content. We had shown previously shown that knockdown of CycA or over-expression of Fzr (Cdh1) induces mitotic follicle cells into precocious endocycles during early oogenesis [36, 42].
We analyzed clones in stage 6, the latest stage of oogenesis during which follicle cells mitotically divide. Based on the known rate of egg chamber maturation, patches of RFP+ follicle cells in these stage 6 egg chambers represent the clonal descendants of single founder follicle cells that were either transit amplifying stem cell daughters or in stage 1 egg chambers at the time of induction three days earlier. Wild type, control clones were comprised of ~28 RFP-positive cells, indicating that they had divided ~4–5 times since FLP-On in the original single founder cell (Fig 6A and 6M). FLP-On of UAS-CycAdsRNA resulted in clones with only one to three cells, indicating that cell division was strongly inhibited, and that they had switched to endocycles during the first or second mitotic cycle after CycA knockdown (Fig 6B and 6M, Table 2). The cells in these clones had a single large nucleus with increased DNA content up to ~16C, indicating that they had endoreplicated, consistent with our previously published results (Fig 6B and 6N, Table 1) [36]. Expression of UAS-MybdsRNA resulted in some clones with reduced cell numbers and larger nuclei, suggesting that they had switched from mitotic divisions to endoreplication, but with variable expressivity among clones (Fig 6C, 6M and 6N, Table 1). A few Myb knockdown cells had two nuclei that were increased in size and DNA content, suggesting that these cells had failed cytokinesis before replicating their DNA again, a type of endomitosis. This variably expressive phenotype is likely the result of partial Myb knockdown by the inefficient UAS-MybdsRNA transgene (S8 Fig). To address this, we compared these UAS-MybdsRNA clones at 25°C to those grown at 29°C, a higher temperature that increases GAL4 activity. The clones at 29°C had a stronger phenotype, with many Myb knockdown cells having very large polyploid nuclei (Fig 6D and 6N). These results are consistent with the results in S2 cells and wing discs, and indicate that knockdown of CycA or Myb is sufficient to induce endoreplication.
The combined RNA-seq and genetic screen results suggested that reduced expression of CPC subunits and other targets downstream of Myb contributes to the switch to endoreplication. Clones expressing the weaker UAS-aurBdsRNA-1 had only two to three cells, indicating that cell proliferation was strongly inhibited, each with variable increases in nuclear size and DNA content (Fig 6E, 6M and 6N, Table 2). A few of these cells had two nuclei of increased size and DNA content, suggesting that UAS-aurBdsRNA-1 impaired cytokinesis followed by endoreplication. FLP-ON expression of the stronger UAS-aurBdsRNA-2 in follicle cells resulted in clones composed of only one to two cells, each with a single, large, polyploid nucleus (Fig 6F, 6M and 6N, Table 2). Many of these nuclei were multi-lobed, with connected chromatin masses composed of large chromosomes that appeared polytene (Fig 6F). These results suggest that mild knockdown of AurB results in cytokinesis failure, whereas a stronger knockdown results in a failure to segregate chromosomes and cytokinesis, followed by endoreplication.
To further test whether reduced CPC activity induces endoreplication, we knocked down expression of its other three subunits: Incenp, Borealin-related (Borr), and Deterin (Det) (fly Survivin ortholog), all of which were expressed at lower levels in iECs and salivary gland devECs (S3 and S5 Tables) [72, 84, 85]. Expression of UAS-IncenpdsRNA did not reduce the number of cells per clone nor increase DNA content, consistent with its lack of effect in the wing discs, an uninformative negative result because this UAS transgene is optimized for expression in the germline but expressed poorly in the soma (Table 2). Knockdown of the other CPC subunits, Borr and Det, had resulted in lethality in the wing screen, whereas their conditional knockdown in follicle cells resulted in clones with very few cells, each with large, polyploid nuclei (Fig 6G, 6H, 6M and 6N, Table 2). As a further test of the importance of the CPC, we knocked down aurB in S2 cells, and compared the effect of its knockdown to another mitotic kinase gene, polo. Similar to wing and ovarian follicle cells, knockdown of aurB in S2 cells resulted in endoreplication, whereas knockdown of polo resulted in a mitotic arrest (S9A and S9B Fig). These results indicate that reduction of CPC activity is sufficient to switch cells from mitotic cycles to endoreplication cycles.
We then addressed which processes downstream of the CPC are crucial for the mitotic cycle versus endoreplication cycle decision in follicle cells [76]. The genes Spc25, tum, and pav encode proteins that function downstream of the CPC [70, 79, 86, 87]. All three of these genes are regulated by the MMB, were expressed at lower levels in iECs and devECs, and were recovered in the wing screen (Figs 4C and 5H–5J). Knockdown of the kinetochore protein Spc25 in follicle cells strongly inhibited cell division and resulted in fewer cells per clone (Fig 6I, 6M and 6N). However, the nuclear size and DNA content of these cells were not increased, indicating that endoreplication was not induced, consistent with results from the wing (Figs 5L and 6I and 6N, Table 2). To further evaluate kinetochore proteins downstream of the CPC, we knocked down Spc105R, a kinetochore protein important for microtubule attachment and the spindle assembly checkpoint (SAC) [88]. Clonal knockdown of Spc105R strongly inhibited cell proliferation, and induced pyknotic nuclei indicative of programmed cell death, but did not result in an increase in DNA content (Fig 6J, 6M and 6N, Table 2). Growing the Spc25 and Spc105R clones at 29°C to enhance knockdown resulted in fewer cells per clone, and more nuclei that appeared pyknotic, but again did not result in enlarged nuclei (Fig 6M and 6N). Thus, despite a strong mitotic arrest phenotype, knockdown of these two kinetochore proteins downstream of the CPC did not result in endoreplication. Knockdown of the cytokinesis proteins Tum or Pav resulted in many fewer cells per clone, with many binucleate, indicating that cytokinesis was inhibited (Fig 6K–6M, Table 2). Unlike kinetochore protein knockdown, however, the binucleate Tum knockdown cells had a significant increase in both nuclear size and DNA content per nucleus (mean ~2 fold, max ~4 fold increase), indicating that they had endoreplicated after a failure of cytokinesis, consistent with the results from the wing disc (Figs 5K, 6K and 6N, Table 2). While some Pav knockdown cells clearly had larger polyploid nuclei (~2 fold), the average was not significantly different from control cell populations, unlike the results from wings where Pav knockdown induced significant polyploidy (Figs 5L, 6L, 6M and 6N, Table 2). Stronger knockdown of pav at 29°C, however, did result in a significant increase in nuclear area and DAPI intensity, consistent with the interpretation that these cells have undergone endoreplication (Fig 6N). Thus, inhibition of cytokinetic, but not kinetochore, proteins downstream of the CPC induces an endoreplication cycle. All together, these results suggest that the status of a CycA—MMB—AurB network determines the choice between mitotic and endoreplication cycles.
We have investigated how the cell cycle is remodeled when mitotic cycling cells switch into endoreplication cycles, and how similar this remodeling is between devECs and experimental iECs. We have found that repression of a CycA—Myb—AurB mitotic network promotes a switch to endoreplication in both devECs and iECs. Although a dampened E2F1 transcriptome of S phase genes is a common property of devECs in flies and mice, we found that repression of the Myb transcriptome is sufficient to induce endoreplication in the absence of reduced expression of the E2F1 transcriptome. Knockdown of different components of the CycA-Myb-AurB network resulted in endoreplication cycles that repressed mitosis to different extents, which suggests that regulation of different steps of this pathway may explain the known diversity of endoreplication cycles in vivo. Overall, these findings define how cells either commit to mitosis or switch to different types of endoreplication cycles, with broader relevance to understanding the regulation of these variant cell cycles and their contribution to development, tissue regeneration, and cancer.
Our findings indicate that the status of the CycA—Myb—AurB network determines the choice between mitotic or endoreplication cycles (Fig 7). These proteins are essential for the function of their respective protein complexes: CycA activates CDK1 to regulate mitotic entry, Myb is required for transcriptional activation of mitotic genes by the MMB transcription factor complex, and AurB is the kinase subunit of the four-subunit CPC. While each of these complexes were previously known to have important mitotic functions, our data indicate that they are key nodes of a network whose activity level determines whether cells switch to the alternative growth program of endoreplication (Fig 7). Our results are consistent with previous evidence in several organisms that lower activity of the Myb transcription factor results in polyploidization, and further shows that repressing the function of the CPC and cytokinetic proteins downstream of Myb also promotes endoreplication [13, 16, 23, 89]. Importantly, our genetic evidence indicates that not all types of mitotic inhibition result in a switch to endoreplication. For example, knockdown of the Spc25 and Spc105R kinetochore proteins or the Polo kinase resulted in a mitotic arrest, not a switch to repeated endoreplication cycles. These observations are consistent with CycA / CDK, MMB, and the CPC playing principal roles in the mitotic network hierarchy and the decision to either commit to mitosis or switch to endoreplication cycles.
While knockdown of different proteins in the CycA-Myb-AurB network were each sufficient to induce endoreplication cycles, these iEC populations had different fractions of cells with multiple nuclei diagnostic of an endomitotic cycle. Knockdown of cytokinetic genes pav and tum resulted in the highest fraction of endomitotic cells, followed by the CPC subunits, then Myb, and finally CycA, with knockdown of this cyclin resulting in the fewest endomitotic cells. These results suggest that knocking down genes higher in this branching mitotic network (e.g. CycA) inhibits more mitotic functions and preferentially promotes G / S endocycles that skip mitosis, whereas inhibition of functions further downstream in the network promote endomitosis (Fig 7). Moreover, we found that different levels of CPC function also resulted in different subtypes of endoreplication. Strong knockdown of AurB inhibited chromosome segregation and cytokinesis resulting in cells with a single polyploid nucleus, whereas a mild knockdown resulted in successful chromosome segregation but failed cytokinesis, suggesting that cytokinesis requires more CPC function than chromosome segregation. It thus appears that different thresholds of mitotic function result in different types of endoreplication cycles. This idea that endomitosis and endocycles are points on an endoreplication continuum is consistent with our evidence that treatment of human cells with low concentrations of CDK1 or AurB inhibitors induces endomitosis, whereas higher concentrations induce endocycles [28]. Our results raise the possibility that in tissues of flies and mammals both conditional and developmental inputs may repress different steps of the CycA—Myb—AurB network to induce slightly different types of endoreplication cycles that partially or completely skip mitosis [5, 90]. Together, our findings show that there are different paths to polyploidy depending on both the types and degree to which different mitotic functions are repressed.
Our findings are relevant to the regulation of periodic MMB transcription factor activity during the canonical mitotic cycle. Knockdown of CycA compromised MMB transcriptional activation of mitotic gene expression, and their physical association suggests that the activation of the MMB by CycA may be direct. The MMB-regulated mitotic genes were expressed at lower levels in CycA iECs, even though Myb protein levels were not reduced. This result is consistent with the hypothesis that CycA / CDK phosphorylation of the MMB is required for its induction of mitotic gene expression. Moreover, misexpression of Myb in CycA knockdown follicle cells did not prevent the switch to endoreplication, further evidence that CycA / CDK is required for MMB activity and mitotic cycles (S10 Fig). While the dependency of the MMB on CycA was not previously known in Drosophila, it was previously reported that in human cells CycA / CDK2 phosphorylates and activates human B-Myb in late S phase, and also triggers its degradation [53, 91]. While further experiments are needed to prove that CycA / CDK regulation of the MMB is direct, interrogation of the results of multiple phosphoproteome studies using iProteinDB indicated that Drosophila Myb protein is phosphorylated at three CDK consensus sites including one, S381 that is of a similar sequence and position to a CDK phosphorylated site on human B-Myb (T447) [92, 93]. We favor the hypothesis that it is CycA complexed to CDK1 that regulates the MMB because, unlike human cells, in Drosophila CycA / CDK2 is not required for S phase, and Myb is degraded later in the cell cycle during mitosis [45, 94]. Moreover, it is known that mutations in CDK1, but not CDK2, induce endocycles in Drosophila, mouse, and other organisms [37, 95]. A cogent hypothesis is that CycA / CDK1 phosphorylates Myb, and perhaps other MMB subunits, to stimulate MMB activity as a transcriptional activator of mitotic genes, explaining how pulses of mitotic gene expression are integrated with the master cell cycle control machinery (Fig 7). It remains formally possible, however, that both CycA / CDK2 and CycA / CDK1 activate the MMB in Drosophila. The early reports that CycA / CDK2 activates B-Myb in human cells were before the discovery that it functions as part of the MMB and the identification of many MMB target genes, and further experiments are needed to fully define how MMB activity is coordinated with the central cell cycle oscillator in fly and human cells [17, 19, 24, 26].
We experimentally induced endocycles by knockdown of CycA to mimic the repression of CDK1 that occurs in devECs. Our data revealed both similarities and differences between these experimental iECs and devECs. Both iECs and SG devECs had a repressed CycA—Myb—AurB network of mitotic genes. In contrast, only devECs had reduced expression of large numbers of E2F1-dependent S phase genes, a conserved property of devECs in fly and mouse [10, 13–15]. In CycA iECs, many of these key S phase genes were not downregulated, including Cyclin E, PCNA, and subunits of the pre-Replicative complex, among others. This difference between CycA dsRNA iECs and SG devECs indicates that repression of these S phase genes is not essential for endoreplication. In fact, none of the E2F1 -dependent S phase genes were downregulated in Myb dsRNA iEC. Instead, the 12 E2F1-dependent genes that were commonly downregulated in Myb dsRNA iEC, CycA dsRNA iEC, and SG devEC all have functions in mitosis (Table 1). These 12 mitotic genes are, therefore, dependent on both Myb and E2F1 for their expression, including the cytokinetic gene tum whose knockdown induced endomitotic cycles. This observation leads to the hypothesis that downregulation of the E2F transcriptome in fly and mouse devECs may serve to repress the expression of these mitotic genes, and that the repression of S phase genes is a secondary consequence of this regulation. These genomic data, together with our genetic evidence in S2 cells and tissues, indicates that in Drosophila the repression of the Myb transcriptome is sufficient to induce endoreplication without repression of the E2F1 transcriptome. The observation that both CycAdsRNA iECs and devECs both have lower CycA / CDK activity, but differ in expression of E2F1 regulated S phase genes, also implies that there are CDK-independent mechanisms by which developmental signals repress the E2F1 transcriptome in devECs.
Our results have broader relevance to the growing number of biological contexts that induce endoreplication. Endoreplicating cells are induced and contribute to wound healing and regeneration in a number of tissues in fly and mouse, and, depending on cell type, can either inhibit or promote regeneration of the zebrafish heart [27, 30–32]. An important remaining question is whether these iECs, like experimental iECs and devECs, have a repressed CycA—Myb—AurB network. If so, manipulation of this network may improve regenerative therapies. In the cancer cell, evidence suggests that DNA damage and mitotic stress, including that induced by cancer therapies, can switch cells into an endoreplication cycle [5, 41, 96, 97]. These therapies include CDK and AurB inhibitors, which induce human cells to polyploidize, consistent with our fly data that CycA / CDK and the CPC are key network nodes whose repression promotes the switch to endoreplication [75, 98]. Upon withdrawal of these inhibitors, transient cancer iECs return to an error-prone mitosis that generates aneuploid cells, which have the potential to contribute to therapy resistance and more aggressive cancer progression [28, 99–101]. Our finding that the Myb transcriptome is repressed in iECs opens the possibility that these mitotic errors may be due in part to a failure to properly orchestrate a return of mitotic gene expression. Understanding how this and other networks are remodeled in polyploid cancer cells will empower development of new approaches to prevent cancer progression.
Drosophila strains were obtained from the Bloomington Stock Center (BDSC, Bloomington, IN), or the Vienna Drosophila Resource Center (VDRC, Vienna Austria). The UAS-mRFP-Myb strain was kindly provided by Dr. Joe Lipsick. Drosophila were raised on BDSC standard cornmeal medium at 25°C unless otherwise indicated. For the genetic screen of Fig 4, fly strains with UAS-dsRNA transgenes were made by the Drosophila RNAi Screening Center (DRSC) and provided by the BDSC. These strains were crossed to dpp-Gal4, UAS-mRFP and multiple progeny of each cross were scored for their adult wing phenotype. Specific details about genotypes and strain numbers can be found in S7 and S8 Tables.
S2 cells were grown at 25°C in M3 + BPYE medium supplemented with 10% Fetal Bovine Serum as described [102]. iECs were supplemented with an additional 2% Fetal Bovine Serum (12% final). Cell proliferation in S2 Fig was quantified by counting cells using a hemocytometer. For RNAi, S2 cells were treated with the indicated dsRNA for 1 hour in serum free medium, followed by culturing for 96 hours at 25°C, as indicated above, and then analyzed as indicated below.
After dsRNA treatment, S2 cells were harvested in PBS and fixed in ethanol. After fixation, cells were incubated in propidium iodide (20 μg/ml) supplemented with RNaseA (250 μg/ml) at 37°C for 30 minutes. Flow cytometry was performed using an LSRII (BD Biosciences) and data were analyzed with Flowjo v7.6.5 software.
Protein extracts were made from S2 cells using a non-denaturing lysis buffer (25mM Tris, pH 7.5, 150mM NaCl, 5mM EDTA, 1% IGEPAL (Sigma-Aldrich), 5% glycerol, complete protease inhibitor cocktail (Sigma-Aldrich), PhosSTOP (Sigma-Aldrich)) and homogenizing the cells on ice. Absolute protein levels were determined by Bradford assays. At least 20 μg protein was separated by SDS-PAGE, electrophoretically transferred to PVDF membranes, and blotted using the following antibodies: anti-Cyclin A (A12, DSHB, concentrate) at 1:1000, anti-Cyclin B (F2F4, DSHB supernatant) at 1:100, anti-HA (Y11, Santa Cruz) at 1:1000, anti-Myb (D3R, provided by J. Lipsick) at 1:1000, anti-Tubulin (E7, DSHB, concentrate) at 1:1000. Blots were labeled with HRP conjugated secondary antibodies and developed using Super Signal West Pico substrate (Thermo Scientific).
Hsp70-Gal4, UAS-mRFP or Hsp70-Gal4, UAS-mRFP-Myb flies were crossed to UAS-CycA (Fig 3C) or UAS-CycA-HA (Fig 3C’) flies. Larvae were heat treated three times at 37°C for 30 minutes over 1.5 days beginning in 2nd instar, and protein extracts made from early 3rd instar larvae by homogenizing in non-denaturing lysis buffer (indicated above) for 1 hour after the final heat treatment. Lysate was quantified using Bradford assays to normalize total protein content among samples. In Fig 3C, extracts were immunoprecipitated using highly-efficient RFP-Trap (Chromotek) single-chain nanobodies made in camelids and conjugated to agarose beads. In Fig 3C’, extracts were immunoprecipitated with anti-HA (F7, Santa Cruz) or normal mouse serum on Protein G Agarose (Invitrogen). Western blots of input and IP were incubated with antibodies against Drosophila Myb (gift of J. Lipsick), Cyclin A (DHSB), DsRed (Takara), and HA (Santa Cruz).
In Fig 2, S2 cells were treated with dsRNA for 96 hours at 25°C, replated on poly-D-lysine coated chamber slide, and allowed to settle for 16–18 hours. Cells were then incubated in EdU (20μM) for 2 hours at 25°C followed by click-it fluorescent labeling according to the manufacturer’s (Invitrogen) protocol. These cells were then labeled with antibodies against (pH3) (Millipore, 06–570) and appropriate fluorescent secondary antibodies. Cells were stained with DAPI (0.5μg/ml) and imaged on a Leica SP5 confocal or Leica DMRA2 fluorescent microscope. The fraction of EdU and pH3 labeled cells and nuclear area were quantified using ImageJ v1.50b software (https://imagej.nih.gov/ij/).
For S7 Fig, wing imaginal discs were dissected from 3rd instar larvae and labeled with antibodies against DsRed (Takara) followed by labelling with anti-rabbit Alexa Fluor 568 (Thermo Fisher). Cells were stained with DAPI (0.5μg/ml) and imaged on a Leica SP5 confocal or Leica DMRA2 fluorescent microscope. Nuclear area and DAPI fluorescence was measured with ImageJ. Nuclear area and DAPI fluorescence of GAL4-expressing, DsRed-positive cells within the wing pouch was normalized to that of DsRed-negative cells in the wing pouch of the same discs.
Hsp70-FLP;Act>cd2>Gal4, UAS-mRFP was crossed to different UAS-dsRNA fly strains. Well-fed 3–5 day old adult G1 females were heat induced at 37°C for 30 minutes and allowed to recover for three days before ovaries were dissected, and labeled with anti-dsRed (Takara) and counterstained with DAPI as previously described [36]. Cell clones in stage 6 egg chambers were imaged on a Leica SP5 confocal and Leica DMRA widefield epifluorescent microscope. Cell number was quantified by counting RFP+ cells. The area and total DAPI fluorescence of nuclei within individual cells of a clone (RFP+) were measured using ImageJ and normalized to the average of wild type cells outside of the clone (RFP-) in the same egg chamber.
mRNA for RT-qPCR was isolated by TRIzol (Invitrogen) according to the manufacturer’s instructions. cDNA was generated using the Superscript III kit (Invitrogen). qPCR was performed using Brilliant III Ultra-Fast SYBR Green qPCR Master Mix (Agilent Technologies) and the primers indicated in S8 Table. Act5C was amplified as an internal reference control. Data were analyzed using LinRegPCR software (ver. 2016.2) the Pfaffl method to determine relative transcript levels [103, 104].
For S2 cell RT-qPCR, RNA was isolated 96 hours after dsRNA knockdown or control GFP dsRNA. Each assay was performed with technical duplicates and biological triplicates.
For quantification of knockdown in discs in S8 Fig, hsp70-GAL4; UAS-dsRNA and control hsp70-GAL4 only larvae were heat treated twice at 37 C for ½ hour over one day, and mRNA was isolated from 3rd instar discs ½ hour after the second heat shock and RT-qPCR performed as described above. Reactions were done in technical and biological duplicates. mRNA levels in the knockdown strains were normalized to levels in the hsp70-GAL4 control strain.
Statistical analysis of Figs 1B, 1C, 2B, 2G, 2H, 3B and 6M, S9A and S9B Fig were performed using two-tailed Student’s t tests using Microsoft Excel (version 15.0.4753.1000). For Fig 2F and S10 Fig a two-tailed Welch’s t test was performed using GraphPad Prism (version 7.04), For Figs 5L, and 6N GraphPad Prism (version 7.04) was used to perform a one-way ANOVA with a two-stage linear step-up procedure of Benjamini, Krieger and Yekutieli post-hoc test [105] to assess statistical difference between control clones and the indicated dsRNA clones.
For RNA-Seq of S2 cells, RNA was prepared from three biological replicates of CycA dsRNA, Myb dsRNA, and GFP dsRNA treated cells. For tissues, RNA was prepared from salivary glands (SG) or brains plus imaginal discs (B-D) from the same feeding early third instar larvae in three biological replicates, as previously described [13]. TruSeq Stranded mRNA Libraries (Illumina) were prepared by the Center for Genomics and Bioinformatics (CGB) of Indiana University according to manufacturer’s protocol. Multiplex sequencing barcodes from TruSeq RNA Single Indexes set A or B (Illumina) were added to the libraries during construction. The barcoded libraries were cleaned by double side beadcut with AMPure XP beads (Beckman Coulter), verified using Qubit3 fluorometer (ThermoFisher Scientific) and 2200 TapeStation bioanalyzer (Agilent Technologies), and then pooled. The pool was sequenced on NextSeq 500 (Illumina) with NextSeq75 High Output v2 kit (Illumina). Single-end 75 bp read sequences were generated. The read sequences were de-multiplexed using bcl2fastq (software versions 1.4.1.2, 1.4.1.2, and 2.1.0.31 for GSF1389, GSF1471, GSF1611).
Read quality was checked with FastQC v0.11.5 [106], and reads were then mapped against the Dmel R6.23 genome assembly and annotation using STAR v2.6.1a [107]. Mapped fragments were assigned to exons via the featureCounts function of the Rsubread v1.24.2 bioconductor package [108], and various pseudogenes and ncRNAs were excluded. Differential gene expression between samples was calculated using DESeq2 v1.14.1 [109]. Gene lists derived from RNA-Seq data sets were categorized as upregulated (Log2 fold-change ≥ 0.5 with an FDR adjusted p ≤ 0.05) or downregulated (Log2 fold-change ≤ -0.5 with an FDR adjusted p ≤ 0.05) [59]. Human ortholog information and DIOPT scores were downloaded from FlyBase on 09-11-2018 [110] and GO terms were retrieved using the Bioconductor package AnnotationHub v2.12.0 with a snapshot date of 04-30-2018 [111]. GO enrichment analysis was performed and plots were generated using clusterProfiler v3.8.1 [112]. The comparisons between the differentially expressed genes in the RNA-seq and the accompanying Venn diagrams were created using custom scripts and the R library VennDiagram [113].
Permutation testing was used to calculate p-values and fold enrichment of the DE gene double overlap between CycA iEC and Myb iEC or triple overlap among CycA iEC, Myb iEC and Sg devEC relative to chance (S3 Fig) [114]. Briefly, either two or three random gene sets were sampled (for the double and triple overlap sets respectively), with total genes sampled equal to the number of DE genes observed for those samples, and the number of overlapping genes between the sampled sets was recorded. This randomization sampling process was repeated 100,000 times. The p-values were calculated by finding the number of permutation samples that resulted in an overlapping number of genes greater than or equal to the observed number of overlapping genes plus one, over the number of permutation samples plus one. For the enrichment plot of S3 Fig, each observed overlap value was converted to fold difference relative to the sampled overlap values, and the median, 5% and 95% quantiles are shown.
|
10.1371/journal.pcbi.1004248 | Predicting Peptide-Mediated Interactions on a Genome-Wide Scale | We describe a method to predict protein-protein interactions (PPIs) formed between structured domains and short peptide motifs. We take an integrative approach based on consensus patterns of known motifs in databases, structures of domain-motif complexes from the PDB and various sources of non-structural evidence. We combine this set of clues using a Bayesian classifier that reports the likelihood of an interaction and obtain significantly improved prediction performance when compared to individual sources of evidence and to previously reported algorithms. Our Bayesian approach was integrated into PrePPI, a structure-based PPI prediction method that, so far, has been limited to interactions formed between two structured domains. Around 80,000 new domain-motif mediated interactions were predicted, thus enhancing PrePPI’s coverage of the human protein interactome.
| Complexes formed between a structured domain on one protein and an unstructured peptide on another are ubiquitous. However, they are often quite difficult to detect experimentally. The development of computational approaches to predict domain-motif interactions is therefore an important goal. We report a method to predict domain-motif interactions using a Bayesian approach to integrate evidence from a variety of sources, including three-dimensional structural and non-structural information. The method was applied to the entire human proteome and showed significant improvement over existing methods. The method was incorporated into PrePPI, a computational pipeline for the prediction of protein-protein interactions that relies heavily on structural information. Approximately 80,000 new interactions were detected. The new PrePPI database provides easy access to about 400,000 human protein-protein interactions and should thus constitute a valuable resource in a variety of biological applications including the characterization of molecular interaction networks and, more generally, in the study of interactions mediated by proteins in families that may not be extensively studied experimentally.
| Mapping the human protein interactome has important implications for understanding basic biology and human disease at the molecular level [1]. High-throughput (HT) experimental techniques such as yeast two-hybrid and tandem affinity purification have been developed and applied to discover protein-protein interactions (PPIs) in multiple organisms on a genome-wide scale [2]. However, these approaches have inherent limitations, leading to a substantial false positive rate [2, 3] with many interactions likely undiscovered due to high rates of false negatives [2, 4, 5]. The development of reliable computational approaches to identify PPIs is therefore an important alternative to HT experimental techniques [6, 7]. Computational predictions of PPIs are based on criteria such as sequence orthology [8], similarity in evolutionary history [9], genomic context [10], and literature curation [11]. Predictions based on detailed structural modeling of PPIs have also been developed [12] and recent approaches [13] that combine low resolution structural modeling with non-structural information have begun to expand the applicability of structure to a genome-wide scale. Interactions determined by HT experiments and computationally have been deposited in databases such as STRING [14] and PrePPI [13].
An important class of PPIs involves interactions between a short peptide motif of one partner, and a structured peptide recognition domain (PRD) from another [15–18]. Discoveries of new domain-motif interactions present unique challenges compared to domain-domain mediated interactions [16, 19]. For a few major PRD families such as PDZ and SH3 domains, HT experimental techniques [19–21] such as phage display have been used to derive binding preferences which can subsequently be used to scan a genome to identify proteins likely to bind a given PRD. Computational modeling has also been used to predict domain-motif interactions [22–27]. In these studies, models for domain-motif complexes are built and evaluated with either physical or statistics–based scoring functions. Despite much progress, the experimental or computational effort involved significantly limits the scope of these studies to a few PRD families so that methods that enable predictions for a larger number of PRD families are needed. Databases such as the eukaryotic linear motif resource [28] (ELM) provide consensus sequence patterns for peptide motifs binding to many different PRD families, and methods such as iELM have been developed to make new predictions based on such information [29]. However, these patterns are often derived from a limited amount of data (e.g. from a few known binding sequences), which necessarily limit their coverage and accuracy. Surveys of available structures of protein-peptide complexes in the PDB have also identified candidate binding motifs [30] as well as generic structural characteristics for binding interfaces [31], but overall structural information has not been widely used in predicting new interactions except for a few PRD families.
In this study we report a computational framework to predict interactions mediated by domain-motif interfaces. The method uses a Bayesian approach to integrate knowledge from the ELM database, domain-peptide structures from the PDB, and non-structural information. We have incorporated the method into PrePPI [13] and found that the addition of domain-motif predictions improves its performance in PPI detection. The new version of PrePPI now contains 400,000 PPI predictions.
Fig 1 outlines the combination of strategies we use to predict PPIs mediated by domain-motif interfaces. The first approach we tried is based on ELM [28], a manually curated database containing more than 200 classes of PRD/motif pairs. In an ELM class a PRD is represented by its Pfam family and a motif is represented by a consensus sequence pattern derived from peptides known to bind to that family. For each ELM class, we identified all pairs of human proteins containing the corresponding PRD/motifs and calculated an interaction likelihood ratio (LR) for each. The LRs were calculated, using a Bayesian approach, as the percentage of pairs of proteins having the PRD/motif match in a true positive set of known human PPIs divided by the same percentage of a true negative set of 1.6 million pairs of proteins that do not interact (see Methods). In this calculation we also considered whether the motifs are located in predicted disordered regions and whether their sequences are conserved evolutionarily (see Methods for details). Sequence conservation and disorder have been shown to be associated with functional motifs [32, 33].
The performance of the PRD/motif predictor in rediscovering known human-human interactions was assessed using 5-fold cross-validation on the true positive and negative sets (see Methods) and compared with the iELM method developed by Weatheritt et al. [29] (Fig 2). iELM is also based on information from the ELM database and uses features such as sequence conservation and disorder incorporated into a support vector machine (SVM). In addition, instead of using Pfam to identify PRDs, Weatheritt et al. constructed their own Hidden Markov Models (HMMs) for each ELM class. As can be seen in Fig 2, PRD/motif performs better than iELM in the lower false positive region but the reverse is true in the higher false positive region. Similar results were obtained when using a precision-recall curve to evaluate performances, with PRD/motif having higher precision in the lower but not the higher recall region (S1 Fig). In what follows, we chose to use PRD/motif when extracting information from the ELM database based on its better performance in the lower false positive region.
Despite its broad scope, certain domain-motif interactions, especially those not belonging to well-studied families, may not be included in the ELM database. To expand our coverage, we used experimentally determined complexes taken from the PepX database [34] as templates to model domain-motif interactions (Fig 1). PepX contains high-resolution structures of protein-peptide complexes in the PDB whose peptide motif length ranges from 5 to 35 amino acids. Structural models for individual human proteins or their subdomains were retrieved from the PDB if available or from one of two homology model databases, ModBase [35] and SkyBase [36]. More than 10,000 human proteins have at least some part of their sequences covered by a structural model [13]. An interaction model for a pair of proteins was constructed if one protein contained a PRD that was structurally similar to a PRD in a given template in PepX and the other protein contained a short sequence motif with sequence similarity (based on BLOSUM62 scores [37], see Methods) to the motif component of the template. We only considered motifs whose BLOSUM scores ranked among the top 0.05% from the entire human proteome to retain a manageable number of candidate peptide motifs.
We again used a Bayesian approach to estimate the likelihood of an interaction given the properties of the model. Sources of evidence integrated into our Bayesian scheme include the sequence similarity score between the candidate motif and the motif in the template, the structural similarity score between the candidate PRD and the PRD in the template, whether the candidate motif is located in predicted disordered regions, and whether sequences around the candidate motif are conserved evolutionarily (see Methods). The resulting predictor, referred to as Struct for Structural information alone (based on the PepX database), performed worse than PRD/motif in our cross validation test. However, a predictor that combines both (PRD/motif+Struct) performs better than PRD/motif alone, showing that structural evidence is adding value to the predictions (Fig 2).
We combined PRD/motif-based and Struct-based LRs with non-structural (NS) evidence that has previously been used to infer PPIs [38]. Specifically, for each pair of proteins, we considered their co-expression level, their gene ontology (GO) functional similarity, and their phylogenetic profile similarity. Derivation of LR scores for these sources of evidence was described previously [13, 38] and the values obtained in our previous study [13] were directly used in the current one. A final score was obtained by multiplying the LR for the predicted domain-motif interface with the LR from non-structural evidence. The resulting integrative predictor, PRD/motif+Struct+NS, was then compared to the method based only on NS evidence in rediscovering known human PPIs (Fig 2). The NS-based method outperforms PRD/motif and Struct, which is not surprising given that NS is not limited to peptide-mediated interactions. However, PRD/motif+Struct+NS offers a significant improvement over NS alone (Fig 2). Furthermore, the combination of methods dramatically increases the number of predicted interactions with LR score > 600 [13, 38], referred to as “strong predictions” in this study. This LR value corresponds to a posterior probability of 0.5 that two proteins interact when assuming a prior odds of 1 in every 600 protein pairs interact. Details of the derivation can be found in Jansen et al. [38]. Using information from PRD/motif, Struct or NS alone led to 1,515; 0; and 15,376 strong predictions, respectively. In contrast, a total of 125,624 predictions were made when combining the three sources of evidence, representing 110,248 new predictions as compared to NS alone. This significant amplification highlights the value of combining independent clues. Notably, a total of 55 true positives can be detected before encountering the first false positive.
To obtain further validation of our approach, we compared our strong predictions to the 257 known human domain-motif interactions found in the ELM database. A total of 75 known interactions were included in our strong predictions (123 when using a LR cutoff of 100), while using only evidence from PRD/motif, Struct or NS alone recovered only 6; 0; and 6 interactions, respectively. Furthermore, when using the combined sources of evidence, the LR scores for more than half of the interactions (40 out of 75) ranked among the top 20% of all strong predictions. These 75 interactions were not dominated by a particular ELM class as they spanned 33 out of the 57 classes that represent the 257 known human interactions. We also examined overlap of our predictions with 160 human domain-motif complexes in PepX. The overlap is only 42 for strong predictions but increases to 86 when using a lower LR cutoff at 100.
As summarized above, we have previously developed PrePPI, a computational PPI prediction method that performs comparably to experimental HT approaches. PrePPI combines structural evidence with non-structural evidence using a Bayesian framework but currently lacks the capability to predict domain-motif mediated PPIs [13], To add this ability we compared the structure-based LR from the original PrePPI and the LR for domain-motif interfaces obtained from PRD/motif+Struct (see Methods). The larger of the two was chosen and multiplied with the LR obtained from NS evidence to generate a final LR for the interaction. As shown in Fig 3, the addition of evidence based on domain-motif interactions (PrePPI_PRD/motif+Struct) results in improvement in performance when compared to the original PrePPI (PrePPI_orig). In this comparison, we use the same true positive set described above but a larger true negative set of non-interacting pairs of proteins which was used in the original PrePPI (performance is nearly identical for both true negative sets, S2 Fig).
PrePPI_PRD/motif+Struct yielded 78,898 additional strong predictions compared to PrePPI_orig. Although more than 40% of the predictions come from the 5 most prevalent PRD families (including SH2, SH3, 14-3-3, nuclear receptors and AGC kinase docking motif), over 130 ELM classes and 150 clusters of PepX template structures contributed to our results. Together with the original 317,814 interactions reported from PrePPI_orig, the new PrePPI which includes domain-motif mediated interactions contains a total of 396,712 predicted human PPIs.
In this study we developed methods to predict PPIs mediated by domain-motif interfaces using both expert knowledge of domain-motif interactions in the ELM database and structures of domain-motif complexes in the PDB. Although there is some overlap between predictions made with the ELM and the structure-based approach (PRD/motif and Struct), differences between them likely led to the observed improvements when two strategies were combined. For example, the Bcl-2 families have multiple structural representatives in the PDB but are not included in ELM. Moreover, the sequence similarity scoring approach of PrePPI_Struct allows the identification of motifs outside of the consensus provided by ELM. For example, the motif sequences for several SH3 and nuclear box receptor complexes in PepX could not be described by consensus patterns from any of their corresponding classes in ELM. Overall, among our new strong predictions, 13,988 are made using motifs that cannot be described by consensus patterns from the corresponding ELM class. On the other hand, the use of consensus patterns as in ELM (and hence PrePPI_PRD/motif) can be effective in capturing the variability of motifs for large, well-studied families even when no structural information is available. In terms of finding PRDs, the structured-based method in PrePPI_Struct applies a filtering criterion to ensure that the candidate PRD aligns well structurally to the template PRD at the binding interface, which is not accounted for by the sequence-based Pfam definition in PrePPI_PRD/motif.
As in the original PrePPI, combining sources of evidence that on their own provide only weak clues has a major impact on overall performance. For example, the consensus sequence patterns used in the PrePPI_PRD/motif approach can be promiscuous as can the use of sequence similarity in PrePPI_Struct, potentially leading to reduced prediction specificity. This can be especially problematic for interactions between candidate PRDs and motifs that interact via similar interfaces. For example, for the structural modeling component in PrePPI_Struct, it is possible that modeled interfaces for many different pairs of proteins share the same sets of clues if they are derived from the same template structure. In this case, prediction specificity would come from non-structural evidence. Moreover, prediction coverage based on the individual sources of evidence can be low as shown in Results, which highlights the importance of combining different sources of orthogonal information implicit in the Bayesian approach.
It is widely appreciated that HT approaches including yeast-two hybrid and tandem affinity purification have limitations in detecting PPIs mediated by protein-peptide interfaces. Apart from issues such as their transient nature and high Kd, they frequently depend on cellular conditions, many of which will never be sampled in an HT experiment [16, 19], potentially resulting in very high false negative rates. Indeed, high-throughput screens focusing on individual PRDs often identify a surprisingly large number of binding partners [19]. Furthermore, it has recently been suggested [17] that the number of putative peptide motifs in the human proteome to be more than a million. The number was based on estimating the extent of disordered regions in the human proteome and the tendencies of these regions to be involved in binding. Motifs that undergo post-translational modification were also included in the estimate, based on their prevalence among a set of well-studied proteins [17]. Although there are certainly false positives in computational predictions, these issues highlight the importance of developing methods such as that described here that can be applied on a genome wide scale and are insensitive to such experimental difficulties. The large number of predictions we make provide hypotheses that can be further refined and tested by more in-depth experimental/computational studies. In addition, the integrative nature of our framework should also help provide the biological context for predicted interactions, further contributing to our understanding of this still largely unexplored portion of the human interactome.
A total of 20,318 unique human protein sequences were downloaded from UniProt [39] and constituted the human proteome in this study.
As of January 2014, a total of 203 ELM classes of motifs that shared similar sequence features and targeted by the same kind of domain were annotated in the ELM database. For each class, a consensus pattern for the motifs and the name of the Pfam family for the interacting PRD were retrieved from the database. Hidden Markov Models (HMMs) for each Pfam family were downloaded from the Pfam [40] website, and the hmmscan utility from the HMMER suite [41] was used to identify domains within each human protein using default cutoffs defined in the downloaded HMM files. Candidate motifs satisfying the consensus pattern were identified using an in-house Perl script.
We obtained domain-motif structures from the PepX database [34] (multimers interacting with a single peptide were excluded). A PRD in PepX was used as a template to model a domain-motif interaction for a given human protein if it is structurally similar to the model for that protein as defined by a protein structural distance (PSD) less than 0.65 calculated with the program Ska [42]. An additional requirement is that in the structural alignment at least 75% of interfacial residues for the template PRD must align to surface residues on the structural model for the protein. Interfacial residues for the template PRD were defined as those with at least one atom located within 4.5 Å of the template peptide motif in the complex structure. Surface residues for the structural models of human proteins were identified using the program SURFace [43], with an accessible surface area cutoff of 10 Å2. PSD scores between candidate domains and the template PRDs were grouped into two bins, [0–0.3] and [0.3–0.65], for the Bayesian classifier.
For a peptide motif in a given template that is x residues long, new potential binding motifs were identified by scanning a x-residue window across the whole human proteome. A sequence similarity score between the sliding window and the template motif was calculated using the BLOSUM62 scoring matrix [37]. Sequence motifs whose scores ranked among the top 0.05% among all such sliding windows were kept as candidate motifs. A cutoff based on percentage but not absolute BLOSUM62 scores enables comparison of motifs across different templates, which can vary greatly in length. For the Bayesian classifier, sequence similarity scores between candidate motifs and the template motifs were grouped into 4 bins: (1) scores within the top 0.0001%, (2) scores between the top 0.0001% and the top 0.001%, (3) scores between the top 0.001% and the top 0.01%, and (4) scores between the top 0.01% and the top 0.05%.
The program IUPred [44] was used to predict if a candidate motif is likely located in a disordered region. A disorder score (ranging from 0 to 1) for each individual residue in the human proteome was obtained by running IUPred on all human protein sequences. For each motif, a score was then obtained by averaging the disorder score for each individual residue in the motif. For the Bayesian scoring, a binary classification of candidate motifs was used: a candidate motif is disordered if the averaged score is larger than 0.5, which is the cutoff recommended by IUPred.
The program GOPHER [45] was used to search for orthologs among the UniProt database for every human protein. Only orthologs belonging to species of the subphylum vertebrata were considered, as including orthologs from a larger range of species (e.g. metazoan) does not significantly impact performance. A multiple sequence alignment of the orthologs was then generated using the program Muscle [46]. A conservation score for each residue in the human protein was calculated as the information content for the corresponding column in the alignment. The score was multiplied by the percentage of non-gap residues in the column. A residue was determined to be conserved locally if its conservation score was higher than the average of such scores for its surrounding residues [47] (up to 31-residue upstream and downstream). For the Bayesian scoring, a binary classification of candidate motifs was used: a motif was classified as locally conserved if all residues in the candidate motif were locally conserved.
A naïve Bayes classifier was used to integrate different sources of evidence into a likelihood ratio (LR) for an interaction between two proteins. The datasets for training the classifier consist of a true positive set that includes 7,409 interactions compiled from a set of 5 databases [48–52] and supported by at least two publications, and a true negative set that contains 206,361,949 interactions not supported by any publication [13]. While one can assume that a non-reported interaction is likely to be non-interacting, the negative set will necessarily contain undiscovered true interactions which are just the ones we would like to detect. The reported FPR at a given LR (which assumes every prediction in the true negative set is wrong) can therefore be viewed as an upper bound on the true value. As constructing a reliable set of non-interacting proteins remains difficult, we proceeded with this simple definition.
For Fig 2, in order to compare to iELM, we used a small set of 1.6 million pairs of proteins randomly sampled from the larger negative set, for which iELM scores were available. Results from the larger set were shown in Fig 3 (performance for both sets was nearly identical). For each property (referred to as a “clue”), ci, of an interaction between protein x and y, the conditional probability that one will observe the clue given that the interaction is in the true positive set or the true negative set is calculated as P(ci|Ixy,TP) and P(ci|Ixy,TN). The probability P(ci|Ixy,TP) is calculated as P(ci|Ixy,TP) = n/NTP, where n is simply the number of interactions in the true positive set with the clue ci, and NTP is the total number of interactions in the true positive set. P(ci|Ixy,TN) is calculated analogously for the true negative set. A LR value can be calculated by dividing these two probabilities, LR(ci) = P(ci|Ixy,TP) / P(ci|Ixy,TN), to reflect how strongly the clue ci is indicative of a true interaction.
For the PRD/motif method based on the ELM database, a total of four clues were used for the domain-motif component: a) whether a domain-motif match from the same ELM class is present (LR(match)); b) the identity of the matching ELM class (LR(class)); c) whether the motif is located in a predicted disordered region (LR(diso)); d) whether the motif is conserved locally in sequence relative to its surrounding regions (LR(consv)). The latter three clues can be assumed to be independent of one another, but they all depend on the first clue being true. Their LR values were therefore normalized by the LR for the first clue, and the final LR for the domain-motif interface is therefore:
LR(DMI)=LR(match)⋅(LR(class)/LR(match))⋅(LR(diso)/LR(match))⋅(LR(consv)/LR(match))
For the Struct method based on the PepX database, a total of five clues were used for the domain-motif component: a) whether a domain-motif match from the same template structure is present (LR(match)); b) The PSD score between the candidate domain and the PRD component in the template (LR(PSD)); c) the sequence similarity score between the candidate motif and the motif component in the template (LR(SIM)); d) whether the motif is located in a predicted disordered region (LR(diso)); e) whether the motif is conserved locally in sequence relative to its surrounding regions (LR(consv)). As above, LRs for the latter four clues were normalized by LR(match) and the final LR for the domain-motif interface is:
LR(DMI)=LR(match)⋅(LR(PSD)/LR(match))⋅(LR(SIM)/LR(match))⋅(LR(diso)/LR(match))⋅(LR(consv)/LR(match))
The LR for the domain-motif interface was then multiplied with LRs for non-structural evidence to obtain a final LR for the interaction. The LR values used in this study are provided as a supplemental table (S1 Table). LR scores for non-structural evidence based on co-expression, similarity in gene ontology, and similarity in phylogentic profiling calculated for the original PrePPI were used in this study [13].
Precision-recall curves were generated using the program AUCCalculator[53].
The iELM scores for the positive set and the randomly generated smaller negative set were kindly provided by Weatheritt et al. Incremental cutoffs of raw iELM scores were used to calculate the TPR and FPRs. If iELM makes multiple PRD/motif predictions for a single pair of protein, only the prediction with the highest score was considered in testing.
Predictions have been incorporated into the PrePPI database and can be downloaded for individual query proteins (https://honiglab.c2b2.columbia.edu/PrePPI/). New predictions are also provided as a supplement (S2 Table).
|
10.1371/journal.pbio.1002155 | The Fitness Consequences of Aneuploidy Are Driven by Condition-Dependent Gene Effects | Aneuploidy is a hallmark of tumor cells, and yet the precise relationship between aneuploidy and a cell’s proliferative ability, or cellular fitness, has remained elusive. In this study, we have combined a detailed analysis of aneuploid clones isolated from laboratory-evolved populations of Saccharomyces cerevisiae with a systematic, genome-wide screen for the fitness effects of telomeric amplifications to address the relationship between aneuploidy and cellular fitness. We found that aneuploid clones rise to high population frequencies in nutrient-limited evolution experiments and show increased fitness relative to wild type. Direct competition experiments confirmed that three out of four aneuploid events isolated from evolved populations were themselves sufficient to improve fitness. To expand the scope beyond this small number of exemplars, we created a genome-wide collection of >1,800 diploid yeast strains, each containing a different telomeric amplicon (Tamp), ranging in size from 0.4 to 1,000 kb. Using pooled competition experiments in nutrient-limited chemostats followed by high-throughput sequencing of strain-identifying barcodes, we determined the fitness effects of these >1,800 Tamps under three different conditions. Our data revealed that the fitness landscape explored by telomeric amplifications is much broader than that explored by single-gene amplifications. As also observed in the evolved clones, we found the fitness effects of most Tamps to be condition specific, with a minority showing common effects in all three conditions. By integrating our data with previous work that examined the fitness effects of single-gene amplifications genome-wide, we found that a small number of genes within each Tamp are centrally responsible for each Tamp’s fitness effects. Our genome-wide Tamp screen confirmed that telomeric amplifications identified in laboratory-evolved populations generally increased fitness. Our results show that Tamps are mutations that produce large, typically condition-dependent changes in fitness that are important drivers of increased fitness in asexually evolving populations.
| Aneuploidy (altered copy number of genomic regions) is observed in the majority of tumors, but it remains unclear whether aneuploidy is a cause or consequence of cancer. Evidence from the yeast Saccharomyces cerevisiae and mammalian cells has shown that aneuploid cells tend to grow more slowly than normal cells; however, aneuploidy has also been shown to promote tumor formation and microbial adaptation. To address this paradox, we took two approaches to study the relationship between fitness—measured as cellular growth—and aneuploidy. First, we examined aneuploid events isolated from laboratory-evolved populations of S. cerevisiae and found that the majority of such events improve cellular fitness, have a large effect-size, and show diverse fitness effects under different conditions. Second, we developed a method to create thousands of aneuploid strains spanning the yeast genome and used pooled competition experiments followed by barcode sequencing to determine their relative fitnesses. These genome-wide data revealed aneuploidy to have effects that were both large and wide-ranging (pleiotropic). We found that both the positive and negative fitness effects are typically driven by a small number of genes within each aneuploidy event. We conclude that aneuploidy is functionally important in the process of adaptation of yeast during laboratory evolution experiments and propose that it has the potential to play an adaptive role during the evolution of cancers.
| Aneuploidy, a class of mutation infamous for its disruption of development [1] and oncogenic connections [2,3], is a genetic alteration that changes the copy number of many genes with a single mutational event (reviewed in [4]). Despite its close connection to cancer, a phenomenon characterized by unchecked cellular proliferation, aneuploidy has been shown to inhibit cellular growth in a variety of model systems. Both trisomic mouse embryonic fibroblasts and disomic strains of Saccharomyces cerevisiae have increased doubling times when compared to their euploid counterparts [3,5]. The fitness cost associated with aneuploidy has been attributed to proteotoxic stress caused by the unbalanced and uncompensated expression of proteins from the regions of altered copy number [6–9].
Despite this general fitness cost, whole-chromosomal aneuploidy and segmental aneusomy, both of which will henceforth be referred to as “aneuploidy” for simplicity, have been commonly observed in the evolution and adaptation of asexually replicating cells [10–20]. Aneuploidy thus has a paradoxical relationship with cellular fitness [21]: while typically decreasing a cell’s fitness, it is nonetheless selected for under a variety of highly selective conditions. By altering the copy number of multiple genes at once, it has been argued that aneuploidy allows a cell to explore a wide fitness landscape [22,23]. Aneuploidy, therefore, may commonly be selected for when cells face novel conditions because this mutation type allows an evolving population to rapidly test many divergent phenotypes. However, the specific fitness effects of aneuploid events have been difficult to directly test and, instead, have typically been inferred from their recurrence between or frequency within evolving populations [13,14,24]. Even in the rare cases in which a fitness advantage is directly associated with a particular aneuploid event, it remains challenging to identify the gene(s) within the aneuploid region whose altered copy number is responsible for the fitness effects observed [25,26]. However, the gene(s) underlying the phenotype(s) associated with an aneuploid event have been identified in a small number of cases [17,20,27].
Aneuploidy’s genetic complexity and the challenges outlined above have made it difficult to draw firm conclusions about the general role aneuploidy plays in fitness, adaptation, and evolution. In this study, we have directly tested the fitness effects of four naturally selected aneuploid events isolated from three laboratory evolution experiments of S. cerevisiae carried out in nutrient-limited chemostats. We have found that while most aneuploid events positively affect fitness, one event actually decreased fitness despite representing a substantial fraction of the population. Unable to draw general conclusions about aneuploidy from the detailed analysis of only a few specific genetically tractable events, we then created a barcoded genomic collection of >1,800 clones each containing a telomeric amplification (Tamp) of a different size. By tiling across the entire yeast genome, this collection allowed us to test the fitness effects of telomeric amplifications genome-wide. Using pooled competition experiments in glucose-, sulfate-, or phosphate-limited chemostats combined with barcode sequencing [28], we have uncovered the fitness profile explored by Tamps under these three conditions. Data from this genome-wide Tamp screen revealed that aneuploidy is typically a large-effect mutation, with condition-specific fitness effects and fitness tradeoffs under alternative conditions. By comparing the Tamp screen data to aneuploid events identified in evolution experiments, we found that most aneuploid events identified in evolution experiments positively affect fitness. The discrete fitness breakpoints in the Tamp fitness profile allowed us to identify candidate driver genes that, in the genetic background of amplification of contiguous genes, were responsible for the fitness effects of each Tamp. We discovered that the fitness effects of most aneuploid events from evolved populations are driven by a small number of driver genes essential for their positive effects on competitive growth. These data are an attempt to systematically define the fitness landscape explored by aneuploidy.
Aneuploidy has been commonly observed in laboratory-evolved populations of S. cerevisiae adapting to nutrient limitation [10,12,29,30]. Our group previously reported that at least one aneuploid clone was observed in 13 out of 24 evolution experiments carried out for over 100 generations (122–328 generations) in nutrient-limiting chemostats [10]. The same study showed that all eight of the evolution experiments carried out under sulfate-limiting conditions contained a recurrent amplification surrounding the high-affinity sulfate transporter SUL1 [10]; two sulfate-limited populations contained aneuploid events in addition to the SUL1 amplification. The direct fitness effects and mechanism of formation regarding the SUL1 amplicon have been examined in detail elsewhere [30–32]; in this study our primary focus was to explore the functional importance of the remaining aneuploid events observed in the 24 evolution experiments.
As a proxy for their direct fitness effects, we first determined the population frequencies of the aneuploid events observed in the 24 evolution experiments [10] using array comparative genomic hybridization (aCGH) of population DNA. We predicted that aneuploid events rising to appreciable population frequencies provided a fitness advantage to the clones carrying them. In the initial description of the evolution experiments examined here, population aCGH was performed on 10 of the 24 evolution experiments [10]; here we performed population aCGH on the remaining 14 evolution experiments (Fig 1A and S1 Table, see GEO Accession GSE67769 for raw data). Given the tandem repeat structure of the SUL1 amplicons [30,32], their clonal copy number was dynamic and prohibited accurate calculation of their population frequency by population aCGH. In order to estimate the SUL1 amplicon population frequency, we defined the SUL1 clonal copy number as the population copy number rounded to the next highest integer. A detailed analysis of the SUL1 amplicon structure and population dynamics has been presented elsewhere [30].
The aneuploid events present in the evolution experiments ranged in size from 5–1,000 kb and were present at frequencies ranging from 6% (our lower limit of detection) to 96% of the population with an average population frequency of 47% (Fig 1A). We confirmed the accuracy of this approach by performing breakpoint PCR across the translocation event in the supernumerary chromosome present in population S8. Both aCGH on S8 population DNA and breakpoint PCR on 98 independent clones isolated from S8 determined this supernumerary chromosome to be present at 13% of the population (see GEO Accession GSE67769).
While 11 of the aneuploid events were unique, seven recurred between populations, both within and between conditions, most notably the amplification on the right arm of chromosome V and the amplification on the left arm of chromosome XIV. The amplification on the right arm of chromosome V recurred in three different evolution experiments carried out under three conditions (Fig 1A), while the amplification of the left arm of chromosome XIV was observed in one of the glucose-limited evolution experiments described here and in two additional glucose-limited evolution experiments previously analyzed [12]. The high population frequencies and the recurrence of aneuploid events between populations supported our hypothesis that the aneuploid events examined here were selected for under the conditions of laboratory evolution.
We next asked whether aneuploid and euploid evolved clones isolated from the final generation of evolution experiments were more fit than their wild-type ancestors. We determined the relative fitness of each evolved clone through chemostat competition experiments against an appropriate green fluorescent protein (GFP)-marked wild-type control clone and under conditions identical to the evolution experiment from which the evolved clone was isolated. Both euploid and aneuploid evolved clones showed a fitness advantage relative to their wild-type ancestor (Fig 1B and S2 Table). Note that clones P3c1 and P3c2 are euploid despite being isolated from an aneuploid population, because the aneuploid events were not fixed in population P3. The relative fitnesses of the evolved clones ranged from 17% to 61% more fit than the wild-type ancestor. Evolved clones isolated from sulfate-limited evolution experiments (n = 8) had significantly greater fitnesses than clones isolated from glucose or phosphate-limited evolution experiments (n = 19) (Wilcoxon Rank-Sum test, p-value = 0.036). While there was a statistical difference in the relative fitnesses between euploid and aneuploid clones (Wilcoxon Rank-Sum test, p-value = 0.0014), this was driven in part by the high fitness conferred by the SUL1 amplicon in all evolved clones isolated from sulfate-limited evolution experiments. However, the relationship between aneuploidy and fitness held true even when we restricted our examination to the eight clones isolated under glucose-limiting conditions: aneuploid clones (n = 4) had a significantly greater fitness than the euploid clones (n = 4) (Wilcoxon Rank Sum test, p-value = 0.029). Although these data demonstrated that evolved aneuploid clones, just like evolved euploid clones, are more fit than their wild-type ancestor, it did not establish whether the aneuploid events themselves or other mutations, such as single-nucleotide variants (SNVs), contributed to the improved fitness of evolved clones. To provide this direct connection we genetically isolated the aneuploid events and SNVs from three evolved aneuploid clones and determined the direct fitness consequences of both the aneuploid events and the SNVs.
To genetically isolate the aneuploid events present in evolved clones we first determined the full repertoire of mutations present in a subset of evolved clones. We chose to study three evolved aneuploid clones: two clones isolated at generations 141 and 217 (P5c3 and P6c1) from phosphate-limited evolution experiments begun with a haploid founder and one clone isolated at generation 250 (S8c2) from a sulfate-limited evolution experiment begun with a diploid founder. P5c3 has two aneuploid events: an extra copy of chromosome XIII and a supernumerary chromosome consisting of the right arm of chromosome VI joined to a telomeric amplicon from the left side of chromosome XVI (VIR t XVIL). P6c1 has a supernumerary chromosome consisting of a telomeric amplicon on the right side of chromosome V joined to the right arm of chromosome VI (VR t VIR) and S8c2 contains a supernumerary chromosome consisting of two copies of a telomeric amplicon from the right side of chromosome V flanking a centromeric segmental amplicon from chromosome X (VR t XCEN). These clones were chosen because they did not contain large deletions, thus making them amenable to backcrossing and tetrad dissection. We performed whole-genome sequencing (WGS) of these clones to an average mapping coverage of 46–68X (S3 Table). Three to seven SNVs were called in each clone (Table 1) and confirmed by Sanger sequencing. We also sequenced the populations from which clones P5c3 and P6c1 were isolated to an average mapping coverage of 39X and 116X, respectively, and determined that the SNVs identified in these clones ranged in frequency from below detection to 98% (Table 1).
To isolate segregants that had a single evolved SNV or aneuploid event in an otherwise ancestral genetic background, we backcrossed the haploid clones P5c3 and P6c1 to their corresponding wild-type strain and directly sporulated the diploid clone S8c2. We identified appropriate segregants by genotyping and then used chemostat competition experiments to determine the independent fitness effects of each evolved mutation (Fig 2 and S2 Table). More than half of the mutations examined showed either neutral/near-neutral (<5%) fitness increase or negative effects on fitness, agreeing with previous reports that genetic hitchhiking is quite important for the spectrum of mutations observed in asexually evolving populations [33,34]. In particular, the supernumerary chromosome isolated from evolved clone S8c2, despite occupying 13% of the S8 population, actually decreased the fitness of clones carrying it by 10% (Fig 2). A minority of evolved mutations, including three large-scale aneuploid events, the amplification of SUL1, and a missense mutation in the high-affinity phosphate transporter PHO84, all increased fitness in the conditions from which they were isolated.
In general, the aneuploid events we examined showed diverse relationships with the overall fitnesses of the evolved clones from which they were isolated. In P6c1, the fitness effect of the VR t VIR supernumerary chromosome added to a second positive-effect mutation, the missense mutation in PHO84, to predict the overall fitness of the original evolved clone. In contrast, in P5c3 the positive fitnesses associated with both aneuploid events in that clone were each similar to the overall fitness of the original evolved clone, suggesting epistasis between these two mutations. Finally, the overall fitness of evolved clone S8c2 was quite similar to the fitness effect of the SUL1 amplification alone, suggesting epistasis between the SUL1 amplicon and the 10% fitness cost conferred by the VR t XCEN supernumerary chromosome and the missense mutation in ADR1 in this clone.
To confirm that we had identified and genetically isolated all functionally important mutations, for P5c3 and P6c1 we isolated and determined the relative fitness of backcrossed segregants that either had all or, in the case of P5c3 alone, none of the mutations present in the original evolved clone. As expected, the backcrossed segregants with all of the evolved mutations had a relative fitness similar to the original evolved clone, while the P5c3 backcrossed segregant with none of the evolved mutations had neutral fitness (Fig 2B and 2C). We were unable to isolate similar backcrossed clones corresponding to S8c2. However, given the negative fitness effects of the VR t XCEN supernumerary chromosome in S8c2, we were particularly interested to see if there was any epistasis, and specifically sign epistasis, between the point mutations and the VR t XCEN supernumerary chromosome in S8c2. To test this, we determined the relative fitness of a backcrossed clone with all of the evolved mutations except for the SUL1 amplicon (“No SUL1 amp” in Fig 2A). This clone had a fitness of -13% which, given the >5% fitness deficit of the VR t XCEN supernumerary chromosome, the HO mutation, and the YNL181W mutation, indicated epistasis between these mutations, although sign epistasis was not observed.
When organisms adapt to a particular environment they may acquire mutations that produce a fitness tradeoff under alternative conditions [35,36]. Aneuploid events have previously been proposed to be pleiotropic mutations that, over the course of a population’s adaptation to a novel environment, are eventually replaced by mutations with fewer non-selective effects and correspondingly fewer fitness tradeoffs [37]. With these observations in mind, we determined the growth rates for 20 of the evolved clones in batch culture in rich media and observed a significant decrease in growth rate relative to wild-type for three of the 20 clones (S1 Fig)
The similar doubling times to wild-type for most of the evolved clones suggested that the majority of evolved clones do not show a fitness tradeoff under typical lab growth conditions. However, comparing monoculture growth rates is an insensitive method to detect small fitness differences between clones. We therefore examined the fitnesses of six evolved aneuploid clones and the four aneuploid events we had previously isolated (Fig 2) using chemostat competition experiments under the two nutrient limitation conditions not previously examined. Each of the four isolated aneuploid events and the SUL1 amplicon showed different fitness effects in the two alternative conditions (Fig 3A). Typically, each aneuploid event decreased or had a small effect (<5%) on fitness under the two alternative conditions tested.
However, both the VR t XCEN supernumerary chromosome from the sulfate-limited population S8 and the VR t VIR supernumerary chromosome from phosphate-limited population P6 increased fitness under glucose-limited conditions. We next tested evolved aneuploid clones under the two nutrient limitation conditions not previously examined and observed results similar to those achieved with the isolated aneuploid events. The evolved aneuploid clones typically had lower-than-wild-type fitness under the two alternative nutrient conditions, although occasionally they had increased fitness under alternative conditions (Fig 3B). Finally, we compared the pleiotropy, defined as the variance in fitness between conditions, of the four isolated aneuploid events to the pleiotropy of single-gene changes in copy number and found aneuploid events to be significantly more pleiotropic than single-gene changes in copy number (unpaired, two-tailed t test, p = 0.049, S2A Fig) These results generally supported previous hypotheses that proposed aneuploidy to be highly pleiotropic [7,37]; however, these results also emphasized that aneuploidy does not always lead to negative fitness tradeoffs but can also have unselected fitness benefits under alternative conditions.
These detailed analyses of evolved aneuploid clones isolated from laboratory evolution experiments demonstrated the varying impact aneuploidy could exert on cellular fitness and proved that aneuploidy can cause fitness improvements in experimental evolution under nutrient limitation. However, this type of rigorous analysis is not scalable, and the limited number of clones we have examined here precluded any conclusions about the general effects of aneuploidy on fitness, adaptation, and evolution. With the dual goals of (1) identifying which aneuploid events in the remaining evolved clones increased fitness and (2) generating sufficient data to approach general questions about aneuploidy’s role in adaptation and evolution, we devised a screen to assay the fitness effects of aneuploidy genome-wide.
In designing our screen, we decided to focus on a particular category of aneuploid event: telomeric amplicons (Tamps), which we defined as a segmental amplification that initiates at a given location in the genome and extends to the proximal telomere. Tamps are a mutation type worthy of focused study as they are frequently observed in our evolved clones (17/36 aneuploid events are Tamps), and Tamps also play a role in human diseases such as cancer and developmental disorders [13,14,38].
To construct a genome-wide collection of Tamps, we returned to a classic method of genetic analysis: chromosome fragmentation [39]. This method was originally used for mapping the physical location of cloned genes. In our case, we were interested in it as an approach to fragment the yeast genome into a series of differently sized Tamps. By targeting our chromosome fragmentation vector (CFV) to the KanMX cassette that replaces each gene in the yeast deletion collection [40], we were able to generate Tamps initiating at selected genomic locations simply by altering the particular deletion collection strain we chose to transform with our CFV (Fig 4A and S3 Fig). Furthermore, as each deletion collection strain already had a unique DNA barcode identifying the genomic location of the KanMX cassette (“Tamp BC” in Fig 4A), we could simply use barcode sequencing (barseq) [28] to determine the location at which the Tamp initiated. The design of our CFV included an additional random 12 base-pair barcode that, upon transformation into the target deletion collection strain, was incorporated into the Tamp and provided a barcode for each independent transformation event of an individual deletion collection strain (Fig 4A “Replicate BC,” see Materials and Methods for details). The ability to track multiple biological replicates of each Tamp allowed us to determine more accurately the fitness for each Tamp.
We chose to build our Tamp pool from the diploid heterozygous deletion collection. We chose this deletion collection for two reasons. First, we wanted to match most closely the diploid background of most of the aneuploid clones isolated from our evolution experiments. Second, we expected there to be fewer suppressor mutations, which are commonly selected for in a homozygous or haploid deletion background to ameliorate the effects of the deleted gene [41,42]. Importantly, we were only able to take advantage of the yeast heterozygous deletion collection in this way because our lab had previously determined a set of 2,254 strains from this collection that have neutral fitness, with a range of relative fitnesses from -0.05 to 0.04, under our standard chemostat growth conditions of sulfate-, glucose-, and phosphate-limitation (S4 Table) [34]. Thus, by restricting our method to these 2,254 deletion collection strains and the limitations under which they have neutral fitness, we can be reasonably confident that any fitness effects we do measure are due to the Tamp itself and not the underlying genetic background.
With the intent of scaling eventually to the entire genome, we first sought to test our method on a small genomic region carrying a known driver gene: specifically, the telomeric 60 kb on the right arm of chromosome II (chr II). We chose to first focus on this region because it contains the high-affinity sulfate transporter SUL1, which our group had previously shown to be advantageous when amplified under sulfate-limiting conditions [10,30]. Furthermore, we demonstrated in the experiments described above that amplification of this region in its native chromosome context is also beneficial (Fig 2, “SUL1 amp”). We chose 60 kb since that amplicon size is the largest we have observed in diploid sulfate-limited evolution experiments [10]. We hypothesized that only Tamps containing SUL1 would increase fitness under sulfate-limiting conditions.
We successfully created 21 Tamp strains, each initiating at a different gene within this 60kb region and extending to the right telomere, by transforming 21 neutral-fitness heterozygous deletion strains with our KanMX-targeted CFV (see Materials and Methods for additional details). Each deletion strain was transformed individually, and the karyotype confirmed by aCGH (see GEO Accession GSE67769, see Materials and Methods for additional details). Pooled competition of these 21 strains for 9–12 generations followed by sequencing of the deletion collection barcode at five different time points allowed us to track the relative frequencies of each Tamp and infer their relative fitnesses (Fig 5A and S5 Table). As expected, our results demonstrated that Tamps containing SUL1 increased fitness under sulfate-limiting conditions (Fig 5A). In addition to SUL1, a second driver gene had previously been identified within this 60 kb region: BSD2 amplification increases fitness under sulfate-limiting conditions [34]. Data from this targeted Tamp screen identified both BSD2 and SUL1 as driver genes under sulfate-limiting conditions. Tamps containing both SUL1 and BSD2 had an average fitness increase of 23%, Tamps containing only SUL1 had an average fitness increase of 15%, and Tamps containing neither SUL1 nor BSD2 had an average fitness decrease of 18%. The same procedure was repeated under glucose-limiting and phosphate-limiting conditions and no increase in fitness relative to wild type was observed (S4 Fig), thus demonstrating the condition-specific fitness effects of the Tamps examined here.
We noticed that both SUL1 and BSD2 were highlighted in our data by a discrete decrease in Tamp fitness, or “Downstep” (Fig 5A). This was due to the fact that Tamps lacking SUL1 or BSD2 had decreased fitness compared to Tamps containing one or both of those genes. We hypothesized that such Downsteps could be used to identify additional driver genes in our genome-wide screen. Similarly, “Upsteps” could be used to identify genes that increased fitness when no longer amplified on a Tamp. Upsteps, therefore, could be used to identify novel “anti-driver” or growth-inhibiting genes in our genome-wide screen.
After we confirmed the validity of our method with the chr II-targeted screen described above, we scaled our approach to the entire genome. The 2,254 neutral-fitness deletion collection strains were pooled and transformed with our KanMX-targeted CFV. To ensure each Tamp was represented by multiple independent transformation events, >42,000 transformant colonies were collected, guaranteeing approximately 20 unique biological replicates for each Tamp. Barseq of the resulting Tamp pool revealed it to be of adequate complexity: 1,802 of 2,254 targeted Tamps (80%) were represented by >0.005% of the total reads, and each Tamp was represented by, on average, 26 independent transformants marked by unique biological replicate barcodes.
We next used our genome-wide pool of Tamps to inoculate three glucose-, phosphate-, and sulfate-limited chemostats, for a total of nine pooled competition experiments. Each competition experiment was carried out for approximately 25 generations, with samples for barseq taken at ten time points throughout (Fig 4B). We were able to track the Tamp frequencies of >100,000 unique biological replicates across ten time points under all three conditions. These data allowed us to, after the filtering steps described below, determine the fitnesses of 1,631, 1,596, and 1,551 Tamps in glucose-, phosphate-, and sulfate-limited conditions, respectively (S6 Table).
Our ability to track independent biological replicates of each Tamp was crucial in obtaining accurate Tamp fitness estimates, as our CFV-based method of generating Tamps had a significant error rate: while 20/25 Tamp strains generated in our chr II-targeted pool had the correct karyotype, only eight of the 16 Tamp strains we tested from our genome-wide pool had the correct karyotype as determined by aCGH (see GEO Accession GSE67769). The abnormal Tamp karyotypes included, most commonly, amplicons initiating at the correct genomic location but not extending to the proximal telomere and, occasionally, contained other large aneuploid events (S7 Table). To mitigate the effect of biological replicates for which fitness was mismeasured due to incorrect Tamp formation or background mutations, we first filtered out Tamps with highly variable fitness estimates between biological replicates: this excluded approximately 20% of all Tamps from subsequent analysis and, as summarized above, left 1,631, 1,596, and 1,551 Tamps in glucose-, phosphate-, and sulfate-limited conditions, respectively, for further analysis (see S1 Text for additional details). Next, we combined data from all biological replicates for a given Tamp to obtain a more accurate estimate of each Tamp’s fitness (S5 Fig, See Materials and Methods). Specifically, for those Tamps with more than 15 biological replicates (approximately 55% of remaining Tamps), we used the mode of the fitness distribution described by all biological replicates as the Tamp fitness; when 15 or fewer biological replicates were available, simply the mean of the biological replicates was used as the Tamp fitness (See S1 Text). We confirmed the overall accuracy of our methods in 24 control experiments competing eight Tamp strains in all three nutrient-limiting conditions in head-to-head competition experiments against an appropriate GFP-marked control strain (S8 Table). We found that the fitnesses determined in our genome-wide screen agreed well with those determined in head-to-head competition experiments of aCGH-validated strains (S2B Fig, adjusted R2 = 0.64).
When we plotted the fitnesses of each Tamp across the genome, we noticed that, similar to the chrII-targeted screen, neighboring Tamps typically had similar fitnesses, which defined plateaus bordered by distinct fitness breakpoints (S6 Fig and S7 Fig). As described above, we hypothesized that “Downsteps” in fitness could be used to identify driver genes that, under the condition tested, increased fitness when amplified in the context of a Tamp. Similarly, sharp increases in fitness, or “Upsteps,” could be used to identify anti-driver genes that, when amplified in the context of a Tamp, decreased fitness under the tested condition.
After we observed the stepwise shape of this fitness profile, we used DNAcopy [43], an analysis program typically applied to aCGH data to identify regions of similar copy number as well as copy number variant (CNV) breakpoints via circular binary segmentation, to define fitness plateaus and fitness breakpoints in our Tamp fitness data (S9 Table, See Materials and Methods). Segmenting our genome-wide fitness data in this way generated a summary view of the fitness effects of Tamps. We believed this analysis approach was well suited to our data because, similar to CNVs analyzed by aCGH, we expected our fitness data to be somewhat noisy and for neighboring Tamps to have similar fitnesses.
As a good example of our analysis approach, Fig 5B visualizes the results of our genome-wide Tamp screen for chromosome II under sulfate-limiting conditions. The top panel of Fig 5B depicts as blue lines the 122 Tamps spanning chromosome II for which we determined fitnesses; each Tamp initiates at a different location along chromosome II and extends to the proximal telomere. The fitness of each Tamp is plotted in the bottom panel of Fig 5B directly below its corresponding vertical blue line (see the red arrow for one example). Segmenting our genome-wide fitness data using DNAcopy defined fitness breakpoints that are outlined with the yellow and teal stacked boxes: yellow boxes enclose Tamps that increased fitness, while teal boxes enclose Tamps that decreased fitness. Just as we observed in our chromosome II–targeted pool, Tamps that amplified the right arm of chromosome II, where SUL1 is located, increased fitness under sulfate-limiting conditions. Note that the Downstep telomeric of SUL1 we observed in the chrII-targeted pool was not observed in the genome-wide pool because we did not include any Tamps initiating between SUL1 and the telomere in our genome-wide Tamp pool.
Although the incorrect karyotypes of individual Tamp biological replicates is an unfortunate by-product of our methodological approach, our analysis pipeline significantly ameliorated this limitation. We are therefore confident that this genome-wide Tamp screen provided an accurate description of the fitness effects of a complex pool of Tamps. As such, our method provides a systematic view of the fitness landscape described by Tamps under multiple selective conditions.
Next, we asked whether the fitness effects of Tamps were always condition dependent or if there were some Tamps that commonly increased or decreased fitness across the conditions we examined. Our genome-wide Tamp screen identified a unique list of fitness breakpoints in each of the three conditions we examined. The union of these three lists thus defines the minimum number of regions showing a change in fitness compared to neighboring regions in at least one condition. Specifically, we identified 175 regions with different fitnesses in at least one condition. We compared the fitnesses of these 175 regions between conditions and generally found little correlation between conditions (Fig 6A–6C). However, a few regions had common fitness effects between conditions; four and seven of the 175 regions increased or decreased fitness by >5%, respectively, in all three conditions.
As examples of our Tamp dataset, the fitnesses of Tamps from four chromosomes are shown in Fig 6D. Similar to Fig 5B, in this figure we have shown stacked boxes that represent groups of Tamps with similar fitness as defined by our segmentation of the genome-wide fitness profile with DNAcopy. However, in this figure we have not plotted raw Tamp data as we did in Fig 5. As we were interested in comparing the fitness effects of Tamps between conditions, we summed the fitness effects of each Tamp under all three conditions. To emphasize regions that have common effects between all three conditions, in the “Summary” section of Fig 6D we displayed as stacked boxes only those Tamps with the same fitness effect under all three conditions (i.e., >5% fitness advantage or disadvantage). The relative fitness of these boxes represents the sum of the fitness effects under all three conditions. Notice that some of the stacked boxes appear to be missing from the”Summary” section of Fig 6D. This is because only a few regions had common fitness effects between all conditions; boxes enclosing regions with different fitness effects under different conditions are excluded from the “Summary” section.
While chromosome II and XIV lacked any region with a common fitness effect across all three conditions, chromosomes V and XI both contained regions that were either universally advantageous or detrimental when amplified. For example, amplification of the left arm of chromosome XI decreased fitness under all three conditions (Fig 6A–6C, grey circles, and Fig 6D). Other Tamps showed common fitness effects in two of the three conditions we tested. For example, amplification of the left arm of chromosome XIV increased fitness not only under glucose-limiting conditions but also under phosphate-limiting conditions (Fig 6B, red arrow, Fig 6D).
Next, we examined two regions recurrently amplified in the set of evolution experiments examined here and those previously described [12]. The right arm of chromosome V was amplified in three different evolution experiments carried out under the three nutrient-limiting conditions. Similarly, the genome-wide Tamp screen predicted a 51 kb Tamp on the right arm of chromosome V to increase fitness by approximately 6%–7% under all three conditions (Fig 6D). However, the Tamps observed in the evolved populations were actually somewhat larger (84–440 kb) than this 51 kb Tamp. It is notable that the chromosome V amplicon in two of the three evolved populations initiated at the closet Ty element centromeric of this 51 kb high-fitness Tamp. The Tamp screen predicted the chromosome V amplicons observed in the evolved populations to affect fitness by +6%, -3%, and -1% under sulfate-, phosphate-, and glucose-limiting conditions respectively (S10 Table). In summary, while our Tamp screen predicted that amplification of 51 kb on the right arm of chromosome V is commonly advantageous, the precise amplifications observed in our evolution experiments were predicted to be neutral under glucose- and phosphate-limiting conditions and to increase fitness only under sulfate-limiting conditions.
The recurrent amplification on the right arm of chromosome XIV has been observed in three independent glucose-limited evolution experiments [10,12]. Consistent with these observations, the genome-wide Tamp screen predicted this event to increase fitness by >20% under glucose-limited conditions (Fig 6D). Interestingly, our genome-wide screen predicted a smaller Tamp on the left arm of chromosome XIV to increase fitness under phosphate-limiting conditions; however, no such amplicon has been yet reported in any phosphate-limited evolution experiment. Chromosome XIV left-arm Tamps were predicted to have a nearly neutral effect on fitness under sulfate-limiting conditions (< 2% fitness increase). A similar rearrangement was also previously identified as yeast “chromosome XVII” because of an aberrant karyotype in the original genetic mapping strains, suggesting this amplification may have fitness benefits in other conditions as well [44].
The dataset from our Tamp screen allowed us to ask general questions about the relationship between aneuploidy, specifically telomeric amplicons, and fitness. First, we compared the fitness of each Tamp to its size in base-pairs and found little correlation (adjusted R2 = 0.05, S2C Fig). Although Tamp truncation was not an insignificant problem in our dataset, our analysis approach, by filtering out Tamps with high intra-replicate variation in fitness and using the mode of the biological replicate fitness distribution to estimate fitness, partially ameliorated the effects of incorrectly sized replicates on each Tamp’s fitness estimate.
Next, we took advantage of data previously generated by our lab that determined the fitness effects of single-gene amplifications genome-wide under the same conditions explored in our Tamp screen [34] (see S1 Text). We compared the fitness distribution defined by our genome-wide Tamp screen to the fitness distribution defined by single-gene amplifications [34] (Fig 7A). We found that the distribution of Tamp fitnesses was much broader than that defined by single-gene amplifications. Additionally, we noted that distribution of Tamp fitnesses appeared bimodal, with one negative fitness peak and a second positive fitness peak. This result supports the hypothesis that aneuploid events are mutations that have large effects, positive and negative, on organisms’ fitness.
Aneuploid events are hypothesized to be highly pleiotropic: a characteristic that may explain their eventual supplantation by more targeted mutation types [37]. To test this hypothesis, we defined pleiotropy as the between-condition variance in fitness. Taking advantage of the same genome-wide single-gene amplification dataset referenced above [34], we compared the density distributions of variance in fitness of Tamps to those of single gene amplifications. We found that Tamps described a much broader distribution than that described by single gene amplification (Fig 7B). These results support the hypothesis that aneuploid events are pleiotropic.
The data from our genome-wide Tamp screen, combined with our lab’s previous data describing the fitness effects of single-gene amplifications genome-wide, also allowed us to explore the genetic basis of aneuploidy’s effects on cellular fitness. First, we asked if the fitness of any given Tamp could be predicted by the average of the fitness effects of all single-gene amplifications within the boundaries of the Tamp. We found that the average of the fitnesses of single-gene amplifications for the genes contained within a Tamp did not predict the fitness of the Tamp itself (S2D Fig) Next, we explored the alternative hypothesis that only a few genes within a Tamp are centrally important for effecting the fitness of the entire amplicon. This hypothesis was additionally supported by the stair-step shape of the Tamp fitness landscape: if many genes within a Tamp contributed to the fitness effects observed, one would expect a smooth fitness profile in which the addition or loss of individual genes from the amplicon produced an incremental change in fitness; instead, the fitness profile produced by our Tamp screen often revealed plateaus in fitness bordered by discrete fitness breakpoints. As discussed above, we hypothesized that Downsteps in the Tamp fitness landscape identified driver genes that increased fitness when amplified, while Upsteps identified anti-driver genes that decreased fitness when amplified. Combining fitness data from all three conditions, we identified 181 fitness breakpoints: 77 Downsteps and 104 Upsteps. As our genome-wide Tamp screen did not contain a Tamp for every gene in the genome, each Downstep or Upstep region necessarily overlapped several genes. We averaged the fitness effects of single-gene amplification for genes overlapping each of the 181 fitness breakpoints and found that the average fitness at Downsteps was significantly greater than at Upsteps (Fig 7C, unpaired, two-tailed t test, p = 0.008, raw data in S17 Table). These results supported the hypothesis that one or few gene(s), located at Upsteps and Downsteps, were primarily responsible for effecting the fitness of each Tamp.
By identifying driver and anti-driver genes respectively, Downsteps and Upsteps can be used to identify potential genetic targets of adaptation. In fact, under both glucose- and phosphate-limiting conditions, but not under sulfate-limiting conditions, the genes overlapping Upsteps were enriched for genes mutated in populations evolved under the corresponding nutrient limitation (Fig 7D, Fisher’s exact test, p = 2.6 x 10–4 and p = 0.027 for glucose- and phosphate-limiting conditions, respectively) [34]. As we expected Upstep genes to decrease fitness when amplified, we might therefore have expected that lower levels of expression of these same genes would increase fitness. Our results thus agree with the recent observation by Kvitek and Sherlock that the majority of mutations selected in haploid yeast evolved under glucose-limited conditions are loss-of-function mutations [45]. There are additional similarities between the glucose-limited Upstep genes identified in our Tamp screen and the genes mutated in glucose-limited evolution experiments. First, glucose-limited Upstep genes are enriched for Gene Ontology (GO) terms closely related to those enriched in the group of recurrently mutated genes identified by Kvitek and Sherlock in glucose-limited evolution experiments (“intracellular signal transduction,” Fisher’s exact test, Holm–Bonferroni corrected p = 0.01, and “regulation of intracellular signal transduction,” Fisher’s exact test, Holm–Bonferroni corrected p = 0.015). Second, the genes located at Upsteps in our glucose-limited Tamp screen are enriched for genes observed by Kvitek and Sherlock to be recurrently mutated in glucose-limited evolution experiments and include: HOG1, IRA2, LCB3, PBS2, PDE2, and SSK2 (Fisher’s exact test, p = 6.4 x 10–6).
Given the large number of genes overlapping several Downsteps (up to 30 genes), we sought to filter the list of Downstep genes and produce a list of high-quality candidate driver genes. We filtered the list of Downstep genes by comparing it to several published datasets: the list of genes commonly up-regulated in clones evolved under glucose-, phosphate-, or sulfate-limiting conditions [10,46]; the list of genes that increased fitness when present on a low-copy-number plasmid under glucose-, phosphate-, or sulfate-limiting conditions [34]; and the list of genes mutated in populations evolved under glucose-, phosphate-, or sulfate-limiting conditions [34]. After this filtering, we identified a total of 100 candidate driver genes important for increasing fitness in the context of a telomeric amplicon under the three nutrient-limiting conditions we examined here (S11 Table). Importantly, our filtered list of candidate driver genes still identified at least one driver gene at most Downsteps (58 out of 77, or 75%).
Although we expected our method to identify driver genes that, when amplified, individually increased fitness, we also expected our method to identify genes that increased fitness only when amplified in the context the Tamp. In fact, 12 of the 73 candidate driver genes identified here have a negative effect on fitness when amplified individually (S11 Table). We hypothesized that these 12 driver genes in particular must synergistically interact with one or more genes coamplified on the Tamp. The synergistic partners of the identified driver genes are probably located between the identified driver gene and the telomere. As these telomeric synergistic partners would only be expected to affect fitness when coordinately amplified with our currently identified driver genes, they would not be expected to produce a step in the fitness profile. Identification of these pairs or groups of synergistically interacting genes remains a target of future research.
With the fitness data from the genome-wide Tamp screen and this list of candidate driver genes in hand, we returned our analysis to the aneuploid events observed in our laboratory-evolved populations as well as aneuploid events previously documented in a similar set of evolution experiments [12]. Fitness data from our Tamp screen predicted that 11 of the 16 telomeric amplicons identified in evolved populations increased fitness under their corresponding conditions, while the remaining five telomeric amplicons were likely passenger mutations (two of these five amplicons represented the chromosome V amplicons observed in sulfate- and phosphate-limited populations and discussed above) (S10 Table). Importantly, our Tamp screen allowed us to predict the fitnesses of telomeric amplicons that are difficult to test by traditional genetic means, as they are linked to large deletions that rendered any haploid spore intermediary inviable. For each telomeric amplicon observed in an evolved population, we estimated the number of driver genes within its length by counting the number of Downsteps it overlapped (S10 Table). Typically, evolved Tamps overlapped one to three Downsteps, suggesting that only a few driver genes were primarily important for determining the fitness increase associated with these aneuploid events. As mentioned above, we have not yet identified the synergistic partners of these driver genes. The main exception to this statement is the amplification of the left arm of chromosome XIV recurrently observed in populations evolved under glucose-limiting conditions (Fig 7E). This large amplification overlapped six Downsteps and was predicted by our screen to increase fitness by >20%. There were multiple candidate driver genes along this segment, including YNL019C, RPL16B, OCA1, RAS2, YNL095C, SKO1, BNI5, YNL162W-A, PEX6, and EGT2.
The data from our Tamp screen proved useful in addressing general questions about the genetic basis for aneuploidy’s effect on cellular fitness and identified potentially novel driver genes that are important for increasing fitness in the context of aneuploidy. Furthermore, we have used data from our Tamp screen to predict the fitness effects of telomeric amplicons observed in evolved populations that are otherwise not amenable to traditional genetic analyses.
Our survey of aneuploid events identified in populations of S. cerevisiae evolved in nutrient-limited chemostats produced circumstantial evidence for aneuploidy’s positive effect on cellular fitness: aneuploid events rose to high population frequencies, and clones isolated with aneuploid karyotypes had fitnesses greater than wild type. In addition, we found that evolved aneuploid clones had a significantly greater relative fitness than evolved euploid clones. However, as the aneuploid and euploid clones were also different with respect to their genetic background, the nutrient-limiting conditions of their evolution experiment, and the number of generations that they were grown in the chemostat [10], there are several possible confounding explanations for their significant difference in fitness.
Three out of four aneuploid events, for which we directly determined the fitness, were sufficient to increase fitness relative to wild type. Each, however, showed a different relationship to the overall fitness of the original corresponding evolved clone, demonstrating that aneuploid events show varying degrees of epistasis with the other mutations acquired over the course of evolution. Interestingly, the VR t XCEN supernumerary chromosome isolated from the sulfate-limited population S8, despite occupying a substantial proportion of the population (13%), decreased fitness under sulfate-limiting conditions. Furthermore, this supernumerary chromosome contained a telomeric amplification of the right arm of chromosome V that was recurrently amplified in three different populations evolved under three different nutrient-limited conditions. Both the population frequency of this event as well as its recurrence were strongly suggestive of its selection under sulfate-limiting conditions. However, the S8c2 VR t XCEN supernumerary chromosome actually decreased fitness by 10% under sulfate-limiting conditions. It is possible that this discrepancy can be explained by the non-transitive relationship of fitness that has previously been observed over the course of laboratory evolution [47]. Epistasis may also explain this result, as the SUL1 amplicon alone from clone S8c2 increased fitness to a similar extent as that observed with the original evolved clone; this suggests that the fitness effects of the VR t XCEN supernumerary chromosome were fairly neutral in the context of a SUL1 amplification. These results argue that the VR t XCEN supernumerary chromosome is a passenger mutation. This is consistent with previous findings that showed genetic hitchhiking to be important to the spectrum of mutations observed in populations of asexually dividing cells [33,48]. Given the strong effects of epistasis and genetic hitchhiking on mutation frequency, these results should offer a strong cautionary message to the sole reliance on recurrence and population frequency for differentiating driver mutations from passenger mutations.
Although the remaining telomeric amplicons observed in the laboratory-evolved populations examined in this study were all concurrent with large deletions, making their genetic isolation difficult using traditional techniques, data from our Tamp screen allowed us to predict that the majority of these laboratory-evolved Tamps increased fitness in the conditions under which they were observed. In contrast, only one out of the 12 non-synonymous mutations we tested, a missense mutation in PHO84 isolated from the phosphate-limited evolved clone P6c1, increased fitness by more than 5%. Consistent with this observation, we observed a broader range of fitness effects in our Tamp screen than in a genome-wide screen for the fitness effects of single-gene amplifications (Fig 7A) [34]. These results show that aneuploid events are important drivers of increased fitness in populations of S. cerevisiae evolving under nutrient limiting conditions. Furthermore, these data are consistent with the hypothesis that aneuploid events allow evolving populations to broadly explore a fitness landscape by prompting large jumps in fitness unattainable by the mutation of single genes [23,37]. Aneuploidy is likely particularly important for the adaptation to novel conditions.
Fitness data from our Tamp screen and from competition experiments with aneuploid events and evolved aneuploid clones confirmed that Tamps, and aneuploidy more generally, are pleiotropic mutations with typically condition-dependent fitness effects; most aneuploid events and clones had decreased fitness under alternative conditions. Occasionally, as observed in both the Tamp screen and in direct fitness assessments with evolved clones and evolved aneuploid events, similar fitness effects were observed between conditions. Notably, similar fitness effects under different conditions were observed with the VR t VIR supernumerary chromosome isolated from the phosphate-limited population P6, which increased fitness to a similar extent under both phosphate- and glucose-limiting conditions. Particularly surprising was the observation that the supernumerary chromosome isolated from the sulfate-limited population S8, which decreased fitness by 10% under sulfate-limiting conditions, actually increased fitness by 11% under glucose-limiting conditions. As all of the competition experiments were carried out under conditions of chemostat growth, it is possible that some of the aneuploid events with common fitness effects across nutrient-limiting conditions affected growth under continuous culture generally.
Mutations such as the S8 supernumerary chromosome might contribute to the increased adaptability of aneuploid cells: an aneuploid event acting as a passenger mutation under a cell’s current condition could provide a dramatic increase in fitness under a novel condition. This conversion of a passenger mutation to a driver mutation may be more likely to occur with aneuploid events than with point mutations or single-gene changes in copy number because of the number of genes affected by a single aneuploid event. For example, although not yet observed in phosphate-limited evolution experiments, our Tamp screen predicted the amplification of the left arm of chromosome XIV to increase fitness under both glucose- and phosphate-limiting conditions (Fig 6D). However, only one of the ten candidate driver genes present within this amplicon is predicted to be responsible for the increased fitness of this amplicon under both phosphate- and glucose-limiting conditions. By affecting the copy number of many genes simultaneously, aneuploid events are necessarily pleiotropic. However, while aneuploid events may also show similar fitness effects under different conditions, these fitness effects are likely mediated through the copy-number change of distinct groups of driver genes. These data emphasize the condition-dependent nature of aneuploidy’s effect on cellular fitness and may help address the “aneuploidy paradox”: the observation that while aneuploidy typically decreases a cell’s proliferative ability, it increases fitness under certain conditions [2,3,5,7].
The data from our Tamp screen allowed us to investigate general aspects of the relationship between aneuploidy and cellular fitness. The data presented here are further support for the current hypotheses that aneuploidy is both a large effect-size mutation and that it is more pleiotropic than single-gene changes in copy number. As aneuploidy generally decreases cells’ proliferative ability, one might have expected larger Tamps to increase fitness to a lesser extent than smaller Tamps as the burden of carrying such a large Tamp outweighed any benefit due to amplification of genes along its length. However, our Tamp data show that there is no overall correlation between size and fitness of the Tamps examined in our screen. Although there was no overall correlation between fitness and Tamp size in our data, there was a distinct negative relationship between size and Tamp fitness on the right arm of chromosome II under sulfate-limiting conditions (slope = -0.0087 relative fitness/kb, adjusted R2 = 0.87). A more detailed analysis of the fitness data produced by our Tamp screen may reveal a more general relationship between Tamp size and fitness. In addition, as it is likely that Tamps initiating far from a telomere are less likely to complete break-induced-replication (BIR), thus resulting in truncated amplicons, our results here, while ameliorated by our analysis pipeline, likely represent an underestimate of any deficit correlated with size.
Our Tamp screen revealed that amplicons with increased fitness could not be differentiated from amplicons with decreased fitness simply by averaging the fitness effects of all single-gene amplifications along their lengths. However, when we focused on fitness breakpoints, we were able to differentiate increases in fitness (Upsteps) from decreases in fitness (Downsteps) by averaging the fitness effects of all single-gene amplifications overlapping the breakpoint region. These results suggested that a minority of genes were responsible for an amplicon’s fitness effects. In fact, we have previously shown this to be true for Tamps of the right arm of chromosome II and the amplification of SUL1 under sulfate-limiting conditions (increase in fitness due to SUL1 on a low-copy number plasmid = 23% [30], increase in fitness due to 60 kb telomeric amplicon overlapping SUL1 = 16%). However, the amplification of SUL1 under sulfate-limiting conditions is a clear outlier, in that it increases fitnesses much more than any other single-gene amplification under the three nutrient-limited conditions examined here. Synergistic effects between the driver genes identified in this study and a small number of as-yet-unknown interaction partners located distally along the amplicon are likely responsible for the fitness effects of most amplicons.
While the models tested here limited the interaction between genes within an amplicon to be simply additive, we acknowledge that the interactions between genes within aneuploid regions are likely to be much more complex and warrant further study. In fact, we have already tested a method to confirm the identity of driver genes and reveal synergistic partners within a Tamp. Focusing once more on the right arm of chromosome II, we created 21 independent strains that each paired a single 60 kb Tamp with deletion of a different gene along this amplicon. As expected, under sulfate-limiting conditions, deletion of SUL1 eliminated the fitness increase due to this 60 kb Tamp (S8 Fig). Genome-wide application of this method, or focused application to amplicons of particular interest, for example the amplification of the left arm of chromosome XIV under glucose-limited conditions, would further illuminate the types of interactions between genes contained within aneuploid regions.
While the data produced by our Tamp screen have allowed us to gain a genome-wide view of the fitness landscape explored by Tamps, it is prudent to highlight some of the limitations of this dataset. First, as noted above, there is a high error rate in the formation of Tamps with the method employed here. Despite our attempts to account for these errors in both our experimental design (i.e., incorporating a biological-replicate barcode into each Tamp strain) and analysis pipeline, future experiments would benefit from an improved experimental approach. One approach would be to identify replicate barcodes that were associated with Tamps of inappropriate sizes and exclude these barcodes from the analysis. This could be accomplished by pairing Pulse-Field Gel Electrophoresis with gel extraction and barseq to determine the actual size of each barcoded Tamp strain. Second, the segmentation approach used to fragment the genome-wide Tamp fitness data into regions of approximately equal fitness may be an oversimplification of these data, and a more detailed examination of the fitness changes across the genome is warranted.
Our genome-wide Tamp pool represents all possible amplicons at 3–4 gene resolution, however, previous studies have shown that rearrangements, including those that produce the Tamps studied here, are often mediated by repetitive elements in the S. cerevisiae genome, such as Ty elements [12]. As such, it becomes informative to compare the telomeric amplicons observed in evolution experiments to our Tamp screen and ask if the most advantageous Tamps are selected during the course of laboratory evolution. The recurrent amplicon of the right arm of chromosome V provides a good example of the benefits of this analysis. As described above, the chromosome V amplicon observed in both phosphate- and sulfate-limited evolution experiments is larger than the highest-fitness Tamp on the right arm of chromosome V identified in our screen. This is likely due to the fact that there are no Ty elements or repetitive regions closer to the fitness breakpoint identified by our Tamp screen than the one employed to form the amplicons observed in the evolution experiments. This provides an example where the genomic context likely restricted the formation of the most advantageous amplicon under these conditions.
Generally, however, the regions commonly amplified in evolution experiments are the highest-fitness Tamps identified by our screen. The top 54 most advantageous Tamps identified by our screen under sulfate-limiting conditions all overlap the SUL1-containing region on the right arm of chromosome II. Similarly, amplification of the left arm of chromosome XIV has been observed repeatedly under glucose-limiting conditions; 15 of the top 16 most advantageous Tamps identified by our screen under glucose-limiting conditions overlap the left arm of chromosome XIV. Unlike sulfate- and glucose-limited evolution experiments, populations evolved under phosphate-limited conditions show no such obvious recurrent amplification. However, the most fit Tamp predicted by our screen is the amplification of the left arm of chromosome XVI, although of a slightly smaller size than that predicted to be advantageous under glucose-limiting conditions (Fig 6). Although this amplification has not yet been observed in any phosphate-limited evolution experiment to date, there are several Ty elements in the region where positive-fitness Tamps initiated in our screen, so its absence cannot easily be explained by a genomic context unfavorable to amplification.
By combining data from a genome-wide telomeric amplicon screen and detailed analyses of clones and aneuploid events isolated from laboratory evolution experiments, we have provided details about the relationship between aneuploidy and cellular fitness. These data identified new candidate driver genes, the copy number changes of which are important for fitness, contribute to our understanding of how aneuploidy acts at the cellular level, and add to our understanding of aneuploidy’s role in adaptation and evolution. Recent advances in the direct targeting of DNA breaks in human cells [49], combined with the wealth of information generated from the sequencing of cancer genomes, may allow a similar comparative approach to be applied to the effects of aneuploidy in cellular proliferation and its role in the evolutions of cancers. We also note that the same experimental design could be applied to conditions in which aneuploidy is known to be detrimental as a way to identify critical dosage-sensitive genes.
The strains, plasmids, and primers used in this study are listed in S12 Table, S13 Table, and S14 Table, respectively. Unless specified below, yeast strains were grown at 30°C and standard media recipes were used.
We generated population DNA from archived glycerol stocks for the 14 evolution experiments not previously examined and determined the population frequency of aneuploid events by aCGH. We confirmed the accuracy of this approach by comparing the population frequencies of aneuploid events in population P7, as determined previously in [10] from fresh population DNA samples, to the frequencies determined by the method described here and found similar results. All aCGH data are available at GEO Accession GSE67769. In addition, we used a PCR assay that amplified the breakpoint of the VR t XCEN translocation event present in population S8 using primers OAS005–0AS0008. This breakpoint PCR assay identified the VR t XCEN supernumerary chromosome in 13 of 98 total clones tested (13%); our population aCGH determined the frequency of VR t XCEN supernumerary chromosome to be 13%.
To determine relative fitness, we competed individual clones of test strains against an appropriate control strain with eGFP integrated at the HO locus in nutrient limited chemostats. We used both large volume (approximately 300 ml) and small volume (20 ml) [50] chemostats for competition experiments. A single colony of each control or test strain was used to start an overnight culture in the same media in which the competition experiment was to be carried out; the overnight culture was then grown at 30°C for approximately 12–36 h. 1 ml of this overnight was used to inoculate each chemostat, which was then allowed to grow at 30°C without dilution for approximately 30 h, at which point fresh media was added to the culture chamber at a rate of 0.17 hour-1. After achieving steady-state, 50% of a control-strain chemostat was mixed with 50% of a test-strain chemostat, resulting in two chemostat replicates for a single competition experiment. Flow-cytometry using a BD Accuri C6 flow cytometer (BD Biosciences) at regular intervals throughout the competition allowed us to track the percentage of GFP-marked control cells over time. The data were plotted with ln[(dark cells/GFP+ cells)] versus generations, and we defined the slope of this relationship as the relative fitness of the test strain. The number of replicate competition experiments, as well as the appropriate control strain, is detailed for all test strains in S2 Table.
Two Tamp strains were constructed individually by direct transformation with a chromosome-fragmentation vector (CFV). 250 bp of homology to the genomic location at which we desired to create a Tamp was cloned into the multiple cloning site of the previously designed CFV YCF4 [39]. The appropriate CFV was then transformed into a haploid FY background strain to create chrIIR-Tamp 1N and chrVR-Tamp 1N and the karyotype confirmed by aCGH (see GEO Accession GSE67769). These haploid strains were backcrossed to create chrII-Tamp 2N and chrVR-Tamp 2N.
DNA samples from evolved clones and populations were prepared for WGS using Illumina Nextera kits according to the provided protocol. Libraries were sequenced on either an Illumina HiSeq or a GAII, generating the number of reads detailed in S3 Table. Reads were aligned with BWA [51] and SNVs were called using samtools [52] after applying standard filters. Population frequency of SNVs from population samples was determined from the allele frequency displayed in Integrative Genome Viewer (IGV) [53]. The clones and populations analyzed here (P6c1, P6, P5c3, P5, S8c2, and S8) were included in a previous analysis [34] and the raw data are deposited at BioProject ID PRJNA248591 and BioSample numbers SAMN02800460 (S8c2), SAMN02800438 (P6c1), SAMN02800436 (P5c3, run 1), SAMN02800435 (P5c3, run 2), SAMN02800403 (CEN.PK WT diploid, run 1), and SAMN02800404 (CEN.PK WT diploid, run 2).
To isolate individual mutations (both SNVs and aneuploid events) identified by WGS of the evolved clones P6c1 and P5c3, we backcrossed each evolved clone to an isogenic wild-type strain of the opposite mating type, sporulated, and dissected tetrads using standard sporulation media and protocols. Evolved clone S8c2, a diploid, was itself sporulated and tetrads dissected using standard sporulation media and protocols. After Sanger-sequencing confirmed the SNVs identified by WGS, tetrads were genotyped by Sanger sequencing and backcrossed repeatedly until each SNV and aneuploid event was isolated into an otherwise wild-type background. Spores isolated from S8c2 with the desired genotype were backcrossed a final time so that each mutation was once again in a diploid background. The karyotypes were confirmed by aCGH for all clones eventually used for relative fitness competition experiments (see GEO Accession GSE67769).
In order to compare the pleiotropic effects of aneuploid events and single-gene changes in copy number we calculated the between-condition variance in relative fitness for each mutation (aneuploid event or single-gene amplification) under the three nutrient-imitated conditions examined. Specifically, for each aneuploid event examined in Fig 3 we determined the between-condition variance in fitness. Next, we performed the same calculation for all single-gene amplifications as determined previously [34]. In this study, Payen et al. determined the fitness effects of single-gene amplifications by pooled competition experiments with a genome-wide collection of yeast ORFs cloned into a low-copy-number (CEN) plasmid [54]. We compared the distribution of fitness differences defined by single-gene changes in copy number to that observed with the aneuploid events examined in Fig 3 (S2A Fig) using an unpaired, two-tailed t test.
In order to create Tamps from deletion collection target strains we constructed two unique CFVs to target the KanMX cassettes that replaced Watson and Crick genes, pABS003 and pABS004, respectively. Primers OAS009 and OAS010 were used to amplify the KanMX cassette region, which was cloned into the BamHI and EcoRI sites of the CFV YCF4 to produce pABS003. Primers OAS011 and OAS012 were used to amplify the KanMX cassette region, which was cloned into the BamHI and EcoRI sites of the CFV YCF4 to produce pABS004. We then transformed 26 heterozygous yeast deletion target strains with a version of the appropriate CFV linearized with NotI.
Overall, 20 of the 26 heterozygous deletion strains yielded transformants with the expected Tamp. For our subsequent experiments, we chose to pool 21 of the Tamp strains, including the ybr282wΔ/+ Tamp strain, which also carried an extra copy of chromosome II. We added to this pool the yal066wΔ/+ heterozygous deletion collection strain to act as a wild-type fitness control; YAL066W is a pseudogene. To make the final pool that was used in subsequent competition experiments, the 21 Tamp strains plus the surrogate wild-type control strain were inoculated in minimal media, grown for approximately 12 h at 30°C, the cell densities were normalized, and all 22 strains were pooled together. 2 ml glycerol stocks made with 1 ml 50% glycerol plus 1 ml pooled culture were saved at -80°C.
To determine the fitness effects of the 21 Tamps in the chrII-targeted pool, we performed chemostat competition experiments with this pool in triplicate under sulfate-, glucose-, and phosphate-limiting conditions. At 5 time points throughout each competition experiment DNA samples were prepared and used to make barseq libraries with the PCR primers OAS013 and OAS014 or OAS029 and OAS030. After purification, the barseq libraries were pooled and loaded onto an Illumina HiSeq. The 6 bp barcode used for multiplexing the samples onto a single lane are indicate in S15 Table. As these reads were obtained from a run that had been multiplexed with other samples unrelated to this study, we have made available tab-delimited files of the raw sequencing data that contain the multiplexing barcode in the first column and the Tamp BC read in the second column. These files can be found at BioProject ID PRJNA257895 with BioSample IDs SAMN02979479 and SAMN02980022 to SAMN029794825. To determine the relative fitness of each of the 21 Tamps in this pool we used an analysis approach that has been successfully used by our lab in a previous publication [34]. Briefly, the frequency of each Tamp at each time point was determined from the barseq reads using a custom pipeline. For each Tamp we then plotted the log2(frequency at time = t / frequency at time = 0) versus generations and the slope of the line was taken as the relative fitness. The relative fitness of the yal066wΔ/+ strain was set at 0 and all the other Tamp fitnesses were normalized to it. The relative fitnesses for all 21 Tamps under all three nutrient-limiting conditions are reported in S5 Table and plotted in S4 Fig. Occasionally, insufficient reads were obtained to calculate the fitness of a particular strain under a particular condition. In this case the fitness is noted as “NA.”
To develop a method that could confirm the identity of driver genes along a Tamp, we tested a method that paired a single large Tamp with single gene deletions along its length. We generated a MATα 60 kb chrII Tamp strain (chrII-Tamp 1N) as described above and crossed it to 22 MATa deletion strains corresponding to genes within this 60 kb region. These MATa deletion strains were from a minimally passaged collection derived from the yeast magic marker collection [40]. We pooled these 22 strains and competed them in the three nutrient-limiting conditions in triplicate as described for the chrII-targeted Tamp pool. Similarly, we performed barseq on these samples using the same protocol as described for the chrII-targeted Tamp pool. These barseq libraries were pooled together and sequenced on an Illumina HiSeq (the 6 bp barcodes used for multiplexing are reported in S15 Table) and 354,545,894 reads were obtained. As these reads were obtained from a run that had been multiplexed with other samples unrelated to this study, we have made available tab-delimited files of the raw sequencing data that contain the multiplexing barcode in the first column and the Tamp BC read in the second column. These files can be found at BioProject ID PRJNA257895 with BioSample IDs SAMN02979479 and SAMN02980022 to SAMN029794825. Fitnesses were determined for each strain as described for the chrII-targeted Tamp pool, except that they were normalized to the fitnesses of the 60 kb chrII amplification alone (strain “chrII Tamp 2N”) instead of yal066wΔ/+ and are reported in S5 Table and plotted in S8 Fig.
To construct the genome-wide Tamp pool, 2,254 neutral fitness strains ([34]; S4 Table) from the yeast heterozygous deletion collection (“Magic Marker” collection, [40]) were grown in YPD + G418 (200μg/ml) + 0.18 μg /ml His (+ 50uM riboflavin when recommended) for approximately 24 h at 30°C. We separated these deletion collection strains into two pools depending on the orientation of the KanMX cassette (S3 Fig): Watson-strand genes on the Left side of the centromere and Crick-strand genes on the Right side of the centromere (wlcr pool) and Watson-strand genes on the Right side of the centromere and Crick-strand genes on the Left side of the centromere(wrcl pool). We designed two CFVs, one for each pool, that were identical except for the orientation of the KanMX cassette: pAS006 (for the wlcr pool) and pAS007 (for the wrcl pool). In order to maintain a high complexity of the 12 bp replicate BC, approximately 20,000–30,000 Escherichia coli colonies transformed with pAS006 and pAS007, respectively, were scraped and used to prepare plasmid DNA (Wizard Miniprep) for yeast transformation.
The wlcr and wrcl yeast heterozygous deletion pools were each transformed with their appropriate CFV. The transformation efficiency with CFVs pABS006 and pABS007 was only about 20% (as determined by a PCR assay), so our pool of scraped colonies included both Tamp strains and original heterozygous deletion strains. However, the design of the PCR primers used to generate our barseq libraries (OAS021 to OAS023) only amplified the strain-identifying barcode from successfully formed Tamp strains. The total number of unique transformants collected was approximately 23,000 and approximately 20,000 for the wlcr and wrcl pools, respectively, and resulted in an average of 26 unique replicates for each Tamp. Given the large number of replicate BCs included in the CFVs pABS006 and pABS007, each transformant was identifiable by a unique combination of the strain-identifying barcode, as derived from the yeast deletion collection barcode (Tamp BC), and the replicate BC (Fig 4A).
To confirm the construction of this pool, we prepared barseq libraries for sequencing using primers OAS021 to OAS023 from aliquots of each pool. These barseq libraries were prepared as described for the chrII-targeted Tamp pool and sequenced on an Illumina MiSeq with sequencing primers OAS024 to OAS027, generating 4,348,080 reads. The fastq files for this barseq experiment are at BioProject ID PRJNA257895 with BioSample IDs SAMN02979480 to SAMN029794821. Additional details about these files are included in S15 Table. Analysis of the barcodes sequenced in this run confirmed that our pool was sufficiently complex to warrant further pooled competition experiments.
As revealed in the construction of our chrII-targeted Tamp pool, generating Tamps using CFVs was not an error-free process and variable karyotypes were sometimes produced. Unfortunately, this problem was exacerbated in the construction of the genome-wide Tamp pool with larger Tamps being more likely to have incorrect karyotypes. The most commonly observed incorrect karyotype was one where the Tamp initiated at the correct genomic location but did not extend all the way to the proximal telomere; this problem was most common for larger Tamps (S7 Table). We adjusted our analysis pipeline to try and correct for these variable karyotypes.
Similar to the chrII-targeted Tamp competition experiments, we inoculated nine total large volume (approximately 300 ml) nutrient-limited chemostats supplemented with 20 mg/L histidine with aliquots of our wlcr and wrcl pools (both pools were inoculated into a single chemostat). We performed pooled competition experiments under the three different nutrient limited conditions (phosphate-, glucose-, and sulfate-limited) in triplicate; chemostat inoculation and growth were the same as described for the chrII-targeted Tamp pool competition experiments. We defined each of the triplicate chemostat competition experiments as a technical replicate. For each of the nine chemostats, ten time points were taken throughout the competition experiment. For each time point, DNA was extracted and two barseq PCR reactions were carried out (one targeting wlcr Tamps and one targeting wrcl Tamps) using primers OAS021 to OAS023 and resulting in a total of 180 barseq samples. These 180 samples were pooled in equal proportions in two pools of 90 samples each. The pool, 6 bp barcodes used for multiplexing, and generations corresponding to each of the 180 samples are recorded in S15 Table. Each pool was sequenced on three lanes of an Illumina HiSeq, generating a total of 752,336,013 reads. These fastq files are deposited at BioProject ID PRJNA257895 with BioSample IDs SAMN02979482 to SAMN02980021. The method we used to determine the fitness of each Tamp in the pools can be found in S1 Text. The relative fitness for each Tamp and its error are plotted for each condition in S7 Fig.
When we plotted the fitnesses for each Tamp across the genome, we observed that parts of the fitness landscape had a stair-step appearance, in which fitness plateaus were bordered by sharp fitness breakpoints. In order to segment the genome into regions defined by Tamps of similar fitness, we applied the copy-number variant prediction software, DNAcopy [43], to our genome-wide fitness data using the following settings: we required a minimum of two adjacent fitness data to define a fitness plateau and a significance of 0.05 to call a fitness breakpoint. This segmentation defined a total of 250 fitness segments across the three different nutrient-limiting conditions (Colored boxes in S7 Fig).
Previously, our lab determined the fitness effects of single-gene amplifications genome-wide using pooled competition experiments followed by barseq of genome-wide ORF collections on both low-copy-number (CEN) and high-copy-number (2 μ) plasmids [34]. We compared these single-gene amplification data to our genome-wide Tamp data in three ways. First, we compared the kernel density estimates for the fitnesses defined by Tamps to the fitnesses defined by single-gene amplifications (Fig 7A). The kernel density estimates were computed in R. Next, we stratified the 250 groups of Tamps defined by DNAcopy as positive or negative and averaged the fitnesses of all single-gene amplifications contained within its length as determined by their low-copy-number (CEN) fitness effects (S2D Fig). Finally, we examined the breaks between fitness plateaus as defined by our DNAcopy segmentation analysis and categorized each break as either an Upstep (i.e., an increase in fitness moving along the chromosome towards the telomere) or a Downstep (i.e., a decrease in fitness moving along the chromosome towards the telomere). We averaged the fitnesses, as determined by their low-copy-number fitness effects, of all single-gene amplifications contained within each breakpoint region plus one gene centromeric of the centromeric border of the breakpoint region (Fig 7A). This extra gene was included simply to compensate for any insensitivity in the DNAcopy segmentation of our fitness data.
As described in the main text, we filtered the list of Downstep genes by comparing it to several published datasets: the list of genes commonly up-regulated in clones evolved under glucose-, phosphate-, or sulfate-limiting conditions [10,46]; the list of genes that increased fitness when present on a low-copy number plasmid under glucose-, phosphate- or sulfate-limiting conditions [34]; and the list of genes mutated in populations evolved under glucose-, phosphate-, or sulfate-limiting conditions [34]. Specifically, for the comparison with the Payen low-copy-number plasmid fitness data, we compared Downstep genes to Payen et al.’s list of outlier fitness genes with fitnesses <-0.10 or >0.10 (denoted as “CEN outlier” in S10 Table and S11 Table) and also to the set of genes with fitnesses greater than two standard deviations more than the mean fitness of that dataset (denoted as “CEN mean + 2SD” in S10 Table and S11 Table). CEN mean + 2SD genes still have extreme fitnesses but did not reach the stringent cutoff imposed in the Payen et al. study to be called “outliers.” For phosphate-limitation this included single-gene amplifications with fitnesses <-0.096 or >0.097, and for glucose-limitation this included single-gene amplifications with fitnesses <-0.052 or >0.050. The list of “outliers” called by Payen et al. already included all mean + 2SD genes for sulfate-limited conditions.
|
10.1371/journal.pntd.0002986 | Serology of Paracoccidioidomycosis Due to Paracoccidioides lutzii | Paracoccidioides lutzii is a new agent of paracoccidioidomycosis (PCM) and has its epicenter localized to the Central-West region of Brazil. Serological diagnosis of PCM caused by P. lutzii has not been established. This study aimed to develop new antigenic preparations from P. lutzii and to apply them in serological techniques to improve the diagnosis of PCM due to P. lutzii. Paracoccidioides lutzii exoantigens, cell free antigen (CFA), and a TCA-precipitated antigen were evaluated in immunodiffusion (ID) tests using a total of 89 patient sera from the Central-West region of Brazil. Seventy-two sera were defined as reactive for P. brasiliensis using traditional antigens (AgPbB339 and gp43). Non-reactive sera for traditional antigens (n = 17) were tested with different P. lutzii preparations and P. lutzii CFA showed 100% reactivity. ELISA was found to be a very useful test to titer anti-P. lutzii antibodies using P. lutzii-CFA preparations. Sera from patients with PCM due to P. lutzii presented with higher antibody titers than PCM due to P. brasiliensis and heterologous sera. In western blot, sera from patients with PCM due to P. lutzii were able to recognize antigenic molecules from the P. lutzii-CFA antigen, but sera from patients with PCM due to P. brasiliensis could not recognize any P. lutzii molecules. Due to the facility of preparing P. lutzii CFA antigens we recommend its use in immunodiffusion tests for the diagnosis of PCM due to P. lutzii. ELISA and western blot can be used as complementary tests.
| Tropical diseases, such as paracoccidioidomycosis, are the most common type of neglected diseases. From 1980 to 1995, 3,181 deaths from paracoccidioidomycosis occurred in Brazil, representing the eighth most common cause of death from predominantly chronic or recurrent types of infectious and parasitic diseases, showing considerable magnitude and low visibility. Paracoccidioidomycosis is traditionally assumed to be caused solely by Paracoccidioides brasiliensis, but a new species, Paracoccidioides lutzii, was discovered in the Central-Western region of Brazil. Thus, new antigenic preparations and tests for an accurate differential diagnosis between these two species appear to be needed. This study aimed to develop new antigenic preparations from P. lutzii isolates to improve the diagnosis of paracoccidioidomycosis. We used patient serum samples predominantly from the Central-Western region of Brazil. Various antigenic preparations were tested, and a cell free antigen derived from P. lutzii was an excellent antigen for serological diagnosis and able to diagnose 100% of sera from patients with PCM due to P. lutzii.
| Paracoccidioidomycosis (PCM) is a mycotic disease caused by species of the genus Paracoccidioides, a group of thermally dimorphic fungi that grow in mycelial form at room temperature and as budding yeasts when cultured at 37°C or in parasitism in host tissues. PCM is limited to Latin American countries, and the most important regions of endemicity are found in Brazil, Colombia, and Venezuela [1]. PCM presents as two major clinical forms: the acute or sub-acute form and the chronic form. In Brazil, PCM is considered the eighth most common cause of death among infectious and parasitic chronic diseases, with a mortality rate of 1.45 per million population [2]. However, PCM may be considered a neglected disease because very few regions of this country have official prevention programs.
The criteria for a definitive diagnosis of PCM are based on demonstrating the presence of the fungus as multiple budding cells in clinical materials. A culture is relatively difficult to obtain, and it is a slow procedure. As an adjunct to clinical and histological findings, serological tests can help establish a diagnosis. The detection of antibodies in serum has been one of the main tools for the diagnosis of PCM and may be useful for monitoring its evolution and response to treatment. Among the different serological techniques, the double immunodiffusion (ID) test is the most commonly used and has a sensitivity of 80 to 95% [3].
Enzyme-linked immunosorbent assay (ELISA) has been employed in the serology of various mycotic diseases [4]–[8]. However, ELISA has important limitations due to cross reactivity. In an attempt to improve the specificity of ELISA for the diagnosis of PCM, Albuquerque and Camargo [9] demonstrated that different procedures are inefficient for eliminating all cross-reacting antibodies and obtaining a specific diagnosis.
Exoantigens from cultures of P. brasiliensis were studied by immunoblotting, and components were detected using serum from patients with PCM. Anti-P. brasiliensis IgG reacted with four major components of 70, 52, 43, and 20–21 kDa. The 43 kDa glycoprotein (gp43) was the predominant IgG reactive antigen and recognized by 100% of the patient sera, and the 70 kDa glycoprotein was recognized by 96% of the tested sera. Both gp43 and gp70 can be considered to be markers for human PCM [10].
Until recently, PCM was assumed to be caused solely by P. brasiliensis and transmitted to humans by inhalation of fungal propagules from the mycelia phase occurring in nature. Recent publications [11]–[12] support the idea of several cryptic Paracoccidioides species being phylogenetically related. The real incidence of each phylogenetic species and their implication on ecology and clinical practice is difficult to establish because there is a lack of information in the literature concerning the distribution of these entities. The Phylogenetic Species Recognition (PSR) method based on genealogical concordance (GCPSR) [13]–[14] has been used to detect limits in many pathogenic fungal species [15]–[17]. Using GCPSR methodology and several isolates of P. brasiliensis from different regions of Latin America, Matute et al. [18] concluded that P. brasiliensis is not a monotypic taxon, but a complex of species: phylogenetic species 1 (S1), broadly distributed through the Latin America [18]–[20], phylogenetic species 3 (PS3) is restricted to clinical cases in Colombia [18]–[20], and a few isolates in the clade phylogenetic species 2 (PS2) have been reported [18]–[20] in Brazil and Venezuela. Phylogenetic analysis of 14 genes in 21 isolates revealed that isolate Pb01 cannot be grouped into any of these species and constitutes a new clade [21]. Seventeen new isolates belonging to this fourth cryptic species (Pb01-like) have been identified, 16 of which originate in the Central-West region of Brazil, and one from Ecuador [12]. Thus, Teixeira et al. [22] proposed a new species, Paracoccidioides lutzii, which was formerly known as ‘Pb01-like’ strains based on phylogenetic and comparative genomics data, recombination analysis, and morphological characteristics (Fig. 1). P. lutzii occurs mainly in the Central-West region of Brazil, but it was also recently described in the northern regions [23]. The Central-West region of Brazil is composed of the states of Mato Grosso, Mato Grosso do Sul, and Goiás, which can be regarded as the areas with highest incidence of PCM caused by P. lutzii. In the South and Southeast regions, P. brasiliensis is the predominant species (Fig. 2).
With the introduction of dissimilar species, PCM serology has been challenging for diagnosis. Several efforts highlight the antigenic variability of this complex, which may have led to a substantial number of false negative results. To date, no studies describing the standardization of P. lutzii antigens for PCM serology are available. In order to improve serological parameters for the diagnosis of PCM caused by P. lutzii, the present study presents a strategy for the immunodiagnosis of PCM caused by P. lutzii by testing three types of antigenic preparations using ID, ELISA, and Western blot.
This study was approved by the Research Ethics Committee with protocol number CAAE: 17177613.6.0000.5541 by Federal University of Mato Grosso (UFMT) and protocol number 1796-10 by Universidade Federal de São Paulo (UNIFESP). Protocol number of Universidade Federal do Mato Grosso do Sul (UFMS): 354.989. All adult subjects provided informed written consent and the study was approved by ethical committee under number 288.250/CEP/HUJM/UFMT.
Fourteen isolates of P. lutzii were obtained from different regions of Brazil (Table 1). Five isolates were obtained from the oral or cutaneous lesions of patients from Júlio Muller University Hospital (HUJM/UFMT) in Cuiabá, MT, in the Central-West region of Brazil. The clinical materials were collected by Dr. R. Hahn. Two isolates were obtained in the North region (Pará state) by Dr. S.H. Marques-da-Silva [23]. Three isolates were isolated in Goiás state (Central-West region) and donated by Dr. C.M.A. Soares, and four isolates were donated by Dr. E. Bagagli. P. lutzii was confirmed by genotyping [12].
Fungal DNA was extracted from suspected colonies and subjected to HSP70 gene amplification using the primers HSPMMT1 (5′-AAC CAA CCC CCT CTG TCT TG-3′) and PLMMT1 (5′-GAA ATG GGT GGC AGT ATG GG-3′) as described by Teixeira et al. [12] in order to identify an exclusive indel region of P. lutzii. P. lutzii Pb01 was used as a positive control and P. brasiliensis B339 as a negative control.
For the ID study, 89 serum samples from PCM patients (75 males and 14 females, age range 23–78 years) were evaluated: 55 from Mato Grosso state (Cuiabá and region, Central-West region) and 34 from Mato Grosso do Sul state (Campo Grande and region, Central-West region). All patients presented with the chronic form of the disease and exhibited clinical and laboratory signs of the disease with pulmonary system involvement and mucosal or mucocutaneous lesions. The diagnosis was endorsed by the clinical experience of the physician responsible for the patient. PCM was confirmed in most patients via direct examination of secretions, such as sputum, oral mucosa lesion samples, and biopsies. However, serological ID tests were not positive for all patients when using the traditional exoantigen from P. brasiliensis B339 (AgPbB339 standard antigen). As a gold standard, we used three sera from PCM patients with positive cultures for P. lutzii (positive control) determined by genotyping; sera and isolates were obtained by R. Hahn). Heterologous sera from patients with histoplasmosis (n = 15), sporotrichosis (n = 15), and aspergillosis (n = 15) were also tested. Finally, 15 serum samples from healthy individuals were also studied (negative control).
For ELISA, a batch of sera from Central-Western Brazil was selected. Twenty-eight serum samples from patients with PCM due to P. lutzii (previously confirmed by ID using P. lutzii antigen) and 28 serum samples from patients with PCM due to P. brasiliensis from São Paulo, southeast region (previously confirmed by ID using AgPbB339) were used in this assay. Among the sera from patients with PCM caused by P. lutzii, three lacked treatment, 16 had 6 to 12 months of therapy, and 9 had 13 to 18 months of treatment. As a gold standard or positive control, three sera from PCM patients with positive cultures for P. lutzii were used. Heterologous sera and negative controls were also studied.
For the Western blot assays, 12 serum samples from patients with PCM due to P. brasiliensis and 12 serum samples from patients with PCM due to P. lutzii (previously positive by ID and ELISA) were used. Heterologous sera and negative control sera were also studied.
Three millimeters of melted 1% agarose (Sigma A-6877) in PBS was poured onto a glass slide (75×25 mm). The pattern for this micro-ID test consisted of a central well surrounded by six wells, each 3 mm in diameter. The central well located 6 mm (edge-to-edge) from the other wells was filled with the antigen solution. Each slide contained two sets of wells. On each slide the two central wells were filled with 10 µl of antigen. Slides were incubated overnight in a moist chamber at room temperature (20–25°C), and then washed for 1 h in 5% sodium citrate and for 24–48 h in saline. The slides were dried, stained for 5 min with 0.15% Coomassie Brilliant blue (Sigma) in ethanol∶ acetic acid∶ water (4∶2∶4; v∶v), and destained in the solvent mixture alone, when necessary. Precipitation bands were recorded by visual observation [30].
Suspected PCM sera were tested first by ID using the traditional P. brasiliensis exoantigen (AgPbB339) and the purified gp43 antigen from P. brasiliensis B339 in order to discriminate sera from patients with PCM caused by P. brasiliensis. The non-reactive PCM sera, which became the focus of this study, were tested with the three different antigenic preparations. Scheme S1 outlines the final strategy to distinguish sera from patients with PCM caused by P. lutzii and P. brasiliensis.
During pilot studies, CFA preparations from different P. lutzii strains were tested and examined by checkerboard titration for antibody detection and P. lutzii strain EPM 208 was chosen for this study. The CFA was used in ELISA at 12.5 µg/ml to detect circulating anti-P. lutzii antibodies. We selected 28 serum samples from patients with PCM due to P. lutzii previously confirmed by ID using CFA from P. lutzii EPM 208, 28 serum samples from patients with PCM due to P. brasiliensis previously confirmed by ID using AgPbB339 and gp43 as antigens, heterologous and negative control sera. Serum samples from patients with PCM due to P. lutzii were from Cuiaba-MT and Campo Grande-MS in the Central-West region of Brazil. As a gold standard, we used three sera from PCM patients with positive cultures for P. lutzii. All sera from patients with PCM due to P. brasiliensis and heterologous sera came from Hospital São Paulo, Escola Paulista de Medicina (São Paulo, SP, Southeast region). All sera were divided into aliquots and stored at −20°C.
Polystyrene flat-bottomed plates (Costar; 96-wells) were coated with CFA (P. lutzii EPM 208) at 12.5 µg/ml diluted in 0.1 M carbonate buffer (pH 9.6; 100 µl/well) and incubated for 2 h at 37°C and overnight at 4°C. The remaining binding sites were blocked with PBS containing 0.1% Tween 20 (PBS-T) and 5% non-fat dry milk (PBS-T-M) (200 µl/well) for 4 h at 37°C. After washing three times with PBS-T, diluted serum (100 µl/well) (1∶50 to 1∶204,800 in PBS-T) was added for 1 h at 37°C. After washing three times with PBS-T, 100 µl of goat anti-human IgG-peroxidase (Sigma) (1∶1000 in PBS-T) was added in each well and incubated for 1 h at 37°C. After three washes with PBS-T, 100 µl of substrate solution (5 mg of o-phenylenediamine in 25 ml of 0.1 M citrate-phosphate buffer pH 5.0 plus 10 µl of 30% H202) was added to each well, and the reaction was interrupted after 8 min in the dark by the addition of 50 µl of 4 N H2SO4. The optical density was read at 492 nm with a Tecan Sunrise 96 well Microplate Reader (Tecan, Grödlg, Austria).
The same sera were tested by ELISA using the classical exoantigen from P. brasiliensis B339 (AgPbB339) at 10 µg/ml as determined by checkerboard titration.
For the immunoblot study we used CFA from P. lutzii strain EPM 208 (10 µg/lane) and CFA from P. brasiliensis B339 (10 µg/lane). Sera that reacted with these two different antigens were compared. The gels and reagents for SDS-PAGE were prepared as described previously [31]; a 10% resolving gel and 3% acrylamide stacking gel were used. After electrophoresis, the gels were stained with Coomassie Blue solution (0.1% Coomassie Blue, 45% methanol, and 10% glacial acetic acid), and excess stain was removed with destaining solution (10% glacial acetic acid and 10% methanol). We selected 12 sera from patients with PCM due to P. lutzii previously confirmed by ID using CFA from P. lutzii strain EPM 208 and 12 sera from patients with PCM due to P. brasiliensis previously confirmed by ID using AgPbB339 and gp43 as antigens. For Western blot, the samples were transferred to nitrocellulose membranes (Milipore) according to Towbin et al. [26]. Gel contents were electrotransferred to nitrocellulose membranes at 400 mA for 1 h in a Trans-Blot cell (Bio-Rad) containing transfer buffer (25 mM Tris-HCl, 192 mM glycine, and 20% methanol [vol/vol]; pH 8.3). Free binding sites in the membranes were blocked by incubation for 2 h in 5% (wt/vol) non-fat dry milk in PBS-T (pH 7.5). Membranes were sliced vertically and the strips were incubated for 1 h at room temperature with diluted serum (1∶500 in PBS-T containing 5% non-fat milk; PBS-T-M). The strips were washed in PBS-T-M four times for 10 min each. The membranes were incubated with peroxidase-conjugated goat anti-human IgG (Sigma) at 1∶1,000 dilution for 1 h and then washed as above. The reactive antigenic molecules were developed by chemiluminescence mix reagents (Millipore, WBKLS0500) onto the membrane which is reactive to the conjugated secondary antibody peroxidase and luminol. To image the blot, we placed the transferred membrane in the transluminator UVITEC imager (Uvitec Cambridge, United Kingdom). The software settings in the Allience 4.7 software were set up to take several images from different time exposures, starting at 2 seconds with a total of ten images between 2 seconds. Then we selected the best time exposure from those 10 pictures taken.
Performance measured analyzed for each test were: sensitivity, specificity, positive predictive value (PPV) and negative predictive value (PNV). The receiver operating characteristics (ROC) curve was drawn to determine the sensitivity and specificity for each antigen preparation (CFA, EXO and TCA) in ID test. The areas under ROC curves (AUC) were calculated to evaluate the diagnostic values of each antigen preparation. We assumed a test without diagnostic power when the ROC curve was linear with AUC of 0.5 (the ROC curve will coincide with the diagonal). A powerful test would give an AUC around 1.0, demonstrating the absence of both false positives and false negatives (the ROC curve will reach the upper left corner of the plot). To measure the degree of concordance of the results of the different assays, the kappa statistic and its 95% confidence interval (95% CI) were calculated. Kappa values were interpreted as follows: 0.00–0.20, poor agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.80, good agreement; 0.81–1.00, very good agreement [32]. A p value of <0.05 was considered to indicate statistical significance. All statistical calculations were performed with the MedCalc Statistical Software version 13.2.0 (MedCalc Software bvba, Ostend, Belgium; http://www.medcalc.org; 2014).
Suspected Paracoccidioides spp. isolates were subjected to HSP70 gene amplification. All clinical isolates had positive amplification and were identified as P. lutzii (Figure S1). During pilot studies, various P. lutzii CFA preparations were tested by ID; three were not reactive (Fig. 3A). P. lutzii strain EPM 208 was chosen for the CFA preparation based on the good and rapid growth of this strain, as well as the good bands observed on the slide. Fig. 3B shows that antigens (CFA) from P. brasiliensis Pb18 and B339 do not react with P. lutzii serum.
In ID tests, 59 (66.2%) of 89 patient sera were reactive to the classical antigen AgPbB339, indicating that these sera were from patients with PCM due to P. brasiliensis. The 30 sera that were not reactive were re-tested with P. brasiliensis-gp43 purified antigen, and 13 (43.3%) were reactive, indicating that they were also from patients with PCM due to P. brasiliensis. Some of these serum samples were positive only for gp43 when tested with different protein concentrations of gp43 (Figure S2). The 17 non-reactive sera were re-tested with three antigenic preparations from P. lutzii strain EPM 208. Ten (58.8%) of the sera were reactive to the exoantigen preparation, but a weak precipitated band was observed (Fig. 3C). Only 3 (17.6%) of the sera were reactive to TCA-precipitated antigen, resulting in bands of low intensity (Fig. 3C). Figure 3D shows that antigens from P. brasiliensis B339 (Ag Exo, TCA or CFA) do not reacted with serum from PCM patient due to P. lutzii. In contrast, all 17 (100%) sera were reactive to the CFA, resulting in sharp bands (1, 2 or 3 bands) (Fig. 3E). Figure 3F shows that antigen from Exo B339 or CFA B339 do not reacted with serum from PCM patient due to P. lutzii and that antigen from Exo-PI or CFA-PI do not reacted with serum from PCM patient due to P. brasiliensis. Among those 17 reactive sera with CFA P. lutzii, 2 (11.7%) of them also reacted weakly with P. brasiliensis CFA antigen, indicating cross-reactivity between these two species. The most sensitive antigenic preparation was CFA-Pl, with a sensitivity 100% (IC 80.5–100) and specificity 100% (IC 94–100), followed by Exo-Pl with a sensitivity 58.82% (32.9–81.6) and specificity 100% (IC 94–100) and TCA-Pl with a sensitivity 17.65% (3.8–43.4) and specificity 100% (IC 94–100). Positive and negative predictive values (PPV, NPV) in ID assay were higher for CFA-Pl (PPV = 100, IC 79.4–100; NPV = 100, IC 94–100), followed by Exo-Pl (PPV = 100, IC 66.3–100, NPV = 89.55, IC 79.6–95.6) and TCA-Pl (PPV = 100, IC 15.8–100; NPV = 22.07, IC 70.3–89.2). Heterologous sera and serum samples from healthy individuals (negative control) did not react with the tested antigens. We used the area under the ROC curve (AUC) to evaluate the discriminatory values of the antigens (comparing subjects with PCM due to P. lutzii and those without the disease). Our results showed that the CFA-Pl antigen afforded better AUC values (AUC = 1.0, IC 0.95–1.0, p<0.0001) than those for the Exo-Pl (AUC = 0.74±0.06, IC 0.68–0.87, p<0.0001) and TCA-Pl (AUC = 0.58±0.04, IC 0.47–0.69, p = 0.0641) (Fig. 4). Judging from the deviating performance of different antigen preparations in ID assays we discarded the Exo-Pl and TCA-Pl antigens in subsequent experiments.
For ELISA, the CFA from P. lutzii strain EPM 208 showed excellent reactivity for antibody detection using PCM sera (P. lutzii and P. brasiliensis) and heterologous sera (histoplasmosis, aspergillosis and sporotrichosis) and normal human sera as negative control. Figure 5A shows the individual serum titers of patients with PCM due to P. lutzii (sera # 1 to 28) and P. brasiliensis (sera # 29 to 58). Among P. lutzii sera, # 1, 2, and 6 represent the gold standard sera and sera from patients lacking treatment; 3, 4, 5, and 7 to 19 represent sera from patients who received 6 to 12 months of therapy; and 20 to 28 represent sera from patients with more than 12 months of therapy. Figure 5B shows these same groups of sera reacting with exoantigen from P. brasiliensis B339 (traditional antigen: AgPbB339). Figure 5C shows the median curve of 28 PCM sera (P. lutzii) compared to the median curve of 28 PCM sera (P. brasiliensis) and heterologous sera (histoplasmosis, aspergillosis and sporotrichosis), besides the NHS used as negative control. Figure 5D shows the results of the same set of sera but expressed by the titers of each serum. The heterologous sera and NHS always showed much lower titers in relation to the PCM sera. On an overall, we observed that the cross-reactivity with heterologous sera and NHS does not interfere with the interpretation of results with homologous sera and P. lutzii CFA antigen. Most of the sera from patients with PCM due to P. lutzii presented with titers from 1∶12,800 to 1∶204,800 (median = 51,200), whereas sera from patients with PCM due to P. brasiliensis presented with maximum titers of 1∶6.400 (median = 1∶ 1,600), indicating a great difference between the sera. Among the heterologous sera the maximum of cross-reactivity was 1∶1,600 for aspergillosis (median = 1∶200), 1∶800 for histoplasmosis (median = 1∶200), 1∶400 for sporotrichosis (median = 1∶200). NHS reacted until 1∶400 (median = 1∶200).
Western blot allowed us to distinguish PCM due to P. lutzii and due to P. brasiliensis, as only sera from patients with PCM due to P. lutzii are able to recognize antigenic molecules from P. lutzii-CFA (Fig. 6A). Sera from patients with PCM due to P. brasiliensis did not recognize any P. lutzii CFA antigens (Fig. 6B). In addition, we used Kappa test to assess the agreement between different serological assays (ID, ELISA and Western blot) based on CFA-Pl antigen. Kappa test (k = 1) revealed a perfect agreement between all pairwise comparisons for different assays.
In order to verify cross-reactions between both Paracoccidioides species, sera from patients with PCM due to P. brasiliensis and due to P. lutzii were tested against CFA from B339 (Fig. 6C and 6D).The results showed that sera from PCM due to P. brasiliensis recognized only gp43 molecule; sera from PCM patients due to P. lutzii recognized various molecules. Considering the main bands evidenced, the reactivity were almost similar to that obtained with P. lutzii sera and CFA-Pl.
The World Health Organization (WHO) estimates that more than 1 billion people, one-sixth of the world's population, suffer from one or more neglected diseases. The diseases are most heavily concentrated in low-income nations in Africa and Latin America, and the most common types of neglected diseases are tropical diseases. Many neglected tropical diseases are caused by parasites, which are spread by insects or contact with contaminated water or soil (http://www.who.int/neglected_diseases/diseases/en/). In this scenario, PCM can be considered a neglected disease. Coutinho et al. [2] analyzed 3,181 deaths from PCM in Brazil based on 16 years of sequential data (from 1980 to 1995). During this period, PCM showed considerable magnitude and low visibility, representing the eighth most common cause of death from predominantly chronic or recurrent types of infectious and parasitic diseases. The study showed that the mortality rate justifies classifying this disease as an important health problem in Brazil. However, no government programs exist for this mycosis, with rare punctual exceptions.
For a century PCM was thought to be caused solely by P. brasiliensis, but recent studies [11], [12], [18]–[21], [33] have shown that other species of Paracoccidioides can also cause PCM. These findings reveal that problems in the serological diagnosis of PCM were due to an incorrect use of antigens from the genus, as only antigens of P. brasiliensis (PS3) have been used for this purpose. With the discovery of P. lutzii, it became clear that new antigens from these new species would be needed to obtain more accurate diagnoses. Recently, a fatal case of PCM due to P. lutzii fungemia was reported [34] and other two cases were reported in the North region of Brazil [23].
Until recently, the serological diagnosis of PCM by ID could be accomplished using only the standard exoantigen from B339, which has a high concentration of gp43 antigen, the immunodominant and specific molecule in ID tests. Early suspicions that this serology was not entirely certain were raised by our group [35] when we observed that an exoantigen produced from an isolate (Pb550B) from the Central-West region of Brazil was capable of reacting with a larger number of PCM sera from that region than the standard antigen from an isolate from São Paulo in the Southeast region (AgPbB339). Nevertheless, when tested with PCM sera from São Paulo, the antigen from the Central-West region of Brazil did not produce satisfactory results. Also, when we used the standard antigen AgPbB339, it reacted very well with sera from São Paulo, but very poorly with sera from the Central-West region. Analyses of antigens obtained from the B339 and 550B isolates showed that the former produced high levels of gp43, which was undetected in the 550B filtrate. This variation in gp43 expression likely influences the low reactivity observed in ID tests using sera from patients from the Central-West region of Brazil. At that time, nothing was known about P. lutzii; therefore, we suggested using regional strains to aid in the diagnosis of PCM for the Central-West region. Importantly, B339 belongs to the S3 species, whereas isolate 550B is from the Central-West region of Brazil, where P. lutzii is more frequently found.
The main fact that spurred this study was the observation that sera from patients with PCM (previously proven by direct examination of secretions or histopathological analysis) were negative by conventional serology using the standard AgPbB339 antigen in ID tests. In order to elucidate the problem with serological diagnosis, we developed a strategy. First, all sera from patients suspected of PCM should be tested against the traditional exoantigen of P. brasiliensis (AgPbB339) and purified gp43 molecule derived from the same standard strain. The reactive sera were then considered to be PCM due to P. brasiliensis. The non-reactive sera were then tested against CFA from P. lutzii EPM 208.
Applying this strategy, we were able to diagnosis 72 patients with PCM due to P. brasiliensis. The remaining non-reactive patients (n = 17) were diagnosed mostly by ID using CFA from P. lutzii and also by ELISA and Western blot. This type of antigen (CFA) is very easy to prepare and is very useful in the diagnosis of PCM due to P. lutzii, exhibiting clearly visible precipitation bands. Among these 17 sera, 2 of them also reacted with the traditional exoantigen B339, but with a weak reactivity, indicating cross reaction between these two species of Paracoccidioides. However, the exoantigen preparation (from P. lutzii) also managed to obtain positive reactions, but the precipitation line was very weak, making it difficult to read the slides. In addition, the antigenic preparations obtained by TCA precipitation were useful, but the bands were quite weak. In both later antigens, the sensitivity of the reactions was lower than that obtained with the CFA preparation.
Thus, the CFA preparation obtained from P. lutzii EPM 208 was the better antigen for use in ID tests to diagnosis PCM due to P. lutzii from the Central-West region of Brazil. Other P. lutzii strains were tested with similar results; however, among 14 CFA preparations, three were unable to precipitate antibodies in sera from P. lutzii-PCM patients, indicating that antigenic differences exist among these strains. Therefore, more studies are necessary to elucidate the antigenic variability among P. lutzii strains. Another curiosity we observed when we tested CFA preparations from P. lutzii in the ID tests was that better results were obtained when the protein concentrations of the antigen were greater than 1200 µg/ml. Protein concentrations less than 1200 µg/ml resulted in very weak bands, hindering the visualization of precipitation reactions. Machado et al. [36] described CFA as being inefficient for the diagnosis of PCM due to P. lutzii.
In ELISA, sera from patients with PCM due to P. lutzii presented with higher antibody titers than those obtained for the sera from patients with PCM due to P. brasiliensis which can perfectly differentiate between patient sera from P. brasiliensis and P. lutzii. Heterologous sera, such as histoplasmosis, aspergillosis and sporotrichosis and normal human sera exhibited very low cross-reactivity and did not interfere with interpretations of the main system. It was expected that the sera from PCM patients due to P. brasiliensis reacted more intensively with P. lutzii antigens because these two species belong to the same genus. Certainly, the antigenic components present in P. lutzii are constituted by specific antigens of this species plus common antigens shared with P. brasiliensis. Probably, antibodies elicited during infection by P. brasiliensis are only generated against antigens from this species whereas antibodies elicited during infection by P. lutzii are generated by specific antibodies against antigens in this species and also against antigens shared with P. brasiliensis. It was also expected that sera from patients with PCM due to P. brasiliensis reacted more intensively than heterologous sera (histoplasmosis, aspergillosis and sporotrichosis), and not almost in the same level. Perhaps, these results may be explained by the presence of common antigens among P. brasiliensis and heterologous fungi, which are reacting in this ELISA system.
In relation to the strategy used for immunoblotting in the present study, we found that sera from patients with PCM due to P. brasiliensis do not recognize any antigen from P. lutzii CFA. In contrast, sera from patients with PCM due to P. lutzii were able to recognize antigens from both antigenic preparations (CFA from P. lutzii or P. brasiliensis). Therefore, P. lutzii is antigenically more complex. These findings suggest that P. lutzii has its own species-specific antigens and antigens common with P. brasiliensis. Thus, we can distinguish between sera from patients with PCM due to P. brasiliensis and those with PCM due to P. lutzii. However, the specific P. lutzii antigen still needs to be identified. The reactivity of PCM sera due to P. brasiliensis predominantly to gp43 of its homologous antigen may be explained by the fact of the great immunogenicity of this molecule, so that, the immune response to other minor molecules is not evidenced.
Based on our results, we estimate that the incidence of PCM due to P. lutzii is much greater than we imagine, even though the Central-West region of Brazil where the disease is prevalent has not invested in acquiring new tools for a specific diagnosis of this entity. In general, laboratories for the diagnosis of infectious diseases are small, have few resources, and perform a limited number of tests per day. Our strategy meets all of the requirements for use in laboratories for the diagnosis of PCM due to P. lutzii, as well as seroepidemiological studies. In addition, the test does not require special skills and can be used for a small number of samples. Although the cost of ELISA and Western blot is higher than that of ID, they are economical if the costs associated with laboratory personnel, quality control, and reagent storage are taken into account.
Surely P. lutzii is not confined to the Central-West region of the country; we have described two cases in the northern region (Pará State) [23], and we have strains of P. lutzii isolated in the Rondônia State (North region) and Paraná (South region) in our fungal collection. With our proposed strategy for the diagnosis of PCM caused by P. lutzii, many new cases may arise throughout Brazil and other South American countries. Studies are underway in our laboratories to identify the species-specific antigen for P. lutzii in order to use it for simple and specific diagnosis of this mycosis.
|
10.1371/journal.pbio.1001624 | A Bacteriophage Tailspike Domain Promotes Self-Cleavage of a Human Membrane-Bound Transcription Factor, the Myelin Regulatory Factor MYRF | Myelination of the central nervous system (CNS) is critical to vertebrate nervous systems for efficient neural signaling. CNS myelination occurs as oligodendrocytes terminally differentiate, a process regulated in part by the myelin regulatory factor, MYRF. Using bioinformatics and extensive biochemical and functional assays, we find that MYRF is generated as an integral membrane protein that must be processed to release its transcription factor domain from the membrane. In contrast to most membrane-bound transcription factors, MYRF proteolysis seems constitutive and independent of cell- and tissue-type, as we demonstrate by reconstitution in E. coli and yeast. The apparent absence of physiological cues raises the question as to how and why MYRF is processed. By using computational methods capable of recognizing extremely divergent sequence homology, we identified a MYRF protein domain distantly related to bacteriophage tailspike proteins. Although occurring in otherwise unrelated proteins, the phage domains are known to chaperone the tailspike proteins' trimerization and auto-cleavage, raising the hypothesis that the MYRF domain might contribute to a novel activation method for a membrane-bound transcription factor. We find that the MYRF domain indeed serves as an intramolecular chaperone that facilitates MYRF trimerization and proteolysis. Functional assays confirm that the chaperone domain-mediated auto-proteolysis is essential both for MYRF's transcriptional activity and its ability to promote oligodendrocyte maturation. This work thus reveals a previously unknown key step in CNS myelination. These data also reconcile conflicting observations of this protein family, different members of which have been identified as transmembrane or nuclear proteins. Finally, our data illustrate a remarkable evolutionary repurposing between bacteriophages and eukaryotes, with a chaperone domain capable of catalyzing trimerization-dependent auto-proteolysis in two entirely distinct protein and cellular contexts, in one case participating in bacteriophage tailspike maturation and in the other activating a key transcription factor for CNS myelination.
| Membrane-bound transcription factors are synthesized as integral membrane proteins, but are proteolytically cleaved in response to relevant cues, untethering their transcription factor domains from the membrane to control gene expression in the nucleus. Here, we find that the myelin regulatory factor MYRF, a major transcriptional regulator of oligodendrocyte differentiation and central nervous system myelination, is also a membrane-bound transcription factor. In marked contrast to most well-known membrane-bound transcription factors, cleavage of MYRF appears to be unconditional. Surprisingly, this processing is performed by a protein domain shared with bacteriophages in otherwise unrelated proteins, where the domain is critical to the folding and proteolytic maturation of virus tailspikes. In addition to revealing a previously unknown key step in central nervous system myelination, this work also illustrates a remarkable example of evolutionary repurposing between bacteriophages and eukaryotes, with the same protein domain capable of catalyzing trimerization-dependent auto-proteolysis in two completely distinct protein and cellular contexts.
| Membrane-bound transcription factors (MBTFs) are a remarkable class of transcription factors that are initially generated as integral membrane proteins. Upon relevant cues, they undergo proteolytic processing, releasing the transcription factor domain from the membrane and allowing it to translocate to the nucleus to control gene expression. Two different broad mechanisms of MBTF proteolytic activation have been observed to date. One class of MBTFs is proteolytically activated by regulated ubiquitin/proteasome-dependent processing (RUP) and includes transcription factors that control membrane fluidity in budding yeast (SPT23 and MGA2) and a fission yeast hypoxic transcription factor (Sre1) [1]–[2]. The second class is activated via regulated intramembrane proteolysis (RIP) and includes sterol regulatory element-binding proteins (SREBPs) [3]–[4], activating transcription factor 6 (ATF6) [5]–[7], and the developmental regulator Notch [8]–[10]. RIP-dependent activation of MBTFs typically requires additional proteases that act outside of the membrane. For example, when cellular cholesterol levels decrease, SREBPs are transported to the Golgi apparatus, where they are cleaved by Site-1 protease, whose active site is located in the lumen of the Golgi. Cleavage by Site-1 protease allows the subsequent intramembrane proteolysis by Site-2 protease [4]. Similarly, following accumulation of misfolded proteins in the endoplasmic reticulum (ER), ATF6 translocates to the Golgi and is proteolyzed sequentially by Site-1 and Site-2 proteases [5],[7]. Recently, many basic leucine zipper proteins homologous to ATF6 have been discovered and appear to play important roles in tissue-specific unfolded protein responses [11]–[12].
Within the human genome, an early genome-wide computational screen suggested the existence of six MBTFs [13]. Since then, the number of characterized DNA-binding domains has increased significantly [14], and prediction methods for the membrane topology of proteins have been improved dramatically [15]–[16], which led us to revisit the search for human MBTFs. We found that C11orf9, the largely uncharacterized human ortholog of mouse Myrf (a key transcriptional regulator of oligodendrocyte (OL) maturation and CNS myelination [17]), was strongly predicted to encode an MBTF. C11orf9 (hereafter referred to as MYRF [CCDS ID: 31579 and RefSeq ID: NP_037411]) and its orthologs were predicted to have a domain homologous to the DNA-binding domain of the yeast transcription factor Ndt80 [18] as well as a single transmembrane (TM) segment.
However, by using algorithms capable of recognizing extremely distant sequence homology, we also observed that MYRF and its orthologs harbor an intramolecular chaperone domain shared with bacteriophage endosialidases [19]–[20], the tailspike proteins essential for bacteriophages to infect bacteria encapsulated with polysaccharides. While the homology of genes between bacteriophages and eukaryotes is not unprecedented, or even the horizontal transfer of genes between the two, it is nonetheless rare, and in general the mechanism of transferred genes is quite different. For example, the GG domain is found in both bacteriophage tail fibers and FAM3 cytokines [21]. In addition, the large nuclear and cytoplasmic viruses, such as Mimivirus, appear to have chimeric origins that include bacteriophages [22].
The tailspike proteins are known to trimerize and to self-process. This raised the hypothesis that this domain in eukaryotes might contribute to a novel method for the formation and function of an MBTF. Indeed, the intramolecular chaperone domain of MYRF facilitates its homo-oligomerization and proteolytically processes it into two halves. The N-terminal trimer, containing the DNA-binding domain, is released from the ER membrane and moves to the nucleus, where it exerts transcriptional effects. Proper processing and translocation of the MYRF N-terminal trimer then contributes to the maturation of OLs. The C-terminal homo-oligomer, containing the TM domain, remains in the ER. These findings not only demonstrate an extraordinary link between a possible endosymbiont or commensal bacteriophage and eukaryotic development, but reveal a novel cleavage mechanism for MBTFs.
Myrf (the mouse ortholog of MYRF) was previously reported to encode a nuclear protein, based on immunofluorescence (IF) microscopy with an N-terminally Myc-tagged construct [17]. However, TOPCONS [15], a state-of-the-art membrane topology prediction program, predicts both MYRF and Myrf to be type-II membrane proteins (Figure S1A). Notably, we identified well-conserved nuclear localization signals (NLSs) in the N-terminus (K245KRK248 and K482KGK485) and potential N-linked glycosylation sites in the C-terminus (Figure 1A).
To determine the precise localization of MYRF in the cell, we expressed epitope-tagged MYRF constructs in HeLa cells. Green fluorescent protein (GFP) tagged to the N-terminus of MYRF (GFP-MYRF, Figure 1A) localized to the nucleus, in agreement with the previous study on Myrf (Figure 1B). However, when GFP was tagged to the C-terminus of MYRF (MYRF-GFP, Figure 1A), the GFP signal co-localized with calnexin (CLX), an ER marker (Figure 1B). A doubly-tagged protein, 3F-MYRF-GFP (Figure 1A), resolved this apparent dichotomy: The FLAG tag at the N-terminus exhibited a nuclear signal, whereas the C-terminal GFP signal co-localized with the ER (Figure 1C).
In order to test if the predicted TM domain mediated the ER localization of the C-terminus of MYRF, we deleted the TM domain from the C-terminally GFP-tagged construct (MYRFΔTM-GFP, Figure 1A). MYRFΔTM-GFP localized to the nucleus of HeLa cells (Figure 1B), confirming the role of the predicted TM domain for ER localization. Similarly, a C-terminally GFP-tagged mutant truncated before the predicted TM domain at L756 (MYRF-1:756-GFP, Figure 1A) also localized to the nucleus (Figure 1B). Control experiments were consistent when using alternate epitope tags (FLAG tag; Figure S1B) and cell lines (CG4 cells, a rat OL cell line that may be used as a model for early OL differentiation [23]–[31]; Figure S1C and S1D). Thus, these localization patterns appear to be intrinsic features of MYRF and not artifacts of the particular tags or cells used.
The microscopy suggested that MYRF is processed in cells, which was further confirmed by Western blot of 3F-MYRF (Figure 2B). The majority of the protein was cleaved into a ∼90 kDa N-terminal fragment from the full length of ∼160 kDa. The latter was further verified by comparing 5M-MYRF-3F protein expressed in cells to that expressed from an in vitro translation system (the in vitro reaction mixture immunoprecipitated with FLAG antibodies and blotted with anti-Myc antibodies) (Figure 2C).
The top band representing full-length MYRF was observed to consist of two closely spaced bands (Figure 2D, arrows), with the upper and lower bands potentially representing glycosylated and unglycosylated full-length MYRF, respectively. Upon MG132 treatment, the lower band became as dominant as the upper one. This suggested either that MG132 treatment alters the degradation of MYRF or that MG132—an inducer of ER stress that decreases glycosylation efficiency [32]—inhibits the glycosylation of full-length MYRF, leading to the accumulation of unglycosylated full-length MYRF. Consistent with the latter possibility, tunicamycin treatment reversed the ratio between the upper and lower bands, with the lower one now dominating (Figure 2D). We note that the 120 kDa isoform (Figure 2B) is most likely a degradation intermediate, as it was inconsistently observed and disappeared upon treatment with MG132 (Figure S2A).
Fractionation of HeLa cells transfected with 3F-MYRF revealed that full-length MYRF could be extracted from membranes by treatment with the detergent SDS, but not with high salt or alkaline pH (Figure 2E), similar to the control protein calnexin, a known integral membrane protein. Thus, the fluorescence microscopy, TM domain mutagenesis, glycosylation analysis, and biochemical fractionation data all demonstrated that full-length MYRF is an integral membrane protein.
Finally, we determined the membrane topology of full-length MYRF by treating cells with digitonin, which selectively permeabilizes the plasma membrane but not organelle membranes (Figure S2B) [33]. When the plasma membrane of HeLa cells expressing GFP-MYRF-3F was selectively permeabilized by digitonin, FLAG IF signals could not be detected, in contrast to a strong signal when membranes were indiscriminately permeabilized by Triton X-100 (Figure 2F), suggesting that the C-terminus of MYRF is oriented to the ER lumen. Additional tests with a point mutant (L690A, detailed below) that blocks the generation of the 90 kDa isoform from full-length MYRF enabled us to probe the subcellular location of the N-terminus of full-length MYRF. FLAG IF signals were detected for 3F-MYRF-L690A-GFP when cell membranes were selectively permeated with digitonin (Figure 2F), indicating that the N-terminus of full-length MYRF is located on the cytoplasmic side of ER membranes. Thus, MYRF is synthesized as a type-II membrane protein and processed into N-terminal and C-terminal portions, localized in the nucleus and on the ER membrane, respectively.
In the course of analyzing the MYRF sequence, we discovered distant but significant homology (16% sequence identity and E-value = 3.1×10−18, as measured by HHpred [34]) between the portion of MYRF that lies between its DNA-binding and TM domains and the intramolecular chaperone domain found in bacteriophage endosialidases, proteins that constitute the tailspikes of many bacteriophages (Figure S3A) [19],[35]. The intramolecular chaperone domain, which we have dubbed an ICA (Intramolecular Chaperone Auto-processing) domain, plays two roles in the maturation of bacteriophage endosialidases. The ICA domain facilitates the protein's folding and trimerization [19],[35]. It then functions as a “folding sensor” and auto-cleaves itself away from the bacteriophage endosialidase [20].
A multiple sequence alignment of MYRF and its orthologs indicated that the ICA domain is a strictly conserved feature (Figure S3D). Further, a multiple sequence alignment of only the ICA domains from eukaryotes, a bacterium, and a phage revealed the absolute conservation of S578 and K583 (following the MYRF numbering, Figure 3A). In bacteriophage endosialidases, the serine and lysine residues equivalent to MYRF S578 and K583 form a catalytic dyad for the auto-cleavage reaction [20]. The correct positioning of these catalytic residues, along with an arginine residue that stabilizes the oxyanion during the peptide bond breakage, is thought to be achieved only upon folding and trimerization of bacteriophage endosialidases [20], enabling the ICA domain to function as a folding sensor. We thus asked if the ICA domain might nonetheless still serve—in a radically altered context as compared to viral tailspikes—as a folding sensor and protease to activate MYRF.
Based on the conservation of the ICA domain, including its catalytic residues, we hypothesized that, once generated as a type-II membrane protein, the ICA domain could potentially facilitate the folding and trimerization of full-length MYRF and then proteolytically process it into two independent trimers (Figure 3B). The N-terminal trimer, containing the DNA-binding domain, might then be released from membranes and enter the nucleus to regulate transcription, while the C-terminal trimer, comprising residues S578-D1111, would remain in the ER membrane.
To test this hypothesis, we mutated the ICA domain of MYRF and assayed the effects on the proteolytic processing of MYRF. Deletions involving the ICA domain (Δ662–752 and Δ538–617) blocked normal processing of MYRF (Figure 3C). Likewise, mutation of the putative catalytic residues S578 and K583 to alanine (3F-MYRF-S578A and 3F-MYRF-K583A) also blocked proteolytic processing (Figure 3C). The FLAG tag at the N-terminus of these mutant constructs remained in the ER membrane of HeLa cells (Figure 3D), demonstrating that the DNA-binding domain of MYRF is retained in the membrane when auto-processing is blocked.
We next asked whether additional residues shown to be important for the function of phage ICA domains are also important for the function of the MYRF ICA domain. As the N912 and G956 residues in the ICA domain of bacteriophage K1F endosialidase are essential for the function of the ICA domain [19], mutation of their corresponding residues in MYRF to alanine (3F-MYRF-D579A and 3F-MYRF-G626A) markedly reduced the proteolytic processing of MYRF (Figure 3C). As an additional control, we expressed a truncated form of MYRF that terminates at P577 (3F-MYRF-1:577), corresponding to the expected N-terminal fragment generated from auto-processing (Figure 3B), and confirmed that it has the same electrophoretic mobility as the processed N-terminal fragment of 3F-MYRF (Figure 3C). Taken together, these results support the hypothesis that the ICA domain mediates the proteolytic processing of MYRF, in a manner similar to bacteriophage endosialidases.
To further investigate the role of the ICA domain in the processing of MYRF, we mapped the amino acid sequence of the MYRF ICA domain (residues N567–R692) onto the crystal structure of the ICA domain of bacteriophage K1F endosialidase (PDB accession code: 3GW6 [20]) (Figure S3A). The homology-derived structure predicted that L683, I687, and L690 of MYRF form a leucine zipper (Figure 3E). Because the leucine zipper appears integral to the trimeric structure of the MYRF ICA domain, we reasoned that its disruption would destabilize the trimer and consequently interfere with proteolytic processing. Site-directed mutagenesis confirmed that the leucine zipper is indeed required for MYRF processing (Figure 3C). IF microscopy of 3F-MYRF-L683A and 3F-MYRF-L690A in HeLa cells confirmed that their localizations matched the catalytic residue mutants (Figure S3B). In contrast, the structure suggested that K596, S599, and L679 would not be essential to either catalytic or structural roles, and all were predicted to face the exterior of the protein (Figure S3C). Consistent with this prediction, mutating each of these residues to alanine (3F-MYRF-K596A, 3F-MYRF-S599A, and 3F-MYRF-L679A) did not affect MYRF processing (Figure 3C). These results confirm that the ICA domain is indeed responsible for the proteolytic processing of MYRF, and that the mechanism of proteolysis is conserved between animals and bacteriophages, in spite of a complete alteration of neighboring protein domains and overall protein function.
The ICA domain is known to function autonomously to proteolyze bacteriophage endosialidases. We therefore asked whether the processing of MYRF was similarly autonomous, testing two specific hypotheses. First, we examined whether the proteolytic processing of MYRF was independent of membrane integration. As shown in Figure 3F, a construct (3F-MYRF-1:756) that was truncated before the TM domain at L756 was normally processed in HeLa cells, but processing was blocked when the catalytic residue S578 was changed to alanine (3F-MYRF-1:756-S578A). Second, we asked whether MYRF is normally processed in heterologous systems, which would support a fully autonomous event. To address this hypothesis, we expressed MYRF in E. coli and yeast cells. Due to the difficulty of expressing full-length MYRF in E. coli, we worked with a truncation construct (MYRF-319:708) that only comprises the DNA-binding and ICA domains of MYRF. This construct was normally processed in HeLa cells (Figure 3F), and its processing was blocked when important residues were mutated to alanine (3F-MYRF-319:708-S578A, 3F-MYRF-319:708-K583A, and 3F-MYRF-319:708-L683A). Figure 3F shows that MYRF-319:708 behaved in the same manner in E. coli, and similarly, full-length MYRF was normally processed in budding yeast (Figure 3F). Taken together, these results indicate that the ICA domain autonomously functions in the proteolytic processing of MYRF.
The ICA domain is known to induce the trimerization of bacteriophage endosialidases as part of its intramolecular chaperone activity [36]. Given the central role of the ICA domain in MYRF auto-processing, we next asked whether it was also promoting trimerization in this context. We first used co-immunoprecipitation experiments of differentially tagged constructs in order to assay homo-oligomerization of the N-terminal fragment generated by the auto-processing of MYRF. As shown in Figure 4B, the N-terminal fragment of N-terminally 5xMyc-tagged MYRF (5M-MYRF; Figure 4A) did not bind beads coated with FLAG antibodies. However, when co-transfected with 3F-MYRF (Figure 4A), the N-terminal fragment of 5M-MYRF robustly bound the FLAG beads, confirming homo-oligomerization of the N-terminal fragment of MYRF. To measure the nature of the homo-oligomer, we employed size exclusion chromatography, which indicated that the N-terminal fragment from the auto-processing of MYRF-319:708 (a construct which only contains the DNA-binding and ICA domains of MYRF) forms a trimer (Figure S4). MrfA, a Dictyostelium ortholog of MYRF, has also been suggested to bind DNA as a trimer in vivo [37].
Given that the MYRF N-terminal fragment forms a trimer, we assayed the state of the C-terminal fragment. We could not directly assay its trimeric state due to expression and purification issues in E. coli; instead, we asked if the C-terminal fragment of MYRF also homo-oligomerized using co-immunoprecipitation. As predicted, the C-terminal fragment of MYRF-6M (Figure 4A) did not bind FLAG beads (Figure 4B). However, co-transfection with MYRF-3F (Figure 4A) induced binding (Figure 4B), confirming that the C-terminal fragment generated from auto-processing is also a homo-oligomer. Because the bacteriophage ICA domain is known to exist autonomously as a trimer [19]–[20] and the ICA domain is part of the C-terminal fragment of MYRF, and we have confirmed the trimeric state of the N-terminal fragment, we expect the MYRF C-terminal fragment to also exist as a trimer.
Because the ICA domain is known to induce trimerization, we expected that a truncated construct encoding only the N-terminal fragment (MYRF-1:577; Figure 4A) should fail to trimerize. When expressed alone in HeLa cells, 5M-MYRF-1:577 did not bind FLAG beads (Figure 4B). When co-transfected with 3F-MYRF-1:577, it still did not bind (Figure 4B), even though co-transfected 3F-MYRF-1:577 robustly bound the FLAG beads (unpublished data). Thus, an intact ICA domain is essential for the formation of the N-terminal trimer.
As bacteriophage ICA domains require the completion of folding and trimerization as a prerequisite to the auto-cleavage reaction [20],[36], we suspected that full-length MYRF should also homo-oligomerize. A test of full-length MYRF, obtained by using a catalytic residue mutant (K583A), confirmed its homo-oligomerization (Figure 4B). Likewise, full-length MYRF obtained by a leucine zipper mutant (L683A) still homo-oligomerized (Figure 4B), in spite of defective auto-processing (Figure 3C). Thus, auto-processing of MYRF apparently requires both trimerization and proper formation of the leucine zipper that includes L683. Notably, the N-terminal fragment generated from the auto-processing of MYRF-1:756 (Figure 4A) also formed a homo-oligomer (Figure 4B), consistent with functional autonomy of the ICA domain.
Upon auto-processing, the N-terminal trimer translocates to the nucleus. To test the roles of the predicted NLSs (K245KRK248 and K482KGK485) in nuclear translocation, we examined the effects of deleting the NLSs on subcellular localization. When either single NLS was deleted (3F-MYRFΔNLS1 or 3F-MYRFΔNLS2), nuclear translocation of the N-terminal trimer was only partially blocked (Figure 4C). Deletion of both NLSs blocked MYRF nuclear translocation to a greater extent (Figure 4C), indicating that both NLSs contribute to the nuclear translocation of the N-terminal trimer.
Once generated as a type-II membrane protein, MYRF is auto-processed into two independent fragments (Figure 3B). The N-terminal trimer enters the nucleus where it is likely to function as a transcription factor, while the C-terminal homo-oligomer remains in the ER, where its function is unknown. In order to assay the transcriptional roles of MYRF, we first identified transcriptional targets of MYRF by performing next-generation RNA sequencing of HeLa cells that were transfected with wild-type MYRF and a catalytic residue mutant. Among genes differentially expressed between the two samples (Table S1), we confirmed Endothelin 2 (Edn2) as a transcriptional target of MYRF in HeLa cells, and thus could use its expression levels measured by quantitative real-time polymerase-chain-reaction (qRT-PCR) as a readout of the transcriptional activity of various MYRF constructs.
Using this assay, we confirmed that auto-processing of MYRF is required for the transcriptional activity of MYRF. Figure 5A shows that the expression level of Edn2 was about 20-fold higher when HeLa cells were transfected with MYRF compared to the empty vector control (pcDNA3). The transcriptional activation of Edn2 by MYRF is most likely due to the direct binding of MYRF to DNA because mutation of R445, a strictly conserved residue essential for the direct binding of MrfA to DNA [37], to alanine ablated its transcriptional effects (Figure 5A). Mutation of R445 to alanine did not affect the auto-processing and localization of MYRF (Figure S5). Blocking the auto-processing of MYRF, by mutating either catalytic residues (S578A and K583A) or a structurally important residue (L683A), abrogated the transcriptional effects of MYRF on Edn2 (Figure 5A).
To test whether the N-terminal trimer is sufficient for this transcriptional activation, we assayed a construct truncated before the TM domain at L756 (MYRF-1:756) for transcriptional activity. MYRF-1:756 properly homo-oligomerizes (Figure 4B) and is normally processed (Figure 3F). Figure 5A shows that MYRF-1:756 is as competent as full-length MYRF in activating the transcription of Edn2, demonstrating that the N-terminal trimer is sufficient for the transcriptional activity of MYRF; notably, the C-terminal homo-oligomer does not significantly contribute to this activity. Although homo-trimeric transcription factors are not common, there is a well-known precedent, heat shock factor 1 [38].
Trimerization also appears to be necessary for full transcriptional effects, as we observed a construct directly encoding the N-terminal fragment of MYRF (MYRF-1:577, Figure 4A) to be only partially functional (Figure 5A); this construct fails to form a trimer (Figure 4B). On the other hand, our observation that MYRF-1:577 is still partially functional is in excellent agreement with a recent demonstration that monomeric MrfA can still bind DNA in vitro, although it appears to function as a trimer in vivo [37].
Given the well-characterized role of Myrf (the mouse ortholog) in OL maturation [17], we examined the functional consequences of MYRF auto-processing on the maturation of CG4 cells. Although the CG4 cell line is widely understood not to be a good model of myelination, it may be used as a model for early OL differentiation. We counted the fraction of transfected CG4 cells that had matured to express myelin basic protein (MBP) or O1, two known OL maturation markers (Figure 5B). MYRF significantly promoted the maturation of CG4 cells: 15% of transfected CG4 cells matured to express MBP when transfected with a vector containing both MYRF and GFP (MYRF, Figure 5C), as compared to less than 2% of cells transfected with a control vector expressing only GFP (IRES-GFP). Consistent with the RT-PCR analysis, the mutation of R445 to alanine abrogated the effects of MYRF on CG4 cell maturation (MYRF-R445A, Figure 5C), suggesting that MYRF directly binds DNA to activate transcription for OL maturation. Auto-processing mutants of MYRF (MYRF-S578A and MYRF-K583A) similarly failed to promote CG4 cell maturation (Figure 5C), indicating that correct processing is required. To test if the N-terminal trimer generated by auto-processing is sufficient for OL maturation, we employed a construct truncated before the TM domain at L756 (MYRF-1:756, Figure 4A). Notably, MYRF-1:756 was much less competent compared to wild-type MYRF in promoting maturation (Figure 5C), in spite of being normally processed (Figure 3F), homo-oligomerizing (Figure 4B), and activating Edn2 expression similarly to wild-type MYRF (Figure 5A) in HeLa cells, suggesting a potential role for the C-terminal domain in OL maturation. Overall, our data confirm that auto-processing is essential for MYRF to promote OL maturation. They also suggest that the maturation of OLs might require both the transcription factor function of the N-terminal trimer and the unknown function of the C-terminal homo-oligomer in the ER.
RIP- and RUP-activated MBTFs are widely observed across organisms, spanning both eukaryotes and prokaryotes. MBTFs are recognized as increasingly common regulatory mechanisms in plants, with many plant MBTFs playing important roles in stress responses and development [39]–[43]. NTM1 (NAC with TM motif1), for example, regulates cell division and growth in Arabidopsis [44]. New examples of MBTFs have also been identified for bacteria [45]–[49]. In fact, ToxR of Vibrio cholerae was the first known MBTF [50], although it is still unclear whether ToxR requires a proteolytic activation step to exert transcriptional effects. Thus, it is likely that many MBTFs remain to be found, and an open question is what other activation mechanisms may be employed. MYRF reveals one such previously unknown activation mechanism.
We show that MYRF is a MBTF that is auto-processed by its ICA domain into two independent homo-oligomers (Figure 6B). The N-terminal trimer, containing a largely disordered protein segment and the Ndt80 DNA-binding domain, is released from the membrane and translocates to the nucleus to regulate gene expression. The disordered N-terminal protein segment presumably functions as a transactivation domain because partial deletions in this region render MYRF nonfunctional in terms of its transcriptional activation of Edn2, although auto-processing and localization are not affected (unpublished data). The N-terminal trimer is both necessary and sufficient for the transcriptional activity of MYRF. The C-terminal homo-oligomer remains in the ER, and may perform an important function there. Our functional assays show that auto-processing is essential for MYRF both to activate transcription and to promote OL maturation.
Importantly, this mechanism is probably conserved across all MYRF family members that include MYRFL (a paralog of MYRF), as all members—found across the animal kingdom, including vertebrates, insects, nematodes, amoeba, and tunicates—are characterized by the domain arrangement shown in Figure 3B. Residues shown to be essential for the direct DNA-binding of the Ndt80 DNA-binding domain [18], including MYRF R445 and R469, are strictly conserved across the family, as are key residues of the ICA domain, such as S578, D579, K583, G626, and I687, and the presence of the leucine zipper. The ICA domain is invariantly followed by a relatively well-conserved TM domain and a poorly conserved ER lumenal domain.
As an intramolecular chaperone, the ICA domain is known to facilitate the trimerization and folding of protein sequences that lie N-terminal to it [36], and it has been best characterized for bacteriophage endosialidases [19]–[20]. Without the ICA domain, the endosialidases apparently fail to fold properly, let alone forming a trimer. For the N-terminal fragment of MYRF, however, the role of the ICA domain seems limited to trimerization. Even though MYRF-1:577 failed to form a trimer (Figure 4B), it did activate the transcription of Edn2, albeit to a lesser degree compared to its trimeric counterpart. Also, the presumably monomeric form of MrfA bound DNA [37]. These results suggest that while the ICA domain has a clear role in trimerization and auto-catalysis, its role in the proper folding of the N-terminal fragment of MYRF may be less critical than for the folding of bacteriophage endosialidases. Consistent with this, sequence alignments show that an extended loop of the bacteriophage K1F ICA domain that comprises T977–H1026, shown to be essential for the folding of bacteriophage endosialidases [20], is missing in most other ICA domains, including those of MYRF and its orthologs (Figure S3A).
From a structural perspective, the ICA domain achieves trimerization by chaperoning a triple β-helix fold [36]. Triple β-helix folds are often associated with very stable trimeric structures such as viral tailspike or fiber proteins, and the bacteriophage K1F endosialidase—whose trimerization and folding are mediated by its ICA domain—is known to form a trimeric structure that is resistant to SDS [19]. An interesting question is whether the N-terminal trimer of MYRF also involves a triple β-helix fold. Figure S3D shows that the region compassing residues W536-P577, where a triple β-helix fold is expected, is indeed well conserved. In PQN-47, a mutation of the glycine residue equivalent to G566 of MYRF renders PQN-47 nonfunctional [51]. The evidence therefore suggests that the N-terminal trimer of MYRF could in principle maintain its trimeric state by having a triple β-helix fold at its C-terminus.
The auto-processing of MYRF appears to be constitutive, which stands in stark contrast to the highly regulated processing of such factors as SREBP [4], ATF6 [5], Notch [9], and STP23 [1]. In fact, we observed normal processing of MYRF in all the systems that we tried, including HEK293 cells, fibroblasts, human umbilical vein endothelial cells, and frog embryos (unpublished data), in addition to the budding yeast and E. coli. Nevertheless, we do not exclude the possibility that the auto-processing of MYRF can in principle be regulated by some mechanism. The apparently constitutive processing of MYRF presents a puzzle, as our data suggest that the processing is not a key step by which MYRF might be externally regulated. Indeed, Myrf (the mouse ortholog) has been shown to be regulated mainly at the transcriptional level [17]. Moreover, another recent study indicates that Myrf is continuously needed throughout adulthood to maintain myelin, consistent with a constitutive process [52]. Why, then, are MYRF and its orthologs MBTFs? While we have demonstrated that the mechanism clearly supports the proper assembly of the N-terminal transcription factor trimer, we speculate that this might additionally represent a mechanism by which the generation of two functionally independent trimers can be mandatorily coupled to coordinate regulation of nuclear and ER processes.
Several pieces of circumstantial evidence support this speculation: First, PQN-47, the C. elegans ortholog of MYRF, has recently been implicated in the regulation of molting—a process involving extensive secretion [51]. Notably, an intact ER lumenal domain, localized outside the nucleus, was shown to be critical for PQN-47's regulation of molting. Second, our CG4 maturation assay shows that the N-terminal trimer alone is not as competent as full-length MYRF in promoting the maturation of CG4 cells, suggesting that the unknown function of the C-terminal homo-oligomer may be as essential for OL maturation as the transcriptional function of the N-terminal trimer. Notably, the physiological processes and genes to which MYRF and its orthologs have been linked all involve the secretory pathway. EcmA (an endogenous target gene of MrfA [37]) is a secreted protein. Myelination, in which Myrf plays an essential role, places heavy demands upon the secretory pathway, as does molting, for which PQN-47 is critical [51]. Finally, a meta-analysis of microarray data in the Gene Expression Omnibus database [53] indicates that MYRF is significantly expressed in secretory tissues including stomach and lung (unpublished data). We speculate that in the context of OL differentiation, while the N-terminal trimer acts in the nucleus to stimulate production of myelin components, the C-terminal homo-oligomer either coordinates their orderly passage through the secretory pathway or functions as part of the UPR pathway to prepare the ER for the increased flux of myelin components. In the future, it will be interesting to explore whether MYRF is truly a dual-functional protein and what function the C-terminal homo-oligomer performs in the ER for OL maturation.
In fact, the recognition of this protein family as MBTFs serves to reconcile apparently contradictory, but likely correct, findings in the prior literature. The report on Myrf ascribed its role to a master transcriptional regulator for OL maturation and CNS myelination [17]. A subsequent study on PQN-47 questioned this role for Myrf, mainly because a PQN-47::GFP translational fusion protein localized outside the nucleus in C. elegans [51]. Based on molting phenotypes, the PQN-47 study concluded that PQN-47 (and Myrf, by implication) might play an important role in the secretory pathway. On the other hand, a recent study in Dictyostelium showed that the DNA-binding domain of MrfA endogenously localizes to the nucleus and binds DNA directly, supporting the conclusion of the report on Myrf [37].
Our finding that MYRF is a MBTF that is auto-processed into two independent homo-oligomers entirely resolves these seemingly conflicting reports: Depending on the location of the epitope tag and whether auto-processing is blocked or not, MYRF and its orthologs can exhibit either nuclear or ER localization, as appropriate; each of these prior studies is consistent with this interpretation. Taken together, MYRF and its orthologs represent a new class of MBTFs that require auto-processing to function in gene transcription and likely also play important roles beyond transcription, including in secretion.
The MYRF cDNA was purchased from Open Biosystems (the 1111-amino-acid-long isoform [CCDS ID: 31579 and RefSeq ID: NP_037411]). HeLa cells were cultured in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum. Cells were maintained in a humidified 5% CO2 incubator at 37°C. Transient transfection was performed using either FuGENE HD or Lipofectamine 2000.
Cells grown on 150 mm culture dishes were rinsed once with PBS, and 500 µL of 2× Cell Lysis Buffer (Cell Signaling) was directly added to the cell layer. Cell lysates were sonicated and then spun down at 14,000× g for 10 min at 4°C. The supernatant was mixed with FLAG antibody-coated beads (Sigma) and incubated for 2 h at 4°C on a rotating plate. The mix was spun down at 7,500× g for 30 s to separate supernatant (“Sup”) from pellet (“Bead”) fractions.
Cells were rinsed once with PBS and then lysed directly in wells with 1× Laemmli Sample Buffer (Bio-Rad). Cell lysates were boiled at 95°C for 5 min. Upon SDS-PAGE, the proteins were transferred to PVDF and probed with primary and horseradish peroxidase (HRP)-conjugated secondary antibodies. The following dilutions were used for immunoblotting: mouse anti-FLAG (1∶1000, Sigma), rabbit anti-c-Myc (1∶250, Santa Cruz Biotechnology), goat anti-actin (1∶400, Santa Cruz Biotechnology), rabbit anti-calnexin (1∶400, Santa Cruz Biotechnology), goat anti-calnexin (1∶400, Santa Cruz Biotechnology), goat anti-mouse HRP-conjugated (1∶10,000, Santa Cruz Biotechnology), mouse anti-FLAG HRP-conjugated (1∶1,000, Sigma), and mouse anti-c-Myc HRP-conjugated (1∶400, Santa Cruz Biotechnology).
Cells cultured in glass bottom six-well plates (In Vitro Scientific) were fixed with 4% formaldehyde, permeabilized in cold 100% methanol or 0.1% Triton X-100, blocked with 1% BSA in 1× PBS with 0.05% Tween, and incubated with primary antibody diluted in blocking buffer at 4°C overnight, followed by incubation with fluorochrome-conjugated secondary antibody. Nuclei were stained with Hoechst 33342 (Invitrogen). Fluorescence was visualized with a Nikon Eclipse TE2000-E fitted with a Plan Apo VC 100×/1.40 oil objective and a digital camera (Cascade II 512; Photometrics) controlled by the NIS Elements software (AR 3.0). To selectively permeabilize plasma and ER membranes, 25 µg/ml digitonin was used to treat cells for 5 min on ice, followed by fixation with 4% formaldehyde. Rhodamine-conjugated donkey anti-goat IgG was from Santa Cruz Biotechnology. Alexa Fluor 594 goat anti-mouse or rabbit IgG and Alexa Fluor 488 goat anti-rabbit IgG were from Invitrogen.
The truncated MYRF (MYRF-319:708) was inserted into pET52b (Novagen) between BamHI and SacI to generate pET52b-StrepII-MYRF-319:708-10xHis. This plasmid was transformed into BL21 Star (DE3) pLysS E. coli (Invitrogen). Cells were cultured at 37°C to OD600 nm 0.4–0.6 in LB and protein expression was induced by 0.5 mM IPTG at 16°C for 16–18 h. Cells were collected and lysed by sonication in lysis buffer (20 mM Tris pH 8.0, 500 mM NaCl, 10 mM β-mercaptoethanol, 10% glycerol). The lysate was clarified by centrifugation at 15,000× g for 30 min at 4°C and the supernatant was loaded onto a Ni-NTA column (Qiagen). The flow-through was loaded onto a Strep-Tactin chromatography column (IBA) to affinity purify the N-terminal fragment of MYRF-319:708 according to the manufacturer's purification protocol. The eluted protein was dialyzed in dialysis buffer (20 mM HEPES pH 7.5, 50 mM NaCl, 10 mM β-mercaptoethanol, 10% glycerol) and concentrated to about 2 mg/ml. The concentrated protein was analyzed by Superdex 200 10/300 GL gel filtration column chromatography.
RNA was extracted from HeLa cells using Trizol (Invitrogen). cDNA was generated using SuperScript First-Strand Synthesis System for RT-PCR (Invitrogen). qRT-PCR was performed using PowerSYBR Green PCR Master Mix (Invitrogen) and ABI ViiA 7 Real-Time PCR System. Primer sequences are available in Table S2.
CG4 cells were maintained in GM [70% of CG4 growth medium (Dulbecco's modified Eagle's medium, 5 µg/ml transferrin, 100 µM putrescine, 20 nM progesterone, 30 nM selenium, 10 ng/ml biotin, and 5 µg/ml insulin) supplemented with 30% of the same medium conditioned by B104 cells]. For maturation assays, CG4 cells were plated on glass bottom six-well plates coated with poly-L-ornithine. 0.4 µg of plasmid DNA was transfected using Lipofectamine 2000 (Invitrogen) for 4 h. After transfection, GM was replaced by DM (CG4 growth medium supplemented with 1% FBS, 40 ng/ml triiodothyronine). CG4 cells were maintained in DM for 4 d before immunostaining for cell counting. Primary antibodies used were 1∶500 mouse anti-O1 (Millipore) and 1∶500 rat-anti-MBP (Millipore). For each sample, cells of at least 50 random fields were counted in a blind fashion.
|
10.1371/journal.pgen.0030119 | Population Bottlenecks as a Potential Major Shaping Force of Human Genome Architecture | The modern synthetic view of human evolution proposes that the fixation of novel mutations is driven by the balance among selective advantage, selective disadvantage, and genetic drift. When considering the global architecture of the human genome, the same model can be applied to understanding the rapid acquisition and proliferation of exogenous DNA. To explore the evolutionary forces that might have morphed human genome architecture, we investigated the origin, composition, and functional potential of numts (nuclear mitochondrial pseudogenes), partial copies of the mitochondrial genome found abundantly in chromosomal DNA. Our data indicate that these elements are unlikely to be advantageous, since they possess no gross positional, transcriptional, or translational features that might indicate beneficial functionality subsequent to integration. Using sequence analysis and fossil dating, we also show a probable burst of integration of numts in the primate lineage that centers on the prosimian–anthropoid split, mimics closely the temporal distribution of Alu and processed pseudogene acquisition, and coincides with the major climatic change at the Paleocene–Eocene boundary. We therefore propose a model according to which the gross architecture and repeat distribution of the human genome can be largely accounted for by a population bottleneck early in the anthropoid lineage and subsequent effectively neutral fixation of repetitive DNA, rather than positive selection or unusual insertion pressures.
| Throughout evolutionary history, fragments of the mitochondrial genome, known as numts (for nuclear mitochondrial sequences), have been inserted into the nuclear genome. These fragments are distinct from all other classes of repetitive DNA found in nuclear genomes, not least because they are incapable of mediating their own proliferation. Taking advantage of their unique evolutionary properties, we have used numts to improve our understanding of the architecture of the human genome with special emphasis on the mechanism of acquisition and retention of repeat sequences, which comprise the bulk of nuclear DNA. We find that numts are unlikely to have any evolutionary benefit driving their retention. Moreover, numts are not acquired randomly during evolutionary time. Instead, their rate of acquisition spikes dramatically around pronounced population bottlenecks, in a manner reminiscent of other repeat classes. Therefore, we propose that the primary driving force of repeat acquisition in the genome is not selection, but random genetic drift, whose force becomes pronounced during profound reductions of population size. Our findings support the theory of neutral evolution, according to which random genetic drift exerts an influence on the acquisition of DNA changes that far outweighs the power of positive selection.
| The present-day human genome arose from the prosimian ancestor through a series of complex chromosomal and local rearrangements. An important feature of our genome, used frequently to understand the adaptive forces that have led to its present-day topology, is the common prevalence of repetitive sequences. Analyses of the Alu family, a 300-bp, primate-specific retrotransposon that represents the most abundant class of repeats [1], have indicated that they underwent a seemingly rapid proliferation at two major evolutionary junctions: the prosimian-anthropoid split some 37–55 million years ago (mya) and the platyrrhine/catarrhine split thereafter [2]. Some studies have pointed to a correlation between retrotransposon expansion and speciation [3,4] and have suggested that the unidirectional proliferation of more than ten copies of the retrotransposon [1,5] might provide a useful marker for tracing phylogeny [6,7].
Despite the apparent importance of repeat expansion to understanding the origins of the human genome, the mechanisms of repeat proliferation are poorly understood. For Alu repeats, a model of increased retrotransposition activity has been proposed [8], but the underlying evolutionary forces behind such a mechanism are unclear.
To investigate the evolutionary forces that might govern the acquisition and retention of repetitive elements in the human genome, we selected an entirely different class of repeat whose mechanisms for insertion, deletion, and selection are so fundamentally different from Alu that any commonality in their evolutionary dynamic is probably due to the fact that they share the same population size, rather than any underlying biological mechanism.
We focused on numts (nuclear mitochondrial sequences/pseudogenes), partial copies of the mitochondrial genome found abundantly in chromosomal DNA. Since the first demonstration of organellar sequence embedded in nuclear DNA [9], numts have been described in several mammalian species, as well as over 70 other eukaryotes [10–12]. The varying level of homology between these sequences and the present-day mitochondrial genome, as well as population and family polymorphisms, indicates that the nuclear transfer of mtDNA is an ongoing process [13–21,28]. In contrast to plants and fungi, in which numts have arisen from both RNA- and DNA-mediated mitochondrial DNA (mt-DNA) transfers [22], the origin of numts in metazoans has been proposed to be DNA- rather than RNA-mediated [23–25]. As such, the numts family of repeats represents a useful tool for evolutionary analysis since its proliferation mechanism is distinct from Alu elements, in that it does not rely on retrotransposition.
We first used the assembled human genomic sequence (Build 36) to investigate the prevalence and distribution of numts in the human genome. Using default sequence alignment selection criteria (e-value <10), we identified 2,329 numts fragments that range in size from <100 bp to 16 kb (Figure 1), a number consistent with previous studies [19,23,26]. Fine-mapping of numts showed many instances in which multiple, seemingly independent, fragments map in close proximity to one another, suggesting a higher-order organization, whereby each numts does not represent an independent integration, but is rather a fossil of a single ancestral integration (Table S1). Clustering of such numts blocks indicated that the human genome likely contains in excess of ∼1,200 numts elements (Table S2). A similar analysis of the mouse and rat assembled genomes showed a marked numts paucity, with 636 and 529 numts fragments, respectively. By contrast, the recent draft of the chimp genome contains numbers comparable to humans, ≥1,280 numts, suggesting that these elements might have undergone a dramatic expansion in the primate lineage (Table S2). These observations are unlikely to be due to inappropriate exclusion of numts sequences from the draft genome assemblies, since analysis of the raw trace data (i.e., all individual preassembly sequence reads) showed a similar percent identity distribution of putative numts, with both sequence collections peaking at 82%–88% identity with the present-day mitochondrial sequence (data not shown).
Prior to further analysis, we corroborated our computational data in two ways. First, we performed flourescent in situ hybridization (FISH) with mtDNA as a molecular probe on interphase and metaphase nuclei of mtDNA-depleted cells as target DNA. Consistent with the predicted abundance of numt in the nuclear genome, we detected fluorescence signals scattered along each chromosome (Figure 2). We observed a similar pattern on chromosomes of mtDNA-depleted lymphoblast cells from chimp, gorilla, and orangutan (Figure 2). These data indicate that the numts element is distributed widely in the genomes of these species and that the actual numts population is probably larger than our computational predictions, potentially reflecting our criteria for numts identification. In addition, amplification from a monochromosomal hybrid panel and subsequent sequencing of 24 randomly selected nucleo–numts junctions, showed that in each case the amplification and sequence data matched exactly with the computationally predicted sequence of each numts (data not shown).
We next investigated numts proliferation. Previous studies have indicated that the mechanism of integration of these repeat elements into the genome is distinct from retroviral insertion or recombination [10], thus enabling us to study the acquisition characteristics of exogenous DNA in a genome context-independent fashion. To identify a subpopulation of numts that arose by independent integrations, rather than a single integration followed by subsequent segmental duplication, we first correlated the positions of all identified numts with the segmental duplication map. In agreement with previous studies founded on numts base substitution rates [13], we determined that although some numts proliferated through chromosomal rearrangements, the majority of numts acquisition of the genome reflects independent integration; some 3%–5% of build 36 has been identified as segmental duplication [27], and only 4% of all numts map to these regions. To further confirm these observations, we compared 500 bp of nuclear sequence on either side of each putative integration and found no similarities among the nuclear junction sequences (data not shown).
We next asked whether numts integration is likely to be genome sequence independent by evaluating the sequence characteristics of nucleo–numts junctions. First, we asked whether there is any observable enrichment for a recognizable element at repeat junctions. A comparison of 1 kb of flanking nuclear junction sequence surrounding 266 numts with the entire human genome showed an initial deficit of repeats, returning to genome-wide levels 500–600 bp past the insertion boundary (Figures 3 and S1). This suggested that: (a) there is no repeat excess at the boundary and (b) the true boundary probably lies 500–600 bp away from our initial prediction. In addition, the possibility of a TE (transposable element) insertional mechanism was also deemed unlikely, since we found no evidence of sequence duplication anywhere within the 1kb region that flanks the boundaries of each numt.
Our data suggest that the human genome has probably acquired a minimum of several hundred numts, most of which arose in an ancestor as independent events, in a process that is still ongoing [28] and can have detrimental effects to gene function [29]. Even though the mechanism of insertion of numts is clearly different from that of Alu elements, especially since numts cannot mediate their own proliferation, similarities or differences in the fitness consequences of those insertions are less obvious. Although numts are unlikely targets for unequal exchange events, they might contain potentially functional genes that could be co-opted into some nuclear role. Thus, we assessed for possible fitness effects of numts insertion by examining their positional preference in the genome, as well as their transcriptional and translational potential.
To interrogate whether numts have positional preference, we determined the relative distribution of all large numts arisen by independent integrations with respect to the coding sequence distribution of the genome. We conducted two tests, one for numts >1 kb (n = 99) and one for numts >500 bp (n = 121). None of the numts considered for the two experiments occurred in exons. In build 36, the fraction of the intronic human genome is ∼28.85%. The percentage of intronic numts is 22.3% (22/99; binomially p = 0.086) for numts >1 kb and 21.5% (26/121; p = 0.042) for the those >500 bp. Thus, numts appear to be distributed relatively randomly in the genome (Figure 4), but a slight statistical tendency towards intergenic intervals was observed, probably underlying the higher potential of intragenic insertions for a deleterious effect. Overall, we conclude that numts position within the genome provides little evidence of its use for transcriptional control.
Next, we considered the possibility that numts might have functionality at the mRNA level. We first examined whether numts are transcribed, by interrogating each numts against dbEST. To reduce the incidence of matches with dbEST due to short segments of sequence, we restricted our queries to numts with length greater than 1 kb and numts longer than 500 bp. Of the 99 numts >1 kb evaluated, (23/99) 23.23% were represented in dbEST, also from the 121 numts >500 bp considered, (33/121) 27.27% were found in dbEST. Reverse-transcriptase PCR (RT-PCR) of 24 randomly selected, nonoverlapping ESTs also indicated that the majority of these sequences represent bona fide transcription, since in 22 instances we amplified successfully the correct fragment from a panel of eight adult human RNA samples by RT-PCR (data not shown). However, we found no positional preference for putatively transcribed numts, suggesting that numts mRNA is unlikely to exert a cis-acting regulatory role.
Finally, we considered the possibility that the introduction of numts into the genome provided the template for new protein sequence, despite the fact that the nuclear and mitochondrial genome have different genetic codes. We therefore examined the translational potential of each numts in all six reading frames (Figure 5). Translating with the nuclear code results in a distribution of open reading frame (ORF) lengths indistinguishable from random sequence (3/64 codons are stop, therefore random sequence will generate ORF sizes with a mean size of ∼20 codons). Although there is a slight excess of long ORFs (suggesting that a small fraction of numts might be translated), the overall distribution of ORF lengths is approximately exponential with a mean length of 5–15 codons.
Cumulatively, our data suggest that there is little evidence for overt functionality for the majority of numts, and although we cannot formally exclude the possibility that some individual repeats have a biological role (and may thus be obvious targets for positive selection), the overall population of this repeat is likely to be on average evolutionarily neutral or deleterious.
To gain a better understanding of the evolutionary dynamics of numts, we sought to determine the most likely time of integration of each numts into the nuclear genome. To do so, we aligned each numts to a collection of complete modern mtDNA sequences spanning the primate radiation. The time of each integration was inferred independently with multiple fossil calibration points [30] under an overdispersed model of molecular evolution, accounting for variation in evolutionary rates within and between numts and the extant mitochondria (Figure 6A) [31]. In contrast to an expectation of progressive numts accumulation during evolutionary time, we were surprised to find an apparent burst of numts integrations at approximately 54 mya. Focusing first on numts >1 kb in length, we found that ∼76% out of the 99 unique integration events, have an estimated time of insertion within 10 mya of 54 mya (Figure 6C). Next, we considered the numts >500 bp, and from 121 unique integration events ∼75% also occurred within 10 mya of 54 mya (Figure 6E). Thus, 75%–80% of all numts integrations appear to have occurred within a relatively narrow window of time around 54 mya, between the New World Monkey and Old World monkey transition (Figure 6B and 6D). Importantly, this estimate is likely to remain true irrespective of assumptions regarding the nucleotide substitution rate of numts versus mtDNA, as judged by a confidence interval plot of the 121 500-bp+ numts (Figure S2).
Most numts appear to have accumulated in a 10-millon-year window centered around 54 mya. Importantly, other repetitive elements show a similar pattern, including Alu repeats [2,32] and processed pseudogenes [33], suggesting a period of intense DNA acquisition in the ancestral genome. Given that numts are markedly distinct from Alu repeats and other retrotransposons in both their mechanism of integration, as well as proliferation (especially since numts lack the ability to self propagate), the force behind the expansion of repeats is likely independent of genome structure. This notion is further supported by the fact that the boundaries of numts integration show no marked enrichment for any sequence elements (Figure 3). It will always remain a formal possibility that numts integration was primarily driven by positive selection for the accumulation of these elements. However, the absence of overt functionality of numts in the present-day genome, and the fact that numts integration is a continuing process [10], principally detected because of its disease phenotype, argues against this hypothesis. Thus, we arrive at three important questions concerning the evolutionary history of numts: (1) Why did so many numts accumulate approximately 54 mya? (2) Why did they stop accumulating? (3) Why does this time period correspond temporally with accumulation of other entirely unrelated genetic elements?
The theory that governs the evolutionary dynamics of TEs can provide important clues about the mechanism of acquisition and retention of numt, Alu, and other repeat elements in the human genome. In an infinite sized population, the change in the mean number of TEs per individual,
, is approximately
where Vn is the variance in copy number between individuals, μ is the rate of new insertions, ν is the rate of new deletions, and
is population mean fitness [34,35]. Thus, in an infinite-sized population, TE copy number is governed by a balance between the effects of new insertion, new deletion, and selection. By contrast, in a finite population, Equation 1 will approximately hold whenever
is much bigger than 1/N, where N is the effective size of the population. If 1/N >
, TE copy number will rise (if the insertion rate is greater than the deletion rate) or fall (if deletion is more frequent than insertion). Thus, a sudden change to TE copy number could reflect a sudden decrease in population size, shifting the balance between selection and mutation forces to one where genetic drift ruled and allowed for unbounded increase in TEs. The Liu et al. hypothesis [8], on the other hand, suggests that the increase in Alu copy number may have resulted from a sudden increase in μ, the rate of insertion.
If we assume that numts integrations are principally weakly deleterious on average (a notion supported by their ongoing contribution to disease), an examination of Equation 1 suggests that a simple population size hypothesis can provide an answer to all three of our questions. We begin by assuming that prior to 54 mya, the effective population size of the primate ancestor was relatively large, leading to an insertion/deletion/selection equilibrium with numts count being few and held stable at that low value (which is consistent with the relative paucity of numts in the mouse and rat lineages). However, if we further assume that at approximately 54 mya, effective population sizes declined dramatically, to a point where 1/N >
, then numts would for evolutionary purposes become effectively neutral, and, during their period of effective neutrality, they would accumulate with little selective check, at a rate proportional to μ−ν (the difference between the insertion and deletion rates of an element). Since population size changes affect everything in the genome, elements with high insertion rates (such as Alu elements) would be expected to accumulate in great abundance (which they do), whereas elements with relatively low insertion rates (such as numts) also accumulated, albeit in fewer numbers. Finally, a subsequent increase in effective population size would shift the population back into an insertion/deletion/selection equilibrium, and the period of accumulation would end.
Clearly, the assumptions of relative numts neutrality and of a population bottleneck at ∼54 mya cannot be proven definitively. Nonetheless, based on observations of the landscape of the present day genome of humans and other species, our proposed evolutionary model has many attractive features. First, it provides a common mechanism (decline in effective population size) for the increase in numbers of unrelated repetitive elements. Second, it explains both the sudden increase in repetitive DNA, and the later cessation of the increase. Third, the timing of the event, occurring immediately prior to the adaptive radiation of monkeys, is highly evocative, reminiscent of a Wrightian/Simpsonian view of speciation: a large population of stem anthropoids splintered into multiple demes. One or more such small deme accumulated repetitive DNA in abundance, which in turn may have served as a post-zygotic reproduction barrier with the original population. This isolated deme ultimately speciated and underwent an adaptive radiation into the anthropoid primates. It is notable (and unlikely to be coincidental) that the timing of the repeat-inferred bottleneck at ∼54 mya coincides with a major environmental disturbance at the Paleocene–Eocene boundary (∼55 mya), which strongly effected global mammalian faunas and corresponds to the first appearance of primates in the fossil record of the northern hemisphere [36].
This hypothesis suggests that human and primate genomic architecture, with its abundance of repetitive elements, arose primarily by evolutionary happenstance; although it remains plausible (and indeed, probable) that some integrons were subsequently co-opted into an interesting use such as X inactivation [37] or perhaps gene regulation [38], these complicated hypotheses do not explain satisfactorily the bulk of human genomic architecture. A simple explanation states that the population that gave rise to primates was quite small, and as a result the genomic architecture of primates may have resulted from effectively neutral integrations of repetitive DNA.
Human mitochondrial genome sequence was compared against genomic sequence with BLAST (NCBI Build 36). The process was repeated for the mitochondrial sequence of chimp, mouse and rat against the following draft builds: chimp Build 2 (October 2005), mouse Build 33 (May 2004; mm5), and rat Version 3.1 (June 2003; rn3). In each case, hits that scored with an expected value <10 were retained. All annotations (repeat classes, gene boundaries, etc.) were taken from the University of California Santa Cruz genome browser, http://genome.ucsc.edu/.
Blast hits were sorted by genomic position, and the differences (“gaps”) between consecutive hits on both the genomic and mitochondrial scales were calculated. Pairs of hits that had a ratio of mitochondrial gap size to genomic gap size between 0.9 and 1.1 were assigned to be in the same block (hand picked). The numts distribution plots were created using Circos (http://mkweb.bcgsc.ca/circos/).
We used high-molecular-weight genomic DNA and highly purified mt-DNA from HeLa cells (kindly provided by Samuel E. Bennett, Oregon State University, Corvallis, Oregon, United States) for PCR. For generating molecular probes in FISH experiments, we used two different PCR products: the complete mitochondrial genome (16.3 kb) amplified with the TaKaRa PCR kit (Fisher Scientific, https://new.fishersci.com/), using conditions as described [39]. Alternatively, we designed seventeen PCR primer sets and amplified overlapping ∼1-kb fragments, covering the entire mt-DNA sequence. Primers and detailed PCR conditions are available upon request.
The nonhuman primate immortalized Epstein–Barr virus–stimulated cell lines of common chimpanzee (Pan troglodytes), lowland gorilla (Gorilla gorilla, CRL 1854), and orangutan (Pongo pygmaeus), were purchased from the American Type Culture Collection (ATCC, http://www.atcc.org/). The pygmy chimp (Pan paniscus) lymphoblast sample was kindly provided by D. Nelson at Baylor College of Medicine, Houston, Texas, United States.
Human and primate lymphoblasts were depleted of mt-DNA according to the slightly modified protocol of King and Attardi [40]. Cells were grown for 5–6 d in DMEM enriched with 10% FCS glucose (4,500 mg/ml), sodium pyruvate (1 mM), uridine (50 μl/ml), and ethidium bromide (50 μl/ml).
Normal and mt-DNA-depleted lymphoblasts were harvested using standard methods. FISH was performed on metaphase and interphase cells as described [41]. Briefly, PCR products were labeled with biotin (Life Technologies-GibcoBRL, http://www.invitrogen.com/) or digoxigenin (Boehringer Mannheim, http://www.roche.com/) by nick translation. Biotin was detected with FITC-avidin DCS (fluoresces green; Vector Labs, http://www.vectorlabs.com/) and digoxigenin was detected with rhodamine-anti-digoxigenin antibodies (fluoresces red; Sigma, http://www.sigmaaldrich.com/). Chromosomes were counterstained with DAPI diluted in Vectashield antifade (Vector Labs). Cells were viewed under a Zeiss Axioskop fluorescence microscope (http://www.zeiss.com/) equipped with appropriate filter combinations. Monochromatic images were captured and pseudocolored using MacProbe 4.2.2/Power Macintosh G4 system (Apple, http://www.apple.com/; Perceptive Scientific Instruments, http://www.perceptive.co.uk/).
The flanking sequence composition of 266 numts was compared to 50,000 randomly chosen sequences drawn uniformly from the human genome. For each flanking sequence, and each randomly drawn sequence, the proportion of the sequence covered by various repeat families (Alu, L1, MALR, etc.) and repeat classes (SINE, LINE, LTR, etc.) was calculated and the repeat composition of each category was evaluated with a t-test.
Once the composition and distribution of numts blocks was established, we designed primers to amplify 250–400-bp junction fragments whereby one primer was anchored at unique nuclear sequence and the other primer was situated at the edge of a numts block. We performed PCR using standard condition on human–rodent monochromosomal hybrids as described [42].
We designed primers from ESTs that matched human numts with >98% identity over 200 bp of sequence. To ascertain their expression patterns, we generated amplicons from eight adult human cDNAs (Clontech, http://www.clontech.com/) according to manufacturer's instructions.
Each numts was translated in all six possible reading frames. An ORF was defined as the sequence between two stop codons, and the frame with the longest mean ORF length was chosen for inclusion in the analysis. Numts were translated using the nuclear genetic codes (stop codons TAA/TAG/TGA).
Each numts was aligned individually with ClustalW (http://www.ebi.ac.uk/clustalw/) to a collection of complete modern mtDNA sequences spanning the primate radiation, rooted by a carnivore outgroup. All pairwise per-site divergences were calculated with the PHYLIP program (http://evolution.genetics.washington.edu/phylip.html) dnadist, using a Kimura 2-parameter substitution model to correct for multiple hits. For each numt, the evolutionary tree was inferred by both parsimony (using the PHYLIP dnapars program) and neighbor-joining (using the PHYLIP program neighbor). In all cases the expected phylogeny [2] of the primate and outgroup was recovered, but the exact position of the numts varied slightly (see below). Once the tree was inferred for each numt, the number of substitutions per branch was estimated by least-squares minimization using the PHYLIP program fitch with default parameters.
To account for any potential uncertainty in the divergence time between extant primates, nonconstancy of evolutionary rates within and among different functional portions of the extant mtDNA, and perhaps vastly different rates of evolution among nuclear pseudogene copies of mtDNA and extant functional mtDNA, the time of each integration was inferred with dating [31], under a stationary substitution model with multiple fossil calibration points [30]. In all cases, the stationary model fit better than the constant rate Poisson model by several orders of magnitude. Confidence intervals for each integration were also calculated [31].
The National Center for Biotechnology Information (NCBI) Genbank (http://www.ncbi.nlm.nih.gov/sites/entrez?db=Nucleotide) accession number for the human mitochondrial genome sequence discussed in this paper is NC_001807. |
10.1371/journal.ppat.1004135 | CD8+ T Cells from a Novel T Cell Receptor Transgenic Mouse Induce Liver-Stage Immunity That Can Be Boosted by Blood-Stage Infection in Rodent Malaria | To follow the fate of CD8+ T cells responsive to Plasmodium berghei ANKA (PbA) infection, we generated an MHC I-restricted TCR transgenic mouse line against this pathogen. T cells from this line, termed PbT-I T cells, were able to respond to blood-stage infection by PbA and two other rodent malaria species, P. yoelii XNL and P. chabaudi AS. These PbT-I T cells were also able to respond to sporozoites and to protect mice from liver-stage infection. Examination of the requirements for priming after intravenous administration of irradiated sporozoites, an effective vaccination approach, showed that the spleen rather than the liver was the main site of priming and that responses depended on CD8α+ dendritic cells. Importantly, sequential exposure to irradiated sporozoites followed two days later by blood-stage infection led to augmented PbT-I T cell expansion. These findings indicate that PbT-I T cells are a highly versatile tool for studying multiple stages and species of rodent malaria and suggest that cross-stage reactive CD8+ T cells may be utilized in liver-stage vaccine design to enable boosting by blood-stage infections.
| Malaria is a disease caused by Plasmodium species, which have a highly complex life cycle involving both liver and blood stages of mammalian infection. To prevent disease, one strategy has been to induce CD8+ T cells against liver-stage parasites, usually by immunization with stage-specific antigens. Here we describe a T cell receptor specificity that recognizes an antigen expressed in both the liver and blood stages of several rodent Plasmodium species. We generated a T cell receptor transgenic mouse with this specificity and showed that T cells from this line could protect against liver-stage infection. We used this novel tool to identify the site and cell-type involved in priming to a recently developed intravenous attenuated sporozoite vaccine shown to have efficacy in humans. We showed that CD8+ T cells with this specificity could protect against liver-stage infection while causing pathology to the blood stage. Finally, we provided evidence that T cells with cross-stage specificity can be primed and boosted on alternative stages, raising the possibility that antigens expressed in multiple stages might be ideal vaccine candidates for generating strong immunity to liver-stage parasites.
| Malaria is a mosquito-transmitted disease found in a range of animals including man, non-human primates and rodents. It is caused by multiple Plasmodium species, several of which may infect the same animal species. For humans, the two most prevalent Plasmodium species are P. falciparum and P. vivax, with the former responsible for the bulk of lethal disease. Mice have been used as a convenient animal model for studying malaria, with three rodent Plasmodium species in use: (i) P. chabaudi, which can cause a disease that shows recrudescence and has many features in common with human malaria including anemia, sequestration of parasites, and metabolic acidosis [1]; (ii) P. yoelii, which has two very closely related strains that differ in their capacity to infect red blood cells and cause lethal disease [2]; and (iii) P. berghei, particularly the ANKA strain (PbA), which has been used as a model for human cerebral malaria [3], [4], [5], a lethal complication of P. falciparum infection. While there is much debate as to the relevance of the PbA rodent infection model to human disease, the pathological processes underlying human cerebral malaria are relatively poorly characterized, making it difficult to accurately compare human and murine diseases. However, like human severe malaria, high parasite burden is required for multi-organ pathology in the PbA model [6], [7], [8]. In itself, the pathological process underlying experimental cerebral malaria (ECM) seen in PbA infections also offers insight into immune-mediated pathology in general, providing a rigorous experimental approach that can be easily manipulated to decipher various cellular and molecular contributions. In this rodent model, various cell types and cytokines have been reported to contribute to lethal ECM, with CD8+ T cells a major and essential contributor [9], [10], [11]. Infection with PbA leads to the activation of parasite-specific T cells that first expand in the spleen and then migrate to the brain, where they cause pathology [11]. Depletion of CD8+ T cells shortly before the onset of ECM prevents disease [11], supporting a role for these cells in the effector phase of disease pathology.
Plasmodium species have a complex life cycle with several distinct stages: a mosquito stage, from which sporozoites emerge to enter the mammalian hosts during a blood meal; a liver-stage where sporozoites enter hepatocytes and eventually develop into a large cohort of merozoites; and a blood stage, where merozoites are released into the blood and cause cyclic infection of erythrocytes. Disease symptoms and immune mediated pathology associated with malaria are limited to the blood-stage of infection, with the preceding liver stage being asymptomatic [12]. Despite this, sporozoite infection is not immunologically silent, with evidence that following pathogen entry via a mosquito bite, the immune response is initiated in the skin draining lymph nodes of mice [13], generating protective immunity that depends on CD8+ T cells and the cytokines TNFα and IFNγ [14]. Sporozoite-specific immunity can control infection in mice [15], non-human primates [16] and humans [17], [18], preventing development of blood-stage infection and its associated disease. As a consequence, researchers have explored the use of live sporozoites attenuated by irradiation or genetic engineering [19], [20], [21] or non-attenuated sporozoites controlled by drug curing, as potential approaches to vaccination [22]. Administration of irradiated cryopreserved sporozoites via the intravenous route was shown to provide superior immunity compared to cutaneous injection in non-human primates and mice [19]. More recently, vaccination of humans by the intravenous route demonstrated protection [21]. The success of the intravenous route was speculated to result from the direct access of parasites to the liver for development of immunity at this site. However, direct examination of where immunity was generated to this effective route of vaccination was not attempted.
During the different life-cycle stages, Plasmodium parasites adopt distinct morphologies and as a consequence express many stage-specific proteins, which are often the focus of immunity and vaccine design. However, many proteins are expressed throughout multiple stages of the life cycle [23] and in the mammalian host may be expected to contribute to immunity across multiple stages. While it has been suggested that blood-stage immunity may impair responses to liver-stage antigens [24], others have shown protection against liver-stage infection by prior blood-stage infection and cure [25], supporting the idea that antigens expressed at both stages may be capable of inducing protective immunity. However, direct demonstration of this capacity was not provided.
Here we describe the development of an MHC I-restricted, T cell receptor (TCR) transgenic murine line specific for PbA. We show that transgenic T cells from this line recognize an antigen expressed in both the blood-stage and the liver-stage of infection, demonstrating the potential for T cells with blood-stage-specificity to protect against sporozoite infection. T cells from this line detect a conserved antigen expressed by several rodent Plasmodium species including P. chabaudi and P. yoelii, rendering it a highly versatile immunological tool for dissecting CD8+ T cell immunity in malaria.
An MHC I-restricted TCR transgenic mouse line specific for blood-stage PbA (termed PbT-I) was generated using TCR genes isolated from a Kb-restricted hybridoma termed B4 (Figure S1) originally derived from a T cell line isolated from a B6 mouse infected with blood-stage PbA. Analysis of spleen and lymph node (LN) cells from PbT-I mice showed a strong skewing towards CD8+ T cells (Figure 1A), with essentially all splenic (Figure 1B) and lymph node (Figure S2) CD8+ T cells expressing the Vα8.3 and Vβ10 transgenes. The few CD4+ T cells detected in the spleen and lymph node also expressed these transgenic receptors, though at a lower level indicative of co-expression of endogenous receptors. There was no reduction in spleen or lymph node cellularity relative to wild-type mice, with CD8+ T cells substituting for the lack of CD4+ T cells (Figure S3). Peripheral skewing towards CD8+ T cells was reflected in the thymus, where a large population of mature CD8+CD4− T cells with high TCR expression was evident (Figure S4). In this case, total thymocyte numbers were reduced to about one third of wild-type (Figure S3), consistent with the cellularity of other TCR transgenic mice we have generated, and likely due to efficient positive selection [26].
To determine if PbT-I cells responded to blood-stage PbA, purified CD8+ T cells from PbT-I mice were labeled with CFSE and then stimulated in vitro with dendritic cells and lysate from either infected red blood cells (iRBC) of mixed stages or enriched as schizonts (Figure S5). This showed a dose-dependent proliferative response to both forms of antigen, though schizont lysate was more efficient.
To test whether PbT-I cells also responded to PbA in vivo, PbT-I cells were labeled with CFSE and adoptively transferred into B6 mice one day before infection with blood-stage PbA. Three or 5 days later, mice were killed and the spleen and blood examined for proliferating PbT-I cells (Figure 2A, B). This revealed a vigorous response by PbT-I cells, which entered the blood from the spleen after day 3. The specificity of PbT-I cells for malarial antigen was demonstrated by their lack of response to intravenous (i.v.) infection with herpes simplex virus type I (HSV-1), an infection that efficiently stimulated viral glycoprotein B-specific transgenic T cells (gBT-I cells) in the same mice (Figure S6).
To more precisely determine where PbT-I cells were activated during the primary response to blood-stage PbA infection, B6 mice were injected with CFSE-labeled PbT-I cells one day before i.v. infection with blood-stage PbA, then various tissues were harvested 2 days later to examine expression of the early activation marker CD69 on PbT-I cells (Figure 2C, D). This showed that blood-stage infection caused T cell activation in the spleen, although some CD69 up-regulation was observed in liver-draining lymph nodes (portal and celiac LNs). Other lymph nodes showed no evidence of T cell activation.
To test whether PbT-I cells induced by blood-stage infection made cytokines and were able to degranulate, as required for lytic activity, mice were adoptively transferred with small numbers of GFP-expressing PbT-I cells and infected i.v. with blood-stage PbA. On day 8 post-infection, PbT-I cells were recovered from the spleen and briefly restimulated with anti-CD3 mAb to test for production of IFNγ, TNFα and CD107a, the latter of which is a surrogate marker for degranulation (Figure S7). This revealed that most PbT-I cells were able to produce both cytokines and degranulate.
As CD8+ T cells have been implicated in the pathology of ECM, we asked whether transfer of PbT-I cells into B6 mice could accelerate this disease. B6 mice were injected with a high (2×106) or low (2×104) number of PbT-I cells or a high number of a herpes simplex virus-specific gBT-I cells, then infected with blood-stage PbA and monitored for disease (Figure 3A). This showed that PbT-I cells significantly accelerated disease onset, though only by about one day. ECM was accompanied by infiltration of PbT-I cells and endogenous CD8+ T cells, but not gBT-I cells into the brain of infected mice on days 5–6 post-infection (Figure 3B and Figure S8).
To determine whether PbT-I cells could themselves cause ECM, endogenous CD8+ T cells were depleted from mice with anti-CD8 mAb and 7 days later replaced with PbT-I cells, control gBT-I cells or no T cells. One day later, these mice were infected with blood-stage PbA and examined for ECM onset. All mice given PbT-I cells developed ECM, while very few other CD8-depleted mice developed disease (Figure 4). Onset of ECM in a small fraction of the latter was likely due to incomplete depletion of endogenous CD8+ T cells in some mice. This could not be avoided because the dose of depleting anti-CD8 antibody had to be sufficient to deplete virtually all endogenous CD8+ T cells while leaving little antibody to persist until adoptively transfer of PbT-I cells a week later (otherwise remaining anti-CD8 mAb would have depleted these PbT-I cells). H&E staining of the brains of mice that received PbT-I cells showed typical features of CM, such as haemorrhages and intravascular accumulation of RBC and leukocytes (Figure S9). These data clearly showed that PbT-I cells were able to cause ECM.
As the precise specificity of PbT-I cells was unknown, we determined whether they recognized other species of Plasmodium. CFSE-labeled PbT-I cells were adoptively transferred into B6 mice that were then infected with blood-stage P. chabaudi AS; 6 or 7 days later proliferation of PbT-I cells was assessed in the spleen (Figure 5). This showed that PbT-I cells could proliferate in response to blood-stage P. chabaudi AS. In a similar set of experiments, PbT-I cells were also shown to respond to blood-stage infection with P. yoelii XNL (Figure S10). These findings indicated that PbT-I cells have specificity for multiple Plasmodium species that cause rodent malaria.
While our PbT-I line was generated to blood-stage parasite infection, a proportion of antigens expressed in the blood stage are also expressed by sporozoites and during the liver-stage of infection [23]. To address whether sporozoites could stimulate PbT-I cells, we adoptively transferred CFSE-labeled PbT-I cells into B6 mice and then injected them i.v. with radiation-attenuated PbA sporozoites (RAS). On day 4 post-infection, proliferating PbT-I cells were detected in the spleen indicating their capacity to respond to sporozoites (Figure 6). Additional mice examined on day 7 did not progress to patency, indicating that day 4 responses were induced by sporozoites and not by break-through blood-stage parasites. Eight days after infection, PbT-I cells harvested from the spleen produced IFNγ, TNFα and CD107a (Figure S11), indicating their development of effector function.
A recent report suggested that the efficiency of intravenous vaccination with irradiated sporozoites relative to subcutaneous vaccination may be because the former route allows more parasites to reach the liver for priming of protective immunity [19]. To test whether priming by irradiate parasites occurred in the liver, we injected irradiated sporozoites intravenously and then 1–4 days later examined the activation (CD69 expression) (Figure 7A, C) and proliferation (Figure 7A, B) of PbT-I cells in the liver and various lymphoid tissues including the spleen and lymph nodes. Upregulation of CD69 was seen as early as one day after infection and was primarily detected in the spleen, with some expression also seen in the liver draining lymph nodes (celiac LN, portal LN and the 1st mesenteric LN) [27]. Proliferation closely followed on day 2, almost entirely in the spleen. These data suggested that PbT-I cells responded to sporozoites by CD69 upregulation and extensive initial proliferation in the spleen and to a lesser extent in the liver-draining lymph nodes, but not in the liver nor other lymph nodes. Divided cells were only evident in the liver once they were present in the blood and had divided extensively, suggesting initiation of proliferation elsewhere, most likely in the spleen.
CD8α+ DC are critical for generating immunity to blood-stage infection [28], [29] and recently the human DC subset equivalent, BDCA3+ DC, have been implicated in severe malaria in humans [30], [31]. To address whether CD8α+ DC also participated in responses to sporozoites, we examined proliferation of PbT-I cells in Batf3-/- mice, which lack this DC subset (Figure 8). The poor proliferation in Batf3-/- mice compared to wild-type mice revealed that this response was dependent on CD8α+ dendritic cells.
It has been reported that blood-stage infection can impair immunity to liver-stage antigens [24], though this is disputed by evidence that there is an equivalent response by liver-stage-specific transgenic T cells to sporozoites in the presence or absence of a subsequent blood-stage infection [32]. To resolve this issue with respect to CD8+ T cell-mediated immunity, we examined the expansion of PbT-I cells after exposing mice to live sporozoites (which will infect the liver then generate blood-stage infection), or irradiated sporozoites alone or followed by blood-stage (iRBC) infection 2 days later, mimicking the time for blood-stage egress after live sporozoite infection (Figure 9). Our results clearly showed that naïve PbT-I cells proliferated to reach greater numbers if additionally exposed to blood-stage infection, indicating that T cells with cross-stage specificity can show cumulative expansion to the liver and bloods stages.
Since sporozoite antigen has been shown to persist in other models [33], and we could demonstrate some proliferation of PbT-I cells transferred 2 days but not 7 days after injection of irradiated sporozoites (Figure S12A, B), indicating at least short-term persistence of the PbT-I antigen, it remained possible that augmented proliferation of PbT-I cells due to blood-stage infection might simply relate to additional inflammatory effects, rather than provision of antigen. To test whether inflammation alone could boost PbT-I expansion to irradiated sporozoites, 20 nmol of 1668 CpG oligonucleotide (CpG) was used as an inflammatory signal on day 2 and its effect on expansion of PbT-I cells examined (Figure S12C). CpG-mediated inflammation failed to induce a significant increase in PbT-I cell numbers in mice given irradiated sporozoites two days earlier, suggesting that antigen provided by blood-stage infection may be important for enhanced proliferation. This did not, however, formally excluding a role for inflammatory signals distinct from CpG that are associated with blood-stage infection.
Because only one parasitized hepatocyte needs to survive to deploy thousands of merozoites into the blood and seed blood-stage infection, it is very difficult to prevent malaria with vaccines directed at pre-erythrocytic stages. It follows that any vaccine targeting pre-erythrocytic stages of infection must generate sterile immunity to be effective. As the antigen recognized by PbT-I cells was expressed by sporozoites, we asked whether this antigen might represent a vaccine candidate capable of eliciting sterile hepatic immunity. To assess this, we asked whether PbT-I cells could provide protective immunity to liver-stage infection. First, we determined an infectious dose of sporozoites that would lead to just under 100% blood-stage infection in the absence of PbT-I cells (Figure 10A). From this we chose 520 sporozoites as our infectious dose. To test the protective capacity of PbT-I cells, these cells or control virus-specific gBT-I cells were activated in vitro and then 7×106 cells adoptively transferred into naïve B6 mice that were subsequently challenged with 520 live sporozoites (Figure 10B). By monitoring these mice for blood parasitemia, we showed that PbT-I cells, but not gBT-I cells, could prevent progression to blood-stage infection, protecting mice from infection. This indicated that the antigen recognized by PbT-I cells has the potential to generate sterilizing immunity to liver-stage infection.
To identify the antigenic determinant recognized by PbT-I cells, we used an octamer combinatorial peptide library scan [34] to identify amino acid residues important for PbT-I activation as measured by MIP1β production (data not shown). These residues were then used to generate a octamer motif (x-x-x-(CD)-(WF)-N-x-(LMIV); where x is any amino acid and residues in brackets are valid for that position) to search the genomes of the three rodent malaria species for which PbT-I cells showed reactivity. 151 peptides fitting this motif were then examined for their capacity to stimulate PbT-I cells either by CD69 expression or MIP1β production (data not shown). Six peptide sequences caused some T cell activation but only one of these (NCYDFNNI (NCY)) was found to act as a target antigen for endogenous killer T cells generated in normal B6 mice infected with PbA (Figure 11A and data not shown). This sequence also induced IFNγ production from endogenous T cells (Figure 11B) and PbT-I T cells (Figure 11C) responding to blood-stage infection. Note that tetramers made with Kb containing NCY were able to stain PbT-I cells, confirming the Kb-restriction of this specificity (data not shown). The NCY peptide was derived from a protein of 745 amino acids (PBANKA_071450), which is now our leading candidate for the antigen responsible for priming PbT-I cells.
Here we characterize a new TCR transgenic mouse that produces CD8+ T cells specific for both the blood and liver stages of rodent malaria. PbT-I cells responded in vivo to the blood-stage of three different rodent Plasmodium species, PbA, P. yoelii XNL and P. chabaudi AS. In addition, PbT-I cells responded to mosquito-derived sporozoites of PbA and were able to provide protection against sporozoite infection. It remains to be tested as to whether PbT-I cells also recognize sporozoites from the other rodent Plasmodium species, but it seems likely that this will be the case given their blood-stage cross-reactivity. Recognition of blood-stage parasites as well as mosquito-derived sporozoites, and the ability to protect against liver-stage infection, suggests that the protein recognized by PbT-I cells is widely expressed throughout the parasite life cycle and is potentially well conserved. Identification of NCYDFNNI as a peptide recognized by PbT-I cells and by endogenous PbA-induced T cells suggests the protein encoded by PBANKA_071450, which is of unknown function and undefined expression pattern, may be the source of the PbT-I epitope. Construction of parasites deficient in this epitope will be required for formal proof.
It is notable that while the source protein is encoded in the genomes of PbA and P. chabaudi, the ortholog appears severely truncated in P. yoelii and consequently lacks the region containing NCYDFNNI found in other species. As PbT-1 cells were able to respond to P. yoelii, the authentic antigen must be present in this species. Whether this invalidates the gene product of PBANKA_071450 as the authentic PbT-I antigen, or is explained by sequencing error within the P. yoelii genome, or has some other basis remains to be established. Whatever the case, the NCYDFNNI epitope is clearly recognized by PbT-I cells and can be used to stimulate these transgenic T cells as well as endogenous T cells specific for PbA.
Evidence that immunization with live blood-stage parasites can protect against the liver-stage infection [25], suggests that multi-stage antigens like that recognized by PbT-I cells can be protective. Our study extends this concept by indicating that CD8+ T cells of a single specificity for a blood-stage antigen can protect against liver-stage infection when the antigen is also expressed during the liver stage. It has been reported that blood-stage infection can impair immunity to liver-stage antigens [24], though this is disputed by the above study, which uses blood-stage infection to induce anti-sporozoite immunity [25] and by another study that shows an equivalent response by liver-stage-specific transgenic T cells to sporozoites in the presence or absence of a subsequent blood-stage infection [32]. The availability of PbT-I cells will give us the opportunity to examine this relationship when the relevant antigen is expressed during both blood- and liver-stages and to determine how antigens presented during the blood-stage might influence the effector function of T cells capable of recognizing liver-stage antigens. Clearly, in our experiments, exposure of cells primed to liver-stage parasites did not impair their capacity to respond to blood-stage parasites, but increased the expansion of PbT-I cells. This raises the possibility that CD8+ T cells specific for antigens expressed in both stages of infection may have a selective advantage for expansion over single stage specific T cells.
The broad cross-reactivity of this TCR transgenic line means that it is suited to exploring the role of CD8+ T cells in several rodent malaria models. For blood-stage infection, this is most relevant to PbA, where ECM is dependent on CD8+ T cells. However, CD8+ T cells have been implicated in protective immunity to blood-stage infection by P. yoelii 17XL [35], raising the possibility that this protective process could be explored using PbT-I cells. These transgenic T cells should also be highly relevant for analysis of liver-stage immunity, as CD8+ T cells are critical for protection at this stage of infection [15], [36]. Here we used PbT-I cells to investigate the site of priming and T cell expansion after intravenous administration of irradiated sporozoites. This study was prompted by the implication that the effectiveness of this route of immunization was related to its capacity to prime in the liver [19]. Our analysis revealed that T cells showed signs of activation in the spleen and in the liver draining lymph nodes, but not in the liver itself, and subsequent examination of T cell proliferation showed that most PbT-I T cell proliferation occurred in the spleen. While our study does not exclude a role for the liver in tailoring the response, it suggests that at least the initial priming steps are unlikely to occur in this site. Thus, efficient priming via this route most likely derives from the large load of irradiated sporozoites deposited in the spleen after intravenous administration and the high frequency of T cells found in this organ. This contrasts infection by mosquito bite, which favors priming within skin draining lymph nodes [13], probably as a consequence of local deposition of sporozoites within the dermis of the skin. Our findings suggest that the spleen is the main site for priming sporozoite specific T cells after intravenous administration of parasites, but they do not formally exclude the liver draining lymph nodes or the liver as important sites of activation for protective immunity.
Initiation of PbT-I proliferation in the spleen in response to intravenous injection of irradiated PbA sporozoites also demonstrated that the sporozoites themselves expressed the antigen recognized by PbT-I cells and that conversion to later liver stages of development was not necessary to provide antigen capable of stimulating these T cells. Furthermore, it showed that the same DC subset as required for priming CD8+ T cell immunity to blood-stage infection, i.e the CD8α+ DC [28], [29], was responsible for inducing CD8+ T cell responses to the liver-stage parasites. Extraction of putative CD8α+ DC from the liver 6 days after sporozoite infection also suggested that these DC might contribute to antigen presentation in the liver at late time points after infection [37], though this idea should be taken with caution as CD8 T cells can express CD11c when activated and can be easily mistaken for DC. This common use of CD8α+ DC probably reflects their dominant capacity to cross-present antigens [38]. The ability of PbT-I cells to protect against infection by PbA sporozoites is encouraging because sterilizing immunity requires destruction of all infected hepatocytes. Our experiments used 7×106 activated PbT-I cells to demonstrate protection, which is a relatively high number of cells but certainly achievable by vaccination. Identification of the antigen recognized by this TCR transgenic line should allow development of vaccination strategies to test the protective power of this potentially conserved antigen expressed in multiple stages of the life cycle. This approach has the potential to be highly effective since both stages of infection are shown to boost responses by CD8+ T cells with such multi-stage specificity. One concern with this type of multi-stage antigen, however, is that priming of T cells by sporozoites may enhance the potential for development of ECM mediated by the same cells during the blood-stage of the infection. While directly relevant for PbA infection where ECM is commonplace, this might not be of relevance to infection models where ECM is not seen e.g. P. yoelii XNL infection. Given the strongly argued lack of adaptive immune involvement in human cerebral malaria, this concern may also be irrelevant for human vaccination approaches. However, caution should be adopted here since our understanding of pathology in human cerebral malaria is still somewhat limited.
In conclusion, the PbT-I TCR transgenic line represents a versatile tool for studying CD8+ T cell immunity to a multitude of rodent Plasmodium species during both the liver- and blood-stages of infection. The current study highlights the spleen as a major organ of priming for intravenously-introduced blood- or liver-stage parasites and suggests that T cells with specificity for antigens expressed in both stages may contribute to pathology or protection, depending on the stage of life cycle.
All procedures were performed in strict accordance with the recommendations of the Australian code of practice for the care and use of animals for scientific purposes. The protocols were approved by the Melbourne Health Research Animal Ethics Committee, University of Melbourne (ethic project IDs: 0810527, 0811055, 1112347, 0911527).
C57BL/6 (B6) mice, B6.Ly5.1 mice, MHC I-/- mice, Kb-/- mice, Batf3-/- mice and the transgenic strains gBT-I [39] and PbT-I were used between 6-12 weeks and were bred and maintained at the Department of Microbiology and Immunology, The University of Melbourne. Batf3-/- mice used in this study had been backcrossed 10 generations to B6. Animals used for the generation of the sporozoites were 4–5 week old male Swiss Webster mice were purchased from the Monash Animal Services (Melbourne, Victoria, Australia) and maintained at the School of Botany, The University of Melbourne, Australia.
Anopheles stephensi mosquitoes (strain STE2/MRA-128 from The Malaria Research and Reference Reagent Resource Center) were reared and infected with PbA as described [40]. Sporozoites were dissected from mosquito salivary glands [41], resuspended in cold PBS, irradiated with 20,000 rads using a gamma 60Co source, and administered to mice i.v. The rodent malaria lines PbA clone 15cy1, P. chabaudi AS and P. yoelii XNL were used in this study.
Transgenic PbT-I mice were generated using the V(D)J segments of the TCRα- and β-genes of a CD8+ T cell hybridoma (termed B4) specific for an unidentified blood-stage PbA antigen. This hybridoma was derived from T cells extracted from the spleen of a B6 mouse at day 7 after infection with PbA. 3×106 splenocytes from a mouse previously infected with PbA were co-cultured with 5×105 conventional DC (extracted from the spleen of FMS-like tyrosine kinase 3 receptor ligand (Flt3-L) treated B6 mice) that were pre-loaded for 2 hours with 2×106 PbA schizont lysate as previously described [42] in complete RPMI at 6.5% CO2, 37°C. One week later, cultured cells were re-stimulated for a week with fresh DC and PbA schizont lysate. To generate PbA-specific hybridomas, in vitro cultured cells were then fused with the BWZ36.GFP fusion partner and exposed to drug selection [43]. This led to isolation of the Kb-restricted B4 hybridoma (Figure S1) from which PbT-I T cell receptor genes were derived.
TCR Vα usage was defined using 5′ RACE PCR on cDNA converted from the RNA of the B4 hybridoma. Sequencing analysis revealed that the TCR α-chain consisted of AV8S6 and Jα17 and Cα2 gene segments. The TCR α region was amplified by PCR from the cDNA of the B4 hybridoma using the forward primer GGATCCAGTGTCATTTCTTCCCT containing a BamHI recognition sequence at the 5′ end, designed to bind the 5′ UTR region of AV8S6, and the reverse primer CAGATCTCAACTGGACCACAG containing a BglII recognition sequence at the 5′ end, specific for the Cα region. The AV8S6-Jα17-Cα2 segment was cloned into the BamHI site of the pES4 cDNA expression vector, comprising the Ig-H chain enhancer, the H2-Kb promoter and the polyadenylation signal sequence of the human β-globulin gene [44]. To prepare the α-chain transgenic construct for microinjection, the pES4-VJC construct was digested with the restriction enzymes ClaI and NotI, and the digested mix was subjected to agarose gel electrophoresis. The ∼5.6 kb transgenic insert containing the VJC sequence, the promoter and enhancer sequences was excised from the gel, purified and quantitated for microinjection.
TCR Vβ usage was defined by PCR on cDNA converted from the RNA of the B4 hybridoma using the forward primer CCTGCCTCGAGCCAACTATGGG specific for the Vβ10 gene and the reverse primer CCAGAAGGTAGCAGAGACCC specific for the Cβ gene. Sequencing analysis revealed that the TCR β-chain consisted of Vβ10 (BV10S1A1), Dβ2 and Jβ2.2. The TCR β-chain was amplified by PCR from the genomic DNA of the B4 hybridoma using the forward primer GATCGATGTCCTAGGCCAGGAGATATGA specific for the Vβ10, incorporating a ClaI restriction enzyme site at the 5′ end, and the reverse primer GATCGATAAGCTCAGTCCAAGA specific for Jβ2.2 and incorporating a ClaI site at the 5′end. The Vβ10 (BV10S1A1), Dβ2 and Jβ2.2 segment was cloned into the ClaI site of the p3A9CβTCR gDNA expression vector, comprising the TCR β-chain enhancer, the 2B4-derived 5′ region and leader sequence and the B3-derived promoter and coding regions [45]. To prepare the construct for microinjection, the p3A9CβTCR VDJ construct was digested with the restriction enzymes ApaI and NotI, and the digested mix was subjected to agarose gel electrophoresis. The larger fragment (∼11 kb) transgenic insert containing the VDJ sequence, Cβ sequence and the promoter and enhancer sequences was excised from the gel, purified and quantitated for microinjection.
Cells were labeled with monoclonal antibodies specific for CD8 (53-6.7), CD4 (RM 4-5), Thy1.2 (30-H12), CD45.1 (A20), Vα8.3 (B21.14), Vβ10 (B21.5) or CD69 (H1.2F3). Dead cells were excluded by propidium iodide staining. Cells were analyzed by flow cytometry on a FACsCanto or Fortessa (BD Biosciences), using the Flowjo software (Tree Star Inc.).
Unless otherwise stated, mice were infected i.v. with 106 PbA infected RBC (iRBC) in 0.2 ml of Hank's balanced salt solution (HBSS). In some experiments, mice were infected i.v. with 105 P.chabaudi iRBC or i.v. with 104 P. yoelii iRBC or with 300, 520, 900, 5×104 or 105 PbA sporozoites as stated in the figure legends. Mice infected with 104 PbA infected RBC were injected i.p. with 0.4 mg chloroquine dissolved in water on days 6 and 7, before being euthanized for analysis on day 8 post-infection.
CD8+ T cells were negatively enriched from the spleens and lymph nodes of transgenic mice and labelled with CFSE as described [46]. Purified cells were injected i.v. in 0.2 ml HBSS. To deplete endogenous CD8+ T cells before adoptive transfer, B6 mice were injected i.p. with 100 µg of anti-CD8 antibody (clone 2.43) 7 days prior to the transfer of PbT-I or gBT-I cells.
1–2×106 CFSE-labelled Ly5.1+ PbT-I cells were adoptively transferred into Ly5.2+ B6 mice a day before mice were infected with blood-stage PbA, P. chabaudi, or P. yoelii or with PbA sporozoites. In other experiments, 5×104 or 1×106 uGFP PbT-I cells labelled with CellTracker Violet stain (Invitrogen) were adoptively transferred into Ly5.2+ B6 or Batf3-/- mice a day before infection with irradiated sporozoites, or 3 days before infection with PbA iRBC. Spleens and other organs were harvested on various days post-infection for the analysis of PbT-I proliferation by flow cytometry.
Dendritic cells were purified from the spleens of mice as previously described (28). Briefly, spleens were finely minced and digested in 1 mg/ml collagenase 3 (Worthington) and 20 µg/ml DNAse I (Roche) for 20 min at room temperature. After removing undigested fragments by filtering through a 70 µm mesh, cells were resuspended in 5 ml 1.077 g/cm3 nycodenz medium (Nycomed Pharma AS, Oslo, Norway), layered over 5 ml nycodenz medium and centrifuged at 1700×g at 4°C for 12 min. The light density fraction was collected and DC were negatively enriched by incubation with a cocktail of rat monoclonal anti-CD3 (clone KT3-1.1), anti-Thy-1 (clone T24/31.7), anti-Gr1 (clone RB68C5), anti-CD45R (clone RA36B2) and anti-erythrocyte (clone TER119) antibodies followed by immunomagnetic bead depletion using BioMag goat anti-rat IgG beads (Qiagen).
5×104 DC extracted from the spleens of naive WT, MHC-I-deficient or Kb-deficient mice and resuspended in complete DMEM medium supplemented with 10% foetal calf serum (FCS) were cultured for 1 h with titrated amounts of lysed whole blood containing mixed stages of PbA parasites before adding 5×104 B4 hybridoma cells. After culture for 40 h at 37°C in 6.5% CO2 supernatants were collected and concentrations of IL-2 were assessed using the Mouse IL-2 ELISA Ready-Set-Go kit (eBiosciences) following manufacturer's instructions.
PbA mixed blood-stages and schizont enriched parasite lysate were prepared as previously described [42]. Conventional DC isolated from the spleen of FMS-like tyrosine kinase 3 receptor ligand (Flt3-L) treated Ly5.2+ B6 mice [42] were incubated with titrated amounts of lysate from either the mixed blood-stages or the schizont-enriched parasites for 2 hours before the addition of CFSE-labeled Ly5.1+ PbT-I cells. After 60 hours of incubation at 6.5% CO2, 37°C, cells were harvested for analysis by flow cytometry.
To detect degranulation and the production of cytokines IFNγ and TNFα from antigen specific cells, splenocytes from mice (either normal B6 mice or those adoptively transferred with PbT-I) infected for 7 days with irradiated sporozoites or 7–8 days with blood-stage PbA were restimulated by 5 µg/ml plate-bound anti-CD3 or 1 µg/ml peptide for 5 hours at 37°C in the presence of 10 µg/ml brefeldin A, monensin and anti-CD107 antibody (clone eBio1D4B). Cells were then surface labeled with antibodies and intracellular cytokine staining was performed to detect intracellular IFNγ and TNFα using Cytofix/Cytoperm Fixation and Permeabilization Solution (BD) according to the manufacturer's instructions. Results were represented in Venn diagrams using the online tool at www.venndiagram.tk.
PbT-I or gBT-I isolated from the spleen and lymph nodes were stimulated with media containing 10% FCS, 10 U/ml IL-2 and 5 µg/ml anti-CD28 in 75 cm2 tissue culture flasks pre-coated with 10 µg/ml anti-CD3 (clone 2c11), anti-CD8 (clone 53-6.7) and anti-CD11a (clone 121/7.7). 40 hours later, cell cultures were divided into two equal volumes and given an equivalent volume of fresh medium before culturing for 24 hours. Cells were then harvested and centrifuged over lymphocyte separation media to remove dead cells. In vitro activated cells generated using this method were routinely >90% pure.
Mice were perfused intracardially with 10 ml PBS prior to harvesting of the brain. Brains were cut into fine fragments, washed once with media and digested with collagenase/DNAse (1 mg/ml collagenase III (Roche); 20 µg/ml DNAse I, (Worthington) for 1 hour at room temperature with rotation. Samples were filtered through 75 µm nylon mesh to remove undigested fragments and then centrifuged once at 596 g, 5 minutes at 4°C. The pellet was resuspended in 7 ml 33% Percoll diluted in media, and centrifuged at 400 g for 20 minutes at room temperature with low brake. The supernatant was discarded and the pellet containing RBC was incubated with 500 µl RBC lysis buffer for 2 minutes on ice. Cells were washed twice with FACS buffer followed by surface staining with various antibodies.
Mice infected with blood-stage PbA were monitored daily for the development of ECM. Mice were considered to have ECM when showing signs of neurological symptoms such as ataxia and paralysis, evaluated as the inability of mice to self-right.
Brains were fixed in 4% paraformaldehyde followed by 70% ethanol overnight and then stained by H&E.
An octamer combinatorial peptide library in positional scanning format [34] was synthesized (Pepscan Presto, Netherlands). For combinatorial peptide library screening, splenocytes from transgenic PbT-I mice were purified, washed and rested overnight in RPMI 1640 containing 100 U/ml penicillin, 100 µg/ml streptomycin, 2 mM L-glutamine and 2% heat inactivated fetal calf serum (all Life Technologies). In 96-well U-bottom plates, 6×104 splenocytes target cells were incubated with 160 library mixtures (at 100 µM) in duplicate for two hours at 37°C. Following peptide pulsing, 3×104 PbT-I splenocytes were added and the assay was incubated overnight at 37°C. The supernatant was then harvested and assayed for MIP-1β by ELISA according to the manufacturer's instructions (R&D Systems).
5×105 GFP-expressing PbT-I lymph node and spleen cells together with 105 B6 spleen cells and peptide (titrated in 10-fold steps from 5–5000 pM) were pelleted together in a 96-well U-bottom plate and incubated for 3 hours at 37°C (6.5% CO2). Cells were then stained with antibodies specific for CD8 and CD69 and the proportion of CD69+CD8+GFP+ cells determined.
To detect peptide-specific lytic activity in vivo, mice were infected for 7 days with 106 blood-stage PbA and cured by chloroquine treatment from day 4–6 before adoptive transfer of target cell populations. In vivo cytotoxicity was performed essentially as described [42], with the modification that target cells were a mixture of CFSElo B6 spleen cells, DsRed-expressing splenocytes and GFP-expressing splenocytes, the latter two populations coated with test peptides at 1 µg/ml. Equal numbers of cells were combined and 2.4×107 cells were injected into host mice and 18 h later spleen cells were harvest for flow cytometric assessment of lysis 18 h later within the spleen.
|
10.1371/journal.pcbi.1005686 | Upregulation of an inward rectifying K+ channel can rescue slow Ca2+ oscillations in K(ATP) channel deficient pancreatic islets | Plasma insulin oscillations are known to have physiological importance in the regulation of blood glucose. In insulin-secreting β-cells of pancreatic islets, K(ATP) channels play a key role in regulating glucose-dependent insulin secretion. In addition, they convey oscillations in cellular metabolism to the membrane by sensing adenine nucleotides, and are thus instrumental in mediating pulsatile insulin secretion. Blocking K(ATP) channels pharmacologically depolarizes the β-cell plasma membrane and terminates islet oscillations. Surprisingly, when K(ATP) channels are genetically knocked out, oscillations in islet activity persist, and relatively normal blood glucose levels are maintained. Compensation must therefore occur to overcome the loss of K(ATP) channels in K(ATP) knockout mice. In a companion study, we demonstrated a substantial increase in Kir2.1 protein occurs in β-cells lacking K(ATP) because of SUR1 deletion. In this report, we demonstrate that β-cells of SUR1 null islets have an upregulated inward rectifying K+ current that helps to compensate for the loss of K(ATP) channels. This current is likely due to the increased expression of Kir2.1 channels. We used mathematical modeling to determine whether an ionic current having the biophysical characteristics of Kir2.1 is capable of rescuing oscillations that are similar in period to those of wild-type islets. By experimentally testing a key model prediction we suggest that Kir2.1 current upregulation is a likely mechanism for rescuing the oscillations seen in islets from mice deficient in K(ATP) channels.
| Pulsatile insulin secretion is important for the proper regulation of blood glucose, and disruption of this pulsatility is a hallmark of type II diabetes. An ion channel was discovered more than three decades ago that conveys the metabolic state of insulin-secreting β-cells to the plasma membrane because it is blocked by ATP and opened by ADP, and thereby controls the activity of these electrically-excitable cells on a rapid time scale according to the prevailing blood glucose level. In addition to setting the appropriate level of insulin secretion, K(ATP) channels play a key role in generating the oscillations in cellular activity that underlie insulin pulsatility. It is therefore surprising that oscillations in activity persist in islets in which the K(ATP) channels are genetically knocked out. In this combined modeling and experimental study, we demonstrate that the role played by K(ATP) current in wild-type β-cells can be taken over by an inward-rectifying K+ current which, we show here, is upregulated in β-cells from SUR1 knockout mice. This result helps to resolve a mystery in the field that has remained elusive for more than a decade, since the first studies showing oscillations in SUR1-/- islets.
| Insulin is secreted from pancreatic islet β-cells in response to elevated blood glucose. Islet activity is oscillatory, with periods ranging from tens of seconds to several minutes, and this is reflected in the reported periods of pulsatile insulin secretion [1–4]. Plasma insulin oscillations play a physiological role in blood glucose regulation [5–8]. A recent study showed that the action of insulin on the liver to lower plasma glucose is more profound when insulin is delivered to the liver in a pulsatile fashion [9], and earlier studies showed that plasma insulin oscillations are disrupted in type II diabetics and their near relatives [10–12].
At stimulatory levels of glucose β-cells exhibit electrical bursting, and Ca2+ that enters the cells during each burst evokes a pulse of insulin secretion [7,13,14]. Several mechanisms have been proposed to explain this bursting electrical activity [15–18]. A recent mathematical model that combines two of these mechanisms can reproduce bursting having a wide range of periods, as seen in experimental studies [19]. One mechanism produces fast oscillations, while the other produces slow oscillations and both can oscillate independently, prompting the name Dual Oscillator Model (DOM). In the DOM, the fast component of bursting results from the negative feedback of Ca2+ on the membrane potential via Ca2+-activated K+ channels and, indirectly, via K(ATP) channel activation. The slow component, in contrast, is due to oscillations in glycolysis that occur as the result of actions of the allosteric enzyme phosphofructokinase (PFK)[20,21]. The subsequent oscillatory ATP production acts through ATP-sensitive K+ channels (K(ATP) channels) to produce oscillations in K(ATP) current, which turns the bursts of electrical activity on and off [22,23].
K(ATP) channels play a crucial role coupling cell metabolism to membrane potential. These channels are comprised of four inwardly rectifying K+ channel subunits (Kir6.2) and four sulfonylurea receptor subunits (SUR1) arranged in an octomeric array (for review see [24]). A mutation in the genes coding for either subunit prevents K(ATP) channels from being trafficked normally to the plasma membrane or alters their sensitivity to adenine nucleotides, leading to persistent hyperinsulinemic hypoglycemia of infancy (PHHI) in humans, a condition characterized by high insulin secretion that occurs even when blood glucose is low [25–27]. High secretion results from the permanent depolarization of the β-cell membrane that is due to the lack of normally hyperpolarizing K(ATP) current. Surprisingly, in SUR1 homozygous knockout mice (SUR1-/- mice), lacking K(ATP) channels, islets typically still exhibit electrical bursting (although the glucose sensitivity of bursting in these islets is largely abrogated), and blood glucose levels are relatively normal unless the animals are metabolically stressed [28,29]. Similarly, islets from Kir6.2 knockout mice exhibit slow Ca2+ oscillations, similar to those observed in wild-type islets which are known to be due to bursting electrical activity [30]. In these mice, compensation must therefore occur to overcome the loss of the large hyperpolarizing K(ATP) current. Indeed, when the K(ATP) channels of wild type islets are acutely blocked by sulfonylurea drugs, β-cells spike continuously from a sustained depolarized level [31–33]. We hypothesized that such compensation could be achieved through the upregulation of another hyperpolarizing K+ channel that impersonates K(ATP) channels in sensing cellular metabolism [34]. In a companion study (Vadrevu et al, manuscript in preparation), we demonstrated that the upregulation of Kir2.1 channel protein in islets from SUR1-/- mice (KO islets) could mediate this compensation. In the current report, we demonstrate that SUR1 KO islets exhibit sustained Ca2+ oscillations at stimulatory levels of glucose, and that the amount of inward rectifying K+ current is increased in these K(ATP) channel KO cells. Using mathematical modeling, we explored the functional role of this current on the electrical activity of islet β-cells when K(ATP) channels are absent. In particular, we investigated whether this inward-rectifying K+ current has the ability to rescue normal electrical bursting pattern in β-cells of SUR1-/- mouse islets.
Kir2.1 channels conduct large inward currents at voltages below the K+ Nernst potential (VK) and smaller outward currents at voltages above VK. This diode-like property, or inward rectification, is caused by blockade of the channels by intracellular ions and polyamines when the cell membrane is depolarized [35–37]. Kir2.1 channels also contain consensus sites for phosphorylation by protein kinase A (PKA) and studies show that PKA potentiates Kir2.1 current [38–40]. One study shows that a phosphatase inhibitor can prevent rundown of the Kir2.1 current that is activated by PKA, which indicates activation of the channels is regulated by protein phosphorylation [41]. Since PKA activity is cAMP-dependent, changes in the cAMP concentration in the β-cell can in principle regulate Kir2.1 channel activity. Recent studies employing FRET-based sensors and TIRF microscopy showed that glucose induces cAMP oscillations in mouse β-cells [42,43], which may be accounted for by oscillations in metabolism [44]. It is therefore possible that, in KO cells, metabolic oscillations drive cAMP oscillations which in turn drive oscillations in Kir2.1 current, and this replaces oscillations in K(ATP) current as the mechanism for bursting electrical activity. We illustrate how this works with the model, and make predictions that are subsequently confirmed experimentally and thereby support the hypothesis that Kir2.1 channel upregulation is a feasible mechanism which can rescue electrical bursting in SUR1-/- mouse islets lacking K(ATP) channels.
The animal protocol used was in accordance with the guidelines of the University of Michigan Institutional Animal Care and Use Committee (IACUC).
Pancreatic islets were isolated from 3–4 month old male Swiss-Webster mice as in [45]. Islets were hand picked into fresh Kreb’s solution and then transferred to culture dishes containing RPMI-1640 supplemented with 10% FBS, glutamine and penicillin-streptomycin. Islets were cultured overnight at 37°C in an incubator. Electrophysiological recordings were made from islets cultured for 72 hours or less.
Patch electrodes were pulled (P-97, Sutter Instrument Co., Novato, CA) from borosilicate glass capillaries (Warner Instrument Inc., Hamden, CT) and had resistances of 8–10 M-ohm when filled with an internal buffer containing (in mM): 28.4 K2SO4, 63.7 KCl, 11.8 NaCl, 1 MgCl2, 20.8 HEPES and 0.5 EGTA at pH7.2. The electrodes were then backfilled with the same solution but containing amphotericin B at 0.3 mg/ml to allow membrane perforation. Islets were transferred from culture dishes into a 0.5 ml recording chamber held at 32–34°C. Islets were visualized using an inverted epifluorescence microscope (Olympus IX50, Tokyo, Japan). Pipette seals obtained were > 1 G-ohms. Recordings were made using an extracellular solution containing (in mM): 135 NaCl, 2.5 CaCl2, 4.8 KCl, 1.2 MgCl2, 20 HEPES, and 11.1 glucose.
After the establishment of a perforated patch, cells were voltage-clamped to a holding potential of -60 mV, and a 2-second voltage ramp from -120 to -50 mV was applied, as in [32]. Evoked currents were digitized at 10 kHz after filtering at 2.9 kHz. The protocols were generated using Patchmaster software (v2x32; HEKA Instruments).
Pancreatic islets were cultured overnight in RPMI medium containing 5 mM glucose and on the day of experiments were transferred to fresh media containing 2.5 μM Fura-PE2-AM for 30 min. Following incubation, islets were loaded into a glass-bottomed chamber mounted onto the microscope stage. The chamber was perfused at 0.3 mL/min with 11 mM glucose solution and the ambient temperature was maintained at 33°C using inline solution and chamber heaters (Warner Instruments). Excitation was provided by a TILL Polychrome V monochromator (TILL Scientific, Germany) with light output set to 10% maximum. Excitation (x) or emission (m) filters (ET type; Chroma Technology, Bellows Falls, VT) were used in combination with an FF444/521/608-Di01 dichroic (Semrock, Lake Forest, IL) as follows: Fura-2, 340/10x and 380/10x, 535/30m (R340x/380x – 535m); A single region of interest was used to quantify the average response of each islet using MetaMorph software (Molecular Devices). In one set of experiments, after three oscillations were recorded, the solution was switched to a solution containing 11 mM glucose with thapsigargin (1 μM). In another set of experiments, the solution was switched to one containing 11 mM glucose and 8-Bromoadenosine 3’,5’-cyclic monophosphate (8-Br-cAMP) (50 μM).
We used an 8-variable model consisting of ordinary differential equations, illustrated in Fig 1. This Dual Oscillator Model (DOM) has electrical, Ca2+, and metabolic components [23,46]. We focus our description on elements of the model that are most important for this study, but all equations and tables of parameter values are given in Supporting Information. (The computer codes, using the CVODE solver implemented in XPPAUT, can be downloaded as freeware from www.math.fsu.edu/~bertram/software/islet.) In the DOM, the fast oscillatory component is based on negative Ca2+ feedback onto the membrane potential through Ca2+-sensitive K+ current (IK(Ca)). This mechanism can drive fast bursting. The second oscillatory component is due to metabolic oscillations, which result from the activity of the allosteric enzyme phosphofructokinase (PFK). In the process of glycolysis, PFK catalyzes the phosphorylation of fructose 6-phosphate (F6P) to fructose 1,6-bisphosphate (FBP). The activity of PFK is increased by its product FBP, so that increased FBP increases the reaction rate and causes a sharp rise in FBP. This eventually depletes the substrate of the reaction, F6P, and turns off flux through PFK, resulting in a reduction in FBP. This allows the substrate, F6P, to recover and the cycle to restart. Oscillatory FBP levels in turn cause oscillations in pyruvate, the end product of glycolysis and the substrate for mitochondrial respiration. The oscillatory glycolytic input results in oscillatory levels of the nucleotide concentrations (ATP, ADP and AMP). The membrane potential is then affected through the action of ATP and ADP on K(ATP) channels, which can drive slow bursting in the model.
Equations for the dynamics of cAMP were recently added to an earlier version of the DOM [44] and it was shown that this version was capable of producing cAMP oscillations in model β-cells. We employed these equations, where the cAMP concentration is determined by the difference between its production by adenylyl cyclase (VAC) and degradation by phosphodiesterases (VPDE):
dcAMPdt=VAC−VPDE
(1)
where,
VAC=v¯AC(αAC+βACc3c3+KACca3)(βampKamp2AMPc2+Kamp2)
(2)
VPDE=v¯PDE(αPDE+βPDEc3c3+KPDEca3)cAMPcAMP+KPDEcamp
(3)
where c is the cytosolic free Ca2+ concentration, which stimulates both AC and PDE. Cytosolic AMP (AMPc) inhibits AC and thus the production of cAMP [47–49]. We modified the VAC equation from the original model to incorporate a higher-order dependence on AMP. In the model, slow cAMP oscillations are the result of AMP oscillations and the accompanying Ca2+ oscillations, which are both the product of glycolytic oscillations. The details of the cAMP dynamics are given in [44].
In the DOM, the rate of change of the membrane potential of a wild type β-cell is given by a conductance-based Hodgkin-Huxley type equation:
dVdt=−(IK+ICa+IK(Ca)+IK(ATP))/Cm
(4)
where, Cm is the membrane capacitance, IK is the delayed rectifier K+ current, ICa is a voltage-sensitive Ca2+ current, IK(Ca) is a Ca2+-sensitive K+ current and IK(ATP) is an ATP-sensitive K+ current. The rate of change of the free cytosolic Ca2+ concentration is:
dcdt=fcyt(−αICa−kpmcac⏞Jmem+kleak(cer−c)−kSERCAc⏞JER)
(5)
where terms labeled by Jmem and JER represent the Ca2+ flux across the plasma membrane and net flux out of the endoplasmic reticulum (ER), respectively. Here, fcyt is the fraction of free to total cytosolic Ca2+, α converts current to flux, kpmca is the Ca2+ pumping rate across the plasma membrane, kleak is the rate of the Ca2+ leak from the ER and kSERCA is the Ca2+ pumping rate into the ER. The ER Ca2+ concentration (cer) is also dynamic and given by:
dcerdt=−ferVcte(kleak(cer−c)−kSERCAc)
(6)
where fer is the ratio of the free Ca2+ in the ER and Vcte is the ratio of the volume of the cytosol to the volume of the ER compartment. The equation for the Ca2+-sensitive K+ current (IK(Ca)) is,
IK(Ca)=gK(Ca)ω(V−VK)
(7)
where, gK(Ca) is the maximal conductance of the current, and ω is the following Ca2+-dependent activation function,
ω=c2c2+Kc2
(8)
where Kc is the affinity constant.
In the KO-cells lacking K(ATP) channels there is no IK(ATP) present. In the model KO-cells, K(ATP) current is replaced by the following Kir2.1-mediated inward-rectifying K+ current:
IKir=gKirk∞c∞(V−VK).
(9)
Here gKir is the maximal Kir2.1 channel conductance, k∞ is the voltage-dependent block of the channel by polyamines which is the cause of the inward rectification, and c∞ is the cAMP-dependent activation of the channels. We use a Boltzmann function to describe k∞:
k∞=11+exp(V−Vkirskir)
(10)
where VKir is the half activation potential and sKir is the slope factor that determines the sensitivity to the voltage. The resulting voltage-dependent k∞ curve is shown in Fig 2A and is parameterized according to [50]. Kir2.1 current has both cAMP dependent and independent components [38], which are incorporated into the activation function c∞ as follows:
c∞=αcamp+βcampcAMP4cAMP4+Kcamp4
(11)
where αcamp is the cAMP independent component, and the cAMP dependency of the current is described by the second term. The c∞ function is illustrated in Fig 2B.
Ca2+ and membrane potential oscillations in SUR1-/- islets lacking functional K(ATP) channels were reported previously [28,51]. Our fura-2 Ca2+ measurements verified that slow cytosolic Ca2+ oscillations persisted in both wild-type (Fig 3A) and KO-islets (Fig 3B) perfused with 11 mM glucose. These data show that our SUR1-/- islets recapitulate the Ca2+ oscillations observed in [28,51].
We recently identified an increase in Kir2.1 channel protein in islets isolated from SUR1-/- mice (Vadrevu et al, manuscript in preparation). To verify the electrophysiological functionality of these channels in the β-cell membrane of KO islets, we measured current-voltage relations of wild-type and KO cells using the perforated patch clamp technique in peripheral islet β-cells. Fig 3C shows current recordings elicited by voltage ramp commands from -120 mV to -50 mV (see Materials and Methods) applied to wild-type islets (black) and K(ATP) KO islets (red).
In wild-type islets, the current-voltage relation is largely linear beyond about -110 mV (Fig 3C, black) (n = 6 islets from 4 mice), while in the KO cells the evoked current was more nonlinear, exhibiting inward rectification (Fig 3C, red). The strong inward rectification is likely due to current from the upregulated Kir2.1 inward-rectifying K+ channels that we report in a companion study (Vadrevu et al, manuscript in preparation), supporting a functional role for the upregulated Kir2.1 channel protein.
Fig 4 illustrates slow bursting produced by the model for the case of wild-type cells. The oscillations in the free Ca2+ concentration observed here (Fig 4A) result from the bursting electrical activity described earlier. The burst timing in this case is controlled by the slow glycolytic oscillations, which are reflected by the FBP time course as shown (Fig 4E). FBP oscillations in turn cause oscillations in downstream metabolic components, including cytosolic AMP and ATP (Fig 4C and 4D). The conductance of K(ATP) channels (gK(ATP)) is dependent on ADP and ATP levels, and oscillations in the concentrations of these nucleotides cause K(ATP) conductance (Fig 4B) and concomitantly K(ATP) current to oscillate and drive slow busting. The slow cAMP oscillations are modulated by Ca2+ and AMP, but in the model of the wild-type β-cells cAMP has no impact on the cell’s electrical activity.
If the key K(ATP) channels are removed, the model cell spikes continuously, as is seen experimentally when a K(ATP) channel blocker like tolbutamide is applied to a wild-type islet [31–33]. The upregulated Kir2.1 conductance shown in Fig 3C would be expected to also provide hyperpolarizing current, but can it rescue the bursting oscillations that are normally driven by K(ATP) current? To answer this, we replaced K(ATP) current in the model with Kir2.1 current to simulate the case for KO cells. The properties of this model current are discussed in Materials and Methods and are shown in Fig 2. A key feature of the Kir2.1 channels is their activation by cAMP [38–40].
In Fig 5 we show that if Kir2.1 is sufficiently up-regulated, it can rescue slow bursting in model cells lacking K(ATP). In the model of the KO condition, slow glycolytic oscillations now drive slow AMPc oscillations (Fig 5C) that cause the cAMP concentration to oscillate (Fig 5A, red). cAMP in turn activates the Kir2.1 channels and results in oscillations in the Kir2.1 conductance (Fig 5B). This causes the membrane potential to switch between the active and silent phases, which drives bursting and Ca2+ oscillations as in the wild-type case (Fig 5A, black). The shape of the burst is largely determined by the details of the V and cAMP dependence of the Kir2.1 channels, which in our model is calibrated by data from a human isoform of the channel expressed in human embryonic kidney cells. Differences of channel properties between mouse and human would change the shape of the burst, but not the burst mechanism (unless channel differences were drastic). A robust property of the burst mechanism is that the cAMP concentration peaks during the silent phase in the KO model cells, unlike the wild-type model cells where cAMP peaks at the beginning of the active phase. This peak in cAMP is reflected in the Kir2.1 conductance. Fig 5B shows the moving average of this conductance, where averaging is done over a window of 6 s to filter out fast V-dependent changes. Like cAMP, the Kir2.1 conductance peaks during the silent phase, and the subsequent decline in this conductance starts the next burst. Although the ATP concentration also oscillates (Fig 5D), it does not affect the membrane potential in this case since there are no K(ATP) channels to sense changes in nucleotides.
In the wild-type model cells, cAMP had no effect on membrane potential or any other components of the model. However, in the model we made of the KO cells, cAMP, acting through Kir2.1 channels, is now the key to slow bursting. To further understand how this occurs, a slow burst is shown in more detail in Fig 6. In this figure, voltage is averaged over the duration of each spike to illustrate mean voltage (Fig 6A, red). This allows us to focus on the slower burst waveform. The figure begins with the system in the silent phase, where Kir2.1 conductance is high (Fig 6D) due to elevated cAMP concentration (Fig 6B, red) and a relatively hyperpolarized voltage (Fig 6A, red). As glycolytic activity declines near the end of the silent phase AMPc slowly increases (Fig 6B, black). This, in turn, reduces the cAMP concentration by inhibiting adenylyl cyclase, thereby reducing Kir2.1 channel activation (Fig 6C, red). The resulting decline in Kir2.1 conductance initiates an active phase of electrical activity, further reducing Kir2.1 conductance due to voltage-dependent channel blockade as the cell depolarizes (Fig 6C, black). Cytosolic Ca2+ now increases due to Ca2+ influx through voltage-dependent Ca2+ channels and this activates Ca2+-ATPase pumps through ATP hydrolysis, further increasing the AMPc. This causes cAMP to decline rapidly. By the middle of the active phase AMP reaches its peak and then starts to decline. This decline, despite the continued rise in c, is due to the upstroke of the glycolytic oscillator, which facilitates the production of ATP at the expense of ADP and AMP. Decreased AMPc disinhibits adenylyl cyclase and cAMP again starts to increase. The cytosolic Ca2+ concentration starts to decrease only after cAMP is elevated enough to significantly activate Kir2.1 current (Fig 6C, red), eventually terminating the active phase.
The KO model relies on the action of cAMP oscillations on Kir2.1 channels to drive electrical bursting and Ca2+ oscillations in the SUR1-/- islets. If cAMP is tonically elevated, then the subsequent tonic activation of Kir2.1 should hyperpolarize the islet, terminating electrical bursting and Ca2+ oscillations, and bringing the intracellular Ca2+ concentration to a low level. We performed this manipulation by adding 8-Bromoadenosine 3’,5’-cyclic monophosphate (8-Br-cAMP) to wild-type and SUR1-/- islets. This is a membrane permeant brominated derivative of cAMP that is resistant to degradation by cAMP phosphodiesterase, and is thus long lasting.
Application of 8-Br-cAMP (50 μM) to wild-type islets (N = 10) had little or no effect on Ca2+ oscillations, as shown in three representative islets (Fig 7A). In contrast, the same concentration applied to SUR1-/- islets terminated Ca2+ oscillations in all islets tested (N = 9), reducing the intracellular Ca2+ level to what is expected from a hyperpolarized islet (Fig 7B). This is consistent with the hypothesis that cAMP activates Kir2.1 channels, and that oscillations in cAMP drive oscillations in Ca2+ in SUR1-/- islets, but not wild-type islets.
To better understand the dynamics of the bursting mechanism, and to help facilitate the design of new experiments, we performed a fast/slow analysis of the Kir2.1 model. Fast/slow analysis separates system variables into component fast and slow subsystems based on their respective time scales [52]. The slow variables are almost constant on the time scale of changes in the fast variables. Therefore, these variables can be treated as slowly-varying parameters of the fast subsystem. In our model, the fast variables are voltage (V), the activation variable for voltage-gated K+ current (n) and cytosolic Ca2+ (c). The variables that change on much slower time scales are fructose 6-phosphate (F6P), fructose 1,6-bisphosphate (FBP), ATPc, AMPc, cAMP and the Ca2+ concentration of the ER (cer). For comparison, Fig 8A shows a fast variable (c) shown together with a slow variable AMPc. At the start of a burst active phase c immediately jumps to a plateau and exhibits small oscillations due to the voltage spikes, and jumps down at the end of the active phase. In contrast, AMPc exhibits a slow rise and fall, with a peak near the middle of the active phase. We start the fast/slow analysis by setting cer to its mean value, since it is not a part of the primary oscillatory mechanism. The slow variables other than cer interact according to the following scheme:
F6P→FBP→ATP→AMP→cAMP
where only cAMP directly affects the fast subsystem, through the cAMP-dependent activation variable of IKir (c∞). We first generate a bifurcation diagram of the fast subsystem with c∞ as the bifurcation parameter (Fig 8B), since the curve is simpler than that obtained using cAMP itself as the bifurcation parameter. For small values of c∞ the system is at a depolarized steady state, since the Kir2.1 current is largely turned off. These stable steady states make up the initial segment of the upper branch of the z-shaped curve (solid line), which we refer to as the z-curve. As c∞ is increased two branches of periodic solutions, one stable (bold solid curve) and one unstable (bold dashed curve), emerge at a saddle node of periodics (SNP) bifurcation. The branch of unstable limit cycles is created at a subcritical Hopf Bifurcation (HB), at which point the branch of stable steady states becomes unstable (dashed curve). The branch of unstable steady states turns at a saddle-node bifurcation (SN1), forming the middle branch of the z-curve. This branch turns at another saddle-node bifurcation (SN2) and forms the stable lower branch of the z-curve. The stable branch of periodic solutions reflects tonic spiking, and the minimum and maximum voltage values during a spike are shown as two separate curves. This branch terminates at the left knee of the z-curve at a saddle-node on invariant circle (SNIC) bifurcation.
The burst trajectory is shown projected into the c∞-V plane in Fig 8C. The left portion of the trajectory reflects the active phase of the burst when the model cell is spiking. When the cell enters the silent phase c∞ first increases and then decreases to start a new active phase. This is the right portion of the trajectory. The burst trajectory is superimposed onto the z-curve in Fig 8D, along the c∞ curve (Eq 11). This curve depends on the cAMP concentration, which has the following steady state function:
cAMPss=kPDEcampVACv¯PDE(αPDE+βPDECiss3Ciss3+KPDEca3)−VAC
(12)
where VAC is the rate of adenylyl cyclase production and is inhibited by AMPc (Eq 2). AMPc changes slowly during a burst (Fig 8A, blue) due to the activity of the glycolytic oscillator. The steady-state cytosolic Ca2+ concentration in Eq 12 (ciss) is given by:
ciss=αICa+kleakcerkpmca+kleak+kSERCA
(13)
where ICa is a function of V and cer is clamped at its mean value. This gives the voltage dependence to the c∞ curve.
During the burst, the glycolytic oscillator moves the c∞ curve back and forth. In Fig 8D the curve is plotted for values of AMPc at its minimum and its maximum during a burst. During a burst AMPc moves between these minimum and maximum values and shifts the c∞ curve back and forth. For small values of AMPc, the c∞ curve is shifted to the right (magenta dashed curve), intersecting the z-curve on the bottom stationary branch. At this point the system is in its hyperpolarized silent phase. As AMPc slowly increases the c∞ curve shifts to the left and the phase point follows it. When the curve passes the knee, the phase point is attracted to the periodic spiking branch, starting the active phase. The phase point follows the periodic branch to the left until AMPc reaches its maximum (green dashed curve). From here AMPc declines and shifts the c∞ curve rightward, bringing the phase point with it. The c∞ curve eventually reaches SN2 again and intersects the stable stationary branch initiating a silent phase. It keeps moving rightward as AMPc continues to decline, bringing the phase point with it. Eventually AMPc begins to rise, restarting the cycle. This is parabolic bursting since the spike frequency during a burst follows a parabolic time course, low at the beginning and the end as the phase point passes near the infinite-period SNIC bifurcation [53]. As the fast subsystem bifurcation diagram lacks a bistable region, the glycolytic oscillations are necessary for the production of bursting in the Kir2.1 model.
To address whether the upregulation of other types of K+ channels might yield effects similar to those of Kir2.1, we examined the effects of replacing K(ATP) current with an alternative hyperpolarizing constant-conductance or “leak” K+ current, instead of Kir2.1 current, and increased the K(Ca) channel conductance (Fig 9). With these modifications, bursting could be produced in the absence of K(ATP) due to Ca2+ feedback onto K(Ca) channels (Fig 9A). In this model, ER Ca2+, which played little or no role in bursting produced using the Kir2.1 model, became absolutely essential in driving the burst.
Glycolytic oscillations are now irrelevant since they do not change the membrane potential or contribute to burst generation in any way. The fast subsystem consists of three variables in this case, V, n, and c, and a slow variable cer, which we consider as a slowly-varying parameter of the fast subsystem. The fast-subsystem bifurcation diagram is shown in Fig 9B. Unlike with the Kir2.1 model (Fig 8), there is a bistable interval in the z-curve, where stable steady states coexist with stable periodic solutions (between the saddle-node bifurcation SN2 and the homoclinic bifurcation HC). The burst trajectory is projected into the cer-V plane in Fig 9C, and superimposed on the fast-subsystem bifurcation diagram in Fig 9D. Also superimposed is the cer nullcline, the curve where the cer derivative is 0. Bursting is produced as the trajectory moves to the left along the bottom stationary branch of the z-curve during the silent phase and to the right along the periodic branch during the active phase, utilizing the fast-subsystem bistability. This is standard square-wave or type 1 bursting that has been described previously for other models of bursting in β-cells and in neurons [52,54].
We have thus far described two possible ways in which the upregulation of hyperpolarizing K+ channels can rescue bursting in SUR1-/- β-cells. As one clear difference between the two alternative mechanisms is their dependence on ER Ca2+ concentration, we explored the consequences of manipulating the ER Ca2+ concentration as a way of determining which model is more likely the correct one. This can be done experimentally by blocking the Ca2+ pumps on the ER membrane (the SERCA pumps) using the agent thapsigargin [55].
In the model, the parameter kSERCA is the Ca2+ pumping rate into the ER from the cytosol. To mimic the effect of thapsigargin we reduced kSERCA by a factor of 4. In the ER bursting model, this greatly lowered cer (Fig 10A, blue trace) and converted slow bursting into fast two-spike bursting (Fig 10A, black trace). In terms of the fast/slow analysis (Fig 9B), the reduction in kSERCA shifts the z-curve and cer nullcline far to the left. In addition, the periodic tonic spiking branch is destabilized through a period doubling bifurcation, and the resulting period doubled branch itself loses stability at a period doubling bifurcation. In fact, there is a period doubling cascade (green curve), leading ultimately to a branch of fast two-spike bursting (blue curve). The trajectory (red curve) moves to this latter curve at the new equilibrium value of cer. Thus, the slow bursting is replaced by very fast 2-spike bursting.
In the Kir2.1 model, in contrast, bursting persisted even when SERCA pumps were inhibited (Fig 10C, black). This is because bursting in this case is driven by the activity of the glycolytic oscillator. Blocking SERCA pumps lowers mean cer, which affects the cytosolic Ca2+ level, but this only modulates the slow bursting pattern rather than abolishing it. Indeed, the fast/slow analysis illustrates that the burst mechanism is very similar in this case to what it was before the reduction in kSERCA (Fig 10D). The main difference is that the period of bursting is now increased, since the c∞ curve moves further to the right during the silent phase (Fig 10D, dashed magenta curve).
These simulations make a testable prediction that can eliminate one or the other of the compensation models. We subsequently tested the predictions in the lab, by treating oscillating SUR1-/- islets with thapsigargin (TG). Fig 11 shows the model prediction obtained with the Kir2.1 model on the top row and the results of the experiments on the bottom three rows (three SUR-/- islets and three wild-type islets are shown). TG application did not terminate slow Ca2+ oscillations in any of the SUR1-/- islets shown (Fig 11B), as predicted by the Kir2.1 model (Fig 11A). In fact, Ca2+ oscillations persisted in all 10 of the KO islets tested, with only a small change in the properties of the oscillations. Before TG treatment the oscillation period was 7.3 ± 1.2 min and the duty cycle (duration of elevated Ca2+ divided by the period) was 0.4 ± 0.08. After TG application there was a slight increase in period to 7.6 ± 1.1 min and the duty cycle increased to 0.6 ± 0.06. The slow Ca2+ decline that occurs at the end of each active phase prior to TG application, characteristic of Ca2+ leaking out of the ER and into the cytosol, was eliminated by the application of TG, as expected [56,57]. The persistence of oscillations when TG is applied is in clear contrast with the wild-type model (the model that has K(ATP) current) and wild-type islets, where in most of the wild-type islets tested TG converted slow oscillations (with period 10.6 ± 0.9 min and duty cycle 0.5 ± 0.06) to continuous spiking or fast bursting with an elevated cytosolic Ca2+ level (in 13 of 14 islets tested) (Fig 11C and 11D). A similar effect of TG on slow Ca2+ oscillations was previously observed in islets [58]. Since the response to TG confirms the prediction of the Kir2.1 model, but not the ER bursting model, we conclude that the Kir2.1 model is a more likely candidate to account for the compensation that occurs in SUR1-/- islets. That is, the data support the hypothesis that bursting observed in KO islets is due to compensatory upregulation of Kir2.1 channels.
The primary aim of this modeling study was to help understand how islet β-cells can compensate for the genetic knockout of K(ATP) channels in SUR1-/- mice. One focus was on Kir2.1 channels, which we found to be upregulated in the SUR1-/- mice (Vadrevu et al, manuscript in preparation). We showed that upregulation of these channels can maintain bursting, even though the K(ATP) channels that normally couple metabolic oscillations to plasma membrane K+ channel activity are missing. This requires that the Kir channels have a dependence on cAMP, as has been reported previously for Kir2.1 channels [38–41]. It has also been reported that cAMP exhibits slow oscillations in insulin-secreting MIN6 cells [43] and in islet β-cells [42,43], a behavior which could reflect oscillations in the nucleotide AMP [44]. Indeed, we were not able to observe bursting in simulations of K(ATP) KO islets if AMP regulation of cAMP was omitted. We did, however, show an alternative mechanism that could produce bursting in the KO islets in a manner that is independent of Kir2.1 current. The two models made very different predictions for the effects of blocking Ca2+ pumps in the ER membrane, and subsequent experiments with the SERCA pump blocker thapsigargin supported the Kir2.1 model over the alternate model. Of course, we do not suggest that these are the only two models that might be capable of mediating bursting in the absence of K(ATP). For example, there are data showing that Mg:ATP can stimulate Kir channels, providing another means by which metabolic oscillations could cause bursting electrical activity [38]. We do show, however, that the two models examined herein are both feasible, and that they are experimentally discernable.
A key hypothesis that we make in the Kir2.1 model is that cAMP regulates Kir2.1 current in SUR1-/- islets, likely through PKA as described in [38–41], rather than direct metabolic regulation of the channels. A consequence of this hypothesis is that manipulations that increase the cAMP level should hyperpolarize the islet and terminate Ca2+ oscillations. Indeed, we found this to be the case. Application of 8-Br-cAMP had no apparent effect on Ca2+ oscillations in wild-type islets (Fig 7A), but terminated oscillations and brought Ca2+ to a resting level in SUR1-/- islets. This is what we predict, since we expect little or no expression of Kir2.1 channels in wild-type islets, but significant expression in SUR1-/- islets (Fig 3). The data of Fig 7 do not preclude the possibility that cAMP activates another type of K+ channel in SUR1-/- islets, but other data show that the upregulated current is an inward-rectifying K+ current (Fig 3). If this upregulated Kir current were regulated directly by metabolism rather than cAMP, it is hard to explain why increasing the cAMP level with membrane permeable 8-Br-cAMP would terminate Ca2+ oscillations and bring Ca2+ to a resting level.
Another hypothesis that we make is that cAMP oscillates in SUR1-/- islets. This has not yet been demonstrated, as it has been in wild-type islets [42,43]. However, we have previously reported that slow NAD(P)H oscillations persist in the SUR1-/- islets (Merrins et al, 2010), indicating the existence of metabolic oscillations which could drive cAMP oscillations as in our model. In glucose-stimulated wild-type islets the glycolytic product fructose 1,6-bisphosphate exhibits oscillations coincident with electrical bursting and Ca2+ oscillations [59], and there are slow oscillations in oxygen consumption [60] and NAD(P)H [61,62]. At present, we do not yet know if the metabolic oscillations in fact result in cyclic AMP oscillations in SUR1-/- islets.
The upregulation of Kir2.1 channels we propose might result from the expected increase in β-cell electrical activity that occurs when K(ATP) channel formation is disrupted by the genetic deletion of SUR1, although when this occurs developmentally is not clear. It is well established that dramatic changes in electrical activity can regulate the expression of ion channels in excitable cells [63–65]. This may result from the increased intracellular Ca2+ concentration that accompanies increased electrical activity, which can enhance gene expression [66,67]. This feedback process would guard against the production of excessive Ca2+ levels in the cell, which can in turn induce apoptosis [68].
One prediction of the Kir2.1 model is that the Ca2+ and cAMP oscillations should be 180° out of phase with one another in the KO cells (Fig 5A). This differs considerably from the wild-type case, where cAMP has a saw-tooth pattern and declines during the burst active phase and then rises during the silent phase (Fig 4A). While cAMP levels have been measured simultaneously with Ca2+ in MIN6 cells and the time course is in general agreement with the model [43], such measurements have not yet been made in SUR1-/- islets. A study performed in MIN6 β-cells in which Ca2+ oscillations were induced with the aid of the K+ channel blocker tetraethylammonium (TEA) showed oscillations in protein kinase A activity that was generally in phase with cAMP oscillations, indicating that the kinase kinetics were sufficiently fast to resolve the roughly 6-min oscillations in the cAMP concentration [69]. Our model would predict this, for both wild-type and SUR1-/- islets. The PKA oscillations could affect islet β-cells in ways other than or in addition to phosphorylation of Kir channels, such as phosphorylation of L-type Ca2+ channels as has been demonstrated in the TC3 β-cell line [70].
Glycolytic oscillations are well established in yeast [71], but until recently there was no direct evidence that they occur in islet β-cells. However, recent studies using a FRET sensor for the glycolytic enzyme pyruvate kinase provide direct evidence for the existence of glycolytic oscillations in islets [72,59]. These metabolic oscillations are readily transmitted to the membrane potential through the cyclic activity of K(ATP) channels [46](Fig 4), and we have now illustrated how these can drive bursting even in the absence of K(ATP) channels by utilizing the cAMP dependence of upregulated Kir2.1 channels. It is not obvious how Kir2.1 channel expression increases to an appropriate level so that bursting is produced when K(ATP) channels are missing, but it is plausible that channel compensation is achieved through the actions of Ca2+ on Ca2+-dependent activators or inhibitors of transcription factors. A further modeling study for the dynamic regulation of Kir2.1 channel expression is currently under way.
|
10.1371/journal.pbio.1000587 | A c-di-GMP Effector System Controls Cell Adhesion by Inside-Out Signaling and Surface Protein Cleavage | In Pseudomonas fluorescens Pf0-1 the availability of inorganic phosphate (Pi) is an environmental signal that controls biofilm formation through a cyclic dimeric GMP (c-di-GMP) signaling pathway. In low Pi conditions, a c-di-GMP phosphodiesterase (PDE) RapA is expressed, depleting cellular c-di-GMP and causing the loss of a critical outer-membrane adhesin LapA from the cell surface. This response involves an inner membrane protein LapD, which binds c-di-GMP in the cytoplasm and exerts a periplasmic output promoting LapA maintenance on the cell surface. Here we report how LapD differentially controls maintenance and release of LapA: c-di-GMP binding to LapD promotes interaction with and inhibition of the periplasmic protease LapG, which targets the N-terminus of LapA. We identify conserved amino acids in LapA required for cleavage by LapG. Mutating these residues in chromosomal lapA inhibits LapG activity in vivo, leading to retention of the adhesin on the cell surface. Mutations with defined effects on LapD's ability to control LapA localization in vivo show concomitant effects on c-di-GMP-dependent LapG inhibition in vitro. To establish the physiological importance of the LapD-LapG effector system, we track cell attachment and LapA protein localization during Pi starvation. Under this condition, the LapA adhesin is released from the surface of cells and biofilms detach from the substratum. This response requires c-di-GMP depletion by RapA, signaling through LapD, and proteolytic cleavage of LapA by LapG. These data, in combination with the companion study by Navarro et al. presenting a structural analysis of LapD's signaling mechanism, give a detailed description of a complete c-di-GMP control circuit—from environmental signal to molecular output. They describe a novel paradigm in bacterial signal transduction: regulation of a periplasmic enzyme by an inner membrane signaling protein that binds a cytoplasmic second messenger.
| Bacteria can live as free swimming cells or attached to surfaces in communities called biofilms. The di-nucleotide c-di-GMP is a key cytoplasmic signal that regulates biofilm formation in a number of bacterial species. Our study, in combination with structural analysis described in the accompanying paper by Sondermann et al., describes key interactions in a c-di-GMP signaling pathway that allows cells of Pseudomonas fluorescens to adapt to changes in the concentration of the nutrient phosphate by regulating biofilm formation. The adhesion protein LapA is localized outside the bacterial cell membrane and is responsible for keeping cells attached to surfaces. We show that under low phosphate conditions levels of c-di-GMP are depleted in cells, and these changes are sensed by LapD, a transmembrane c-di-GMP receptor protein. When c-di-GMP levels are low, the LapD protein is kept in an “off” state that allows LapG, a periplasmic protease, to interact with LapA and cleave the N-terminal domain of this adhesion, releasing LapA from the cell surface and promoting biofilm detachment. Under abundant phosphate conditions, LapD binds c-di-GMP in the cytoplasm and binds to and sequesters LapG in the periplasm, promoting cell adhesion via maintenance of LapA on the cell surface.
| Bacteria can be exquisitely tuned to sense and respond to changes in their environment. A single cell may possess an immense repertoire of signal transduction systems capable of receiving sensory input and directing physiological adaptation. The recent groundswell of studies on the intracellular second messenger cyclic dimeric GMP (c-di-GMP) has added a new dimension to bacterial signaling. c-di-GMP controls major lifestyle transitions for bacteria, promoting the shift from motile to sessile modes of growth through impacts on diverse physiological outputs. This molecule is synthesized by diguanylate cyclases (DGCs) [1], proteins that contain the GGDEF domain, and can be degraded by specific phosphodiesterases (PDEs) containing either the EAL or HD-GYP domain [2],[3]. Such domains are ubiquitous in bacterial genomes, and occur in combination with an array of sensory input and output modules [4].
A substantial body of work has identified specific DGCs and PDEs that impact cell adhesion and biofilm formation in diverse bacteria. The phenotypic effects of these signaling proteins include changes in exopolysaccharide (EPS) production, motility, and transcription [5]. Assigning c-di-GMP signaling activity to many proteins, sometimes dozens within a single bacterium, has highlighted the complexities of c-di-GMP signaling networks, and has exacerbated the task of connecting specific environmental signals to discrete outputs.
A key, recent advance in our understanding c-di-GMP's role in bacteria has been the identification of c-di-GMP receptors with defined outputs. Receptors, or effector proteins, identified thus far utilize a range of c-di-GMP binding mechanisms to impact EPS synthesis [6],[7],[8], motility [9],[10],[11],[12], transcription [13],[14],[15], and sub-cellular [16] or cell-surface protein localization [17]. In a few cases, molecular details of the effector's output have been determined. c-di-GMP binding to the PilZ domain of YcgR stimulates its interaction with the flagellar complex of E. coli, resulting in a counter-clockwise rotational bias and reduced motility [18],[19],[20]. In V. cholerae, c-di-GMP binds the transcription factor VpsT causing a change in its oligomerization and activity, inversely regulating genes for rugosity and motility [15]. PopA of C. crescentus undergoes dynamic localization to the cell pole upon c-di-GMP binding, recruiting a cell cycle regulator for degradation [16]. In addition to binding effector proteins, c-di-GMP has also been shown to bind riboswitches [21],[22],[23]. The diversity of these control mechanisms, and their varied targets, highlights the scope and intricacy of c-di-GMP signaling. Despite the significant progress these studies represent, in most cases the environmental or cellular inputs controlling the DGCs and/or PDEs that regulate these effectors have yet to be defined.
Stable surface attachment and subsequent biofilm formation by Pseudomonas fluorescens Pf0-1 requires a large adhesive protein, LapA. This ∼520 kD protein is secreted to the surface of the outer membrane by an ABC transporter encoded by the lapEBC genes [24]. LapA's maintenance on the cell surface is controlled post-translationally by the c-di-GMP binding protein LapD [17]. When c-di-GMP levels are high, LapD binds c-di-GMP and promotes biofilm formation via accumulation of LapA on the cell surface. In the absence of c-di-GMP binding to LapD, LapA is released from the cell rendering it unable to attach [17].
In a prior study, our group characterized LapD, reporting genetic and biochemical evidence that LapD binds c-di-GMP through its cytoplasmic EAL domain and controls biofilm formation via a periplasmic output domain [17]. The structure/function analysis presented by Newell et al. suggested that LapD controls LapA localization by a unique inside-out signaling mechanism: binding c-di-GMP in the cytoplasm and transmitting this signal through the inner membrane to the periplasm via a HAMP domain. Such a mechanism could account for how changes in cytoplasmic c-di-GMP levels control LapA's stability on the cell surface post-translationally. However, the mechanism by which LapD's periplasmic domain impacted LapA localization was unknown.
The availability of inorganic phosphate (Pi) is an important environmental signal that governs biofilm formation by P. fluorescens Pf0-1 via a c-di-GMP-dependent mechanism. When Pi is limiting, the c-di-GMP PDE RapA is expressed and depletes cellular c-di-GMP, suppressing biofilm formation [25]. One effect of RapA's activity is the loss of the adhesin LapA from the cell surface. While our previous study showed that the effects of Pi starvation and RapA expression on biofilm require signaling through LapD [17], the specific contribution of LapD to changes in LapA localization in this signaling pathway was not known.
Here we uncover how LapD controls LapA localization and provide biochemical data describing its function as an inside-out signaling protein. When bound to c-di-GMP, LapD inhibits the activity of a periplasmic protease, LapG. In the absence of c-di-GMP binding to LapD, LapG is free to cleave the N-terminus of LapA, releasing the adhesin from the cell and preventing biofilm formation. Upon Pi starvation, the LapD-LapG system responds to c-di-GMP depletion by RapA and promotes biofilm detachment. These data, in combination with the companion study by Navarro et al. 26 presenting a structural analysis of LapD's signaling mechanism, describe a key connection in a complete c-di-GMP control circuit that links environmental signal to cellular output.
Note: The Supporting Information section includes an expanded alignment of LapA-like proteins (Figure S1), a graphical depiction of data describing inhibition of LapG activity by c-di-GMP and additional data on LapG activity in the presence of detergents (Figure S2), data describing the localization of control proteins in the presence and absence of c-di-GMP (Figure S3), and images of representative biofilm assays from the dataset depicted graphically in Figure 7B (Figure S4).
In this study, our objective was to determine the mechanism by which the c-di-GMP effector LapD controls LapA localization. In an effort to identify additional players in this pathway, we investigated the function of a gene immediately upstream of lapD, designated lapG. We deleted the lapG gene (Pfl_0130) and determined the effects of this mutation on irreversible surface attachment and biofilm formation. After 6 h in a static culture, the lapG mutant (ΔlapG) showed a hyper-adherent biofilm phenotype, accumulating twice as much biomass on the culture well as the WT (Figure 1A). These strains were examined by microscopy under similar, static growth conditions. After a 1-h incubation, irreversibly attached ΔlapG cells covered twice as much of the substratum as compared to the WT (Figure 1B). Through longer incubation times, ΔlapG continued to show about twice as many attached cells as WT (unpublished data). These results suggest that increased cell attachment accounts for the biofilm phenotype of ΔlapG.
To complement the lapG mutant we reintroduced the gene on a multi-copy plasmid. This caused total loss of biofilm formation, shown and discussed in more detail below. A second approach was employed: restoring the lapG gene to its native locus in ΔlapG using allelic replacement. The resulting strain, lapGREST, showed similar levels of biofilm formation and surface attachment as WT (Figure 1A,B). When a lapG allele carrying an internal HA epitope tag (lapG-HA) was introduced into the lapG locus, this also restored the WT phenotype (Figure 1A).
The adhesin LapA is the primary factor required by P. fluorescens for attachment to surfaces under these conditions [17],[24],[25]. We hypothesized that increased expression or cell surface localization of LapA might account for the biofilm phenotype of ΔlapG. To test these hypotheses, we examined LapA levels in cell extracts and culture supernatants by Western blot, and on the surface of intact cells by dot blot. Cell extracts of WT and ΔlapG showed similar levels of LapA, suggesting comparable levels of LapA protein expression in these strains (Figure 1C; 0.98±0.05-fold change from WT, n = 3). Interestingly, the lapG mutant had a unique LapA localization phenotype: there was no detectable LapA in the supernatant and a 2-fold increase in LapA on the cell surface (Figure 1C,D). These data suggest that lapG is involved in the release of LapA from the cell surface. The ΔlapG phenotypes are consistent with previous data showing that cell-surface localization of LapA has a direct and proportional stimulatory effect on biofilm formation [17],[27]. Restoration of either the WT or lapG-HA alleles to the lapG locus of ΔlapG restored a WT LapA localization phenotype (Figure 1C,D).
If increased adhesion by ΔlapG is caused by the aberrant accumulation of LapA on the cell surface, then a mutation in lapA should be epistatic to lapG. Introduction of a null mutation in lapA into the ΔlapG mutant completely eliminated biofilm formation (Figure 1E), a phenotype identical to that of a lapA mutant.
LapG contains a putative Sec secretion signal (probability 0.85; www.cbs.dtu.dk/services/signalP) but no transmembrane domains, and thus is predicted to be periplasmic. To test this proposed localization of LapG, we used osmotic shock to release periplasmic proteins and compared the relative proportion of LapG in this periplasmic fraction versus the remaining spheroplasts. We utilized the lapGHAREST strain expressing GFP to provide a control for tracking cytoplasmic proteins. GFP localized exclusively to the spheroplasts, where it was enriched relative to whole cell lysates (Figure 1F). In contrast, LapGHA was enriched in the periplasmic fraction and depleted in the spheroplasts. These data suggest that LapG resides in the periplasm.
Noting the absence of LapA in the supernatant and its accumulation on the cell surface of the lapG mutant, we hypothesized that LapG functions to modify and release LapA from the cell. A smaller variant of LapA, Mini-LapA, was generated to assess modification of this large protein (∼520 kDa). Mini-LapA consists of the N- and C-termini of LapA, flanking an internal myc-epitope tag (Figure 2A, top). Mini-LapA was introduced on a multi-copy plasmid to WT and ΔlapG strains and its secretion and cell-surface localization were assessed. Mini-LapA localized to the cell-associated (Figure 2B) and supernatant fractions (unpublished data), indicating that it is still secreted. However Mini-LapA does not appear to be functional for adhesion, as it cannot complement a lapA mutant (unpublished data). We observed a difference in the apparent size of secreted Mini-LapA isolated from strains with or without LapG: in the presence of LapG (i.e., the WT genetic background), Mini-LapA migrates at a smaller size, approximately 130 kDa. In contrast, Mini-LapA in ΔlapG migrates at its predicted size of 145 kDa (Figure 2B). This result suggests that LapG is required for a change in molecular weight of Mini-LapA. Given the predicted function of LapG, as a cysteine protease [28], the modification to Mini-LapA was likely proteolytic cleavage.
We developed an assay to assess the necessity of LapG for Mini-LapA modification. Unmodified Mini-LapA was prepared from a cell extract of the ΔlapG mutant overexpressing Mini-LapA. This Mini-LapA substrate was incubated at room temperature for 30 min with cell extracts prepared from the ΔlapG mutant, the ΔlapG mutant carrying an empty vector, and the ΔlapG mutant carrying a plasmid overexpressing LapG. Reactions were then analyzed by Western blotting to reveal that Mini-LapA modification only occurs in the presence of LapG (Figure 2C). This result suggests a model in which LapG functions to modify the LapA protein through proteolytic cleavage of 10–15 kDa from LapA.
We hypothesized that LapG cleaves 10–15 kDa from the N-terminus of Mini-LapA, as the C-terminus of LapA contains residues necessary for type I secretion. To test this hypothesis, we constructed another surrogate LapG substrate, N-Term-LapA, consisting of the first 235 amino acids of LapA with a 6xhistidine (6H) epitope tag at its C-terminus (Figure 2A, bottom). N-Term-LapA and LapG-6H were each purified by nickel affinity chromatography, then incubated together for 15, 60, and 120 min. Subsequent Western blotting revealed that LapG-6H is necessary and sufficient for N-Term-LapA modification in vitro (Figure 2D). Modification occurs in a time-dependent manner and results in a 10–15 kDa reduction in the apparent molecular weight of N-Term-LapA, consistent with the observations of Mini-LapA above.
LapG contains a conserved domain of unknown function (DUF920), proposed to constitute a family of Bacterial Transglutaminase-like Cysteine Proteinases (BTLCPs) [28]. In the study identifying BTLCPs, the authors note that BTLCPs contain a conserved C-H-D catalytic triad. We tested the requirement for the cysteine of LapG's catalytic triad for LapA modification, by mutating C135 of LapG to alanine. Even after incubation with N-Term-LapA for 2 h, purified LapG-C135A did not modify N-Term-LapA. As a control, the WT LapG completely converted the N-Term-LapA substrate in this time (Figure 2E).
Given the inactivity of LapG-C135A, we predicted that this mutation would disrupt LapG's function in vivo. We expressed this mutant on a multi-copy plasmid in the ΔlapG strain and assessed the effect of biofilm formation, relative to the WT allele. As mentioned above, expressing WT LapG from a plasmid resulted in a loss of biofilm formation (Figure 2F). The strain expressing the C135A mutant showed a hyper-adherent biofilm phenotype, comparable to that of the ΔlapG mutant. Together, these data show that LapG's cysteine residue is required for N-Term-LapA modification and that LapG's activity is required for WT biofilm formation. This suggests a model in which cleavage of the LapA protein by LapG is necessary for release of the adhesin from the cell.
To identify the site where N-Term-LapA is cleaved by LapG, modified and unmodified N-Term-LapA samples were purified and sequenced by Edman degradation. N-terminal sequencing revealed that the first 10 amino acids of modified N-Term-LapA are AGPSAAGTGG. These residues correspond to residues 109–118 of unmodified N-Term-LapA and chromosomally encoded LapA. Therefore, LapG functions to proteolytically cleave 108 amino acids from the N-terminus of N-Term-LapA (Figure 3A). A BlastP search with the LapG sequence helped us identify a number of LapA-like proteins encoded near LapG homologs in other bacteria. Upon aligning the N-termini of these putative adhesins, we found some residues were conserved at the site where LapA is cleaved, including alanines 108 and 109 that flank the site, as well as the position of this site relative to the N-terminus (Figures 3B and S1).
To test if conserved residues in LapA are important for recognition and/or cleavage by LapG, we constructed a mutant N-Term-LapA replacing both alanines 108 and 109 with arginine (AA-RR). Cellular extracts were prepared from WT and ΔlapG strains expressing WT or mutant N-Term-LapA variants and N-Term-LapA cleavage was assessed by Western blot. We observed that LapG is unable to cleave N-Term-LapA-AA-RR variant (Figure 3C), suggesting that the alanines at positions 108 and/or 109 are critical for LapG cleavage of N-Term-LapA in vitro.
The phenotype of C135A suggests that LapG-dependent cleavage of the first 108 amino acids from the N-terminus of LapA is required to release LapA from the cell surface in vivo. We therefore hypothesized that an AA-RR mutation in full-length LapA would block LapG activity in vivo and result in a hyper-adherent biofilm phenotype due to accumulation of LapA at the cell surface. We introduced the AA-RR mutation into the chromosomal copy of lapA by allelic replacement and assessed the biofilm phenotype. The strain expressing LapA-AA-RR forms a hyper-adherent biofilm compared to the WT, although less so than that observed for the lapG mutant (Figure 3D). Next we examined the effect of the AA-RR mutation on LapA accumulation. Quantitative dot blot analysis showed much higher LapA levels on the cell surface of the lapA-AA-RR strain compared to the WT, approaching the abundance observed for the ΔlapG mutant (Figure 3E). Cell extracts of these strains showed similar levels of LapA, suggesting comparable levels of LapA protein expression (Figure 3F). We saw a reduction in LapA in the culture supernatant of lapA-AA-RR relative to WT (down 27% ± 11% SD, n = 4) but not a complete loss, as observed in the ΔlapG mutant (Figure 3F). These results suggest that while the AA-RR mutation eliminates cleavage of N-Term-LapA in vitro, this mutation only partially blocks LapG cleavage of LapA in vivo. In support of this interpretation, introducing a lapG mutation into the lapA-AA-RR strain background yielded a hyper-adherent biofilm indistinguishable from the ΔlapG mutant phenotype. Importantly, these results support a model in which cleavage of the first 108 amino acids from the N-Terminus of LapA by LapG is the mechanism required to release LapA from the cell surface in vivo.
The effects of the lapG deletion on cell attachment and LapA localization are precisely opposite those of a lapD mutant (Figure 4B,C). Our previous work showed that lapD is required for maintenance of LapA on the cell surface; conversely, gain-of-function mutations in LapD result in biofilm and LapA localization phenotypes similar to that of a lapG mutant [17]. Given that LapD and LapG play opposing roles in regulating attachment via LapA, we predicted that they might function in the same pathway, and thus analyzed their genetic relationship.
The lapG and lapD genes occur in a putative operon adjacent to the genes encoding LapA and LapEBC, the ABC transporter required for LapA secretion (Figure 4A). We made a clean deletion of lapG-lapD and tested this strain for biofilm formation. As shown in Figure 4B and C, the lapG lapD double mutant (ΔlapGD) has a hyper-adherent biofilm phenotype and increased cell surface LapA, indicating that lapG is epistatic to lapD and that LapG likely acts downstream of LapD in controlling LapA localization. Introduction of both genes on a plasmid (pLapGD) to ΔlapGD was sufficient to restore WT biofilm and cell surface LapA levels (Figure 4B,C). A plasmid on which each ORF was epitope-tagged (pLapGHA-LapD6H) also complemented ΔlapGD (unpublished data) and was used for protein interaction studies described below.
To further explore the opposing effects of lapG and lapD on LapA localization and biofilm formation, we overexpressed each gene individually, then both simultaneously in the WT strain. Overexpressing either lapG or lapD individually phenocopied the mutant phenotype of the other gene in our biofilm assay (Figure 4D). That is, overexpression of lapG eliminated biofilm formation, while overexpression of lapD increased biofilm to levels near those of a lapG mutant. Importantly, expression of both genes together from the same plasmid caused no change in biofilm formation by the WT strain, indicating that the relative dosage of each protein causes the observed effects on biofilm (Figure 4D). We next examined the localization of the LapA protein in each strain. Overproduction of LapD increased LapA in the cell and at the cell surface, but decreased the amount in the supernatant (Figure 4E). Overproduction of LapG had the opposite effect: reducing cellular and cell surface LapA, while increasing the amount in the supernatant. Overexpression of both genes had no effect. Taken together, with the mutant and epistasis analyses, these data confirm that lapG and lapD exert opposing forces on the maintenance of LapA on the cell surface and suggest that they act in the same pathway.
The genetic relationship between lapG and lapD predicts a pathway in which LapD controls LapA localization through regulation of LapG's protease activity. Our previous work suggested a model whereby LapD is a transmembrane signaling protein that binds c-di-GMP via an EAL domain in the cytoplasm, and transmits this signal through a HAMP domain to a periplasmic output domain (Figure 5A) [17]. We reasoned that, when bound to c-di-GMP, LapD might inhibit LapG activity through an interaction in the periplasm.
To test this model, we first analyzed the effect of c-di-GMP on LapG's cleavage of N-Term-LapA in vitro. Addition of c-di-GMP to the lysis buffer in which cell extracts were prepared showed a dose-dependent inhibitory effect on LapG activity, consistent with first order binding kinetics (Figure 5B). Three replicate data sets were obtained for this assay and quantitative densitometry was used to determine the percentage of substrate cleaved at each concentration of c-di-GMP. A curve was fit to each data set and we estimated an average IC50 of LapG for c-di-GMP: 2.3±1.1 µM (Figure S2A). This value is similar to the estimated affinity of the LapD protein for c-di-GMP, a Kd of 5.5+2.8 µM, obtained using different methodology [17].
We next tested if inhibition of LapG activity by c-di-GMP requires LapD. Cell extracts from the lapD mutant carrying the empty vector pMQ72, pLapD, or pLapD mutant variants were assayed for LapG activity in the presence or absence of 50 µM c-di-GMP. Consistent with the results obtained with WT extracts, c-di-GMP addition eliminated LapG activity in extracts with functional LapD (ΔlapD pLapD; Figure 5C). However, in extracts that lacked LapD (ΔlapD pMQ72 strain), there was no effect of c-di-GMP addition on LapG activity.
To test the functional requirements for inhibition of LapG by LapD, we compared the effects of three previously characterized LapD mutants (shown in Figure 5A; [17]). A mutation in the EAL domain of LapD, E617A, shows a severe reduction in c-di-GMP binding. In a cell extract containing this LapD mutant protein, there was no inhibition of LapG by c-di-GMP (Figure 5C). The periplasmic mutation L152P reduces signaling output from LapD, and this LapD variant showed little LapG inhibition upon c-di-GMP addition (Figure 5C). Finally, the ΔH1 mutation in the HAMP domain of LapD results in constitutive signaling output regardless of c-di-GMP binding. Extracts with LapD-ΔH1 showed a severe reduction in LapG activity irrespective of c-di-GMP addition (Figure 5C). These data are fully consistent with a model in which LapD inhibits LapG activity in response to binding c-di-GMP. The effects of each LapD mutation on biofilm formation and LapA retention at the cell surface (Figure 5A) [17] are well explained by their ability to inhibit LapG.
We observed that adding a number of different detergents relieved inhibition of LapG in cell extracts (Figure S2B), suggesting that membrane integrity is important for LapD to inhibit LapG. Given this observation, we hypothesized that inhibition of LapG activity by LapD is a consequence of LapD sequestering LapG to the membrane in a c-di-GMP-dependent manner. To test this idea, we looked to see if addition of c-di-GMP during cell extract preparation affected LapG localization to the inner membrane fraction. We prepared cell extracts of the lapG-HAREST strain, which carries a chromosomal copy of LapG-HA at the lapG locus, in buffer with 0, 1 and 10 µM c-di-GMP. Soluble and inner membrane fractions were isolated as described [24]. Addition of c-di-GMP promoted re-localization of LapG from the soluble fraction to the inner membrane fraction in a dose-dependent manner (Figure 6A), at concentrations consistent with the concentrations needed to inhibit LapG activity in cell extracts prepared under identical conditions (Figure 5B). c-di-GMP addition did not affect LapD's localization (exclusively in the inner membrane), nor did it change the localization of the cytoplasmic protein GFP (see Figure S3 for localization controls).
To determine if LapD was necessary for LapG re-localization, we disrupted the lapD gene in the lapG-HAREST strain. When cell fractions were prepared from the resulting strain, addition of c-di-GMP had no effect on LapG localization (Figure 6B). Interestingly, some LapG was still detected in the IM. Reintroduction of LapD on a plasmid restored c-di-GMP-dependent re-localization of LapG to the IM (Figure 6B). We also tested the functional requirements for LapD's effect on LapG by reintroducing the three LapD variants utilized above. LapD E617A is defective for c-di-GMP binding, shows no inhibition of LapG activity (Figure 5C), and nearly eliminated recruitment of LapG to the IM—even with addition of c-di-GMP (Figure 6B). The L152P mutation to LapD reduces its output [17] and LapG inhibition (Figure 5C) and also reduced LapG recruitment to the IM (Figure 6B). Lastly, the ΔH1 allele of LapD is constitutively active and strongly inhibits LapG activity; this allele promotes almost exclusive IM localization of LapG irrespective of c-di-GMP addition (Figure 6B).
To further substantiate a direct interaction between the LapG and LapD proteins, we assessed the ability of LapG and LapD to co-precipitate. First, immunoprecipitation (Ip) of HA tagged LapG was performed, and we looked for enrichment of LapD. Cell extracts were prepared from the ΔlapGD strain carrying pLapGHA-LapD6H, in buffer with 5 µM c-di-GMP, and 0.8% Thesit to solubilize membranes. Ip of LapGHA by the addition of anti-HA antibody and Protein A resin resulted in co-Ip of LapD6H (Figure 6C). When the assay was performed with a nearly identical strain lacking only the HA epitope on LapG, Ip of LapD6H was eliminated (Figure 6C).
We next utilized a nickel resin to pull down LapD6His and look for LapGHA co-precipitation. Precipitations were performed under the same conditions (5 µM c-di-GMP, 0.8% Thesit) with the addition of 10 mM Imidazole to reduce non-specific binding to the resin. Pull down of LapD6H enriched for LapGHA (Figure 6D). Importantly, omission of the 6H epitope from LapD eliminated precipitation of LapGHA (Figure 6D).
To examine the dependence of LapG-LapD interaction on c-di-GMP, we performed reciprocal pull down assays with 0, 0.5, or 5 µM c-di-GMP. We observed little co-precipitation in the absence of c-di-GMP but saw a dose-dependent increase when the nucleotide was added (Figure 6E). Importantly, the concentrations of c-di-GMP required to promote this interaction are similar for both types of co-precipitation. These concentrations are also on par with what is needed to recruit LapG to the inner membrane (Figure 6A), inhibit LapG activity (Figure 5B), and are consistent with the affinity of LapD for c-di-GMP.
Lastly, we introduced the E617A, L152P, and ΔH1 mutations into the pLapGHA-LapD6H plasmid to test the functional requirements for LapG-LapD interactions. The E617A LapD6H variant was not expressed at as high a level as the other alleles in this construct (Figure 6F). This is in contrast to the wild-type level of expression we have seen for this mutant from the plasmid used in our prior experiments (Figures 5C,6B; unpublished data). Despite this, we still detected some LapGHA pull down by LapD E617A, yet co-precipitation was not stimulated by c-di-GMP addition. This is consistent with the E617A mutant's defect in c-di-GMP binding (Figure 6B). LapD L152P showed a reduced ability to pull down LapGHA relative to the WT, both with and without c-di-GMP addition, consistent with reduced signaling output in this mutant (Figure 6D). Lastly, the ΔH1 mutation resulted in increased co-precipitation of LapGHA by LapD in the absence or presence of c-di-GMP, in full agreement with other data showing this allele to be constitutively active. The effects of these three mutations on LapD's ability to pull down LapG are consistent with their effects on recruitment of LapG to the inner membrane and inhibition of LapG activity. Collectively, these data describe an interaction between LapG and LapD that requires c-di-GMP binding by LapD's EAL domain, signaling through its HAMP domain and a functional periplasmic output domain.
In prior publications, our group has shown that extracellular Pi is an important signal governing biofilm formation by P. fluorescens [17],[25]. In the absence of sufficient Pi, biofilm formation is inhibited by expression of the Pho regulon (Pho). Pho expression blocks LapA-mediated attachment in two ways: it inhibits secretion of the LapA from the cytoplasm to the cell surface, and promotes the release of the adhesin from the cell surface to the culture supernatant. Pho's effects on LapA's cell surface localization require depletion of cellular c-di-GMP by the Pho-regulated PDE RapA [25], and signaling through LapD [17]. In this study, we uncovered evidence that the output of LapD signaling is control of LapG. We next sought to test the necessity and sufficiency of the RapA-LapD-LapG signaling pathway for control of biofilm formation via LapA localization and secretion, in low Pi conditions.
First, we assessed the ability of the ΔlapG mutation to suppress the effects of constitutive Pho regulon expression on biofilm formation. The pst mutant constitutively expresses Pho irrespective of Pi levels [29]; this mutation causes inhibition of biofilm formation even in high Pi medium (Figure 7A) [25]. Deletion of rapA in the pst mutant partially restores biofilm formation, to ∼70% that of the WT (Figure 7A). Pho regulon control of biofilms involves lapD, as the ΔpstΔrapA lapD mutant cannot form a biofilm, and constitutively active lapD ΔH1 suppresses the pst mutation. Finally, deletion of lapG in the pst mutant leads to a hyper-adherent biofilm phenotype, despite constitutive Pho expression in this strain (Figure 7A). The level of biofilm formation by the ΔlapG pst strain is less than that of ΔlapG, but comparable to what is observed in the ΔlapD pst pΔH1 strain (Figure 7A). These results show that LapG plays a critical role in suppression of biofilm formation by the Pho regulon.
To gain insight into the dynamic response of LapD and LapG to changes in cellular c-di-GMP concentration, we evaluated the effects of physiological Pi-starvation on pre-formed biofilms. In this assay, biofilm formation proceeded identically in high Pi and low Pi media up to 5.5 h post-inoculation. At this time (designated t = 0) biofilms in low Pi medium began to disperse, while those in high Pi persisted at a relatively constant level for the duration of the assay (Figure 7B; biofilm images are in Figure S4). After 90 min, the WT strain showed a 70% reduction in attached biomass in low Pi, relative to the high Pi condition. The ΔphoB mutant showed no reduction in biofilm in low Pi, consistent with biofilm detachment requiring the activation of the Pho regulon. The rapA mutation partially rescues biofilm formation in low Pi [25], and here showed only 40% reduction in biofilm after 90 min in low Pi (Figure 7B). Both the ΔlapG and ΔlapD pΔH1 biofilms were unaffected by Pi starvation, showing no detachment in low Pi. These data show that Pho regulon induction leads to detachment of biofilms from the surface, and that this process requires RapA, LapD, and LapG.
Pho induction inhibits secretion of LapA from the cytoplasm to the outer membrane, and also promotes its release from the cell surface into the culture supernatant [25]. To test if the RapA-LapD-LapG pathway is genetically sufficient to explain these effects, we monitored LapA localization under high and low Pi conditions in the WT, rapA, lapD, and lapG mutants. Consistent with our prior work, the WT strain accumulated LapA in the cellular and supernatant fractions under low Pi conditions (Figure 7C). These changes were accompanied by an 80% reduction in LapA on the cell surface when cells are grown in low Pi (Figure 7D). In contrast, the rapA mutant showed no apparent differences in LapA secretion between high and low Pi, and had ≈WT levels of cell surface LapA in both conditions (Figure 7C,D). These observations corroborate previous data implicating rapA in Pho control of both secretion and cell surface localization of LapA [25]. They suggest that c-di-GMP depletion by RapA impacts LapA in two ways: inhibiting its secretion from the cytoplasm to the outer membrane, and promoting release from the cell.
The lapD mutant exhibited little cell surface LapA, and abundant accumulation of LapA in the supernatant fraction irrespective of Pi concentration (Figure 7C,D). In high Pi, ΔlapD shows reduced LapA levels in the cellular fraction relative to WT, as reported [17]. Despite this, ΔlapD still accumulated intracellular LapA in low Pi (Figure 7C). This implies that, in contrast with the necessity of LapD for regulating LapA release from the cell surface, RapA controls LapA secretion in a LapD-independent manner.
In low Pi, the ΔlapG strain showed hyper-accumulation of LapA at the cell surface, comparable to that seen in high Pi (Figure 7D). While ΔlapG did not release LapA into the supernatant fraction in either high or low Pi, it did show some increase in cellular LapA in low Pi (Figure 7C). Taken together, these data suggest signaling through LapD and LapG is required for release of LapA from the cell surface in response to c-di-GMP depletion by RapA (detailed in Figure 8). Our data also suggest that c-di-GMP depletion inhibits LapA secretion by a yet-unidentified LapD-independent mechanism.
c-di-GMP plays a key role in integrating cellular and environmental signals into a bacterium's decision to swim or stick. Recent studies highlight that c-di-GMP can impact varied outputs by binding to effector proteins, including transcription [13],[14],[15], protein localization [16], flagellar motility [9],[10],[11],[12], and EPS synthesis [6],[7],[8]. While the ubiquity and diversity of c-di-GMP signaling pathways is evident, the details of how c-di-GMP effector proteins sense and respond to their ligand are just beginning to emerge.
In P. fluorescens we observed release of the LapA adhesin from the cell surface in response to phosphate limitation [25]. Here we have closed a key gap in the c-di-GMP signaling pathway responsible for this effect. Together with our previous work, this study shows that LapA release depends on c-di-GMP depletion by the PDE RapA, signaling from the cytoplasm to the periplasm by the c-di-GMP effector LapD, and cleavage of the N-terminus of LapA by the protease LapG. To our knowledge, this is the most complete description of a c-di-GMP signaling “circuit” to date, linking a molecular chain of events from environmental signal to output.
Relay of a second messenger signal across the inner membrane to affect an extra-cytoplasmic output is a new paradigm in bacterial signal transduction. In the companion manuscript, Navarro et al. [26] describe structural and functional analyses of LapD, providing significant mechanistic insight into how inside-out signaling works. LapD has two stable conformations, autoinhibited and activated, and c-di-GMP binding drives conversion from the one state to the other. In the autoinhibited conformation, the “empty” EAL domain interacts with the other cytoplasmic domains, likely applying some force on the periplasmic domain and preventing it from interacting with LapG. Mutations that disrupt autoinhibtion cause hyper-adherent phenotypes in vivo, akin to that seen for the ΔH1 mutant described here (data in [26]). Our data are consistent with the ΔH1 mutation uncoupling autoinhibition from the output domain, thus causing constitutive interaction with LapG.
The L152P mutation in LapD causes reduced biofilm formation and impairs interaction with LapG, underscoring the importance of the periplasmic domain for LapD's output. How this mutation may alter LapD's conformation is not clear, as it is C-terminal to the periplasmic domain crystal structure (Navarro et al., [26]). Structure/function analyses of interactions between the purified periplasmic domain of LapD and LapG in vitro demonstrate that this domain is necessary and sufficient for LapG binding (Navarro et al., companion manuscript).
How does LapD inhibit LapG activity? One model is that LapD simply sequesters LapG at the inner membrane from its outer membrane substrate, LapA (Figure 8). It is also possible that LapD inhibits LapG enzymatic activity through allosteric or competitive means. We found no support for the latter hypothesis, observing cleavage of N-Term-LapA under in vitro conditions in which we demonstrate LapD and LapG interact (e.g. in the presence of c-di-GMP). Also, addition of excess LapD output domain had no effect on LapG activity in vitro assays with purified components (our unpublished data). These data argue for a simple sequestration model, though additional regulation cannot be ruled out.
We predict that LapG cleaves LapA in the periplasm (Figure 8). This prediction would require the N-terminus of LapA to span the outer membrane, a possibility that has yet to be investigated. LapA contains RTX motifs, which, in other proteins, can mediate interaction with and insertion into membranes [30] lending some credence to this idea. The LapG cleavage site appears to be conserved in the N-termini of other putative adhesins (Figures 3B,S1) suggesting that adhesin modification is a conserved function of LapG homologs. Additional bioinformatic analyses indicate that LapD and LapG homologs are co-conserved in putative operons, near ABC transporters and their substrates, indicating that this effector system is likely to regulate adhesin localization in many other bacteria (in [26]). A recent study on a homologous Lap system in Pseudomonas putida presents genetic evidence in support of this hypothesis [31].
Here we observe that activation of LapG's protease activity under low Pi conditions leads to dissolution of established P. fluorescens biofilms. Pho regulon induction in planktonic cells also inhibits their ability to initiate biofilms, likely due to release of LapA from the cell surface [25]. This deficiency does not impact a cell's propensity to contact the surface, however. Instead loss of LapA specifically blocks the transition from a reversible association to more stable, “irreversible” attachment [25]. Our data put c-di-GMP signaling through the LapD-LapG system at the crux of this regulatory step. The extreme phenotypes that can result from mutations to LapD, ranging from biofilm defective to hyper-adherent (Figure 5A), suggest that regulation of LapA localization by LapD-LapG sets an equilibrium between stable attachment and detachment. Loosely attached cells receiving signals that an environment is favorable may accumulate enough c-di-GMP to inhibit LapG, and initiate and maintain stable attachment via LapA. Cells that do not receive favorable signals, or firmly attached cells that sense environmental/nutritional cues that “life” is getting worse can activate LapG, allowing the cell to pick up and leave. The involvement of the LapD-LapG system in regulating both attachment to and detachment from surfaces is unique among described biofilm pathways. Whether the intrinsic reversibility of this system is common to other c-di-GMP signaling systems that regulate biofilm formation remains to be seen.
The Supporting Information section includes additional materials and methods information (Text S1).
Strains and plasmids were constructed using standard molecular biology techniques and are listed in Table 1. Oligonucleotides used in this study are listed in Text S1. Detailed descriptions of strain and plasmid construction procedures can be found in the Supporting Information (Text S1).
Strains were grown statically for 6 h in K10T-1 (high Pi) medium, and biomass was stained with 0.1% crystal violet and quantified as described [25]. Data presented are means ± standard deviation (SD), n = 12, unless noted otherwise. For microscopy of surface attachment, strains were grown in K10T-1, and the air liquid interface imaged by phase contrast microscopy. Percent surface coverage was estimated by density measurements of digital images using ImageJ software (NIH.gov). A detailed description of the imaging and analysis procedure is in Supporting Information (Text S1). To analyze the effects of Pi starvation on biofilms, low Pi medium (K10T-π) medium was used [25].
For Western and dot blots of the LapA protein, we utilized strains with an internal 3xHA tag in chromosomal lapA [25]. Overnight cultures were diluted 1∶75 into K10T-1, grown for 6 h at 30°C, shaking at 230 rpm. Preparation and analysis of samples for LapA localization were performed as described previously [25]. Detection of cell surface LapA by dot blot was performed on aliquots of whole cells from the same cultures grown for LapA localization; blotting and quantification were performed as described [17]. In experiments monitoring the effects of Pi starvation on LapA localization, cultures were grown in high and low Pi media for 6.5 h.
Cultures were grown in the same manner as for LapA localization. The periplasmic fraction was obtained by incubation in osmotic shock buffer (50 mM Tris pH8, 20% sucrose, 2 mM EDTA) for 20 min at RT, followed by 10 min centrifugation at 15,000× g to pellet spheroplasts. For tracking the effect of c-di-GMP on LapG, periplasmic fractions were not prepared. Instead, clarified cell lysates were separated into soluble and membrane fractions by ultracentrifugation, 1 h at 100,000× g, and inner membranes were isolated by solubilization in 1% sarkosyl as described [24].
Purification of the histidine-tagged LapG, LapG-C135A, and N-Term-LapA from E. coli was performed using standard Nickel affinity chromatography techniques, as described [32].
Cell extract preparation. Bacterial cultures were grown in the same manner as for LapA localization. Clarified cell extracts were prepared by sonication (4×10 s on ice) in resuspension buffer, followed by centrifugation 12 min at 15,000× g.
Activity assays with Mini-LapA and N-Term-LapA. To assess cleavage of Mini-LapA, cell extracts from ΔlapG pMini-LapA were mixed 1∶1 with cell extracts from strains with and without LapG to test their activity, and incubated at RT for 30 min. Activity assays with purified protein were performed in resuspension buffer: 750 ng N-Term-LapA (∼30 pmol; est. 95% pure) were incubated with 750 ng of LapG (est. purity: 50%) in 37.5 ul, at RT, for 15–120 min. In assays testing the site specificity of LapG, neither N-Term-LapA AA-RR nor N-term LapA were purified. Instead cleavage by endogenous LapG was assayed in cell extracts.
Inhibition by c-di-GMP. Chemically pure c-di-GMP (GLSynthesis Inc.) was added at various concentrations to identical aliquots of resuspended cells prior to sonication for cell extract preparation. LapG activity in cell extracts was assessed by addition of 30 pmol of purified N-Term-LapA, and incubation for 100 min at RT.
Proteins were precipitated from clarified lysates prepared in the same manner as for LapA localization. Immunoprecipitations, the lysis buffer contained 20 mM Tris pH 8, 10 mM MgCl2, and 0.8% Thesit (Sigma). The same buffer was used for nickel resin pull downs, with the addition of 10 mM imidazole. Each immunoprecipitation contained 400 µl lysate, 40 µl Protein A sepharose (Genscript), and 0.5 µl monoclonal, mouse anti-HA 11.1 antibody (Covance). Each nickel resin precipitation contained 400 µl lysate, and 40 µl resin. After incubating pull downs at 4°C for 90 min, the nickel resin (Invitrogen) was washed 2×5 min at RT with gentle shaking, then a third time briefly prior to SDS-PAGE. Pull downs with c-di-GMP added were washed with buffer containing the same concentration(s) of c-di-GMP.
Edman degradation was performed by the Dartmouth College Proteomics Core. Details on sample preparation are included in the Supporting Information (Text S1).
|
10.1371/journal.ppat.1004241 | Innate Immune Responses and Rapid Control of Inflammation in African Green Monkeys Treated or Not with Interferon-Alpha during Primary SIVagm Infection | Chronic immune activation (IA) is considered as the driving force of CD4+ T cell depletion and AIDS. Fundamental clues in the mechanisms that regulate IA could lie in natural hosts of SIV, such as African green monkeys (AGMs). Here we investigated the role of innate immune cells and IFN-α in the control of IA in AGMs. AGMs displayed significant NK cell activation upon SIVagm infection, which was correlated with the levels of IFN-α. Moreover, we detected cytotoxic NK cells in lymph nodes during the early acute phase of SIVagm infection. Both plasmacytoid and myeloid dendritic cell (pDC and mDC) homing receptors were increased, but the maturation of mDCs, in particular of CD16+ mDCs, was more important than that of pDCs. Monitoring of 15 cytokines showed that those, which are known to be increased early in HIV-1/SIVmac pathogenic infections, such as IL-15, IFN-α, MCP-1 and CXCL10/IP-10, were significantly increased in AGMs as well. In contrast, cytokines generally induced in the later stage of acute pathogenic infection, such as IL-6, IL-18 and TNF-α, were less or not increased, suggesting an early control of IA. We then treated AGMs daily with high doses of IFN-α from day 9 to 24 post-infection. No impact was observed on the activation or maturation profiles of mDCs, pDCs and NK cells. There was also no major difference in T cell activation or interferon-stimulated gene (ISG) expression profiles and no sign of disease progression. Thus, even after administration of high levels of IFN-α during acute infection, AGMs were still able to control IA, showing that IA control is independent of IFN-α levels. This suggests that the sustained ISG expression and IA in HIV/SIVmac infections involves non-IFN-α products.
| Chronic inflammation is considered as directly involved in AIDS pathogenesis. The role of IFN-α as a driving force of chronic inflammation is under debate. Natural hosts of SIV, such as African green monkeys (AGMs), avoid chronic inflammation. We show for the first time that NK cells are strongly activated during acute SIVagm infection. This further demonstrates that AGMs mount a strong early innate immune response. Myeloid and plasmacytoid dendritic cells (mDCs and pDCs) homed to lymph nodes; however mDCs showed a stronger maturation profile than pDCs. Monitoring of cytokine profiles in plasma suggests that the control of inflammation in AGMs is starting earlier than previously considered, weeks before the end of the acute infection. We tested whether the capacity to control inflammation depends on the levels of IFN-α produced. When treated with high doses of IFN-α during acute SIVagm infection, AGMs did not show increase of immune activation or signs of disease progression. Our study provides evidence that the control of inflammation in SIVagm infection is not the consequence of weaker IFN-α levels. These data indicate that the sustained interferon-stimulated gene induction and chronic inflammation in HIV/SIVmac infections is driven by factors other than IFN-α.
| Chronic immune activation during HIV infection is considered as the main driver of CD4+ T cell depletion and AIDS, and early T cell activation is a better predictor of the outcome of the infection than viral load [1]. Recent observations suggest that inflammation is even more important than T cell activation to predict disease progression and mortality [2], [3]. Already in the acute primary phase of HIV-1 infection, the levels of soluble inflammatory mediators, such as IP-10 (CXCL10), were predictive of disease progression [4], [5].
Type I IFN (IFN-I), such as IFN-α, is an important component of innate immunity providing a first-line defense to viral infections, as well as bridging the innate and adaptive immune systems. This cytokine is mainly produced by plasmacytoid dendritic cells (pDCs) in viral infections. These cells interact with myeloid dendritic cells (mDCs), NK cells, monocytes, T and B cells and contribute to the orchestration of the immune response. IFN-α production is critical for the activation of NK cells, enhancing IFN-γ secretion and their cytotoxicity. Reciprocally, NK cells can affect pDC maturation and function [6]. Thus, upon infection, a crosstalk is engaged between NK cells, pDCs and mDCs, an interplay that involves IFN-I activity coupled with the release of other soluble factors [7].
Upon recognizing HIV-1, pDCs become activated, secreting high amounts of IFN-α and inflammatory cytokines, such as TNF-α [8]. This leads to bystander maturation of mDCs [9]. Both pDCs and mDCs are reduced in number and function in HIV-1 infected individuals in the circulation [10]. PDCs have been shown to migrate to lymph nodes (LNs), gut and spleen and accumulate there [11]–[14]. As a matter of fact, the diminished responses seen in disease progressors might be explained by pDC exhaustion or trafficking to tissues [13], [15]. Moreover, a defect in the pDC-NK cell cross-talk, due in large part to impaired NK cell responsiveness to IFN-α, has been described in HIV-1 infection [16], [17]. Still the role of IFN-α in HIV infection is controversial. On the one hand, IFN-α may delay disease progression by inhibiting viral replication through the induction of cellular restriction factors and by stimulating various components of the immune response involved in the control of HIV [18], [19]. A beneficial effect of IFN-α is also suggested by the observation of higher levels of pDCs and IFN-α production by TLR9-stimulated pDCs in HIV-infected long-term non-progressors [20]. On the other hand, IFN-α levels and type I interferon-stimulated gene (ISG) are markedly increased and sustained in progressors as compared to long-term non-progressors [21], [22]. Indeed, in HIV-untreated patients, high levels of ISG, such as IP-10, were associated with a more rapid CD4+ T cell depletion [4], [23]. Thus, it has been suggested that IFN-α might exert deleterious effects through various mechanisms. It could fuel chronic immune activation by the induction of ISGs including chemokines able of attracting target cells to the site of viral replication [24]. It could also stimulate innate immune cells, such as NK cells, which will in turn produce cytokines (IFN-γ,…) and chemokines, and indirectly contribute to the activation of other cell types. Moreover, the up-regulation of the ISG TRAIL may induce apoptosis of uninfected CD4+ T lymphocytes [25]. Chronic high levels of IFN-α could also induce defects in the thymopoiesis and bias in T cell selection, thereby accelerating disease progression [26].
Fundamental clues regarding the role of inflammation in AIDS and the mechanisms that protect against it may lie in natural hosts of SIV, such as African green monkeys (AGMs) and sooty mangabeys (SMs), which are asymptomatic carriers of SIV [27], [28]. This protection against AIDS is seen despite virus replication levels in blood and gut similar to HIV-1 infected humans and SIVmac-infected macaques [29]. It is associated with an absence of chronic immune activation, lacking both chronic T cell activation and chronic inflammation [27], [30]–[32]. This is not due to ignorance of the virus or to a functional defect of pDCs in sensing the virus [33]–[36]. Indeed, a vigorous innate immune response is triggered upon infection [34], [36]–[40]. Thus, the acute phase of SIVagm infection is characterized by the recruitment of pDCs to LNs, IFN-α production, induction of ISG and corresponding protein (ISP) expression [34], [36]–[40]. The levels of ISP strongly correlated with IFN-α levels during the acute phase of SIVagm infection [36]. However, there are major differences as compared to SIVmac infection: the levels of IFN-α produced in blood and LN were lower than those observed in SIVmac infection [35], [36], [38]. Moreover, in some reports, most cytokines were produced only to moderate levels in natural hosts and several pro-inflammatory cytokines were not induced at all, in contrast to the cytokine storm seen during pathogenic HIV-1/SIVmac infections [30], [35], [36], [38], [41]–[43]. Finally, ISGs, cytokines and T cell activation are down-regulated by the end of the acute phase in natural hosts and maintained as such. Thus, while in HIV/SIVmac pathogenic infections, immune activation persists, in natural hosts there are mechanisms that either prevent the onset of sustained inflammation or mechanisms that rapidly and efficiently turn them off.
In this report, we investigated the effect of SIVagm infection in AGM on innate immune cell compartments, in particular pDCs, mDCs and NK cells, and tested whether exogenous administration of IFN-α would modify the development of antiviral responses, promote chronic inflammation and/or alter clinical parameters.
The innate immune responses were followed in six SIVagm.sab92018-infected AGMs between days 2 and 547 post-infection (pi). We analyzed both blood and LNs. It is crucial indeed to study LNs because these are the sites where T and B cell responses are induced, shaped, and regulated and where correlates of protection were identified [44], [45]. Consistent with previous reports, AGMs displayed high levels of SIV replication with a peak on day 9 pi coinciding with a transient decline in CD4+ T cell levels (Figure 1A and B) [30], [36], [46]. We monitored T cell proliferation and confirmed that the primary phase of SIVagm infection in AGMs is associated with a transient increase in the percentages of Ki-67+ T cells in blood and LNs (Figure 1D and E) [30]. The peak of Ki-67+ CD4+ T cells was observed between day 7 and 9 pi (at day 9, p = 0.031), while the percentage of Ki-67+ CD8+ T cells reached a plateau on day 11 pi in blood (p = 0.008) and a peak on day 25 pi in LN (p = 0.016). The Ki-67+ CD4+ and DN T cell frequencies subsequently decreased on day 11 and that of Ki-67+ CD8+ T cells after day 31 pi.
To better understand the trafficking and function of mDCs, pDCs and NK cells in AGMs, we investigated the early changes in activation, maturation, function and homing markers of these cells in blood and LNs (Figures 2, 3 and 4, respectively). The gating strategy used for flow cytometry analysis is depicted in Figure S1. We first confirmed previous data on pDC and mDC dynamics during SIVagm infection (data not shown) [38], [47], [48]. We then studied two homing receptors for DCs: the homing inflammatory chemokine receptor CXCR3, which is the receptor for CXCL9, IP-10 and CXCL11 and CCR7, which is a receptor for chemokines that are expressed constitutively in secondary lymphoid organs. In line with the increase of mDC frequency in LNs, expression of CXCR3 increased on these cells in blood and LNs during acute infection (Figure 2A and B). Moreover, the percentage of CCR7+ mDCs was transiently increased in blood (p = 0.039 at day 9 pi) (not shown) while CCR7 levels on mDC surface did not increase (Figure 2C and D). For pDCs, the expression levels of CCR7 were significantly increased up to day 14 pi, while the CXCR3 levels were not increased (Figure 3A–D). Hence, CXCR3 and CCR7 showed opposite expression profiles on pDCs and mDCs. Still, both mDCs and pDCs showed an increase expression of one homing marker, concomitant with their increases in LNs [38], [47], [48]
MDCs showed up-regulations of the maturation markers CD80 and CD86 in blood and LN at early time points of primary infection (in blood: p = 0.008 at day 4 pi for CD86 and p = 0.008 at day 2 pi for CD80, in LNs: p = 0.031 at day 9 for CD80) (Figure 2F, G, I and J). In contrast, the expression of CD86 was not modulated on pDCs (Figure 3E and F).
It was surprising to see this discrepancy in maturation profiles between mDCs and pDCs. To confirm these findings, we analyzed the maturation profiles of mDCs and pDCs in the blood of another group of 8 AGMs infected with SIVagm. In this group, we further distinguished CD16+ mDCs (inflammatory) and CD16− mDCs (Figure 2E, H and K). In addition to the maturation markers CD80 and CD86, we also measured HLA-DR expression. The two CD16+ and CD16− mDC subsets were present in similar frequencies in blood (not shown). Both mDC subsets displayed increases in the expression of CD80, CD86 and HLA-DR. These maturation markers were more significantly increased on the CD16+ than on the CD16− subset (Figure 2E, H and K).
We confirmed the pDC phenotype in these 8 additional animals by staining for BDCA-2 (Figure 3G–H). We choose to follow HLA-DR and CD40 as these markers are well known to be up-regulated when pDCs mature and CD40 expression is increased on pDCs in SIV/HIV pathogenic infection [49], [50]. On AGM pDCs, the expression of HLA-DR was significantly down-regulated during the acute phase and the expression of the activation/maturation marker CD40 was not modulated (Figure 3G and H). The expression of CCR7 was transiently increased (p = 0.031 at day 4 pi) in this group of AGM too (not shown). These analyses thus confirm that mDCs show a more pronounced maturation profile than pDCs during SIVagm infection.
The two main functional NK cell subsets (cytolytic versus cytokine producers) were analyzed. These two subsets were differentiated based on the expression of CD16, the CD16+ subset being the predominant one in blood, as in humans and macaques (Figure 4A). Similar to cynomolgus macaques, the CD56 marker cannot be used in AGM to differentiate the NK cell subsets [51]. Thus, NK cells were defined as CD3− CD20− HLA-DR− CD8α+ NKG2A+ CD16+/− (Figure S1B and C), as in other studies on NK cells from macaques and SMs [39], [52]. A significant transient decline of both subsets was observed in blood (p = 0.031 at day 2 pi) (Figure 4A). NK cell numbers then progressively increased to reach 249% of the pre-infection levels at the end of primary infection (day 25–31 pi) for the major CD16+ subset, and 154% for the CD16− subset. They returned to baseline levels in the chronic phase (not shown). In LNs, only few NK cells were detectable and most corresponded to the CD16− subset, similar to humans [53]. A significant decrease in the percentage of the CD16− subset in LNs was observed (Figure 4B). The levels of CD16+ cells in LNs were too low to be followed. Thus, CD16+ NK cells in the blood exhibited maximal increase at the time of transition between acute and chronic phase, similar to what has been observed in SMs [39].
We monitored the activation profiles of CD16+ and CD16− NK cells in blood and of the CD16− NK cells in LNs. As shown in Figure 4C and E, the frequencies of Ki-67+ and CD69+ NK cells were markedly enhanced upon SIVagm infection in blood with a peak on day 11 pi. The activation profiles of CD16− NK cells in blood followed a similar kinetic than that of CD16+ NK cells (not shown). The percentage of activated NK cells also highly increased in the LNs (Figure 4D and F).
To evaluate NK cell function, the surface expression of CD107a (a surrogate marker for cytolytic function) and intracellular expression of IFN-γ (cytokine production) were measured (Figure 4G–J). The NK cell cytolytic activity was significantly increased only in LNs (Figure 4G and H) and no significant increase of IFN-γ production was observed in either blood or LNs (Figure 4I and J).
NK activation in blood and LNs was correlated with viral replication (blood CD16+Ki-67%: Rs = 0.37, p<0.001; LN CD16−Ki-67%: Rs = 0.54, p = 0.002; blood CD16+CD69%: Rs = 0.4, p<0.001; LN CD16−CD69%: Rs = 0.53, p = 0.002). The NK cell cytolytic activity in LNs was also correlated with viral load (CD16−CD107a%: Rs = 0.54, p = 0.003). Thus, in SIVagm primary infection, NK cells were strongly activated and cytolytic NK cells increased in LNs. These increases were positively associated with viremia levels.
Both NK cell activation and cytotoxic activity are stimulated by IFN-I, which is driven by virus. IL-15 plays a pivotal role in the development, survival and function of NK cells. We quantified IFN-I and IL-15 concentrations in blood and tissues. In line with previous reports for SIVagm infection [36], [38], [42], the IFN-α levels in plasma were transiently increased during primary infection. In addition, we reveal an increase of IL-15 production. The animals displayed two peaks of IFN-α and IL-15 production, on days 2 and 9 pi, day 9 corresponding to the peak of plasma viremia (Figure 5A and C). By day 11 pi, these levels were already decreased and below detection limit after day 14 pi for IFN-α. IFN-α and IL-15 were also measured in LNs ex vivo by collection of supernatants from the LN cell preparations (Figure 5B and D). The IFN-α concentrations in these supernatants were increased between days 2 and 11 pi. The limited number of LNs that could be collected did not allow for the same close monitoring frequency as in blood and it is unclear whether two peaks of expression were present in the LN compartment as well.
We found that NK cell activation in blood was correlated with the IFN-α levels (CD69%: Rs = 0.39, p<0.001; Ki-67%: Rs = 0.24, p = 0.008) and the IL-15 levels in plasma (CD69%: Rs = 0.25, p = 0.009; Ki-67%: Rs = 0.28, p = 0.003). At the level of LNs, IFN-α was correlated with NK cell activation (CD69%: Rs = 0.35, p = 0.032; Ki-67%: Rs = 0.41, p = 0.012) and cytotoxicity (CD107a%: Rs = 0.5, p = 0.002), while IL-15 was only correlated with NK cell proliferation (Ki-67%: Rs = 0.43, p = 0.007). The IFN-α levels also correlated with Ki-67+ CD4+ T cell levels (Rs = 0.3, p = 0.002). Altogether, IFN-α and IL-15 correlated with NK activation and IFN-α with NK cytotoxic activity in LNs.
We quantified thirteen additional cytokines in the plasma for the two AGM groups (Figure 5 and S2) to determine the earliest kinetics of cytokines and search for differences with the cytokine storm reported in HIV-1 and SIVmac infections. Cytokines reported to be increased early during HIV-1 infection were selected, such as MCP-1 as well as “innate” cytokines, such as IL-12. The early collection time points were chosen at very short intervals starting at 6 hours pi. Among the 15 cytokines studied in total, 8 were significantly up-regulated and 6 displayed a first peak on day 2 as well as a second increase on day 7 and/or 9 pi: IL-15, IFN-α, IP-10, MCP-1, IFN-γ and IL-18 (Figure 5 and S2). IL-15, IP-10 and MCP-1 are inducible by IFN. Their profiles strongly correlated with IFN-α levels (IL-15: Rs = 0.53, p<0.001; IP-10: Rs = 0.73, p<0.001; MCP-1: Rs = 0.6, p<0.001) (Figure 5). IL-8 was modestly increased at day 9 pi, while IL-12 was up-regulated only later on, on days 14 and 28 pi and even downregulated at time points just after the IFN-α peaks (Figure S2). This might be due to the fact that IL-12 is inhibited by IFN-α [54]. As a matter of fact, previous reports in SIVmac and HIV-1 infection showed that IL-12 levels increase late in primary infection, once the IFN-α levels decrease [43], [55]. Strikingly, in SIV-infected AGMs, most of the other pro- and anti-inflammatory proteins and ISPs measured (IL-6, sTRAIL, TNF-α, IL-17, TGF-β) were not modulated (Figure S2).
We wondered whether the first peak (day 2 pi) is specific for natural hosts. For those cytokines, which showed increases already on day 2 pi in AGMs, we also measured cytokines in two rhesus macaques infected with SIVmac251 (Figure S3). Also in these monkeys, a peak of cytokine production was observed on day 2 pi depending on the cytokine and the animal studied, suggesting that the early peak is not unique to AGMs. Most of studies conducted so far don't include such early time points. However, a similar early induction (day 2–4 pi) of IFN-α and some ISGs has been observed at mucosal sites of orally infected macaques [56]. We had already noticed such an early peak in previous studies [36], [38]. The AGMs were infected here with a purified virus which therefore excludes that the early peak is due to contaminants in the inoculum. Finally, the data confirm previous reports showing that IFN-α levels are lower in acutely infected AGMs compared to macaques (Figure S3B) [35], [36], [38].
Altogether, the close monitoring of fifteen soluble factors showed that cytokines, which are known to be produced early during SIVmac infection in macaques or HIV-1 infection in humans, i.e. before, during or shortly after the viral peak, were all induced in SIVagm infection. In contrast, cytokines that are induced late during the acute phase of pathogenic infection were not or only moderately induced in AGMs.
We tested whether the lower levels of IFN-α in SIVagm infection might dictate the outcome of infection, in particular with respect to the resolution of the inflammation. We therefore administered high doses of recombinant IFN-α (r-mamu-IFN-α) during the acute phase of SIVagm infection in an attempt to perturb the control of inflammation and abolish its resolution, which would be characterized by uncontrolled expression of ISGs and chronic immune activation.
The r-mamu-IFN-α used for the in vivo treatment was first tested for its efficacy on AGM cells in vitro and in vivo (Figure 6A–D and F). The same cytokine has been previously used in SMs without inducing any anti-IFN-α antibodies [57]. AGM PBMCs exposed to r-mamu-IFN-α in vitro up-regulated the expression of ISGs, such as Mx1 or IP-10, to similar levels as in macaque PBMCs (Figure 6A). Low doses of the r-mamu-IFN-α were already highly efficient for ISG induction, in line with previous data [36]. After a single in vivo injection of 5×105 IU of r-mamu-IFN-α, high levels of IFN-α were observed in plasma 1 hour post-treatment (Figure 6B), leading to strong up-regulation of ISGs, such as IP-10 (Figure 6C). Since the half-life of a similar human recombinant IFN-α has been estimated at 2–5 h in AGMs in vivo [58], this could explain why the levels of IFN-α and IP-10 mRNA were already low at 24 h after administration despite the r-mamu-IFN-α being an IgG fusion protein [57], [59]. In order to maintain robust levels of IFN-α and constantly high expression of ISGs, we injected r-mamu-IFN-α daily with a 10% increment every 2 days for 16 days. The safety and efficacy of such treatment were verified on a SIV-chronically-infected AGM. The latter displayed a 2 log10 decrease of the chronic viral load during the treatment (Figure 6D), but no major difference in T cell activation (Figure 6F), similar to data reported for chronically SIV-infected SMs [57]. Anti-IFN-α antibodies were not detected at any time point during or after treatment (data not shown).
Since r-mamu-IFN-α was efficient in cells from uninfected and chronically infected AGMs, we then tested whether such treatment would affect the resolution of immune activation during primary infection. The treatment was started on day 9 pi because after day 9 pi, endogenous IFN-α levels started to decrease, concomitantly with the diminution of other cytokines and ISPs such as IP-10 and MCP-1 and the decrease of activated NK cells and Ki-67+ CD4+ T cells. Also, based on data from the literature, we could not exclude that an initial inflammation during the first week of infection is necessary to establish infection. Finally, we did not want to interfere with the efficacy of the initial viral replication. Two AGMs were injected daily with r-mamu-IFN-α between day 9 and 24 pi. The virological and immunological profiles in the IFN-α-treated AGMs were compared to those of the 6 infected but untreated AGM (Figures 6E, 6G, 6H, 7, S3 and S4).
The administration of r-mamu-IFN-α during the acute phase of infection had no major effect on viral load (Figure 6E), even if a slight but not significant decrease was observed as compared to untreated animals at the first time points after treatment. Body temperatures were elevated during the treatment period with IFN-α (Figure S5). The expression of the ISGs CXCL9, IP-10 and CXCL11 were however comparable between treated and untreated AGMs in both PBMCs and LN cells (Figure 7). The treatment also did not result in a persistent T cell activation (Figure 6G and H) or CD4+ T cell loss over time (data not shown). As IFN-α is known to exert direct and indirect effects on innate immune cells, we also investigated whether in the absence of a change in disease outcome, the IFN-α treatment would still have had an impact on NK cells, mDCs and pDCs (Figure S4). The administration of r-mamu-IFN-α in the context of the SIV primary infection did not affect their frequencies and did not induce an increase of maturation or activation of these innate immune cells.
In summary, in spite of daily administration of high doses of IFN-α post peak of SIV replication, AGMs were still able to resolve inflammation and immune activation.
We aimed to study if the lower levels of IFN-α described during SIVagm infection as compared to SIVmac infections matter in the resolution of the inflammation in AGMs. The early host immune responses are an essential factor in determining the subsequent clinical course of disease. In mice, an early innate alteration significantly compromises the following immune responses and the host's ability to counteract the virus/parasite spread [60], [61]. We tested here whether by artificially increasing IFN-α related inflammation during the acute phase of SIVagm infection, one can overcome the intrinsic control of immune activation in this natural host. In order to study the control of immune activation and not interfere with the establishment of viral infection, the treatment was administered between the plasma viral peak and the end of the acute phase of infection. Surprisingly, the treatment did not affect viral dynamics, control of inflammation or T cell activation. This is not due to a lack of sensitivity of AGM cells to the recombinant IFN-α used here. Indeed, the r-mamu-IFN-α molecule was functional in vitro and in vivo in healthy and chronically-infected AGMs. Moreover, when administered during the chronic phase of infection, our results paralleled those described in chronically-infected SMs treated with the same molecule, namely a reduction in viral load and increase in ISG expression in the absence of major increases of T cell activation [57]. Although the number of animals was low, the analyses show that r-mamu-IFN-α was fully active on AGM cells.
It is possible that the lack of changes after IFN-α treatment during primary SIVagm infection was due to tolerance to the injected IFN-α. In SMs chronically infected with SIVsmm and treated with IFN-α, the effect of IFN-α was transient likely due to the induction of tolerance to such treatment as also reported in humans [62], [63]. Here, the treatment was short, but refractoriness could have been induced by the previous response to endogenous high levels of IFN-α. Of note, we treated the animals starting from day 9 pi, corresponding to the peak of endogenous IFN-α production. Had we started the treatment on the day of infection, we cannot exclude that we might have seen an effect on ISGs or viral load. However, such protocol might have lowered the initial viral replication, which was in opposition with the aim of our study. Altogether, administration of IFN-α in the mid and late part of the acute phase did not change the outcome, suggesting that the resolution of inflammation in AGMs is not due to a difference in the levels of IFN-α production during primary infection. It has been suggested that the combination of antiretroviral therapy and interferon given during acute HIV infection may potentiate both innate and adaptive immune responses against HIV replication and/or reservoir levels [64]. Our study shows that in primary infection, IFN-α, when administered after the peak of viremia, does not affect viral replication or innate responses. It reduces viral load during chronic phase. Whether this is true for pathogenic infection, remains to be determined. However, the timing is very important and should be considered when such treatment is envisaged.
We previously showed that during the acute phase of SIV infection the levels of ISGs, such as of IP-10, strictly correlate with IFN-α levels [36]. Here, treatment with high doses of IFN-α did not lead to sustained ISG expression. This suggests that, at the transition of acute to chronic phase other factors than IFN-α predominantly drive ISG expressions in macaques and humans. Elevated expression levels of ISGs in chronic infection are associated with uncontrolled viremia and disease progression [23], [65]. It could be that not the IFN-I production, but constant ISG expression are deleterious for the host. IP-10 has been reported to be an excellent marker of inflammation and disease progression [4], [65]–[68]. IP-10 is inducible not only by IFN-I and IFN-γ, but also by other pro-inflammatory cytokines (TNF-α, IL-1β, IL-18) [69]–[71]. Hence, it is possible that IFN-α alone is not sufficient, but that a combination with other factors, TNF-α for example, is also required to induce ISG expression. Moreover, even though ISGs are IFN-inducible, some were shown to be directly up-regulated through recognition of viral or bacterial products by pattern-recognition receptors in an IFN-independent manner [72]–[75]. Finally, an expansion of the enteric virome and microbial translocation are observed in chronic HIV-1 and SIVmac infections [76], [77]. It could thus explain why macaques and humans maintain ISG expression but not AGMs who do not display microbial translocation or virome expansion. Other or additional factors might also play a role in the maintenance of ISG expressions during HIV-1/SIVmac infections. For instance, distinct IFN-α subtypes differently induce ISG expressions in vivo, while here, only IFN-α2, which is considered the most abundant in viral infections, was used [69], [78]–[80].
Altogether, our study indicates that the non-pathogenic outcome of SIVagm infection is not due to differences in IFN-α levels between AGM and macaques or humans. It does not exclude that the difference in outcome is related to different levels of ISG expression. It indicates however, that the mechanisms, which maintain high levels of ISG expression, are due to other or additional factors than IFN-α.
Several studies have debated whether natural hosts display lower or similar immune activation levels during primary infection as compared to pathogenic infections. Some studies have reported weaker levels of T cell activation and cytokine concentrations during the acute phase of SIVagm or SIVsmm infection, whereas in other studies the levels were equivalent to those observed in SIVmac-infected macaques [30], [36], [41], [81]. We performed a detailed follow-up of T cell activation and, to attempt to reconcile the discrepant cytokine profiles, we deciphered here the early production of cytokines in AGM. We included in the study the cytokines known to be induced very early during pathogenic infection, such as IL-15 [19], [43]. Of note, the acute phase of SIVagm infection resulted in significant increases of early cytokines, including IL-15, IP-10 and MCP-1, similar to pathogenic infection. However, salient differences were observed for cytokines known to be produced in later stages of the acute phase of HIV-1/SIVmac infections. They were either not or only weakly induced in AGMs. In particular, IL-6 and TNF-α were not up-regulated (Figure S2). We hypothesize that the early cytokines, which are produced in AGMs during the first two weeks pi, confer a benefit to both the virus and the host. It would be beneficial to the virus as inflammation attracts target cells to the sites of infection. For the host, the induction of early innate responses (restriction factors, NK cells, mDCs), would allow the development of antiviral innate and adaptive responses for partial control of viral replication. AGMs might have found a way to allow early inflammation resulting in productive infection while blocking the cytokine storm that takes place following the viral peak. This would avoid the sustained inflammatory environment.
The dual pattern of cytokines that we observed might be explained by a differential susceptibility to activation by the innate cells. While pDCs show a normal sensing of SIVagm [36], [37], [40] leading to the production of IFN-α, other cells, for instance myeloid cells, such as mDC and macrophages, might not produce any cytokine. Indeed, a recent study reported that in contrast to SIVmac and HIV-1 infections, mDCs mature but do not show spontaneous production of pro-inflammatory cytokines such as TNF-α in primary SIVagm infection [48].
To understand what might be the key events of the innate response in natural hosts that allows them to maintain the inflammation under control, we investigated the effect of SIVagm infection in AGM on innate immune cell compartments, in particular pDCs, mDCs and NK cells. Little is known about those decisive early cellular responses in AGMs, in particular regarding pDC maturation and NK cell activation. In addition, for the first time the activation profiles of these three types of innate cells were analyzed concomitantly in the same animals. We analyzed the maturation and homing patterns of two sub-populations of mDCs, the CD16− and CD16+ subsets. These correspond to two major subsets of mDCs in humans [82], [83]. The CD16+ mDCs displayed higher levels of activation and maturation than the CD16− subpopulation. Whether these cells play a distinct role in T cell activation or tolerance is unclear. PDCs displayed lower levels of maturation than mDCs. IFN-α production by pDCs is associated with their maturation stage. It has been shown that HIV skews pDCs toward a partially matured and persistently IFN-α-secreting phenotype which allows their survival [84]. Eventually, the partial maturation of pDCs in AGMs might be associated with their capacity of efficient IFN-α production during the acute phase of SIVagm infection. Egress of pDC precursors from bone marrow could then account for the return of IFN-α levels to baseline [85]. This is supported by the decrease of HLA-DR expression on the pDC's surface. On the contrary, a preferential maturation process at the expense of cytokine secretion might be occurring at the level of mDCs in AGMs, especially in presence of IFN-I, since IFN-I induces mDC maturation rather than cytokine secretion [83].
We also analyzed NK cells for the first time in the context of SIVagm infection. We observed a strong increase in proliferation and activation of NK cells during the acute phase of SIVagm infection. Our data support the observations reported in SMs on earlier and stronger NK cell responses than in SIVmac-infected macaques [39]. The rapid and strong increase in NK cell proliferation in AGMs might be a direct consequence of the early and robust production of IL-15 and IFN-α during primary SIVagm infection. No production of IFN-γ by NK cells was observed, while NK cell cytotoxicity was induced. It has been shown that IFN-α and IL-15 promote NK cell proliferation and survival, while IFN-α is able to increase NK cell cytotoxicity, and IL-12 to augment the secretion of IFN-γ [86]. This is in accordance with the fact that modest levels of IL-12 and high levels of IFN-α were detected, playing thus a putative role in establishing such protective NK cell responses. It was surprising to detect increases in NK cytotoxic function in LNs. One hallmark of SIV infection in natural hosts is the high viral load in blood and intestinal tissues, but low viral burden in LNs in the chronic phase of infection [29], [87]–[90]. It has been suggested for SMs that the rapid and dramatic control of viral replication in LNs is associated with CD8+ T cell responses [89]. However, it is tempting to speculate that at least in the early stage of SIVagm infection in AGMs, NK cells could significantly contribute to the control of viral replication in LNs which in consequence could contribute to limit immune activation [91].
Altogether, our study provides evidence that the control of immune activation in SIVagm infection is not a consequence of lower levels of IFN-α production. We show that AGMs mount strong early innate immune responses as exemplified by the significant NK cell activation and production of early cytokines, such as IL-15 and MCP-1. Our study indicates that the sustained ISG production in HIV/SIVmac infections is likely driven by additional or factors other than IFN-α, among which could be elevated pro-inflammatory cytokine levels, enteric virome expansion and microbial translocation. The data also suggest that mechanisms controlling inflammation are in place before the transition of the acute to the chronic phase, thus earlier than previously considered. Whether this is due to the establishment of inhibitory or tolerance mechanisms after the viral peak, or to a distinct susceptibility to infection or immune activation by specific immune cell subsets, needs to be further investigated.
Animals were housed in the facilities of the CEA (“Commissariat à l'Energie Atomique”, Fontenay-aux-Roses, France) and Institut Pasteur (Paris, France) (CEA permit number: A 92-032-02, Institut Pasteur permit number: A 78-100-3). All experimental procedures were conducted in the CEA animal facility and in strict accordance with the international European guidelines 2010/63/UE about protection of animals used for experimentation and other scientific purposes (French decree 2013-118) and with the recommendations of the Weatherall report. The monitoring of the animals was under the supervision of the veterinarians in charge of the animal facilities. All efforts were made to minimize suffering, including efforts to improve housing conditions and to provide enrichment opportunities (e.g., 12∶12 light dark schedule, provision of monkey biscuits supplemented with fresh fruit and constant water access, objects to manipulate, interaction with caregivers and research staff). All procedures were performed under anesthesia using 10 mg of ketamine per kg body weight. For deeper anesthesia required for lymph node removal a mixture of ketamine and xylazine was used. Paracetamol was given after the procedure. Euthanasia was performed prior to the development of any symptoms of disease (e.g., for macaques when the biological markers indicated progression towards disease, such as significant CD4+ T cell decline and increases of viremia). Euthanasia was done by IV injection of a lethal dose of pentobarbital. The CEA is in compliance with Standards for Human Care and Use of Laboratory of the Office for Laboratory Animal Welfare (OLAW, USA) under OLAW Assurance number #A5826-01. Animal experimental protocols were approved by the Ethical Committee of Animal Experimentation (CETEA-DSV, IDF, France) (Notification number: 10-051b).
Eighteen Caribbean-origin African green monkeys (Chlorocebus sabaeus) and two Chinese rhesus macaques (Macaca mulatta) were used in the study. AGMs were infected by intravenous inoculation with 250 TCID50 of purified SIVagm.sab92018, and macaques with 5000 AID50 of SIVmac251, as previously described [90]. SIVagm.sab92018 has been purified by sucrose density gradient centrifugation and on Vivaspin 20 columns (Vivaproducts). Neither IFN-α nor endotoxin (LAL QCL-1000 Kit, Lonza) were detected in the two viral stocks. Four AGMs were treated with the r-mamu IFN-α-IgFc by subcutaneous injection (Resource for NHP Immune Reagents, Emory University, Atlanta, GA): 2 were used for the establishment of the treatment protocol and control of its efficiency in AGM, and 2 were treated during the acute phase of infection. When daily injections of 5×105 IU for over a period of 16 days were performed, the dose was increased by 10% every second day.
Whole blood was collected from all AGMs. For the initial groups of monkeys, baseline blood collections were performed at 4 to 6 time points before infection (days −30, −28, −23, −21, −19 and −16) to mimic the sampling of the acute phase and measure any difference linked to the sampling. No variation due to the sampling was observed. Blood was then collected during primary infection (on days 2, 4, 7, 9, 11, 14 and 25) and during the chronic phase (days 31, 59, 85, 122, 183, 241, 354, 456, 547 or euthanasia). AGM peripheral LNs were obtained by excision before infection (days −15 and/or −10) and after infection at the following days: 2 (3 AGMs), 7 (1 AGM), 9 (8 AGMs), 11 (7 AGMs), 14 (8 AGMs), 25 (7 AGMs) and 547 pi or euthanasia (5 AGM). In a second group of 8 AGMs, blood was collected at 3 to 4 time points before infection (days −40, −30, −20, −10), at very early time points during primary infection (6 hours post-infection and at days 1, 2, 4, 7, 9, 11, 14, 28 pi) and during the chronic phase (days 42 and 63 pi).
Panels of fluorochrome-labeled monoclonal antibodies (mAbs) that have been shown to be cross-reactive with AGMs, were used to label fresh whole blood and LN cells and were purchased from BD Biosciences unless otherwise stated: CD3 (SP34-2), CD4 (L200), CD8 (Sk1), CD20 (2H7, ebioscience), HLA-DR (L243), CD16 (3G8), NKG2A (Z199, Coulter), CD107a (H4A3), CXCR3 (1C6), CD69 (FN50, ebioscience), CD123 (7G3), BDCA-2 (AC144, Miltenyi), CD86 (FUN-1), CD40 (HB14, Caltag), CCR7 (3D12, ebioscience), CD11c (S-HCL-3), CD80 (L307.4), IFN-γ (45-15, Miltenyi), CD45 (D058-1283), Ki-67 (MIB-1, Dako) as well as isotype controls. FcR Blocking Reagent (Miltenyi) was used to block unwanted binding of antibodies and increase the staining specificity of cell surface antigens. For detection of IFN-γ and CD107a, cells were pre-incubated for 4 hours with brefeldin A and monensin at 37°C prior to labeling with surface-binding antibodies and then fixed and permeabilized prior to incubation with IFN-γ antibody. Cells were run on a BD LSR-II flow cytometer system, collected with BD FACS Diva 6.0 software, and analyzed with FlowJo 8.8.7 (TreeStar).
Cytokines were measured in plasma and LN cell supernatants. LN cell supernatant consists of the medium in which the biopsy was collected and kept for 2–3 hours at 4°C. Cells were prepared by homogenization in the same medium and the supernatant was collected after centrifugation. Titers of bioactive IFN-α were determined as previously described [38]. The same test was used to search for plasma IFN-α antibodies that might have developed in response to the treatment. The other cytokines were quantified using the following ELISA kits: MONKEY IFN-gamma, IL-6, IL-8, IL-10, IL-12/23p40, TNF-alpha (U-Cytech), Human IL-15, CXCL10/IP-10, CCL2/MCP-1, TRAIL/TNFSF10 Quantikine Kits (R&D), Human IL-17A Ready-SET-Go (eBiosciences), Human IL-18 Kit (MBL), Simian IFN-beta Kit (USCN), TGF-β1 Multispecies Kit (Invitrogen). To verify the cross-reactivity of the antibodies used in the ELISA kit for cytokines that have never been tested on AGM [30], [36], AGM PBMCs were stimulated in vitro and cytokines were measured in the supernatants (data not shown).
Plasma viral load was determined by real-time PCR [90]. Quantification of ISG transcripts was performed by real-time RT-PCR in triplicate using Taqman gene expression assays (Life technologies). The expression of each gene was normalized against that of 18S rRNA [30], [36].
To characterize each marker's progression (figures 2–4), a linear mixed effect model was used to account for multiple measurements within each AGM. Firstly, we graphically assessed that the marker's distribution was Gaussian; if not, a logarithmic transformation was used. Secondly, a LOWESS (locally weighted scatterplot smoothing) curve was used to assess whether the marker's trajectory looked linear or piecewise linear. Based on these trajectories, we introduced or not slopes. We indicated at which time-point the change of slope occurred. Finally, a mixed effect linear or piecewise-linear model was applied. When two slopes were introduced into the model, the difference between the two slopes was tested using Wald's test. The Wilcoxon matched-pairs signed rank test was used to evaluate whether there was a statistically significant difference in the level of one given marker at a given time point following inoculation when compared to the baseline medians (day 0), using Prism (GraphPad). Baseline medians in blood consisted of 3 to 6 pre-infection values per animal for the flow cytometry and gene expression analysis, and 4 total pre-infection values per animal for the plasma cytokines study. In LNs, it consisted of 1 to 2 pre-infection values per animal for all the measurements. Finally, the Spearman rank test was used to assess the correlation between 2 continuous variables.
|
10.1371/journal.pgen.1000158 | Mutant Screen Distinguishes between Residues Necessary for Light-Signal Perception and Signal Transfer by Phytochrome B | The phytochromes (phyA to phyE) are a major plant photoreceptor family that regulate a diversity of developmental processes in response to light. The N-terminal 651–amino acid domain of phyB (N651), which binds an open tetrapyrrole chromophore, acts to perceive and transduce regulatory light signals in the cell nucleus. The N651 domain comprises several subdomains: the N-terminal extension, the Per/Arnt/Sim (PAS)-like subdomain (PLD), the cGMP phosphodiesterase/adenyl cyclase/FhlA (GAF) subdomain, and the phytochrome (PHY) subdomain. To define functional roles for these subdomains, we mutagenized an Arabidopsis thaliana line expressing N651 fused in tandem to green fluorescent protein, β-glucuronidase, and a nuclear localization signal. A large-scale screen for long hypocotyl mutants identified 14 novel intragenic missense mutations in the N651 moiety. These new mutations, along with eight previously identified mutations, were distributed throughout N651, indicating that each subdomain has an important function. In vitro analysis of the spectral properties of these mutants enabled them to be classified into two principal classes: light-signal perception mutants (those with defective spectral activity), and signaling mutants (those normal in light perception but defective in intracellular signal transfer). Most spectral mutants were found in the GAF and PHY subdomains. On the other hand, the signaling mutants tend to be located in the N-terminal extension and PLD. These observations indicate that the N-terminal extension and PLD are mainly involved in signal transfer, but that the C-terminal GAF and PHY subdomains are responsible for light perception. Among the signaling mutants, R110Q, G111D, G112D, and R325K were particularly interesting. Alignment with the recently described three-dimensional structure of the PAS-GAF domain of a bacterial phytochrome suggests that these four mutations reside in the vicinity of the phytochrome light-sensing knot.
| Adapting to the light environment, plants have evolved several photoreceptors, of which the phytochromes are specialized in perceiving the red and far-red light region of the spectrum. Although phytochrome was first discovered in plants, the phytochrome species are present in several organisms, including bacteria. The mechanisms by which phytochromes transduce light signals to downstream components are most well studied in plants. Upon light activation, phytochromes translocate from the cytoplasm into nucleus and regulate the gene expression network through interaction with nuclear transcription factors. The phytochrome molecule can be divided into two major domains: the N-terminal moiety, which is responsible for the light perception, and the C-terminal moiety. Although the C-terminal moiety was though to be involved in signal transduction, it has recently been shown that the N-terminal moiety has a role not only in the light perception, but also in light signal transfer to the downstream network. However, no signaling motifs have been found in the N-terminal moiety. In this study, we analyzed intragenic mutations derived from a genetic screen and found a cluster of residues necessary for signal transduction in a small region neighboring the light-sensing chromophore moiety on the three-dimensional structure. This is an important step towards understanding how a major plant photoreceptor, phytochrome, intramolecularly processes the light signal to trigger diverse physiological responses.
| To adapt to fluctuating environmental conditions, plants obtain and interpret information from light. These light sensing processes utilize at least three classes of photoreceptors [1]–[3] of which phytochromes are well characterized with respect to molecular structure and biological function. Phytochromes are unique pigments whose function is mediated through photoreversible conformational changes between two spectrally distinct forms: an inactive red-light (R)-absorbing form (Pr) and an active far-red-light (FR)-absorbing form (Pfr). R converts Pr to Pfr, and FR converts Pfr back to Pr. In addition, Pfr is gradually converted back to Pr in darkness by a thermally driven process called “dark reversion”. In Arabidopsis the phytochrome family consists of five members [4]. Two members of the family, phytochrome A (phyA) and B (phyB) are the most important in seedling development. PhyA and phyB have different photosensory specificities. PhyA mediates de-etiolation under continuous FR (cFR), whereas phyB mediates de-etiolation under continuous R (cR) [5].
Phytochromes, which are soluble proteins, are synthesized in the Pr form and reside in the cytoplasm in darkness. Upon light activation, phytochromes translocate to the nucleus [6]–[9] where they regulate gene expression [10]–[12]. Phytochromes interact with nuclear basic helix-loop-helix proteins such as PIF3 in a light-dependent manner [13]–[15]. These interactions are thought to induce alterations in the expression of target genes [16],[17].
Phytochromes in solution exist as dimers of approximately 120 kD subunits, each of which binds a single open tetrapyrrole chromophore responsible for the absorption of visible light. Each phytochrome monomer consists of a chromophore-bearing N-terminal moiety of about 70 kD and a C-terminal moiety of about 55 kD. The N-terminal moiety is highly conserved among members of the phytochrome family. The N-terminal moiety alone can bind the chromophore and show photoreversible conformational changes. On the other hand, the C-terminal moiety is required for dimerization [18] and nuclear localization [19]. Although the C-terminal moiety had long been presumed to transduce the signal to downstream components, we have shown that the N-terminal moiety of phyB alone can transduce the signal in the nucleus in response to light stimuli [20]. The data indicate therefore, that the N-terminal moiety has not only a light perception function but also a signal transferring function.
Although phytochromes were originally discovered in plants, recent analyses have demonstrated that phytochrome-related molecules are found in various bacteria [21]. Based on sequence analysis, four domains are recognized in the N-terminal moiety of phytochromes: the N-terminal extension, the N-terminal Per/Arnt/Sim (PAS)-like domain (PLD), the cGMP phosphodiesterase/adenyl cyclase/FhlA domain (GAF), and the phytochrome domain (PHY) [21]. The N-terminal extension is found in higher plant phytochromes but not in bacteriophytochromes. GAF has bilin lyase activity and covalently binds the chromophore [22]. PHY stabilizes Pfr [23]. Although the crystal structure of plant phytochromes has not been determined yet, that of the PAS-GAF domain of Deinococcus radiodurans bacteriophytochome (DrCBD) has been determined [24]. Interestingly, an unusual three dimensional structure, designated the light sensing knot [24], is found between the PAS and GAF domains in DrCBD.
To identify regions of the protein important for signal transduction by phytochromes, several deletion derivatives have been examined for their biological activities [23], [25]–[29]. According to those studies, a phyB derivative that lacks the N-terminal 103 amino acid extension exhibits reduced but significant biological activity [29]. Similarly, the PHY subdomain is dispensable for the signaling activity [23]. Hence, the core region of phyB responsible for signal transduction activity can be narrowed down to the region composed of PLD and GAF. However, critical amino acid residues necessary for signaling have not been identified.
Mutational analyses have been adopted for the study of the phytochrome signal transduction mechanism. Several amino acid substitutions within phytochrome molecules have been identified that reduce the biological activity of the molecule without affecting either the amount of protein accumulation or the photochemical properties of the protein. Although this kind of mutational analysis led to identification of the Quail-box, which resides in the C-terminal moiety [30], it has been later shown that some of the mutations in this region impair the subcellular dynamics of phytochromes [8],[20]. On the other hand, as the N-terminal moiety retains dual functions (a light perception and a signal transferring function) the amino acid substitutions, which reduce the biological activity, within the N-terminal moiety may be expected to fall into two classes: (1) one consisting of those that are defective in photoperception and/or the maintenance of the active Pfr form, and (2) the other containing those that are normal in photoperception and the maintenance of active Pfr form, but defective in regulatory activity.
Of the above two classes, the latter class of mutations is thought to directly disrupt the signal transfer to components downstream of phyB. Although altogether 8 mutations have been reported within the N-terminal moiety of phyB [23], [31]–[34], none have been fully investigated. This may be because the signal transferring function of N-terminal moiety had not been established until the recent evidence that our engineered N-terminal moiety of phyB can complement the phyB mutation [20]. In addition, the number of mutations reported within the N-terminal moiety is too few and the distribution throughout the N-terminal moiety is too disperse to indicate regions important for signal transduction of phyB (Figure 1A, Table 1).
Here, to first identify the critical amino acid residues necessary for signal transfer, we performed a large scale genetic screen for long hypocotyl mutants under dim cR. In this screen, we mutagenized Arabidopsis thaliana expressing the engineered N-terminal moiety of phyB in order to focus on this moiety. Our data identify two classes of residues with functionally distinct roles, respectively, in photosensory perception and signal propagation to downstream targets.
The N-terminal 651 amino acid fragment of phyB (N651), fused in tandem to green fluorescent protein (GFP), β-glucuronidase and a nuclear localization signal (NLS) (N651G-GUS-NLS), is fully functional in all phyB responses examined, and exhibits hypersensitivity to cR for various phyB responses [20] except for root greening under red light [35]. To identify amino acid residues that are important for N651 function, an Arabidopsis line expressing N651G-GUS-NLS in the phyB mutant background was mutagenized with ethyl methanesulfonate (EMS), and the M2 seedlings were screened for the long hypocotyl phenotype under weak cR (0.05 µmol m−2 sec−1).
At least 1,000,000 M2 seedlings derived from 200,000 M1 plants were subjected to screening. Putative mutant lines were examined further in the M3 generation. GFP fluorescence was severely reduced in more than 90% of these lines. The lines in which GFP fluorescence was not reduced were further examined with respect to the hypocotyl phenotype under cFR. We selected 69 lines that showed the long hypocotyl phenotype only under cR. These 69 lines were crossed with the phyB mutant. Subsequent segregation analysis in the F2 generation revealed that 19 of them were linked to the N651G-GUS-NLS gene, indicating that they were intragenic mutants. Sequence analysis of these 19 lines revealed an amino acid substitution within the N651 moiety in each line. These 19 lines yielded 14 distinct substitutions representing, therefore, 14 different variants of the N651 gene (Figure 1A, Table 1). None of these mutations has been reported previously [30]–[34]. We confirmed that no mutations were found in the GFP-GUS-NLS moiety in these lines. GFP fluorescence was observed exclusively in the nucleus in each line (data not shown), verifying their expected constitutive nuclear localization.
The hypocotyl lengths of these lines compared to the phyB null mutant and the parental N651G-GUS-NLS 4-1 and N651G-GUS-NLS 3–8 lines, under two intensities of cR, are shown in Figure 1B. As we described previously [20],[23], the lower of these two intensities of cR (0.05 µmol m−2 sec−1) is already saturating for inhibition of hypocotyl elongation in these parental N651G-GUS-NLS lines, so that no difference in hypocotyl length between the two intensities was observed for these two lines. Each of the mutant lines, on the other hand, exhibited a long hypocotyl phenotype, to varying degrees compared to the N651G-GUS-NLS lines, with some displaying cR-intensity responsiveness, and others not. The hypocotyls of D64N, R110Q, G111D, P309L, and S370F lines were almost as long as those of the phyB mutant under both intensities, indicating severe or complete loss of phyB activity. The remaining nine variants showed an intermediate hypocotyl-length phenotype between the phyB parent and the N651G-GUS-NLS transgenic rescue lines. Of these, six (G248E, P304L, R313K, R322Q, V401L and S584F) showed a greater or lesser degree of reduced responsiveness to the lower compared to the higher cR intensity, whereas the remaining three (G112D, P149L and R352K) did not show such a difference in hypocotyl responsiveness to the cR intensity. We confirmed that the long hypocotyl phenotype was observed neither under cFR nor in darkness (Figure 1B).
Immunoblot blot analysis of light grown seedlings showed that the mutant-variant lines contain levels of the phyB fusion-protein similar to or higher than the parental N651G-GUS-NLS 4-1 line, in most cases (Figure 1C). Although the levels were reduced in some lines, they were still higher than that in another N651G-GUS-NLS line, 3–8 (Figure 1C), in which the full response to cR was observed (Figure 1B). Concordant results were obtained from measuring GUS activity in these lines (data not shown). These data indicate that the reduced responsiveness to cR is due to reduced intrinsic activity of the mutant phyB rather than reduced levels of expression.
In addition to the 14 mutations described above, 8 missense mutations within the N651 moiety that reduce the function of phyB have been reported previously [23], [31]–[34]. These 22 missense mutations in the N-terminal moiety of phyB are detailed in Figure 1A and Table 1. The spectral characteristics had been examined for only two of these mutations [23],[36] prompting us to examine the entire cohort for spectral integrity. Spectrally active phyB derivatives were reconstituted in vitro using phycocyanobillin (PCB) as the chromophore [37]. Wild type and mutated N651 fragments fused to intein and chitin binding domain (CBD) were expressed in E. coli and subjected to chromophore incorporation analysis [23]. The crude extracts from E.coli were mixed with phycocyanobilin (PCB) and examined by the Zn blot assay (Figure 2). The results showed that 17 mutants displayed normal PCB incorporation. Of the remaining 5 mutants, PCB incorporation was not detected in G118R and S134G, was reduced in G284E and P309L, and was markedly reduced in S370F.
We examined whether the mutations affected the spectral properties of N651. Twenty mutants that allowed chromophore incorporation (Table 1) were tested for the Pr-Pfr difference spectrum (Figure 3). The spectrum for the wild type N651 fragment exhibited an absorption maximum around 650 nm and minimum around 710 nm as previously described [23]. Of these 20 mutants, 14 mutants exhibited normal difference spectra. The remaining 6 mutants, G284E, P309L, R322Q, S370F, A372T and S584F, exhibited abnormal difference spectra. The G284E and P309L mutants exhibited a bleached spectrum in which the trough in the far-red region was much shallower compared with the peak in the red region. The S584F mutant exhibited a similar defect but to a lesser extent. In addition, a substantial blue-shift of the difference spectrum minimum was observed in this mutant. In R322Q and A372T, a red-shift of the difference spectrum maximum was observed. Conversely, a blue-shift of the difference spectrum maximum was observed in S370F.
The Pfr form of phytochrome is thermally unstable, and it spontaneously converts back to Pr in darkness by a process called ‘dark reversion’. This dark reversion is an important process to regulate the level of Pfr in vivo. Hence, we compared the dark reversion rates in the wild type and the N651 mutants (Figure 4A). Those mutants that were severely deficient in chromophore incorporation (G118R, S134G, G284E, P309L and S370F) were excluded from this analysis. As has been reported previously, the wild type N651 exhibited a relatively slow dark reversion rate, with more than 80% remaining as Pfr 1 hr after a pulse of R (pR). Eight out of the 17 mutants exhibited similar dark reversion rates to that in wild type N651 (Figure 4A, Table 1). The other 9 mutants, to various extents, exhibited an increase in the dark reversion rate. Three mutants in particular, S584F, R322Q and A372T, exhibited a very fast dark reversion rate with only 40% remaining as Pfr 1 hr after pR.
The hypocotyl response to intermittent pR depends very much on the stability of Pfr in darkness [23]. Hence, we examined how the mutant plants responded to cR and pR (Figure 4B). This was done only in the 14 mutants that were obtained in the present study (Table 1). As expected, S584F and R322Q, in which the dark reversion rates were very fast in vitro (Figure 4A), exhibited reduced responses to pR. Similar differences were observed in R313K and V401, both of which exhibited relatively fast dark reversion rates. In addition, we observed smaller but significant differences in G284E, P304L and P309L. Of these, the dark reversion rate was not measured in G284E and P309L because severe reduction in chromophore incorporation (Figure 2) and bleached difference spectra (Figure 3) were observed. Exceptionally, P304L did not exhibit any significant phenotype with respect to spectral properties in vitro.
Mutations D64N, R110Q, G111D, G112D, P149L, I208T, P304L and R352K reduced the biological activity of N651G-GUS-NLS without affecting the spectral properties in vitro (Table 1). Especially interesting are R110Q, G111D, G112D and R352K because alignment of the Arabidopsis phyB sequence with that of DrCBD (Figure 5) suggested that these residues would reside in the vicinity of the light sensing knot (for detail, see discussion). Hence, we examined the biological activities of the full-length phyB carrying these mutations in transgenic Arabidopsis.
The mutated full-length phyB-GFP fusion proteins (PBG) carrying R110Q, G111D, G112D or R352K were expressed in the phyB mutant background under the control of the cauliflower mosaic virus 35S promoter. Immunoblot analysis revealed that the expression levels were comparable to or higher than those in PBG18 (Figure 6A). The long hypocotyl phenotype under cR was observed in PBG(R110Q), PBG(G111D) and PBG(R352K) mutants (Figure 6A). Exceptionally, the phenotype was less clear in PBG(G112D). This was probably due to the residual activity in this mutant. Indeed, the phenotype was weaker in the original N651(G112D)G-GUS-NLS mutant than the other 3 mutants (Figure 1B). It is not clear why these mutations showed weaker phenotypes in PBG background, compared to N651G-GUS-NLS background (Figure 1B). This was probably because of the fact that PBG line has a higher expression level than N651G-GUS-NLS line[20]. However, it is possible that the C-terminal moiety may acquire regulatory activity in conjunction with the photoactive N-terminal moiety, despite the observation that the C-terminal moiety alone does not show any apparent biological activity [20],[29].
Based on recent reports that early and late phases of phyB-regulated seedling deetiolation may involve different modes of regulation [38]–[40], we examined the effect of the R110Q, G111D and R352K mutations in the full-length PBG molecule on the cR-induced expression of three early-response genes, ELF4 (At2g40080), SAUR-LIKE (At4g38840) and AMYLASE (At4g17090), shown previously, in time-course experiments, to be robustly phyB-dependent [11],[12],[41]. Twelve hr of cR exposure was selected for this experiment because the differential in expression between wild-type and phyB-null-mutant seedlings was found to be maximal at that time-point [11], providing maximal sensitivity for detecting reductions in cR sensitivity in our phyB-mutant variants. Although a small number of other genes had been reported to exhibit differences in expression at 1 hr of cR between the wild-type and phyB-null mutant by microarray analysis [12], none of these were found to display sufficiently robust differences at 1 hr cR by qPCR in our present analysis to permit reliable assessment of the effects of the point mutants identified here. Our data show that all three selected genes exhibit a similar pattern. Whereas the wild-type PBG sequence fully rescues the reduced cR-induced expression of the phyB mutant, all three mutant phyB variants fail to a greater or lesser extent to reinstate full induction of expression (Figure 6B). This pattern parallels the behavior of these variants in failing to complement the long-hypocotyl phenotype of the phyB mutant (Figure 6A), indicating a loss of phyB function in both early and late phases of the seedling deetiolation process.
We confirmed that the intracellular localization of PBG was not affected by these mutations (Figure 7). The wild-type PBG as well as its mutated derivatives were detected not only in the cytoplasm but also in the nucleus in most of the cells in the etiolated seedlings. After 2 min irradiation with white light, early PBG speckles [42] were observed in the nuclear region in all derivatives. After 24 hr treatment with cR, nuclear accumulation and formation of late nuclear speckles were observed in all of the lines. The normal dynamics of these mutant PBG derivatives as regards subcellular localization indicates that these mutants are normal in photoperception. We also found that PBG formed both early and late speckles even on the phyAphyB double mutant background (Figure 7). Hence, formation of both early and late speckles was independent of the phyA function.
We recently demonstrated that phyB lacking a C-terminal moiety is still capable of robustly transducing a light signal to regulate normal seedling development [20]. Those results prompted us to elucidate the structural basis of this observation. Hence, we screened for long hypocotyl mutants to identify missense mutations that reduced the biological activity of phyB within the N-terminal domain of phyB. Prior to the present work, several missense mutations had been identified in phyA and phyB [23], [30]–[34],[43]. Of these, 8 mutations reside in the N-terminal moiety of phyB (Table 1), but the consequences of these residue substitutions to the molecular functions of the photoreceptor had only been examined for two of these. In the present study we identified 14 additional missense mutations and examined them in detail for functional relevance.
To identify as many novel mutations as possible, we modified the screening procedure, compared to previous studies. First, to focus on the N-terminal moiety of phyB, we used the N651G-GUS-NLS line as a parental line for mutagenesis. Second, the seedlings were grown under dim cR, which allowed us to detect smaller reductions in activity. Combined with a large scale screening of at least 1,000,000 M2 seedlings derived from 200,000 M1 plants, we successfully identified 14 novel missense mutations within the N-terminal moiety of phyB (Figure 1A and Table 1). It remains unclear why the present set of mutants did not overlap with the known ones. This might be because the N651G-GUS-NLS line rather than the full-length phyB line was used in the present study.
The 14 mutations found in the present study, together with the 8 previously described mutations [23], [31]–[34] were characterized with respect to their spectral properties in vitro, resulting in the identification of two principal classes of defects. One consists of the spectral mutants, which are defective in chromophore incorporation, photoconversion and/or stability of Pfr. The other comprises signaling mutants, which are normal in spectral properties but defective in biological activity. 14 mutations out of the total of 22 were classified as spectral mutants and the remaining 8 as signaling mutants.
As the loss of spectral integrity directly affects the amount or overall structure of the active Pfr form of the photoreceptor, the reduced biological activity of the spectral mutants is simply explained by the low amount of or aberrant Pfr form. These mutants are, therefore, defective in normal light signal perception. It is well established that mutation at the chromophore attachment site (C357S of phyB), preventing chromophore ligation, shows loss of biological activity [29], and the N-terminal 450 amino acid-fragment of phyB which exhibits an aberrant Pfr form and fast dark reversion has reduced biological activity [23]. Of the fourteen mutants newly studied here, seven (G284E, P309L, R313K, R322Q, S370F, V401L and S584F) are photoperception mutants. Of these, two (P309L and S370F) display essentially complete loss of photosensory activity in vivo (Figure 1B), consistent with the absence or severe loss of chromophore ligation capacity (Figure 2), whereas the remainder display reduced photosensory activity, consistent with varying degrees of spectral aberration (Figures 3 and 4).
By contrast, the remaining seven of the fourteen mutants studied here (D64N, R110Q, G111D, G112D, P149L, P304L and R352K) retain spectral integrity (Figures 2,3 and 4), indicating that they are normal in light signal perception, but defective in signal transfer to downstream components of the phyB transduction chain. The retention of normal spectral properties by these mutant molecules is a strong indication that they retain the broad structural integrity of the N-terminal moiety, because of the well-established evidence that deletion of any of the major subdomains causes aberrant spectral properties and altered biological activity [44]. In addition, of the four signaling mutants examined here in the context of the full-length phyB protein, all showed nuclear localization and normal intranuclear dynamics upon light activation (Figure 7). This result strongly indicates that these mutations specifically disturb the signal transferring function without reducing other functions of phyB.
It is notable that of the nine phyB-variant lines showing an intermediate phenotype (intermediate hypocotyl length between the phyB and the N651G-GUS-NLS transgenic rescue lines), six (G248E, P304L, R313K, R322Q, V401L and S584F) show some degree of reduced responsiveness to the lower compared to the higher intensity. With the exception of P304L, these are all photoperception mutants, compromised in their spectral activity, consistent with the prediction that they will have reduced photosensory sensitivity. The remaining three (G112D, P149L and R352K) do not show such a difference in hypocotyl responsiveness to the cR intensity. This is also not unexpected, because these are signal-transfer mutants. These exhibit normal photoperception, but reduced regulatory activity in inhibiting hypocotyl elongation. This behavior is consistent with the prediction that these mutants will retain the same equal sensitivity as the parent N651G-GUS-NLS molecule to the two cR intensities, but have reduced capacity to transduce the perceived light signal (this second step being independent of the intensity of the signal at saturation).
All 22 mutations were mapped within the phyB amino acid sequence (Figure 1A). These mutations were more or less evenly distributed throughout the N651 moiety, suggesting that all subdomains are important for the normal function of N651. However, the different types of mutations distributed differently. The spectral mutations are distributed mainly in the GAF and PHY subdomains (Table 1). By contrast, the signaling mutations tend to cluster in both the N-terminal extension and PLD. This observation thus defines the roles of the subdomains in the N-terminal moiety: GAF and PHY are apparently responsible for light-signal input (photoperception and/or maintenance of the Pfr form), whereas the N-terminal extension and PLD are mainly involved in signal transduction by phyB. This conclusion is consistent with the fact that GAF forms the chromophore pocket [24] and PHY stabilizes phyB in the Pfr form [23]. Similarly, it has been shown that deletion of the N-terminal extension reduces the biological activity of phyB [29]. Although the importance of PLD to the signal transfer function of phytochrome has not been reported, many PAS domains are known to be involved in protein-protein interactions [45], implying that PLD may be directly involved in the interaction with downstream signaling components such as PIF3 [13]–[15].
Recently, the three dimensional structure of the bacterial phy DrCBD has been determined [24]. The data show that the PAS domain of DrCBD exhibits a typical PAS fold while the GAF domain constitutes the chromophore-binding pocket in which the phytochromobilin chromophore is buried. Especially interesting is an unusual three dimensional structure, the proposed “light sensing knot”, found between the PAS and GAF domains. Alignment of the phyB sequence with that of DrCBD allowed us to predict the positions of the mutated residues in the three dimensional model (Figure 5). The chromophore is surrounded by a β-sheet consisting of β6-11 strands and two α-helices (α6 and 7) in the DrCBD chromophore pocket [24]. All of the mutations in the GAF domains except R352K were predicted to be within this region (Figure 5). These amino acid residues are highly conserved among diverse phytochromes. Of the PLD mutations, R110Q, G111D and G112D were predicted to be within or in the vicinity of the β1′-strand, which is one of the partners for β3′ in formation of the knot [24]. G118R, S134G, P149L and I208T were located between β1′ and β1, at the end of β2, between α1 and α2, and at the end of β4, respectively (Figure 5). These amino acid residues mutated in PLD are, for the most part, not highly conserved among phytochromes, with the exception of G118 and S134, which reduce the chromophore incorporation.
We employed an in vitro reconstitution system [37] to examine the spectral properties of mutant N651 derivatives. Zn-blot analysis effectively identified mutants that were deficient in chromophore incorporation (Figure 2). Chromophore incorporation was severely impaired in the G118R, S134G and S370F mutants. In addition, reduced chromophore incorporation was observed in G284E and P309L, both of which also exhibited abnormal difference spectra (Figure 3). In another subclass of mutants, which included R322Q, A372T and S584F, chromophore incorporation was normal but the difference spectrum was altered (Figure 3).
Alignment of the phyB sequence with that of DrCBD allowed us to predict the positions of the mutated residues in the three dimensional model (Figure 5). In the following description, amino acid residues in DrCBD are shown in parentheses. As expected, many of the chromophore incorporation and difference spectrum mutations mapped to the vicinity of the chromophore. Indeed, close interactions of S370(S272) and A372(S274) with the chromophore in DrCBD has been reported [24] (Figure 5). In addition, G284(G184), P309(P209) and R322(R222) are situated in the vicinity of the chromophore.
It remains unclear why mutations in G118(G39)R and S134(S55)G severely disturbed chromophore incorporation. These residues reside in PLD. In the three dimensional model, these residues are spatially separated from the chromophore pocket in DrCBD [24] (Figure 5). However, there are reports that indicate the involvement of PLD in chromophore incorporation. The N-terminal 225 amino acid deletion abolishes chromophore incorporation in Arabidopsis phyA[46]. The I80 residue of pea phyA, which corresponds to I114(I35) of Arabidopsis phyB, is critical for chromophore binding [47]. Insight into the means by which these residues in PLD contribute to chromophore binding awaits elucidation of the three dimensional structure of higher plant phytochrome.
The dark reversion rate, which reflects the stability of Pfr in darkness, is an important process regulating the level of Pfr in vivo. Mutants defective in Pfr stability are thus compromised in normal light-signal perception. Indeed, a faster dark reversion rate has been shown to reduce the physiological activity of phyB [23]. We observed faster dark reversion in 9 of 17 mutants examined (Figure 4A, Table 1). It is known that PHY stabilizes Pfr [22],[23],[48]. Concordantly, each of the three PHY mutants (G564A, S584F, A587T) produced higher dark reversion rates.
The other mutants that exhibited faster dark reversion (H283T, R313K, R322Q, C327Y, A372T, V401I) were found to be mutations in GAF. This is not surprising because GAF constitutes the chromophore binding pocket [24]. In the DrCBD three dimensional structure, A372(S274) directly interacts with the C-ring of the chromophore molecule. In addition, R322(R222) and V401(A288) reside in the chromophore pocket. H283(T183), R313(R213) and C327(T227) are a little more distant but still in the vicinity of the chromophore pocket.
Including I208T identified in a previous study [32], eight mutants that exhibited reduced biological activity with no effect on spectral activity are defined as signaling mutants (Table 1). One mutation (D64N) was found in the N-terminal extension consistent with the reports that, although no structural information is yet available, the N-terminal extension is important for the signal transduction activity of phyB [29]. Two mutations, P304L and R352K, were found in GAF. R352 is particularly interesting because it is presumed to reside in the vicinity of both the chromophore and the light sensing knot (see below). The reason why P304L reduced the signaling activity is less clear. However, P304(P204) is next to Y303(F203), which interacts with ring D of the chromophore in DrCBD [24] suggesting that P304(P204) might affect signaling activity through an interaction with the ring D.
The other 5 signaling mutants were found in PLD, suggesting that this domain is important for the signal transduction activity. Particularly interesting are the three successive mutations, R110Q, G111D and G112D. Interestingly, R110(I31), G111(P32) and G112(G33) partly overlap with the β1′ strand which, together with the β2′ and β3′ strands in DrCBD, participates in the formation of the light sensing knot [24] (Figure 5). Hence, the present data are consistent with the idea that the light sensing knot plays a critical role in phytochrome signal transduction. Two additional mutants, P149L and I208T, were found in PLD. The I208(V118) residue is at the end of the β4 strands and faces the knot in the DrCBD structure [24]. The P149(R70) reside is in the loop connecting the α1 and α2 helices and faces the knot as well.
The R352(R254) residue forms salt bridges through its two amines with the carbonyl oxygen of the ring B propionate of the chromophore in DrCBD [24]. Since one of these amines is missing in the R352K mutant, the mutation would be expected to weaken the interaction between ring B and the polypeptide moiety. Because of the tight connection with the chromophore, the R352K mutant might be expected to have negatively affected photochemical properties. Indeed, the substitution to E of R318 in pea phyA, and that to K of R254 in cph1, which correspond to R352 of Arabidopsis phyB, altered their photochemical property [49],[50]. Nevertheless, abnormal spectral properties of R352K were less clear in the N-terminal moiety of phyB (Figures 2–4). This may be because of the different phytochrome species involved. Furthermore, PBG(R352K) accumulated in the nucleus and formed speckles in a light-dependent manner (Figure 7), which strongly indicates that PBG(R352K) was spectrally active in vivo.
One surprising feature of R352(R254) is its proximity to the light sensing knot. In the DrCBD structure, R352(R254) is on the β3′ strand, which is a component of the knot (Figure 5). The three successive R110(I31), G111(P32) and G112(G33) residues are partly included in β1′, which is one of the partners for β3′ in formation of the knot [24]. Considering the possible tight connection of R352(R254) with the chromophore, these four amino acid residues may constitute a route to relay the conformational changes in the chromophore to the surface of the molecule. It should be noted here that the model presented here is based on the DrCBD structure. Unfortunately, the homology is not particularly high between higher plant phyB and DrCBD within PLD (Figure 5). Consequently, the three dimensional structure of phyB may be different from that of DrCBD. To answer the question definitively, the three dimensional structure of phyB needs to be determined.
It is notable, that the disruption of the signal transfer capacity of the phyB molecule by mutations in the light-sensing knot region have parallel deleterious effects on both early and late phases of seedling deetiolation regulated by phyB. This suggests that these amino acids have a central role in the primary signaling function of the photoreceptor molecule.
The Arabidopsis thaliana mutant, phyB-5, is a null allele on the Landsberg erecta background [34]. The PBG [9] and N651G-GUS-NLS (originally NG-GUS-NLS) [20] lines on the phyB-5 background and the PBG18 line on the phyA-201phyB-5 double mutant background [23] have been described elsewhere.
Seeds were surface-sterilized and sown on 0.6% agar plates containing Murashige-Skoog (MS) medium with or without 2% (w/v) sucrose. The plates were kept in the dark at 4°C for 72 hr and then irradiated with continuous white light (cW) for 3 hr at 22°C to induce germination. The plates were then placed under various light conditions, as specified in the figure legends. The light sources were as described previously [23]. For hypocotyl length measurements, the seedlings were grown on MS agar plates without sucrose for 5 days at 22°C and then pressed gently onto the surface of agar medium before photographs were taken. Hypocotyl length was determined by the NIH image software (Bethesda, ND). For immunoblot analysis, the seedlings were grown on MS agar plates with 2% (w/v) sucrose for 1 week at 22°C in cW (45 µmol m−2 sec−1).
Seeds of the N651G-GUS-NLS expressing Arabidopsis line, 4-1, were mutagenized with 0.3% EMS. Approximately 600 seeds were sown directly onto soil in individual pots. Growth in each pot, which consisted of about 300 plants, was considered an M1 family. From each M1 family, M2 seeds were collected. One to two thousand M2 seeds were then subjected to screening. Seedlings were screened visually for tall phenotype after 5 days under weak cR (0.05 µmol m−2 sec−1). M3 seedlings were then examined for hypocotyl lengths in weak cR and cFR (10 µmol m−2 sec−1). Lines that were taller only in cR were backcrossed to the phyB mutant. The long hypocotyl phenotype was examined in both F1 and F2 generations to determine if the mutation was linked to the N651G-GUS-NLS locus. The light sources employed have been described elsewhere [23].
Crude plant DNA was prepared from the M3 plants. The N651 fragment of N651G-GUS-NLS was amplified using PCR primers complementary to the cauliflower mosaic virus 35S promoter and GFP regions. Purified PCR products were sequenced using BigDyeTerminator V3.1 Cycle Sequencing Kit (Applied Biosystems).
Protein extraction, SDS-polyacrylamide gel electrophoresis, protein blotting, and immunodetection were performed as described [9]. Antibodies used were a monoclonal anti-phyB mBA1 antibody [51], an anti-GFP monoclonal antibody (SIGMA) and antiserum against chitin binding domain (New England Biolabs).
For N651 protein expression, the N651 fragment was cloned into the pTYB2 vector containing Intein/CBD (New England Biolabs) [23]. Mutations were introduced into N651 using the QuikChange Site-Directed Mutagenesis Kit (Stratagene). Escherichia coli transformation and expression of wild type and mutant N651-Intein/CBD fusion proteins were performed as previously described [23]. Intact holoproteins were reconstructed using PCB as a chromophore [37]. The resultant holoproteins were subjected to spectral analyses.
The Zn blot, difference spectra, and dark reversion analyses were essentially as described previously [23]. For Zn blot analysis, extracts containing equal amounts of N651-Intein/CBD protein were loaded onto the gel. To ensure equal sample loading, immunodetection of Intein/CBD fusion proteins was performed in advance.
To generate mutant PBG constructs, mutations were introduced into PBG using the QuikChange Site-Directed Mutagenesis Kit (Stratagene). Mutant PBGs were inserted between the cauliflower mosaic virus 35S promoter and the Nos terminator of pPZP211/35S-nosT, which is itself derived from pPZP211 [52]. The phyB-5 mutant was used as the host for transformation by the Agrobacterium-mediated floral dip method [53]. Transformed plants were selected on MS medium containing 25 µg mL−1 kanamycin and 166 µg mL−1 claforan (Hoechst) and by microscopic observation of GFP fluorescence.
RNA isolation, cDNA synthesis and the real-time PCR were performed essentially as described [38]. The specific primer sequences were as follows: ELF4-F, 5′-CGACAATCACCAATCGAGAATG-3′, ELF4-R, 5′-AATGTTTCCGTTGAGTTCTTGAATC-3′, SAUR-like-F, 5′-TTCTTCACTGCAAGGGATTGTG-3′ SAUR-like-R, 5′-AAAGGCAGAGGAAGAGTTTGGA-3′ AMYLASE-F, 5′-AAAGCACGGTCTCAAACTCC-3′, and AMYLASE-R, 5′-CACAGAATCACATCCCAAGG-3′. The gene PP2A (At1g13320), which is expressed at similar level in darkness or red light (data not shown), was used as a normalization control [54]. Each PCR was repeated three times. Gene expression data were represented relative to the average value for the wild type grown in darkness in each experiment, after normalization to the control. The experiment was performed with three independent biological replicates.
Seedlings were grown on MS agar plates without sucrose for 3 days at 22°C in darkness. Seedlings were set on the stage of a confocal laser microscope (Olympus) and nuclei were located under green safe light by conventional microscopic observation. Seedlings were scanned once to observe GFP fluorescence [9] and then irradiated with the microscope white lamp for 2 min. After irradiation, the seedlings were scanned again. For long-term irradiation, seedlings were treated with cR of 44 µmol m−2 sec−1 for 24 hr.
|
10.1371/journal.pntd.0005296 | Neuropathogenesis of Zika Virus in a Highly Susceptible Immunocompetent Mouse Model after Antibody Blockade of Type I Interferon | Animal models are needed to better understand the pathogenic mechanisms of Zika virus (ZIKV) and to evaluate candidate medical countermeasures. Adult mice infected with ZIKV develop a transient viremia, but do not demonstrate signs of morbidity or mortality. Mice deficient in type I or a combination of type I and type II interferon (IFN) responses are highly susceptible to ZIKV infection; however, the absence of a competent immune system limits their usefulness for studying medical countermeasures. Here we employ a murine model for ZIKV using wild-type C57BL/6 mice treated with an antibody to disrupt type I IFN signaling to study ZIKV pathogenesis. We observed 40% mortality in antibody treated mice exposed to ZIKV subcutaneously whereas mice exposed by intraperitoneal inoculation were highly susceptible incurring 100% mortality. Mice infected by both exposure routes experienced weight loss, high viremia, and severe neuropathologic changes. The most significant histopathological findings occurred in the central nervous system where lesions represent an acute to subacute encephalitis/encephalomyelitis that is characterized by neuronal death, astrogliosis, microgliosis, scattered necrotic cellular debris, and inflammatory cell infiltrates. This model of ZIKV pathogenesis will be valuable for evaluating medical countermeasures and the pathogenic mechanisms of ZIKV because it allows immune responses to be elicited in immunologically competent mice with IFN I blockade only induced at the time of infection.
| Research addressing the severe clinical complications associated with ZIKV infection, including GBS and congenital ZIKV syndrome, are urgently needed. Key to this effort is development of well-characterized animal models that recapitulate human disease. Adult wild-type mice infected with ZIKV can develop viremia in some instances, but they do not emulate the disease associated with the severe congenital and adult neuropathology. Several groups have recently described type I or type II IFN-deficient murine models that are permissive for viral replication in several organs including the brain. The major limitation of these models is they utilize immunodeficient knockout mice lacking key components of the innate antiviral response. We describe the use of a lethal murine model for ZIKV where the innate response of immunocompetent mice is suppressed only at the time of infection. We show that the mice develop severe neurological disease similar to that previously demonstrated in mice deficient in the type I or II IFN response. Using this model, we provide a detailed description of the ZIKV-associated pathologic changes, which mirrors the neuropathogenic properties of ZIKV in humans. These studies provide a baseline for assessment of medical countermeasures that can prevent or treat such pathogenic effects caused by ZIKV infection.
| Zika virus (ZIKV, Flaviviridae, Flavivirus) is an arthropod-borne virus (arbovirus) that is closely related to dengue, West Nile, Japanese encephalitis and yellow fever viruses [1,2]. ZIKV was first isolated in Uganda in 1947 from a febrile sentinel rhesus monkey in the Zika forest [3,4]. No significant outbreaks of ZIKV infection involving more than a few persons were detected until 2007, when ZIKV caused an explosive outbreak in Micronesia [5–8] where approximately 75% of the population on the island of Yap became infected during a four-month period [5]. In subsequent years, ZIKV continued to spread throughout Oceania [9–12]. In early 2015, ZIKV first emerged in the Western Hemisphere with an outbreak detected in Brazil [13,14]. The virus spread rapidly throughout Latin America and the Caribbean and within one year most countries in the region reported local transmission [15,16]. ZIKV is expected to continue to spread and imported cases from travelers returning from Latin America and the Caribbean have already been reported in several countries including the U.S. and Europe [15,17–19]. In fact, numerous locally acquired mosquito-borne cases have recently been reported in Florida.
Historically, infections with ZIKV are asymptomatic and have been associated with a self-limiting febrile illness with no long-term sequelae, but more severe complications have become apparent during the recent outbreaks in the South Pacific and Latin America. In particular, significant concern is growing about the association of ZIKV infection and the development of fetal abnormalities such as microcephaly. ZIKV was isolated from the brains and cerebrospinal fluid of neonates born with microcephaly and identified in the placental tissue of mothers who had symptoms consistent with ZIKV infection during pregnancy [20–22]. An additional concern is the association of ZIKV infection and Guillain-Barré syndrome (GBS). GBS is an autoimmune polyradiculoneuropathy that can result in weakness, paralysis, and death [23–25], and was first associated with ZIKV infection during the 2013–2014 outbreak in French Polynesia. Cases of a diffuse demyelinating disorder consistent with GBS that are temporally associated with ZIKV infection have been reported in Brazil, El Salvador, Colombia, and Venezuela [26]. As ZIKV continues to spread, so does concern about the association of ZIKV infection and the development of severe clinical complications. Therefore, the development of medical countermeasures for ZIKV is a high research priority.
Animal models are needed to better understand the pathogenic mechanisms of ZIKV and to evaluate candidate medical countermeasures. Early ZIKV mouse models have relied on the use of juvenile animals and/or intracerebral inoculations [3,4,27–35]. These initial studies suggest that in mice ZIKV can replicate and cause injury in cells of the CNS. In contrast, other animals to include cotton rats, guinea pigs, rabbits, and rhesus monkeys did not develop CNS disease even when infected by intracerebral inoculation [3]. In mice, neuronal degeneration and cellular infiltration were observed in regions of the spinal cord and brain [3]. Neuronal injury was also evident in the pathological evaluation of a human fetus infected in utero with ZIKV. Diffuse astrogliosis and activation of microglia were observed and damage extended to the brain stem and spinal cord [22]. Recently, mice deficient in the type I or type II interferon response developed severe neurological disease due to ZIKV infection [36–40]. ZIKV-infected Ifnar1-/- mice (C57BL/6 background mice lacking the IFN α/β receptor) developed disease that was associated with high viral titers in the brain and spinal cord [39]. Similar results were described for A129 mice (129Sv/Ev background mice lacking the IFN α/β receptor), which were highly susceptible to ZIKV and developed neurological disease [36,38]. AG129 mice (129Sv/Ev background mice lacking the IFN α/β and γ receptors) were found to be more susceptible to ZIKV-induced disease compared to A129 mice [36]. Collectively, these efforts underscore the importance of innate immunity in modulating ZIKV infection and disease outcome.
The major limitation of these recently described ZIKV mouse models is that they utilize immunodeficient mice. These mouse models lack a key component of antiviral immunity which impairs comprehensive evaluation of medical countermeasures. In an attempt to produce infection models that do not rely upon knockout mice, several groups, including ours, have explored the temporal blockade of IFN-I in immune intact mice using polyclonal and monoclonal antibodies targeting either IFN-Is directly or the IFN-I receptor. The major advantage of this approach is that it allows immune responses to be elicited in immunologically competent mice with IFN I blockade only induced at the time of infection. A murine non-cell depleting monoclonal antibody (MAb) that efficiently targets the IFNAR-1 subunit of the mouse IFN-α/β receptor (MAb-5A3) was developed and shown to prevent type I IFN-induced intracellular signaling in vitro and to inhibit antiviral, antimicrobial, and antitumor responses in mice [41]. MAb-5A3 has been used to explore the role of IFN-I in the infection and pathogenesis of several viruses including West Nile virus (WNV) [42], lymphocytic choriomeningitis virus, and vesicular stomatitis virus (VSV) [41]. For VSV and WNV, treatment of mice with this antibody results in a severe and lethal infection model similar to that produced in IFN-I receptor knockout mice. Here, we describe the use of MAb-5A3 antibody to block IFN-I signaling in immune intact, wild-type mice at the time of ZIKV infection. We demonstrate that these mice develop severe ZIKV-mediated disease accompanied by significant neuroinflammation and mortality when infected by multiple exposure routes. While we were completing this study, another group also reported the use of this system for ZIKV studies [39,43]. Although in the first report, these authors were unable to demonstrate ZIKV lethality, in the second study, they did find lethality using an African lineage strain that was derived from a brain homogenate by passage of the virus in Rag1-/- mice. Our report not only expands on those findings, but also provides the first comprehensive description of the pathologic changes associated with ZIKV infection using this model. This model of ZIKV pathogenesis will be valuable for evaluating medical countermeasures because it allows an immune response to be elicited in immunocompetent mice and infection is enhanced at the time of virus challenge.
This work was supported by an approved USAMRIID IACUC animal research protocol. Research was conducted under an IACUC approved protocol in compliance with the Animal Welfare Act, PHS Policy, and other Federal statutes and regulations relating to animals and experiments involving animals. The facility where this research was conducted is accredited by the Association for Assessment and Accreditation of Laboratory Animal Care, International and adheres to principles stated in the Guide for the Care and Use of Laboratory Animals, National Research Council, 2011. Approved USAMRIID animal research protocols undergo an annual review every year. Animals are cared for by a large staff of highly qualified veterinarians, veterinary technicians, and animal caretakers. All personnel caring for and working with animals at USAMRIID have substantial training to ensure only the highest quality animal care and use. Humane endpoints were used during all studies and mice were humanely euthanized when moribund according to an endpoint score sheet. Mice were euthanized by terminal exsanguination or CO2 exposure using compressed CO2 gas followed by cervical dislocation. However, even with multiple observations per day, some animals died as a direct result of the infection.
ZIKV strain DAK AR D 41525 isolated in 1984 from Aedes africanus mosquitoes in Senegal and was obtained from the World Reference Center for Emerging Viruses and Arboviruses (R. Tesh, University of Texas Medical Branch) where it was amplified once in AP61 and C6/36 cells, and two times in Vero cells prior to our receipt. We then amplified the virus once more in Vero cells (ATCC, CCL-81) prior to use in this study and sequenced [44]. Female C57BL/6 mice (n = 10/group; Jackson Laboratories) five weeks of age were injected IP with a total of 3.0 mg (2.0 mg first dose, 0.5 mg subsequent doses) of MAb-5A3 (produced by Leinco Technologies, St. Louis, MO) [41,45] or PBS on day -1, day +1, and day 4. A prior study characterizing this antibody indicated that a large bolus is needed to saturate the IFNAR-1 receptor pool and the half-life of a 2.0 mg injected dose is 5.2 days [41]. On day 0, mice were infected with 6.4 log10 PFU of ZIKV strain DAK AR D 41525 by the SC (in between the shoulder blades) or IP exposure route in a total volume of 200 μL. Mice were monitored for signs of disease and bled on day 4 post-infection (PI) or when euthanized to evaluate viremia. A cohort of Mab-5A3-treated, uninfected control mice (n = 3) was included for histopathology assessment. These control mice were treated with the antibody as described above and euthanized on day 10.
Mouse serum samples were inactivated using a 3:1 ratio of TRIzol LS Reagent (Thermo Fisher Scientific, Waltham, MA). Tissues were homogenized in 1X Minimum Essential Medium with Earle’s Salts and L-glutamine (MEM) with 1% penicillin/streptomycin and 5% heat-inactivated fetal bovine serum (FBS-HI) using a gentleMACS dissociator (Miltenyi Biotec, San Diego, CA) followed by centrifugation at 10,000 x g for 10 minutes and the supernatant was stored at -80°C until further evaluation. Supernatant was inactivated using a 3:1 ratio of TRIzol LS. Total nucleic acid from all samples was purified using the EZ1 Virus Mini Kit v 2.0 (Qiagen, Valencia, CA) and the EZ1 Advanced XL robot (Qiagen) according to the manufacturer’s recommendations. Samples were eluted in 60μL. Viral load was determined using a real-time RT-PCR assay specific to the 5’-untranslated region of ZIKV. Specific amplification detection was accomplished using a forward primer (5’-GARTCAGACTGCGACAGTTCGA), reverse primer (5’-CCAAATCCAAATTAAACCTGTTGA), and probe (5’-ACTGTTGTTAGCTCTCGC–MGBNFQ). A standard curve was generated using serial dilutions of the challenge virus having PFU/mL titers determined by plaque assay. Five μL of extracted nucleic acid were run in triplicate on the LightCycler 480 (Roche Diagnostics, Inc., Indianapolis, IN) using SuperScript One-Step RT-PCR (Thermo Fisher Scientific), and samples were considered negative if the cycle of quantification (Cq) was greater than 40 cycles. This Cq cutoff value was selected because when Cq's are greater than 40 cycles, you are outside of the linear dynamic range of real-time PCR, and thus, can negatively impact data reproducibility. The virus titers were calculated using the standard curve and the LightCycler 480 software, and the final PFU equivalents/mL (PFUe/mL) calculations were determined based on the sample input volumes and the upfront sample dilutions.
Vero cells were plated at 3x105 cells/well in a six-well plate and incubated overnight at 37°C, 5% CO2. Serial dilutions of samples were made in 1X MEM with 1% penicillin/streptomycin and 5% heat-inactivated FBS. Uninfected-control and serially-diluted samples were incubated with the Vero cells for one hour at 37°C, 5% CO2 for virus adsorption. The inoculum was removed and a 1:1 mixture of 0.8% (w/v) Seaplaque agarose and 2X Basal Medium Eagle with Earle’s Salts (EBME) solution containing 2X EBME, 10% FBS-HI, 2% penicillin/streptomycin, 50 μg/mL gentamicin, and 2.5 μg/mL Fungizone/Amphotericin B was added. After addition, the 0.4% Seaplaque agarose/2X EBME overlay was incubated at room temperature for 30 minutes to allow the overlay to solidify. Vero cells were incubated with the overlay at 37°C, 5% CO2 for five days before the overlay was removed. Cells were fixed and plaques were visualized by a 20 minute addition of 10% formalin with 50% Crystal Violet solution followed by a wash with water.
A necropsy was completed to collect the spleen, liver, head (to include brain), heart, kidney and spinal cord. All collected tissues were immersion fixed in 10% neutral buffered formalin for at least 2 days. The tissues were trimmed and processed according to standard protocols [46]. Histology sections were cut at 5 to 6 μM on a rotary microtome, mounted onto glass slides, and stained with hematoxylin and eosin (HE). Unblinded histological examination was performed by a board-certified veterinary pathologist.
In situ hybridization was performed using RNAscope 2.5 HD RED kit according to the manufacturer’s recommendations (Advanced Cell Diagnostics, Hayward, CA). Briefly, 20 ZZ probes set targeting the 1550–2456 fragment of the ZIKV polyprotein gene [47] were synthesized. After deparaffinization and peroxidase blocking, the sections were heated in antigen retrieval buffer and then were digested by proteinase. The sections were covered with ISH probes and incubated at 40°C in a hybridization oven for two hours. They were rinsed and the ISH signal was amplified by applying Pre-amplifier and Amplifier conjugated with HRP. A red substrate-chromogen solution was applied for 10 minutes at room temperature. The slides were further stained with hematoxylin, air dried, and mounted.
Formalin-fixed, paraffin-embedded mouse brain sections on slides were deparaffinized in xyless and rehydrated through graded ethanol (100%, 95%, 90%, and 70%). Antigen was retrieved by citric acid-based antigen unmasking solution (Vector Laboratories) during 10 minute boiling. After three washes with PBS (pH 7.4), the sections were blocked with 10% normal donkey serum in PBS-tween (0.1%; PBS-T) for one hour at room temperature. The sections were incubated with primary antibodies, goat anti-Iba1 (3 μl/mL; Novus Biotechnology) and Rabbit anti-GFAP (1:5000; Abcam), diluted in 10% normal donkey serum in PBS-T overnight at 4°C. After washing in PBS-T (3x5 min), the sections were incubated for 2 h at room temperature with secondary antibodies diluted in 10% normal donkey serum in PBS-T. For standard immunofluorescence, the secondary antibodies were donkey anti-goat Alexa Fluor 488 (1:300; Invitrogen) and donkey anti-rabbit Rhodamine-Red-X (1:200; Jackson ImmunoResearch). The nuclei were stained with Hoecht’s. For infrared analysis, the secondary antibodies were donkey anti-goat IRDye 680RD (1:1500; Li-cor Biosciences) and donkey anti-rabbit IRDye 800CW (1:1500; Li-cor Biosciences). The sections were subsequently washed in PBS-T (3x10 min), PBS (3x5 min), and water (2x5 minutes). For immunofluorescence, the sections were cover slipped with Fluoromount-G (SouthernBiotechnology). For double-fluorescence labeling, primary rabbit anti-Zika Envelope (E) glycoprotein (1:400, IBT Bioservices) and mouse anti-alpha-Smooth Muscle Actin (1:200 Clone 1A4, R&D Systems) and secondary Alexa Fluor 488 conjugated goat anti-rabbit and Alexa Fluor 561 conjugated goat anti-mouse antibody were used. Rabbit IgG isotype (1:500; cat# MA5-16384, ThermoFisher Scientific) was used as an immunofluorescence staining control. The nuclei were stained with 4′,6-diamidino-2-phenylindole (DAPI). Images were captured on a Zeiss LSM 780 confocal system and processed with Zen 2011 confocal, Photoshop, or ImageJ software. Sections for infrared analysis were air-dried overnight. A Licor-Odyssey CLx (Li-cor Biosciences) scanned sections at 21 μm/pixel resolution. The average intensities of GFAP and Iba1 on each slide were obtained from fields-of-interest draw around each section with the Li-cor-Odyssey analysis software on at least two sections per slide and three slides per brain were scanned. Negative control staining, for which the primary antibodies were omitted, showed no detectable labeling in immunofluorescence or infrared imaging.
Survival analysis was completed by Kaplan Meier estimate. A one-way ANOVA by Kruskal-Wallis test with Dunn’s multiple comparisons was used to analyze differences in viremia. An unpaired t-test was used to compare Iba1 and GFPA expression. SAS version 9.1.3 (SAS Institute Inc., Cary NC) was used for all analyses.
Immunocompetent mice were treated with MAb-5A3 to block IFN-I signaling or PBS prior to and after challenge with 6 log10 PFU of ZIKV given by IP or SC injection. All of the PBS-treated mice challenged with ZIKV by either route survived and no apparent signs of disease, including weight loss were observed (Fig 1A and 1B). Mice exposed to ZIKV by the IP route and treated with MAb-5A3 began to succumb or were euthanized on day 7 PI at which time the mice presented with ruffled fur, hunched posture, and were poorly responsive. On day 8 PI, a mouse exposed SC and another mouse exposed IP exhibited right-side, hind-limb paralysis and were euthanized. Mice treated with MAb-5A3 and exposed to ZIKV by the IP route continued to succumb or were euthanized through day 12 PI where 100% (n = 10/10) mortality was observed. MAb-5A3-treated mice challenged with ZIKV by the SC route continued to succumb through day 19 PI and resulted in 40% (n = 4/10) mortality. The mean times-to-death (MTD) for mice exposed IP was 9.7 days and for mice exposed SC was 14.75 days, which were significantly different (P<0.0001). Weight loss corresponded with survival for both challenge routes for mice treated with MAb-5A3. Mice treated with MAb-5A3 and exposed to ZIKV IP or SC began to lose weight on day 4 PI and 6 PI, respectively. Weight loss continued for MAb-5A3treated mice exposed to ZIKV IP through day 12 PI when all mice had succumbed or were euthanized. MAb-5A3-treated mice exposed to ZIKV SC continued to lose weight until day 8 PI and then slowly began to regain weight and return to baseline values relative to day 0 PI by day 19 PI when mortality was no longer observed. However, we cannot determine if weight loss occurred only in mice succumbing from SC ZIKV infection because mice were weighed by group and not individually. These findings indicated that ZIKV can cause high mortality in mice with intact immune systems when IFN-I is blocked and that IP exposure was more lethal than SC exposure.
The viral titers in the sera were determined on day 4 PI and when mice were euthanized (Fig 1C and 1D). All mice treated with MAb-5A3 and exposed to ZIKV developed viremia, however these levels were slightly (but not significantly) higher on average (0.6 log10 PFUe/mL higher) in IP exposed animals versus SC exposed animals. While all PBS-treated mice exposed to ZIKV by IP injection developed viremia, it was significantly (p<0.0001) lower on average (3.6 log10 PFUe/mL lower) compared to MAb-5A3-treated mice exposed to ZIKV IP. Most mice treated with PBS and exposed SC to ZIKV had viremia below the limit of detection for our assay (5/10), however, one mouse had a titer of >5 log10 PFUe/ml. There was no significant difference in the viremia of mice treated with PBS and exposed SC vs. IP. In addition to analyzing day 4 viremia in all mice, we also measured the viremia in animals succumbing to infection. At the time of euthanasia, viremia was generally lower compared to levels observed on day 4 PI. At time of euthanasia or death, unperfused tissues were collected and viral titers in the liver, spleen, heart, kidney, brain, and spinal cord were measured by qRT-PCR (Fig 2). Virus was detected in all tissues collected in mice treated with MAb-5A3 and exposed IP and SC to ZIKV. The tissues were collected from moribund mice on different days PI, which does not warrant direct comparison of the results by statistics. The viral titers here represent the detection of viral RNA and not infectious virus, so we completed plaque assays on a subset of samples and confirmed the presence of infectious virus in the brain (a key target tissue in this study) where 3.9 or 3.8 PFU/g was detected in antibody treated mice exposed IP or SC to ZIKV, respectively. Collectively, these findings demonstrated that ZIKV infection in mice with IFN-I blockade results in viremia and tissue titers. However, these mice were not perfused so some of the virus detected in the tissues may be from the blood.
We completed ISH and IFA coupled with histopathological analysis in tissues from ZIKV-infected mice that succumbed or were euthanized due to severe disease. Significant histopathological changes occurred in the CNS of all ZIKV-infected animals treated with the IFN-I blocking MAb-5A3 (Fig 3, S1 Table). The most notable microscopic lesions attributable to ZIKV infection in these mice included evidence of minimal to mild inflammation and necrosis in the brain and spinal cord of all animals that succumbed on days 7 and 8, and to a lesser extent, in at least 5 of the 6 animals that succumbed between days 11 and 12. The presence and /or extent of CNS lesions were difficult to ascertain in a few of the animals (not included in the assessment) due to artifactual damage to the tissue (S1 Table) and autolysis precluded the assessment of three animals that succumbed after day 12 PI and were removed from the histopathology results. Findings in the CNS included perivascular infiltrates of mononuclear cells (“perivascular cuffing") and multifocal to diffuse gliosis with activated microglia (Fig 3A). ZIKV RNA was frequently detected in the same regions by ISH (Fig 3B). Perivascular cuffing was a more consistent finding in mice that succumbed earlier (days 7–8) than in those that died or were euthanized later in the course of disease. In the sections examined, 5 of 5 animals from days 7–8, and 2 of 5 animals from days 11–12 exhibited perivascular cuffing in the brain; 4 of 4 animals from days 7–8, but none from days 11–12 exhibited perivascular cuffing in the spinal cord. In most animals, the perivascular cuffing was subtle, consisting of few mononuclear inflammatory cells; in 2 animals exposed to ZIKV IP that succumbed or were euthanized on day 7 PI, few neutrophils were admixed with the mononuclear cells. Microgliosis was a fairly consistent finding that was observed in the examined CNS sections of all animals from days 7 through 12. The severity was mild to occasionally moderate in animals that succumbed on days 7–8; whereas, microgliosis was minimal to occasionally mild in animals that succumbed on days 11–12 (Fig 3C).
Additional findings in the CNS included multifocal areas of neuropil vacuolation (edema) and scattered necrotic cellular debris, most notably in the cerebrum (Fig 3C), hippocampus (Fig 4A) and thalamus. Again, ZIKV RNA was detected by ISH in the same regions as these histopathological changes (Figs 3D and 4B), which is in contrast to the uninfected control mice (Figs 3E, 3F, 4E and 4F). Multifocal areas of edema occurred in 5/5 animals from days 7–8, and in 2/5 animals from days 11–12. Edema frequently surrounded blood vessels and was most prevalent in areas with the most abundant gliosis and neuronal necrosis. Scattered necrotic cellular debris often occurred adjacent to blood vessels and was observed in the brain of 5/5 animals from days 7–8, and in 5/5 animals from days 11–12. Although necrotic debris was often observed in multiple CNS sections, it was most frequently observed in the hippocampus, thalamus, cerebrum and less so in the cerebellum, pons and spinal cord. It is difficult to determine the cell type from which this necrotic debris originated—possibilities include neurons, resident glia, or infiltrating leukocytes.
An additional notable finding in the CNS was neuronal necrosis, characterized by hypereosinophilic neurons and pyknosis, karyolysis, and replacement of neurons with necrotic debris. Neuronal necrosis occurred most frequently and extensively in the hippocampal pyramidal and granule layers and less frequently in the thalamus, cerebrum, and least frequently in the cerebellum, pons and spinal cord (Fig 4C). Once again, RNA was consistently detected in corresponding regions of the brain by ISH (Fig 4D). In the sections examined, neuronal necrosis was observed in the brain of 4/5 animals from days 7–8 and 3/5 animals from day 11–12. Neuronal necrosis was also observed in the spinal cord of 4/4 animals examined from days 7–8 and 2/3 animals examined from day 11–12. Less frequently there was neuronal degeneration and satellitosis. Another, although less consistent finding, was minimal neutrophilic infiltrates scattered within the brain parenchyma, which was observed in only 2/5 animals from days 7–8, and 3/5 animals from days 11–12.
Histopathological analysis of the spinal cord showed evidence of one or more of the following: gliosis, neuronal satellitosis, and perivascular inflammatory infiltrates, in all mice observed (Fig 5A). The detection of ZIKV RNA by ISH in the spinal cord suggests that these lesions are also due to infection (Fig 5B and 5C). In general, cells in the spinal cord were only occasionally observed to be positive for ZIKV RNA by ISH; however, in some animals the spinal cord was more severely affected, and massive ZIKV infection was detected by ISH in some animals as depicted in Fig 5C. Additionally, a focally extensive area of necrosis affecting a spinal ganglion was observed in a mouse exposed to ZIKV IP that succumbed on day 7 PI (Fig 5D). The noted pathologic changes and ISH findings in the spinal cord of ZIKV-infected mice is in contrast to what was observed in uninfected control mice (Fig 5E and 5F).
It has been known that both ionized calcium binding adaptor molecule 1 (Iba1) expressed by microglia and glial fibrillary acidic protein (GFAP) expressed by astrocytes are upregulated in activated microglia and astrocytes, respectively, during neuroinflammation [48,49]. Infrared and immunofluorescent imaging was used to assess the effect of ZIKV on neuroinflammation by detecting the expression of Iba1 and GFAP in brain sections from 10 of the mice that succumbed to ZIKV infection (brains from four of the mice were not analyzed due to artifactual damage). Compared to brains of control mice treated with IFNAR1-blocking MAb-5A3, Iba1 and GFAP levels were significantly increased in the brains of ZIKV-infected mice (Fig 6A and 6B) and these findings were confirmed by immunofluorescent labeling of Iba1 and GFAP in these brains (Fig 6C). Collectively, the results suggest that in our model, CNS infection by ZIKV results in significant neuroinflammation.
Other significant findings outside of the CNS include necrotic and/or apoptotic cellular debris within the splenic lymphoid follicles (white pulp) interpreted as lymphocytolysis (Fig 7A). This was a fairly consistent finding in these mice, with similar lesions being observed in all (8/8) spleens examined (S1 Table). The severity of the lymphocytolysis varied from mild in the earliest stages (3/3 animals from day 7) to minimal in all subsequent animals. Additionally, several animals (2 from day 7, 1 from day 8 and 1 from day 11) exhibited minimal lymphoid hyperplasia in the spleen, presumably in response to ZIKV infection. ZIKV RNA was also consistently detected by ISH in cells in the white pulp of the spleens from these animals (Fig 7B), which is in contrast to the histopathology and ISH findings observed in uninfected control mice (Fig 7C and 7D). No significant histopathological findings were observed in the liver, kidney, and heart of antibody treated, ZIKV-infected mice; however, ISH staining was observed in the smooth muscle cells within the tunica media of blood vessels in the kidney and heart, which was not observed in uninfected control mice (S1 Fig). IFA confirmed the presence of ZIKV in the smooth muscle of blood vessels in the kidney, which was not observed in uninfected control mice or by isotype control antibody staining (S1 Fig). The lack of concurrent microscopic evidence of tissue injury despite the presence of ZIKV in these mice suggests that, although virus is present in the smooth muscle, no direct damage to these cells is occurring; however, further studies are needed to fully elucidate the significance of ZIKV presence in smooth muscle of these blood vessels. Also, it is unclear why ZIKV appears to be present in the tunica media of vessels of these two organs but not in other organs examined.
Additional findings of interest outside the CNS include the observation of degeneration, inflammation, and less consistently, regeneration, of skeletal muscles of the head and vertebral column. Myocyte degeneration, inflammation, and nuclear rowing were evident by hematoxylin and eosin staining (S2 Fig). However, ZIKV was not detected in the skeletal muscle by ISH when the mice were moribund. Interestingly, ZIKV RNA was detected in a small subset of mice (n = 5) from an additional independent study where the mice were euthanized on day 3 PI (S2 Fig). ZIKV RNA was not detected by ISH in uninfected control mice where the histopathology appears normal (S2 Fig). It appears that ZIKV RNA is present in the skeletal muscle before mice become moribund, but is then cleared at later time points and inflammation persists. Inflammation and degeneration was a fairly common finding, and was noted in at least one anatomic site from all 11 animals examined. Regeneration was observed less frequently, with minimal regeneration occurring in 8 animals. Although lesions were observed in skeletal muscles from multiple anatomic sites, no similar lesions were observed in cardiac muscle. The skeletal muscle observed in this study was limited to the head and vertebral column regions. Future investigations should include examination of the skeletal muscles from the rear limbs, particularly since hind limb paralysis was observed in two of the mice.
The unexpectedly frequent and severe clinical complications of ZIKV infection, including GBS and congenital ZIKV syndrome, have prompted intense research on host-virus interactions. Key to these efforts is the development of well-characterized animal models that recapitulate human disease. Adult wild-type mice infected with ZIKV can develop viremia in some instances [39,50], but do not reflect the severe neurological disease seen in humans. However, transplacental and vaginal infection has been described in wild-type mice [51,52]. Knockout mice, deficient in type I or type II IFN responses, are permissive for viral replication in several organs including the brain [36–40]; however, infection in these mice does not provide an adequate means to test the efficacy of medical countermeasures or to study pathogenic events after infection. For example, in addition to its role in controlling viral infection through antiviral gene induction, type I IFN plays a role in priming of B and T cell responses [reviewed in [53,54]. Therefore, a more immunologically competent mouse model is essential for evaluating vaccines and treatments for ZIKV. The major advantage of this approach is that it allows immune responses to be elicited in immunologically competent mice with IFN I blockade only induced at the time of infection.
We developed and characterized a partially immunocompetent murine model of ZIKV infection replicating the pathologic changes noted in genetically modified mice through antibody blockade of the type I IFN receptor. We are using this model to evaluate vaccines and therapeutics and as reported herein, have used the model to establish a baseline of ZIKV pathogenesis. As we were completing our pathogenesis study, another group reported a similar model. In these studies, a non-lethal model of ZIKV infection was described in wild-type mice injected IP with 1 mg or 2 mg of the same IFN-1 blocking antibody that we used and then exposed SC to 3 log10 focus-forming units of ZIKV strain H/PF/2013, which is a human isolate from the 2013 French Polynesia outbreak [39] or strain Paraíba 2015, an isolate from Brazil [55]. ZIKV replication was observed in several organs of mice treated with the antibody; however, the mice did not lose weight, succumb to infection, or develop neuropathology. The same group established an in utero transmission model of ZIKV infection using wild-type mice treated with this IFNAR-1 blocking monoclonal antibody [56]. In another report, IFN blockade by the same monoclonal antibody followed by SC infection with a higher dose of an African lineage strain of ZIKV (Dakar 41519) that was derived from a brain homogenate via passage in Rag1-/- mice resulted in lethal disease [43]. Our mouse model results are similar to those in this report in that we found that antibody blockade of the type I IFN receptor recapitulates the severity of ZIKV disease observed in Ifnar1-/- mice. Further, we demonstrated that route of exposure is a factor in lethality in this mouse model in that mice exposed by the IP route began to succumb on day 7 PI and 100% mortality was observed by day 12 PI. In contrast, 40% mortality was observed when mice were exposed SC and they succumbed as late as day 19 PI. However, the disease was similar in mice that succumbed to ZIKV regardless of the route of exposure where significant pathologic changes were observed in the CNS. Since vaccination studies requiring boosting would utilize older mice, we confirmed that lethality is observed in mice that are 10 weeks old where 80–100% mortality was observed in mice exposed to 6 log10 PFU IP of ZIKV strain DAK AR D 41525 used in this study. These results are being submitted in a separate publication that also evaluates the susceptibility of this mouse model to multiple ZIKV strains. Zhao et al. report higher lethality by SC infection which may be due in part to passaging their virus in mice [43]. Our results and those described by Zhao et al. and Lazear et al. suggest that the virus strain (African vs. Asian lineage), passage history, and exposure route can affect susceptibility to infection. More studies are needed to evaluate the pathogenesis of African vs. Asian lineage strains (and the effect of passage history) in the various infection models.
Mortality is not a common feature of human infection with ZIKV, which is generally asymptomatic and self-limiting in most individuals. However, severe disease in humans is characterized by neurological complications associated with ZIKV infection. In adults, reported neurological complications include GBS [24,25] or in a few cases, encephalopathy [57], meningoencephalitis [58], and acute myelitis [59] have been described. In the fetus, intrauterine infection can cause congenital abnormalities to include severe fetal brain injury [22]. The neuropathology that we observed in our studies offers a model for dissecting the pathological consequences of ZIKV infection. We observed significant pathologic changes in the CNS of all mice that succumbed to ZIKV infection. Overall, the CNS lesions in these mice represent an acute to subacute encephalitis/encephalomyelitis that is characterized by neuronal death, astrogliosis, microgliosis, scattered necrotic cellular debris, and a minimal to mild mononuclear (and less frequently neutrophilic) inflammatory cell infiltrate. In the brain, lesions were most evident in the hippocampus, particularly affecting the pyramidal and granule cell layers, followed by thalamus and cerebrum, and less often affecting the cerebellum and pons. The presence of ZIKV RNA as detected via ISH suggests these lesions are attributable to ZIKV infection. It is likely that encephalitis/encephalomyelitis contributed to the morbidity and mortality in these animals, particularly those that succumbed early, between days 7 to 11 PI. Lesions, particularly neuronal necrosis and inflammation, appear to be less severe in animals from day 12; however, it is possible that CNS injury played a role in the deaths of these animals as well. The neurotropism of ZIKV was demonstrated in early studies where neonatal mice intracerebrally infected with ZIKV showed evidence of nuclear fragmentation, perivascular cuffing, and degenerative cells in the hippocampus of the brain [3,28]. Bell et al. also observed enlarged astrocytes with extended processes and containing cytoplasmic virus factories throughout the cortex of ZIKV-infected mouse brains [28]. More recent studies characterizing ZIKV murine models in immunodeficient mice also showed that microscopic lesions resulting from ZIKV infection were found primarily in the brain [37,38,40]. Our findings are most consistent with those reported by Dowall et al. [38] where inflammatory and degenerative changes were observed in the brains of A129 mice challenged with ZIKV. The CNS lesions observed in ZIKV-infected mice may be relevant for brain-related pathologies in some ZIKV-infected humans. However, the histopathology of ZIKV-associated microcephaly has been limited to only a few reports thus far (reviewed in [60]). The major findings have mostly been in the brain and include diffuse grey and white matter involvement consisting of dystrophic calcifications, gliosis, microglial nodules, neuronophagia, and scattered lymphocytes [60]. Astrocyte pathology was observed in post-mortem analysis of a neonatal brain with microcephaly associated with ZIKV infection where diffuse astrogliosis was apparent with focal astrocytic outburst into the subarachnoid space. Activated microglial cells were also found to be present throughout most of the cerebral gray and white matter [22].
Outside of the CNS, the most consistent histopathological lesions were observed in the spleen. We observed lymphocytolysis in the splenic white pulp which correspond to areas with ZIKV ISH signal. Dowall et al. also noted similar lesions in the spleen although detection of viral RNA was not described [38]. Other histopathological observations indicated a significant inflammatory response resulting from ZIKV infection. Lymphoid hyperplasia and myeloid (granulocytic) hyperplasia were noted and are indicative of an inflammatory response to antigenic stimulation and an increased demand for leukocytes. Although these are both somewhat non-specific findings, in these cases they most likely represent a systemic immune reaction to ZIKV infection and are related to the significant neuroinflammation noted in the animals. A previous study indicated that ZIKV infection in AG129 mice led to a systemic inflammatory response, where multiple pro-inflammatory cytokines were found to be increased in the sera [40]. It is unknown how this relates to human disease since the cytokine response during the acute phase of ZIKV infection in humans has not been studied.
Inflammation was also observed in the skeletal muscles of the head and vertebral column. Skeletal muscle degeneration and inflammation observed in these animals are thought to be attributable to ZIKV infection since direct infection of the myocytes was observed on day 3 PI. ZIKV is presumably cleared following day 3 PI, but inflammation persists in the skeletal muscle. Interestingly, Ailota et al. described similar pathologic changes in the musculature from the posterior rear limb of a ZIKV-infected mouse where multi-focal myofiber degeneration and necrosis with inflammatory cell infiltration, nuclear rowing, and attempted regeneration were observed. Further investigation into whether the skeletal muscle inflammation is viral or immune mediated is warranted.
In summary, we described a murine model for ZIKV that mimics the severe neurological disease previously described in mice deficient in the type I or II IFN response [37,38,40]. Our detailed description of the ZIKV-associated pathology in this model, much of which mirrors what is known about neuropathogenesis in humans, will provide a baseline for evaluating medical countermeasures to prevent or treat ZIKV infections.
|
10.1371/journal.pgen.1007709 | LytTR Regulatory Systems: A potential new class of prokaryotic sensory system | The most commonly studied prokaryotic sensory signal transduction systems include the one-component systems, phosphosignaling systems, extracytoplasmic function (ECF) sigma factor systems, and the various types of second messenger systems. Recently, we described the regulatory role of two separate sensory systems in Streptococcus mutans that jointly control bacteriocin gene expression, natural competence development, as well as a cell death pathway, yet they do not function via any of the currently recognized signal transduction paradigms. These systems, which we refer to as LytTR Regulatory Systems (LRS), minimally consist of two proteins, a transcription regulator from the LytTR Family and a transmembrane protein inhibitor of this transcription regulator. Here, we provide evidence suggesting that LRS are a unique uncharacterized class of prokaryotic sensory system. LRS exist in a basal inactive state. However, when LRS membrane inhibitor proteins are inactivated, an autoregulatory positive feedback loop is triggered due to LRS regulator protein interactions with direct repeat sequences located just upstream of the -35 sequences of LRS operon promoters. Uncharacterized LRS operons are widely encoded by a vast array of Gram positive and Gram negative bacteria as well as some archaea. These operons also contain unique direct repeat sequences immediately upstream of their operon promoters indicating that positive feedback autoregulation is a globally conserved feature of LRS. Despite the surprisingly widespread occurrence of LRS operons, the only characterized examples are those of S. mutans. Therefore, the current study provides a useful roadmap to investigate LRS function in the numerous other LRS-encoding organisms.
| The ability to sense stimuli triggered by the extracellular environment is a fundamental requirement of all cellular life. For prokaryotes, there are a variety of recognized classes of sensory systems that are used to detect and respond to environmental stimuli. In the current study, we provide the first evidence for the existence of a potentially new class of prokaryotic sensory system, which we refer to as LytTR Regulatory Systems (LRS). Here, we show that LRS are broadly distributed among prokaryotes and are distinct from the other commonly studied sensory systems like two-component signal transduction systems and ECF sigma factor systems. Presently, there are only two characterized examples of LRS, both from Streptococcus mutans. We employ these LRS as models to first define the key features of LRS and then demonstrate how some of these characteristics are likely universally conserved among the plethora of uncharacterized LRS in other organisms. Based upon these data, we further describe how these sensory systems are likely to function in diverse species and illustrate how to identify and investigate the function of novel LRS.
| The capacity of bacteria to sense and respond to stimuli triggered by the extracellular environment is fundamental for survival, particularly in highly dynamic and/or competitive niches. Prokaryotes currently have several recognized classes of sensory signal transduction systems that are used specifically for this purpose. The most diverse class consists of the one-component systems, which contain single protein fusions of a signal-sensing input domain and a transcription regulatory output domain [1]. The vast majority of one-component systems are soluble proteins that utilize a diverse array of small molecules to modulate their transcription factor activity [1]. Among the best characterized classes of prokaryotic sensory systems are the phosphosignaling systems, exemplified by two-component signal transduction systems (TCSTS) and eukaryotic-like serine-threonine kinases/phosphatases (eSTK/P). Phosphosignaling systems respond to environmental stimuli using sensor proteins containing integrated kinase/phosphatase domains, which alter the phosphorylation status of downstream proteins involved in the signaling pathway. For TCSTS, phosphorylation typically controls the sequence-specific DNA binding affinity of one or more cognate transcription regulators [2–5], whereas eSTK/P usually regulate the phosphorylation status of a broad assortment of proteins [6–8]. The next major class of prokaryotic sensory systems is the extracytoplasmic function (ECF) sigma (σ) factors. Unlike TCSTS and eSTK/P, ECF systems do not typically encode enzymatic domains within sensor proteins; rather, gene expression is regulated through the production of alternative σ factors that dictate the promoter affinity of RNA polymerase [9, 10]. ECF σ factors are normally maintained in an inactive state through direct interactions with cotranscribed cognate anti-σ factors that are typically embedded within the cell membrane [11, 12]. ECF systems can be classified into 50 distinct subgroups [13, 14] and are activated when the anti-σ factor is inhibited via regulated proteolysis, protein-protein interactions, or through a signal-induced conformational change, thus liberating the σ factor to assemble within the RNA polymerase holoenzyme [15]. Finally, bacteria (and many other organisms) also utilize a variety of purine-derived second messenger systems to transduce sensory information via molecules such as cAMP, (p)ppGpp, cyclic di-GMP (c-di-GMP), cyclic di-AMP (c-di-AMP), and cyclic GMP-AMP (c-GAMP) [16]. With the exception of (p)ppGpp, these second messenger systems are generally regulated through the action of two classes of proteins: cyclases that create the second messengers and the phosphodiesterases that degrade them [16–21]. For (p)ppGpp, its synthesis is catalyzed by RelA-SpoT family enzymes [22]. Once created, these second messengers can bind directly to their target proteins or RNAs to modulate their functions [20, 23, 24].
Recently, we have been examining the regulatory function of two related signal transduction systems in Streptococcus mutans, which we previously named HdrRM and BrsRM. Both systems share a variety of features and appear to be distinct from the aforementioned signal transduction system paradigms. Homologs of these two S. mutans systems, which we broadly refer to as LytTR Regulatory Systems (LRS), can be found in various bacteria, particularly within the Firmicutes phylum [25]. Despite their widespread distribution, all putative LRS in other organisms remain uncharacterized. Thus, our current knowledge of LRS is presently limited to our previous studies of the HdrRM and BrsRM LRS [25–29]. These two LRS are both arranged within 2-gene operons with the first gene encoding a transcription regulator from the LytTR Family [30] and the adjacent downstream gene encoding a transmembrane protein inhibitor of the LRS regulator [25]. Under normal laboratory growth conditions, the HdrRM and BrsRM LRS are both maintained in a basal inactive state, due to the function of their cognate membrane inhibitor proteins [26, 27, 29]. Thus, the membrane proteins presumably serve as the proximal switches responsible for LRS activation, much like the analogous role of two-component system sensor kinases or ECF system anti-σ proteins. By mutating either of the membrane inhibitors HdrM or BrsM, it is possible to forcibly activate both LRS and examine their effect upon downstream gene expression. Surprisingly, the HdrRM and BrsRM LRS both contain largely overlapping regulons, which includes natural competence and bacteriocin genes in addition to both LRS operons [26–29]. Thus, these two LRS appear to be both autoregulatory and coregulatory. Furthermore, activation of bacteriocin gene expression by the LRS regulators HdrR and BrsR is critically dependent upon their interaction with direct repeat sequences found upstream of the bacteriocin gene promoters [27, 29]. These direct repeat sequences conform to a broadly defined consensus recognized by members of the LytTR Family [29, 30]. While the actual signals responsible for HdrRM and BrsRM activation are currently unknown, both LRS operons are induced by a rapid switch to high cell density growth conditions [26]. Intriguingly, HdrRM and BrsRM also jointly control a potent suicide-like cell death pathway, which underscores their potential ecological significance for S. mutans and perhaps other species [29]. Overall, it is clear that the HdrRM and BrsRM LRS are not cryptic regulators, rather they control distinct regulons that are integrated into a variety of genetic networks. In the current study, we sought to define the key characteristics and global distribution of LRS. We provide evidence that HdrRM, BrsRM, and several other previously unrecognized S. mutans LRS are actually members of a large family of analogous regulatory systems found amongst both bacteria and archaea. The conserved features of these systems indicate that LRS may comprise a previously unrecognized class of prokaryotic signal transduction system.
Our previous investigations of S. mutans LRS have focused upon the HdrRM and BrsRM LRS. However, it was unclear whether additional uncharacterized LRS might also exist in this species. Therefore, we began by searching the S. mutans genome for all of the transcription regulators containing putative LytTR Family DNA binding domains, which identified a total of seven genes. Two of these are obvious TCSTS response regulators (ComE and LytR), two are known LRS regulators (HdrR and BrsR), and the remaining three are uncharacterized hypothetical genes (SMU_294, SMU_433, and SMU_1070c). Inspection of the three uncharacterized genes revealed that all are arranged in apparent polycistronic operons and are upstream of open reading frames (ORFs) encoding putative transmembrane proteins (Fig 1A). This is highly reminiscent of the hdrRM and brsRM LRS operons, except that each of the uncharacterized operons also includes additional ORFs that are likely cotranscribed, whereas the hdrRM and brsRM operons are simply 2-gene operons. The SMU_294/295 genes are located between a conserved hypothetical gene (SMU_293) and an ORF encoding a putative ketopantoate reductase (SMU_296), while the SMU_433/434 and SMU_1070c/1069c genes are both likely cotranscribed with ABC transporter genes (Fig 1A). A key feature of the HdrRM and BrsRM LRS is their autoregulatory ability, which can be activated by mutagenesis of their respective membrane inhibitor proteins [26, 28, 29]. As shown in Fig 1B, each of the putative membrane proteins from all five operons was required to repress transcription of their respective operons indicating that the membrane proteins all similarly serve as inhibitors of an endogenous autoregulatory ability. The levels of induction triggered by the membrane protein deletions did vary widely however, with the SMU_294/295, SMU_433/434, and hdrRM operons all exhibiting ~50 to 60-fold maximum induction, while the SMU_1070c/1069c and brsRM operons exhibited <20-fold and >500-fold induction, respectively (Fig 1B). Overall, the expression characteristics of the operons were quite similar, except for the brsRM LRS, which has only a slightly lower maximum expression but a substantially lower basal expression. Thus, the dynamic range of inducibility for each of these operons seems primarily dependent upon the stringency of operon repression, rather than its maximum expression. In our previous studies, we also observed cross-regulation between the HdrRM and BrsRM LRS [28, 29]. Thus, we were interested to determine whether this is a unique feature of the HdrRM and BrsRM LRS or if other LRS might also exhibit cross-regulation of other LRS operons. To test this, we mutated each LRS membrane inhibitor protein and examined its resulting impact upon the other four non-cognate LRS luciferase reporter strains. To simplify the analysis, we deleted all but the two LRS of interest for each reporter to test every pairwise combination of LRS. With the exception of the SMU_433/434 LRS, all other LRS were found to trigger ≥2-fold change in reporter activity for one or more non-cognate LRS operons (Fig 1C). Several cross-regulatory interactions were quite strong, such as the opposing roles of the SMU_1070c/1069c LRS as both a potent activator of SMU_294/295 LRS operon expression and as an inhibitor of SMU_433/434 LRS expression (Fig 1C and 1D). The SMU_1070c/1069c LRS was also found to be particularly promiscuous, as it is the lone LRS capable of regulating all other LRS operons (Fig 1D). From these results, we can conclude that the activation of one LRS can influence the production of another, possibly as part of a regulatory network to modulate the kinetics associated with non-cognate LRS activation and/or the control of non-cognate LRS regulons.
As mentioned previously, the regulatory function of the HdrRM and BrsRM LRS in S. mutans is strongly indicative that they are not simply cryptic regulators. In further support of this notion, we examined whether the five S. mutans LRS operons are likely to be components of its core genome. 25 randomly selected S. mutans genomes were examined for the presence of all five operons and indeed all were present in every strain examined (Table 1). It should be noted that there were four strains in which brsM was either not annotated or annotated as a pseudogene due to the presence of apparent frameshift mutations within a poly-A region near the 3’ of the brsM ORF (Table 1). If such a mutation were truly present in brsM, it should constitutively activate BrsR in these strains. There was also a single instance in which hdrR was simply not annotated, even though the complete ORF is present (Table 1).
Our previous transcriptomic analyses of the HdrRM and BrsRM LRS indicated that both systems are autoregulatory and coregulatory, as we observed potent induction of both LRS operons due to deletions of either of the LRS inhibitor proteins HdrM or BrsM [28, 29]. The same results could also be recapitulated via ectopic overexpression of either of the LRS regulator genes hdrR or brsR [28, 29]. As members of the LytTR Family of transcription regulators, both HdrR and BrsR would be predicted to recognize direct repeat sequences conforming to a broadly defined consensus [30]. Accordingly, LytTR Family consensus direct repeats are essential for HdrR and BrsR activation of bacteriocin gene expression [27, 29, 31–33]. However, a previous in silico analysis of the S. mutans genome failed to detect LytTR Family direct repeats in any of the LRS operon promoter regions [31]. Thus, we were curious whether the autoregulatory activity of LRS is mediated directly by the LRS regulators or via an indirect mechanism. As a test case, we first scanned the intergenic region upstream of the hdrRM operon to identify potential promoters. A strong candidate containing a putative extended -10 sequence was identified in this region in addition to a pair of direct repeats located 8 nucleotides upstream of the putative -35 sequence (Fig 2A). The spacing and length of the direct repeats are identical to those found in the multiple bacteriocin promoters regulated by HdrR and BrsR, but the operon direct repeat sequence diverges from the reported LytTR Family consensus [30–33]. This likely explains why it had not been previously detected. To further examine the identified operon promoter and direct repeats, we created two separate transcription fusion reporter strains, one in which a luciferase ORF replaced the hdrRM ORFs (i.e. ΔhdrRM) and another in which the luciferase ORF was inserted immediately downstream of the hdrRM ORFs (i.e. wild-type hdrRM). Using the ΔhdrRM reporter strain, we mutagenized the putative extended -10 sequence in the operon promoter, which resulted in substantially lower reporter activity compared to the parent strain (Fig 2B). In addition, the -10 deletion created a dominant phenotype that could not be suppressed even via ectopic hdrR overexpression, strongly supporting the role of this sequence as part of the operon promoter. To determine whether the upstream direct repeats might comprise an HdrR binding site, we performed electrophoretic mobility shift assays (EMSAs) using full-length recombinant HdrR and a small DNA fragment encompassing the hdrRM direct repeat region upstream of the -35. Sequence-specific mobility shifts were both detectable and critically dependent upon the identified direct repeats (Fig 2C). Next, we further assayed the same direct repeat mutations shown in Fig 2C using an hdrRM reporter strain containing a luciferase ORF inserted immediately downstream of the operon ORFs. A double mutation of hdrM and the direct repeats in this reporter confirmed that the direct repeat mutations are similarly dominant, as they resulted in reporter activity below that of the parent strain (Fig 2D). This indicated that the operon direct repeats further increase the basal expression of the operon via HdrR. It is worth noting that the basal luciferase activity of the ΔhdrRM reporter strain in Fig 2B is lower than that of the wild-type hdrRM reporter in Fig 2D (S1 Fig). We attributed this difference to modest levels of HdrR autoactivation upon the hdrRM operon promoter in the wild-type reporter strain and the lack of such regulation in the ΔhdrRM reporter. As further support for this notion, we created an ectopic hdrRM overexpression strain and observed an identical dependence upon the operon direct repeats to maintain the parental level of basal expression (Fig 2E). Thus, in addition to its role in bacteriocin production and natural competence development [27, 28], we can conclude that HdrR also directly serves as an autoregulatory transcription activator, triggering positive feedback autoregulation upon its own operon via two 9 bp direct repeat sequences located just upstream of the operon promoter. Next, we scanned the brsRM operon as well as the three other putative S. mutans LRS operons for similar promoter elements as those found in hdrRM. Like the hdrRM operon, we found that each of the other four operons indeed contain similar direct repeats located 4–11 bp upstream of their operon -35 sequences (Table 2). With the exception of the SMU_1070c/1069c LRS, each set of direct repeats is separated by 12 bp of intervening sequence. For the SMU_1070c/1069c LRS, the repeats are separated by 11 bp. For all five LRS, the locations of the direct repeats immediately upstream of the -35 sequences indicate they share similar regulatory mechanisms utilizing positive feedback autoactivation of their respective operons.
LRS share some analogous features that are highly reminiscent of TCSTS and ECF σ factor systems. In fact, while searching for novel LRS operons in S. mutans and other species, we noticed a number of instances in which uncharacterized LRS regulators are erroneously annotated as LytTR Family response regulators. This would imply that such genes encode members of TCSTS, perhaps as orphan response regulators. While the LytTR Family does include numerous TCSTS response regulators, most members of this family are not [5, 30]. We compared the domain architectures of the two S. mutans response regulators containing LytTR Family DNA binding domains (ComE and LytR) with each of the five S. mutans LRS regulators. While the sizes of all of the LytTR Family DNA binding domains are comparable, the response regulators ComE and LytR are larger proteins overall due to the additional presence of signal receiver domains (Fig 3A), which are key features found in canonical response regulators [5] and are notably absent from the LRS regulators. Likewise, response regulators encode strictly conserved aspartate residues that are essential for phosphosignaling (S2A Fig), yet these are also absent from LRS regulators (S2B Fig). Obvious differences are similarly apparent when comparing TCSTS sensor kinases with LRS membrane inhibitors. The cognate sensor kinases for ComE and LytR (ComD and LytS, respectively) are considerably larger proteins due to the presence of various sensory domains and/or ATPase domains (Fig 3B), which are key features essential for sensor kinase function [34]. No predicted kinase domains or any other putative enzymatic functions are detectable in the five LRS membrane proteins, although four of these proteins do encode either of two Domains of Unknown Function (DUF3021 or DUF2154) (Fig 3B).
Like TCSTS, ECF σ factor systems are a major class of prokaryotic multi-protein sensory signal transduction system that share some analogous characteristics of LRS. One of the defining features of ECF systems is their utilization of ECF σ factors, which are distinct from those in the σ70 family, due to their lack of the conserved sigma 3 region (Fig 3A) [9, 35]. Both conserved domain analyses (Fig 3A) and DNA binding characteristics (Fig 2A–2E) clearly indicate that LRS regulators are bona fide transcription factors rather than σ factors, thus precluding them from being part of true ECF systems. Regardless, LRS membrane proteins do share some basic characteristics with most ECF anti-σ factors, as they are similarly sized membrane proteins, lack obvious enzymatic domains, and serve as inhibitors (Figs 1B and 3B) [11, 12]. Interestingly, after screening the genome sequence data of a phylogenetically diverse group of ECF system-encoding bacteria, we identified at least 10 separate Domains of Unknown Function encoded by ECF anti-σ factors, but we were unable to identify a single instance of anti-σ factors encoding either DUF3021 or DUF2154. Thus, this could be one major distinction between anti-σ factors and LRS membrane proteins.
Given the highly conserved features of S. mutans LRS operons, we expanded our search for putative LRS in other species and were surprised to discover that LRS are encoded by a far broader diversity of organisms than previously recognized (Fig 4 and S3 Table). Using a multi-tiered search strategy modeled on the five S. mutans LRS, we were able to identify >4600 putative LRS operons spread amongst the genomes of numerous Gram positive and Gram negative bacteria as well as some archaea (S3 Table). Overall, the majority of identified LRS are encoded within the Firmicutes phylum, which agrees with previous observations [25]. Of the five S. mutans LRS, the BrsRM-type LRS exhibits the most diverse distribution and is the most commonly encoded (Fig 4). In all cases, the identified LRS operons are arranged similarly as in S. mutans with the LRS regulator encoded upstream of the membrane inhibitor (S3 Table). We also observed a conservation of ABC transporter genes linked to the SMU_433/434-like and SMU_1070c/1069c-like LRS of other species (Fig 5). The conserved co-occurrence of LRS and ABC transporter genes suggests that the respective encoded proteins all function together in related genetic pathways. However, this was not the case for the genes surrounding the SMU_294/295-type LRS, as only the very closely related species Streptococcus troglodytae contained a similar 4-gene operon (Fig 5). Therefore, the 4-gene operon structure of the S. mutans SMU_294/295 LRS (Fig 1A) is presumably either incidental or a niche-specific adaptation. Intriguingly, the LRS operons of other organisms all share highly analogous promoter regions to those of S. mutans LRS indicating that they similarly function via positive feedback autoregulation. Table 3 illustrates some of the diversity of LRS operon promoter elements that can be identified in both bacteria and archaea. Similar to S. mutans, most LRS operon direct repeat sequences are separated by 12 bp, but a minority is separated by either 11 bp or 13 bp. It is also evident there is a particularly strong bias for the direct repeats to be oriented 10 bp upstream of -35 sequences. The S. mutans LRS operons are somewhat unusual in this regard, as only the SMU_433/434 LRS operon contains direct repeats located exactly 10 bp upstream of the operon promoter. We also used Protter [36, 37] to illustrate the predicted topologies of S. mutans LRS membrane proteins to their corresponding weakest homology examples shown in Fig 5 and all yielded highly similar structures despite their limited sequence similarities (S3A–S3E Fig). Overall, the data indicate that most of the identified LRS in S3 Table are highly likely to be orthologs of the S. mutans proteins. While searching for putative LRS in other species, we also encountered a number of potentially novel LRS-types that are not found in S. mutans. The LRS listed in Table 3 for Staphylococcus aureus, Listeria monocytogenes, and Treponema bryantii have characteristics that are all nearly identical to the LRS found in S. mutans, except that their LRS membrane proteins exhibit no obvious homology to those of S. mutans. For the S. aureus membrane protein SACOL_RS12400, its predicted topology is also obviously distinct from the five S. mutans LRS membrane proteins (S3F Fig). Furthermore, members of the Bacteroides fragilis group, such as B. thetaiotaomicron and B. ovatus, encode “LRS-like” operons (Btheta7330_RS19920/RS19915 and Bovatus_RS21370/RS21375) that exhibit a number of distinctions from S. mutans LRS. These Bacteroides operons encode the membrane proteins upstream of the LytTR Family regulators. Unlike S. mutans LRS, the encoded membrane proteins contain two conserved domains, an NfeD-like domain in addition to DUF2154, which is the same domain found in the S. mutans LRS membrane protein HdrM (Figs 3B and S3G). The LytTR Family regulators encoded in the Bacteroides operons are also unusual, as they contain multiple transmembrane segments before the DNA binding domains, whereas all of the S. mutans-type LRS encode soluble transcription regulators (Fig 3A). The intergenic regions of the Bacteroides LRS-like operons also contain 11 bp direct repeats separated by 11 bp of intervening sequence with the repeats located 11 bp upstream of the operon promoters [38, 39] (Table 3). Presumably, these repeats similarly function in autoregulatory transcription activation of the operons. The presence of these distinct LRS-like operons indicates that additional uncharacterized varieties of LRS are likely to exist.
As mentioned previously, little is known about the environmental and/or cellular signals that naturally activate LRS from their basal inactive states. Given the broad distribution and conservation of LRS, it was of interest to gain further insight into LRS activation, as similar mechanisms may exist in other organisms. We created a mariner transposon library of >10,000 mutants to screen for mutations that could trigger activity from a transcription fusion brsRM-gusA β-glucuronidase reporter strain. We selected the brsRM LRS for several reasons: 1) we have previously studied the BrsRM LRS [29], 2) BrsRM is the most stringently regulated LRS (Fig 1B), and 3) BrsRM is the most broadly distributed LRS (Fig 4). Prior to transposon mutagenesis, we deleted all other LRS from the brsRM-gusA reporter strain to eliminate any potential impact of cross-regulation between LRS (Fig 1C and 1D). After screening the library, we initially identified 49 transposon mutants that exhibited various levels of β-glucuronidase activity. We retransformed these mutations into the parent brsRM-gusA reporter to assess reproducibility and then identified the insertion sites of clones exhibiting β-glucuronidase reporter activity (S4 Fig). The final list of 11 distinct brsRM-inducing mutations is shown in Table 4. We next introduced these same mutations into a brsRM-gusA transcription fusion reporter strain in which the brsRM ORFs were replaced by gusA (i.e. ΔbrsRM). In the ΔbrsRM background, only the rgpD and SMU_2060–2061 intergenic region (IGR) mutants still exhibited obvious reporter activity (Table 4), suggesting these two mutations increase brsRM operon expression independent of BrsR autoregulation (i.e. the BrsRM LRS is not required). The remaining 9 mutations in Table 4 do require BrsRM to induce reporter activity and are therefore likely to function via the activation of the BrsRM LRS. Of these, we were next interested to determine whether there is any common theme or pathway among them that might yield clues as to the source of their BrsRM activation phenotypes. After testing various hypotheses, ultimately, it was purine metabolism that proved to be a key aspect of BrsRM activation. Since several of the genes listed in Table 4 have either verified or predicted roles in purine metabolic processes (tilS, mnmE, and SMU_1297), purines were among the numerous reagents tested for brsRM-gusA reporter activity using chemically defined medium agar plates. As shown in Fig 6A, in adenine/guanine drop-out medium, the reporter strain exhibited no obvious response after four days of incubation. In contrast, low concentrations of adenine and guanine both served as potent activators of the reporter. Interestingly, reporter activity increased concomitantly with adenine concentration, whereas the opposite was observed with guanine (Fig 6A). We repeated the purine experiment using the mutant strains listed in Table 4 and all but the SMU_1297 mutant exhibited obvious reporter activity after incubating for only two days in the presence of adenine, and to a lesser extent, guanine as well (Fig 6B–6E). Despite the lack of reporter activity from the SMU_1297 mutant, this strain still exhibited an intriguing response to adenine, as it was the only mutant likely exhibiting adenine auxotrophy (Fig 6D and 6E). Thus, SMU_1297 is presumably an unrecognized key component of purine metabolism. Similarly, both the rpoB and rgpD mutants grew poorly on defined medium in the absence of purine supplementation, whereas both grew normally on complex medium. It is worth noting that the rpoB mutant likely encodes a partially functional RpoB protein, as the transposon insertion occurred near to the 3’ of the rpoB ORF (S4 Fig). This reduced functionality is apparently problematic for growth on chemically defined medium, as only a fraction of the rpoB mutant cells was able to grow in this condition (Fig 6B–6E). Despite this, the rpoB mutant as well as the tilS mutant were the only ones to exhibit obvious brsRM expression in the absence of purines, although purine supplementation could still further augment their reporter activity like most of the other mutants (Fig 6B–6E). Overall, these results support a major role for purines (especially adenine) as mediators of BrsRM activation.
The current study provides the first insights into a widely conserved, but almost entirely uncharacterized group of prokaryotic sensory systems. In S. mutans, these systems, termed LytTR Regulatory Systems, are included within its core genome (Table 1) and control diverse regulons as well as a cell death pathway [28, 29]. The key features of LRS are distinct from the other 2-protein sensory systems (TCSTS and ECF σ factor systems) (Fig 3A and 3B) suggesting LRS possibly represent a novel class. Despite the large number of putative LRS operons we identified amongst both bacteria and archaea, the true breadth and diversity of LRS is likely to be underestimated, as our analyses were performed using S. mutans LRS as model systems, due to the current lack of relevant studies in other species. For example, in the MRSA strain S. aureus COL, the two-gene operon SACOL_RS12395/RS12400 encodes a putative LytTR Family regulator upstream of a DUF3021-containing membrane protein and the operon contains typical LRS repeats located 9 bp upstream of the operon -35 sequence (Table 3). However, the LRS membrane protein SACOL_RS12400 lacks significant sequence similarity to those of S. mutans LRS and it exhibits a distinct predicted topology as well (S3F Fig). Despite this, the putative SACOL_RS12395/RS12400 LRS is widely encoded among the staphylococci and many other Gram positive species. A similar result can be observed from the lmo0984 –lmo0987 operon of L. monocytogenes, except this operon also includes an ABC transporter much like those associated with the SMU_433/434 and SMU_1070c/1069c LRS of S. mutans (Fig 5). Whether these LRS are weak homology orthologs of S. mutans LRS or represent entirely distinct varieties of LRS remains to be determined. However, protein topology predictions suggest the latter scenario is more likely to be the case (S3A–S3F Fig). Furthermore, we have also encountered a number of “LRS-like” operons that are analogous, but clearly distinct from those of S. mutans or the aforementioned unclassified LRS from S. aureus and L. monocytogenes. Such operons can be found among members of the Bacteroides fragilis group, such as B. thetaiotaomicron and B. ovatus, and exhibit a unique operon arrangement encoding transcription regulators and membrane proteins unlike those of S. mutans LRS (Table 3 and S3G Fig). Despite the unique qualities of these operons, the obvious parallels to S. mutans LRS suggest that LRS likely exist in a greater variety than is currently recognized.
One of the key features defining LRS control in S. mutans is the autoregulatory positive feedback regulation encoded within the operons. For the HdrRM LRS, this is mediated directly by HdrR and is critically dependent upon its recognition of the direct repeats located upstream of the hdrRM operon promoter (Fig 2A–2E). It is now evident that these direct repeats are not only key to LRS function in S. mutans, but they appear to be a defining feature of most, if not all, LRS encoded by a wide diversity of prokaryotes (Table 3). Among the putative orthologous LRS found in other species, there is low overall sequence conservation of the individual direct repeats, whereas the direct repeat lengths, their spacing, and their locations immediately upstream of LRS operon promoters are all highly conserved (Table 3). Another conserved characteristic of S. mutans LRS is the inhibitory function of LRS membrane proteins, which play key roles in dictating the basal expression levels of LRS operons (Fig 1B). Presumably, it is the inhibitory equilibrium maintained between an LRS membrane protein and its cognate regulator, which is the principal determinant of LRS operon basal expression. The inhibitory function of LRS membrane proteins can also yield misleading results when performing genetic studies of unrecognized LRS, since single mutations of LRS regulators or double mutations of cognate LRS regulators and membrane proteins are both likely to result in wild-type phenotypes [26]. To observe LRS-related phenotypes, one must solely mutate the LRS membrane protein to constitutively activate the system.
Based upon these conserved features of LRS, several inferences can be made regarding their functionality. Firstly, LRS exist in a basal inactive state. A variable, but limited amount of autoregulation is permitted under normal growth conditions (Figs 1B, 2B, 2D and 2E), which would ensure that the cell maintains a minimal abundance of LRS for the detection of relevant stimuli. Upon signal detection, LRS abundance should quickly increase due to positive feedback autoregulation, thus amplifying both the signal detection apparatus as well as the downstream transcriptional response. Secondly, LRS presumably respond to unusual growth conditions and/or environmental stress. This is supported by several observations: 1) LRS exist in a basal inactive state, 2) the HdrRM LRS responds to a rapid switch to high cell density growth conditions [26], 3) purines, which mediate activation of the BrsRM LRS (Fig 6A–6E) are also central transducers of environmental stress signals [19, 21, 22], and 4) DUF2154, which is found in HdrM, is encoded by proteins responding to cell envelope damage [40–42]. These features are also highly reminiscent of ECF systems. Like LRS, ECF systems are maintained in a basal inactive state, due to the inhibitory function of cognate anti-σ factors. Furthermore, ECF systems are similarly dispensable under normal growth conditions [43, 44] and their activation is typically dependent upon positive feedback autoregulation, ultimately triggered by environmental stress [11, 12, 15, 45]. The lack of shared domains between ECF anti-σ proteins and LRS membrane proteins (Fig 3B) as well as the obvious distinctions between σ factors and transcription regulators suggest that ECF systems and LRS evolved independently, although it is conceivable that both systems could be products of convergent evolution.
When examining the distribution of LRS, it is evident that these systems are encoded by a phylogenetically diverse group of Gram positive and Gram negative bacteria and even some archaea (Fig 4). However, their distribution appears highly biased as well with a subset of Firmicutes encoding the majority of LRS, especially the Lactic Acid Bacteria (Fig 4 and S3 Table). It is currently unclear why such a bias exists. This could be partly due to the utility of some LRS for the regulation of bacteriocin genes. Lactic Acid Bacteria are particularly rich sources of diverse bacteriocins that are regulated by LytTR Family-like repeats upstream of the bacteriocin gene -35 sequences [25, 29, 31, 46–51]. Another possibility that is not mutually exclusive with the former could be that LRS are a fairly recent evolutionary innovation originating within the Firmicutes phylum. In which case, a biased overrepresentation in these species would be expected [52]. Certainly, it is also possible, if not likely, that our current view of LRS distribution is reflective of only a subset of LRS as a consequence of our comparisons to S. mutans. In this case, an apparent skewed overrepresentation among the Lactic Acid Bacteria might be simply due to their close phylogenetic relatedness to S. mutans. As mentioned previously, the presence of LRS-like operons in other distantly related organisms hints at the possibility of a greater diversity of LRS than is currently recognized. Further clarity should arise once additional functional data are available from other LRS-encoding species.
All bacterial strains used in this study are listed in S1 Table and were either grown in an anaerobic chamber containing 85% N2, 10% CO2, and 5% H2 at 37°C, a 5% CO2 incubator at 37°C, or cultured with aeration at 37°C. The S. mutans strain UA140 [53] was used as the parent wild-type for all experiments. S. mutans strains were cultured using Todd Hewitt medium supplemented with 0.3% wt vol-1 yeast extract (THYE, Difco) or in chemically defined medium [54], while E. coli strains were cultured with Lennox LB (LB, Difco) medium. For antibiotic selection, cultures were supplemented with the following antibiotics: S. mutans–(10 μg ml-1 erythromycin, 1 mg ml-1 spectinomycin, 0.02 M p-chlorophenylalanine [4-CP], and 800 μg ml-1 kanamycin) and E. coli–(100 μg ml-1 ampicillin, 50 μg ml-1 chloramphenicol, 250 μg ml-1 erythromycin, and 100 μg ml-1 spectinomycin).
All primers used for strain construction are listed in S2 Table. All PCR reactions employed Phusion DNA Polymerase (NEB). PCR amplicons were purified using the Zymo Research DNA Clean & Concentrator-25. All constructs were assembled using an overlap extension PCR (OE-PCR) strategy.
The S. mutans luciferase reporter strains used in Fig 1B were created by inserting the green renilla luciferase ORF immediately downstream of the LRS operons. Briefly, the luciferase open reading frame (ORF) containing the S. mutans ldh (lactate dehydrogenase) ribosome binding site was amplified from the strain ldhRenGSm [55] using the primer pair RenG-F/RenG-R. The ermAM erythromycin resistance cassette was PCR amplified from the plasmid pJY4164 [56] using the primer pair (RenG) erm-F/erm-R. Primers used to amplify the respective upstream and downstream homologous fragments for each reporter construct are as follows: wild-type SMU_294/295 LRS [SMU294-LF/SMU295(RenG)-R and (erm)SMU295-RF/SMU295-RR], SMU_294/Δ295 LRS [SMU294-LF/SMU294(RenG)-R and (erm)SMU294-RF/SMU295-RR], wild-type SMU_1070c/1069c LRS [SMU1070c-LF/SMU1069c(RenG)-R and (erm)SMU1069c-RF/ SMU1070c-RR], SMU_1070c/Δ1069c LRS [SMU1070c-LF/SMU1070c(RenG)-R and (erm)SMU1070c-RF/SMU1070c-RR], wild-type SMU_1854/1855 (hdrRM) LRS [hdrRM159-LF/hdrM(RenG)-R and (erm)hdrM-RF/hdrRM159-RR-2], SMU_1854/Δ1855 (hdrRΔM) LRS [hdrRM159-LF/hdrR(RenG)-R and (erm)hdrR-RF/hdrRM159-RR-2], SMU_2080/2081 (brsRM) LRS [brsM-LF/brsM(RenG)-R and (erm)brsM-RF/brsM-RR], SMU_2080/Δ2081 (brsRΔM) LRS [brsM-LF/brsR(RenG)-R and (erm)brsR-RF/brsM-RR]. All PCR amplicons were purified and mixed in equal molar concentrations and then subjected to a 4-fragment OE-PCR reaction using the respective upstream forward/downstream reverse primer pairs. The assembled PCR amplicons were transformed into S. mutans strain UA140 and selected on agar plates supplemented with erythromycin to obtain the following strains: 294-295-RenG, 294-RenG, 1070c-1069c-RenG, 1070c-RenG, hdrRM-RenG, hdrR-RenG, brsRM-RenG, and brsR-RenG. The wild-type SMU_433/434 and SMU_433/Δ434 LRS luciferase reporter constructs were PCR amplified from strains 01-luc and 01-luc-434. The resulting PCR amplicons were then transformed into S. mutans strain UA140 and selected on agar plates supplemented with spectinomycin to obtain the strains 433-434-RenG and 433-RenG.
To create markerless in-frame deletions of all 5 LRS in S. mutans UA140, we first deleted SMU_433/434 using our previously described markerless mutagenesis protocol [57]. Two fragments corresponding to the upstream and downstream regions of the SMU_433/434 operon were amplified with the primer pairs SMU433-LF/(IFDC2)smu433-LR and (IFDC2)smu434-RF/SMU434-RR, respectively. The IFDC2 cassette was amplified from the plasmid pIFDC2 [57] using the primer pair ldhF/ermR. The three fragments were mixed and used as templates for OE-PCR with the primer pair SMU433-LF/SMU434-RR. The resulting OE-PCR product was transformed into UA140 and selected on medium containing erythromycin to isolate transformants containing the IFDC2 cassette. Next, DNA fragments containing the SMU_433 upstream region and SMU_434 downstream region were amplified with the primer pairs SMU433-LF/smu433-LR2 and smu434-RF2/SMU434-RR. The two fragments were mixed and assembled with OE-PCR using the primer pair SMU433-LF/SMU434-RR. The OE-PCR amplicon was then transformed into the IFDC2-containing strain and selected on the medium containing p-chlorophenylalanine (4-CP) to remove the IFDC2 cassette and obtain the markerless deletion mutant. This strain was then used as a recipient for the sequential deletion of SMU_1070c/1069c, SMU_294/295, hdrRM, and brsRM using the same approach to obtain the final 5 LRS deletion strain ifdLRS.
Genomic DNA from strains 294-295-RenG, 1070c-1069c-RenG, hdrRM-RenG, brsRM-RenG, and 433-434-RenG were transformed into strain ifdLRS and selected on THYE plates contains erythromycin or spectinomycin to obtain the single LRS luciferase reporter strains ifdLRS/294-295-RenG, ifdLRS/1070c-69c-RenG, ifdLRS/hdrRM-RenG, ifdLRS/brsRM-RenG, and ifdLRS/433-434-RenG.
To examine potential cross-regulation between different LRS, ORFs encoding LRS membrane proteins were replaced by a kanamycin resistance cassette using the single LRS luciferase reporter strains as recipients. Briefly, upstream and downstream homologous fragments of SMU_295 were amplified using the primer pairs SMU294-LF/(kan)smu295-LR and (kan)smu295-RF/SMU295-RR as well as UA140 genomic DNA as a template. The kanamycin resistance gene was amplified using the primer pair kan-F/kan-R and plasmid pWVTKs [58] as the template. Three fragments were mixed and assembled with OE-PCR using the primer pair SMU294-LF/SMU295-RR. The OE-PCR amplicon was transformed into the single luciferase reporter strains ifdLRS/1070c-69c-RenG, ifdLRS/hdrRM-RenG, ifdLRS/brsRM-RenG and ifdLRS/433-434-RenG to obtain d295/1070c-69c-RenG, d295/hdrRM-RenG, d295/brsRM-RenG and d295/433-434-RenG. A similar approach was used to delete hdrM and brsM in each of the single LRS reporter strains. The SMU_434 and SMU_1069c mutations were PCR amplified from d-smu434/UA140 and d-smu1069/UA140 and then transformed into the single LRS reporter strains.
The S. mutans firefly luciferase reporter strains used in Fig 2 were created using a markerless mutagenesis approach. To create the markerless replacement of the hdrRM ORFs with that of luciferase, we first created an allelic replacement of the hdrRM ORFs with the counterselectable IFDC2 cassette [57]. Using UA140 genomic DNA as a template, two fragments corresponding to the upstream and downstream regions of the hdrRM operon were amplified with the primer pairs hdrRupF/hdrRupR-ldh and hdrMdnF-erm/hdrMdnR, respectively. The IFDC2 cassette was amplified using the primer pair ldhF/ermR. The three fragments were mixed and used as template for OE-PCR with the primer pair hdrRupF/hdrMdnR. The resulting OE-PCR product was transformed into UA140 and selected on medium containing erythromycin to obtain strain RMIFDC2. Next, a DNA fragment containing the hdrR upstream region and firefly luciferase ORF was amplified with the primer pair hdrRupF/lucR-1856 and strain LZ89-luc [26] as a template. Using strain UA140 as a template, a fragment corresponding to the hdrM downstream region was amplified with the primer pair 1856F-luc/hdrMDnR. The two fragments were mixed and assembled with OE-PCR using the primer pair hdrRupF/hdrMdnR. The OE-PCR amplicon was transformed into strain RMIFDC2 and selected on medium containing p-chlorophenylalanine (4-CP) to obtain strain RpLuc. To create strains Rp+1luc and Rp-10mluc, the upstream and downstream regions of the hdrRM operon were amplified from strain UA140 with the primer pairs hdrRupF/(luc)hdrRp-R or hdrRupF/(luc)hdrRp-10-R and (lucR)hdrMdn-F/hdrMDn-R, respectively. The luciferase ORF was amplified from strain RpLuc with the primer pair lucF/lucR. The three fragments were mixed and used as template for OE-PCR with the primer pair hdrRupF/hdrMdnR. OE-PCR products were transformed into RMIFDC2 and selected on medium containing 4-CP to obtain the strains Rp+1luc and Rp-10mluc. Strains Rp+1luc and Rp-10mluc were both transformed with the plasmid pHdrRoe [27] to create the strains Rp+1lucROE and Rp+1lucROE-10. Using the genomic DNA from strain RpLuc as a template, two fragments were amplified with the primer pairs hdrRupF/(repeat-m)hdrR-LR and (repeat-m)hdrR-RF/hdrMDnR. The two PCR amplicons were mixed with hybridized EMSA-hdrRpm-F/R primers and assembled using OE-PCR with the primer pair hdrRupF/hdrMdnR. The OE-PCR amplicon was transformed into strain RMIFDC2 and selected on medium containing 4-CP to create the strain RpDRmluc. To create the hdrR ectopic overexpression plasmid pJYROE, a fragment containing the hdrR ORF fused to the ldh promoter was first amplified from pHdrRoe using the primer pair ldhF-bamHI/hdrRR-hindIII. The resulting PCR amplicon was digested with BamHI and HindIII and then ligated to pJY4164 to obtain the suicide vector pJYROE. To create the hdrM ectopic overexpression plasmid pMOE, an ldh promoter-hdrM transcription fusion was assembled by first PCR amplifying the ldh promoter and hdrM ORF using the primer pairs ldhF-BamHI/ldhR-SpeI and hdrMF-SpeI/hdrMR-EcoRI as well as UA140 gDNA as a template. The resulting amplicons were then digested with BamHI/SpeI and SpeI/EcoRI and subsequently ligated to the BamHI/EcoRI restriction sites of the E. coli-Streptococcus shuttle vector pDL278 [59] to create the plasmid pMOE. The suicide vector pJYROE was transformed into strain RpLuc or RpDRmluc to create the strains ROE or ROE/DR-, while the shuttle vector pMOE was transformed into strain ROE to obtain the strain RMOE.
To insert the luciferase ORF downstream of the hdrRM ORFs, a DNA fragment containing the hdrR upstream region and IFDC2 were PCR amplified from strain RMIFDC2 with the primer pair hdrRupF/ermR-lucf. Using the genomic DNA of RpLuc as a template, the luciferase ORF was amplified with the primer pair lucF-erm/lucmR. The two amplicons were assembled using OE-PCR and the primer pair hdrRupF/lucmR. The resulting overlapping PCR products were transformed into RpLuc strain and selected on medium containing erythromycin to obtain the strain RMlucIFDC2. Next, two fragments encompassing the hdrRM locus were amplified from strain UA140 with the primer pair hdrRupF/MterR-luc, while the luciferase ORF was amplified from strain RpLuc with the primer pair lucF-Mter/lucmR. The PCR amplicons were mixed and assembled by OE-PCR using the primer pair hdrRupF/lucmR. The resulting OE-PCR amplicon was transformed into strain RMlucIFDC2 and selected on plates supplemented with 4-CP to obtain the strain hdrRMluc. To mutate hdrM in strain hdrRMluc, three fragments were amplified from this strain using the primer pairs hdrRupF/(spec)smu1853R, (spec)smu1853-hdrR-LF2/hdrM(TAA)R, and hdrM(TAA)F/lucmR. The spectinomycin resistance cassette aad9 was amplified from the E. coli-Streptococcus shuttle vector pDL278 [59] using the primer pair specF/specR. The four amplicons were mixed and assembled by OE-PCR using the primer pair hdrRupF/lucmR. The resulting OE-PCR amplicon was transformed into strain hdrRMluc to obtain the strain dhdrMluc. To mutate the direct repeats upstream of the hdrRM promoter in strain dhdrMluc, two fragments were amplified from this strain using the primer pair hdrRupF/(repeat-m)hdrR-LR and (repeat-m)hdrR-RF/lucmR. The two PCR amplicons were mixed with hybridized EMSA-hdrRpm F/R primers and assembled using OE-PCR and the primers hdrRupF/lucmR. The resulting OE-PCR amplicon was transformed into strain hdrRMluc to obtain the strain dhdrMdDRluc.
To create markerless gusA transcription fusions to the brsRM operon, a brsRM upstream homologous fragment was amplified from strain UA140 or ifdLRS using the primer pair brsRM-LF/(gusA)brsRM-LR, while the brsRM downstream homologous fragment was amplified from strain UA140 using the primer pair (gusA)brsRM-RF/brsRM-RR. The gusA ORF was amplified from plasmid pZX7 [60] using the primer pair GusA-F/GusA-R. The three amplicons were assembled via OE-PCR with the primer pair brsRM-LF/brsRM-RR. The two resulting OE-PCR amplicons were then transformed into the strain ifdLRS/brsRM(IFDC2) and selected on the medium containing 4-CP to obtain the strains ifdLRS/brsRM-gusA and ifdLRS/brsRMp-gusA respectively.
The ifdLRS/brsRM-gusA reporter strain transposon library was generated by a previously described transposon mutagenesis protocol [61]. Briefly, the primer pair MmeI-MGL-erm-F/MmeI-MGL-erm-R was used to amplify the erythromycin resistance cassette from plasmid pJY4164. Sequences at the 5’ ends of both primers add repeat sequences recognized by the himar transposon onto both ends of the PCR amplicon. The resulting amplicon was then ligated to the pGEM®-T vector (Promega) to obtain pT-MGL-erm. In vitro transposon mutagenesis was performed by combining MarC9 transposase, genomic DNA from strain ifdLRS, and plasmid pT-MGL-erm and then incubating at 30°C for 1 h. Transposon junctions were subsequently repaired and then the transposition reaction was transformed into strain ifdLRS/brsRM-gusA. Transposon mutants were selected on THYE plates containing erythromycin and 5-bromo-4-chloro-3-indolyl-β-D-glucuronic acid (X-gluc, 200 μg ml-1). After 5 days of incubation, blue colonies were selected. Transposon insertion sites were mapped according to the published protocol [61], except that PCR amplicons were ligated into the pGEM®-T vector, transformed into E.coli DH5α, and then the resulting plasmid inserts were sequenced. PCR was used to confirm the expected locations of transposon insertions sites in each of the mutant strains. Genomic DNA from confirmed transposon mutants was also transformed into strain ifdLRS/brsRMp-gusA (ΔbrsRM) to compare its reporter activity with the corresponding transposon mutants obtained in the ifdLRS/brsRM-gusA (brsRM+) background.
The hdrR ORF was amplified from strain UA140 using the primer pair hdrRF-NdeI/HdrRR-Hind. The amplicon was then digested with NdeI/HindIII and ligated to the expression vector pET29b to create the plasmid pEcROE.
Recombinant HdrR was purified using pET29b and the E. coli BL21(DE3) pLysS expression system. Cultures were grown to OD600 0.6 at 37°C with aeration before adding 0.1 mM IPTG and culturing for an additional 12 hr. at 20°C. Cells were harvested by centrifugation (6000 x g, 5 min, 4°C), washed twice with binding buffer (20 mM Tris, 300 mM NaCl, 5 mM imidazole, 10% glycerol, pH 7.9) and then resuspended in 20 ml of the same buffer. Next, the cells were chilled on ice, lysed by sonication, centrifuged to recover supernatants (20,130 x g, 20 min, 4°C), and then HdrR-His6 was purified using Ni-NTA agarose chromatography (Novagen). Proteins were eluted with 4 ml elution buffer (20 mM Tris, 300 mM NaCl, 500 mM imidazole, 10% glycerol, pH 7.9) and concentrated by ultrafiltration (Millipore membrane, 3 kDa cut-off size). Purified proteins were stored in 10% glycerol at -80°C.
EMSAs were performed similarly as previously described [62]. Briefly, double-stranded probes were obtained by annealing equal molar concentrations of two oligonucleotides (S2 Table) in 50 mM Tris-HCl (pH 8.0), 10 mM MgCl2, 50 mM NaCl and 1 mM EDTA, with the forward primer 5′-end labeled with digoxigenin-11-ddUTP (Roche). The oligonucleotide pair EMSA-hdrRp-F/EMSA-hdrRp-R served as the wild-type probe, while the oligonucleotide pair EMSA-hdrRpm-F/EMSA-hdrRpm-R served as the direct repeat mutant probe. 1 ng of DNA probe was incubated individually with various concentrations of HdrR-His6 at 25°C for 20 min in a 20 μl reaction volume. After incubation, the reaction mixtures were separated by electrophoresis and electro-transferred to nylon membranes. Images were detected using chemiluminescence and X-ray films. For competition experiments, 50- and 200-fold excess of unlabeled probes (S2 Table) were added to the binding reactions before performing electrophoresis and imaging as described above.
Assays of firefly and green renilla luciferase activity were performed using a previously described methodology [55] with mid-log phase cultures. Reporter data were normalized by dividing luciferase values by their corresponding optical density (OD600) values. Luciferase activity was measured with a GloMax Discover 96-well luminometer (Promega).
To identify homologs of LRS membrane proteins, we searched the NCBI non-redundant nucleotide collection (nr/nt) and whole-genome shotgun (wgs) databases using tBLASTn (E-value <10, >25% positives). These putative LRS membrane proteins (except for SMU_295 homologs) were then refined contingent on containing either DUF3021 or DUF2154 domains, as determined by NCBI RPS-tBLASTn (E-value <1). Qualifying LRS membrane protein results were further filtered based upon the presence of adjacent upstream LytTR Family transcription regulator homologs identified using tBLASTn (E-value <0.1).
To assess the effect of purines on the BrsRM LRS, overnight cultures of ifdLRS/brsRM-gusA and isogenic transposon mutants were harvested by centrifugation, washed thrice with an equal volume of 0.9% NaCl, and spotted on adenine/guanine-replete or adenine/guanine drop-out chemically defined medium (CDM) agar plates [54]. Different concentrations of adenine (0 mM, 0.075 mM, 0.15 mM, 0.3 mM and 0.6 mM) or guanine (0 mM, 0.066 mM, 0.132 mM, 0.264 mM and 0.53 mM) were added to the CDM medium and plates were incubated at 37°C with 5% CO2 for 4 days. To assay the impact of purines on the transposon mutants of ifdLRS/brsRM-gusA, adenine and/or guanine was added to the CDM at a final concentration of 0.15 mM and/or 0.132 mM, respectively. Plates were incubated at 37°C with 5% CO2 for 2.5 days.
All statistical analyses were performed using GraphPad Prism software to calculate significance via two-tailed Student’s t-tests with Welch’s correction. Statistical significance was assessed using a cutoff value of P < 0.05.
|
10.1371/journal.ppat.1006091 | A Precise Temperature-Responsive Bistable Switch Controlling Yersinia Virulence | Different biomolecules have been identified in bacterial pathogens that sense changes in temperature and trigger expression of virulence programs upon host entry. However, the dynamics and quantitative outcome of this response in individual cells of a population, and how this influences pathogenicity are unknown. Here, we address these questions using a thermosensing virulence regulator of an intestinal pathogen (RovA of Yersinia pseudotuberculosis) as a model. We reveal that this regulator is part of a novel thermoresponsive bistable switch, which leads to high- and low-invasive subpopulations within a narrow temperature range. The temperature range in which bistability is observed is defined by the degradation and synthesis rate of the regulator, and is further adjustable via a nutrient-responsive regulator. The thermoresponsive switch is also characterized by a hysteretic behavior in which activation and deactivation occurred on vastly different time scales. Mathematical modeling accurately mirrored the experimental behavior and predicted that the thermoresponsiveness of this sophisticated bistable switch is mainly determined by the thermo-triggered increase of RovA proteolysis. We further observed RovA ON and OFF subpopulations of Y. pseudotuberculosis in the Peyer’s patches and caecum of infected mice, and that changes in the RovA ON/OFF cell ratio reduce tissue colonization and overall virulence. This points to a bet-hedging strategy in which the thermoresponsive bistable switch plays a key role in adapting the bacteria to the fluctuating conditions encountered as they pass through the host’s intestinal epithelium and suggests novel strategies for the development of antimicrobial therapies.
| The ability of pathogens to sense temperature changes when they enter their mammalian hosts from the environment is crucial to optimize their fitness and adjust expression of their virulence programs. Until now it has been assumed that all cells within a population participate in the thermo-triggered adaptive response. Here, we show that a small subpopulation of an enteric pathogen does not follow thermo-induced reprogramming when the bacteria pass the intestinal epithelial layer. Observed heterogeneity is promoted by a new type of bistable switch, implicating a highly precise, thermoresponsive control element. Moreover, we demonstrate that this regulatory implement is important for virulence as it prepares the pathogen for sudden, unpredictable fluctuations encountered during host entry and exit.
| Temperature is a prominent signal used by pathogens to adjust their virulence and host survival programs during infection. Different biomolecules can act as thermosensors, including DNA, RNA and regulatory proteins. They all detect changes in temperature through thermally induced conformational changes [1–3]. The velocity and reversibility of thermosensors enable rapid adaptation to the temperature shifts encountered when transitioning between different hosts or environments. The precise thermosensation mechanism of several molecular thermometers was uncovered using population level analyses. However, bulk-scale methods are insufficient for characterizing key features of this process, such as sensor dynamics and quantitative outcome in individual cells. Here, we addressed these features by single-cell level analyses using the Yersinia regulator protein RovA as an example for a thermosensing molecule that controls virulence [4, 5].
This approach is important as during transition processes genetically identical populations can generate phenotypic heterogeneity, which supports persistence of pathogens in fluctuating environments (bet-hedging) via fitness improvement of the whole population by cooperativity or division of labor [6–11]. One example is bistability, in which isogenic bacteria exist in two distinct phenotypic states (ON or OFF) driven by divergent gene expression profiles in response to nutrient shifts and stress conditions [6, 10, 12–14]. A binary distribution of phenotypes can be generated by feedback-based circuitry in combination with non-linear responses, e.g. by cooperativity in DNA binding of a regulator, [6, 14], a characteristic also observed for the thermoresponsive virulence regulator RovA. RovA is active and autoregulated at moderate temperatures (20–25°C) and binds cooperatively to a high-affinity site upstream of the distal rovA promoter (P2) and activates rovA and invA transcription. When the RovA amount has reached a certain threshold, RovA binds to a low affinity site downstream of the proximal rovA promoter (P1) to prevent uncontrolled rovA induction (Fig 1A). An upshift to 37°C induces a reversible conformational change in RovA that leads to a strong reduction of its DNA-binding capacity and renders this regulator susceptible to proteolysis by the Lon protease [4, 5, 15] (Fig 1A). Since autoregulatory features, which can generate a bistable output of a genetic system, are combined with a thermosensing element [12–14], we hypothesized a novel type of ‘thermo-controllable’ bistable switching device for the control of Yersinia virulence.
To prove our hypothesis we first tested for the occurrence of distinct bacterial subpopulations by measuring rovA expression at a single-cell level. The rovA promoter was fused to egfpLVA, encoding a green fluorescence protein derivative (eGFPLVA) with a brighter fluorescence but reduced stability. Upon shifting from 25°C to 37°C, eGFPLVA-expressing Yersinia demonstrated successive reduction in eGFPLVA synthesis, that corresponded to the average RovA level (Fig 1B–1D). Two distinct subpopulations showing no (OFF) or high (ON) eGFPLVA production at growth temperatures between 30°C and 34°C were detected in the wild-type (Fig 1B–1D). No ON subpopulation could be detected when rovA-eGFPLVA was expressed in a rovA mutant, confirming that expression of the reporter depends on active RovA (S1A Fig). Immunofluorescently labeled RovA-dependent adhesin InvA [16] exhibited a similar bimodal staining pattern (S1B and S1C Fig). Time-lapse microscopy revealed that individual bacteria can spontaneously switch (average time of 2–3 h at 32°C) from one state to the other, demonstrating reversibility of the switching process (Fig 1E, S1–S3 Videos). We quantified the switching dynamics by measuring rovA expression and intracellular RovA amounts under stable physiological conditions in chemostat cultures over many generations. Quantification of bacteria in the RovA ON and OFF state after transitioning between 25°C and 37°C revealed a dependence of the system’s output on present and past inputs (hysteretic behavior) and showed that activation and deactivation of RovA synthesis occurred at strikingly different times scales (Fig 2A and 2B). Thermal upshift caused a rapid decrease in rovA expression with a bimodal RovA distribution and a continuous decrease in the RovA+ subpopulation over 3–4 h (Fig 2A and 2B). In contrast, activation of rovA was delayed and the RovA+ population increased very slowly upon thermal downshifting, indicating that the remaining amount of RovA at 37°C was insufficient to allow rapid autoinduction. In summary, this demonstrated the presence of a new, highly precise thermoresponsive bistable switch with an exceptional hysteretic behavior.
We devised mathematical models to derive information about the underlying drivers dictating the temperature-dependent bistability of RovA (Fig 3A–3C, S1 Text, S2–S4 Figs). Our deterministic model is based on ordinary differential equations for the temporal change in RovA concentrations (dr/dt) in response to temperature (Τ) in a continuous deterministic manner. The temporal change of RovA concentration was described by a sigmoidal regulation function with a basal permanent RovA production rate α0 and a RovA-induced RovA production rate α. The feedback loops were coupled and influenced by an activating DNA-binding constant ka and a repressive DNA-binding constant kr. Cooperative RovA binding was included by the Hill coefficients ha and hr. The RovA degradation rate was included as δ. Our experimental results revealed that the DNA binding constants and the degradation rate of RovA were temperature-dependent, and thus a function of the temperature, described by Τ (ka(Τ), kr(Τ), and δ(Τ)) which leads to the resulting model:
drdt=α0+α⋅rhaka(Τ)ha+rhakr(Τ)hrkr(Τ)hr+rhr−δ(Τ)⋅r
We used experimentally determined kinetic parameters to calculate the corresponding values for all temperatures and nonlinear regression to estimate the DNA-binding constants, Hill coefficients and degradation rates. To obtain the production rates α0 and α, we carried out stochastic modeling to fit data obtained by temperature shift experiments (S1 Text, S3 Fig).
A stochastic, individual-based version of the deterministic model was used to elucidate the mechanisms determining hysteresis (Fig 3A). The parameters obtained from chemostat experiments (α = 0.7 nM/min and α0 = 0.002 nM/min) predict a number of approximately 35 nM of free RovA per cell (≈25 RovA dimers), which contribute to rovA regulation (S1 Text, S3 Fig). Determination of RovA molecule numbers in Y. pseudotuberculosis expressing the ProvA-egfpLVA fusion at 25°C revealed an average of 400 RovA molecules per cell, which corresponds to approximately 275 nM RovA (S3B Fig). A higher concentration of RovA molecules than the predicted 35 nM is expected, since not all RovA molecules within the bacterial cell are available for autoregulation, as (i) only a fraction of RovA molecules is in the active form and (ii) a certain number of RovA dimers is also likely to be bound at different locations on the bacterial chromosome. Furthermore, the bacterial population is still in the ON state at 30°C (Fig 1D), while RovA amounts are considerably decreased compared to 25°C (40–50%). This indicates that less than 275 nM of RovA is sufficient to trigger RovA autoinduction in the entire population.
Our model further predicts that net production of RovA tends to be zero at 37°C as a result of the loss of DNA-binding and increased degradation rate. Consequently, RovA amounts are rapidly (within 3 h) reduced to only a few molecules. Upon downshift to 25°C, protein activation combined with the positive feedback loop can reactivate RovA synthesis, but the positive circuit is active only when sufficient RovA molecules cooperate. A simulation of the autoactivation circuit with six RovA molecules (S1 Text) correlates perfectly with the experimental RovA data (S3E Fig). When less than six active RovA molecules are present per bacterial cell and the net production rate is very low, the population basically follows a neutral birth-death process until the critical number of active RovA molecules is produced through stochastic processes, including random fluctuations and transcriptional noise. Once this threshold is reached, the bacterial cell switches rapidly to the RovA ON state due to the positive feedback loop. Because of the time interval required to reach the critical RovA number by stochastic forces, a significantly longer time period is needed to drive the population from the RovA OFF into the RovA ON state (Figs 2 and 3A, S1 Text, S3 Fig).
A stimulus-response diagram generated by calculating the steady-state concentration of RovA at various temperatures revealed bistable response behavior with hysteresis from 27°C to 37°C highly similar to the experimental data (Fig 3C). The model predicted that bistability was mainly caused by the positive feedback loop, whereby the inhibitory RovA binding site reduced only the response time and degree of bistability (Fig 1A, S1 Text, S4 Fig). A numerical approach was used to describe the influence of δ, α and α0 on the bistable behavior (S1 Text, Fig 3D). The model predicted that the thermally induced increase in degradation was crucial for temperature-responsiveness, and that α0 was critical for RovA bistability, whereby either extremely high or extremely low degradation and production rates abolished bistability and maintained the system in a monostable state.
To challenge our analysis and mathematical predictions, we first proved whether the rovA regulatory region is essential for bistability. We replaced the rovA promoter in ProvA-egfpLVA with the constitutive Prho promoter, analyzed eGFPLVA expression in Y. pseudotuberculosis strain YPIII and found that the substitution of ProvA by Prho eliminated bistability and resulted in a unimodal population with strong eGFPLVA production from 25°C to 37°C (S5A–S5C Fig). Moreover, we tested the influence of different RovA mutant proteins on bistability and found that RovA variants carrying amino acid substitutions in the thermosensing region (G116A, SG127/128IK), the Lon protease recognition site (P98S) and their combination (P98S/SG127/128IK/G116A) [5] did not abrogate bimodal rovA expression. The overall eGFPLVA intensity of the different ON subpopulations was comparable (S5D Fig), but the temperature range for bistability was broader and shifted toward higher temperatures (Figs 1D and 4). Notably, a mutation eliminating the Lon recognition site (P98S) of RovA had no or only a very weak influence on bimodal rovA expression at higher temperatures. This can be explained by the fact that at 37°C the majority of RovA dimers targeted by Lon is inactive, i.e. RovA is partially defolded which abolishes its DNA-binding functions and autoactivation [4].
As shown in Fig 1C and 1D Yersiniae harboring the rovA-eGFPLVA reporter were predominantly in the OFF state at 37°C in vitro. However, rovA transcripts were identified in infected lymphatic tissues of mice by in vivo RNA-Seq analysis, in particular during their persistence stage in the caecum [26], indicating that additional parameters induce rovA transcription during infection. It is known that rovA expression is strongly affected by changes in carbon source availability involving the carbon storage regulator system (Csr) and the cAMP receptor protein Crp, which are transmitted through the LysR-type regulator RovM (Fig 5A) [17, 18]. We therefore tested whether a deletion of rovM influences the distribution of RovA ON and OFF cells. Strikingly, bimodal expression of rovA was fully preserved but shifted toward higher temperatures (Fig 5). In the absence of RovM a significantly higher amount of RovA ON cells was observed in particular at temperatures ranging from 34°C to 36°C (Fig 5B and 5C). Moreover, a very small fraction of RovA ON cells was detectable at 25°C, but not at all tested higher temperature when RovM was overexpressed (Fig 5D and 5E).
Obviously, the observed bistable phenotype correlates with our mathematical models and appears very robust, as temperature-responsive switching was not abolished by fundamental changes in RovA stability. Moreover, the RovA ON/OFF cell ratio is adjustable by a temperature-independent regulator (RovM). This allows the pathogen to modulate the outcome in host tissues at constant temperatures according to nutrients.
Observed robustness of the bistable switch suggested a bimodal expression of rovA during infection. To obtain direct evidence for phenotypic heterogeneity in vivo, mice were orally challenged with Y. pseudotuberculosis using a dual fluorescence reporter system (Ptet-mCherry, ProvA-egfpLVA). We observed two subpopulations in the Peyer’s patches and the caecal lymph nodes, with low numbers of RovA ON bacteria randomly distributed within microcolonies within tissue lesions (Fig 6A and 6B, S6A and S6B Fig). There was a statistically significant increase in the RovA ON cell population in the caecum when bacteria expressed the thermotolerant variant RovAP98S/SG127/128IK/G116A (Fig 6B), verifying a shift of bistability towards higher temperatures in vivo. In contrast, no eGFPLVA-expressing bacteria were detected in the absence of the rovA promoter or in a rovA mutant strain, demonstrating that in vivo expression of the reporter depends on RovA (S6C and S6D Fig).
Mice were then infected with a lethal dose (2×108 bacteria) of Y. pseudotuberculosis wild-type or mutants producing the more stable RovA variants. Mice infected with wild-type bacteria displayed typical signs of the infection (e.g. weight loss, piloerection and lethargy) after 5–10 days. In contrast, none of the mice infected with a rovA-deficient strain or strains producing stabilized RovA variants developed severe disease symptoms, with 40–60% of the mice still alive after 15 days (Fig 6C). Infections with all mutants resulted in a statistically significant reduction in tissue colonization (Figs 6D and S7). This difference was pronounced for the mesenteric lymph nodes (Fig 6D) from which >1000-fold fewer bacteria producing RovAP98S/SG127/128IK/G116A were recovered. Smaller but statistically significant effects were observed in mice infected with mutants producing moderately stable RovA variants RovAP98S and RovAG116A (Fig 6D) indicating that variation of RovA bistable properties, which increase RovA+ subpopulations reduces pathogenicity. RovA+ cells, which express the colonization factor invasin, can efficiently invade lymphatic tissues [19–21], but its presence also renders the bacteria more susceptible to immune responses [22, 23]. A transcriptome analysis further revealed that also other surface-exposed pathogenicity factors, e.g. the afimbrial adhesin PsaA as well as lipopolysaccharide synthesis genes are activated by RovA of Y. pseudotuberculosis [24]. Although beneficial for the initiation of the infection, they are likely to trigger innate immunity-mediated antimicrobial responses when expressed in deeper tissues. In addition, several general stress adaptation genes (ibpAB, uspA, cspB,C1-3,D,E) are activated by RovA, which could support survival in the lumen of the intestine and/or in the external environment. The analysis of the RovA regulon further uncovered different metabolic programs for the wild-type and a rovA mutant [24], which may endow the RovA OFF population with a better fitness within lymphatic tissues. In fact, multiple enzymes of the pyruvate-TCA cycle (icdA, sucDCB, gltA, acnAB, aceE,F) are down-regulated in a rovA mutant, whereas several enzymes of the amino acid and nucleotide transport and metabolism are induced [24]. Different metabolic programs in the RovA ON and OFF population could contribute to the beneficial effect of bistable RovA expression as they adapt the bacterial metabolism to the distinct nutritional conditions in the intestinal tract or the lymphatic tissues. In summary, the discovered thermo-responsive bistable switch enables expression of an alternative virulence program in a small subpopulation within a single infection site. Transcriptional specialization supports survival and pathogenesis as it primes the bacteria to environmental uncertainty encountered at two critical stages when they cross the intestinal layer: (i) shortly after host entry, when Yersinia colonizes the intestinal tract, of which only a subset invades the Peyer’s patches [25], and (ii) during persistence in the caecum, which is a potential reservoir from which the bacteria re-emerge in the intestinal lumen after expulsion from damaged tissues (Fig 6E) [26, 27]. This new form of bet-hedging complements other types of heterogeneous host-pathogen interactions (i.e. slow-growing variants which are more resistant to antibiotics, or populations subsets formed within the complex tissue landscape as a response to varying local conditions faced outside and inside a bacterial microcolony [28–33]), and (ii) opposes recent approaches to targeting virulence traits such as adhesion and virulence-relevant regulatory processes to combat bacteria-mediated diseases [34–37]. Based on our study, detailed knowledge of present pathogen subsets and their distinct virulence programs including single-cell expression profiles of potential virulence targets in infected tissues are imperative for the development of successful anti-microbial therapies.
The strains used in this work are listed in S1 Table. For batch culture experiments bacteria were routinely grown in Luria-Bertani (LB) broth to exponential growth phase (OD600nm = 0.5–0.6) at temperatures ranging from 25°C to 37°C under aerobic conditions. If necessary, antibiotics were added at the following concentrations: carbenicillin 100 μg ml-1, chloramphenicol 30 μg ml-1 and kanamycin 50 μg ml-1.
All DNA manipulations, transformations, restriction digestions and ligations were performed using standard genetic and molecular methods. The plasmids used in this work are listed in S1 Table. Oligonucleotides used for PCR and sequencing were purchased from Metabion and are listed in S2 Table.
Plasmid DNA was isolated using QIAprep Spin Miniprep Kit (Qiagen). DNA-modifying enzymes and restriction enzymes were purchased from Roche or New England Biolabs. PCRs were done in a 50 μl mix for 29 cycles using Phusion High-Fidelity DNA polymerase (New England Biolabs). Purification of PCR products was routinely performed using the QIAquick PCR Purfication Kit (Qiagen). All constructed plasmids were sequenced by the in-house facility.
For construction of a RovA-dependent gfp reporter, the rovA promoter region along with 170 nts of rovA coding region (-622 to +170) and the egfpLVA gene were PCR amplified from plasmid pYPL using primers 158 and II525. The PCR product was digested with SalI and NotI and ligated with T4-DNA ligase (NEB) into pFU76 of the pFU vector series [38], yielding plasmid pKH87. This plasmid was subsequently digested with SacI and AvrII for the exchange of the R6K origin of replication against the origin 29807 from plasmid pFU33, resulting in plasmid pKH70.
To generate a plasmid for constitutive egfpLVA expression the promoter region of the rho gene was PCR amplified from Y. pseudotuberuclosis YPIII genomic DNA from nucleotide -433 (primer IV490) to nucleotide -21 (primer IV491) and egfpLVA was amplified with primers II525/IV483 from plasmid pKH70. The two fragments were inserted into the same backbone as pKH70 using AatII and NotI yielding plasmid pFS5. To perform Quick-change mutagenesis (Stratagene) on rovA, the rovA+ plasmid pFS6 was generated. To do so, rovA was amplified with primers III784/III947 and ligated into SalI/SphI sites of the pJet1.2 cloning vector (Thermo Scientific). Quick-change mutagenesis of pFS6 was performed using primer pairs II379/II380 and II624/II625 resulting in plasmids pFS7 and pFS14, respectively. The modified versions of the rovA gene from pFS7 and pFS14 were transferred into the suicide mutagenesis plasmid pDM4 using SalI and SphI, yielding pFS8 and pFS16. Furthermore, Quick-change mutagenesis with pFS7 was performed with primers II624/II625 to generate plasmid pFS23, which was subsequently used to perform Quick-change mutagenesis with primers II626/II627 to obtain plasmid pFS24. Subsequently, pFS24 was digested with SalI/SphI and the insert was ligated into the suicide plasmid pDM4 to obtain pFS28. For constitutive expression of Ptet-mCherry, plasmid pFU76 was cut with KpnI/AvrII and ligated into pZE21 resulting in plasmid pFS42. mCherry was amplified with primers V842/V843 from plasmid pTB23, cut with SalI/NotI and ligated into pFS42 generating plasmid pFS43. The origin of replication p15a was amplified from plasmid pAKH120 with primers V521/V522 and cut with AvrII/SacI. Additionally, the chloramphenicol cassette was amplified from pFU228 with primers V519/V520, cut with AatII/SacI and both fragments were ligated into pFS43 to obtain pFS48.
Construction of the Y. pseudotuberculosis rovA mutant strains YP269, YP270 and YP287 was performed by integration of the suicide plasmids pFS8, pFS16 or pFS28 in the rovA locus of strain YP107. E. coli strain S17-1λpir harbouring the plasmids were used for conjugation and the resulting transconjugants were identified by plating on Yersinia selective agar (Oxoid) supplemented with chloramphenicol. Expression of the sacB gene, which is also encoded on the integrated plasmids, is induced when the bacteria are plated on LB agar with 10% sucrose. This results in a growth reduction of the bacteria. Derivatives which have lost the plasmid due to a second recombination event were identified as more rapidly growing clones on 10% sucrose plates and presence of the individual rovA mutant genes was verified by PCR and sequencing with primers 135/151 as described [16].
Batch cultures: Y. pseudotuberculosis YPIII harboring a RovA-dependent egfpLVA reporter (pKH70) was grown over night at different temperatures ranging from 25°C to 37°C in liquid LB broth. A fresh culture was started by inoculating pre-warmed medium with over night culture in a 1:50 dilution and incubated at identical temperatures until the cultures reached an OD600nm = 0.6. Subsequently, 1 ml of culture was harvested by centrifugation for western blotting and flow cytometry. For flow cytometry cell pellets were rapidly fixed in 4% para-formaldehyde for 20 min at 25°C. Pellets were washed twice with 1 x PBS and at least 100.000 cells were analyzed by a LSRII flow cytometer (BD Biosciences). Data were acquired with the FACS Diva software (BD Biosciences) and further analyzed with FlowJo v9.7.2 (Treestar).
Continuous culture: YPIII pKH70 was aerobically grown in Vario 500 mini-bioreactors (Medorex). The bacteria were pre-cultured for 12 h at 25°C under aerobic conditions in LB medium containing carbenicillin (100 μg ml-1). The bioreactor was filled with 210 ml LB medium containing carbenicillin. Antifoam 204 (Sigma-Aldrich), an entirely organic antifoaming agent, was added to the medium at a concentration of 0.02% (vol/vol). Pre-cultures were washed twice in fresh LB medium (25°C). The bioreactor was inoculated with the pre-culture (final OD600nm: 0.2) and run in batch mode at 25°C under continuous stirring (400 rpm). Cultivation was switched to continuous mode (25°C) at a growth rate of μ = 0.32h-1. After crucial processing parameters, i.e. (i) OD600nm: 4.6 ± 0.35; (ii) pO2:—45–55%; and (iii) pH 8.0 remained constant, samples were taken for western blotting and flow cytometry. Subsequently, temperature was shifted to 37°C for an 8 h period and shifted back to 25°C for additional 18 hours. Samples were taken in 1 h intervals for western blotting and flow cytometry. At least 105 fixed cells were analyzed by a LSRII (BD Biosciences) flow cytometer and the data were extracted and analyzed as described above.
Batch cultures: Poly-L-lysine solution (Sigma) was diluted 1:10 in sterile filtered 1 x PBS and spotted onto acid washed microscopy slides (VWR). After 2 hours incubation at room temperature the slides were rinsed with ultra pure water and air dried over night. In 4% para-formaldehyde fixed Yersinia cell suspensions (OD600nm 0.6) were diluted 1:10 in 1 x PBS. This dilution was spotted onto the poly-L-lysine-coated microscopy slides and incubated for 30 min at room temperature. The slides were washed 3 times with 1 x PBS and cells were blocked for 1 h in 1 x PBS containing 2% BSA. For immunostaining of invasin (InvA), slides were then washed with 1 x PBS before addition of the anti-InvA42 monoclonal mouse IgG (1:1.000 in 1 x PBS containing 1% BSA). After 1 h incubation at room temperature the slides were washed 3 times with 1 x PBS. The secondary antibody (goat anti-mouse IgG, Cy5 conjugate, Invitrogen) was added in a 1:1.000 dilution and slides were incubated for an additional hour at room temperature. Slides were washed 3 times in 1 x PBS. Coverslips were mounted using SlowFade Gold (Life Technologies), covered with a glass slide and analysed with an Axiovert II fluorescence microscope (Zeiss) with an Axiocam HR digital charge-coupled device (CCD) camera (Zeiss) and the AxioVision program (Zeiss) and the software ImageJ (https://imagej.nih.gov/ij/).
Infected tissue: Y. pseudotuberculosis YPIII and YP287 harboring a ProvA::egfpLVA fusion (pKH70) and a Ptet::mCherry expression construct (pFS48) as well as Y. pseudotuberculosis YPIII harboring only pFS48, which served as negative control, were grown in LB medium at 25°C overnight. Mice were infected orally with 2 x 108 bacteria. After three days mice were sacrificed by CO2 asphyxiation. For cryosections, the Peyer’s patches and caeca were frozen in Tissue-Tek OCT freezing medium (Sakura Finetek) on dry ice. Sections of 6–10 μm were prepared using a Microm HM 560 cryostat (Thermo Scientific) and mounted on SuperFrost Plus slides (Thermo Scientific). Air-dried sections were fixed for 20 min in ice-cold 4% para-formaldehyde and washed twice with PBS. For visualization of nuclei in the fixed tissue, samples were stained and mounted with Roti Mount Flour Core 49,6-diamidino-2-phenylindole (DAPI, Roth). Tissues were imaged and localization of Yersiniae in the infected tissues was analyzed using a fluorescence microscope (Axiovert II, Zeiss) with 25 x and 40 x objectives, an Axiocam HR digital charge-coupled device (CCD) camera (Zeiss) and the ZEN program (Zeiss). Total number of mCherry-positive bacteria and the number of the cells expressing also eGFPLVA was counted in 40 randomly chosen tissue sections of the Peyer’s patches and the caecum of three infected mice, and the percentage of cells expressing eGFPLVA was calculated.
For time-lapse microscopy bacteria were grown over night in LB medium at 32°C in the presence of carbenicillin. A fresh culture was started by inoculating pre-warmed LB medium with an over night culture in a 1:50 dilution and was grown at the same temperature to OD600nm of 0.6. Subsequently, 1 μl of bacterial culture was distributed on a microdish (IBIDI), overlaid with a thin agarblock (2% LB-agar with carbenicillin) and covered with a glass slide.
For live cell imaging, shutters were computer-controlled, synchronized with the HR camera and opened only during exposure time to reduce photobleaching of eGFPLVA and photodamage of the cells. Starting from single cells or cell doublets, eGFPLVA fluorescence was recorded over several generations. Each imaging cycle consisted of one fluorescence frame to track eGFPLVA expression, followed by one phase-contrast frame to monitor also those cells, which do not express eGFPLVA. The temperature of the microscope chamber was controlled by the Heating Unit XL S and the Incubator XL S (Zeiss). A stable focus was ensured over several hours of imaging by using the Definite-Focus system (Zeiss). Captured images were processed using the Axiovision or ZEN software (Zeiss) and the software ImageJ (https://imagej.nih.gov/ij/).
For the detection of RovA, RovM and H-NS, bacterial whole cell extracts were prepared from equal amounts of bacteria and separated on SDS-polyacrylamide gels, and blotted onto nitrocellulose membranes. Subsequently, membranes were blocked in 1 x TBST containing 3% BSA (blocking buffer). Primary anti-RovA [4], anti-RovM [17] and anti-H-NS [15] antibodies were added in a 1:4.000 dilution in blocking buffer. The secondary antibody, anti-rabbit IgG conjugated with horseradish peroxidase, was supplied in a 1:8.000 dilution in blocking buffer and the immunological detection of the proteins was performed as described previously[15, 17].
His-tagged RovA was overexpressed with BL21λDE3 pLW2 and purified as described earlier [4].
EMSAs were performed as described [4]. The DNA fragments of the rovA regulatory regions including either RovA binding site I or II were amplified with the primer pairs 153/296 and 178/V99. For competitive EMSAs DNA fragments containing either RovA binding site I or II were mixed in equimolar amounts. Pre-incubation of recombinant RovA with the DNA fragments and native gel electrophoresis were performed at 25°C and 37°C, respectively.
To determine the amount of RovA molecules per cell Y. pseudotuberculosis YPIII harbouring a RovA-dependent egfpLVA reporter (pKH70) was grown over night at 25°C in liquid LB broth. A fresh culture was started by inoculating pre-warmed medium with over night culture in a 1:50 dilution and incubated at identical temperatures until the cultures reached an OD600nm = 0.6. Subsequently, 1 ml of culture was harvested by centrifugation for western blotting. The bacterial pellet was resuspended in 60 μl 1 x SDS loading dye, heated to 95°C for 10 min, cooled on ice and centrifuged for 5 min at 10.000 g. 10 μl of supernatant (bacterial cell extract from approx. 108 bacteria) were loaded onto 15% polyacrylamide SDS gels. In parallel 1 and 3 ng of recombinant RovA were loaded. Western blotting was performed as described.
For survival and organ burden experiments, 6–7 week old female Balb/c mice were purchased from Janvier (Saint Berthevin Cedex, France) and housed under specific pathogen-free conditions in the animal facility of the Helmholtz Centre for Infection Research, Braunschweig. After 16 hours of starvation, mice were orally infected with approximately 2 x 108 colony forming units (cfu) of Y. pseudotuberculosis YPIII or the different isogenic rovA mutant strains using a gavage needle. Bacteria were grown over night in LB medium at 25°C, washed and resuspended in PBS. For survival experiments infected mice were monitored for 14 days on a daily basis to determine the survival rate, the body weight and health status. For organ burden experiments, mice were euthanized by CO2 asphyxiation three days after infection. Peyer’s patches, caecum, MLNs, liver and spleen were isolated. Subsequently, all organs were weighed and homogenized in PBS at 30.000 rpm for 30 sec using a Polytron PT 2100 homogenizer (Kinematica, Switzerland). To determine the bacterial load of the organs serial dilutions of the homogenates were plated on LB plates with triclosan (Calbiochem). The cfu were counted and are given as cfu per g organ/tissue. To assure presence of the reporter plasmids during infection serial dilutions of Peyer’s patches and caecum of four infected mice were plated in parallel on LB plates containing either triclosan (total bacteria) or a combination of triclosan, chloramphenicol and carbenicillin. The cfu were counted and are given as percentage of cfu, normalized to the amount of total bacteria.
Animal housing and all animal experiments were performed in strict accordance with the German Recommendations of the Society of Laboratory Animal Science (GV-SOLAS) and the European Health Recommendations of the Federation of Laboratory Animal Science Associations (FELASA). The animal care and use protocols adhered to the German Animal Welfare Act, Tierschutzgesetz (TierSchG) and were approved by the Niedersächsisches Landesamt für Verbraucherschutz und Lebensmittelsicherheit: animal licensing committee permission no. 33.9.42502-04-12/1010. Animals were handled with appropriate care and all efforts were made to minimize suffering.
Statistical tests were performed with Prism 5.0c (GraphPad Software). Mann-Whitney test was used to compare wild-type and the rovA mutants in the organ burden experiments. The survival was statistical analyzed by the log-rank (Mantel-Cox) test. The amount of green bacteria in microcolonies in the infected tissues was compared between wild-type and the rovA mutant using Students t-test. p values < 0.05 were considered significant.
|
10.1371/journal.ppat.1006031 | The Ebola Interferon Inhibiting Domains Attenuate and Dysregulate Cell-Mediated Immune Responses | Ebola virus (EBOV) infections are characterized by deficient T-lymphocyte responses, T-lymphocyte apoptosis and lymphopenia. We previously showed that disabling of interferon-inhibiting domains (IIDs) in the VP24 and VP35 proteins effectively unblocks maturation of dendritic cells (DCs) and increases the secretion of cytokines and chemokines. Here, we investigated the role of IIDs in adaptive and innate cell-mediated responses using recombinant viruses carrying point mutations, which disabled IIDs in VP24 (EBOV/VP24m), VP35 (EBOV/VP35m) or both (EBOV/VP35m/VP24m). Peripheral blood mononuclear cells (PBMCs) from cytomegalovirus (CMV)-seropositive donors were inoculated with the panel of viruses and stimulated with CMV pp65 peptides. Disabling of the VP35 IID resulted in increased proliferation and higher percentages of CD4+ T cells secreting IFNγ and/or TNFα. To address the role of aberrant DC maturation in the IID-mediated suppression of T cell responses, CMV-stimulated DCs were infected with the panel of viruses and co-cultured with autologous T-lymphocytes. Infection with EBOV/VP35m infection resulted in a significant increase, as compared to wt EBOV, in proliferating CD4+ cells secreting IFNγ, TNFα and IL-2. Experiments with expanded CMV-specific T cells demonstrated their increased activation following co-cultivation with CMV-pulsed DCs pre-infected with EBOV/VP24m, EBOV/VP35m and EBOV/VP35m/VP24m, as compared to wt EBOV. Both IIDs were found to block phosphorylation of TCR complex-associated adaptors and downstream signaling molecules. Next, we examined the effects of IIDs on the function of B cells in infected PBMC. Infection with EBOV/VP35m and EBOV/VP35m/VP24m resulted in significant increases in the percentages of phenotypically distinct B-cell subsets and plasma cells, as compared to wt EBOV, suggesting inhibition of B cell function and differentiation by VP35 IID. Finally, infection with EBOV/VP35m increased activation of NK cells, as compared to wt EBOV. These results demonstrate a global suppression of cell-mediated responses by EBOV IIDs and identify the role of DCs in suppression of T-cell responses.
| The extensive investigation of interferon antagonism mediated by Ebola virus (EBOV) over the last 16 years resulted in identification of two interferon inhibiting domains (IIDs) located in the VP24 and VP35 proteins of the virus and of multiple mechanisms by which the domains disable the innate immune system and promote replication of the virus. However, the effects of these domains on cell-mediated immune response had not been investigated. To determine the effects of IIDs on cell-mediated responses, we used a panel of recombinant strains of EBOVs with point mutations disabling the VP24 and/or VP35 IIDs. The viruses were used for infection of peripheral blood mononuclear cells (PBMCs) or dendritic cells (DCs), which were subsequently co-cultured with T cells. We found that IIDs block activation and proliferation of T cells as a result of their functional role in suppressing maturation of DCs and limiting the formation of immunological synapses. Similarly, IIDs were demonstrated to suppress activation and differentiation of B cells, and skew activation of NK cells present in infected PBMCs. These data provide evidence of previously unknown effects of IIDs on the adaptive and innate cell-mediated immune responses and identify a novel mechanism of “immune paralysis” during EBOV infections.
| The 2013–2016 outbreak of Ebola virus (EBOV) in West Africa claimed the lives of 11,300 people [1]. EBOV infections are characterized by ‘immune paralysis’, the profound immune deficiency resulting in uncontrolled viral replication [2]. A characteristic feature of EBOV infections is lymphopenia, which is observed in both humans and experimentally infected nonhuman primates (NHPs) [3–10] and is particularly pronounced during fatal human cases [9–11]. Fatal human cases and studies with EBOV-infected NHPs also demonstrated apoptosis of T cells accompanied by upregulation of tumor necrosis factor related apoptosis inducing ligand (TRAIL) and Fas/FasL [11, 12]. Moreover, EBOV infection of macaques resulted in depletion of T-cells, NK-cells but not CD20+ B cells, and no detectable activation of T-cell [4]. The lack of T cell activation in infected macaques contrasts a recent study of EBOV survivors, which received EBOV-specific antibody treatment and demonstrated a substantial immune activation of T and B cells [13]. Thus, the available information on the effect of EBOV on cell-mediated response is incomplete and controversial.
Type I interferons (IFN-I) are well-characterized inflammatory mediators whose interaction with IFNα/β receptors (IFNAR) is critical for controlling viral infections [reviewed in reference[14]. IFNAR induces the Janus activated kinase-signal transducer that results in activation of transcription JAK-STAT pathway in the majority of cells, along with other pathways, some of which are cell type-specific, which jointly transcriptionally control expression of hundreds of IFN-stimulated genes (ISG) [15]. IFN-I directly regulates activation of numerous immune cell types including dendritic cells (DCs), T-lymphocytes, B-lymphocytes and NK cells [16–22]. IFN-I has been shown to affect monocyte and macrophage functions and differentiation [14, 23, 24]. Furthermore, IFN-I stimulates antibody-dependent cytotoxicity of macrophages while exerting both positive and negative regulation of secreted inflammatory mediators [14, 25]. IFN-I triggers macrophages to upregulate nitric oxide synthase 2, resulting in enhanced IFNγ-induced oxidative response and eventually enhanced phagocytosis [17, 26, 27]. With regards to DCs, IFN-I has multiple effects, including differentiation from monocytes, maturation and migration [23, 28–32] and enhancing their antigen presentation capacity [14, 16, 23, 24, 28–30, 33]. The presence of IFN-I during antigen-dependent maturation of DCs has been shown to strongly enhance their capacity to induce human antibody responses and CTL expansion [18, 19, 34, 35]. IFNα/β stimulation of immature DCs leads to a rapid upregulation of cell surface markers associated with the initiation of an adaptive immune response including MHC class I, MHC class II, CD40, CD80, CD86 and CD83 [16, 24, 33]. INFα has been shown to promote expression of chemokine receptors such CC chemokine receptor type 7 (CCR7) while both IFNα and IFNβ are required for the migration of plasmacytoid DCs to the marginal zones occupied by T-lymphocytes in vivo [36].
Signaling through IFNAR in T-lymphocytes is critical to the development of their effector functions. As noted above, IFN-I increases presentation of MHC-associated antigenic peptides on the surface of antigen presenting cells (APCs) that in turn results in antigen-specific activation of T-lymphocytes. IFN-I can exert it effects on immune cells either directly, through IFNAR signaling, or indirectly by the induction of chemokines, which promote recruitment of immune cells to the site of infection and result in the release of a second wave of immune modulatory cytokines [36]. Studies in mice have shown that IFNα promotes efficient cross-priming of antigen-specific CD8+ T-cells and secretion of IFNγ [18, 37]. IFNα has been shown to be a critical regulator of genes involved in the CTL responses [38, 39]. IFN-I stimulation of naïve CD4+ T-cells results in their differentiation into IFNγ-producing Th1 cells [19]. IFN-I directly enhances the functional role of CD4+ T-cells in the development of antibody responses [20]. Furthermore, IFN-I have been shown to directly inhibit regulatory T-cell function thereby promoting optimal antiviral T-cell responses during acute infection [21]. Conversely, IFN-I, along with IFN-III, may directly suppress proliferation of CD4+ T cells in context of viral infection [40]. Hence, IFN-I exhibits complex, pleotropic effects that are T lymphocyte subset specific.
IFN-I also has a profound effect on immune cells other than DCs and T cells. In addition to the ability to enhance antibody responses through the effects on DC and CD4+ T cells mentioned above, IFN-I also directly stimulates the ability of B cells to secrete antibodies [20]. IFN-I is critically associated with the production of all subtypes of immunoglobulin G (IgG) and the development of long-lived plasma cells and immunological memory [20, 41–43]. IFN-I was shown to enhance secretion of IFNγ and numerous other cytokines by NK cells through an autocrine IFNγ-dependent activity and enhance cytolytic activity [22, 44]. IFN-I also promotes both expansion and survival of proliferating NK cells via IFN-I/STAT1-dependent production of IL-15 [14, 45, 46]. Due to these pleiotropic effects, viruses have evolved targeted subversion strategies aimed at blocking IFN-I signal transduction by targeting the JAK-STAT pathway.
One of the characteristic features of EBOV infection is the strong antagonism of IFN-I responses by IFN-inhibiting domains (IID) located in the viral proteins VP24 and VP35, including the suppression of cytosolic sensing of double stranded RNA by VP35 IID and the subversion of IFN-induced signaling by both VP35 and VP24 IIDs [reviewed in reference[47]. Another important feature of EBOV infections is the lack of maturation of DCs despite their susceptibility to the virus [48, 49]. We recently demonstrated that a point mutation disabling the VP35 IID effectively unblocks maturation of DCs exposed to the virus, while a mutation disabling VP24 IID promotes partial maturation [50]. These mutations result in a global modulation of infected DCs transcriptome profiles with only partial overlapping being observed; hence, VP24 and VP35 IIDs are associated with distinct antagonistic mechanisms [51].
In this study we attempted to determine the effects of EBOV VP24 and VP35 IIDs on global adaptive and innate cell-mediated responses. We used recombinant viruses carrying point mutations disabling IID in VP24 (EBOV/VP24m), VP35 (EBOV/VP35m) or both (EBOV/VP35m/VP24m) previously generated in our lab [50, 51] each expressing green fluorescent protein (GFP) to visualize the infection of susceptible cells. For comparisons, GFP-expressing virus with no mutations in IIDs [52] which otherwise is identical to the mutated viruses, referred here as wild type (wt) EBOV, was used. Many important findings on EBOV pathogenesis utilize a mouse model; however, IFN-I has different effects on human and mouse lymphocytes [53, 54]. We therefore used only primary human immune cells in these studies. We found that IIDs result in a global attenuation and dysregulation of cell-mediated responses, including T, B and NK cells.
To determine the extent of EBOV-mediated suppression of cell-mediated immune response in our experimental systems, we assessed the ability of mock and EBOV-infected DCs to stimulate CMV-specific T-lymphocytes isolated from healthy individual donors seropositive for cytomegalovirus (CMV). CD14+ monocytes were isolated from peripheral blood mononuclear cells (PBMCs) using magnetic bead-based methods and differentiated into DCs for 7 days. Differentiated immature DCs were infected with wt EBOV expressing enhanced green fluorescent protein (GFP) from an added gene referred here as wild type (wt) EBOV [52] and simultaneously pulsed with pooled peptides overlapping CMV pp65 immunodominant protein followed by incubation for an additional 18–24 hours. In parallel, autologous PBMCs were stimulated with CMV peptides for 48 hours, at which point CMV-specific CD137+ T-lymphocytes were isolated and expanded for 8 days. Expanded responder cells were added at day 8 at a 1:1 ratio with mock or EBOV-infected DCs. Following overnight stimulation, cells were stained as described in the Materials and Methods and analyzed for markers of activation. Non-peptide stimulated DCs were used as a control to determine specific induction of CMV responders over background. A significant increase in percentages of T-lymphocytes positive for Ki67, a nuclear marker of proliferation [55], was detected as compared to no peptide control (Fig 1A). When T-lymphocyte responders were cultured in the presence of EBOV-infected DCs, percentages of Ki67+ cells were significantly reduced as compared to mock. Intracellular cytokine staining paralleled these findings as infection of DCs with EBOV significantly reduced percentages of T cells positive for activation markers IFNγ, IL2 or TNFα, as well as for the marker of degranulation CD107α+ [56] (Fig 1A and 1B). Despite a reduction in the proportion of activated T cells in response to EBOV infection of DCs, we also observed a moderate increase of the proportion of activated T cells over no peptide controls. This is likely due to the presence of DCs with the lack of high level EBOV replication, which are expected to display MHC-CMV peptides, as infection of DCs with EBOV results in different levels of viral replication in individual cells present in the population [50, 51]. Overall, these findings are highly correlative with previous in vivo data on the deficient T cell response to EBOV infection.
To determine the effects of IID on T cell response, we utilized a panel of recombinant EBOV mutants, each expressing GFP from an added gene to visualize the infection, which included EBOV/VP24m with the VP24 IID disabled by the mutation K124A [50], EBOV/VP35m with the VP35 IID disabled by the mutation R312A, and EBOV/VP35m/VP24m with both mutations [51] (Fig 2A). PBMCs from CMV-seropositive donors were infected with the panel of recombinant EBOVs and simultaneously stimulated with CMV pp65 peptides for 7 days, and then re-stimulated with the peptides for 6 hours (Fig 2B). Multi-color flow cytometry analysis of CD4+ T cells secreting IFNγ, IL-2 or TNFα as markers of activation demonstrated that disabling of the VP35 IID, but not the VP24 IID, resulted in significantly increased percentages of IFNγ+ and TNFα+ cells (Figs 2C and S1, S1 Table). The observed effects were more prominent in the dividing 5, 6-carboxyfluorescein diacetate succinimidyl ester (CFSE)-negative cell population. Since the functionality of T cells generally correlates with the number of cytokines and other markers of activation simultaneously expressed by individual cells [57, 58], we next quantified CD4+ T cells positive or negative for all 32 possible combinations of IFNγ, IL-2, IL-4, IL-17a or TNFα by Boolean gating. We found two abundant populations: IFNγ+ and IFNγ+TNFα+ (S2 Fig). Comparison of these populations and IFNγ+TNFα+IL-2+, expected to be the most highly activated, demonstrated that disabling of VP35 IID results in the increase of IFNγ+TNFα+ cell populations (Fig 2D, S2 Table). We also detected the increase in the percentages of single-positive IFNγ+ cells in the majority of donors, but the effect did not reach statistical significance. Disabling of any of the two IIDs did not significantly affect the percentage of GFP+ cells, although some reduction was observed when both IIDs were disabled (S3A Fig). These data suggest that the effects of mutations on activation of T cells are not related to changes in viral replication. We next determined the levels of IL-4, IL-5 and IL-13, which are markers of Th2 response, in supernatants. Wt EBOV strongly induced the expression of IL-5 and IL-13 (but not IL-4), while disabling of the VP35 IID resulted in their significant reduction (Fig 2E). These findings suggest that VP35 IID strongly reduces Th1 response.
Since T cells are resistant to EBOV infection [3], mechanisms associated with the VP35 IID-induced suppression of the Th1 response may be related to infection of multiple non-lymphoid immune cells susceptible to EBOV. We hypothesized that the effect results from the deficient maturation of DCs associated with VP35 IID [50]. To test the hypothesis, we established a co-cultivation system, which included monocyte-derived DCs from CMV-positive donors, which were infected with the panel of viruses in the presence of CMV pp65 peptides and subsequently cultured with autologous purified CD4+ T cells (Fig 3A). Following 7 days of co-cultivation, cells were re-stimulated with CMV peptides for 6 hours, and CD4+ T cells were analyzed by multi-color flow cytometry. We found that disabling the VP35 IID strongly increased the proportion of total CD4+ T cells populations secreting IFNγ, TNFα or IL-2 (Figs 3B and S4, S3 Table). The effect was more pronounced in the actively dividing CFSE- population. Once again, disabling of any of the two IIDs did not significantly affect the percentage of GFP+ cells, and a small reduction was observed with both IIDs disabled (S3B Fig), suggesting the effects of the mutations are not due to changes in the viral replication. To check if the time of peptide stimulation affects the observed effects, we compared simultaneous infection and stimulation used in our experiments, with stimulation 24 hours after infection. We found that the time of stimulation did not affect the percentage of GFP+ DCs (S5A Fig) or the percentage of activated IFNγ+ CD4+ T cells (S5B Fig).
We next quantified CD4+ T cells positive or negative for IFNγ, IL-2, IL-4, IL-17a and TNFα by Boolean gating, resulting in 32 possible combinations. The vast majority of cells exposed to the panel of viruses or SEB belonged to the IFNγ+, TNFα+ and IL-4+ single-positive populations, and IFNγ+TNFα+ double-positive populations (Figs 3C, S6 and S7, S4 Table). Disabling of VP35 IID resulted in a significantly increased percentages of IFNγ+, IFNγ+TNFα+ and IFNγ+TNFα+IL-2+ cells. The increases were observed both in the dividing (CFSE-) and the total cell populations. To determine the effects of IID under conditions when the majority of CD4+ T cells respond to stimulation, CMV-specific CD4+ T cells from three donors were expanded and used in co-culture assays with autologous DCs (Fig 3D). Under these conditions, not only the disabling of VP35 IID, but also VP24 IID resulted in significant increases of cytokine secreting T cells compared to wt EBOV (Fig 3E). We next quantitated cytokines and chemokines in the medium of expanded CD4+ T cell responders cultured for 24 hours with CMV-pulsed DCs pre-infected with the panel of viruses (Fig 4, S5 Table). Infection of DCs with wt EBOV resulted in a significant reduction of the majority of cytokines and chemokines analyzed in comparison to mock-infected DCs, with a few exceptions, including IL-4, IL-5 and IL-13 (Fig 4A). These data further confirm suppression of T cell responses in general, and Th1 responses, by EBOV demonstrated in Figs 1 and 2. Disabling of either VP24 IID or VP35 IID resulted in an increase, as compared to mock-treated cells, of the majority of the cytokines analyzed in donors 2 and 3, and a lesser number of cytokines in donor 1. The effects of the two IIDs were not identical; disabling of the both of them increased almost all cytokines analyzed in donors 2 and 3, and approximately half in donor 1. Remarkably, the levels of IL-4, IL-5 and IL-13 did not increase compared to wt EBOV-infected DC. The effects of the mutations on chemokine expression was even stronger as the majority of those analyzed were upregulated in response to either or both mutations (Fig 4B).
We next determined if the induction of Th1 response associated with disabling of VP35 IID is related to soluble factors, by assessing the ability of conditioned media to activate naïve CD4+ T-cells. We therefore infected DCs with the panel of viruses or mock-infected and pulsed with CMV-peptides, incubated for 5 days, collected cell-free supernatants and transferred them to naïve CD4+ T cells. Following overnight stimulation, brefeldin A and monenesin were added to media for an additional 6 hours that was followed by intracellular staining of IFNγ. Culture of CD4+ T-cells in conditioned media from DCs infected with EBOV/VP35m or EBOV/VP35m/VP24m resulted in significantly increased percentages of IFNγ+ CD4+ T cells (Fig 4C and 4D). To further demonstrate the direct role of VP35 IID in the restriction of a Th1-response, lentiviral vectors expressing wild type or mutant VP35 were used to transduce DCs. Co-culture of CD4+ T-cells in the presence of DCs transduced with the mutant VP35 resulted in a significant increase in the percentages of IFNγ+ CD4+ T-cells compared to wild type VP35 (S8 Fig). Similar results were obtained both in the presence or absence of CMV peptides. These data suggest that the restriction of Th1 response by VP35 IID at least in part is mediated by soluble factors, and the observed effects are not related to changes in biological properties of the virus due to the introduced mutation. Taken together, the results demonstrate that the VP35 IID, and in a lesser degree VP24 IID, suppress activation of CD4+ T cells as a result of the IID-associated deficient maturation of DC.
Previous reports have indicated that DCs require IFN-I response to mature [59]; however, induction of the IFN-I signaling pathway, but not release of IFN was reported to be a requirement for maturation of DCs following induction by negative-strand RNA viruses [60]. We therefore determined whether the observed suppression of DCs maturation by VP35 IID is a consequence of the suppression of IFN-I signaling and whether released IFN has any effect on this phenotype. First, we blocked IFN receptor 2 (IFNAR2), since the anti-viral activities of IFN-I correlate well with its binding to IFNAR2, rather than IFNAR-I [61]. For this purpose, IFNAR2 blocking antibodies were added to DCs at a concentration 30 μg/ml; we previously demonstrated that a 1,000-fold lesser dose of the antibody suppresses the expression of the IFN-inducible genes Mx1 and ISG56 in T cells [62]. Cells were incubated with IFNAR2 blocking antibodies at 37°C for 1 hour, followed by inoculation with the viruses as indicated previously. At 40 hours post infection, we examined the expression levels of markers associated with DC maturation including CD86, CD80 and CD54 (Figs 5A and S9A). Even though wt EBOV only induced low expression of CD86, IFNAR2 blockade further reduced it to the level of mock-infected DCs. No significant effects of the blockade were detected for CD80, and the effect on CD54 was similar to that of CD86, but somewhat less pronounced. The reduction of the level of CD86 in DCs infected with wt EBOV by the IFNAR2 blockade is likely a result of the synergistic effect of the blockade with the effects of the VP35 and VP24 IID present in the virus. In contrast, in DCs infected with EBOV/VP35m, the blockade did not result in significant changes in the expression of any of the three markers of maturation in both GFP+ and GFP- cells. Since EBOV/VP35m has the intact VP24 IID, which inhibits IFN-I signaling, and VP35 IID suppresses not only double stranded RNA cytosolic sensing but also inhibits other host defense pathways [reviewed in reference[47], these findings suggest that DC maturation associated with disabling of VP35 IID is only partially dependent on IFN-I. This hypothesis is supported by the release of numerous cytokines including TNFα following infection with EBOV/VP35m (Fig 4A), which could stimulate DC maturation by induction of pathways other than the IFN-I pathway. Furthermore, these findings are consistent with the increased Th1 response by conditioned media from DCs infected with EBOV/VP35m (Fig 4C and 4D) and with the lack of difference between maturation of infected (GFP+) and uninfected (GFP-) DCs exposed to EBOV/VP35m we reported earlier [50]. These results differ from the previous observation with an HIV gag vaccine in mice, when blocking of IFN-I receptors prevented effective DCs maturation [59].
In separate experiments, the effects of exogenously added IFNα2 or IFNβ were determined. Due to the anti-viral effects of IFNs, DCs were infected and incubated for 24 hours before IFNα2 or IFNβ was added at the concentrations 1,000 and 800 IU, respectively. This was followed by an additional 20 hour-long incubation, and analysis of the expression of the maturation markers CD86 (Fig 5B), CD80, and CD54 (S9B Fig). No significant effect of exogenous IFN on the expression of CD86 and CD80 was found. Contrary to the expectations that exogenous IFNs would facilitate DCs maturation, the expression of CD54 was reduced, rather than increased in both the infected and mock-infected cells. In addition, no significant differences in the expression of the three maturation markers were found between GFP+ and GFP- cells. These results support the previously published observation that secreted IFN-I is irrelevant for the induction of DC maturation by viruses [60]. Taken together, our results suggest that the suppression of IFN-I signaling and the prevention of IFN-I release by EBOV per se play only a limited role in the suppression of DCs maturation by the IID that is consistent with their diverse and highly redundant mechanisms of the suppression of the innate immune response by EBOV IIDs [47].
Stimulation of T cells by DCs is a key step, which affects the magnitude of the immune response to a viral infection [63]. To determine if VP35 IID interferes with the ability to form immunological synapses between DCs and T-lymphocytes, we co-cultured purified CD4+ T-cells with autologous CMV-pulsed DCs infected with wt EBOV or EBOV/VP35m. Twenty four hours following infection, autologous CD4+ T-cells were added at a 1:1 ratio, cultured for an additional 4 hours, fixed, and stained as described in the Materials and Methods. Immunological synapses between DCs and T cells were visualized by confocal microscopy, which demonstrated co-localization of HLA-DR and CD3ε in mock-infected cultures. Infection of DCs with wt EBOV resulted in a significantly reduced number of immunological synapses as compared to mock-infected DCs, while disabling of VP35 IID significantly increased the number in comparison to wt EBOV (Fig 6A and 6B). In addition, the intensity of the staining of the immunological synapses and their sizes were greater in DCs infected by EBOV/VP35m than in cells infected with wt EBOV. Furthermore, we noted that infection with wt EBOV reduced the HLA-DR expression compared to mock-infected cells, while no reduction was observed with EBOV/VP35m, the finding further confirmed by flow cytometry analysis (Fig 6C). Taken together these data suggest that EBOV VP35 IID interferes with IS formation by reducing DC maturation, thereby limiting the capacity of DCs to effectively generate an adaptive immune response.
We next examined the effects of VP24 and VP35 IID on proliferation of T cells. Total PBMC from four healthy CMV-seropositive donors were labeled with CFSE, inoculated with the panel of viruses with or without simultaneous pulsing with CMV peptides, and cultured for 7 days (Figs 7A and S10, S6 Table). Infection with wt EBOV did not affect the percentages of the dividing CFSElow cells without peptide stimulation, but reduced it, compared to uninfected cells, in presence of peptides. Infection with EBOV/VP35m resulted in an increase, as compared to wt EBOV, in proliferation of both CD4+ and CD8+ T cells in PBMCs from each donor, with or without peptides, although the effect was weak for CD4+ T-cells with peptides, and the overall effect did not reach statistical significance due to the high donor-to-donor variability. In contrast, no consistent effect was observed in PBMCs infected with EBOV/VP24m or EBOV/VP35m/VP24m.
Based on these data, we hypothesized that IID also suppresses expression of CD69, which is involved in lymphocyte proliferation as a signal transmitting receptor [64], and Ki67, a nuclear marker of proliferation. We utilized the enriched CMV responder assay described in Fig 3D, in which DCs and autologous isolated CMV-specific CD137+ CD4+ T-lymphocytes were cultured for 8 days, infected with the panel of the viruses, and pulsed with CMV peptides. Next, expanded CMV-specific T-lymphocyte responders were combined at a 1:1 ratio with the infected DCs and after 24 hours long co-cultivation, analyzed by flow cytometry. In T-cells cultured with wt EBOV infected DCs, the percentages of both CD69+ and Ki67+ CD4+ T cells were reduced as compared to T-cells cultured with uninfected DCs (Fig 7B and 7C) that suggests suppression of T cell activation/proliferation by EBOV and is consistent with the data shown in Figs 1 and 7A. Again, disabling of VP35 IID completely reversed the suppressive effect of wt EBOV. Unexpectedly disabling of VP24 IID, in addition to VP35 IID, or both, also completely reversed the suppressing effect of IIDs in this system. Taken together, these data demonstrate the suppressive effect of both VP35 and VP24 IID on T cell proliferation.
Previous studies have demonstrated that DCs and other APCs are vital to the survival of CD4+ T-cells due to MHC-TCR complex-dependent signal transduction that require a direct cellular contact [65–67]. We therefore sought to examine the effects of IID on signal transduction using co-cultures of unstimulated DCs and autologous CD4+ T-cells. Following a four-day co-culture with mock or EBOV-infected DCs, the phosphorylation cascades associated with TCR signaling were analyzed by Western blotting (Fig 8A). Specifically, we examined phosphorylation of TCR complex-associated adapters and downstream signal molecules, which have previously been shown to remain phosphorylated in the absence of antigen-dependent activation, and whose phosphorylation depends on DC-T-cell contact [65, 68–70]. In the absence of DCs, a limited phosphorylation of signal transduction mediators was detected in CD4+ T-cells. In contrast, CD4+ T-cells cultured in the presence of mock-infected DCs exhibited phosphorylation of molecules activated at both early and late stages of TCR-mediated signaling. Presumably, the observed phosphorylation status represents basal phosphorylation events that while supporting survival of CD4+ T-cells, remain below the threshold required for activation. Infection with wt EBOV resulted in the increase in phosphorylation of Lck; however, phosphorylation of additional adapters including ZAP70, PLCγ1 and SLP76 appeared to be blocked. Consistent with the data demonstrating reduced formation of immunological synapse (Fig 6), the absence of downstream phosphorylation events may be the result of impaired synaptic formation and/or aberrant signal transduction. Phosphorylation of ZAP70 is dependent on the formation of TCR microclusters that form scaffolding for downstream signaling events at the intracellular sites of immunological synapse formations [71]. Thus the absence of phosphorylated ZAP70 in co-culture of T-cells with wt EBOV-infected DCs strongly indicates at suboptimal engagement of TCR. In agreement with that, infection of DCs with wt EBOV significantly reduced expression of HLA-DR (Fig 6C). It is not immediately clear why a relatively strong induction of Lck phosporylation was observed following wt EBOV infection; however, it is plausible that secondary signal transduction may be suboptimal resulting in impaired downstream signaling.
Disabling of VP35 IID unblocked phosphorylation of ZAP70 and PLCγ1, and to a lesser degree, SLP76, while disabling of VP24 IID unblocked only ZAP70, but to a much greater degree than disabling of VP35 IID. Unexpectedly, infection with EBOV/VP24m resulted in an increase in the relative amount of phosphorylated ZAP70 despite the lack of phosphorylation of the downstream signaling molecules PLCγ1 and SLP76; however, this is consistent with the lack of effective T cell activation by this mutant. This discrepancy may result from differential kinetics associated with altered signaling events, the absence of co-activating signals or upregulation of negative regulators of PLCγ1 and SLP76 or other inhibitory factors downstream ZAP70. Surprisingly, disabling of both IIDs resulted in only weak phosphorylation of PLCγ1 and SLP76, but not ZAP70. We note however, the absence of CD3ζ and Lck phosphorylation in EBOV/VP35m/VP24m co-cultures, suggesting that the phosphorylation of PLCγ1 and SLP76 may be associated with alternate signaling pathways. Although not examined in these studies, it is possible that both phosphorylation and dephosphorylation kinetics may be variable following co-culture.
To further characterize the capacity of infected DCs to transmit survival signals, we determined the relative levels of the anti-apoptotic Bcl-2 family members. These proteins are involved in the survival of T cells and are upregulated in response to survival signals [72–74]; on the other hand, survival signal transduction was previously associated with reduced Bcl-2 levels [75–77]. We also determined phosphorylation of Src, which has previously been shown to be essential for naïve T-cell survival and also TCR-dependent [78] (Fig 8B). Bcl-XL was undetectable in control CD4+ T-cells alone, but was readily detectable at relatively similar levels in CD4+ T-cells cultured with mock, wt or mutant EBOV infected DCs. Bcl-2 was readily detectable in CD4+ T-cells cultured with mock or wt EBOV-infected DCs, but greatly reduced when VP35 and/or VP24 IID were mutated. Furthermore, the phosphorylated Src was detectable in T cells cultured with mock-infected DCs and DCs infected with the mutated but not wt EBOV, showing correlation with activation of infected DCs (Figs 2–7). Taken together, these data demonstrate that cultivation of T cells with EBOV-infected DCs blocks phosphorylation of ZAP70, PLCγ1 and SLP76 involved in TCR signaling and the pro-survival molecule P-Src, and that the levels of P-Src correlate with T cell activation, while that of Bcl-2 correlate with T cell survival in the autologous system. Furthermore, these data identify the role of VP35 IID in blocking phosphorylation of these molecules, and role of VP24 IID in blocking phosphorylation of ZAP70. Paradoxically, the two IIDs demonstrated an opposite (i.e. stimulating) effect on phosphorylation of Lck, as well as on expression of the prosurvival molecule Bcl-2. A schematic which illustrates the positive and negative signals associated with IIDs as they relate to the observed phosphorylation status of adapter molecules in Fig 8A and 8B is presented in Fig 8C.
To date, only limited data concerning the effects of EBOV infection on B and NK cell function have been published. To determine the effects of EBOV infection and IID on B-cells, PBMCs were infected with wt or mutated EBOVs at MOI 1.0 PFU/cell, incubated for 7 days, and analyzed for markers of activation, maturation and those associated with specific subsets. The percentages of naïve B-cells (CD19+CD27-IgD+) were reduced after infection with wt EBOV or the mutants except the double mutant (S11A Fig). The percentages of memory B-cells (CD19+CD27+IgD-) did not change after infection with wt EBOV, while disabling of VP35 (but not VP24) IID increased this population (Fig 9A). Of note, these results inversely correlated with observed changes in the percentage of naïve B-cell populations (S11A Fig). We next examined the effects of IID on the percentages of class-switched memory B-cells (CD19+CD27-IgD-IgM-CD20+CD38++) (Figs 9B, S11B and S11C). Infection with wt EBOV slightly reduced the percentages of this population, while disabling of VP35 IID significantly increased it. Examination of post-class switched memory B-cells (CD19+CD27+/DimIgD-IgM-CD20-CD38-) revealed a similar effect (Fig 9C); again, the decrease in SEB-treated cells is consistent with an increase in other subsets in samples treated with SEB (Fig 9B and 9D). These results suggest that VP35 IID may, by a yet to be determined mechanism, block the development of post-class switched memory cells. Analysis of plasma cells (CD19+CD38+CD138+) demonstrated the lack of effect of wt EBOV, but as much as 4-fold increase in their percentage after infection with EBOV/VP35m (Fig 9D). Interestingly, infection with the double mutant also increased the percentages of this cell population, but only ~1.5-fold, suggesting that VP24 IID reduces the effect of VP35 IID. These data demonstrate that VP35 IID suppresses induction of memory B cells, their class switching, and their differentiation into plasma cells.
The primary role of NK cells is the continuous surveillance of “stressed” cells, which is regulated by detection of both activating and inhibitory signals, as well as by cytokines. One of the major molecules, whose expression affects activation of NK cells, is MHCI; reduced MHCI expression levels in virus-infected cells are sensed by NK receptors, which typically result in activation of NK cells [79, 80]. In addition, IFN-I is required for the optimal NK cell response and promotes the activation and effector functions of these cells [81]. As noted above, our previous study demonstrated that disabling of IID effectively unblocks maturation of EBOV-infected DCs, as evidenced by increased expression of multiples maturation markers, although expression of MHCI was not tested, and increased expression of IFN-I [50]. We therefore examined the effects of IID on NK function in PBMCs by analyzing expression of both activation and inhibitory markers on CD56+CD3- NK cells. We tested expression of the two activation markers: CD38 (Fig 10A), which triggers cytolytic response [82] and NKp46 (Fig 10B), which is an immunoglobulin-like natural cytotoxicity receptor [83], and three inhibitory receptors: CD27 (Fig 10C), which is expressed by mature cytotoxic effector NK cells [84], CD158 killer immunoglobulin-like receptor (Fig 10D), which inhibits NK cytotoxicity [85] and the lectin-like receptor KLRG1 (Fig 10E) [86]. Paradoxically, infection with wt EBOV increased the percentages of NK cells positive for the activating marker NKp46 and all inhibitory markers tested, as well as the percentage of dead NK cells (Fig 10F). Disabling of VP24 IID resulted in a strong increase in the percentages of cells expressing NKp46, and a limited non-statistically significant increase of CD38, and reduced the percentages of dead cells. On the other hand, disruption of VP35 IID resulted in strongly (10.1-fold versus wt EBOV) increased expression of CD38 but not NKp46, which showed an opposite trend perhaps due to altered signaling events. Furthermore, it reduced expression of all three inhibitory receptors. These data suggest, based on multiple activating and inhibitory receptors, that both IIDs suppress activation of NK cells, with the exception of the effect of VP35 IID on the activating receptor NKp46.
Our findings provide new insights into functional role of IID in EBOV pathogenesis and identify their role as suppressors or modulators of both the innate and adaptive cell-mediated immune responses. We show that EBOV infection of PBMCs resulted in only a limited activation of T cells, while disabling of VP35 IID significantly increased their activation (Fig 2). As lymphocytes, along with NK cells, are refractory to EBOV infection, we hypothesized that the IID-associated effects are indirect and result from interaction of these cells with other cells susceptible to the virus, such as DCs. To identify the role of DCs, two co-culture systems were used, which included DCs infected with the panel of viruses and CMV-stimulated T cells (Fig 3A–3C), and its modified version with expanded CMV responder T-lymphocytes (Fig 3D and 3E). Analysis of supernatants of CMV-peptide stimulated DCs infected with wt EBOV co-cultured with expanded responder T-lymphocytes demonstrated that most of the cytokines and chemokines analyzed were below that observed in mock-infected control (Figs 2E and 4A). In addition, analysis of supernatants of wt EBOV-infected cultures where indicative of Th2 skewed response, as IL-5 and IL-13 were elevated. This is consistent with the induction of Th2 response in EBOV patients [87] and in macaques infected with Marburg virus [88], which similarly to EBOV belongs to the family Filoviridae and causes human infections with high case fatality rates. Supernatants of EBOV/VP24m-infected cells also demonstrated elevated levels of IL-13. However, flow cytometry analysis of T cell phenotypes of CD4+ T-cells from CMV-responder assays indicated that disabling of VP24 IID promotes expression of Th1-associated cytokines (Fig 3E) in addition to Th2, suggesting induction of a mixed Th1/Th2 response. Strikingly, flow cytometry analysis also demonstrated that disabling of VP35 IID results in the significant expansion of activated Th1 population suggesting that VP35 IID limits the capacity of EBOV-infected DCs to initiate the Th1 response. Consistent with that, analysis of EBOV/VP35 IID supernatants demonstrated several fold reduced levels of IL-5 and IL-13 indicating induction of a uniformly Th1-like response. Overall, the cytokine analysis provides a clear indication regarding the inhibitory effects of VP35 and VP24 IIDs, as their disruption by point mutations dramatically increased cytokine/chemokine secretion. A clear cumulative effect was observed in EBOV/VP35m/VP24m cultures as the levels of virtually all cytokines and chemokines were elevated in comparison to either single mutant. This finding suggests that activation of immune cells and their migration to sites of infection may be severely impaired in vivo due to the presence of IIDs. We note that in T cell activation analyses, that the effect of VP24 IID in the double mutant often countered the effects of the VP35 IID in a single mutant (Figs 2 and 3), which is consistent with our recent transcriptome analysis of DCs infected with the panel of viruses [51]. In general, the EBOV/VP24m mutant exhibited a phenotype consistent with more moderate effects in comparison with EBOV/VP35m. Interestingly, while the effects of VP35 IID on activation and proliferation of T cells in the DC-T cell co-cultivation system (Figs 2C, 2D, 3B, 3C and 7A) were strong, the effects of VP24 IID were minimal. We therefore expanded CMV-specific CD4+ T cells that resulted in not only better identification of the effects of VP35 IID, but also demonstrated the effects of VP24 IID (Figs 3E and 7B).
The limited formation of immunological synapses in co-culture experiments with wt infected DCs (Fig 6) demonstrated the mechanism by which VP35 effectively blocks the development of an adaptive immune response. As noted above, previous studies demonstrated that infection of DCs with wt EBOV results in their aberrant maturation. Hence, DC-associated ligands required for the formation of synapses are likely inadequately expressed resulting in fewer immunological synapse formations between DCs and T-cells. This reduction results in an inability to reach a signal transduction threshold necessary for cellular activation. This presumption is consistent with both the limited T cell activation observed in co-cultures experiments and the altered phosphorylation cascade profiles when T-cells were cultured with wt EBOV-infected DCs (Fig 8). The block in immunological synapses formation and the limited T cell activation were highly correlative with the presence of functional VP35 IID in wt EBOV, as disabling the VP35 IID reversed the suppressive effects.
The EBOV VP35 IID-mediated suppression of cytosolic sensing and induction of IFN-I response, which otherwise would induce an antiviral state are well established [reviewed in reference [47]. Thus both the lack of proper DC maturation and impairment in the development of an antiviral state blunt the initiation of a T cell response to EBOV infection. Interestingly, the suppression of IFN-I signaling and the prevention of IFN-I release per se during EBOV infection appeared to play only a limited role in the suppression of DCs maturation by the IID (Fig 5). On the other hand, transfer of conditioned media from DCs infected with EBOV/VP35m, but not wt EBOV, to DC-T cell co-cultures effectively stimulated secretion of IFNγ by T cells (Fig 4C and 4D), despite the block of IFN signaling by intact VP24 IID present in the virus. These data suggest that that some cytokines, such as TNFα whose expression was unblocked by disabling VP35 IID (Fig 4A) could contribute the observed DC maturation. More studies are required to mechanistically connect the IFN-inhibiting effects of EBOV IIDs with their effects on maturation of DCs and ultimately phosphorylation of TCR-associated adaptors and downstream signaling molecules.
These studies also demonstrate aberrant B cell and NK cell activation by EBOV, thus suggesting the global impairment of the adaptive and innate cell-mediated immune responses by IIDs. As indicated above, B-cell function and maturation are highly affected by IFN-I stimulation in the presence of cognitive antigen (35, 40–42). Previous findings have indicated that survivors of EBOV infection develop and maintain antigen specific adaptive immune responses, which included both CTL and humoral responses but did not include direct examination of antigen-specific B-cell function [89, 90]. We demonstrated an overall increase in the percentages of class switched and post class-switched memory B-cells in response to disruption of VP35 IID (Fig 9). These data suggest a suppressive effect of VP35 IID on class switching and identify the role of IFN-I in B-cell differentiation. Disruption of VP35 IID also led to an increase in the overall percentage of B-cells positive for markers associated with plasma cells. Furthermore, consistently with in vivo data demonstrating loss of NK cells in EBOV-infected macaques [4], infection of PBMCs with wt EBOV appeared to increase the rate of NK cell death. Similarly to our findings regarding increased functional activity, proliferation and differentiation of lymphocytes associated with disruption of VP35 IID, the mutation altered the expression of several NK cell markers in a direction generally corresponding to a greater cytotoxicity, and also reduced NK cell death (Fig 10). Intriguingly, disruption of VP24 IID dramatically increased the percentage of NK cells expressing the natural cytotoxicity receptor NKp46.
We note that this study was entirely performed with primary human immune cells from donors, which allowed us to detect the magnitude of the effects on cells with different genetic background, as can be seen by a relatively high donor-to-donor variability, and avoid producing skewed results related to a specific genetic background of inbred mice, such as the Th2-skewed response in BALB/c mice [91]. Despite that, the inhibition of activation of T cells co-cultured with EBOV-infected DCs contrasts the T cell activation in EBOV patients [13]. The most obvious difference is that most of CD8+ T cells in patients were positive for Ki-67, while only 12% of CD4+ T cells in our study were Ki67+ (Fig 7C). This discrepancy can be explained by a much greater increase in the numbers of Ki67+ CD8+ T cells as compared to Ki-67+ CD4+ T cells, as demonstrated with human immunodeficiency virus infection [92]. In addition, activation, including non-antigen specific activation, of T cells in infected patients by other types of cells not included in our studies, or by stimulation related to high doses of EBOV-specific antibodies administered to patients may contribute to the observed in vivo activation through an unknown mechanism.
These studies provide evidence of the dual role of the VP35 and VP24 IIDs in the pathogenesis of EBOV. While IIDs are intimately linked to the ability of EBOV to block IFN-I production and signaling, they also counter the ability of DCs to initiate the adaptive immune response. The lack of DC maturation following EBOV-infection presents a significant obstacle in development of the adaptive immune response but also may render immune cell populations susceptible to premature cell death due to aberrant signal transduction and/or the absence of survival signals. Importantly, the suppressive effect of IIDs is not limited to T and B cells, which are the central components of the adaptive response, but also extends to NK cells, the key players in cell-mediated innate immune response. Taken together, these findings suggest global suppressive effects of EBOV IIDs on cell mediated response, and also indicate the potential benefits of blocking the immunosuppressive effects of IIDs as a potential therapy for EBOV-infection.
All work with EBOV was performed in BSL-4 facilities of the Galveston National Laboratory. Flow cytometry was performed either in BSL-4 using the Canto-II instrument (BD Biosciences), or cells were treated with 4% paraformaldehyde in PBS for 48 hours according the UTMB standard operating procedure and removed from BSL-4 for analysis with LSRII Fortessa flow cytometer (BD Biosciences) available at the UTMB Flow Cytometry Core Facility. Cells for confocal microscopy were placed on slides, stained, fixed in 4% paraformaldehyde for 24 hours, which was replaced with a fresh solution, incubated for additional 48 hours, and taken out of BSL-4. To remove supernatants of EBOV-infected cells from BSL-4, they were gamma-irradiated with the 5 Mrad dose according the UTMB standard operating procedure protocol. Staining and mounting procedures are described below. The staff had the U.S. government permissions and appropriate training for work with EBOV.
Generation of the recombinant EBOVs carrying the mutation R312A in the VP35 IID or K142A in the VP24 IID, or both, each expressing GFP from an added gene, and the control GFP-expressing EBOV with no mutations was described previously [50, 51]. The viruses were propagated on Vero-E6 monolayers and quantitated by plaque titration as previously described [93].
Peripheral blood nuclear cells (PBMC) were obtained from buffy coats from anonymous healthy adult blood donors from the UTMB blood bank according to a clinical protocol approved by the UTMB Institutional Review Board. Study population included both CMV-positive and CMV-negative individuals as tested using the Beckman Coulter PK CMV-PA System for qualitative detection of IgG and IgM antibodies to CMV.
A pool of 138 CMV peptides (15-mers overlapping by 11 amino acid residues) was obtained from the NIH AIDS Research and Reference Reagent Program. Peptides were reconstituted at 1 mg/ml in DMSO and stored at -70°C in 200 μl aliquots and used to stimulate PBMC or dendritic cells at final concentration of 2 μg/ml.
Total PBMC from CMV+ donors were resuspended at 1x106 per ml using 50 ml conical tubes and inoculated at an MOI of 2 PFU/cell with the recombinant strains of EBOV depicted in Fig 2A and simultaneously stimulated with 2 μg/ml of CMV pp65 peptides in media containing 10% human serum (Corning, Celgro) at 37°C, 5% CO2. Staphylococcal Enterotoxin B (SEB) (Sigma Aldrich) was used as a positive control at a final concentration of 2 μg/ml. Additional controls included mock-infected cells and cells stimulated with 15-mer CMV pp65 peptides only. After 4 hours of incubation, cells were washed twice by centrifugation at 200 x g for 5 min with 2% human serum media and cultured at 1x106 per ml in Advanced RPMI 1640 medium (Gibco, Life technologies) supplemented with 10% human serum (Gemini Bio-Products), 2 mM L-glutamine, 200 IU/ml penicillin, and 200 μg/ml streptomycin sulfate (Invitrogen) in 6-well plates.
PBMC were isolated by density gradient centrifugation (histopaque; Sigma life science). CD14+ monocytes were purified by positive selection using anti-CD14 monoclonal antibody-coated magnetic microbeads according to the manufacturer’s instructions (Quadro Macs; Miltenyi Biotech). CD14+ monocytes were cultured in T-225 flasks (Corning Incorporated, Costar) at 6 x 105 cells per ml in Advanced RPMI 1640 medium (Gibco, Life Technologies) supplemented with 10% heat-inactivated bovine serum (Quality Biologicals), 2 mM l-glutamine (Invitrogen), 0.05 mM β-mercaptoethanol, 50 ng/ml granulocyte-macrophage colony-stimulating factor (R&D Systems), 16 ng/ml interleukin-4 (R&D Systems), 200 IU/ml penicillin, and 200 μg/ml streptomycin sulfate (Invitrogen). The cells were incubated for 7 days at 37°C, 5% CO2 as described previously [50].
Immature DCs generated as described above were harvested, infected with the panel of EBOVs and stimulated with CMV pp65 peptides either simultaneously or 24 hours following infection. After 4 hours of peptide stimulation, cells were washed twice and co-cultured with CFSE (Molecular Probes, Life Technologies)-labeled autologous PBMC or purified CD4+ T cells at 37°C, 5% CO2 for 7 days. The DC: lymphocyte ratio was 1:10 (2 x 105 DC: 2 x 106 lymphocytes) in 2 ml of Advanced RPMI 1640 medium. For assays involving purified CD4 or CD8 T cells, these cell populations were purified by negative selection using a primary cocktail of antibodies conjugated to biotin and secondary anti-biotin antibody conjugated to magnetic microbeabs in order to deplete non-CD4+ or non-CD8+ T cells including γ/δ T cells, B cells, NK cells, DC, monocytes, granulocytes and erythroid cells.
Lentiviral vectors encoding wt and R312A mutant VP35 were prepared as previously described [94]. Briefly, plasmids encoding the lentivirus, HIV-Gag-pol, VSV-G and SIV-Vpx were transfected into 293T using PEI (Sigma Aldrich) transfection reagent. Cell monolayers were incubated overnight, fresh medium was added, and monolayers were incubated for additional 48 hours. Thereafter cell-free supernatants were collected and lentiviral vectors were titrated in 293T cell monolayers. DCs were transduced twice with ~8 hour culture period between the addition of lentiviral stocks. Following three days of culture, mock or medium containing CMV peptides were added. Autologous CD4+ T-cells were added to transduced T-cells at a 1:1 ratio, and the percentages of IFNγ+ CD4+ T-cells were determined 24 hours after initiation of co-culture.
After 7 days of culture, cells were harvested, washed and 2 x 106 cells were stimulated for 6 hours in culture medium with 10 μg/ml Brefeldin A (Sigma-Aldrich), 0.7 μg/ml monensin (GolgiStop, BD Biosciences), 1 μg/ml anti-CD28 (BD Biosciences), 1 μg/ml anti-CD49d (BD Biosciences), 20 μg/ml DNase (Calbiochem) and 2 μg/ml of 15-mer CMV pp65 peptides. PMA (Sigma Aldrich) at 20 μg/ml and ionomycin (EMD Chemicals) at 1 μg/ml were also used as an additional positive control. Following stimulation, cells were washed 2 x with wash buffer (PBS, 1% FBS, 0.02% sodium azide) followed by PBS. CD4+ T cells were stained extracellularly with anti-CD3 antibodies labeled with anti-CD3 BD Horizon Brilliant Ultraviolet 395 (clone UCHT1, BD Biosciences) and anti-CD4 PE-CF594 (clone RPA-T4, BD Biosciences); for analysis of activation, cells were also stained with CD69-PE/Dazzle (clone FN50, BioLegend) or anti-Ki67-brilliant violet 421 (Clone B56, BD Biosciences). CD8+ T cells were stained extracellularly with anti-CD3 BD Horizon Brilliant Ultraviolet 395 (clone UCHT1, BD Biosciences) and anti-CD8 BD Horizon PE-CF594 (clone RPA-T, BD Biosciences). Both CD4+ and CD8+ T cell subsets were also stained extracellularly with Live/Dead Fixable Aqua or Near-Infra Red (Invitrogen) to discriminate dead cells by flow cytometry for 30 minutes at 4°C. Following extracellular staining, cells were washed, fixed and permeabilized with CytofixCytoperm (BD Biosciences) according to manufacturer’s instructions. CD4+ and CD8+ T cells were stained intracellularly with the following antibodies: anti-IFNγ-PE (clone B27), anti-IL2-allophycocyanin (APC) (clone MQ1-17H12), anti-IL-4-peridinin chlorophyll protein PerCP)/Cy5.5 (clone 8D4-8), anti-IL17a-BD Horizon V450 (clone N49-653) and -TNFα-Alexa Fluor 700 (clone Mab11); all from BD Biosciences. Flow cytometry analysis of DCs for HLA-DR was performed with anti-HLA-DR-PE/Dazzle 594 (clone L243, BioLegend). B-cell subset staining was performed as follows: anti-CD19-PerCP/Cy5.5 (clone HIB19, BioLegend), anti-IgD-PE/CF594 (clone IA6-2, BD Biosciences), anti-IgM-brilliant violet 786 (clone R6-60.2, BD Biosciences), anti-CD27-brilliant violet 510 (clone M-T271, BioLegend), anti-CD24-PE (clone ML5, BioLegend), anti-CD20-Alexa Fluor 700 (clone 2H7, BioLegend), anti-CD38-APC (clone HIT2, BioLegend), anti-IgG-Brilliant Violet 421 (clone G18-145, BD Bioscience), anti-CD138-fluorescein isothiocyanate (FITC) (clone DL-101, BioLegend). To analyze CD56+CD3- NK cell function/activation, the following antibodies were used: anti-CD38-Alexa Fluor 488 (clone HIT2, BioLegend), anti-NKp46-Brilliant Violet 786 (clone 9E2, BD Bioscience), anti-CD27-Brilliant Violet 510 (clone M-T271, BioLegend), anti-CD158B-APC (clone DX27, BioLegend), anti-KLRG1-PE-CF594 (clone 14C2A07, BioLegend), anti-CD56-Brilliant Violet 421 (clone HCD56, BioLegend) and anti-CD3-Brilliant Ultraviolet 395 (clone UCHT1, BD Biosciences). Cells were subsequently washed twice with Perm/Wash and one time with PBS, fixed with 4% paraformaldehyde (Polysciences) and taken out of BSL-4 according to an approved protocol. Cells were washed, resuspended in PBS and 200,000 to 500,000 events were acquired on the BD LSR II flow cytometer (BD Biosciences). For data analysis, FlowJo version 10 (Tree Star) and SPICE (National Institute of Allergy and Infectious Diseases) software was used to create Boolean gate arrays that allowed us to determine the frequency of 32 possible response patterns based on the five cytokines tested.
PBMCs from CMV-positive donors were resuspended in RPMI media supplemented with 2% human AB serum (Corning, Celgro), and labeled with 5 μM CFSE (Molecular Probes, Life Technologies). CFSE-labeled PBMCs were inoculated with the recombinant strains of EBOV at an MOI of 2 PFU/cell, with or without simultaneous stimulation with 2 μg/ml of CMV pp65 peptides, SEB treated, or mock treated for 4 hours. PBMCs were washed twice to remove virus inoculum and cultured at a concentration of 2 x 106 cells/ml for 7 days in Advanced RPMI medium supplemented with 10% human AB serum, 2 mM L-glutamine, 200 IU/ml penicillin, and 200 μg/ml streptomycin sulfate (Invitrogen) in 6 well plates. PBMCs were harvested, stained extracellularly with the following antibodies: anti-CD3 PE-Cy7 (clone SK7, BD Biosciences), anti-CD4-PerCP/Cy5.5 (clone SK3, BD Biosciences), anti-CD8 APC-Cy7 (clone SK1, BD Biosciences), and Live/Dead-Far Red Dead cell stain (Molecular Probes, Life Technologies). Cells were subsequently washed, fixed with 4% paraformaldehyde and removed from the BSL-4 according to an approved protocol. Cells were washed with PBS and analyzed by flow cytometry on BD LSR II.
PMBCs were pulsed with CMV peptides as previously described. Following 24–48 hours, CD137+ T-lymphocytes were isolated using magnetic bead separation in accordance with the manufacturer’s protocol (Miltenyi). Cells were expanded for 8 days in complete RPMI (Gibco)/10% Human serum (Cellgro) media supplemented with recombinant IL-2 (10 ng/ml) and IL-7 (10 ng/ml) (R&D Systems) and Dynabeads Human T-Activator CD3/CD28/CD137 beads (ThermoFisher Scientific) to promote the expansion of CMV-specific responders. DCs were infected at an MOI of 3 and pulsed overnight with CMV peptides. Expanded responders were cultured with 2x105 autologous DCs at a 1:1 ratio for an additional 24 hours. Intracellular staining was performed as described previously.
Frozen CD4+ T-cells were recovered at 37°C for 2 hours prior to co-culture with autologous DCs infected overnight with wt or mutant EBOVs (MOI of 3) at a 1:1 ratio. CD4+ T-cells were harvested after 4 days of co-culture, and analyzed by Western Blotting. Briefly, cell pellets were lysed in Laemmli lysis buffer (Invitrogen), separated on 4–12% SDS-PAGE gradient gels (ThermoFisher Scientific) and transferred to nitrocellulose membranes using the I-blot system (ThermoFisher Scientific). Thereafter blots were incubated with primary antibodies provided in the T-cell Signaling Antibody Sampler and the Pro-Survival Bcl-2 Family Antibody Sampler Kits (Cell Signaling Technologies). GAPDH was used as an internal loading control (Cell Signaling Technologies). HRP-conjugated secondary antibodies (Santa Cruz Biotech) were used for chemiluminescent detection on Hyperfilm (Amersham).
For analysis of cytokines and chemokines in CMV-responder—DC co-culture assays, cell-free supernatants were irradiated at 5 Mrad, stored at -80°C and shipped to Eve Technologies on dry ice. Undiluted samples were analyzed using Human Cytokine/Chemokine Array 65-Plex Panel. Heatmaps of values normalized to mock-infected samples were generated using GENE-E software (Broad Institute). In some experiments, conditioned media were collected from DC-T-cell co-culture plates, clarified of cellular debris by low-speed centrifugation and stored at -80°C. Following overnight recovery, CD4+ T-cells were plated in 96-well plates in 100 μl of RPMI1640 medium supplemented by 10% FBS, followed by the addition of 150 μl of conditioned media. Cells were incubated overnight followed by the addition of brefeldin A and monensin for an additional 5 hours. Cells were then stained as previously described.
In experiments requiring IFNAR2 blockade, DC were pre-incubated with 30 μg/ml of blocking antibody specific to IFNAR subunit 2 (CD118; PBL Assay Science) or mouse immunoglobulin G2a (IgG2a; R&D Systems) as an isotype control (the endotoxin levels in the two antibody preparations were <1 and 0.1 endotoxin units per 1 μg of antibody, respectively), for 1 hour before virus inoculation. In the experiments involving exogenous IFNs, recombinant human IFNα2a or IFNβ1a (PBL Assay Science) was added to DCs at 24 hours after infection at a final concentration of 1,000 IU/ml or 800 IU/ml, respectively, and after additional 20 hours, cells were analyzed by flow cytometry. DC were analyzed for cell surface expression of markers of maturation by flow cytometry at 40 hours post-infection. Most DCs were collected by pipetting, and the remaining cells, which were attached to the bottom of the plates, were collected by applying staining buffer (PBS containing 2% fetal bovine serum and 2 μM EDTA). Cells were pelleted by centrifugation at 200×g at 4°C for 5 minutes, buffer was removed, and the pellet was re-suspended in staining buffer. DCs were incubated for 20 min on ice in the dark with the monoclonal antibodies anti-CD86-PerCP-Cy5.5 (clone FUN-1), anti-CD80-APC-H7 (clone L307.4), and anti-CD54-PE (clone HA58). In addition, IgG2a-PerCp-Cy5.5, IgG1-APC-H7, and IgG1-PE were used as the respective isotype control antibodies (all from BD Biosciences). Following incubation, cells were washed three times with the staining buffer and re-suspended in 200 μl of the same buffer. The Far Red fluorescent dye (Invitrogen) was used to evaluate cell viability by flow cytometry. Data were acquired using a flow cytometer FACSCanto II in BSL-4 facility or fixed with formalin as described above, taken out of BSL-4 and analyzed with BD LSR flow cytometer. Data were analyzed using FlowJo 7.6.1 software (Tree Star).
Suspensions of DCs and lymphocytes were fixed with 4% paraformaldehyde for 15 minutes and washed 3 times with PBS. Cells were then pelleted by centrifugation at 600xg for 6 minutes and resuspended in 50 μL of PBS. Cells were loaded on charged slides and dried overnight at 4°C. For staining, cells were rehydrated with PBS and permeabilized with 0.5% Triton X100 solution (ThermoFisher Scientific) in PBS for 15 minutes, washed with PBS, incubated with 0.5 M glycine in PBS for 10 minutes at room temperature, and washed 3 times with PBS. Antigen blocking was performed using 5% donkey serum (Sigma Aldrich) diluted in stain buffer (1% BSA and 0.1% Triton X100 in PBS) for 30 minutes. Mouse monoclonal antibody targeting HLA-DR (ThermoFisher Scientific, clone 7.3.19.1, 1:10 dilution) and goat polyclonal targeting CD3ε (Santa Cruz, 1:50 dilution) were used as primary antibodies and diluted in stain buffer as indicated. After 1 hour incubation at 37°C, slides were washed 3 times in washing buffer (0.1% Triton X100 in PBS), incubated with a mixture of two secondary antibodies: donkey anti-mouse conjugated with Alexa Fluor 647 (ThermoFisher Scientific) and donkey anti-goat conjugated with Alexa Fluor 594 (ThermoFisher Scientific) both diluted at 1:200 in stain buffer for 1 hour. Next, washed cells were incubated with 6-diamin-2-phenylindole dihydrochloride (DAPI) (ThermoFisher Scientific) at 1 μg/ml for 2 minutes, and washed 3 times in PBS. Slides were then fixed in 4% paraformaldehyde and removed from BSL-4. Slides were washed in 0.5 M glycine, washed in PBS, and mounted with coverslips using PermaFluor mounting medium (ThermoFisher Scientific). Infected cells were identified by expression of GFP encoded by the recombinant viruses. Slides were analyzed by laser scanning confocal microscopy using Olympus FV1000 confocal microscope. Lasers with 405 nm wavelengths were used for DAPI excitation, 488 nm for GFP, 543 for Alexa Fluor 594, and 635 nm for Alexa Fluor 647. This was followed by counting of immunological synapses, defined as colocalizations of CD3 and MHC-II (HLA-DR). For each experimental group, the data presented are based on count of at least 10 confocal imaging acquisitions. Results of the count were expressed as numbers of synapses per 100 cells.
Statistical analyses and generations of graphs were performed using GraphPad Prism version 6.05 (GraphPad Software). Statistical significances were calculated using a paired T-test. Statistical significance was set at p < 0.05.
|
10.1371/journal.pcbi.1003067 | Theta Coordinated Error-Driven Learning in the Hippocampus | The learning mechanism in the hippocampus has almost universally been assumed to be Hebbian in nature, where individual neurons in an engram join together with synaptic weight increases to support facilitated recall of memories later. However, it is also widely known that Hebbian learning mechanisms impose significant capacity constraints, and are generally less computationally powerful than learning mechanisms that take advantage of error signals. We show that the differential phase relationships of hippocampal subfields within the overall theta rhythm enable a powerful form of error-driven learning, which results in significantly greater capacity, as shown in computer simulations. In one phase of the theta cycle, the bidirectional connectivity between CA1 and entorhinal cortex can be trained in an error-driven fashion to learn to effectively encode the cortical inputs in a compact and sparse form over CA1. In a subsequent portion of the theta cycle, the system attempts to recall an existing memory, via the pathway from entorhinal cortex to CA3 and CA1. Finally the full theta cycle completes when a strong target encoding representation of the current input is imposed onto the CA1 via direct projections from entorhinal cortex. The difference between this target encoding and the attempted recall of the same representation on CA1 constitutes an error signal that can drive the learning of CA3 to CA1 synapses. This CA3 to CA1 pathway is critical for enabling full reinstatement of recalled hippocampal memories out in cortex. Taken together, these new learning dynamics enable a much more robust, high-capacity model of hippocampal learning than was available previously under the classical Hebbian model.
| We present a novel hippocampal model based on the oscillatory dynamics of the theta rhythm, which enables the network to learn much more efficiently than the Hebbian form of learning that is widely assumed in most models. Specifically, two pathways, Tri-Synaptic and Mono-Synaptic, alternate in strength during theta oscillations to provide an alternation of encoding vs. recall bias in area CA1. The difference between these two states and the unaltered cortical input representation creates an error signal, which can drive powerful error-driven learning in both Tri-Synaptic and Mono-Synaptic pathways. Furthermore, the presence of these alternating modes of network behavior (encoding vs. recall) provide an intriguing target for future work examining how prefrontal control mechanisms can manipulate the behavior of the hippocampus.
| Over the past half century the hippocampus has provided fertile ground for the work of mechanistic computational models to inform empirical research. From the earliest investigations into Long Term Potentiation to the complex dynamics of place cells, models of hippocampal function have enabled a greater understanding of how learning and memory emerges from more basic neural mechanisms in this remarkable brain area. The paradigmatic theoretical model guiding this work is the Hebb-Marr framework [1]–[3], which features the core idea that Hebbian learning wires together neurons that are firing together as part of a memory or engram representation, e.g., in the central area CA3 of the hippocampus. With these connections strengthened, the ability to pattern complete a partial memory cue to a full representation of the original memory is enhanced. For this pattern completion within CA3 to actually drive full memory recall, it must trigger a chain reaction of pattern completion throughout the cortex — although central to most theoretical accounts, the critical role of the CA1 in this larger pattern completion process has not been as widely recognized. Specifically, learning between CA3 and CA1 neurons must take place at memory encoding, to enable the CA1 to then drive entorhinal cortex (EC), which then drives the higher-level association cortex areas that are bidirectionally interconnected with it. This plasticity at the CA3 to CA1 synapses indeed may be the most important factor for subsequent memory recall [4]. It is the nature of this plasticity, and the learning that takes place in the bidirectional connections between EC and CA1, that is the focus of this paper.
We argue that, by taking into account the phase differences of firing for these areas within the overall theta cycle of the hippocampus [5], [6], a powerful error-driven form of learning emerges, which can result in much higher storage capacity than the standard Hebbian learning mechanism. Furthermore, these phase dynamics within the EC – CA1 bidirectional connections enable the CA1 to very naturally learn to be a sparse, invertible auto-encoder of the EC inputs, which has long been an important but somewhat implausibly implemented feature of our computational models [7]–[10]. Thus, this new model, which we refer to as the theta-phase hippocampus model, in reference to the theta oscillation, provides a more unified and computationally powerful model of hippocampal function. This model also enables us to make more direct contact with a large base of evidence, in both humans and rodents, relating hippocampal EEG oscillations to learning and memory. Much of the progress within this literature has been made in animal electrophysiology targeting hippocampal representation during spatial navigation and recall, while evidence from human EEG and intracranial recordings of oscillatory interactions also shows connections to episodic memory.
Modeling work, originally developed within the spatial navigation literature, suggested that connectivity between hippocampal subregions is coordinated via the 3 to 8 Hz EEG theta oscillation [5], [6]. This work has also been extended into a more general framework of hippocampal function including a proposed extension from spatial navigation into episodic memory [8], [11]. These investigations provide the foundation for the theta-phase model described in the current work, in terms of establishing the existence and functional role of the oscillatory coordination of hippocampal subregions within an encoding and retrieval dynamic. We build upon this foundation by showing how these dynamics can lead to error-driven learning, and a concomitant increase in overall storage capacity for the system. The implementation of this theta-phase model is based directly on the Complimentary Learning Systems neural network model of the hippocampus [7], [9], [10], which is implemented within the Local, Error-driven, and Associative, Biologically Realistic Algorithm (Leabra) framework [12], [13]. We assess the impact of the theta-phase error-driven learning mechanisms by comparing it with an otherwise identical model that uses a Hebbian learning rule, while varying the number of units within the Dentate Gyrus (DG) and area CA3, and measuring the models' recall on a varying number of learned patterns. These learned patterns are presented at test with 25 percent of the pattern missing, and the models are compared on their ability to complete the missing portion of the pattern. Results show the error-driven signal performs significantly better than the Hebbian learning rule.
The model used in the current work is built upon a series of structural and functional hypotheses based on anatomical and physiological data, which have been captured in the complementary learning systems (CLS) model of the hippocampus [7], [9]. The Entorhinal Cortex (EC) in the model is assumed to be the cortical gateway to the hippocampus. This gateway feeds through the trisynaptic pathway (TSP) to the Dentate Gyrus (DG), CA3, and then to CA1. Similarly, there is a parallel connection through the monosynaptic pathway (MSP) from the EC to the CA1 (and back) (Figure 1).
The TSP connections via the perforant path from EC to DG and CA3 are broadly diffuse, and support the conjunctive binding of various distributed pieces of information into an overall episodic memory representation in the CA3. The CA3 has sparse and highly separable patterns of activity (which are further pattern-separated via the very sparse DG layer), resulting in substantially reduced interference from synaptic weight changes, thus enabling rapid learning of novel episodic or conjunctive information [7]. To recall existing memories, the recurrent connections in CA3, along with plasticity in the EC to CA3, as well as the DG to CA3 connections, support pattern completion of missing information from retrieval cues.
For pattern completion in CA3 to have any effect on the rest of the brain, there must be a way to map the CA3 representation back out to the neocortex. This occurs via connections from CA3 to CA1 (the Schaffer collateral pathway), and then from CA1 back to EC, which then projects back out to the cortex to fill in the full memory representation in the cortical areas where it can actually be used in further cognitive processing. This Schaffer collateral pathway is a key focus of the theta-phase model, where we can train synapses in this pathway according to an error-driven learning signal, instead of the standard Hebbian signal assumed in other existing models.
The MSP between EC and CA1 is also essential for supporting memory retrieval, in a way that is often under-appreciated in the literature. This pathway is topologically organized, not diffuse, which we capture by organizing the simulated neurons in EC and CA1 into mutually interconnected slots, presumably encoding different separable elements across all the cortical areas that converge on the EC [14]. This slot architecture (Figure 1) enables the MSP to develop separable invertible pathways where a given EC input pattern can be encoded over a sparser representation in the corresponding CA1 slot, and this CA1 representation can in turn recover the full original EC slot pattern. The topographic nature of this CA1 representation is important for providing a mapping from cortex into the hippocampus and back out again. Weight adjustments along the TSP form conjunctive representations that bind information across the topography of EC and are important for recreating a previously experienced state from incomplete inputs (i.e., pattern completion). The Schaffer collaterals (the connection between CA3 and CA1) provide the translation between these two types of representations, allowing the conjunctive representations learned in the TSP to influence the topographic representations within CA1, and subsequently back out to EC. In our previous CLS models, we have trained these topographic slot mapping weights between EC and CA1 in an offline manner prior to training the full hippocampal network. The new theta-phase learning mechanisms now enables us to train this important MSP pathway in a very natural manner, at the same time as the rest of the hippocampal system learns.
To summarize, after learning, the model recollects studied items by reactivating the original patterns via the trained weighted connections between areas. The accuracy of this recall is scored as a simple comparison between the originally studied pattern and the recollected pattern. If the input pattern corresponds to a non-studied pattern, or even if individual components of the pattern were previously studied, but not together, the conjunctive nature of the CA3 representations will minimize the extent to which recall occurs. Conversely, when previously studied patterns are presented in an incomplete or noisy input format, these weights allow the hippocampus to recall the originally studied pattern.
As noted previously, the original Complementary Learning Systems (CLS) hippocampal model pretrained the invertible mapping between EC and CA1 on a vocabulary of possible patterns for a single slot [9]. The resulting weights for the connections within this individual slot network were then replicated across all EC–CA1 slots (see Figure 1 where an individual slot is highlighted) in the MSP. This restricts the space of inputs possible to the vocabulary of patterns in which the slot network was trained.
The alternative approach adopted in this work utilizes simultaneous, independent learning along both the MSP and the TSP. This dual-pathway learning is motivated by physiological recordings within the subfields of rat hippocampi, along with mathematical models of hippocampal function [6], in terms of the 3–8 Hz oscillatory EEG signal known as theta. The theta oscillation can be found throughout the hippocampus and surrounding cortex, however it is strongest and most consistent when recorded within the region separating CA1 and DG known as the hippocampal fissure. For this reason all references to theta oscillations will be referring to the EEG signal measured at the hippocampal fissure.
Figure 2A shows an illustration of hippocampal subfield dynamics in relation to the fissure recorded theta oscillation shown in red. This cartoon, derived from current source density analysis [6], [15], shows the current sinks into area CA1 alternatively originating from either area CA3 in blue or EC layers II and III in green. At the trough of fissure recorded theta, EC sources into CA1 are at their peak and area CA3 is at its minimum. This implies that EC has a strong influence over synaptic potentials within area CA1 at this time. At the peak of fissure recorded theta, CA3 sources are at their peak and EC influence has diminished. This again suggests that CA3 input to area CA1 is now the dominant influence, and EC is less so as compared to the trough of the theta oscillation.
These dynamics are modeled within the neural network as, for any given input pattern, three distinct time points of activation: Theta Trough (TT), Theta Peak (TP), and Theta Plus (+) as shown in Figure 2B. These three time points are modeled as three independent settling processes across simulated neurons within the differential equation described in eq. (1). The patterns of activation that arise from these three time points are used to train the weighted connections along the MSP and the TSP, where the equations for the error-driven weight changes at these synapses are shown in eqs. (6) and (7).
Specifically, input patterns are projected onto which is then allowed to project to CA1 and subsequently to area , while CA3 input to CA1 is inhibited. This creates a pattern of activation dominated by the MSP which is then used to drive learning within these connections. This is denoted as superscript TT in eq. (6) for “Theta Trough”, as this time point is analogous to the connectivity dynamics at the trough of theta oscillations, where EC strongly influences CA1, and CA3 influence is relatively low. Following this, CA3 input onto CA1 is released from inhibition while the influence from onto CA1 is diminished. This corresponds to the Theta Peak (denoted as TP in eq. (7)); a time point that reflects strong influence from CA3 onto CA1. This time point is analogous to the peak of the fissure recorded theta oscillation where EC input to CA1 is weak, while CA3 input is strong.
The final plus stage of activation (denoted with the + in eqs. (6) and (7)) corresponds to projecting onto and area CA1, and projecting back onto CA1. The representations within and will remain relatively static due to the direct connection between them, which then forces CA1 to settle into a representation that respects this symmetric mapping between and . This provides the veridical ground truth in the error-driven learning signal. In reference to eq. (3), this pattern of activation is used for the plus stage learning signal in contrast to the MSP's TT and the TSP's TP minus stage.
The alteration of these connections' strength are manipulated in the model by simply denying information flow through specific subregion projections at select points in the settling process of the differential equation shown in eq. (1). The three particular projections that are manipulated in the model are , and where the pattern of manipulation that these projections are subjected to are highlighted in Table 1. All other connections within the network have no error-driven component to their weight adjustments, only Hebbian, as seen in eq. (8).
The validation process adopted in this work is to compare the theta-phase learning model described above with a simple Hebbian learning model. The critical connections that utilize an error-driven learning signal within the theta-phase model are the Mono-Synaptic Pathway (), as well as the Schaffer collaterals (). In contrast, these connections in the comparison model use a purely Hebbian learning rule. The task run across both models is a simple capacity test such that each model is trained for 15 repetitions of an input pattern set (referred to as 15 epochs), and the performance of the two models is then tested by measuring the accuracy of the recalled patterns of activation given an input cue which has 25 percent of the trained pattern missing.
We explored three training regimes to contrast error-driven vs. Hebbian learning. First, both the MSP and TSP utilized an error-driven learning signal and was compared to a full Hebbian network. We then compared the contribution of these two pathways by using error-driven learning within either the MSP and not the TSP, or conversely within the TSP and not the MSP. Finally, to better compare against earlier models where the MSP pathway was pretrained in advance, we compared pretrained vs. non-pretrained MSP. In the pretrained MSP, only the MSP pathway was trained for 15 epochs (on the same patterns used for the overall training), followed by integrated training of both TSP and MSP as described above. In the non-pretrained MSP, both pathways were trained in the integrated fashion from the start.
The question of how network performance scales is addressed by varying the training set size, and network size across multiple levels of these two variables. The size of the input pattern set is varied from 40 to 800 patterns to get a measure of model performance across small and large training sets, with the assumption that better performance on larger training sets is more reflective of hippocampal function. Similarly, the size of the network itself was varied by increasing the number of units within the CA3 and DG layers, while holding a constant ratio between them. This is done to try and maintain a connection to the original biological constraints of the hippocampal circuit, and for this reason a ratio of 5 DG units to 1 CA3 unit was adopted, as this generally reflects the ratio in the human hippocampus [14],[16]. Maintaining this ratio, the total units within CA3 were varied from 10 to 100 units, which in turn corresponds to a varying of DG units from 50 to 500.
Finally, input patterns were constructed, and memory retrieval performance measured, based on the slot topology in the EC layers (as highlighted in Figure 1). This slot structure is intended to capture the modality segregation within EC, and within each slot we assume there is a vocabulary of different patterns, which reflect the representational repertoire within those modalities. We generated a vocabulary of 100 distributed activity patterns, with a minimum hamming distance of 10 between each vocabulary pattern generated. A complete input pattern used in the model validation process was then constructed by selecting a single pattern from these 100 vocabulary patterns for each of the EC slots. With 8 slots, a total of 1008 (npatterns raised to the nslots power) unique patterns are possible, however only 800 were used in the testing of these models. These vocabulary patterns were similarly used to estimate error within the networks' output by comparing, within a given slot, the output pattern of activation with all other vocabulary patterns. If, for the given input pattern, the slots' output at the layer is closest to the vocabulary pattern it was trained on, it is considered correct, and otherwise considered incorrect. This closest-pattern calculation is done across each of the slots for every input pattern, and if any slot shows an incorrect response the network output for that input pattern is counted as incorrect. This measurement is referred to as Name Error in the results section, and is thought to better represent the potential for clean up of hippocampal output as compared to more standard measures such as Sum Squared Error (SSE). It also has the advantage of not requiring any further threshold or other parameterization. It should be noted that this measure of error, compared to a SSE, deemphasizes single unit based errors in output in favor of an emphasis on distributed patterns of error across groups of units.
The model is implemented in the Leabra framework which uses a combination of supervised and Hebbian learning [13]. What follows is a coarse description of the essential components within this framework necessary for understanding the current work. The activation function for a given unit is a threshold based neuronal model with continuous valued spike rate as output. Each neuron's membrane potential () is updated using the following differential equation:(1)Here, 3 channels () summed across in the membrane potential calculation are: excitatory input, inhibitory input, and leak current. Excitatory input is calculated as the average over all weighted inputs coming into a unit (), where is the activity of sending unit and is the weighted connection between sending unit and receiving unit . All principal weights between units are excitatory while local circuit inhibition controls positive feedback loops. Leabra assumes a winner take all dynamic through a set-point inhibitory current (), producing a kWTA (k-Winners-Take-All) dynamic. kWTA is computed via a uniform level of inhibitory current for all units within a layer. Finally leak current () is a constant value set to 0.1
Activation of communication ( for a given unit ) with other units is a thresholded function of membrane potential:(2)Here is the gain factor which is set to a constant value of 100, and is the firing threshold value which is set to a constant of 0.5 within a units dynamic range of 0 to 1.
The Leabra framework utilizes a biologically plausible error-driven learning algorithm which is equivalent to Contrastive Hebbian Learning (CHL) [17]. Leabra uses two stages of activation; the stage is the initial activation or expected output of the network, while the stage is the provided target output activation. The Leabra weight updating component between sending units () and receiving units () is thus calculated as:(3)The + and − superscripts represent the plus and minus phase components respectively. In addition to the error-driven learning of CHL, a pure form of Hebbian learning is also used. Here the weight change is calculated using only the target, or plus phase, activations(4)This learning rule can be seen as computing the expected value of the sending unit's activity conditional on the receiver's activity [13]. Finally these two learning rules are proportionally weighted () along with a learning rate parameter, , for the combined learning rule used in this work:(5)
The theta-phase learning approach uses the learning framework described above within a particular dual-pathway architecture. The target, or plus, component of the error signal (superscript ‘+’ in eqs. (6) and (7)) is activation acquired from the layer projected onto , and allowed to propagate back on to area CA1 which settles into a pattern of activation constrained by static representations in and . Similarly projects along the TSP providing a plus phase activation within DG and CA3, however projections from CA3 onto CA1 are inhibited. Error signals used in weight adjustment are then calculated by taking the difference between this plus phase activation and the two distinct time points within the theta cycle (peak and trough), yielding two distinct error signals. Specifically, the MSP connections are adjusted according to an error signal acquired from the difference between plus phase activation and activation patterns acquired during the trough of theta (superscript TT in eq. (6)). It is critical to remember in the trough of theta there is no influence on representations from area , while in the plus phase projects onto . The difference in CA1 activation patterns at the conclusions of these two phases allows for the calculation of an error signal that is used to adjust the weighted connections within the MSP. Similarly, the TSP connections are adjusted according to an error signal acquired from the difference between the plus phase activation and the activation during the peak of theta (superscript TP in eq. (7)). In the peak of theta CA3 has a strong influence on CA1, while in the plus phase CA1 is influenced solely by the MSP. This change in CA1 representations allows for a error signal tailored to best adjust the TSP connections to more closely match the stimulus driven representation of the plus phase activations. All other connections, within the network, i.e. to DG and CA3, DG to CA3, and recurrent connections within CA3, have no error-driven component to their weight adjustment (eq. (8)).(6)(7)(8)
Settling dynamics within the network are dictated by the temporal evolution of Equation 1. This dynamic process, within every unit, is allowed 30 time steps to settle into its equilibrium state for each of the three phases within the theta cycle, thus yielding a total of 90 time steps for each full theta oscillation. All activation values within the network are reset to 0 at the onset of theta trough but are allowed to be carried over from trough to peak and finally from peak to the plus phase without alteration. All manipulations of Hebbian vs. error-driven learning where done via the lmix parameter as shown in eq. (5). Values used to instantiate full Hebbian learning implies a lmix value of 1, while error-driven learning used a lmix value 0f 0.001. This implies that error-driven networks also used a very small amount of Hebbian weight adjustment which we believe is implicit in normal neural circuitry.
Figure 3 shows the comparison of various network configurations. In panel A the theta-phase network with error-driven learning in the MSP and TSP is compared with a fully Hebbian learning network across various network sizes and trained input pattern set sizes. Plots are shown as a function of network size, where the number of CA3 units are shown on the x-axis which implies that the number of DG units for that network are 5 times that of CA3. Training set size, shown on the y-axis refers to the number of patterns a given network was trained and tested on. Surface plots of the average Name Error, on the z-axis, across the full training set are shown on the left for both the theta-phase and the Hebbian network. Each cyan dot in the surface plots represents a measured data point where both network types were tested in the network-size by training-set-size space. Each data point is the average within network type across 5 random weight initializations. These points were then fit to a 3D surface for visualization. The difference between theta-phase and Hebbian surfaces is shown on the right. These differences are compared using a random bootstrap method where Name Error values are sampled with replacement from both network types into two groups and a distribution of difference values is calculated to produce a null hypothesis. Data points with p values less than 0.005 are shown in the difference plot with an asterisk.
Similarly, in panel B of Figure 3 a more fine grained follow up test of performance shows a network with error-driven learning in the TSP and Hebbian learning in the MSP (labeled TSP ErrDrv) compared against a network with Hebbian learning in the TSP and error-driven learning in the MSP. This secondary test attempts to evaluate the relative importance of error-driven learning within these two pathways on overall performance. These results are then further tested by comparing pretrained MSP connections to non-pretrained MSP connections. These results are not shown in a similar style as Figures 3A and B as these pretrained networks yielded results nearly identical to non-pretrained networks. Figure 3C shows performance plots from individual networks within these three comparisons; here the overlap in pretrained and non-pretrained results can be seen in comparison to the other two comparisons shown in Figures 3A and B.
Results, shown in panel A of Figure 3, of the comparison between network models shows that the error-driven learning, provided by the theta coordination of subfield influence on CA1, out-performs the purely Hebbian based learning network. Investigating this relationship further, in panel B of Figure 3 it is shown that the crucial connection that leads to this benefit is between CA3 and CA1 along the TSP. There is little difference in performance when the TSP uses Hebbian learning (plot labeled MSP ErrDrv) compared to when the full network is exclusively using Hebbian learning. Conversely, when the TSP (specifically the connection between CA3 and CA1) takes advantage of the error-driven learning signal, performance is dramatically increased, and approaches Name Error levels achieved when both TSP and MSP are using error-driven learning signals (shown in Figure 3B and C). Contrasting performance from the full ThetaPhase network with the TSP error-driven network shows that there is indeed some performance benefit in the ThetaPhase network compared to TSP error-driven network, suggesting some synergy between the TSP and MSP over and above the benefit from the TSP error-driven network alone.
Our comparison of the effects of pretraining on the MSP, as was done in our earlier models, revealed very little difference as shown in Figure 3C. This is of considerable practical benefit, as it is often difficult to anticipate the full range of input pattern variability needed for pretraining, and it also increases the overall plausibility of our model, by eliminating any need for this extra step in the model.
These results provide insight as to how these learning signals compare across multiple network sizes and varying training set sizes. Looking at the difference in network performance we can see a divergence towards better error-driven performance as training size and network size increases. Many hippocampal models used within the literature test on relatively small training sets and with small network sizes; usually of the size required for the task or phenomena being modeled. Results from the current work suggest that Hebbian performance may not scale with these dimensions as expected, and that a more robust learning signal such as that provided by error-driven learning may be necessary to provide realistic performance in more ecologically valid network sizes and training set sizes. Given the significant performance advantages of the error-driven learning mechanism, and its biological support in the theta-phase coordination process, it would be surprising if the biological hippocampus did not also leverage this form of error-driven learning. In sum, we argue that this model represents a significant advantage over the existing Hebbian-based models of hippocampal learning, and can provide a predictive framework for future empirical studies.
The idea of temporal differentiation between Mono-Synaptic and the Tri-Synaptic pathways along the theta wave, as shown in previous hippocampal modeling work [5], [6], provides a well founded framework for how theta oscillations interact with behavior. The key contribution of this work to these models is a demonstration that the invertible mapping in and out of area CA1 along the Mono-Synaptic Pathway can be learned in tandem with the connections along the Tri-Synaptic Pathway, and that these oscillatory dynamics enable a form of powerful error-driven learning. Further, these results suggest that error-driven learning in the Schaffer collaterals connecting CA3 to CA1 are a crucial component in stabilizing this invertible mapping in the Mono-Synaptic Pathway, and providing the performance advantage shown in Figure 3.
The mapping of distributed representations into and out of area CA1 is a problem that has not been adequately addressed in previous models. Many models have used a simplified symmetric representation between hippocampal subregions [5], [6]. This allows for a transparent interpretation of subregion processing, however it reduces the ecological validity of the model's processing. An early model of episodic memory allowed for learning within this invertible mapping between EC and CA1, however the representations used were relatively small and simplified [8]. The current work shows that error-driven learning is a key component behind the requirement of relatively complex representational transformation between subregions. The attempt to match the hippocampal architecture and representational complexity within this work provides insight into these more subtle issues that are often assumed in other models of the hippocampus. The simulations done in this work show that the representational transformation into and out of the hippocampus is a non-negligible problem, and that more robust learning signals than the standard Hebbian model are required for accurate recall within large training data and small network sizes.
The current model provides a simplified version of oscillatory processes within a discretized time frame, as compared to previous models [5], [6]. The peak and trough time points being modeled in the current work can be thought of as stimulus driven at the trough of theta, and recall driven at the peak of theta [6], however these processes are implemented within the model as two relatively discontinuous patterns of activation that get integrated together when calculating the weight changes in the learning algorithm. Additionally, the plus phase of activation, i.e. the ground truth within the error calculation, is proposed as a projection of the superficial EC layers onto the deep layers of the EC. Computationally within the model this is implemented after both the trough and peak of the theta oscillation have completed, however we conceptualize the theta cycle to begin on the trough of the oscillation where the MSP is strongly active, and we therefore speculate that this plus phase projection would occur within the descending theta cycle following the peak but just before the trough. In Figure 2 we show the plus phase to occur at the trough of theta, however the model predicts that the plus phase would occur anytime between theta peak and theta trough. In some sense the plus phase is a transition from theta peak to theta trough where the onset of the plus phase is marked by the inhibition of the TSP and a projection from the superficial layers of EC to the deep layers. This allows for the error-driven contrasting of this plus phase pattern of activation with the preceding theta trough and theta peak patterns. Indeed, laminar recordings from Entorhinal Cortex support this theta phase reversal in deep layers compared to superficial [18], and a recent investigation into the microcircuits within EC layers supports the increased firing from superficial EC to deep EC just preceding the trough of ongoing theta oscillations [19]. Future electrophysiological work could test these temporal dynamics further by stimulating at these various stages of the theta wave to try and disrupt or enhance this theoretical cascade of activation.
Previous models have labeled activation patterns associated with theta peak and trough as Encoding during the trough, and Retrieval during the peak, which our model also captures [5], [6], [20]. This separation of functionality between the two pathways might allow for other systems to interact with the nominal theta cycle to influence these processes and thereby bias the hippocampus towards one process over the other. A growing base of empirical evidence within the rodent literature suggests that oscillatory coherence within the theta band between frontal regions and the hippocampus is correlated with successful retrieval [21]–[23]. In humans these interactions could provide the framework for some form of volitional control over either encoding or retrieval. Future empirical work in humans could probe this relationship between encoding and retrieval within the hippocampus as well as its interaction with other systems. The current model would suggest that disruption of the theta oscillation during the trough of theta would alter the encoding of new experiences, while disruption at the peak of theta would alter the retrieval of previous experiences.
The question of how incoming stimuli align to these phase dynamics is somewhat unclear, however constraints from previous empirical work do exist. There is evidence suggesting that theta oscillations show a phase resetting approximately 200 ms after stimulus onset [24]–[26]. The entry point into the theta wave on these phase resets, however, show a difference in study vs. test items where test items enter on the descending wave of theta while study items enter on the ascending wave. Our model suggests that there would be a plus phase following the descending theta wave, and would be evident through the projection from the superficial layers of EC to the deep layers. This task dependent phase reset could help to target this plus phase dynamic, and potentially determine whether it is more associated with start of a given theta oscillation or with the end.
There are many limitations within the current work in regards to the scope of biological components, and we do not mean to suggest that this model accurately reflects all aspects of hippocampal function. For example, the discrete nature of the two time points modeled, i.e. trough and peak, within the theta cycle could be better approximated by having a continuous change of activation after the plus phase. The current work simplifies the more continuous change of activation at the end of a Theta cycle by resetting activation after the plus phase. Additionally, there are hippocampal subfields, in particular the Subiculum [27], which are not included within this model. We are currently exploring the addition of a Subiculum layer within our model which modulates the learning rate of connections into CA1. The Subiculum-mediated modulation focuses on increasing the learning rate for novel stimuli, and reducing the learning rate for well learned Tri-Synaptic Pathway (TSP) representations, theoretically allowing for the reduction of interference in the otherwise purely Hebbian learning in the TSP (e.g., in perforant pathway projections from EC to CA3). Although no current explorations are underway, area CA2 could also provide an augmentation to our model of the MSP [28]. This area would fit in as a intermediary between the and CA1, providing a non-topographic representation across the slots of Entorhinal Cortex, and potentially increasing the learning capacity along this pathway.
In conclusion, within the subfields modeled, we have accurately represented the known connectivity and topology using a biologically motivated neural network framework. Further, we have included coordination between those subfields through the currently understood inhibitory processes as modulated by theta oscillations. Building upon this framework in future projects can provide a strong foundation in the known biological constraints, and representational complexity of the hippocampal circuit.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.